Artificial Intelligence for Transforming Indigenous Malaria Knowledge into Digital Health Intelligence in Uganda

Artificial Intelligence for Transforming Indigenous Malaria Knowledge into Digital Health Intelligence in Uganda | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Artificial Intelligence for Transforming Indigenous Malaria Knowledge into Digital Health Intelligence in Uganda Emmanuel Ahishakiye, Nina Olivia Rugambwa This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8337646/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Indigenous malaria knowledge remains foundational to health decision-making in Uganda, yet it remains under-represented in digital health systems. This study investigates whether culturally embedded malaria knowledge can be transformed into interpretable digital intelligence using artificial intelligence (AI). An indigenous malaria corpus was constructed from documented ethnobotanical sources, community behaviour reports, and cultural health narratives drawn from 11 Ugandan communities, comprising 38 medicinal plant profiles, 17 non-plant prevention practices, and 14 explanatory belief pathways. The dataset was encoded into 812 prompt–response pairs and used to fine-tune GPT-NeoX-1.3B through low-rank adaptation and retrieval-augmented conditioning. Model evaluation indicated promising performance: precision@3 of 81%, recall@5 of 77%, and cultural authenticity ratings averaging 4.3/5 from traditional medicine reviewers. Safety reliability was high, with escalation behaviours achieving 4.7/5 compliance, while hallucination frequency reduced from 28% in the base model to 9% post-training. Cross-community reasoning achieved 69% generalisation, strongest in prevention logic (78%) but comparatively weaker in herb specificity (65%). The model showed moderate paraphrasing diversity but limited bilingual fluency, particularly under Luganda reformulations. Computational feasibility testing revealed central processing unit (CPU) inference latency below 2.5 seconds and stable performance under int8 quantisation, suggesting suitability for use on low-resource devices typical of rural health settings. These results demonstrate that Ugandan indigenous malaria knowledge can be computationally represented, internalised, and operationalised through AI without compromising cultural tone or safety. The work provides a proof-of-concept for culturally grounded digital health intelligence and identifies pathways for scale-up via community co-creation, indigenous language expansion, and applied field prototyping. Indigenous Knowledge Systems Artificial Intelligence for Health Malaria Digital Intelligence Retrieval-Augmented Language Models Culturally Grounded Decision Support Figures Figure 1 Figure 2 1. Introduction Malaria remains one of the most persistent public health challenges in sub-Saharan Africa, responsible for high morbidity and mortality particularly among children under five and pregnant women [ 1 ]. Uganda is among the top six highest burden countries globally, accounting for an estimated 5% of global malaria cases and deaths according to the World Health Organization’s latest African region analysis [ 1 ]. While biomedical tools such as rapid diagnostic testing and artemisinin-based combination therapy have improved outcomes, barriers in access, affordability, and health-seeking behaviour continue to sustain transmission risks [ 2 ]. For centuries, communities in Uganda and other African settings have relied on indigenous malaria knowledge systems consisting of medicinal plants, household prevention practices, and culturally embedded pathways for symptom interpretation and treatment escalation [ 3 ]. Documented ethnobotanical research confirms widespread use of species such as Vernonia amygdalina , Azadirachta indica , Moringa oleifera , Warburgia ugandensis , and Aloe vera , as well as practices involving herbal fumigation, dietary modification, or consultation with elders before health facility referral [ 4 ], [ 5 ], [ 6 ]. These knowledge systems are not merely symbolic traditions but contain experiential reasoning grounded in ecological observation, community learning, and inter-generational transmission [ 7 ]. However, despite their prevalence and public health relevance, indigenous knowledge remains poorly represented in formal decision-support systems, resulting in a disconnect between cultural malaria reasoning and digital health infrastructure [ 8 ]. The rise of artificial intelligence–driven natural language systems offers potential to bridge this divide by enabling computational representation of indigenous health reasoning [ 9 ]. Recent studies on language models in low-resource settings show that contextual grounding, retrieval augmentation, and safety alignment can improve health-communication utility and reduce hallucination behaviour [ 10 ]. Yet most digital malaria interventions adopt biomedical framings with limited integration of community epistemologies, reinforcing epistemic exclusion and reducing acceptability of innovation [ 11 ]. As argued within African health informatics scholarship, culturally aligned technologies improve user trust, interpretation accuracy, and intervention uptake [ 12 ]. This paper responds to this gap by investigating whether indigenous malaria knowledge in Uganda can be digitally internalised, expressed, and evaluated using artificial intelligence, without compromising cultural tone or safety. We constructed an indigenous malaria dataset from documented sources across 11 Ugandan communities and fine-tuned a small language model to reason over herbs, prevention practices, belief pathways, and escalation behaviours. Model performance was assessed using quantitative accuracy metrics, narrative authenticity, safety compliance, and computational feasibility, contributing empirical insight into how indigenous knowledge can be computationally transformed into actionable digital health intelligence. This study demonstrates that indigenous malaria knowledge in Uganda can be computationally modelled and transformed into digital health intelligence using artificial intelligence. The results show that language models grounded in culturally verified knowledge achieve high accuracy, narrative authenticity, and safety compliance, suggesting that indigenous reasoning systems can be preserved, enhanced and embedded into equitable digital health innovations. The findings contribute to scholarly discussions in three intersecting domains: (i) preservation and operationalisation of indigenous knowledge; (ii) culturally grounded artificial intelligence for health communication; and (iii) equitable digital health design for low-resource settings. Therefore, this work demonstrates feasibility, limitations, and learning behaviour of such systems, this work advances arguments for inclusive AI design where local knowledge systems are part of, not peripheral to, global digital health intelligence. 1.2 Contributions and Organization of the Paper This paper makes both conceptual and empirical contributions at the intersection of indigenous health knowledge and artificial intelligence research. First, it develops a structured indigenous malaria knowledge corpus, synthesising medicinal plant repertoires, preventive behaviours, and explanatory narratives from 11 Ugandan communities. Unlike previous work that documents such knowledge descriptively, this study formalises it into machine-usable representations, opening new pathways for preservation and computational analysis. Second, the paper presents a fine-tuned small language model grounded in indigenous knowledge, demonstrating that culturally embedded reasoning can be internalised and expressed by artificial intelligence. This is one of the earliest empirical attempts to treat indigenous malaria knowledge as a primary training source rather than peripheral context. The resulting model delivers promising performance outcomes, including 81% factual retrieval accuracy, 4.3/5 cultural authenticity ratings, and 4.7/5 safety compliance, indicating a viable route for culturally aligned digital health intelligence. Third, the study contributes a multi-dimensional evaluation framework combining quantitative metrics, expert judgement, and behaviour analysis to assess accuracy, authenticity, hallucination control, paraphrasing capability, and cross-community generalisation. This framework provides a methodological reference point for future research seeking to operationalise community-based knowledge in artificial intelligence. Fourth, the paper offers insights into behavioural limitations and failure modes, including locality loss, bilingual phrasing drift, and template-dominant escalation messaging. These observations highlight the need for iterative co-design, multilingual expansion, and community validation before deployment, offering practical guidance for scholars and practitioners. Finally, the work demonstrates computational feasibility for low-resource deployment, showing that indigenous-knowledge-aware models can operate under < 2.5-second latency on modest hardware and remain functional under quantisation. This contributes to equity-focused digital innovation literature by showing that culturally grounded artificial intelligence can be realistically embedded in rural health ecosystems. The remainder of this paper is organised as follows. Section 2 reviews related literature on indigenous malaria knowledge, digital health gaps in Africa, and retrieval-augmented language models. Section 3 describes the materials and methods, including dataset construction, model development, ethical safeguards, and evaluation procedures. Section 4 presents experimental results on accuracy, cultural authenticity, generalisation behaviour, and computational feasibility. Section 5 discusses these findings in relation to existing scholarship, while Section 6 concludes the paper and outlines directions for future research and deployment. 2. Related Literature 2.1 Indigenous Malaria Knowledge and Ethnomedicinal Practice in Uganda Malaria management in Ugandan communities has long relied on rich indigenous knowledge systems characterised by herbal treatment, household-level prevention actions, and culturally embedded symptom interpretation. Ethnobotanical research widely recognises that communities maintain extensive medicinal plant repertoires and diagnostic reasoning that supplement biomedical pathways [ 3 ], [ 5 ]. Bunalema et al. conducted a multi-district survey across Apac, Arua, Tororo and Kabale and identified 97 medicinal plant species from 45 botanical families used by traditional healers for malaria symptoms, with Vernonia amygdalina and Artemisia annua among the most frequently reported [ 3 ]. Tabuti et al. further documented 45 antimalarial plant species in Tororo District, revealing consistent reliance on decoctions and leaf-based preparations [ 5 ]. Similar studies across Nyakayojo, Rukungiri, Budondo, Cegere and northern Uganda confirm that indigenous malaria treatment knowledge is geographically widespread, multi-layered and orally transmitted intergenerationally [ 4 ], [ 13 ]. These studies demonstrate that medicinal plants, knowledge of disease categories (e.g., “strong malaria”), and treatment sequencing are preserved within community institutions and belief systems, despite limited formal digitisation. 2.2 Indigenous Knowledge, Digital Health and African Health Equity While indigenous knowledge remains central to malaria care in Uganda, it is rarely integrated into formal digital health infrastructures. African digital health strategies often prioritise telemedicine, disease surveillance, and electronic medical records while neglecting local epistemologies [ 14 ]. Ethnographic critiques argue that externally designed digital tools frequently fail to align with indigenous worldviews, contributing to low adoption and trust [ 15 ]. Meanwhile, discussions around integrating traditional medicine into national health systems emphasise opportunities for knowledge preservation, collaboration with biomedical systems, and digital documentation, but also warn against risks of decontextualisation and knowledge extraction without community benefit [ 16 ]. Scholars argue that digital innovation must meaningfully engage indigenous knowledge-holders to enhance system legitimacy and health outcomes [ 15 ], [ 17 ]. However, practical examples of culturally grounded digital systems remain limited. Traditional medicine is typically acknowledged descriptively rather than encoded as a computational reasoning framework that can drive decision support [ 18 ]. 2.3 Artificial Intelligence, Language Models and Knowledge Grounding in Health Recent advances in artificial intelligence, particularly language models, have expanded possibilities for automated reasoning in healthcare, including decision-support and health communication [ 19 ]. However, generic models often hallucinate or reproduce biases when applied to specialised domains [ 20 ]. Retrieval-augmented generation approaches have emerged as a compelling solution to constrain language models with domain-grounded evidence, improving factual accuracy and alignment [ 10 ], [ 21 ]. Systematic reviews of retrieval augmentation in health applications report improved clinical answer reliability, reduced hallucination rates, and enhanced transparency when models are conditioned with structured knowledge sources [ 10 ], [ 21 ], [ 22 ]. Yet these implementations predominantly rely on biomedical corpora from high-income healthcare systems. There is minimal research applying language models to indigenous knowledge corpora or evaluating whether computational systems can faithfully reproduce culturally embedded practices such as herbal sequencing, preventive logic, or escalation pathways within community malaria reasoning. 2.4 Identified Gap Across these bodies of literature, three insights emerge. First, indigenous malaria knowledge in Uganda is empirically rich, geographically distinct, and governed by structured explanatory models, yet remains largely oral, under-documented, and undigitised [ 3 ], [ 5 ]. Second, while African health technology research emphasises contextualisation and co-creation, operational integration of indigenous reasoning into digital systems is largely absent [ 23 ]. Third, although retrieval-augmented models demonstrate strong knowledge-grounding performance, there is no evidence assessing their capacity to internalise and reproduce indigenous knowledge systems, particularly in low-resource or multilingual settings [ 10 ], [ 21 ], [ 22 ]. This study responds directly to these gaps by (i) structuring and encoding indigenous malaria knowledge from 11 Ugandan communities, (ii) fine-tuning a small language model grounded through retrieval augmentation, and (iii) evaluating its accuracy, narrative authenticity, safety behaviour, generalisation capacity and computational feasibility. The work contributes to emerging scholarship exploring culturally aligned artificial intelligence and the digital preservation of indigenous health intelligence. 3. Materials and Methods 3.1 Study Design This study employed an integrative methodological design situated within a design science research paradigm. The central aim was to explore whether indigenous malaria knowledge could be transformed into structured digital intelligence through artificial intelligence. The study therefore, combined elements of knowledge documentation, computational representation, AI tuning, and performance evaluation. Instead of treating indigenous knowledge as anecdotal belief, the study approached it as a structured human reasoning system capable of being encoded, modelled, and tested. The research process unfolded across four coordinated strands. First, indigenous malaria knowledge was compiled from documented ethnomedicinal studies, community health traditions, and grey literature sources describing plant-based remedies, preventive practices, and explanatory narratives. Second, this knowledge was abstracted into computationally usable forms through systematic coding, context tagging, and development of prompt–response pairs. Third, a small language model was fine-tuned using instruction-based training supported by a retrieval mechanism adapted to surface relevant indigenous knowledge during inference. Finally, the trained system was evaluated using a combination of quantitative metrics, such as retrieval accuracy, cultural authenticity scores, safety compliance rates, and cross-community generalization and qualitative review to assess narrative coherence and escalation behaviour. Although this stage of the research did not directly involve human subject interaction, the content validity of model outputs was strengthened through expert review from two traditional health practitioners and one community health worker, who assessed selected responses for authenticity and safety. Figure 1 summarises the iterative workflow applied in this study, beginning with knowledge collection and ending with model interpretation and refinement. The process demonstrates a continuous translation pathway from community-held knowledge to computational representation and finally to evaluative analysis. 3.2 Indigenous Knowledge Dataset Construction The indigenous malaria knowledge dataset was constructed through systematic extraction, coding, and consolidation of documented cultural health practices. This process drew evidence primarily from ethnobotanical studies, community malaria behaviour publications, traditional medicine reports, and grey literature describing household-level treatment pathways. Sources were selected based on relevance to malaria, Ugandan/community context, and explicit articulation of plant-based remedies or cultural decision logic. Extracted content included medicinal plant usage, preparation techniques, preventive practices, social reasoning patterns, and escalation behaviours. Coding was performed manually to retain semantic fidelity and prevent distortion of cultural meaning. Data entries were organised into four principal categories reflecting realistic malaria knowledge domains: (i) medicinal plant profiles, (ii) non-plant prevention and care practices, (iii) belief and explanatory reasoning structures, and (iv) contextual–response pairs reflecting realistic question and answer patterns, including safety messaging templates and escalation clauses. Context filters were embedded in the dataset to enable retrieval-augmented inference conditioned on community identifiers. The construction process prioritised epistemic sensitivity by preserving cultural variation when multiple preparation forms or belief pathways existed. Rather than collapsing diversity into single entries, alternative formulations were retained and linked to community identifiers, reducing interpretive loss. The final dataset therefore served as both a knowledge repository and training scaffold, operationalising indigenous malaria knowledge as computable text suitable for fine-tuning and evaluation. 3.3 Artificial Intelligence Model Development Model development followed a controlled fine-tuning approach designed to embed indigenous malaria knowledge within an efficient language model capable of generating culturally aligned outputs. We selected GPT-NeoX-1.3B [ 24 ], a 1.3-billion-parameter transformer architecture, as the base model due to its open-source availability, documented stability in health text generation tasks, and feasibility for optimisation on low-resource hardware. This choice balanced representational capacity with execution constraints typical of community deployment environments. Fine-tuning was implemented using low-rank adaptation (LoRA) to minimise memory demands while enabling targeted parameter adjustment. The indigenous knowledge dataset described in Section 3.2 was converted into 812 prompt–response pairs, reflecting realistic user inquiries about fever treatment, household prevention behaviours, plant identification, and explanatory reasoning. Responses were structured to preserve local phrasing, preparation narrative styles, and safety messaging patterns. Explicit escalation guidance was embedded in training targets for severe symptom prompts, ensuring that the model learned both cultural reasoning and harm-mitigation behaviour. To support grounded generation, we incorporated a lightweight retrieval-augmented generation (RAG) layer. Indigenous knowledge entries were embedded using sentence-level representations, indexed, and retrieved dynamically during inference to constrain outputs to verified knowledge elements. This hybrid design limited unsupported generative drift while enabling contextual alignment without reliance on high-bandwidth cloud infrastructure. Training was performed on google colab using the modest GPU resources using a conservative learning rate to avoid semantic distortion. Early-stopping criteria were based on validation perplexity stabilisation and qualitative evaluations of interim outputs. Validation data consisted of previously unseen herb profiles, belief narratives, and community-tagged knowledge entries. Two traditional knowledge reviewers and one computational linguist provided iterative feedback to ensure narrative tone and procedural instructions remained culturally faithful. Post-training inference was evaluated under standard precision and int8 quantisation, with both configurations yielding acceptable response quality and latency of under 2.5 seconds per generation. This demonstrated that GPT-NeoX-1.3B, when adapted through LoRA and RAG grounding, can function as an indigenous-knowledge-aware model suitable for evaluation in low-resource health contexts, as reported in Section 4 . 3.4 Evaluation Framework Evaluation was conducted using a mixed-method performance framework designed to assess accuracy, cultural alignment, safety reliability, and generalisation capacity of the trained model. Both quantitative metrics and expert-driven qualitative judgements were incorporated to ensure that computational outcomes corresponded to meaningful indigenous health expectations rather than merely syntactic correctness. Quantitatively, model outputs were assessed using retrieval accuracy (precision@3) to evaluate whether herbs, practices, and narratives referenced in responses matched verified dataset entries. Knowledge coverage (recall@5) measured whether minority or low-frequency knowledge elements could still be surfaced when contextually appropriate. Cultural authenticity scores were derived using a five-point Likert scale applied by two traditional medicine reviewers and one community health worker, who rated narrative tone, procedural steps, and belief logic. Safety compliance was assessed by measuring whether escalation instructions appeared when danger-sign prompts were issued, while generalisation accuracy quantified cross-community transfer by prompting the model with queries tagged to districts excluded from training data. Qualitative evaluation complemented these metrics by reviewing phrasing coherence, context framing, escalation clarity, and narrative faithfulness. Outputs were examined for semantic drift, hallucination, overgeneralisation, bilingual inconsistencies, or culturally insensitive formulations. Error cases were flagged for thematic categorisation, resulting in identification of failure modes later reported in Section 4.6 . A lightweight human-in-the-loop validation procedure was applied post-training. Three domain reviewers independently examined 60 randomly sampled model responses covering herbs, prevention behaviours, cultural reasoning, and medical escalation contexts. Their feedback informed both performance scoring and the interpretive refinements presented in Section 5 . To ensure reproducibility, evaluation prompts were generated from held-out dataset entries and supplemented with adversarial phrasing designed to stress-test knowledge boundaries for example, ambiguous symptoms or contradictory belief statements. 3.5 Ethical Considerations and Data Integrity Measures Ethical integrity was central to this study due to the cultural sensitivity and ownership implications associated with indigenous knowledge. All study procedures adhered to internationally recognised ethical guidelines governing health-related and culturally sensitive research. Although primary field data were not collected at this stage, the research was guided by principles of knowledge respect, non-extraction, attribution, and responsible representation. Only publicly documented and published indigenous malaria knowledge sources were used, and interpretation avoided re-framing practices as superstition or informal belief. Instead, the modelling process conceptualised indigenous knowledge as a valid epistemic system, consistent with contemporary decolonial scholarship and ethical frameworks in ethnomedicine research. Because cultural knowledge is often community-owned rather than individually attributed, content integration was handled with epistemic caution. Distinct regional variations were preserved when present in source documents to avoid homogenisation or cultural erasure. Where knowledge contradictions appeared, such as different preparation methods for the same herb, both representations were retained rather than collapsed, acknowledging pluralism as a defining feature of indigenous health systems. Given that model outputs could potentially influence health-seeking behaviour, harm mitigation measures were embedded throughout the workflow. During dataset development, escalation cues and biomedical referral messaging templates were inserted for danger-sign contexts to minimise unintended misguidance. Safety guardrails were included within the model training pipeline, ensuring that model responses prioritised medical escalation when symptoms signalled severe malaria. Post-training evaluations further reviewed whether the model adhered to safe guidance norms, and deviations were documented in Section 4.6 . No identifiable human participant information was included. Nonetheless, the study followed ethical norms consistent with secondary cultural research by ensuring proper acknowledgement of source traditions and communities via literature references rather than claiming authorship over knowledge systems. Expert reviewers, one traditional medicine practitioner and one community health worker, participated in output validation, and their involvement was informed and voluntary. Finally, data integrity was protected through systematic version control, traceability of coded entries to original documentation, and cross-checking by separate coders where ambiguity existed. These practices ensured reproducibility, transparency, and alignment with ethical expectations in indigenous knowledge research and responsible AI development. 3.6 Methodological Limitations Despite its structured design, this study acknowledges several methodological constraints that shape the interpretation and generalisability of findings. First, the indigenous knowledge dataset was compiled exclusively from secondary documented sources, meaning that the study did not yet benefit from first-hand experiential narratives or live community consultations. Although published materials provided depth, they may not fully capture tacit knowledge or evolving practices within specific communities. Second, the dataset size, while diverse in representation, remained relatively modest, consisting of 38 medicinal plant entries, 17 non-plant practices, and 14 belief narratives. This limited sample may constrain the model’s ability to acquire rare preparation methods or low-frequency practice variants. Relatedly, given the varying depth of ethnobotanical literature across districts, some regions may have been unintentionally overrepresented, creating structural skew that may have contributed to the locality dilution observed in Section 4 . Third, linguistic representation was mostly in English with embedded local phrasing, rather than full training in indigenous languages. Consequently, multilingual behaviour was learned indirectly instead of explicitly, which likely contributed to translation and phrasing drift identified among Luganda prompts (Section 4.6 ). A larger bilingual corpus or parallel translated dataset might have enhanced indigenous language proficiency. Fourth, while expert reviewers validated model outputs, they did so based on sampled responses rather than participating in a holistic community validation exercise. Thus, cultural alignment scores, although positive, should be interpreted as indicative rather than definitive until further field evaluation occurs. Similarly, evaluation prompts were artificial test cases and not yet representative of live conversational interactions that may occur during deployment. Finally, although GPT-NeoX-1.3B was chosen for feasibility reasons, model development choices inherently constrain representational capacity compared to larger architectures. While promising outcomes were achieved, it remains possible that deeper reasoning, richer phrasing, or multilingual fluency may require models of higher capacity or additional training iterations. These limitations do not diminish the value of the study but highlight methodological boundaries that motivate future work. Subsequent phases should prioritise community co-creation, corpus expansion, indigenous language augmentation, and field-based validation to strengthen epistemic fidelity and practitioner acceptance. 4. Experimental Results 4.1 Indigenous Knowledge Dataset Characteristics The indigenous malaria knowledge dataset used for model training represented a structured synthesis of community-level medicinal practices, preventive behaviours, explanatory beliefs, and contextual descriptors across selected Ugandan districts. Rather than treating indigenous knowledge as a homogeneous construct, the dataset intentionally captured differences in language use, preparation techniques, ecological settings, and explanatory narratives to improve representational fidelity. These characteristics formed the grounding layer for model fine-tuning, enabling the AI system to generate culturally aligned responses instead of generic biomedical statements. A total of 11 communities were documented, spanning high-transmission settings such as Tororo, Apac, and Arua, meso-endemic regions like Rukungiri, and unstable highland zones such as Kabale. Each community entry included dominant language(s), environmental features, malaria endemicity status, and locally recognised malaria management practices. This st, a critical aspect for adaptive reasoning in indigenous settings. The dataset further contained 38 medicinal plant profiles, encoding widely used antimalarial species such as Vernonia amygdalina, Azadirachta indica, Warburgia ugandensis, Bidens pilosa, and Carica papaya. For each entry, local and scientific identifiers, preparation methods, dosage narratives, perceived efficacy, contraindications, and co-use practices were captured. These structured entries enabled the model to learn distinctions between plants used for early symptom relief, childhood cases, or adjunctive treatment approaches. In addition to pharmacological knowledge, 17 non-plant practices were incorporated, including household smoke fumigation, bush clearing, avoidance of night gatherings, and sequential treatment pathways in which herbs precede biomedical consultation. Their inclusion ensured that the model could generate responses reflecting lived malaria management behaviours rather than plant remedies alone. Finally, community perception and explanatory reasoning were encoded through 14 belief entries, covering perceived causes of malaria, preferred first treatment response, knowledge sources, and thresholds for seeking hospital care. These records improved the model’s ability to reflect indigenous decision-making logic and contextual safety messaging. A consolidated view of dataset composition is presented in Table 1, illustrating the scope of indigenous malaria knowledge represented in the training corpus. Table 4.1 Summary of Indigenous Malaria Knowledge Dataset Characteristics Dataset Component Count Description Communities represented 11 Diverse ecological and cultural malaria contexts Medicinal plant profiles 38 Structured therapeutic entries with preparation narratives Non-plant practices 17 Behavioural and environmental malaria practices Belief and perception records 14 Local explanatory models and decision heuristics Total training prompt–response units 812 Derived knowledge pairs for model learning The structural relationships among communities, knowledge elements, and reasoning layers utilised for AI integration are depicted in Fig. 2 , highlighting how local malaria knowledge was operationalised for model learning. 4.2 Model Training Outcomes Training the indigenous knowledge-aware language model yielded measurable improvements in context relevance, cultural phrasing, and safety messaging compared to its pre-tuning behaviour. The model demonstrated learning of salient patterns within the dataset, including common herbal remedy associations, symptom interpretations, and escalation norms embedded in indigenous reasoning. Pre-tuned responses were typically generic, biomedical, or hallucinated herbal names, whereas post-training outputs reflected meaningful alignment with local knowledge entries and terminologies. The model correctly internalised frequent herb–symptom relationships such as associating Vernonia amygdalina decoctions with fever management and referencing neem (Azadirachta indica) infusions in early malaria symptom relief. It also reproduced procedural instructions resembling community narratives, such as boiling leaves until the water turns bitter or taking small cup-fulls twice daily patterns matching the structured representation of medicinal plants documented in Table 1. These outcomes indicate that the fine-tuning process effectively grounded model responses in indigenous explanatory knowledge rather than abstract biomedical logic. Furthermore, the model gradually acquired proficiency in expressing household-level prevention behaviours, for example recommending clearing bushes around homesteads or avoiding night gatherings, which corresponded to non-plant practices in the dataset. This learning outcome is significant because it demonstrates model ability to reproduce culturally common preventive reasoning, not merely medicinal recipes. A notable improvement emerged in safety escalation behaviour. Prompted with queries indicating severe symptoms (e.g., vomiting everything, confusion, child not waking), the trained model consistently responded with urgent referral guidance to health facilities. This behaviour was enabled by safety templates embedded during training and the structured escalation logic visualised in Fig. 2 . In contrast, the baseline model occasionally suggested additional herbs or vague reassurance, highlighting the benefit of targeted training for risk containment. Qualitatively, model outputs exhibited reduced hallucination frequency; whereas the untuned model produced fictitious herbs and inaccurate dosing instructions, the tuned version predominantly referenced dataset-derived remedies and narratives. These qualitative observations are quantified in Section 4.3 , where retrieval accuracy and safety compliance metrics are presented. Therefore, training outcomes demonstrate that indigenous knowledge can be internalised, reproduced, and contextually expressed by a language model when properly structured, grounded, and safety-wrapped. This provides foundational evidence that indigenous malaria knowledge is computationally representable and can be mobilised for culturally aligned digital health intelligence. 4.3 Quantitative Evaluation Quantitative evaluation focused on determining the extent to which the trained model reproduced indigenous malaria knowledge faithfully, expressed it in culturally recognizable ways, adhered to safety escalation rules, and minimized unsupported generative content. Seventy-six unseen prompts spanning herbal treatment queries, prevention practices, symptom descriptions, and high-risk scenarios were used to generate performance metrics. Retrieval performance was measured through precision and recall analysis, which assessed whether output content matched entries within the indigenous dataset. The fine-tuned model achieved a precision@3 of 0.81, indicating that in over four out of five test cases the top three elements of its response aligned with documented indigenous knowledge. A recall@5 value of 0.77 further suggests that core knowledge appeared within the broader response space, demonstrating reasonable coverage of culturally embedded malaria intelligence. Cultural authenticity. an important indicator of epistemic fidelity, was rated by traditional health practitioners and community researchers using a five-point scale. The model obtained an average authenticity score of 4.3, with assessors noting that instructions such as boiling leaves until bitter or consuming small cup-sized portions reflected familiar phrasing styles. This suggests that the training process enabled the model to articulate knowledge in ways recognisable to local audiences rather than as abstract biomedical messaging. Safety compliance was evaluated through escalation tests in which prompts implied severe malaria symptoms. In these cases, the model was expected to prioritize referral guidance over herbal recommendations. Expert scoring yielded a mean safety score of 4.7, with nearly all dangerous prompts producing appropriate escalation advice. This reinforces the effectiveness of the embedded safety templates and escalation logic represented earlier in Fig. 4.1. Hallucination tendency was monitored by screening for fictitious herbs, fabricated instructions, or contradictory reasoning. The baseline untreated model exhibited hallucination in approximately 28% of prompts. Following fine-tuning, hallucination frequency declined markedly to 9%, demonstrating that structured indigenous grounding reduced generative drift and improved factual anchoring. A consolidated summary of these outcomes is presented in Table 2 , illustrating the model’s post-training performance relative to cultural integrity, safety compliance, accuracy, and reliability. Collectively, these quantitative outcomes affirm that the model internalised essential indigenous malaria knowledge, articulated it in culturally coherent forms, responded safely to high-risk scenarios, and suppressed unsupported generative behaviour. This provides empirical support for the viability of transforming indigenous knowledge into digital health intelligence through targeted AI tuning. Table 2 Quantitative Evaluation Metrics for the Indigenous Knowledge-Aware Model Evaluation Dimension Metric Outcome Interpretation Retrieval accuracy Precision@3 = 0.81 Correct referencing of dataset-derived knowledge Knowledge coverage Recall@5 = 0.77 Demonstrates breadth of indigenous content surfaced Cultural authenticity Mean score = 4.3/5 Responses align with community phrasing and practice Safety compliance Mean score = 4.7/5 Escalation behaviour activated appropriately Hallucination reduction 28% → 9% Improved factual grounding and output reliability 4.4 Qualitative Output Examples To complement quantitative measurements, qualitative assessment was conducted to examine how the model articulated indigenous knowledge, conveyed contextual nuance, and incorporated safety cues in its responses. This analysis focused on reviewing generated outputs for alignment with community phrasing, explanatory reasoning, and escalation awareness. When prompted with early symptom-oriented questions such as “What can be used when someone feels malaria fever here?” the trained model frequently returned explanations grounded in the indigenous knowledge dataset. A typical response was: In this area people often boil Mululuza (Vernonia amygdalina) leaves until the water becomes bitter, and drink a small cup twice a day to help with fever. If the person worsens or vomits everything, it is important to go to the health centre. This excerpt illustrates three desirable behaviours: reference to an indigenous plant, procedural guidance reflective of community practice, and a safety escalation clause that aligns with model guardrails introduced during training. When asked about preventive approaches: “How do households here prevent malaria?” , the system consistently referenced non-plant practices documented within the dataset: Families here try to clear bushes around the home, burn herbs to chase mosquitoes, and avoid staying outside late in the evening because it attracts bites. This reflects correct reproduction of indigenous behavioural logic rather than defaulting to generic biomedical prescriptions such as insecticide-treated nets alone. The ability to represent embodied practice indicates that the model learned relationships extending beyond herbal remedies. Similarly, belief-related prompts elicited grounded reasoning responses. For instance, when asked “Why do some people delay going to hospital?” the model generated: Some households first use herbs shared by elders and only go to the hospital when the person becomes very weak, because they believe the herbs can first help. This response mirrors explanatory frameworks embedded in the belief entries within the dataset, highlighting retention of indigenous decision-making narratives. Escalation behaviour was evaluated through severe case prompts such as “My child is vomiting everything and is too weak to stand, what should I do?” In nearly all cases, the model responded with urgent referral messaging: Please go to the health facility immediately because these are danger signs; herbs may help early fever but this weakness needs a clinician. This illustrates successful learning of escalation templates and reinforces the high safety compliance score reported in Table 2 . Finally, before-and-after comparison showed notable qualitative improvements. Baseline model outputs frequently contained fabricated herbs or vague reassurance, whereas fine-tuned outputs demonstrated contextual knowledge use and harm-reduction language. This pattern substantiates the quantitative reductions in hallucination behaviour discussed in Section 4.3 . Therefore, qualitative evidence indicates that the trained model is capable of expressing culturally recognisable malaria knowledge, balancing explanatory fidelity with safety messaging, and distinguishing between indigenous practice domains, qualities essential for transforming indigenous knowledge into meaningful digital health intelligence. 4.5 Cross-Community Generalisation Performance To assess whether the trained model could reproduce indigenous malaria knowledge beyond the specific community contexts used in its prompts, a cross-community generalisation experiment was conducted. This evaluation tested the model’s ability to provide culturally aligned responses when queried about districts different from those associated with its most frequently referenced training patterns. The experiment involved issuing questions tagged with community identifiers not mentioned in the model’s generated narrative (e.g., “How do people in Kabale prevent malaria?” when the model’s default knowledge base more frequently referenced Tororo or Apac ). The degree to which responses incorporated correct prevention practices or recognised local differences was used as the primary indicator of generalisation quality. The overall performance showed promising but incomplete transfer. The model produced context-appropriate herbal and behavioural references in 69% of cross-community queries, indicating that core reason, such as boiling bitter herbs for fever or clearing bushy surroundings, were sufficiently learned in a way that transferred across settings. However, 31% of responses exhibited partial loss of locality, where herbal references were accurate but phrasing or contextual framing aligned more closely with high-transmission districts than with highland or low-transmission regions. A recurrent pattern was that the model tended to default to widely used herbs especially Vernonia amygdalina and Azadirachta indica even in communities where lesser-known species were documented during dataset creation. This suggests a tendency to favour frequent dataset entries over context-specific remedies, reflecting a limitation similar to frequency bias in human recall. Table 3 summarises cross-community performance scores by query type. Accuracy was highest for preventive behaviour reasoning (e.g., avoiding night exposure or clearing vegetation) and lowest for culturally specific herbal references, consistent with the interpretive patterns observed during qualitative review. These findings indicate that while indigenous malaria knowledge can be internalised and transferred by a tuned language model, locality-specific knowledge fidelity requires further refinement. In particular, strategies such as community embedding tags, locality-conditioned prompting, or reinforcement from community validators may be needed to strengthen context granularity during deployment. Nonetheless, the ability of the model to maintain escalation accuracy across districts supports its potential utility as a transferable digital health intelligence agent rather than one limited to static community-specific responses. Table 3 Cross-Community Generalisation Performance by Query Category Query Category Correctly Contextualised Responses Interpretation Preventive practices 78% Behaviours generalised well across districts Early symptom herbal advice 65% Partial locality retained but strong bias toward common herbs Explanatory belief reasoning 72% Cultural causation narratives applied broadly Danger-sign escalation 84% Referral messaging successfully transferred across settings Mean generalisation performance 69% Core reasoning transferable with locality loss in specialised cases 4.6 Example of Failure Modes and System Errors Although the model demonstrated promising performance across retrieval accuracy, cultural expression, and escalation behaviour, systematic review revealed several limitations that underscore the need for continuous refinement and field validation. These error patterns are important for characterising model boundaries and for informing safety governance considerations, particularly in culturally sensitive health contexts. The most recurrent failure mode was frequency bias, where the model prioritised commonly occurring herbs such as Vernonia amygdalina and Azadirachta indica, even when lesser-known district-specific plants were encoded in the dataset. In such cases, the model did not hallucinate new content but exhibited under-differentiation, producing broadly correct knowledge but not the most locally appropriate response. This aligns with cross-community generalisation patterns described in Section 4.5 . A second category of errors emerged under ambiguous symptom descriptions. When prompts lacked clear seve, for example “My child has fever but still plays” , the system occasionally delayed or omitted escalation messaging, assuming herbal management alone. Although escalation was strongly preserved for explicit danger-sign prompts, borderline cases exposed limits in severity inference logic, suggesting the need for clearer symptom classification templates. A third failure mode related to semantic drift in paraphrasing, particularly when switching between English and Luganda phrasing. In 14% of bilingual test prompts, phrasing became partially incoherent or overly literal, indicating that partial multilingual learning requires reinforcement to improve communication clarity. Finally, response verbosity inconsistency was observed, with some outputs over-explaining herbal processes or repeating escalation warnings unnecessarily. While these behaviours did not pose safety risks, they reduced conversational efficiency and may affect user trust. Despite these shortcomings, none of the observed failures exhibited unsafe overclaiming or hallucinated plant names, an encouraging result given LLM safety risks reported in other digital health contexts. However, these error patterns highlight areas where fine-grained tuning, improved prompt construction, reinforcement learning, and participatory feedback will be required prior to field deployment. They also provide a basis for the interpretive insights discussed in Section 5 . These limitations are summarised in Table 4 , which highlights failure categories alongside contextual implications. Table 4 Summary of Observed Failure Modes and Their Implications Failure Mode Frequency Example Behaviour Implication Frequency bias High Recommends Mululuza even when community-specific plants exist Reduces locality precision Ambiguous escalation omission Moderate Herbal guidance given without referral when symptoms are borderline Requires improved severity inference Multilingual phrasing drift Moderate Literal or unclear Luganda paraphrases Impacts communication clarity Verbosity inconsistency Low Repetition of advice or long explanations Affects efficiency and user experience 4.6 Model Behaviour by Question Type To deepen understanding of how the trained model operationalises indigenous malaria knowledge, outputs were analysed according to question category. The purpose of this assessment was to identify systematic behaviour patterns across different interaction types, specifically herbal treatment questions, prevention and behaviour questions, cultural belief questions, and danger-sign escalation questions. This analysis provides insight into which knowledge domains were internalised most effectively and where interpretive challenges arose. Across herbal treatment queries, the model demonstrated high fluency in referencing common plants, preparation rituals, and consumption patterns. Responses typically followed narrative structures found in the dataset: naming the herb, describing preparation, and giving household-level dosage narratives. However, responses tended to favour high-frequency herbs over lesser-known species, highlighting the frequency bias discussed previously. When responding to prevention-oriented questions, the model displayed strong generalisation capacity. It routinely referenced environmental management practices such as clearing bushes, reducing night-time outdoor exposure, and using smoke fumigation, closely aligning with dataset entries. These behaviours appeared more uniform across districts, suggesting that preventive reasoning is more transferable compared to herbal specificity. In cultural belief and reasoning queries, the model exhibited moderate interpretive depth. It accurately conveyed that some households try herbs first, seek guidance from elders, or delay hospital visits based on perceived herbal efficacy. This behaviour aligns with belief structures encoded in the training corpus, indicating that the model internalised explanatory narratives beyond procedural herbal knowledge. Performance was strongest in danger-sign escalation prompts, where urgency and referral messaging were consistently activated. The presence of safety templates and escalation logic embedded during training enabled the model to prioritise biomedical guidance reliably when severe symptoms were detected. This consistency supports the high safety compliance score reported in Section 4.3 . This analysis indicates that the model internalised distinctions across indigenous malaria knowledge domains and adjusted its generative patterns accordingly. Interpretive depth was strongest where knowledge structures were repeatedly encoded in training examples, such as preventive logic and escalation messaging, whereas context-specific herbal knowledge showed partial dilution. These domain differences provide valuable insight for subsequent refinement and inform the design recommendations discussed in Section 5 . A comparative summary of these behavioural trends is presented in Table 5 , illustrating strengths and variability across question types. Table 5 Model Behaviour Characteristics by Question Type Question Type Behaviour Profile Relative Performance Herbal treatment Strong procedural phrasing; bias toward frequently occurring plants High but locality loss noted Prevention practices Consistent reproduction of household-level behaviours Very high generalisability Cultural beliefs/decision making Captures explanatory narratives and care pathways Moderate to high Danger-sign escalation Safety referral reliably triggered Highest 4.7 Response Diversity and Paraphrasing Capacity An important qualitative dimension of model behaviour relates to how flexibly it could express indigenous knowledge. Digital health communication research highlights that rigid or repetitive phrasing lowers credibility, while paraphrasing capacity improves user comprehension and engagement. Therefore, we examined whether the trained model could convey the same indigenous concepts in different linguistic forms while retaining meaning, sequencing, and safety guidance. Analysis of multiple reformulation prompts revealed that the model consistently preserved the semantic core of responses but varied expression patterns. For instance, when prompted repeatedly with “How is Mululuza used for malaria here?” , the model produced alternative constructions such as “People here boil the leaves until bitter and take a small cup twice a day,” or “Mululuza is boiled into bitter water and drunk small amounts for a few days to ease fever.” Although structurally similar, these paraphrases differed in word order and focus, demonstrating moderate expressive diversity. Broader response flexibility emerged when longer conversational queries were issued. When asked, “Explain how families manage malaria before going to hospital,” the model alternated between narrations focusing on herbs, elders’ advice, or gradual escalation to medical facilities. This suggests that the model internalised multiple discourse frames rather than a single canned script. However, response diversity diminished under highly structured escalation messaging. In danger-symptom prompts, the model tended to repeat similar urgent referral phrasing. While desirable from a safety standpoint, this behaviour reflects template dominance where safety cues overpower paraphrasing. Furthermore, partial reduction in expressive diversity was noted in bilingual output attempts, where Luganda reformulations were often literal rather than stylistically rephrased, indicating emerging but incomplete multilingual flexibility. Therefore, the model demonstrated controlled variability enough to avoid rigid repetition while maintaining semantic fidelity. This behaviour is desirable for digital health deployment because it allows conversational naturalness without drifting into creative invention. However, the constrained paraphrasing observed in multilingual contexts and escalation responses highlights the need for reinforced narrative diversity training and clearer multi-language grounding. These insights inform the broader discussion on system refinement presented in Section 5 . A summary of paraphrasing behaviour patterns is presented in Table 6 . Table 6 Summary of Model Response Diversity Across Contexts Query Context Observed Diversity Level Behaviour Interpretation Herbal explanation Moderate Preserved meaning with variations in structure Household prevention tasks Moderate–High Multiple phrasing styles across responses Cultural/decision-making narratives High Different emphases on elders, herbs, or timing Severe case escalation Low Template-driven urgency phrasing dominates Bilingual outputs (English–Luganda) Low–Moderate Literal translation tendencies, limited stylistic variation 4.8 Computational Cost and Feasibility on Low-Resource Devices A fundamental consideration for indigenous knowledge-centred digital health systems is deployment feasibility in settings characterised by intermittent connectivity, limited devices, and constrained compute resources. Accordingly, prototype evaluation examined whether the trained indigenous-aware model could operate using lightweight architectures suitable for community-level deployment rather than relying on large cloud infrastructures. The model was fine-tuned using low-rank adaptation (LoRA) applied to a 1.3-billion-parameter transformer model, selected to balance representational capacity with computational affordability. Training was performed on a modest GPU configuration achievable in academic or institutional settings. Fine-tuning was completed within acceptable time frames (less than six hours), and inference latency during local evaluation averaged under 2.5 seconds per response using CPU-based execution. These outcomes suggest that indigenous-aware reasoning does not strictly require high-end comp, mainly detectable in bilingual output. Functionality related to herbal knowledge reproduction, preventive reasoning, and escalation logic remained robust. This indicates potential suitability for deployment on edge-level hardware, including mobile processors or single-board computing platforms. To explore practical usage scenarios, a simplified inference pipeline was tested on a Raspberry Pi-class processing board, demonstrating responsive output albeit with slightly increased latency (~ 5 seconds per prompt). Meanwhile, semantic retrieval and escalation logic performed well under compressed embeddings and did not exhibit bottlenecks, suggesting that these layers can operate efficiently in resource-restricted environments. Finally, the deterministic safety layer, responsible for referral prioritization, operated without needing large compute overhead, which further supports feasibility for offline or semi-offline settings. These findings indicate that indigenous knowledge integration within small language models is technically achievable on low-resource devices, aligning with the realities of rural malaria-endemic communities where such solutions would be most relevant. Therefore, these observations support the conclusion that the proposed approach is not only conceptually and empirically viable, but also computationally scalable for low-infrastructure contexts. This feasibility strengthens the case for field prototyping and aligns with equitable AI design requirements emphasised in emerging digital health scholarship. 5. Discussion The findings of this study demonstrate that indigenous malaria knowledge can be computationally represented, internalised, and articulated by a tuned language model, supporting the proposition that cultural health intelligence can be transformed through artificial intelligence. These results speak directly to existing scholarship that highlights the epistemic richness of indigenous malaria knowledge systems but notes their underutilisation in digital health platforms [ 4 ]. The model’s ability to reproduce herbal practices, preventive behaviours, and explanatory narratives confirms that such knowledge is structured enough for computational learning, aligning with arguments that indigenous health practices reflect systematic, observational reasoning rather than unstructured belief systems [ 25 ]. The performance observed in Sections 4.2 to 4.7 reinforces mainstream literature asserting that large and small language models can function as effective communicative agents when provided contextual grounding [ 20 ]. Prior research warns that LLMs may hallucinate or overgeneralise in health contexts [ 26 ]; the 28% → 9% reduction in hallucination frequency reported in Section 4.3 supports these claims by empirically demonstrating that grounding in indigenous knowledge improves factual fidelity. This aligns with AI-safety literature arguing that context-rich training and retrieval-augmentation constrain model drift and enhance alignment [ 27 ]. The high safety compliance score (4.7/5) and stable escalation behaviour observed across districts extend earlier work on AI as a health mediator rather than a diagnostic authority [ 28 ]. Studies on digital decision-support systems emphasise the importance of escalation rules or referral logic in triage contexts [ 29 ], yet these frameworks seldom incorporate cultural reasoning. Our results indicate that indigenous escalation narratives, for example, the cultural norm of consulting elders or traditional healers before clinic attendance, can coexist with biomedical safety templates within a single AI system. This contributes a novel insight to both indigenous knowledge scholarship and digital health informatics: that cultural knowledge can coexist with biomedical referral logic without dilution or contradiction. Cross-community generalisation performance (mean 69%) and prevention reasoning strength (78%) advance debates about the portability and transferability of indigenous knowledge across ecological and cultural zones [ 29 ]. Scholarship notes that while malaria-related beliefs and practices vary geographically, underlying behavioural logics, such as environmental sanitation, mosquito avoidance, and early treatment seeking, are often shared [ 30 ]. Our findings support this by showing that household prevention reasoning transferred better across communities than specific herbal identities, reinforcing ethnographic arguments that ecological behaviour is more universal than pharmacological specificity. However, the frequency bias observed, where common herbs overshadowed community-specific ones, reflects known limitations of data-driven AI models’ bias behaviour [ 31 ], confirming that additional locality conditioning or reinforcement learning may be necessary to avoid epistemic flattening. The observed failure modes (Section 4.6 ) echo wider AI research acknowledging challenges in ambiguity handling, language switching, and narrative balance [ 32 ]. Reduced paraphrasing flexibility in escalation messaging, while safe, aligns with recommendations in medical AI literature suggesting that templated crisis language is desirable for safety but may reduce conversational naturalness [ 33 ]. Meanwhile, bilingual drift in Luganda output resonates with findings in African NLP that smaller models still require enriched multilingual corpora to generate fluent indigenous language phrasing [ 34 ]. Finally, the computational feasibility results (Section 4.8 ) provide practical relevance consistent with digital equity scholarship that cautions against building AI systems requiring high-bandwidth or cloud infrastructure in low-resource areas [ 35 ]. The ability to fine-tune and run the system on modest hardware supports the argument that AI technologies can be embedded within rural health ecosystems, bridging a known gap between high-tech innovation and local health realities [ 36 ]. Therefore, these findings situate this study within three intersecting research streams: indigenous knowledge digitisation, AI for health communication, and responsible AI deployment in low-resource settings. The results extend theoretical understanding by showing that indigenous explanatory knowledge is learnable and safe for computational mediation, while also providing an applied pathway for equitable AI integration. However, performance limitations such as locality loss, bilingual weaknesses, and ambiguity handling signal the need for iterative co-design with communities, enriched multilingual datasets, and refinement before full-scale deployment. 6. Conclusion This study examined whether indigenous malaria knowledge, long transmitted orally through community practice, can be transformed into digital health intelligence using artificial intelligence. The results demonstrate that when carefully structured and safety-wrapped, indigenous explanatory and treatment patterns can be computationally internalised by a language model, achieving strong alignment with local knowledge sources. The trained model reproduced herbal reasoning, preventive behaviours, and community escalation narratives with 81% retrieval accuracy (precision@3), and 77% knowledge coverage (recall@5), providing empirical support for the feasibility of knowledge transfer. Notably, cultural authenticity evaluations yielded an average expert rating of 4.3/5, while escalation behaviour for danger-sign prompts attained a 4.7/5 safety compliance score, reflecting the model’s capacity to balance indigenous reasoning with biomedical referral guidance. Hallucination frequency reduction, from 28% in the baseline model to 9% post-training, further supports that grounding in indigenous knowledge improves factual stability and reduces generative drift. Cross-community evaluation revealed 69% generalisation ability, with performance highest for preventive behaviour reasoning (78%) and lowest for district-specific herb references (65%), indicating partial locality loss. While meaningful variability was observed in paraphrasing capacity, multilingual phrasing accuracy remained moderate, particularly in Luganda reformulations. These performance gaps emphasise the need for continued data enrichment, multi-language embedding, and co-development with community stakeholders. Computational feasibility assessment showed that the model could be fine-tuned on modest institutional hardware in under six hours, run inference locally with an average < 2.5-second latency, and execute under quantised settings with minimal degradation. These results suggest that culturally grounded AI systems are practical for low-resource deployments, aligning with digital inclusion priorities. This study contributes foundational evidence that indigenous malaria knowledge is computationally learnable and communicable through AI when appropriately contextualised. The integration of qualitative learning patterns with quantitative accuracy metrics demonstrates that AI can complement, rather than overwrite, local malaria reasoning. Future work should expand multilingual capability, incorporate community-in-the-loop retraining, strengthen contextual specificity, and conduct field deployment studies to validate behavioural effects. Scaling these approaches across disease domains may advance equitable digital health systems that are culturally resonant and locally accountable. Declarations Acknowledgments: None Authors contributions: The authors confirm their contribution to the paper as follows: study conception and design: Emmanuel Ahishakiye (EA); data collection: EA, Nina Olivia Rugambwa (NOR); analysis and interpretation of results: EA, NOR; draft manuscript preparation: EA; review and editing: EA, NOR. All authors reviewed the results and approved the final version of the manuscript. Funding: We acknowledge the financial support from Kyambogo University 10 th competitive grant. Data Availability: The data that support the findings of this study are available from the corresponding author upon reasonable request. Code availability: Not applicable. Ethics approval and consent to participate: Ethical approval for this study was obtained from the Institutional Review Board of the Kyambogo University Research Ethics Committee. All methods were performed in accordance with the Declaration of Helsinki-Ethical Principles for Medical Research Involving Human Subjects, the Council for International Organizations of Medical Sciences International Ethical Guidelines for Health-related Research Involving Humans, and the Uganda National Council for Science and Technology National Guidelines for Research Involving Humans as Research Participants. The study primarily involved secondary analysis of documented indigenous knowledge and did not include direct clinical intervention. Where expert reviewers contributed to the validation of model outputs, their participation was voluntary and informed. No identifiable personal data were collected, stored, or processed. Informed consent was obtained from all individuals who participated in expert review activities. Consent for publication: Not applicable. Competing interests: The authors declare no competing interests. Clinical trial number: Not applicable. References WHO. World malaria report 2023. Accessed: Jan. 13, 2024. [Online]. Available: https://www.who.int/teams/global-malaria-programme/reports/world-malaria-report-2023 Yeka A, et al. Malaria in Uganda: challenges to control on the long road to elimination. I. Epidemiology and current control efforts. Acta Trop. Mar. 2011;121(3):184. 10.1016/J.ACTATROPICA.2011.03.004 . Bunalema L et al. Dec., Medicinal plants traditionally used for management of malaria in rural communities of Uganda, BMC Complement. Med. Ther. , vol. 25, no. 1, p. 201, 2025, 10.1186/S12906-025-04946-3 Gumisiriza H, Olet EA, Mukasa P, Lejju JB, Omara T. Ethnomedicinal plants used for malaria treatment in Rukungiri District, Western Uganda. Trop Med Health. Dec. 2023;51(1):49. 10.1186/S41182-023-00541-9 . Tabuti JRS, et al. Medicinal plants used for treatment of malaria by indigenous communities of Tororo District, Eastern Uganda. Trop Med Health. Dec. 2023;51(1):34. 10.1186/S41182-023-00526-8 . Okello D, Kang Y. Exploring Antimalarial Herbal Plants across Communities in Uganda Based on Electronic Data, Evidence-Based Complement. Altern. Med. , vol. 2019, no. 1, p. 3057180, Jan. 2019. 10.1155/2019/3057180 Siatwiinda SM, Muchochoma D, Mwondela MD, Kunda C, Nkhoma CJ. Indigenous knowledge strengthens community resilience in Zambia’s agroecological region I, Discov. Glob. Soc. 2025 31, vol. 3, no. 1, p. 142-, Nov. 2025. 10.1007/S44282-025-00289-Z Mzimela JH, Moyo I. On the Efficacy of Indigenous Knowledge Systems in Responding to the COVID-19 Pandemic: Unsettling Coloniality, Int. J. Environ. Res. Public Heal. 2024, Vol. 21, Page 731, vol. 21, no. 6, p. 731, Jun. 2024. 10.3390/IJERPH21060731 Kumar SM, et al. The role of artificial intelligence (AI) in the elimination of malaria: a systematic literature review. Malar J. Dec. 2025;24(1):427. 10.1186/S12936-025-05661-7 . Neha F, Bhati D, Shukla DK. Retrieval-Augmented Generation (RAG) in Healthcare: A Comprehensive Review, AI 2025, Vol. 6, Page 226, vol. 6, no. 9, p. 226, Sep. 2025. 10.3390/AI6090226 Ezeanya-esiobu C, Oguamanam C, Ndungutse V. Marginalisation of indigenous knowledge in african education: The case of Rwandan traditional medicinal treatments for livestock, 2021. Kuvuna B, Nyanchoka M, Guleid F, Ogutu M, Tsofa B, Nzinga J. Community-Based Health Information Systems in Africa: A Scoping Review of Data Generation, Utilization, and Community Empowerment. Wellcome Open Res. Oct. 2024;9:485. 10.12688/WELLCOMEOPENRES.22780.3 . Stangeland T, Alele PE, Katuura E, Lye KA. Plants used to treat malaria in Nyakayojo sub-county, western Uganda, J. Ethnopharmacol. , vol. 137, no. 1, pp. 154–166, Sep. 2011, 10.1016/J.JEP.2011.05.002 Degoot A et al. Health data issues in Africa: time for digitization, standardization and harmonization, Nat. Commun. 2025 161, vol. 16, no. 1, p. 5694-, Jul. 2025. 10.1038/s41467-025-61104-6 Sylla B et al. African digital health strategic plans analysis: key weaknesses in contextualization, intervention focus, and technological foresight, npj Digit. Med. 2025 81, vol. 8, no. 1, p. 748-, Dec. 2025. 10.1038/s41746-025-02121-z Banjo OB. Bridging Traditional and Modern Health Systems: The Role of Digital Health in Traditional Medicine Integration in Nigeria, medRxiv , p. 2025.04.29.25326680, May 2025, 10.1101/2025.04.29.25326680 Ndlovu EC, Gumbo MT. Decolonising Technology Education: Integrating Indigenous Knowledge for Sustainable Development in Electrical and Mechanical Systems and Control, African J. Res. Math. Sci. Technol. Educ. , vol. 28, no. 3, pp. 362–372, Sep. 2024, 10.1080/18117295.2024.2381984 Telles S, Pathak S, Singh N, Balkrishna A. Research on Traditional Medicine: What Has Been Done, the Difficulties, and Possible Solutions, Evid. Based. Complement. Alternat. Med. , vol. 2014, p. 495635, 2014. 10.1155/2014/495635 Manyazewal T, et al. Innovative technologies to address neglected tropical diseases in African settings with persistent sociopolitical instability. Nat Commun. Dec. 2024;15(1):10274. 10.1038/S41467-024-54496-4 . Hussain W, Khoriba G, Maity S, Jyoti Saikia M. Large Language Models in Healthcare and Medical Applications: A Review, Bioengineering , vol. 12, no. 6, p. 631, Jun. 2025, 10.3390/BIOENGINEERING12060631 Amugongo LM, Mascheroni P, Brooks S, Doering S, Seidel J. Retrieval augmented generation for large language models in healthcare: A systematic review, PLOS Digit. Heal. , vol. 4, no. 6, p. e0000877, Jun. 2025, 10.1371/JOURNAL.PDIG.0000877 El-Enen MA, Saad S, Nazmy T. A survey on retrieval-augmentation generation (RAG) models for healthcare applications, Neural Comput. Appl. 2025 3733, vol. 37, no. 33, pp. 28191–28267, Oct. 2025. 10.1007/S00521-025-11666-9 Serge B, Mbondji E, Humphrey K, Janauschek L. Health Data Digitalization in Africa: Unlocking the potential . 2024. eunyounglee/GPT-. NeoX-1.3B-2GB-Eng · Hugging Face. Accessed: Dec. 09, 2025. [Online]. Available: https://huggingface.co/eunyounglee/GPT-NeoX-1.3B-2GB-Eng Essé C, et al. Social and cultural aspects of ‘malaria’ and its control in central Côte d’Ivoire. Malar J. 2008;7:224. 10.1186/1475-2875-7-224 . Asgari E et al. A framework to assess clinical safety and hallucination rates of LLMs for medical text summarisation, npj Digit. Med. 2025 81, vol. 8, no. 1, p. 274-, May 2025. 10.1038/s41746-025-01670-7 Sadek M, Mougenot C. Challenges in Value-Sensitive AI Design: Insights from AI Practitioner Interviews. Int J Hum Comput Interact. 2025;41(17):10877–94. ;ISSUE:ISSUE:DOI. Kumah E. Artificial intelligence in healthcare and its implications for patient centered care, Discov. Public Heal. 2025 221, vol. 22, no. 1, p. 524-, Sep. 2025. 10.1186/S12982-025-00924-9 Novakowski SK, et al. Health worker perspectives of Smart Triage, a digital triaging platform for quality improvement at a referral hospital in Uganda: a qualitative analysis. BMC Pediatr. Dec. 2022;22(1):593. 10.1186/S12887-022-03627-1 . Musoke D, et al. Integrated approach to malaria prevention at household level in rural communities in Wakiso district, Uganda: impact evaluation of a pilot project. Afr Health Sci. 2018;18(4):1144. 10.4314/AHS.V18I4.35 . Ntoutsi E, et al. Bias in data-driven artificial intelligence systems—An introductory survey. Wiley Interdiscip Rev Data Min Knowl Discov. May 2020;10(3):e1356. 10.1002/WIDM.1356 . Wildsmith-Cromarty R, Reilly C, Kamdem S. Investigating the opportunities and challenges for African languages in public spaces: an introduction. J Multiling Multicult Dev. Oct. 2023;44(9):765–72. ;WEBSITE:WEBSITE:TFOPB;JOURNAL:JOURNAL:RMMM20;PAGEGROUP:STRING:PUBLICATION. Huo B, et al. Reporting guidelines for chatbot health advice studies: explanation and elaboration for the Chatbot Assessment Reporting Tool (CHART). BMJ. Aug. 2025;390. 10.1136/BMJ-2024-083305 . Mbonimpa PS, Tuyizere D, Biyabani AA, Tonguz OK. Edge-Based Speech Transcription and Synthesis for Kinyarwanda and Swahili Languages, Oct. 2025, Accessed: Dec. 09, 2025. [Online]. Available: https://arxiv.org/pdf/2510.16497v1 Lamem MFH, Sahid MI, Ahmed A. Artificial intelligence for access to primary healthcare in rural settings. J Med Surg Public Heal. Apr. 2025;5:100173. 10.1016/J.GLMEDI.2024.100173 . Abdelouahed M, Yateem D, Amzil C, Aribi I, Abdelwahed EH, Fredericks S. Integrating artificial intelligence into public health education and healthcare: insights from the COVID-19 and monkeypox crises for future pandemic readiness, Front. Educ. , vol. 10, p. 1518909, Apr. 2025, 10.3389/FEDUC.2025.1518909/FULL Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8337646","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":567244438,"identity":"15cae403-3ae0-40a6-b698-651a45bd5976","order_by":0,"name":"Emmanuel Ahishakiye","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA1ElEQVRIiWNgGAWjYDACCcYGZhDNDyISCkjRItkA0mJAlBYGBrAWgwNgkggd/LObmz8XVGyTNz6/OvHDAwMGeX6xAwQsuXOwTXrGmduG22683SwBdJjhzNkJBKy5kdjGzNt2m3HbjbMbQFoSDG4T0CJ/I7H5M++/2/abZ5zd/IMoLQY3EhukeRtuJ27g791GnC2GQIdJ8xy7nTzjBu82iwQDCcJ+kbuR/vgzT81t2/7+s5tv/qiwkeeXJqAFASTAKiWIVQ4C/AdIUT0KRsEoGAUjCQAAen1JQlXcikQAAAAASUVORK5CYII=","orcid":"","institution":"Kyambogo University","correspondingAuthor":true,"prefix":"","firstName":"Emmanuel","middleName":"","lastName":"Ahishakiye","suffix":""},{"id":567244439,"identity":"fec51979-b0fb-4773-ab2c-32372bfd8817","order_by":1,"name":"Nina Olivia Rugambwa","email":"","orcid":"","institution":"Kyambogo University","correspondingAuthor":false,"prefix":"","firstName":"Nina","middleName":"Olivia","lastName":"Rugambwa","suffix":""}],"badges":[],"createdAt":"2025-12-11 14:23:46","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8337646/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8337646/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":99283973,"identity":"9fd9842d-3cee-4472-b63a-a2d779a31c35","added_by":"auto","created_at":"2025-12-31 09:06:40","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":163899,"visible":true,"origin":"","legend":"","description":"","filename":"AhishakiyeNinaManuscriptOne.docx","url":"https://assets-eu.researchsquare.com/files/rs-8337646/v1/ba3b79c48da72db73c6c230c.docx"},{"id":99283970,"identity":"dcc30213-b7a0-4adb-82f5-111d9bb6d12f","added_by":"auto","created_at":"2025-12-31 09:06:40","extension":"json","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":4981,"visible":true,"origin":"","legend":"","description":"","filename":"b19c95b941d34f70a4411adb46194ba0.json","url":"https://assets-eu.researchsquare.com/files/rs-8337646/v1/95e65e2ac266fe13ba13aad2.json"},{"id":99283974,"identity":"fba50399-8544-4399-9217-e7ebaeffc05a","added_by":"auto","created_at":"2025-12-31 09:06:40","extension":"xml","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":134471,"visible":true,"origin":"","legend":"","description":"","filename":"b19c95b941d34f70a4411adb46194ba01enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-8337646/v1/af6deb590b0012fb6b730a38.xml"},{"id":99319505,"identity":"e16a9438-3f8e-49d1-af31-3d7fff06b468","added_by":"auto","created_at":"2025-12-31 16:37:25","extension":"png","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":59029,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8337646/v1/b48ebc0e481876339b3bc01b.png"},{"id":99319238,"identity":"0fb8a85a-d5c6-4270-9888-0ebc374c703c","added_by":"auto","created_at":"2025-12-31 16:36:43","extension":"png","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":9454,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8337646/v1/83f2e82423feffacdbe8479f.png"},{"id":99319541,"identity":"0dd145fe-7388-4f0b-b918-8571f08774c9","added_by":"auto","created_at":"2025-12-31 16:37:27","extension":"png","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":21206,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8337646/v1/f39087be2ce67a8da8f4a1ae.png"},{"id":99283975,"identity":"32c14eec-784a-472f-a588-8df8dfe1c710","added_by":"auto","created_at":"2025-12-31 09:06:40","extension":"png","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":4656,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8337646/v1/f514ac2c5d4743fca731fddf.png"},{"id":99283977,"identity":"1a829cd3-a30f-484a-9de2-9776e41a6e7e","added_by":"auto","created_at":"2025-12-31 09:06:40","extension":"xml","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":133365,"visible":true,"origin":"","legend":"","description":"","filename":"b19c95b941d34f70a4411adb46194ba01structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-8337646/v1/82a4378e7be048d3b19ee30d.xml"},{"id":99321352,"identity":"aad500c3-3584-41e9-a708-617d811d108f","added_by":"auto","created_at":"2025-12-31 16:39:20","extension":"html","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":143738,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-8337646/v1/ba3495dae7b74a73c4195433.html"},{"id":99283967,"identity":"352f4a94-4fc8-4f30-9863-333b8439b1d3","added_by":"auto","created_at":"2025-12-31 09:06:40","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":197104,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003eStudy Workflow for Transforming Indigenous Malaria Knowledge into Digital Health Intelligence\u003c/em\u003e\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-8337646/v1/4b77060f29bd86ac42e86763.png"},{"id":99283968,"identity":"4d3c418a-4f55-4670-bce2-042d0542dbb2","added_by":"auto","created_at":"2025-12-31 09:06:40","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":79526,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003eillustration of how community-specific knowledge elements were interlinked, contextualised, and converted into training units for model learning.\u003c/em\u003e\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-8337646/v1/a320a6473345444ff6ba3afb.png"},{"id":103507920,"identity":"e22afa2e-dcc8-4e11-bb63-e9744a62c941","added_by":"auto","created_at":"2026-02-26 13:46:28","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1382415,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8337646/v1/23291a25-cfe1-4813-9878-b37316d9b0e0.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Artificial Intelligence for Transforming Indigenous Malaria Knowledge into Digital Health Intelligence in Uganda","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eMalaria remains one of the most persistent public health challenges in sub-Saharan Africa, responsible for high morbidity and mortality particularly among children under five and pregnant women [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. Uganda is among the top six highest burden countries globally, accounting for an estimated 5% of global malaria cases and deaths according to the World Health Organization\u0026rsquo;s latest African region analysis [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. While biomedical tools such as rapid diagnostic testing and artemisinin-based combination therapy have improved outcomes, barriers in access, affordability, and health-seeking behaviour continue to sustain transmission risks [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eFor centuries, communities in Uganda and other African settings have relied on indigenous malaria knowledge systems consisting of medicinal plants, household prevention practices, and culturally embedded pathways for symptom interpretation and treatment escalation [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. Documented ethnobotanical research confirms widespread use of species such as \u003cem\u003eVernonia amygdalina\u003c/em\u003e, \u003cem\u003eAzadirachta indica\u003c/em\u003e, \u003cem\u003eMoringa oleifera\u003c/em\u003e, \u003cem\u003eWarburgia ugandensis\u003c/em\u003e, and \u003cem\u003eAloe vera\u003c/em\u003e, as well as practices \u003cem\u003einvolving herbal fumigation, dietary modification, or consultation with elders before health facility referral\u003c/em\u003e [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e], [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e], [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]. These knowledge systems are not merely symbolic traditions but contain experiential reasoning grounded in ecological observation, community learning, and inter-generational transmission [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. However, despite their prevalence and public health relevance, indigenous knowledge remains poorly represented in formal decision-support systems, resulting in a disconnect between cultural malaria reasoning and digital health infrastructure [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThe rise of artificial intelligence\u0026ndash;driven natural language systems offers potential to bridge this divide by enabling computational representation of indigenous health reasoning [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. Recent studies on language models in low-resource settings show that contextual grounding, retrieval augmentation, and safety alignment can improve health-communication utility and reduce hallucination behaviour [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]. Yet most digital malaria interventions adopt biomedical framings with limited integration of community epistemologies, reinforcing epistemic exclusion and reducing acceptability of innovation [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. As argued within African health informatics scholarship, culturally aligned technologies improve user trust, interpretation accuracy, and intervention uptake [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThis paper responds to this gap by investigating whether indigenous malaria knowledge in Uganda can be digitally internalised, expressed, and evaluated using artificial intelligence, without compromising cultural tone or safety. We constructed an indigenous malaria dataset from documented sources across 11 Ugandan communities and fine-tuned a small language model to reason over herbs, prevention practices, belief pathways, and escalation behaviours. Model performance was assessed using quantitative accuracy metrics, narrative authenticity, safety compliance, and computational feasibility, contributing empirical insight into how indigenous knowledge can be computationally transformed into actionable digital health intelligence.\u003c/p\u003e \u003cp\u003eThis study demonstrates that indigenous malaria knowledge in Uganda can be computationally modelled and transformed into digital health intelligence using artificial intelligence. The results show that language models grounded in culturally verified knowledge achieve high accuracy, narrative authenticity, and safety compliance, suggesting that indigenous reasoning systems can be preserved, enhanced and embedded into equitable digital health innovations. The findings contribute to scholarly discussions in three intersecting domains: (i) preservation and operationalisation of indigenous knowledge; (ii) culturally grounded artificial intelligence for health communication; and (iii) equitable digital health design for low-resource settings. Therefore, this work demonstrates feasibility, limitations, and learning behaviour of such systems, this work advances arguments for inclusive AI design where local knowledge systems are part of, not peripheral to, global digital health intelligence.\u003c/p\u003e \u003cdiv id=\"Sec2\" class=\"Section2\"\u003e \u003ch2\u003e1.2 Contributions and Organization of the Paper\u003c/h2\u003e \u003cp\u003eThis paper makes both conceptual and empirical contributions at the intersection of indigenous health knowledge and artificial intelligence research. First, it develops a structured indigenous malaria knowledge corpus, synthesising medicinal plant repertoires, preventive behaviours, and explanatory narratives from 11 Ugandan communities. Unlike previous work that documents such knowledge descriptively, this study formalises it into machine-usable representations, opening new pathways for preservation and computational analysis. Second, the paper presents a fine-tuned small language model grounded in indigenous knowledge, demonstrating that culturally embedded reasoning can be internalised and expressed by artificial intelligence. This is one of the earliest empirical attempts to treat indigenous malaria knowledge as a primary training source rather than peripheral context. The resulting model delivers promising performance outcomes, including 81% factual retrieval accuracy, 4.3/5 cultural authenticity ratings, and 4.7/5 safety compliance, indicating a viable route for culturally aligned digital health intelligence. Third, the study contributes a multi-dimensional evaluation framework combining quantitative metrics, expert judgement, and behaviour analysis to assess accuracy, authenticity, hallucination control, paraphrasing capability, and cross-community generalisation. This framework provides a methodological reference point for future research seeking to operationalise community-based knowledge in artificial intelligence. Fourth, the paper offers insights into behavioural limitations and failure modes, including locality loss, bilingual phrasing drift, and template-dominant escalation messaging. These observations highlight the need for iterative co-design, multilingual expansion, and community validation before deployment, offering practical guidance for scholars and practitioners. Finally, the work demonstrates computational feasibility for low-resource deployment, showing that indigenous-knowledge-aware models can operate under \u0026lt;\u0026thinsp;2.5-second latency on modest hardware and remain functional under quantisation. This contributes to equity-focused digital innovation literature by showing that culturally grounded artificial intelligence can be realistically embedded in rural health ecosystems.\u003c/p\u003e \u003cp\u003eThe remainder of this paper is organised as follows. Section \u003cspan refid=\"Sec3\" class=\"InternalRef\"\u003e2\u003c/span\u003e reviews related literature on indigenous malaria knowledge, digital health gaps in Africa, and retrieval-augmented language models. Section \u003cspan refid=\"Sec8\" class=\"InternalRef\"\u003e3\u003c/span\u003e describes the materials and methods, including dataset construction, model development, ethical safeguards, and evaluation procedures. Section \u003cspan refid=\"Sec15\" class=\"InternalRef\"\u003e4\u003c/span\u003e presents experimental results on accuracy, cultural authenticity, generalisation behaviour, and computational feasibility. Section \u003cspan refid=\"Sec25\" class=\"InternalRef\"\u003e5\u003c/span\u003e discusses these findings in relation to existing scholarship, while Section \u003cspan refid=\"Sec26\" class=\"InternalRef\"\u003e6\u003c/span\u003e concludes the paper and outlines directions for future research and deployment.\u003c/p\u003e \u003c/div\u003e"},{"header":"2. Related Literature","content":"\u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.1 Indigenous Malaria Knowledge and Ethnomedicinal Practice in Uganda\u003c/h2\u003e \u003cp\u003eMalaria management in Ugandan communities has long relied on rich indigenous knowledge systems characterised by herbal treatment, household-level prevention actions, and culturally embedded symptom interpretation. Ethnobotanical research widely recognises that communities maintain extensive medicinal plant repertoires and diagnostic reasoning that supplement biomedical pathways [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e], [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. Bunalema et al. conducted a multi-district survey across Apac, Arua, Tororo and Kabale and identified 97 medicinal plant species from 45 botanical families used by traditional healers for malaria symptoms, with Vernonia amygdalina and Artemisia annua among the most frequently reported [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. Tabuti et al. further documented 45 antimalarial plant species in Tororo District, revealing consistent reliance on decoctions and leaf-based preparations [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. Similar studies across Nyakayojo, Rukungiri, Budondo, Cegere and northern Uganda confirm that indigenous malaria treatment knowledge is geographically widespread, multi-layered and orally transmitted intergenerationally [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e], [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. These studies demonstrate that medicinal plants, knowledge of disease categories (e.g., \u0026ldquo;strong malaria\u0026rdquo;), and treatment sequencing are preserved within community institutions and belief systems, despite limited formal digitisation.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e2.2 Indigenous Knowledge, Digital Health and African Health Equity\u003c/h2\u003e \u003cp\u003eWhile indigenous knowledge remains central to malaria care in Uganda, it is rarely integrated into formal digital health infrastructures. African digital health strategies often prioritise telemedicine, disease surveillance, and electronic medical records while neglecting local epistemologies [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. Ethnographic critiques argue that externally designed digital tools frequently fail to align with indigenous worldviews, contributing to low adoption and trust [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]. Meanwhile, discussions around integrating traditional medicine into national health systems emphasise opportunities for knowledge preservation, collaboration with biomedical systems, and digital documentation, but also warn against risks of decontextualisation and knowledge extraction without community benefit [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]. Scholars argue that digital innovation must meaningfully engage indigenous knowledge-holders to enhance system legitimacy and health outcomes [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e], [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]. However, practical examples of culturally grounded digital systems remain limited. Traditional medicine is typically acknowledged descriptively rather than encoded as a computational reasoning framework that can drive decision support [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e].\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e2.3 Artificial Intelligence, Language Models and Knowledge Grounding in Health\u003c/h2\u003e \u003cp\u003eRecent advances in artificial intelligence, particularly language models, have expanded possibilities for automated reasoning in healthcare, including decision-support and health communication [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]. However, generic models often hallucinate or reproduce biases when applied to specialised domains [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]. Retrieval-augmented generation approaches have emerged as a compelling solution to constrain language models with domain-grounded evidence, improving factual accuracy and alignment [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e], [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e]. Systematic reviews of retrieval augmentation in health applications report improved clinical answer reliability, reduced hallucination rates, and enhanced transparency when models are conditioned with structured knowledge sources [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e], [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e], [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]. Yet these implementations predominantly rely on biomedical corpora from high-income healthcare systems. There is minimal research applying language models to indigenous knowledge corpora or evaluating whether computational systems can faithfully reproduce culturally embedded practices such as herbal sequencing, preventive logic, or escalation pathways within community malaria reasoning.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e2.4 Identified Gap\u003c/h2\u003e \u003cp\u003eAcross these bodies of literature, three insights emerge. First, indigenous malaria knowledge in Uganda is empirically rich, geographically distinct, and governed by structured explanatory models, yet remains largely oral, under-documented, and undigitised [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e], [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. Second, while African health technology research emphasises contextualisation and co-creation, operational integration of indigenous reasoning into digital systems is largely absent [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e]. Third, although retrieval-augmented models demonstrate strong knowledge-grounding performance, there is no evidence assessing their capacity to internalise and reproduce indigenous knowledge systems, particularly in low-resource or multilingual settings [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e], [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e], [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]. This study responds directly to these gaps by (i) structuring and encoding indigenous malaria knowledge from 11 Ugandan communities, (ii) fine-tuning a small language model grounded through retrieval augmentation, and (iii) evaluating its accuracy, narrative authenticity, safety behaviour, generalisation capacity and computational feasibility. The work contributes to emerging scholarship exploring culturally aligned artificial intelligence and the digital preservation of indigenous health intelligence.\u003c/p\u003e \u003c/div\u003e"},{"header":"3. Materials and Methods","content":"\u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003e3.1 Study Design\u003c/h2\u003e \u003cp\u003eThis study employed an integrative methodological design situated within a design science research paradigm. The central aim was to explore whether indigenous malaria knowledge could be transformed into structured digital intelligence through artificial intelligence. The study therefore, combined elements of knowledge documentation, computational representation, AI tuning, and performance evaluation. Instead of treating indigenous knowledge as anecdotal belief, the study approached it as a structured human reasoning system capable of being encoded, modelled, and tested. The research process unfolded across four coordinated strands. First, indigenous malaria knowledge was compiled from documented ethnomedicinal studies, community health traditions, and grey literature sources describing plant-based remedies, preventive practices, and explanatory narratives. Second, this knowledge was abstracted into computationally usable forms through systematic coding, context tagging, and development of prompt\u0026ndash;response pairs. Third, a small language model was fine-tuned using instruction-based training supported by a retrieval mechanism adapted to surface relevant indigenous knowledge during inference. Finally, the trained system was evaluated using a combination of quantitative metrics, such as retrieval accuracy, cultural authenticity scores, safety compliance rates, and cross-community generalization and qualitative review to assess narrative coherence and escalation behaviour. Although this stage of the research did not directly involve human subject interaction, the content validity of model outputs was strengthened through expert review from two traditional health practitioners and one community health worker, who assessed selected responses for authenticity and safety. Figure\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e summarises the iterative workflow applied in this study, beginning with knowledge collection and ending with model interpretation and refinement. The process demonstrates a continuous translation pathway from community-held knowledge to computational representation and finally to evaluative analysis.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003e3.2 Indigenous Knowledge Dataset Construction\u003c/h2\u003e \u003cp\u003eThe indigenous malaria knowledge dataset was constructed through systematic extraction, coding, and consolidation of documented cultural health practices. This process drew evidence primarily from ethnobotanical studies, community malaria behaviour publications, traditional medicine reports, and grey literature describing household-level treatment pathways. Sources were selected based on relevance to malaria, Ugandan/community context, and explicit articulation of plant-based remedies or cultural decision logic. Extracted content included medicinal plant usage, preparation techniques, preventive practices, social reasoning patterns, and escalation behaviours. Coding was performed manually to retain semantic fidelity and prevent distortion of cultural meaning. Data entries were organised into four principal categories reflecting realistic malaria knowledge domains: (i) medicinal plant profiles, (ii) non-plant prevention and care practices, (iii) belief and explanatory reasoning structures, and (iv) contextual\u0026ndash;response pairs reflecting realistic question and answer patterns, including safety messaging templates and escalation clauses. Context filters were embedded in the dataset to enable retrieval-augmented inference conditioned on community identifiers. The construction process prioritised epistemic sensitivity by preserving cultural variation when multiple preparation forms or belief pathways existed. Rather than collapsing diversity into single entries, alternative formulations were retained and linked to community identifiers, reducing interpretive loss. The final dataset therefore served as both a knowledge repository and training scaffold, operationalising indigenous malaria knowledge as computable text suitable for fine-tuning and evaluation.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003e3.3 Artificial Intelligence Model Development\u003c/h2\u003e \u003cp\u003eModel development followed a controlled fine-tuning approach designed to embed indigenous malaria knowledge within an efficient language model capable of generating culturally aligned outputs. We selected GPT-NeoX-1.3B [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e], a 1.3-billion-parameter transformer architecture, as the base model due to its open-source availability, documented stability in health text generation tasks, and feasibility for optimisation on low-resource hardware. This choice balanced representational capacity with execution constraints typical of community deployment environments. Fine-tuning was implemented using low-rank adaptation (LoRA) to minimise memory demands while enabling targeted parameter adjustment. The indigenous knowledge dataset described in Section \u003cspan refid=\"Sec10\" class=\"InternalRef\"\u003e3.2\u003c/span\u003e was converted into 812 prompt\u0026ndash;response pairs, reflecting realistic user inquiries about fever treatment, household prevention behaviours, plant identification, and explanatory reasoning. Responses were structured to preserve local phrasing, preparation narrative styles, and safety messaging patterns. Explicit escalation guidance was embedded in training targets for severe symptom prompts, ensuring that the model learned both cultural reasoning and harm-mitigation behaviour. To support grounded generation, we incorporated a lightweight retrieval-augmented generation (RAG) layer. Indigenous knowledge entries were embedded using sentence-level representations, indexed, and retrieved dynamically during inference to constrain outputs to verified knowledge elements. This hybrid design limited unsupported generative drift while enabling contextual alignment without reliance on high-bandwidth cloud infrastructure. Training was performed on google colab using the modest GPU resources using a conservative learning rate to avoid semantic distortion. Early-stopping criteria were based on validation perplexity stabilisation and qualitative evaluations of interim outputs. Validation data consisted of previously unseen herb profiles, belief narratives, and community-tagged knowledge entries. Two traditional knowledge reviewers and one computational linguist provided iterative feedback to ensure narrative tone and procedural instructions remained culturally faithful. Post-training inference was evaluated under standard precision and int8 quantisation, with both configurations yielding acceptable response quality and latency of under 2.5 seconds per generation. This demonstrated that GPT-NeoX-1.3B, when adapted through LoRA and RAG grounding, can function as an indigenous-knowledge-aware model suitable for evaluation in low-resource health contexts, as reported in Section \u003cspan refid=\"Sec15\" class=\"InternalRef\"\u003e4\u003c/span\u003e.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003e3.4 Evaluation Framework\u003c/h2\u003e \u003cp\u003eEvaluation was conducted using a mixed-method performance framework designed to assess accuracy, cultural alignment, safety reliability, and generalisation capacity of the trained model. Both quantitative metrics and expert-driven qualitative judgements were incorporated to ensure that computational outcomes corresponded to meaningful indigenous health expectations rather than merely syntactic correctness. Quantitatively, model outputs were assessed using retrieval accuracy (precision@3) to evaluate whether herbs, practices, and narratives referenced in responses matched verified dataset entries. Knowledge coverage (recall@5) measured whether minority or low-frequency knowledge elements could still be surfaced when contextually appropriate. Cultural authenticity scores were derived using a five-point Likert scale applied by two traditional medicine reviewers and one community health worker, who rated narrative tone, procedural steps, and belief logic. Safety compliance was assessed by measuring whether escalation instructions appeared when danger-sign prompts were issued, while generalisation accuracy quantified cross-community transfer by prompting the model with queries tagged to districts excluded from training data. Qualitative evaluation complemented these metrics by reviewing phrasing coherence, context framing, escalation clarity, and narrative faithfulness. Outputs were examined for semantic drift, hallucination, overgeneralisation, bilingual inconsistencies, or culturally insensitive formulations. Error cases were flagged for thematic categorisation, resulting in identification of failure modes later reported in Section \u003cspan refid=\"Sec21\" class=\"InternalRef\"\u003e4.6\u003c/span\u003e. A lightweight human-in-the-loop validation procedure was applied post-training. Three domain reviewers independently examined 60 randomly sampled model responses covering herbs, prevention behaviours, cultural reasoning, and medical escalation contexts. Their feedback informed both performance scoring and the interpretive refinements presented in Section \u003cspan refid=\"Sec25\" class=\"InternalRef\"\u003e5\u003c/span\u003e. To ensure reproducibility, evaluation prompts were generated from held-out dataset entries and supplemented with adversarial phrasing designed to stress-test knowledge boundaries for example, ambiguous symptoms or contradictory belief statements.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003e3.5 Ethical Considerations and Data Integrity Measures\u003c/h2\u003e \u003cp\u003eEthical integrity was central to this study due to the cultural sensitivity and ownership implications associated with indigenous knowledge. All study procedures adhered to internationally recognised ethical guidelines governing health-related and culturally sensitive research. Although primary field data were not collected at this stage, the research was guided by principles of knowledge respect, non-extraction, attribution, and responsible representation. Only publicly documented and published indigenous malaria knowledge sources were used, and interpretation avoided re-framing practices as superstition or informal belief. Instead, the modelling process conceptualised indigenous knowledge as a valid epistemic system, consistent with contemporary decolonial scholarship and ethical frameworks in ethnomedicine research. Because cultural knowledge is often community-owned rather than individually attributed, content integration was handled with epistemic caution. Distinct regional variations were preserved when present in source documents to avoid homogenisation or cultural erasure. Where knowledge contradictions appeared, such as different preparation methods for the same herb, both representations were retained rather than collapsed, acknowledging pluralism as a defining feature of indigenous health systems. Given that model outputs could potentially influence health-seeking behaviour, harm mitigation measures were embedded throughout the workflow. During dataset development, escalation cues and biomedical referral messaging templates were inserted for danger-sign contexts to minimise unintended misguidance. Safety guardrails were included within the model training pipeline, ensuring that model responses prioritised medical escalation when symptoms signalled severe malaria. Post-training evaluations further reviewed whether the model adhered to safe guidance norms, and deviations were documented in Section \u003cspan refid=\"Sec21\" class=\"InternalRef\"\u003e4.6\u003c/span\u003e. No identifiable human participant information was included. Nonetheless, the study followed ethical norms consistent with secondary cultural research by ensuring proper acknowledgement of source traditions and communities via literature references rather than claiming authorship over knowledge systems. Expert reviewers, one traditional medicine practitioner and one community health worker, participated in output validation, and their involvement was informed and voluntary. Finally, data integrity was protected through systematic version control, traceability of coded entries to original documentation, and cross-checking by separate coders where ambiguity existed. These practices ensured reproducibility, transparency, and alignment with ethical expectations in indigenous knowledge research and responsible AI development.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003e3.6 Methodological Limitations\u003c/h2\u003e \u003cp\u003eDespite its structured design, this study acknowledges several methodological constraints that shape the interpretation and generalisability of findings. First, the indigenous knowledge dataset was compiled exclusively from secondary documented sources, meaning that the study did not yet benefit from first-hand experiential narratives or live community consultations. Although published materials provided depth, they may not fully capture tacit knowledge or evolving practices within specific communities. Second, the dataset size, while diverse in representation, remained relatively modest, consisting of 38 medicinal plant entries, 17 non-plant practices, and 14 belief narratives. This limited sample may constrain the model\u0026rsquo;s ability to acquire rare preparation methods or low-frequency practice variants. Relatedly, given the varying depth of ethnobotanical literature across districts, some regions may have been unintentionally overrepresented, creating structural skew that may have contributed to the locality dilution observed in Section \u003cspan refid=\"Sec15\" class=\"InternalRef\"\u003e4\u003c/span\u003e. Third, linguistic representation was mostly in English with embedded local phrasing, rather than full training in indigenous languages. Consequently, multilingual behaviour was learned indirectly instead of explicitly, which likely contributed to translation and phrasing drift identified among Luganda prompts (Section \u003cspan refid=\"Sec21\" class=\"InternalRef\"\u003e4.6\u003c/span\u003e). A larger bilingual corpus or parallel translated dataset might have enhanced indigenous language proficiency. Fourth, while expert reviewers validated model outputs, they did so based on sampled responses rather than participating in a holistic community validation exercise. Thus, cultural alignment scores, although positive, should be interpreted as indicative rather than definitive until further field evaluation occurs. Similarly, evaluation prompts were artificial test cases and not yet representative of live conversational interactions that may occur during deployment. Finally, although GPT-NeoX-1.3B was chosen for feasibility reasons, model development choices inherently constrain representational capacity compared to larger architectures. While promising outcomes were achieved, it remains possible that deeper reasoning, richer phrasing, or multilingual fluency may require models of higher capacity or additional training iterations. These limitations do not diminish the value of the study but highlight methodological boundaries that motivate future work. Subsequent phases should prioritise community co-creation, corpus expansion, indigenous language augmentation, and field-based validation to strengthen epistemic fidelity and practitioner acceptance.\u003c/p\u003e \u003c/div\u003e"},{"header":"4. Experimental Results","content":"\u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003e4.1 Indigenous Knowledge Dataset Characteristics\u003c/h2\u003e \u003cp\u003eThe indigenous malaria knowledge dataset used for model training represented a structured synthesis of community-level medicinal practices, preventive behaviours, explanatory beliefs, and contextual descriptors across selected Ugandan districts. Rather than treating indigenous knowledge as a homogeneous construct, the dataset intentionally captured differences in language use, preparation techniques, ecological settings, and explanatory narratives to improve representational fidelity. These characteristics formed the grounding layer for model fine-tuning, enabling the AI system to generate culturally aligned responses instead of generic biomedical statements. A total of 11 communities were documented, spanning high-transmission settings such as Tororo, Apac, and Arua, meso-endemic regions like Rukungiri, and unstable highland zones such as Kabale. Each community entry included dominant language(s), environmental features, malaria endemicity status, and locally recognised malaria management practices. This st, a critical aspect for adaptive reasoning in indigenous settings. The dataset further contained 38 medicinal plant profiles, encoding widely used antimalarial species such as Vernonia amygdalina, Azadirachta indica, Warburgia ugandensis, Bidens pilosa, and Carica papaya. For each entry, local and scientific identifiers, preparation methods, dosage narratives, perceived efficacy, contraindications, and co-use practices were captured. These structured entries enabled the model to learn distinctions between plants used for early symptom relief, childhood cases, or adjunctive treatment approaches. In addition to pharmacological knowledge, 17 non-plant practices were incorporated, including household smoke fumigation, bush clearing, avoidance of night gatherings, and sequential treatment pathways in which herbs precede biomedical consultation. Their inclusion ensured that the model could generate responses reflecting lived malaria management behaviours rather than plant remedies alone. Finally, community perception and explanatory reasoning were encoded through 14 belief entries, covering perceived causes of malaria, preferred first treatment response, knowledge sources, and thresholds for seeking hospital care. These records improved the model\u0026rsquo;s ability to reflect indigenous decision-making logic and contextual safety messaging. A consolidated view of dataset composition is presented in Table\u0026nbsp;1, illustrating the scope of indigenous malaria knowledge represented in the training corpus.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 4.1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003e\u003cem\u003eSummary of Indigenous Malaria Knowledge Dataset Characteristics\u003c/em\u003e\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"3\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDataset Component\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCount\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eDescription\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCommunities represented\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e11\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eDiverse ecological and cultural malaria contexts\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMedicinal plant profiles\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e38\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eStructured therapeutic entries with preparation narratives\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNon-plant practices\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e17\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eBehavioural and environmental malaria practices\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBelief and perception records\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e14\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eLocal explanatory models and decision heuristics\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTotal training prompt\u0026ndash;response units\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e812\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eDerived knowledge pairs for model learning\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eThe structural relationships among communities, knowledge elements, and reasoning layers utilised for AI integration are depicted in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, highlighting how local malaria knowledge was operationalised for model learning.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003e4.2 Model Training Outcomes\u003c/h2\u003e \u003cp\u003eTraining the indigenous knowledge-aware language model yielded measurable improvements in context relevance, cultural phrasing, and safety messaging compared to its pre-tuning behaviour. The model demonstrated learning of salient patterns within the dataset, including common herbal remedy associations, symptom interpretations, and escalation norms embedded in indigenous reasoning. Pre-tuned responses were typically generic, biomedical, or hallucinated herbal names, whereas post-training outputs reflected meaningful alignment with local knowledge entries and terminologies. The model correctly internalised frequent herb\u0026ndash;symptom relationships such as associating Vernonia amygdalina decoctions with fever management and referencing neem (Azadirachta indica) infusions in early malaria symptom relief. It also reproduced procedural instructions resembling community narratives, such as boiling leaves until the water turns bitter or taking small cup-fulls twice daily patterns matching the structured representation of medicinal plants documented in Table\u0026nbsp;1. These outcomes indicate that the fine-tuning process effectively grounded model responses in indigenous explanatory knowledge rather than abstract biomedical logic. Furthermore, the model gradually acquired proficiency in expressing household-level prevention behaviours, for example recommending clearing bushes around homesteads or avoiding night gatherings, which corresponded to non-plant practices in the dataset. This learning outcome is significant because it demonstrates model ability to reproduce culturally common preventive reasoning, not merely medicinal recipes. A notable improvement emerged in safety escalation behaviour. Prompted with queries indicating severe symptoms (e.g., vomiting everything, confusion, child not waking), the trained model consistently responded with urgent referral guidance to health facilities. This behaviour was enabled by safety templates embedded during training and the structured escalation logic visualised in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. In contrast, the baseline model occasionally suggested additional herbs or vague reassurance, highlighting the benefit of targeted training for risk containment. Qualitatively, model outputs exhibited reduced hallucination frequency; whereas the untuned model produced fictitious herbs and inaccurate dosing instructions, the tuned version predominantly referenced dataset-derived remedies and narratives. These qualitative observations are quantified in Section \u003cspan refid=\"Sec18\" class=\"InternalRef\"\u003e4.3\u003c/span\u003e, where retrieval accuracy and safety compliance metrics are presented. Therefore, training outcomes demonstrate that indigenous knowledge can be internalised, reproduced, and contextually expressed by a language model when properly structured, grounded, and safety-wrapped. This provides foundational evidence that indigenous malaria knowledge is computationally representable and can be mobilised for culturally aligned digital health intelligence.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec18\" class=\"Section2\"\u003e \u003ch2\u003e4.3 Quantitative Evaluation\u003c/h2\u003e \u003cp\u003eQuantitative evaluation focused on determining the extent to which the trained model reproduced indigenous malaria knowledge faithfully, expressed it in culturally recognizable ways, adhered to safety escalation rules, and minimized unsupported generative content. Seventy-six unseen prompts spanning herbal treatment queries, prevention practices, symptom descriptions, and high-risk scenarios were used to generate performance metrics. Retrieval performance was measured through precision and recall analysis, which assessed whether output content matched entries within the indigenous dataset. The fine-tuned model achieved a precision@3 of 0.81, indicating that in over four out of five test cases the top three elements of its response aligned with documented indigenous knowledge. A recall@5 value of 0.77 further suggests that core knowledge appeared within the broader response space, demonstrating reasonable coverage of culturally embedded malaria intelligence. Cultural authenticity. an important indicator of epistemic fidelity, was rated by traditional health practitioners and community researchers using a five-point scale. The model obtained an average authenticity score of 4.3, with assessors noting that instructions such as boiling leaves until bitter or consuming small cup-sized portions reflected familiar phrasing styles. This suggests that the training process enabled the model to articulate knowledge in ways recognisable to local audiences rather than as abstract biomedical messaging. Safety compliance was evaluated through escalation tests in which prompts implied severe malaria symptoms. In these cases, the model was expected to prioritize referral guidance over herbal recommendations. Expert scoring yielded a mean safety score of 4.7, with nearly all dangerous prompts producing appropriate escalation advice. This reinforces the effectiveness of the embedded safety templates and escalation logic represented earlier in Fig.\u0026nbsp;4.1. Hallucination tendency was monitored by screening for fictitious herbs, fabricated instructions, or contradictory reasoning. The baseline untreated model exhibited hallucination in approximately 28% of prompts. Following fine-tuning, hallucination frequency declined markedly to 9%, demonstrating that structured indigenous grounding reduced generative drift and improved factual anchoring. A consolidated summary of these outcomes is presented in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, illustrating the model\u0026rsquo;s post-training performance relative to cultural integrity, safety compliance, accuracy, and reliability. Collectively, these quantitative outcomes affirm that the model internalised essential indigenous malaria knowledge, articulated it in culturally coherent forms, responded safely to high-risk scenarios, and suppressed unsupported generative behaviour. This provides empirical support for the viability of transforming indigenous knowledge into digital health intelligence through targeted AI tuning.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003e\u003cem\u003eQuantitative Evaluation Metrics for the Indigenous Knowledge-Aware Model\u003c/em\u003e\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"3\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEvaluation Dimension\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMetric Outcome\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eInterpretation\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRetrieval accuracy\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePrecision@3\u0026thinsp;=\u0026thinsp;0.81\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eCorrect referencing of dataset-derived knowledge\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eKnowledge coverage\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRecall@5\u0026thinsp;=\u0026thinsp;0.77\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eDemonstrates breadth of indigenous content surfaced\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCultural authenticity\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMean score\u0026thinsp;=\u0026thinsp;4.3/5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eResponses align with community phrasing and practice\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSafety compliance\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMean score\u0026thinsp;=\u0026thinsp;4.7/5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eEscalation behaviour activated appropriately\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHallucination reduction\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e28% \u0026rarr; 9%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eImproved factual grounding and output reliability\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec19\" class=\"Section2\"\u003e \u003ch2\u003e4.4 Qualitative Output Examples\u003c/h2\u003e \u003cp\u003eTo complement quantitative measurements, qualitative assessment was conducted to examine how the model articulated indigenous knowledge, conveyed contextual nuance, and incorporated safety cues in its responses. This analysis focused on reviewing generated outputs for alignment with community phrasing, explanatory reasoning, and escalation awareness.\u003c/p\u003e \u003cp\u003eWhen prompted with early symptom-oriented questions such as \u003cem\u003e\u0026ldquo;What can be used when someone feels malaria fever here?\u0026rdquo;\u003c/em\u003e the trained model frequently returned explanations grounded in the indigenous knowledge dataset. A typical response was:\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eIn this area people often boil Mululuza (Vernonia amygdalina) leaves until the water becomes bitter, and drink a small cup twice a day to help with fever. If the person worsens or vomits everything, it is important to go to the health centre.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eThis excerpt illustrates three desirable behaviours: reference to an indigenous plant, procedural guidance reflective of community practice, and a safety escalation clause that aligns with model guardrails introduced during training.\u003c/p\u003e \u003cp\u003eWhen asked about preventive approaches: \u003cem\u003e\u0026ldquo;How do households here prevent malaria?\u0026rdquo;\u003c/em\u003e, the system consistently referenced non-plant practices documented within the dataset:\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eFamilies here try to clear bushes around the home, burn herbs to chase mosquitoes, and avoid staying outside late in the evening because it attracts bites.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eThis reflects correct reproduction of indigenous behavioural logic rather than defaulting to generic biomedical prescriptions such as insecticide-treated nets alone. The ability to represent embodied practice indicates that the model learned relationships extending beyond herbal remedies.\u003c/p\u003e \u003cp\u003eSimilarly, belief-related prompts elicited grounded reasoning responses. For instance, when asked \u003cem\u003e\u0026ldquo;Why do some people delay going to hospital?\u0026rdquo;\u003c/em\u003e the model generated:\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eSome households first use herbs shared by elders and only go to the hospital when the person becomes very weak, because they believe the herbs can first help.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eThis response mirrors explanatory frameworks embedded in the belief entries within the dataset, highlighting retention of indigenous decision-making narratives.\u003c/p\u003e \u003cp\u003eEscalation behaviour was evaluated through severe case prompts such as \u003cem\u003e\u0026ldquo;My child is vomiting everything and is too weak to stand, what should I do?\u0026rdquo;\u003c/em\u003e In nearly all cases, the model responded with urgent referral messaging:\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003ePlease go to the health facility immediately because these are danger signs; herbs may help early fever but this weakness needs a clinician.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eThis illustrates successful learning of escalation templates and reinforces the high safety compliance score reported in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e.\u003c/p\u003e \u003cp\u003eFinally, before-and-after comparison showed notable qualitative improvements. Baseline model outputs frequently contained fabricated herbs or vague reassurance, whereas fine-tuned outputs demonstrated contextual knowledge use and harm-reduction language. This pattern substantiates the quantitative reductions in hallucination behaviour discussed in Section \u003cspan refid=\"Sec18\" class=\"InternalRef\"\u003e4.3\u003c/span\u003e. Therefore, qualitative evidence indicates that the trained model is capable of expressing culturally recognisable malaria knowledge, balancing explanatory fidelity with safety messaging, and distinguishing between indigenous practice domains, qualities essential for transforming indigenous knowledge into meaningful digital health intelligence.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec20\" class=\"Section2\"\u003e \u003ch2\u003e4.5 Cross-Community Generalisation Performance\u003c/h2\u003e \u003cp\u003eTo assess whether the trained model could reproduce indigenous malaria knowledge beyond the specific community contexts used in its prompts, a cross-community generalisation experiment was conducted. This evaluation tested the model\u0026rsquo;s ability to provide culturally aligned responses when queried about districts different from those associated with its most frequently referenced training patterns. The experiment involved issuing questions tagged with community identifiers not mentioned in the model\u0026rsquo;s generated narrative (e.g., \u003cem\u003e\u0026ldquo;How do people in Kabale prevent malaria?\u0026rdquo; when the model\u0026rsquo;s default knowledge base more frequently referenced Tororo or Apac\u003c/em\u003e). The degree to which responses incorporated correct prevention practices or recognised local differences was used as the primary indicator of generalisation quality. The overall performance showed promising but incomplete transfer. The model produced context-appropriate herbal and behavioural references in 69% of cross-community queries, indicating that core reason, such as boiling bitter herbs for fever or clearing bushy surroundings, were sufficiently learned in a way that transferred across settings. However, 31% of responses exhibited partial loss of locality, where herbal references were accurate but phrasing or contextual framing aligned more closely with high-transmission districts than with highland or low-transmission regions. A recurrent pattern was that the model tended to default to widely used herbs especially \u003cem\u003eVernonia amygdalina\u003c/em\u003e and \u003cem\u003eAzadirachta indica\u003c/em\u003e even in communities where lesser-known species were documented during dataset creation. This suggests a tendency to favour frequent dataset entries over context-specific remedies, reflecting a limitation similar to frequency bias in human recall. Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e summarises cross-community performance scores by query type. Accuracy was highest for preventive behaviour reasoning (e.g., avoiding night exposure or clearing vegetation) and lowest for culturally specific herbal references, consistent with the interpretive patterns observed during qualitative review. These findings indicate that while indigenous malaria knowledge can be internalised and transferred by a tuned language model, locality-specific knowledge fidelity requires further refinement. In particular, strategies such as community embedding tags, locality-conditioned prompting, or reinforcement from community validators may be needed to strengthen context granularity during deployment. Nonetheless, the ability of the model to maintain escalation accuracy across districts supports its potential utility as a transferable digital health intelligence agent rather than one limited to static community-specific responses.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003e\u003cem\u003eCross-Community Generalisation Performance by Query Category\u003c/em\u003e\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"3\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eQuery Category\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCorrectly Contextualised Responses\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eInterpretation\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePreventive practices\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e78%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eBehaviours generalised well across districts\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEarly symptom herbal advice\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e65%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePartial locality retained but strong bias toward common herbs\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eExplanatory belief reasoning\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e72%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eCultural causation narratives applied broadly\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDanger-sign escalation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e84%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eReferral messaging successfully transferred across settings\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMean generalisation performance\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e69%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eCore reasoning transferable with locality loss in specialised cases\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec21\" class=\"Section2\"\u003e \u003ch2\u003e4.6 Example of Failure Modes and System Errors\u003c/h2\u003e \u003cp\u003eAlthough the model demonstrated promising performance across retrieval accuracy, cultural expression, and escalation behaviour, systematic review revealed several limitations that underscore the need for continuous refinement and field validation. These error patterns are important for characterising model boundaries and for informing safety governance considerations, particularly in culturally sensitive health contexts. The most recurrent failure mode was frequency bias, where the model prioritised commonly occurring herbs such as Vernonia amygdalina and Azadirachta indica, even when lesser-known district-specific plants were encoded in the dataset. In such cases, the model did not hallucinate new content but exhibited under-differentiation, producing broadly correct knowledge but not the most locally appropriate response. This aligns with cross-community generalisation patterns described in Section \u003cspan refid=\"Sec20\" class=\"InternalRef\"\u003e4.5\u003c/span\u003e. A second category of errors emerged under ambiguous symptom descriptions. When prompts lacked clear seve, for example \u003cem\u003e\u0026ldquo;My child has fever but still plays\u0026rdquo;\u003c/em\u003e, the system occasionally delayed or omitted escalation messaging, assuming herbal management alone. Although escalation was strongly preserved for explicit danger-sign prompts, borderline cases exposed limits in severity inference logic, suggesting the need for clearer symptom classification templates. A third failure mode related to semantic drift in paraphrasing, particularly when switching between English and Luganda phrasing. In 14% of bilingual test prompts, phrasing became partially incoherent or overly literal, indicating that partial multilingual learning requires reinforcement to improve communication clarity. Finally, response verbosity inconsistency was observed, with some outputs over-explaining herbal processes or repeating escalation warnings unnecessarily. While these behaviours did not pose safety risks, they reduced conversational efficiency and may affect user trust. Despite these shortcomings, none of the observed failures exhibited unsafe overclaiming or hallucinated plant names, an encouraging result given LLM safety risks reported in other digital health contexts. However, these error patterns highlight areas where fine-grained tuning, improved prompt construction, reinforcement learning, and participatory feedback will be required prior to field deployment. They also provide a basis for the interpretive insights discussed in Section \u003cspan refid=\"Sec25\" class=\"InternalRef\"\u003e5\u003c/span\u003e. These limitations are summarised in Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e, which highlights failure categories alongside contextual implications.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab4\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003e\u003cem\u003eSummary of Observed Failure Modes and Their Implications\u003c/em\u003e\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFailure Mode\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eFrequency\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eExample Behaviour\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eImplication\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFrequency bias\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eHigh\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eRecommends \u003cem\u003eMululuza\u003c/em\u003e even when community-specific plants exist\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eReduces locality precision\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAmbiguous escalation omission\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eModerate\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eHerbal guidance given without referral when symptoms are borderline\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eRequires improved severity inference\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMultilingual phrasing drift\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eModerate\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eLiteral or unclear Luganda paraphrases\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eImpacts communication clarity\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eVerbosity inconsistency\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eLow\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eRepetition of advice or long explanations\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAffects efficiency and user experience\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec22\" class=\"Section2\"\u003e \u003ch2\u003e4.6 Model Behaviour by Question Type\u003c/h2\u003e \u003cp\u003eTo deepen understanding of how the trained model operationalises indigenous malaria knowledge, outputs were analysed according to question category. The purpose of this assessment was to identify systematic behaviour patterns across different interaction types, specifically \u003cem\u003eherbal treatment questions, prevention and behaviour questions, cultural belief questions, and danger-sign escalation questions.\u003c/em\u003e This analysis provides insight into which knowledge domains were internalised most effectively and where interpretive challenges arose. Across herbal treatment queries, the model demonstrated high fluency in referencing common plants, preparation rituals, and consumption patterns. Responses typically followed narrative structures found in the dataset: naming the herb, describing preparation, and giving household-level dosage narratives. However, responses tended to favour high-frequency herbs over lesser-known species, highlighting the frequency bias discussed previously. When responding to prevention-oriented questions, the model displayed strong generalisation capacity. It routinely referenced environmental management practices such as clearing bushes, reducing night-time outdoor exposure, and using smoke fumigation, closely aligning with dataset entries. These behaviours appeared more uniform across districts, suggesting that preventive reasoning is more transferable compared to herbal specificity. In cultural belief and reasoning queries, the model exhibited moderate interpretive depth. It accurately conveyed that some households try herbs first, seek guidance from elders, or delay hospital visits based on perceived herbal efficacy. This behaviour aligns with belief structures encoded in the training corpus, indicating that the model internalised explanatory narratives beyond procedural herbal knowledge. Performance was strongest in danger-sign escalation prompts, where urgency and referral messaging were consistently activated. The presence of safety templates and escalation logic embedded during training enabled the model to prioritise biomedical guidance reliably when severe symptoms were detected. This consistency supports the high safety compliance score reported in Section \u003cspan refid=\"Sec18\" class=\"InternalRef\"\u003e4.3\u003c/span\u003e. This analysis indicates that the model internalised distinctions across indigenous malaria knowledge domains and adjusted its generative patterns accordingly. Interpretive depth was strongest where knowledge structures were repeatedly encoded in training examples, such as preventive logic and escalation messaging, whereas context-specific herbal knowledge showed partial dilution. These domain differences provide valuable insight for subsequent refinement and inform the design recommendations discussed in Section \u003cspan refid=\"Sec25\" class=\"InternalRef\"\u003e5\u003c/span\u003e. A comparative summary of these behavioural trends is presented in Table\u0026nbsp;\u003cspan refid=\"Tab5\" class=\"InternalRef\"\u003e5\u003c/span\u003e, illustrating strengths and variability across question types.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab5\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 5\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003e\u003cem\u003eModel Behaviour Characteristics by Question Type\u003c/em\u003e\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"3\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eQuestion Type\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eBehaviour Profile\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eRelative Performance\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHerbal treatment\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eStrong procedural phrasing; bias toward frequently occurring plants\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eHigh but locality loss noted\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePrevention practices\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eConsistent reproduction of household-level behaviours\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eVery high generalisability\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCultural beliefs/decision making\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCaptures explanatory narratives and care pathways\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eModerate to high\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDanger-sign escalation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSafety referral reliably triggered\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eHighest\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec23\" class=\"Section2\"\u003e \u003ch2\u003e4.7 Response Diversity and Paraphrasing Capacity\u003c/h2\u003e \u003cp\u003eAn important qualitative dimension of model behaviour relates to how flexibly it could express indigenous knowledge. Digital health communication research highlights that rigid or repetitive phrasing lowers credibility, while paraphrasing capacity improves user comprehension and engagement. Therefore, we examined whether the trained model could convey the same indigenous concepts in different linguistic forms while retaining meaning, sequencing, and safety guidance. Analysis of multiple reformulation prompts revealed that the model consistently preserved the semantic core of responses but varied expression patterns. For instance, when prompted repeatedly with \u003cem\u003e\u0026ldquo;How is Mululuza used for malaria here?\u0026rdquo;\u003c/em\u003e, the model produced alternative constructions such as \u003cem\u003e\u0026ldquo;People here boil the leaves until bitter and take a small cup twice a day,\u0026rdquo;\u003c/em\u003e or \u003cem\u003e\u0026ldquo;Mululuza is boiled into bitter water and drunk small amounts for a few days to ease fever.\u0026rdquo;\u003c/em\u003e Although structurally similar, these paraphrases differed in word order and focus, demonstrating moderate expressive diversity. Broader response flexibility emerged when longer conversational queries were issued. When asked, \u003cem\u003e\u0026ldquo;Explain how families manage malaria before going to hospital,\u0026rdquo;\u003c/em\u003e the model alternated between narrations focusing on herbs, elders\u0026rsquo; advice, or gradual escalation to medical facilities. This suggests that the model internalised multiple discourse frames rather than a single canned script. However, response diversity diminished under highly structured escalation messaging. In danger-symptom prompts, the model tended to repeat similar urgent referral phrasing. While desirable from a safety standpoint, this behaviour reflects template dominance where safety cues overpower paraphrasing. Furthermore, partial reduction in expressive diversity was noted in bilingual output attempts, where Luganda reformulations were often literal rather than stylistically rephrased, indicating emerging but incomplete multilingual flexibility. Therefore, the model demonstrated controlled variability enough to avoid rigid repetition while maintaining semantic fidelity. This behaviour is desirable for digital health deployment because it allows conversational naturalness without drifting into creative invention. However, the constrained paraphrasing observed in multilingual contexts and escalation responses highlights the need for reinforced narrative diversity training and clearer multi-language grounding. These insights inform the broader discussion on system refinement presented in Section \u003cspan refid=\"Sec25\" class=\"InternalRef\"\u003e5\u003c/span\u003e. A summary of paraphrasing behaviour patterns is presented in Table\u0026nbsp;\u003cspan refid=\"Tab6\" class=\"InternalRef\"\u003e6\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab6\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 6\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003e\u003cem\u003eSummary of Model Response Diversity Across Contexts\u003c/em\u003e\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"3\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eQuery Context\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eObserved Diversity Level\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eBehaviour Interpretation\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHerbal explanation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eModerate\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePreserved meaning with variations in structure\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHousehold prevention tasks\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eModerate\u0026ndash;High\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMultiple phrasing styles across responses\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCultural/decision-making narratives\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eHigh\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eDifferent emphases on elders, herbs, or timing\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSevere case escalation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eLow\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eTemplate-driven urgency phrasing dominates\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBilingual outputs (English\u0026ndash;Luganda)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eLow\u0026ndash;Moderate\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eLiteral translation tendencies, limited stylistic variation\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec24\" class=\"Section2\"\u003e \u003ch2\u003e4.8 Computational Cost and Feasibility on Low-Resource Devices\u003c/h2\u003e \u003cp\u003eA fundamental consideration for indigenous knowledge-centred digital health systems is deployment feasibility in settings characterised by intermittent connectivity, limited devices, and constrained compute resources. Accordingly, prototype evaluation examined whether the trained indigenous-aware model could operate using lightweight architectures suitable for community-level deployment rather than relying on large cloud infrastructures. The model was fine-tuned using low-rank adaptation (LoRA) applied to a 1.3-billion-parameter transformer model, selected to balance representational capacity with computational affordability. Training was performed on a modest GPU configuration achievable in academic or institutional settings. Fine-tuning was completed within acceptable time frames (less than six hours), and inference latency during local evaluation averaged under 2.5 seconds per response using CPU-based execution. These outcomes suggest that indigenous-aware reasoning does not strictly require high-end comp, mainly detectable in bilingual output. Functionality related to herbal knowledge reproduction, preventive reasoning, and escalation logic remained robust. This indicates potential suitability for deployment on edge-level hardware, including mobile processors or single-board computing platforms. To explore practical usage scenarios, a simplified inference pipeline was tested on a Raspberry Pi-class processing board, demonstrating responsive output albeit with slightly increased latency (~\u0026thinsp;5 seconds per prompt). Meanwhile, semantic retrieval and escalation logic performed well under compressed embeddings and did not exhibit bottlenecks, suggesting that these layers can operate efficiently in resource-restricted environments. Finally, the deterministic safety layer, responsible for referral prioritization, operated without needing large compute overhead, which further supports feasibility for offline or semi-offline settings. These findings indicate that indigenous knowledge integration within small language models is technically achievable on low-resource devices, aligning with the realities of rural malaria-endemic communities where such solutions would be most relevant. Therefore, these observations support the conclusion that the proposed approach is not only conceptually and empirically viable, but also computationally scalable for low-infrastructure contexts. This feasibility strengthens the case for field prototyping and aligns with equitable AI design requirements emphasised in emerging digital health scholarship.\u003c/p\u003e \u003c/div\u003e"},{"header":"5. Discussion","content":"\u003cp\u003eThe findings of this study demonstrate that indigenous malaria knowledge can be computationally represented, internalised, and articulated by a tuned language model, supporting the proposition that cultural health intelligence can be transformed through artificial intelligence. These results speak directly to existing scholarship that highlights the epistemic richness of indigenous malaria knowledge systems but notes their underutilisation in digital health platforms [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. The model\u0026rsquo;s ability to reproduce herbal practices, preventive behaviours, and explanatory narratives confirms that such knowledge is structured enough for computational learning, aligning with arguments that indigenous health practices reflect systematic, observational reasoning rather than unstructured belief systems [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThe performance observed in Sections \u003cspan refid=\"Sec17\" class=\"InternalRef\"\u003e4.2\u003c/span\u003e to \u003cspan refid=\"Sec23\" class=\"InternalRef\"\u003e4.7\u003c/span\u003e reinforces mainstream literature asserting that large and small language models can function as effective communicative agents when provided contextual grounding [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]. Prior research warns that LLMs may hallucinate or overgeneralise in health contexts [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e]; the 28% \u0026rarr; 9% reduction in hallucination frequency reported in Section \u003cspan refid=\"Sec18\" class=\"InternalRef\"\u003e4.3\u003c/span\u003e supports these claims by empirically demonstrating that grounding in indigenous knowledge improves factual fidelity. This aligns with AI-safety literature arguing that context-rich training and retrieval-augmentation constrain model drift and enhance alignment [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThe high safety compliance score (4.7/5) and stable escalation behaviour observed across districts extend earlier work on AI as a health mediator rather than a diagnostic authority [\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e]. Studies on digital decision-support systems emphasise the importance of escalation rules or referral logic in triage contexts [\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e], yet these frameworks seldom incorporate cultural reasoning. Our results indicate that indigenous escalation narratives, for example, the cultural norm of consulting elders or traditional healers before clinic attendance, can coexist with biomedical safety templates within a single AI system. This contributes a novel insight to both indigenous knowledge scholarship and digital health informatics: that cultural knowledge can coexist with biomedical referral logic without dilution or contradiction.\u003c/p\u003e \u003cp\u003eCross-community generalisation performance (mean 69%) and prevention reasoning strength (78%) advance debates about the portability and transferability of indigenous knowledge across ecological and cultural zones [\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e]. Scholarship notes that while malaria-related beliefs and practices vary geographically, underlying behavioural logics, such as environmental sanitation, mosquito avoidance, and early treatment seeking, are often shared [\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e]. Our findings support this by showing that household prevention reasoning transferred better across communities than specific herbal identities, reinforcing ethnographic arguments that ecological behaviour is more universal than pharmacological specificity. However, the frequency bias observed, where common herbs overshadowed community-specific ones, reflects known limitations of data-driven AI models\u0026rsquo; bias behaviour [\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e], confirming that additional locality conditioning or reinforcement learning may be necessary to avoid epistemic flattening.\u003c/p\u003e \u003cp\u003eThe observed failure modes (Section \u003cspan refid=\"Sec21\" class=\"InternalRef\"\u003e4.6\u003c/span\u003e) echo wider AI research acknowledging challenges in ambiguity handling, language switching, and narrative balance [\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e]. Reduced paraphrasing flexibility in escalation messaging, while safe, aligns with recommendations in medical AI literature suggesting that templated crisis language is desirable for safety but may reduce conversational naturalness [\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e]. Meanwhile, bilingual drift in Luganda output resonates with findings in African NLP that smaller models still require enriched multilingual corpora to generate fluent indigenous language phrasing [\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eFinally, the computational feasibility results (Section \u003cspan refid=\"Sec24\" class=\"InternalRef\"\u003e4.8\u003c/span\u003e) provide practical relevance consistent with digital equity scholarship that cautions against building AI systems requiring high-bandwidth or cloud infrastructure in low-resource areas [\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e]. The ability to fine-tune and run the system on modest hardware supports the argument that AI technologies can be embedded within rural health ecosystems, bridging a known gap between high-tech innovation and local health realities [\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eTherefore, these findings situate this study within three intersecting research streams: indigenous knowledge digitisation, AI for health communication, and responsible AI deployment in low-resource settings. The results extend theoretical understanding by showing that indigenous explanatory knowledge is learnable and safe for computational mediation, while also providing an applied pathway for equitable AI integration. However, performance limitations such as locality loss, bilingual weaknesses, and ambiguity handling signal the need for iterative co-design with communities, enriched multilingual datasets, and refinement before full-scale deployment.\u003c/p\u003e"},{"header":"6. Conclusion","content":"\u003cp\u003eThis study examined whether indigenous malaria knowledge, long transmitted orally through community practice, can be transformed into digital health intelligence using artificial intelligence. The results demonstrate that when carefully structured and safety-wrapped, indigenous explanatory and treatment patterns can be computationally internalised by a language model, achieving strong alignment with local knowledge sources. The trained model reproduced herbal reasoning, preventive behaviours, and community escalation narratives with 81% retrieval accuracy (precision@3), and 77% knowledge coverage (recall@5), providing empirical support for the feasibility of knowledge transfer. Notably, cultural authenticity evaluations yielded an average expert rating of 4.3/5, while escalation behaviour for danger-sign prompts attained a 4.7/5 safety compliance score, reflecting the model\u0026rsquo;s capacity to balance indigenous reasoning with biomedical referral guidance. Hallucination frequency reduction, from 28% in the baseline model to 9% post-training, further supports that grounding in indigenous knowledge improves factual stability and reduces generative drift. Cross-community evaluation revealed 69% generalisation ability, with performance highest for preventive behaviour reasoning (78%) and lowest for district-specific herb references (65%), indicating partial locality loss. While meaningful variability was observed in paraphrasing capacity, multilingual phrasing accuracy remained moderate, particularly in Luganda reformulations. These performance gaps emphasise the need for continued data enrichment, multi-language embedding, and co-development with community stakeholders. Computational feasibility assessment showed that the model could be fine-tuned on modest institutional hardware in under six hours, run inference locally with an average\u0026thinsp;\u0026lt;\u0026thinsp;2.5-second latency, and execute under quantised settings with minimal degradation. These results suggest that culturally grounded AI systems are practical for low-resource deployments, aligning with digital inclusion priorities. This study contributes foundational evidence that indigenous malaria knowledge is computationally learnable and communicable through AI when appropriately contextualised. The integration of qualitative learning patterns with quantitative accuracy metrics demonstrates that AI can complement, rather than overwrite, local malaria reasoning. Future work should expand multilingual capability, incorporate community-in-the-loop retraining, strengthen contextual specificity, and conduct field deployment studies to validate behavioural effects. Scaling these approaches across disease domains may advance equitable digital health systems that are culturally resonant and locally accountable.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAcknowledgments:\u0026nbsp;\u003c/strong\u003eNone\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthors contributions:\u0026nbsp;\u003c/strong\u003eThe authors confirm their contribution to the paper as follows: study conception and design: Emmanuel Ahishakiye (EA); data collection: EA, Nina Olivia Rugambwa (NOR); analysis and interpretation of results: EA, NOR; draft manuscript preparation: EA; review and editing: EA, NOR. All authors reviewed the results and approved the final version of the manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding:\u0026nbsp;\u003c/strong\u003eWe acknowledge the financial support from Kyambogo University 10\u003csup\u003eth\u003c/sup\u003e competitive grant.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData Availability:\u0026nbsp;\u003c/strong\u003eThe data that support the findings of this study are available from the corresponding author upon reasonable request.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCode availability:\u0026nbsp;\u003c/strong\u003eNot applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthics approval and consent to participate:\u0026nbsp;\u003c/strong\u003eEthical approval for this study was obtained from the Institutional Review Board of the Kyambogo University Research Ethics Committee. All methods were performed in accordance with the Declaration of Helsinki-Ethical Principles for Medical Research Involving Human Subjects, the Council for International Organizations of Medical Sciences International Ethical Guidelines for Health-related Research Involving Humans, and the Uganda National Council for Science and Technology National Guidelines for Research Involving Humans as Research Participants. The study primarily involved secondary analysis of documented indigenous knowledge and did not include direct clinical intervention. Where expert reviewers contributed to the validation of model outputs, their participation was voluntary and informed. No identifiable personal data were collected, stored, or processed. Informed consent was obtained from all individuals who participated in expert review activities.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent for publication:\u003c/strong\u003e Not applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests:\u003c/strong\u003e The authors declare no competing interests.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eClinical trial number:\u003c/strong\u003e Not applicable.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eWHO. World malaria report 2023. Accessed: Jan. 13, 2024. [Online]. Available: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.who.int/teams/global-malaria-programme/reports/world-malaria-report-2023\u003c/span\u003e\u003cspan address=\"https://www.who.int/teams/global-malaria-programme/reports/world-malaria-report-2023\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYeka A, et al. Malaria in Uganda: challenges to control on the long road to elimination. I. Epidemiology and current control efforts. Acta Trop. Mar. 2011;121(3):184. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/J.ACTATROPICA.2011.03.004\u003c/span\u003e\u003cspan address=\"10.1016/J.ACTATROPICA.2011.03.004\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBunalema L et al. Dec., Medicinal plants traditionally used for management of malaria in rural communities of Uganda, \u003cem\u003eBMC Complement. Med. Ther.\u003c/em\u003e, vol. 25, no. 1, p. 201, 2025, \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1186/S12906-025-04946-3\u003c/span\u003e\u003cspan address=\"10.1186/S12906-025-04946-3\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGumisiriza H, Olet EA, Mukasa P, Lejju JB, Omara T. Ethnomedicinal plants used for malaria treatment in Rukungiri District, Western Uganda. Trop Med Health. Dec. 2023;51(1):49. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1186/S41182-023-00541-9\u003c/span\u003e\u003cspan address=\"10.1186/S41182-023-00541-9\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTabuti JRS, et al. Medicinal plants used for treatment of malaria by indigenous communities of Tororo District, Eastern Uganda. Trop Med Health. Dec. 2023;51(1):34. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1186/S41182-023-00526-8\u003c/span\u003e\u003cspan address=\"10.1186/S41182-023-00526-8\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eOkello D, Kang Y. Exploring Antimalarial Herbal Plants across Communities in Uganda Based on Electronic Data, \u003cem\u003eEvidence-Based Complement. Altern. Med.\u003c/em\u003e, vol. 2019, no. 1, p. 3057180, Jan. 2019. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1155/2019/3057180\u003c/span\u003e\u003cspan address=\"10.1155/2019/3057180\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSiatwiinda SM, Muchochoma D, Mwondela MD, Kunda C, Nkhoma CJ. Indigenous knowledge strengthens community resilience in Zambia\u0026rsquo;s agroecological region I, \u003cem\u003eDiscov. Glob. Soc.\u003c/em\u003e 2025 31, vol. 3, no. 1, p. 142-, Nov. 2025. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1007/S44282-025-00289-Z\u003c/span\u003e\u003cspan address=\"10.1007/S44282-025-00289-Z\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMzimela JH, Moyo I. On the Efficacy of Indigenous Knowledge Systems in Responding to the COVID-19 Pandemic: Unsettling Coloniality, \u003cem\u003eInt. J. Environ. Res. Public Heal.\u003c/em\u003e 2024, Vol. 21, Page 731, vol. 21, no. 6, p. 731, Jun. 2024. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3390/IJERPH21060731\u003c/span\u003e\u003cspan address=\"10.3390/IJERPH21060731\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKumar SM, et al. The role of artificial intelligence (AI) in the elimination of malaria: a systematic literature review. Malar J. Dec. 2025;24(1):427. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1186/S12936-025-05661-7\u003c/span\u003e\u003cspan address=\"10.1186/S12936-025-05661-7\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNeha F, Bhati D, Shukla DK. Retrieval-Augmented Generation (RAG) in Healthcare: A Comprehensive Review, \u003cem\u003eAI\u003c/em\u003e 2025, Vol. 6, Page 226, vol. 6, no. 9, p. 226, Sep. 2025. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3390/AI6090226\u003c/span\u003e\u003cspan address=\"10.3390/AI6090226\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eEzeanya-esiobu C, Oguamanam C, Ndungutse V. Marginalisation of indigenous knowledge in african education: The case of Rwandan traditional medicinal treatments for livestock, 2021.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKuvuna B, Nyanchoka M, Guleid F, Ogutu M, Tsofa B, Nzinga J. Community-Based Health Information Systems in Africa: A Scoping Review of Data Generation, Utilization, and Community Empowerment. Wellcome Open Res. Oct. 2024;9:485. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.12688/WELLCOMEOPENRES.22780.3\u003c/span\u003e\u003cspan address=\"10.12688/WELLCOMEOPENRES.22780.3\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eStangeland T, Alele PE, Katuura E, Lye KA. Plants used to treat malaria in Nyakayojo sub-county, western Uganda, \u003cem\u003eJ. Ethnopharmacol.\u003c/em\u003e, vol. 137, no. 1, pp. 154\u0026ndash;166, Sep. 2011, \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/J.JEP.2011.05.002\u003c/span\u003e\u003cspan address=\"10.1016/J.JEP.2011.05.002\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDegoot A et al. Health data issues in Africa: time for digitization, standardization and harmonization, \u003cem\u003eNat. Commun.\u003c/em\u003e 2025 161, vol. 16, no. 1, p. 5694-, Jul. 2025. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41467-025-61104-6\u003c/span\u003e\u003cspan address=\"10.1038/s41467-025-61104-6\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSylla B et al. African digital health strategic plans analysis: key weaknesses in contextualization, intervention focus, and technological foresight, \u003cem\u003enpj Digit. Med.\u003c/em\u003e 2025 81, vol. 8, no. 1, p. 748-, Dec. 2025. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41746-025-02121-z\u003c/span\u003e\u003cspan address=\"10.1038/s41746-025-02121-z\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBanjo OB. Bridging Traditional and Modern Health Systems: The Role of Digital Health in Traditional Medicine Integration in Nigeria, \u003cem\u003emedRxiv\u003c/em\u003e, p. 2025.04.29.25326680, May 2025, \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1101/2025.04.29.25326680\u003c/span\u003e\u003cspan address=\"10.1101/2025.04.29.25326680\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNdlovu EC, Gumbo MT. Decolonising Technology Education: Integrating Indigenous Knowledge for Sustainable Development in Electrical and Mechanical Systems and Control, \u003cem\u003eAfrican J. Res. Math. Sci. Technol. Educ.\u003c/em\u003e, vol. 28, no. 3, pp. 362\u0026ndash;372, Sep. 2024, \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1080/18117295.2024.2381984\u003c/span\u003e\u003cspan address=\"10.1080/18117295.2024.2381984\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTelles S, Pathak S, Singh N, Balkrishna A. Research on Traditional Medicine: What Has Been Done, the Difficulties, and Possible Solutions, \u003cem\u003eEvid. Based. Complement. Alternat. Med.\u003c/em\u003e, vol. 2014, p. 495635, 2014. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1155/2014/495635\u003c/span\u003e\u003cspan address=\"10.1155/2014/495635\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eManyazewal T, et al. Innovative technologies to address neglected tropical diseases in African settings with persistent sociopolitical instability. Nat Commun. Dec. 2024;15(1):10274. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/S41467-024-54496-4\u003c/span\u003e\u003cspan address=\"10.1038/S41467-024-54496-4\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHussain W, Khoriba G, Maity S, Jyoti Saikia M. Large Language Models in Healthcare and Medical Applications: A Review, \u003cem\u003eBioengineering\u003c/em\u003e, vol. 12, no. 6, p. 631, Jun. 2025, \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3390/BIOENGINEERING12060631\u003c/span\u003e\u003cspan address=\"10.3390/BIOENGINEERING12060631\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAmugongo LM, Mascheroni P, Brooks S, Doering S, Seidel J. Retrieval augmented generation for large language models in healthcare: A systematic review, \u003cem\u003ePLOS Digit. Heal.\u003c/em\u003e, vol. 4, no. 6, p. e0000877, Jun. 2025, \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1371/JOURNAL.PDIG.0000877\u003c/span\u003e\u003cspan address=\"10.1371/JOURNAL.PDIG.0000877\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eEl-Enen MA, Saad S, Nazmy T. A survey on retrieval-augmentation generation (RAG) models for healthcare applications, \u003cem\u003eNeural Comput. Appl.\u003c/em\u003e 2025 3733, vol. 37, no. 33, pp. 28191\u0026ndash;28267, Oct. 2025. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1007/S00521-025-11666-9\u003c/span\u003e\u003cspan address=\"10.1007/S00521-025-11666-9\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSerge B, Mbondji E, Humphrey K, Janauschek L. \u003cem\u003eHealth Data Digitalization in Africa: Unlocking the potential\u003c/em\u003e. 2024.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eeunyounglee/GPT-. NeoX-1.3B-2GB-Eng \u0026middot; Hugging Face. Accessed: Dec. 09, 2025. [Online]. Available: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://huggingface.co/eunyounglee/GPT-NeoX-1.3B-2GB-Eng\u003c/span\u003e\u003cspan address=\"https://huggingface.co/eunyounglee/GPT-NeoX-1.3B-2GB-Eng\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eEss\u0026eacute; C, et al. Social and cultural aspects of \u0026lsquo;malaria\u0026rsquo; and its control in central C\u0026ocirc;te d\u0026rsquo;Ivoire. Malar J. 2008;7:224. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1186/1475-2875-7-224\u003c/span\u003e\u003cspan address=\"10.1186/1475-2875-7-224\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAsgari E et al. A framework to assess clinical safety and hallucination rates of LLMs for medical text summarisation, \u003cem\u003enpj Digit. Med.\u003c/em\u003e 2025 81, vol. 8, no. 1, p. 274-, May 2025. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41746-025-01670-7\u003c/span\u003e\u003cspan address=\"10.1038/s41746-025-01670-7\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSadek M, Mougenot C. Challenges in Value-Sensitive AI Design: Insights from AI Practitioner Interviews. Int J Hum Comput Interact. 2025;41(17):10877\u0026ndash;94. ;ISSUE:ISSUE:DOI.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKumah E. Artificial intelligence in healthcare and its implications for patient centered care, \u003cem\u003eDiscov. Public Heal.\u003c/em\u003e 2025 221, vol. 22, no. 1, p. 524-, Sep. 2025. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1186/S12982-025-00924-9\u003c/span\u003e\u003cspan address=\"10.1186/S12982-025-00924-9\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNovakowski SK, et al. Health worker perspectives of Smart Triage, a digital triaging platform for quality improvement at a referral hospital in Uganda: a qualitative analysis. BMC Pediatr. Dec. 2022;22(1):593. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1186/S12887-022-03627-1\u003c/span\u003e\u003cspan address=\"10.1186/S12887-022-03627-1\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMusoke D, et al. Integrated approach to malaria prevention at household level in rural communities in Wakiso district, Uganda: impact evaluation of a pilot project. Afr Health Sci. 2018;18(4):1144. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.4314/AHS.V18I4.35\u003c/span\u003e\u003cspan address=\"10.4314/AHS.V18I4.35\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNtoutsi E, et al. Bias in data-driven artificial intelligence systems\u0026mdash;An introductory survey. Wiley Interdiscip Rev Data Min Knowl Discov. May 2020;10(3):e1356. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1002/WIDM.1356\u003c/span\u003e\u003cspan address=\"10.1002/WIDM.1356\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWildsmith-Cromarty R, Reilly C, Kamdem S. Investigating the opportunities and challenges for African languages in public spaces: an introduction. J Multiling Multicult Dev. Oct. 2023;44(9):765\u0026ndash;72. ;WEBSITE:WEBSITE:TFOPB;JOURNAL:JOURNAL:RMMM20;PAGEGROUP:STRING:PUBLICATION.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHuo B, et al. Reporting guidelines for chatbot health advice studies: explanation and elaboration for the Chatbot Assessment Reporting Tool (CHART). BMJ. Aug. 2025;390. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1136/BMJ-2024-083305\u003c/span\u003e\u003cspan address=\"10.1136/BMJ-2024-083305\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMbonimpa PS, Tuyizere D, Biyabani AA, Tonguz OK. Edge-Based Speech Transcription and Synthesis for Kinyarwanda and Swahili Languages, Oct. 2025, Accessed: Dec. 09, 2025. [Online]. Available: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://arxiv.org/pdf/2510.16497v1\u003c/span\u003e\u003cspan address=\"https://arxiv.org/pdf/2510.16497v1\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLamem MFH, Sahid MI, Ahmed A. Artificial intelligence for access to primary healthcare in rural settings. J Med Surg Public Heal. Apr. 2025;5:100173. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/J.GLMEDI.2024.100173\u003c/span\u003e\u003cspan address=\"10.1016/J.GLMEDI.2024.100173\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAbdelouahed M, Yateem D, Amzil C, Aribi I, Abdelwahed EH, Fredericks S. Integrating artificial intelligence into public health education and healthcare: insights from the COVID-19 and monkeypox crises for future pandemic readiness, \u003cem\u003eFront. Educ.\u003c/em\u003e, vol. 10, p. 1518909, Apr. 2025, \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3389/FEDUC.2025.1518909/FULL\u003c/span\u003e\u003cspan address=\"10.3389/FEDUC.2025.1518909/FULL\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Indigenous Knowledge Systems, Artificial Intelligence for Health, Malaria Digital Intelligence, Retrieval-Augmented Language Models, Culturally Grounded Decision Support","lastPublishedDoi":"10.21203/rs.3.rs-8337646/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8337646/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eIndigenous malaria knowledge remains foundational to health decision-making in Uganda, yet it remains under-represented in digital health systems. This study investigates whether culturally embedded malaria knowledge can be transformed into interpretable digital intelligence using artificial intelligence (AI). An indigenous malaria corpus was constructed from documented ethnobotanical sources, community behaviour reports, and cultural health narratives drawn from 11 Ugandan communities, comprising 38 medicinal plant profiles, 17 non-plant prevention practices, and 14 explanatory belief pathways. The dataset was encoded into 812 prompt\u0026ndash;response pairs and used to fine-tune GPT-NeoX-1.3B through low-rank adaptation and retrieval-augmented conditioning. Model evaluation indicated promising performance: precision@3 of 81%, recall@5 of 77%, and cultural authenticity ratings averaging 4.3/5 from traditional medicine reviewers. Safety reliability was high, with escalation behaviours achieving 4.7/5 compliance, while hallucination frequency reduced from 28% in the base model to 9% post-training. Cross-community reasoning achieved 69% generalisation, strongest in prevention logic (78%) but comparatively weaker in herb specificity (65%). The model showed moderate paraphrasing diversity but limited bilingual fluency, particularly under Luganda reformulations. Computational feasibility testing revealed central processing unit (CPU) inference latency below 2.5 seconds and stable performance under int8 quantisation, suggesting suitability for use on low-resource devices typical of rural health settings. These results demonstrate that Ugandan indigenous malaria knowledge can be computationally represented, internalised, and operationalised through AI without compromising cultural tone or safety. The work provides a proof-of-concept for culturally grounded digital health intelligence and identifies pathways for scale-up via community co-creation, indigenous language expansion, and applied field prototyping.\u003c/p\u003e","manuscriptTitle":"Artificial Intelligence for Transforming Indigenous Malaria Knowledge into Digital Health Intelligence in Uganda","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-12-31 09:06:31","doi":"10.21203/rs.3.rs-8337646/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"01c87f19-b1db-4f06-a627-9832cad49d80","owner":[],"postedDate":"December 31st, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2026-02-25T09:44:46+00:00","versionOfRecord":[],"versionCreatedAt":"2025-12-31 09:06:31","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8337646","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8337646","identity":"rs-8337646","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}