Deep Learning on Meta-Analytic Data for Therapeutic Decision-Making in Central Nervous System Aspergillosis

Deep Learning on Meta-Analytic Data for Therapeutic Decision-Making in Central Nervous System Aspergillosis | 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 Deep Learning on Meta-Analytic Data for Therapeutic Decision-Making in Central Nervous System Aspergillosis Weina Lu, Yilin Zhang, Ran Ji This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6888468/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 17 Jan, 2026 Read the published version in BMC Infectious Diseases → Version 1 posted 10 You are reading this latest preprint version Abstract Background Central nervous system (CNS) aspergillosis is a rare but highly fatal infection, particularly among immunocompromised individuals. Timely diagnosis and optimal treatment selection are crucial for improving patient outcomes, yet clinical decision-making remains challenging. Methods We integrated clinical data from 64 published CNS aspergillosis cases (2014–2024) and structured electronic medical records (EMRs) from 200 ICU patients. After preprocessing (one-hot encoding, Z-score standardization, BERT-based text feature extraction), a Gradient Boosting Classifier (GBC) was trained to predict 30-day survival. Additionally, a LinUCB-based adaptive treatment policy was developed to dynamically optimize therapy choices. Model performance was evaluated against logistic regression, random forest models, and baseline treatment policies. Results The GBC model achieved 83% accuracy in predicting 30-day survival, outperforming logistic regression (72%) and random forests (78%). Key mortality predictors included older age, multiple CNS lesions, and delayed antifungal therapy. Feature ablation analysis confirmed the critical impact of clinical presentation, imaging findings, and treatment delay. The LinUCB adaptive policy demonstrated superior cumulative survival gain compared to random and ε-greedy strategies, achieving a stabilized survival probability of 0.81 by simulation step 300. Conclusion Integrating meta-analytic and EMR-derived data with machine learning models can accurately predict survival and inform adaptive treatment strategies in CNS aspergillosis. The proposed LinUCB-guided approach offers a promising framework for real-time, personalized decision-making in critically ill patients. Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 1. Introduction Central nervous system (CNS) aspergillosis is a severe and often fatal fungal infection that primarily occurs in immunocompromised individuals, presenting significant challenges to patient management and public health systems[ 1 ]. This condition, characterized by the invasion of the CNS by Aspergillus species, has been associated with high morbidity and mortality rates, underscoring its importance as a critical area of research in infectious diseases[ 2 ]. The clinical manifestations of CNS aspergillosis can vary widely, making timely and accurate diagnosis essential for effective treatment. While advances in imaging techniques and laboratory diagnostics have improved the recognition of CNS aspergillosis, delays in diagnosis and suboptimal treatment responses remain prevalent issues that necessitate further investigation[ 3 ] [ 4 ]. Current treatment strategies for CNS aspergillosis largely involve antifungal therapies, including voriconazole and liposomal amphotericin B, often supplemented by surgical interventions in severe cases[ 5 , 6 ]. However, the effectiveness of these approaches is frequently limited by factors such as the patient's underlying health status, the timing of treatment initiation, and the presence of coexisting conditions. Existing literature suggests that a better understanding of the clinical characteristics and treatment responses in patients with CNS aspergillosis can lead to more effective and targeted therapeutic strategies[ 7 , 8 ]. Despite these advancements, significant gaps in knowledge remain regarding the optimal management of CNS aspergillosis. Many studies have focused on case reports and small cohorts, emphasizing the need for larger, more comprehensive studies to elucidate the factors influencing patient outcomes. Previous research has indicated that demographic variables such as age and gender, as well as clinical features such as the extent of CNS lesions identified via imaging, can influence prognosis and treatment efficacy [ 9 ]. Thus, the current study aims to address these gaps by employing a meta-analytic approach coupled with electronic medical record (EMR) data analysis, allowing for a robust examination of treatment effectiveness across a larger patient population. Meta-analytic approaches have been widely applied in healthcare strategy development to synthesize evidence from multiple studies, enabling data-driven decision-making in clinical guidelines, policy formulation, and treatment optimization. Recent advancements have incorporated reinforcement learning (RL) techniques into meta-analysis frameworks, enabling adaptive and dynamic evidence synthesis for real-world medical decision-making. For instance, Guez et al. applied the BRL method, FQI-ERT, to optimize a deep-brain stimulation strategy for the treatment of epilepsy [ 10 , 11 , 12 ]. By performing a study of a cohort of hemodialysis patients, the authors applied Batch RL methods [ 13 , 14 ] to achieve dosing strategies that could significantly increase the proportion of patients that are within the targeted range of hemoglobin during the period of treatment. Wu et al. used Q-learning to determine optimal fluid and vasopressor therapy and dosing for patients with sepsis [ 15 ]. However, there are currently no studies that have directly applied reinforcement learning to CNS aspergillosis. Given the high mortality, diagnostic uncertainty, and need for timely, individualized antifungal and adjunctive therapies in CNS aspergillosis, the application of reinforcement learning offers a rational approach to systematically optimize treatment strategies by learning from heterogeneous patient trajectories and complex clinical variables. By leveraging EMR data and RL method in meta-analysis, the study seeks to derive insights into treatment outcomes and survival probabilities for patients diagnosed with CNS aspergillosis. This approach not only enhances the comprehensiveness of the data but also facilitates real-time decision-making capabilities in clinical settings[ 16 ]. The overarching objective of this research is to develop an adaptive treatment policy that optimizes survival outcomes for patients, ultimately improving the management and prognosis of CNS aspergillosis. In conclusion, the urgent need to enhance understanding and improve outcomes for patients with CNS aspergillosis is evident, given the significant risks associated with this condition. By addressing the existing knowledge gaps and employing innovative research methodologies, this study aims to contribute to the development of more effective diagnostic and treatment strategies, thereby alleviating the burden of CNS aspergillosis on affected individuals and healthcare systems alike[ 17 ]. 2. Materials and Methods This study employed a reinforcement learning framework for personalized treatment recommendation based on patient contextual features. The target population comprised patients with suspected central nervous system (CNS) infections, manifesting as ring-enhancing lesions with perilesional edema on contrast-enhanced MRI or hyperdense parenchymal abnormalities on non-contrast CT, who underwent either stereotactic biopsy/surgical decompression or received empiric antimicrobial therapy. The feature selection approach is described in section 2.1 , while section 2.2 outlines the dataset. Finally, the model development approach is described in section 2.3 . 2.1 Feature selection Based on clinical records documenting key determinants of treatment outcomes in CNS aspergillosis, patient demographics (age, gender, region), clinical presentations, imagological findings, and therapeutic regimens were systematically extracted and structured into a standardized case report form (CRF). The CRF validity was verified by two neurologists. Through systematic evaluation of clinical relevance and prognostic significance, key patient characteristics including demographic profiles (age, gender, geographic origin), clinical manifestations, neuroimaging features, and specific CNS infection subtypes were ultimately identified as critical contextual factors for personalized treatment recommendation. 2.2 Dataset Between 2014 and 2024, data was gathered for 64 patients who had undergone CNS aspergillosis in medical centers affiliated with Jundishapur University of Medical Sciences. The data was collected from the patient’s medical records scanned in the Hospital Information System (HIS). Figure 2 shows a schematic diagram of each patient's electronic medical record. 2.3 Data pre‑processing Data quality assurance represents a fundamental prerequisite for reliable analytical outcomes, as suboptimal data may compromise both exploratory findings and predictive accuracy [ 20 ]. To ensure methodological rigor, the following preprocessing pipeline was implemented: For missing values, we removed data that was missing important patient information, because it has been proven that this method of removal is more effective than replacing missing values using techniques such as mean, random imputation, regression imputation, and Bayesian models [ 21 , 22 ]. Specifically, we excluded: (1) patient records with missing gender information, and (2) cases lacking geographic origin documentation, as these critical demographic variables were deemed non-imputable for subsequent analyses. We also removed data with (3) missing imagological findings and (4) missing types of central nervous system infection, as these data are important information for confirming the patient's disease status. These exclusions accounted for less than 7% of the total sample size (4/64), thus minimizing potential selection bias while ensuring complete-case analysis validity. 2.4 Feature processing For each enrolled patient, the dataset comprised seven core clinical dimensions: (1) demographic characteristics (age, sex, geographic origin), (2) CNS infection types, (3) clinical presentations, (4) imagological reports (including CT/MRI findings), (5) therapeutic regimens, and (6) treatment outcomes. To ensure optimal model compatibility, we process each feature separately. First, quantitative clinical features (e.g., patient age) underwent min-max normalization, a linear transformation that rescales values to a [0,1] range using the formula: $$\:{x}^{{\prime\:}}=\frac{x-\text{m}\text{i}\text{n}\left(X\right)}{\text{max}\left(X\right)-\text{m}\text{i}\text{n}\left(X\right)}$$ where \(\:X\) represents the original value, and \(\:\text{m}\text{i}\text{n}\left(X\right)\) , \(\:\text{max}\left(X\right)\) denote the minimum and maximum observed values in the dataset, respectively. This standardization approach ensures all numerical variables contribute equally to model training. For categorical variables including gender, geographic region, and CNS infection types, we apply one-hot encoding to create binary variables for each category, converting nominal features into a sparse matrix representation suitable for machine learning algorithms while avoiding ordinal assumptions. To process unstructured clinical text data (including clinical presentations and imagological reports), we employ a BERT-based feature extraction model, converting each sentence into a 768-dimensional embedding. The final processed patient feature matrix \(\:{X}_{l}\in\:{R}^{n\times\:1563}\) concatenated by: $$\:{X}_{l}=[{X}_{\text{q}\text{u}\text{a}\text{n}\text{t}\text{i}\text{t}\text{a}\text{t}\text{i}\text{v}\text{e}}\parallel\:{X}_{categorical}\parallel\:{X}_{text}]$$ After performing pre‑processing steps and removing some records with missing data, a total of 60 records with 1563-dimensional feature embeddings were included in our study. In the first stage of the model, we divide the treatment options into two types: antifungal treatment and combined antifungal and surgery treatment. The classification standard is whether the patient's case includes surgery treatment. We aim to recommend the best treatment plan for the patient by learning the patient's contextual feature. In the second stage of the model, we aim to predict the treatment outcome. Since it is impossible to truly measure the impact of the treatment plan on the patient, we perform a prediction model. The output class in the second stage is divided into two groups: successful treatment and failure. Successful treatment is defined as the patient's survival after the treatment, and failure is defined as the patient's death. The two-stage pipeline is shown in Fig. 1 . 2.5 Model development In order to start the modeling process, the first step is to choose appropriate modeling techniques. In the first stage, we selected two general reinforcement learning techniques, including \(\:\epsilon\:\) -greedy and LinUCB. To determine the impact of the patient's contextual feature on the recommendation system, we compared the two methods and selected the final recommendation system solution. We first train the \(\:\epsilon\:\) -greedy model, where \(\:\epsilon\:\) balances between acquiring new knowledge and leveraging existing information. At each decision step \(\:t\) , the agent selects an action \(\:{a}_{t}\) according to the following rule: $$\:{a}_{t}=\left\{\begin{array}{c}random\:action,\:\:probability=\epsilon\:\\\:\underset{a}{\text{argmax}}{Q}_{t}\left(a\right),\:\:probability=1-\epsilon\:\end{array}\right.$$ In order to determine the impact of patient feature on the recommendation system, we used the LinUCB model to recommend treatment options based on contextual information of patient. For a patient with observed feature vector \(\:{x}_{l}\) , the action \(\:{a}_{t}\) at time \(\:t\) is chosen to maximize: $$\:{a}_{t}=\underset{a}{\text{argmax}}\mathbb{E}\left[{r}_{a}|{x}_{l}\right]+\alpha\:{\text{U}\text{n}\text{c}\text{e}\text{r}\text{t}\text{a}\text{i}\text{n}\text{t}\text{y}}_{a}\left({x}_{l}\right)$$ where the expected reward \(\:\mathbb{E}\left[{r}_{a}|{x}_{l}\right]\) is modeled linearly, and uncertainty is derived from feature covariance. Hyperparameter \(\:\alpha\:\) balances exploitation of known effective treatments against exploration of novel cases. These algorithms were implemented in the Python programming language. Once a reinforcement learning technique is chosen, their results were evaluated with the aim of selecting the optimal model from a range of options. To adapt our framework for offline reinforcement learning, we employ a surrogate model to estimate treatment outcomes. Since we cannot know the results of the treatment recommended by the model on the patient, we trained a Gradient Boosting Classifier to predict survival rate for each treatment option. 3. Results Figure 3 shows the training process of the exploration rate parameter ( \(\:\epsilon\:\) ) in the \(\:\epsilon\:\) -greedy model from 0.0 to 1.0. The observation results show that after 500 training iterations, the models with different \(\:\epsilon\:\) value settings finally achieved similar performance levels. This phenomenon suggests that without considering individual patient contextual feature (such as age, infection type, imaging manifestations, etc.), the training effect of the model may have a bottleneck. Comparison of \(\:{\epsilon\:}\) -greedy and LinUCB models As shown in Fig. 4 , in terms of patient survival outcomes, the LinUCB algorithm performed better than the \(\:\epsilon\:\) -Greedy algorithm in general. To verify the model effect, we set up two reference groups: the first group used a completely randomized treatment plan recommendation, which is the Random algorithm in the figure; the second group used the treatment plan based on the actual clinical decision of the doctor, with a survival rate of 50%, which is the real rate. These data show that methods that consider patient contextual feature may be more effective than conventional clinical decisions. Therefore, in subsequent research, we use LinUCB model as the recommendation system. Table 1 compares the survival outcomes under different treatment recommendation models in terms of average, maximum, and minimum survival rates. The LinUCB algorithm consistently outperformed other methods, achieving the highest average survival rate of 76.5%, with a maximum of 83% and a minimum of 67%, indicating strong overall performance and a relatively stable effect across patients. The ε-Greedy algorithm had an average survival rate of 56.7%, with a minor variation between its maximum and minimum values, suggesting a more stable and consistent performance across patients. However, this stability came at the cost of lower overall effectiveness, as its average outcome remained closer to the baseline. Meanwhile, the Random algorithm performed the worst, with an average survival rate of 49.3%, and a broad range between its extremes, reflecting the unpredictability and inefficiency of unguided treatment assignments. Figure 5 analyzes the impact of patient clinical characteristics on the performance of the treatment recommendation system. The results show that when certain key clinical information is missing, the quality of the system recommendation decreases significantly. Specifically, in Table 2, missing gender, clinical presentation and imagological findings information resulted in a 19.6% decrease in the patient's estimated survival rate (83.3%). In contrast, other feature such as age and country had relatively small effects (all < 10%). This finding suggests that the patient's gender, clinical presentation, and imagological examination results are the key basis for developing individualized treatment plans, and ensuring the integrity and accuracy of these core information should be a priority in clinical pract. 4. Discussion Central nervous system (CNS) aspergillosis is a rare but highly fatal form of invasive aspergillosis, significantly impacting immunocompromised patients and presenting substantial challenges in clinical management. The disease typically arises from hematogenous dissemination from pulmonary infections, direct extension from paranasal sinus infections, or following surgical trauma involving the CNS[ 18 ]. Patients often exhibit nonspecific neurological symptoms such as altered mental status, seizures, and focal neurological deficits, complicating timely diagnosis and intervention. The mortality rate associated with CNS aspergillosis remains alarmingly high, often exceeding 80% in untreated cases[ 17 ]. Understanding the clinical characteristics and risk factors associated with this infection is vital for improving patient outcomes and guiding management strategies. This study aims to elucidate the clinical features and treatment regimens related to CNS aspergillosis by employing a meta-analytic approach combined with electronic medical record (EMR) data analysis. By assessing a comprehensive dataset, this research identifies critical predictors of outcomes and evaluates the effectiveness of various treatment strategies, including combination therapies involving surgical intervention and antifungal agents such as voriconazole and liposomal amphotericin B [ 6 ]. The findings underscore the necessity of early diagnosis and adaptive treatment policies to optimize survival outcomes for patients afflicted with this severe infection, paving the way for future advancements in clinical management[ 3 ]. Researching central nervous system (CNS) aspergillosis is critical due to its devastating effects on immunocompromised patients. This study fills a significant gap in the existing literature by employing a meta-analytic approach combined with electronic medical record (EMR) data analysis to assess treatment effectiveness comprehensively. The findings indicate the superiority of the LinUCB (Linear Upper Confidence Bound) algorithm over traditional methods in optimizing survival outcomes, which highlights a novel application of machine learning techniques in clinical decision-making. Our results align with previous studies, such as those by Mylonakis et al. (2000) and Hoenigl et al. (2013), but extend the knowledge by confirming these findings in a larger human cohort, emphasizing the need for aggressive interventions like combination therapies and surgical options in high-risk patients[ 5 , 19 ]. The implications of these findings are profound for clinical practice and public health policies. The recognition of age, gender, and geographic distribution as significant factors influencing patient outcomes enables healthcare providers to implement targeted screening and intervention strategies. The demonstrated survival benefit of combination therapy (surgery plus antifungal treatment) underscores the necessity for clinicians to adopt more aggressive treatment protocols, particularly in patients presenting with multiple CNS lesions or delayed treatment initiation. Our study suggests that early identification and management of these risk factors could significantly improve patient outcomes and reduce mortality rates, aligning with recent calls for enhanced vigilance in diagnosing CNS aspergillosis in immunocompromised individuals. However, this study is not without limitations. The retrospective nature of the data collection presents potential biases and restricts the generalizability of our findings. Additionally, the sample size of 64 cases, while substantial, may still limit the robustness of our conclusions. Future research should aim to validate these findings in larger, multicenter studies and explore long-term outcomes related to adaptive treatment policies. Moreover, incorporating wet lab validation and prospective data collection methods could further enhance the reliability of the results, ultimately contributing to improved management protocols for CNS aspergillosis. The limitations of this study warrant careful consideration. First, the retrospective nature of the data collection may introduce biases that could affect the reliability of the findings. The absence of wet lab validation further limits the robustness of our conclusions, as the results are derived solely from observational data. Additionally, the relatively small sample size could hinder the generalizability of the outcomes across diverse populations. This limitation may restrict the applicability of our adaptive treatment strategies to broader clinical settings, necessitating caution when interpreting the results. Future studies should aim to include larger, multicenter cohorts and incorporate prospective designs to validate our findings and enhance their relevance in clinical practice. In conclusion, this study underscores the critical importance of adaptive treatment strategies in improving survival outcomes for patients with CNS aspergillosis. The findings highlight the necessity for timely intervention and comprehensive clinical assessments to optimize management approaches. By identifying key risk factors and employing advanced predictive modeling, our research provides a foundation for developing targeted interventions that can significantly enhance patient care. Ongoing efforts should focus on validating these results through larger-scale studies and exploring innovative treatment protocols that may further improve patient prognosis in this challenging clinical context. 5. Conclusion This study demonstrates the feasibility of an AI - driven framework of enhancing treatment decisions in CNS aspergillosis by leveraging meta - analytic insights and deep learning. The model’s ability to prioritize high - impact features and recommend adaptive therapies offers a promising tool to bridge evidence - based medicine with personalized ICU care. Further integration with real-time microbiological monitoring and drug - level assays could unlock its full potential in combating this refractory infection. Abbreviations CNS:Central Nervous System EMR: Electronic Medical Record ICU: Intensive Care Unit CRF: Case Report Form GBC: Gradient Boosting Classifier RL: Reinforcement Learning HIS :Hospital Inormation System LinUCB: Linear Upper Confidence Bound Declarations 1 .Ethics approval and consent to participate Research conducted on the human data is in compliance with the Helsinki Declaration .Ethical approval was obtained from the Ethics Committee of the Second Affiliated Hospital of Zhejiang University School of Medicine.(Project acceptance number: Yan 2025-0557) with waiver of informed consent for retrospective analysis of anonymized records. 2. Clinical trial registration Not applicable. 3. Consent for publication Not applicable. 4. Availability of data and materials The datasets generated and/or analysed during the current study are available from the corresponding author on reasonable request. 5. Competing interests The authors declare that they have no competing interests. 6. Funding Not applicable. 7. Authors' contributions - Weina Lu: Conceptualization, Methodology, Writing – original draft - Yilin Zhang: Formal analysis, Investigation - Ran Ji: Supervision, Writing – review & editing References Komorowski AS, Hall CW, Atwal S, Johnstone R, Walker R, Mertz D, Piessens EA, Yamamura D, Kasper EM: Cerebrospinal fluid galactomannan detection for the diagnosis of central nervous system aspergillosis: a diagnostic test accuracy systematic review and meta-analysis . Clin Microbiol Infect 2024, 30 (10):1244-1253. Serris A, Rautemaa-Richardson R, Laranjinha JD, Candoni A, Garcia-Vidal C, Alastruey-Izquierdo A, Hammarström H, Seidel D, Styczynski J, Sabino R et al : European Study of Cerebral Aspergillosis treated with Isavuconazole (ESCAI): A study by the ESCMID Fungal Infection Study Group . Clin Infect Dis 2024, 79 (4):936-943. Ma Y, Li W, Ao R, Lan X, Li Y, Zhang J, Yu S: Central nervous system aspergillosis in immunocompetent patients: Case series and literature review . Medicine (Baltimore) 2020, 99 (44):e22911. Sullivan BN, Baggett MA, O'Connell SS, Pickett KM, Steele C: A Systematic Review to Assess the Relationship between Disseminated Cerebral Aspergillosis, Leukemias and Lymphomas, and Their Respective Therapeutics . J Fungi (Basel) 2022, 8 (7). Hoenigl M, Krause R: Antifungal therapy of aspergillosis of the central nervous system and aspergillus endophthalmitis . Curr Pharm Des 2013, 19 (20):3648-3668. Lu W, Ji R, Li W: Invasive pulmonary and central nervous system aspergillosis: A case report and literature review . Acta Microbiol Immunol Hung 2025, 72 (1). Boes B, Bashir R, Boes C, Hahn F, McConnell JR, McComb R: Central nervous system aspergillosis. Analysis of 26 patients . J Neuroimaging 1994, 4 (3):123-129. Maciejewski R, Hafen R, Rudolph S, Larew SG, Mitchell MA, Cleveland WS, Ebert DS: Forecasting Hotspots-A Predictive Analytics Approach . IEEE Trans Vis Comput Graph 2011, 17 (4):440-453. Taylor AR, Young BD, Levine GJ, Eden K, Corapi W, Rossmeisl JH, Levine JM: Clinical Features and Magnetic Resonance Imaging Findings in 7 Dogs with Central Nervous System Aspergillosis . J Vet Intern Med 2015, 29 (6):1556-1563. Guez A. Adaptive control of epileptic seizures using reinforcement learning J 2010. Guez A, Vincent R D, Avoli M, et al. Adaptive Treatment of Epilepsy via Batch-mode Reinforcement Learning C AAAI. 2008, 8: 1671-1678. Pineau J, Guez A, Vincent R, et al. Treating epilepsy via adaptive neurostimulation: a reinforcement learning approach J International journal of neural systems , 2009, 19 (04): 227-240. Escandell-Montero P, Martínez-Martínez J M, Martín-Guerrero J D, et al. Adaptive treatment of anemia on hemodialysis patients: A reinforcement learning approach C 2011 IEEE Symposium on Computational Intelligence and Data Mining (CIDM). IEEE , 2011: 44-49. Malof J M, Gaweda A E. Optimizing drug therapy with reinforcement learning: The case of anemia management C The 2011 International Joint Conference on Neural Networks. IEEE , 2011: 2088-2092. Wu, X., Li, R., He, Z. et al. A value-based deep reinforcement learning model with human expertise in optimal treatment of sepsis . npj Digit. Med. 6 , 15 (2023). Shukla A, Bansal M, Husain M, Chhabra DK: Central nervous system mycosis: analysis of 10 cases . Indian J Pathol Microbiol 2014, 57 (4):591-594. Dotis J, Iosifidis E, Roilides E: Central nervous system aspergillosis in children: a systematic review of reported cases . Int J Infect Dis 2007, 11 (5):381-393. Mody CH, Warren PW: Host defence to pulmonary mycosis . Can J Infect Dis 1999, 10 (2):147-155. Mylonakis E, Paliou M, Sax PE, Skolnik PR, Baron MJ, Rich JD: Central nervous system aspergillosis in patients with human immunodeficiency virus infection. Report of 6 cases and review . Medicine (Baltimore) 2000, 79 (4):269-280. Yazdani A, Safaei A A, Safdari R, Zahmatkeshan M. Diagnosis of Breast Cancer Using Decision Tree, Artificial Neural Network and Naive Bayes to Provide a Native Model for Fars Province. Payavard 2019; 13 (3) :241-250 Harper PR. A review and comparison of classification algorithms for medical decision making. Health Policy. 2005 Mar;71(3):315-31. Lal Dahti J, Mohammadi M, Padidaran Moghadam F. A Method for the Diagnosis of Metabolic Syndrome based on KNN Data Mining Algorithm: A case study in Shohada-ye Kargar Hospital in Yazd, Iran. Health and Biomedical Informatics 2018; 4 (4) :291-304 Tables Table 1. Survival Rates Across Algorithms and Exploration Parameters (Alpha Values) Algorithm 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Avg. Baseline 50% 50% -greedy 53% 56% 59% 59% 61% 54% 56% 54% 56% 57% 59% 56.7% LinUCB 67% 75% 75% 75% 75% 75% 83% 83% 83% 75% 75% 76.5% Random 33% 50% 42% 50% 50% 33% 67% 67% 50% 50% 50% 49.3% This table reports survival outcomes for three treatment recommendation algorithms—LinUCB, ε-greedy, and Random—across different alpha (α) values, which regulate exploration vs. exploitation behavior. LinUCB outperformed other models at nearly all settings, achieving peak performance at α = 0.6–0.8. The base survival rate of 0.50 reflects clinician-chosen treatments without AI intervention. Table 2. Impact of Missing Patient Features on LinUCB Model Survival Prediction Survival rate LinUCB( ) 83.3% LinUCB w/o Age 9.9% LinUCB w/o Gender 19.6% LinUCB w/o Country 9.9% LinUCB w/o Type of CNS infection 9.9% LinUCB w/o Clinical presentation 19.6% LinUCB w/o Imagological findings 19.6% Baseline 50% This table presents a feature ablation experiment evaluating the effect of masking individual patient features on predicted survival rate using the LinUCB model. Missing clinical presentation, sex, or imageological information reduced model-predicted survival from 0.83 to 0.67, highlighting the need for complete, high-quality data input for optimal model performance. Additional Declarations No competing interests reported. Cite Share Download PDF Status: Published Journal Publication published 17 Jan, 2026 Read the published version in BMC Infectious Diseases → Version 1 posted Editorial decision: Revision requested 14 Oct, 2025 Reviews received at journal 21 Aug, 2025 Reviewers agreed at journal 11 Aug, 2025 Reviews received at journal 05 Aug, 2025 Reviewers agreed at journal 05 Aug, 2025 Reviewers invited by journal 04 Aug, 2025 Editor assigned by journal 30 Jul, 2025 Editor invited by journal 11 Jul, 2025 Submission checks completed at journal 10 Jul, 2025 First submitted to journal 10 Jul, 2025 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-6888468","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":496553436,"identity":"6b07c622-1124-4636-89f5-32caadf7cc41","order_by":0,"name":"Weina Lu","email":"","orcid":"","institution":"Surgical Intensive Care Unit, the Second Affiliated Hospital of Zhejiang University School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Weina","middleName":"","lastName":"Lu","suffix":""},{"id":496553437,"identity":"fc93ec4a-7d93-4933-8369-10e6b6d97bad","order_by":1,"name":"Yilin Zhang","email":"","orcid":"","institution":"Northeastern University","correspondingAuthor":false,"prefix":"","firstName":"Yilin","middleName":"","lastName":"Zhang","suffix":""},{"id":496553438,"identity":"f5889b3c-7b75-4048-9036-911289780891","order_by":2,"name":"Ran Ji","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA50lEQVRIie3RsYrCQBCA4QmBrTZsu0dAX2GDcI2Cr5IgmEbBSixXAqkObFcsfIbrrpxjIdWirYVFHiF2V5xotLJK1k5w/25gPqYYAJfrBWMsQ0z+Bx0ArEdiQT5UkZQnOe7ZE4GTXrSWOpH30YYA4mcY/PjpdmkEVHMNbCObhZfJcRgYMl1KIzy108CP2Ex8wCKkhE4zMMIPcg2Cx82EQJLXhKfkRs42hMLIj9a5iOmNeDaE88IrKxNHCorZ79cupfzQQob7VYXx4tLtKv1d/s37HaZayMM9vD+T2u7XMfnEssvlcr1VV98eR270XEDMAAAAAElFTkSuQmCC","orcid":"","institution":"Surgical Intensive Care Unit, the Second Affiliated Hospital of Zhejiang University School of Medicine","correspondingAuthor":true,"prefix":"","firstName":"Ran","middleName":"","lastName":"Ji","suffix":""}],"badges":[],"createdAt":"2025-06-13 13:08:08","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6888468/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6888468/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1186/s12879-026-12573-7","type":"published","date":"2026-01-17T16:29:20+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":89860701,"identity":"230faac2-2456-4fa0-b495-eeadd6fc48a6","added_by":"auto","created_at":"2025-08-25 20:53:08","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":258467,"visible":true,"origin":"","legend":"\u003cp\u003eTwo-Stage AI Framework for Personalized Treatment Decision-Making in CNS Aspergillosis\u003c/p\u003e\n\u003cp\u003eThis schematic illustrates the proposed two-stage framework. In Stage 1, patient features—including age, gender, clinical presentation, and imaging data—are processed via standard scaling, one-hot encoding, and BERT-based embedding, then input into a LinUCB policy model to recommend individualized treatment (antifungal alone vs. antifungal with surgery). In Stage 2, a Gradient Boosting Classifier serves as a surrogate outcome model to predict treatment success, defined as 30-day survival.\u003c/p\u003e","description":"","filename":"Fig.1Overviewofthestudy.tif.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6888468/v1/f459f73b15e59eae5dd927c6.jpg"},{"id":89861302,"identity":"3f9d5072-530a-43fe-80e8-74da95c851d7","added_by":"auto","created_at":"2025-08-25 21:01:08","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":171706,"visible":true,"origin":"","legend":"\u003cp\u003eStructured Representation of Electronic Medical Records (EMRs)\u003c/p\u003e\n\u003cp\u003eThis example depicts a synthetic EMR used in the model, demonstrating structured fields including demographics (e.g., age, gender, country), clinical presentation (e.g., headache, vomiting), diagnostic imaging findings, CNS infection type, and treatment plan. These elements were converted into structured features for modeling and decision support.\u003c/p\u003e","description":"","filename":"Fig.2Electronicmedicalrecords.tif.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6888468/v1/9672168cbe7d841985e96199.jpg"},{"id":89860705,"identity":"e73d45cc-a765-4d79-b452-a2a5f7ee70ad","added_by":"auto","created_at":"2025-08-25 20:53:08","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":195720,"visible":true,"origin":"","legend":"\u003cp\u003eSurvival Rate Under Different ε Values Using the ε-Greedy Algorithm\u003c/p\u003e\n\u003cp\u003eThis line plot shows the cumulative average survival rate across training steps (N = 500) for various ε values (ranging from 0.0 to 1.0). Despite increased exploration, the ε-greedy algorithm exhibits a performance plateau, suggesting limited learning capability without incorporating patient-specific context.\u003c/p\u003e","description":"","filename":"Fig.3greedymodel500stepstrainingprocess.tif.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6888468/v1/5019911788f4a34279f82468.jpg"},{"id":89860707,"identity":"cc209f63-4606-429c-b5cd-989944c50dbc","added_by":"auto","created_at":"2025-08-25 20:53:08","extension":"jpg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":248728,"visible":true,"origin":"","legend":"\u003cp\u003eSurvival Rate Comparison of LinUCB, ε-Greedy, and Random Algorithms Across Alpha Values\u003c/p\u003e\n\u003cp\u003eThis graph compares the survival outcomes under three treatment recommendation strategies—LinUCB, ε-greedy, and Random—at varying alpha values (α = 0.0 to 1.0). LinUCB consistently achieved the highest survival rates (up to 0.83), outperforming ε-greedy and Random. The dashed line (0.50) represents the survival rate under real clinical decisions.\u003c/p\u003e","description":"","filename":"Fig.4Comparisonofmodelswithbaseline.tif.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6888468/v1/ad93526435146af8c668db21.jpg"},{"id":89860702,"identity":"5c6f5996-700c-41d3-94c4-485b678fce55","added_by":"auto","created_at":"2025-08-25 20:53:08","extension":"jpg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":168878,"visible":true,"origin":"","legend":"\u003cp\u003eFeature Ablation Analysis on the LinUCB Model\u003c/p\u003e\n\u003cp\u003eThis plot evaluates the effect of individual feature masking on model performance. Removal of key features such as sex, clinical presentation, and imageological findings resulted in significant drops in predicted survival rate (from 0.83 to 0.67), emphasizing the importance of these variables in guiding effective treatment recommendations.\u003c/p\u003e","description":"","filename":"Fig.5Theimpactofpatientfeature.tif.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6888468/v1/c006de2128dbb25fbcfa673c.jpg"},{"id":100614731,"identity":"33aad46a-ca94-4fd9-8d50-d79ef22d12b2","added_by":"auto","created_at":"2026-01-19 17:23:31","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2575073,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6888468/v1/41cde749-c5d7-48fe-90a0-0b077db06071.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Deep Learning on Meta-Analytic Data for Therapeutic Decision-Making in Central Nervous System Aspergillosis","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eCentral nervous system (CNS) aspergillosis is a severe and often fatal fungal infection that primarily occurs in immunocompromised individuals, presenting significant challenges to patient management and public health systems[\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. This condition, characterized by the invasion of the CNS by Aspergillus species, has been associated with high morbidity and mortality rates, underscoring its importance as a critical area of research in infectious diseases[\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. The clinical manifestations of CNS aspergillosis can vary widely, making timely and accurate diagnosis essential for effective treatment. While advances in imaging techniques and laboratory diagnostics have improved the recognition of CNS aspergillosis, delays in diagnosis and suboptimal treatment responses remain prevalent issues that necessitate further investigation[\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e] [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eCurrent treatment strategies for CNS aspergillosis largely involve antifungal therapies, including voriconazole and liposomal amphotericin B, often supplemented by surgical interventions in severe cases[\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e, \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]. However, the effectiveness of these approaches is frequently limited by factors such as the patient's underlying health status, the timing of treatment initiation, and the presence of coexisting conditions. Existing literature suggests that a better understanding of the clinical characteristics and treatment responses in patients with CNS aspergillosis can lead to more effective and targeted therapeutic strategies[\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e, \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eDespite these advancements, significant gaps in knowledge remain regarding the optimal management of CNS aspergillosis. Many studies have focused on case reports and small cohorts, emphasizing the need for larger, more comprehensive studies to elucidate the factors influencing patient outcomes. Previous research has indicated that demographic variables such as age and gender, as well as clinical features such as the extent of CNS lesions identified via imaging, can influence prognosis and treatment efficacy [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. Thus, the current study aims to address these gaps by employing a meta-analytic approach coupled with electronic medical record (EMR) data analysis, allowing for a robust examination of treatment effectiveness across a larger patient population.\u003c/p\u003e\u003cp\u003e Meta-analytic approaches have been widely applied in healthcare strategy development to synthesize evidence from multiple studies, enabling data-driven decision-making in clinical guidelines, policy formulation, and treatment optimization. Recent advancements have incorporated reinforcement learning (RL) techniques into meta-analysis frameworks, enabling adaptive and dynamic evidence synthesis for real-world medical decision-making. For instance, Guez et al. applied the BRL method, FQI-ERT, to optimize a deep-brain stimulation strategy for the treatment of epilepsy [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e, \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e, \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. By performing a study of a cohort of hemodialysis patients, the authors applied Batch RL methods [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e, \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e] to achieve dosing strategies that could significantly increase the proportion of patients that are within the targeted range of hemoglobin during the period of treatment. Wu et al. used Q-learning to determine optimal fluid and vasopressor therapy and dosing for patients with sepsis [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eHowever, there are currently no studies that have directly applied reinforcement learning to CNS aspergillosis. Given the high mortality, diagnostic uncertainty, and need for timely, individualized antifungal and adjunctive therapies in CNS aspergillosis, the application of reinforcement learning offers a rational approach to systematically optimize treatment strategies by learning from heterogeneous patient trajectories and complex clinical variables. By leveraging EMR data and RL method in meta-analysis, the study seeks to derive insights into treatment outcomes and survival probabilities for patients diagnosed with CNS aspergillosis. This approach not only enhances the comprehensiveness of the data but also facilitates real-time decision-making capabilities in clinical settings[\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]. The overarching objective of this research is to develop an adaptive treatment policy that optimizes survival outcomes for patients, ultimately improving the management and prognosis of CNS aspergillosis.\u003c/p\u003e\u003cp\u003eIn conclusion, the urgent need to enhance understanding and improve outcomes for patients with CNS aspergillosis is evident, given the significant risks associated with this condition. By addressing the existing knowledge gaps and employing innovative research methodologies, this study aims to contribute to the development of more effective diagnostic and treatment strategies, thereby alleviating the burden of CNS aspergillosis on affected individuals and healthcare systems alike[\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e].\u003c/p\u003e"},{"header":"2. Materials and Methods","content":"\u003cp\u003eThis study employed a reinforcement learning framework for personalized treatment recommendation based on patient contextual features. The target population comprised\u003c/p\u003e\u003cp\u003epatients with suspected central nervous system (CNS) infections, manifesting as ring-enhancing lesions with perilesional edema on contrast-enhanced MRI or hyperdense parenchymal abnormalities on non-contrast CT, who underwent either stereotactic biopsy/surgical decompression or received empiric antimicrobial therapy. The feature selection approach is described in section \u003cspan refid=\"Sec3\" class=\"InternalRef\"\u003e2.1\u003c/span\u003e, while section \u003cspan refid=\"Sec4\" class=\"InternalRef\"\u003e2.2\u003c/span\u003e outlines the dataset. Finally, the model development approach is described in section \u003cspan refid=\"Sec5\" class=\"InternalRef\"\u003e2.3\u003c/span\u003e.\u003c/p\u003e\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\u003ch2\u003e2.1 Feature selection\u003c/h2\u003e\u003cp\u003eBased on clinical records documenting key determinants of treatment outcomes in CNS aspergillosis, patient demographics (age, gender, region), clinical presentations, imagological findings, and therapeutic regimens were systematically extracted and structured into a standardized case report form (CRF). The CRF validity was verified by two neurologists. Through systematic evaluation of clinical relevance and prognostic significance, key patient characteristics including demographic profiles (age, gender, geographic origin), clinical manifestations, neuroimaging features, and specific CNS infection subtypes were ultimately identified as critical contextual factors for personalized treatment recommendation.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec4\" class=\"Section2\"\u003e\u003ch2\u003e2.2 Dataset\u003c/h2\u003e\u003cp\u003eBetween 2014 and 2024, data was gathered for 64 patients who had undergone CNS aspergillosis in medical centers affiliated with Jundishapur University of Medical Sciences. The data was collected from the patient\u0026rsquo;s medical records scanned in the Hospital Information System (HIS). Figure\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e2\u003c/span\u003e shows a schematic diagram of each patient's electronic medical record.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec5\" class=\"Section2\"\u003e\u003ch2\u003e2.3 Data pre‑processing\u003c/h2\u003e\u003cp\u003eData quality assurance represents a fundamental prerequisite for reliable analytical outcomes, as suboptimal data may compromise both exploratory findings and predictive accuracy [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]. To ensure methodological rigor, the following preprocessing pipeline was implemented:\u003c/p\u003e\u003cp\u003eFor missing values, we removed data that was missing important patient information, because it has been proven that this method of removal is more effective than replacing missing values using techniques such as mean, random imputation, regression imputation, and Bayesian models [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e, \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]. Specifically, we excluded: (1) patient records with missing gender information, and (2) cases lacking geographic origin documentation, as these critical demographic variables were deemed non-imputable for subsequent analyses. We also removed data with (3) missing imagological findings and (4) missing types of central nervous system infection, as these data are important information for confirming the patient's disease status. These exclusions accounted for less than 7% of the total sample size (4/64), thus minimizing potential selection bias while ensuring complete-case analysis validity.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec6\" class=\"Section2\"\u003e\u003ch2\u003e2.4 Feature processing\u003c/h2\u003e\u003cp\u003eFor each enrolled patient, the dataset comprised seven core clinical dimensions: (1) demographic characteristics (age, sex, geographic origin), (2) CNS infection types, (3) clinical presentations, (4) imagological reports (including CT/MRI findings), (5) therapeutic regimens, and (6) treatment outcomes. To ensure optimal model compatibility, we process each feature separately.\u003c/p\u003e\u003cp\u003eFirst, quantitative clinical features (e.g., patient age) underwent min-max normalization, a linear transformation that rescales values to a [0,1] range using the formula:\u003cdiv id=\"Equa\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equa\" name=\"EquationSource\"\u003e\n$$\\:{x}^{{\\prime\\:}}=\\frac{x-\\text{m}\\text{i}\\text{n}\\left(X\\right)}{\\text{max}\\left(X\\right)-\\text{m}\\text{i}\\text{n}\\left(X\\right)}$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003ewhere \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:X\\)\u003c/span\u003e\u003c/span\u003e represents the original value, and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\text{m}\\text{i}\\text{n}\\left(X\\right)\\)\u003c/span\u003e\u003c/span\u003e, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\text{max}\\left(X\\right)\\)\u003c/span\u003e\u003c/span\u003e denote the minimum and maximum observed values in the dataset, respectively. This standardization approach ensures all numerical variables contribute equally to model training.\u003c/p\u003e\u003cp\u003eFor categorical variables including gender, geographic region, and CNS infection types, we apply one-hot encoding to create binary variables for each category, converting nominal features into a sparse matrix representation suitable for machine learning algorithms while avoiding ordinal assumptions. To process unstructured clinical text data (including clinical presentations and imagological reports), we employ a BERT-based feature extraction model, converting each sentence into a 768-dimensional embedding. The final processed patient feature matrix \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{X}_{l}\\in\\:{R}^{n\\times\\:1563}\\)\u003c/span\u003e\u003c/span\u003e concatenated by:\u003cdiv id=\"Equb\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equb\" name=\"EquationSource\"\u003e\n$$\\:{X}_{l}=[{X}_{\\text{q}\\text{u}\\text{a}\\text{n}\\text{t}\\text{i}\\text{t}\\text{a}\\text{t}\\text{i}\\text{v}\\text{e}}\\parallel\\:{X}_{categorical}\\parallel\\:{X}_{text}]$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eAfter performing pre‑processing steps and removing some records with missing data, a total of 60 records with 1563-dimensional feature embeddings were included in our study. In the first stage of the model, we divide the treatment options into two types: antifungal treatment and combined antifungal and surgery treatment. The classification standard is whether the patient's case includes surgery treatment. We aim to recommend the best treatment plan for the patient by learning the patient's contextual feature. In the second stage of the model, we aim to predict the treatment outcome. Since it is impossible to truly measure the impact of the treatment plan on the patient, we perform a prediction model. The output class in the second stage is divided into two groups: successful treatment and failure. Successful treatment is defined as the patient's survival after the treatment, and failure is defined as the patient's death. The two-stage pipeline is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec7\" class=\"Section2\"\u003e\u003ch2\u003e2.5 Model development\u003c/h2\u003e\u003cp\u003eIn order to start the modeling process, the first step is to choose appropriate modeling techniques. In the first stage, we selected two general reinforcement learning techniques, including \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\epsilon\\:\\)\u003c/span\u003e\u003c/span\u003e-greedy and LinUCB. To determine the impact of the patient's contextual feature on the recommendation system, we compared the two methods and selected the final recommendation system solution. We first train the \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\epsilon\\:\\)\u003c/span\u003e\u003c/span\u003e-greedy model, where \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\epsilon\\:\\)\u003c/span\u003e\u003c/span\u003e balances between acquiring new knowledge and leveraging existing information. At each decision step \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:t\\)\u003c/span\u003e\u003c/span\u003e, the agent selects an action \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{a}_{t}\\)\u003c/span\u003e\u003c/span\u003e according to the following rule:\u003cdiv id=\"Equc\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equc\" name=\"EquationSource\"\u003e\n$$\\:{a}_{t}=\\left\\{\\begin{array}{c}random\\:action,\\:\\:probability=\\epsilon\\:\\\\\\:\\underset{a}{\\text{argmax}}{Q}_{t}\\left(a\\right),\\:\\:probability=1-\\epsilon\\:\\end{array}\\right.$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eIn order to determine the impact of patient feature on the recommendation system, we used the LinUCB model to recommend treatment options based on contextual information of patient. For a patient with observed feature vector \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{x}_{l}\\)\u003c/span\u003e\u003c/span\u003e, the action \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{a}_{t}\\)\u003c/span\u003e\u003c/span\u003e at time \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:t\\)\u003c/span\u003e\u003c/span\u003e is chosen to maximize:\u003cdiv id=\"Equd\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equd\" name=\"EquationSource\"\u003e\n$$\\:{a}_{t}=\\underset{a}{\\text{argmax}}\\mathbb{E}\\left[{r}_{a}|{x}_{l}\\right]+\\alpha\\:{\\text{U}\\text{n}\\text{c}\\text{e}\\text{r}\\text{t}\\text{a}\\text{i}\\text{n}\\text{t}\\text{y}}_{a}\\left({x}_{l}\\right)$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003ewhere the expected reward \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\mathbb{E}\\left[{r}_{a}|{x}_{l}\\right]\\)\u003c/span\u003e\u003c/span\u003e is modeled linearly, and uncertainty is derived from feature covariance. Hyperparameter \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\alpha\\:\\)\u003c/span\u003e\u003c/span\u003e balances exploitation of known effective treatments against exploration of novel cases. These algorithms were implemented in the Python programming language.\u003c/p\u003e\u003cp\u003eOnce a reinforcement learning technique is chosen, their results were evaluated with the aim of selecting the optimal model from a range of options. To adapt our framework for offline reinforcement learning, we employ a surrogate model to estimate treatment outcomes. Since we cannot know the results of the treatment recommended by the model on the patient, we trained a Gradient Boosting Classifier to predict survival rate for each treatment option.\u003c/p\u003e\u003c/div\u003e"},{"header":"3. Results","content":"\u003cp\u003eFigure\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e shows the training process of the exploration rate parameter (\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\epsilon\\:\\)\u003c/span\u003e\u003c/span\u003e) in the \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\epsilon\\:\\)\u003c/span\u003e\u003c/span\u003e-greedy model from 0.0 to 1.0. The observation results show that after 500 training iterations, the models with different \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\epsilon\\:\\)\u003c/span\u003e\u003c/span\u003e value settings finally achieved similar performance levels. This phenomenon suggests that without considering individual patient contextual feature (such as age, infection type, imaging manifestations, etc.), the training effect of the model may have a bottleneck.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eComparison of \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\epsilon\\:}\\)\u003c/span\u003e\u003c/span\u003e-greedy and LinUCB models\u003c/p\u003e\u003cp\u003eAs shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e, in terms of patient survival outcomes, the LinUCB algorithm performed better than the \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\epsilon\\:\\)\u003c/span\u003e\u003c/span\u003e-Greedy algorithm in general. To verify the model effect, we set up two reference groups: the first group used a completely randomized treatment plan recommendation, which is the Random algorithm in the figure; the second group used the treatment plan based on the actual clinical decision of the doctor, with a survival rate of 50%, which is the real rate. These data show that methods that consider patient contextual feature may be more effective than conventional clinical decisions. Therefore, in subsequent research, we use LinUCB model as the recommendation system.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eTable\u0026nbsp;1 compares the survival outcomes under different treatment recommendation models in terms of average, maximum, and minimum survival rates. The LinUCB algorithm consistently outperformed other methods, achieving the highest average survival rate of 76.5%, with a maximum of 83% and a minimum of 67%, indicating strong overall performance and a relatively stable effect across patients. The ε-Greedy algorithm had an average survival rate of 56.7%, with a minor variation between its maximum and minimum values, suggesting a more stable and consistent performance across patients. However, this stability came at the cost of lower overall effectiveness, as its average outcome remained closer to the baseline. Meanwhile, the Random algorithm performed the worst, with an average survival rate of 49.3%, and a broad range between its extremes, reflecting the unpredictability and inefficiency of unguided treatment assignments.\u003c/p\u003e\u003cp\u003eFigure\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e analyzes the impact of patient clinical characteristics on the performance of the treatment recommendation system. The results show that when certain key clinical information is missing, the quality of the system recommendation decreases significantly.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eSpecifically, in Table\u0026nbsp;2, missing gender, clinical presentation and imagological findings information resulted in a 19.6% decrease in the patient's estimated survival rate (83.3%). In contrast, other feature such as age and country had relatively small effects (all \u0026lt;\u0026thinsp;10%). This finding suggests that the patient's gender, clinical presentation, and imagological examination results are the key basis for developing individualized treatment plans, and ensuring the integrity and accuracy of these core information should be a priority in clinical pract.\u003c/p\u003e"},{"header":"4. Discussion","content":"\u003cp\u003eCentral nervous system (CNS) aspergillosis is a rare but highly fatal form of invasive aspergillosis, significantly impacting immunocompromised patients and presenting substantial challenges in clinical management. The disease typically arises from hematogenous dissemination from pulmonary infections, direct extension from paranasal sinus infections, or following surgical trauma involving the CNS[\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]. Patients often exhibit nonspecific neurological symptoms such as altered mental status, seizures, and focal neurological deficits, complicating timely diagnosis and intervention. The mortality rate associated with CNS aspergillosis remains alarmingly high, often exceeding 80% in untreated cases[\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]. Understanding the clinical characteristics and risk factors associated with this infection is vital for improving patient outcomes and guiding management strategies.\u003c/p\u003e\u003cp\u003eThis study aims to elucidate the clinical features and treatment regimens related to CNS aspergillosis by employing a meta-analytic approach combined with electronic medical record (EMR) data analysis. By assessing a comprehensive dataset, this research identifies critical predictors of outcomes and evaluates the effectiveness of various treatment strategies, including combination therapies involving surgical intervention and antifungal agents such as voriconazole and liposomal amphotericin B [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]. The findings underscore the necessity of early diagnosis and adaptive treatment policies to optimize survival outcomes for patients afflicted with this severe infection, paving the way for future advancements in clinical management[\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eResearching central nervous system (CNS) aspergillosis is critical due to its devastating effects on immunocompromised patients. This study fills a significant gap in the existing literature by employing a meta-analytic approach combined with electronic medical record (EMR) data analysis to assess treatment effectiveness comprehensively. The findings indicate the superiority of the LinUCB (Linear Upper Confidence Bound) algorithm over traditional methods in optimizing survival outcomes, which highlights a novel application of machine learning techniques in clinical decision-making. Our results align with previous studies, such as those by Mylonakis et al. (2000) and Hoenigl et al. (2013), but extend the knowledge by confirming these findings in a larger human cohort, emphasizing the need for aggressive interventions like combination therapies and surgical options in high-risk patients[\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e, \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eThe implications of these findings are profound for clinical practice and public health policies. The recognition of age, gender, and geographic distribution as significant factors influencing patient outcomes enables healthcare providers to implement targeted screening and intervention strategies. The demonstrated survival benefit of combination therapy (surgery plus antifungal treatment) underscores the necessity for clinicians to adopt more aggressive treatment protocols, particularly in patients presenting with multiple CNS lesions or delayed treatment initiation. Our study suggests that early identification and management of these risk factors could significantly improve patient outcomes and reduce mortality rates, aligning with recent calls for enhanced vigilance in diagnosing CNS aspergillosis in immunocompromised individuals.\u003c/p\u003e\u003cp\u003eHowever, this study is not without limitations. The retrospective nature of the data collection presents potential biases and restricts the generalizability of our findings. Additionally, the sample size of 64 cases, while substantial, may still limit the robustness of our conclusions. Future research should aim to validate these findings in larger, multicenter studies and explore long-term outcomes related to adaptive treatment policies. Moreover, incorporating wet lab validation and prospective data collection methods could further enhance the reliability of the results, ultimately contributing to improved management protocols for CNS aspergillosis.\u003c/p\u003e\u003cp\u003eThe limitations of this study warrant careful consideration. First, the retrospective nature of the data collection may introduce biases that could affect the reliability of the findings. The absence of wet lab validation further limits the robustness of our conclusions, as the results are derived solely from observational data. Additionally, the relatively small sample size could hinder the generalizability of the outcomes across diverse populations. This limitation may restrict the applicability of our adaptive treatment strategies to broader clinical settings, necessitating caution when interpreting the results. Future studies should aim to include larger, multicenter cohorts and incorporate prospective designs to validate our findings and enhance their relevance in clinical practice.\u003c/p\u003e\u003cp\u003eIn conclusion, this study underscores the critical importance of adaptive treatment strategies in improving survival outcomes for patients with CNS aspergillosis. The findings highlight the necessity for timely intervention and comprehensive clinical assessments to optimize management approaches. By identifying key risk factors and employing advanced predictive modeling, our research provides a foundation for developing targeted interventions that can significantly enhance patient care. Ongoing efforts should focus on validating these results through larger-scale studies and exploring innovative treatment protocols that may further improve patient prognosis in this challenging clinical context.\u003c/p\u003e"},{"header":"5. Conclusion","content":"\u003cp\u003eThis study demonstrates the feasibility of an AI - driven framework of enhancing treatment decisions in CNS aspergillosis by leveraging meta - analytic insights and deep learning. The model\u0026rsquo;s ability to prioritize high - impact features and recommend adaptive therapies offers a promising tool to bridge evidence - based medicine with personalized ICU care. Further integration with real-time microbiological monitoring and drug - level assays could unlock its full potential in combating this refractory infection.\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003col\u003e\n \u003cli\u003eCNS:Central Nervous System\u003c/li\u003e\n \u003cli\u003eEMR: Electronic Medical Record\u003c/li\u003e\n \u003cli\u003eICU: Intensive Care Unit\u003c/li\u003e\n \u003cli\u003eCRF: Case Report Form\u003c/li\u003e\n \u003cli\u003eGBC: Gradient Boosting Classifier\u003c/li\u003e\n \u003cli\u003eRL: Reinforcement Learning\u003c/li\u003e\n \u003cli\u003eHIS :Hospital Inormation System\u003c/li\u003e\n \u003cli\u003eLinUCB: Linear Upper Confidence Bound\u003c/li\u003e\n\u003c/ol\u003e"},{"header":"Declarations","content":"\u003cp\u003e1\u003cstrong\u003e.Ethics approval and consent to participate\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eResearch conducted on the human data is in compliance with the Helsinki Declaration .Ethical approval was obtained from the Ethics Committee of the Second Affiliated Hospital of Zhejiang University School of Medicine.(Project acceptance number: Yan 2025-0557) with waiver of informed consent for retrospective analysis of anonymized records. \u0026nbsp;\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e2.\u003cstrong\u003eClinical trial registration\u0026nbsp;\u003c/strong\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eNot applicable. \u0026nbsp;\u003c/p\u003e\n\u003cp\u003e3.\u003cstrong\u003eConsent for publication\u0026nbsp;\u003c/strong\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eNot applicable. \u0026nbsp;\u003c/p\u003e\n\u003cp\u003e4.\u003cstrong\u003eAvailability of data and materials\u003c/strong\u003e \u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe datasets generated and/or analysed during the current study are available from the corresponding author on reasonable request. \u0026nbsp;\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e5.\u003cstrong\u003eCompeting interests \u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare that they have no competing interests. \u0026nbsp;\u003c/p\u003e\n\u003cp\u003e6.\u003cstrong\u003eFunding \u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable. \u0026nbsp;\u003c/p\u003e\n\u003cp\u003e7.\u003cstrong\u003eAuthors\u0026apos; contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e- Weina Lu: Conceptualization, Methodology, Writing \u0026ndash; original draft \u0026nbsp;\u003c/p\u003e\n\u003cp\u003e- Yilin Zhang: Formal analysis, Investigation \u0026nbsp;\u003c/p\u003e\n\u003cp\u003e- Ran Ji: Supervision, Writing \u0026ndash; review \u0026amp; editing \u0026nbsp;\u0026nbsp;\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eKomorowski AS, Hall CW, Atwal S, Johnstone R, Walker R, Mertz D, Piessens EA, Yamamura D, Kasper EM: \u003cstrong\u003eCerebrospinal fluid galactomannan detection for the diagnosis of central nervous system aspergillosis: a diagnostic test accuracy systematic review and meta-analysis\u003c/strong\u003e. \u003cem\u003eClin Microbiol Infect \u003c/em\u003e2024, \u003cstrong\u003e30\u003c/strong\u003e(10):1244-1253.\u003c/li\u003e\n\u003cli\u003eSerris A, Rautemaa-Richardson R, Laranjinha JD, Candoni A, Garcia-Vidal C, Alastruey-Izquierdo A, Hammarstr\u0026ouml;m H, Seidel D, Styczynski J, Sabino R\u003cem\u003e et al\u003c/em\u003e: \u003cstrong\u003eEuropean Study of Cerebral Aspergillosis treated with Isavuconazole (ESCAI): A study by the ESCMID Fungal Infection Study Group\u003c/strong\u003e. \u003cem\u003eClin Infect Dis \u003c/em\u003e2024, \u003cstrong\u003e79\u003c/strong\u003e(4):936-943.\u003c/li\u003e\n\u003cli\u003eMa Y, Li W, Ao R, Lan X, Li Y, Zhang J, Yu S: \u003cstrong\u003eCentral nervous system aspergillosis in immunocompetent patients: Case series and literature review\u003c/strong\u003e. \u003cem\u003eMedicine (Baltimore) \u003c/em\u003e2020, \u003cstrong\u003e99\u003c/strong\u003e(44):e22911.\u003c/li\u003e\n\u003cli\u003eSullivan BN, Baggett MA, O\u0026apos;Connell SS, Pickett KM, Steele C: \u003cstrong\u003eA Systematic Review to Assess the Relationship between Disseminated Cerebral Aspergillosis, Leukemias and Lymphomas, and Their Respective Therapeutics\u003c/strong\u003e. \u003cem\u003eJ Fungi (Basel) \u003c/em\u003e2022, \u003cstrong\u003e8\u003c/strong\u003e(7).\u003c/li\u003e\n\u003cli\u003eHoenigl M, Krause R: \u003cstrong\u003eAntifungal therapy of aspergillosis of the central nervous system and aspergillus endophthalmitis\u003c/strong\u003e. \u003cem\u003eCurr Pharm Des \u003c/em\u003e2013, \u003cstrong\u003e19\u003c/strong\u003e(20):3648-3668.\u003c/li\u003e\n\u003cli\u003eLu W, Ji R, Li W: \u003cstrong\u003eInvasive pulmonary and central nervous system aspergillosis: A case report and literature review\u003c/strong\u003e. \u003cem\u003eActa Microbiol Immunol Hung \u003c/em\u003e2025, \u003cstrong\u003e72\u003c/strong\u003e(1).\u003c/li\u003e\n\u003cli\u003eBoes B, Bashir R, Boes C, Hahn F, McConnell JR, McComb R: \u003cstrong\u003eCentral nervous system aspergillosis. Analysis of 26 patients\u003c/strong\u003e. \u003cem\u003eJ Neuroimaging \u003c/em\u003e1994, \u003cstrong\u003e4\u003c/strong\u003e(3):123-129.\u003c/li\u003e\n\u003cli\u003eMaciejewski R, Hafen R, Rudolph S, Larew SG, Mitchell MA, Cleveland WS, Ebert DS: \u003cstrong\u003eForecasting Hotspots-A Predictive Analytics Approach\u003c/strong\u003e. \u003cem\u003eIEEE Trans Vis Comput Graph \u003c/em\u003e2011, \u003cstrong\u003e17\u003c/strong\u003e(4):440-453.\u003c/li\u003e\n\u003cli\u003eTaylor AR, Young BD, Levine GJ, Eden K, Corapi W, Rossmeisl JH, Levine JM: \u003cstrong\u003eClinical Features and Magnetic Resonance Imaging Findings in 7 Dogs with Central Nervous System Aspergillosis\u003c/strong\u003e. \u003cem\u003eJ Vet Intern Med \u003c/em\u003e2015, \u003cstrong\u003e29\u003c/strong\u003e(6):1556-1563.\u003c/li\u003e\n\u003cli\u003eGuez A. \u003cstrong\u003eAdaptive control of epileptic seizures using reinforcement learning\u003c/strong\u003e \u003cem\u003eJ\u003c/em\u003e 2010.\u003c/li\u003e\n\u003cli\u003eGuez A, Vincent R D, Avoli M, et al. \u003cstrong\u003eAdaptive Treatment of Epilepsy via Batch-mode Reinforcement Learning\u003c/strong\u003e\u003cem\u003e C AAAI. \u003c/em\u003e2008, 8: 1671-1678.\u003c/li\u003e\n\u003cli\u003ePineau J, Guez A, Vincent R, et al. \u003cstrong\u003eTreating epilepsy via adaptive neurostimulation: a reinforcement learning approach\u003c/strong\u003e \u003cem\u003eJ International journal of neural systems\u003c/em\u003e, 2009, \u003cstrong\u003e19\u003c/strong\u003e(04): 227-240.\u003c/li\u003e\n\u003cli\u003eEscandell-Montero P, Mart\u0026iacute;nez-Mart\u0026iacute;nez J M, Mart\u0026iacute;n-Guerrero J D, et al. \u003cstrong\u003eAdaptive treatment of anemia on hemodialysis patients: A reinforcement learning approach\u003c/strong\u003e \u003cem\u003eC 2011 IEEE Symposium on Computational Intelligence and Data Mining (CIDM). IEEE\u003c/em\u003e, 2011: 44-49.\u003c/li\u003e\n\u003cli\u003eMalof J M, Gaweda A E. \u003cstrong\u003eOptimizing drug therapy with reinforcement learning: The case of anemia management\u003c/strong\u003e \u003cem\u003eC The 2011 International Joint Conference on Neural Networks. IEEE\u003c/em\u003e, 2011: 2088-2092.\u003c/li\u003e\n\u003cli\u003eWu, X., Li, R., He, Z. \u003cem\u003eet al.\u003c/em\u003e \u003cstrong\u003eA value-based deep reinforcement learning model with human expertise in optimal treatment of sepsis\u003c/strong\u003e. \u003cem\u003enpj Digit. Med.\u003c/em\u003e \u003cstrong\u003e6\u003c/strong\u003e, 15 (2023).\u003c/li\u003e\n\u003cli\u003eShukla A, Bansal M, Husain M, Chhabra DK: \u003cstrong\u003eCentral nervous system mycosis: analysis of 10 cases\u003c/strong\u003e. \u003cem\u003eIndian J Pathol Microbiol \u003c/em\u003e2014, \u003cstrong\u003e57\u003c/strong\u003e(4):591-594.\u003c/li\u003e\n\u003cli\u003eDotis J, Iosifidis E, Roilides E: \u003cstrong\u003eCentral nervous system aspergillosis in children: a systematic review of reported cases\u003c/strong\u003e. \u003cem\u003eInt J Infect Dis \u003c/em\u003e2007, \u003cstrong\u003e11\u003c/strong\u003e(5):381-393.\u003c/li\u003e\n\u003cli\u003eMody CH, Warren PW: \u003cstrong\u003eHost defence to pulmonary mycosis\u003c/strong\u003e. \u003cem\u003eCan J Infect Dis \u003c/em\u003e1999, \u003cstrong\u003e10\u003c/strong\u003e(2):147-155.\u003c/li\u003e\n\u003cli\u003eMylonakis E, Paliou M, Sax PE, Skolnik PR, Baron MJ, Rich JD: \u003cstrong\u003eCentral nervous system aspergillosis in patients with human immunodeficiency virus infection. Report of 6 cases and review\u003c/strong\u003e. \u003cem\u003eMedicine (Baltimore) \u003c/em\u003e2000, \u003cstrong\u003e79\u003c/strong\u003e(4):269-280.\u003c/li\u003e\n\u003cli\u003eYazdani A, Safaei A A, Safdari R, Zahmatkeshan M. \u003cstrong\u003eDiagnosis of Breast Cancer Using Decision Tree, Artificial Neural Network and Naive Bayes to Provide a Native Model for Fars Province. \u003c/strong\u003e\u003cem\u003ePayavard \u003c/em\u003e2019; 13 (3) :241-250\u003c/li\u003e\n\u003cli\u003eHarper PR. \u003cstrong\u003eA review and comparison of classification algorithms for medical decision making. \u003c/strong\u003e\u003cem\u003eHealth Policy.\u003c/em\u003e 2005 Mar;71(3):315-31.\u003c/li\u003e\n\u003cli\u003eLal Dahti J, Mohammadi M, Padidaran Moghadam F. \u003cstrong\u003eA Method for the Diagnosis of Metabolic Syndrome based on KNN Data Mining Algorithm: A case study in Shohada-ye Kargar Hospital in Yazd, Iran.\u003c/strong\u003e \u003cem\u003eHealth and Biomedical Informatics\u003c/em\u003e 2018; 4 (4) :291-304\u003c/li\u003e\n\u003c/ol\u003e"},{"header":"Tables","content":"\u003cp\u003eTable 1. Survival Rates Across Algorithms and Exploration Parameters (Alpha Values)\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"539\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 76px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAlgorithm\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 38px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.0\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 38px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.1\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 38px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.2\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 38px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.3\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 38px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.4\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 38px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.5\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 38px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.6\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 38px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.7\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 38px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.8\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 38px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.9\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 38px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e1.0\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 47px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAvg.\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 76px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eBaseline\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"11\" valign=\"top\" style=\"width: 416px;\"\u003e\n \u003cp\u003e50%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 47px;\"\u003e\n \u003cp\u003e50%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 76px;\"\u003e\n \u003cp\u003e\u003cimg width=\"7\" height=\"15\" src=\"data:image/png;base64,R0lGODlhCgAXAHcAMSH+GlNvZnR3YXJlOiBNaWNyb3NvZnQgT2ZmaWNlACH5BAEAAAAALAAACQAKAAoAhAAAAAAAAAAAOgAAZgA6kABmtjoAADo6Ojo6kDpmkDpmtmYAAGaQtma225A6AJBmOpDb/7aQZrb//9uQOtu2Ztv///+2Zv/bkP//tv//2wECAwECAwECAwECAwECAwECAwU3ICBSR2AKiTiZDXAtBZA5AVGJOGAFApSLmYjBNJCILEPFDYdZ1Ja4yzD2m/F8QcQt8xgGDoxbCAA7\" alt=\"image\"\u003e\u003cstrong\u003e-greedy\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 38px;\"\u003e\n \u003cp\u003e\u003cu\u003e53%\u003c/u\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 38px;\"\u003e\n \u003cp\u003e56%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 38px;\"\u003e\n \u003cp\u003e59%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 38px;\"\u003e\n \u003cp\u003e59%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 38px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e61%\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 38px;\"\u003e\n \u003cp\u003e54%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 38px;\"\u003e\n \u003cp\u003e56%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 38px;\"\u003e\n \u003cp\u003e54%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 38px;\"\u003e\n \u003cp\u003e56%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 38px;\"\u003e\n \u003cp\u003e57%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 38px;\"\u003e\n \u003cp\u003e59%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 47px;\"\u003e\n \u003cp\u003e56.7%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 76px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eLinUCB\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 38px;\"\u003e\n \u003cp\u003e\u003cu\u003e67%\u003c/u\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 38px;\"\u003e\n \u003cp\u003e75%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 38px;\"\u003e\n \u003cp\u003e75%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 38px;\"\u003e\n \u003cp\u003e75%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 38px;\"\u003e\n \u003cp\u003e75%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 38px;\"\u003e\n \u003cp\u003e75%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 38px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e83%\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 38px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e83%\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 38px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e83%\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 38px;\"\u003e\n \u003cp\u003e75%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 38px;\"\u003e\n \u003cp\u003e75%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 47px;\"\u003e\n \u003cp\u003e76.5%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 76px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eRandom\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 38px;\"\u003e\n \u003cp\u003e\u003cu\u003e33%\u003c/u\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 38px;\"\u003e\n \u003cp\u003e50%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 38px;\"\u003e\n \u003cp\u003e42%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 38px;\"\u003e\n \u003cp\u003e50%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 38px;\"\u003e\n \u003cp\u003e50%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 38px;\"\u003e\n \u003cp\u003e\u003cu\u003e33%\u003c/u\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 38px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e67%\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 38px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e67%\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 38px;\"\u003e\n \u003cp\u003e50%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 38px;\"\u003e\n \u003cp\u003e50%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 38px;\"\u003e\n \u003cp\u003e50%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 47px;\"\u003e\n \u003cp\u003e49.3%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eThis table reports survival outcomes for three treatment recommendation algorithms\u0026mdash;LinUCB, \u0026epsilon;-greedy, and Random\u0026mdash;across different alpha (\u0026alpha;) values, which regulate exploration vs. exploitation behavior. LinUCB outperformed other models at nearly all settings, achieving peak performance at \u0026alpha; = 0.6\u0026ndash;0.8. The base survival rate of 0.50 reflects clinician-chosen treatments without AI intervention.\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eTable 2. Impact of Missing Patient Features on LinUCB Model Survival Prediction\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 424px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 106px;\"\u003e\n \u003cp\u003eSurvival rate\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 424px;\"\u003e\n \u003cp\u003eLinUCB(\u003cimg width=\"43\" height=\"15\" src=\"data:image/png;base64,R0lGODlhQQAXAHcAMSH+GlNvZnR3YXJlOiBNaWNyb3NvZnQgT2ZmaWNlACH5BAEAAAAALAAABgBBAA0AhQAAAAAAAAAAOgAAZgA6ZgA6kABmtjoAADpmZjpmkDpmtjqQ22YAAGYAOmY6AGZmkGaQ22a2tma222a2/5A6AJBmOpDb/7ZmALZmOraQOrbb/7b//9uQOtu2Ztu2kNv///+2Zv/bkP/btv//tv//2wECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwbYQIBwSCwaj0hj5xAIKD5HUHPaHGyS2GyWI5AAPAfDkUOdirXoNGDEOAMugYmRsyCOGha1HiuVC6V1WwVQQyIVTQoAIQeBe0kXAnlCi25IbH6TYR8kFBMXVmNloo1FkJKKB4NYHKpCbK0cCAeYjkimQ6+ER4u0ZJhklbVHt64MrcOgQpzJb8zCw3FDlEmLpLmZwUVkolSkRL5DgI/O15wB3s9F00Jcp+qMRcsahhEMCyIPuulEcF6LiWsYMPt0RZ2DAAQ0AMAQIIHCfUZIZGAiAAIugQXZZAsCADs=\" alt=\"image\"\u003e)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 106px;\"\u003e\n \u003cp\u003e83.3%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 424px;\"\u003e\n \u003cp\u003eLinUCB w/o Age\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 106px;\"\u003e\n \u003cp\u003e\u003cimg width=\"6\" height=\"15\" src=\"data:image/png;base64,R0lGODlhCQAXAHcAMSH+GlNvZnR3YXJlOiBNaWNyb3NvZnQgT2ZmaWNlACH5BAEAAAAALAAABQAJAA4AhAAAAAAAAAA6kABmtjoAADpmtjqQkDqQ22YAAGY6AGY6ZmaQkGa2/5Db/7ZmOtu2Ztv///+2Zv/bkP//tv//2wECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwUvICBOCCOeAGmiacm2K6q+M1vLrp3j8aQ0KorjAHgEDIhFQgARRQiBwKB5khAKpxAAOw==\" alt=\"image\"\u003e\u0026nbsp;9.9%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 424px;\"\u003e\n \u003cp\u003eLinUCB w/o Gender\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 106px;\"\u003e\n \u003cp\u003e\u003cimg width=\"6\" height=\"15\" src=\"data:image/png;base64,R0lGODlhCQAXAHcAMSH+GlNvZnR3YXJlOiBNaWNyb3NvZnQgT2ZmaWNlACH5BAEAAAAALAAABQAJAA4AhAAAAAAAAAA6kABmtjoAADpmtjqQkDqQ22YAAGY6AGY6ZmaQkGa2/5Db/7ZmOtu2Ztv///+2Zv/bkP//tv//2wECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwUvICBOCCOeAGmiacm2K6q+M1vLrp3j8aQ0KorjAHgEDIhFQgARRQiBwKB5khAKpxAAOw==\" alt=\"image\"\u003e\u0026nbsp;19.6%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 424px;\"\u003e\n \u003cp\u003eLinUCB w/o Country\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 106px;\"\u003e\n \u003cp\u003e\u003cimg width=\"6\" height=\"15\" src=\"data:image/png;base64,R0lGODlhCQAXAHcAMSH+GlNvZnR3YXJlOiBNaWNyb3NvZnQgT2ZmaWNlACH5BAEAAAAALAAABQAJAA4AhAAAAAAAAAA6kABmtjoAADpmtjqQkDqQ22YAAGY6AGY6ZmaQkGa2/5Db/7ZmOtu2Ztv///+2Zv/bkP//tv//2wECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwUvICBOCCOeAGmiacm2K6q+M1vLrp3j8aQ0KorjAHgEDIhFQgARRQiBwKB5khAKpxAAOw==\" alt=\"image\"\u003e\u0026nbsp;9.9%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 424px;\"\u003e\n \u003cp\u003eLinUCB w/o Type of CNS infection\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 106px;\"\u003e\n \u003cp\u003e\u003cimg width=\"9\" height=\"15\" src=\"data:image/png;base64,R0lGODlhDgAXAHcAMSH+GlNvZnR3YXJlOiBNaWNyb3NvZnQgT2ZmaWNlACH5BAEAAAAALAAABQAJAA4AhAAAAAAAAAA6kABmtjoAADpmtjqQkDqQ22YAAGY6AGY6ZmaQkGa2/5Db/7ZmOtu2Ztv///+2Zv/bkP//tv//2wECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwUvICBOCCOeAGmiacm2K6q+M1vLrp3j8aQ0KorjAHgEDIhFQgARRQiBwKB5khAKpxAAOw==\" alt=\"image\"\u003e9.9%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 424px;\"\u003e\n \u003cp\u003eLinUCB w/o Clinical presentation\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 106px;\"\u003e\n \u003cp\u003e\u003cimg width=\"6\" height=\"15\" src=\"data:image/png;base64,R0lGODlhCQAXAHcAMSH+GlNvZnR3YXJlOiBNaWNyb3NvZnQgT2ZmaWNlACH5BAEAAAAALAAABQAJAA4AhAAAAAAAAAA6kABmtjoAADpmtjqQkDqQ22YAAGY6AGY6ZmaQkGa2/5Db/7ZmOtu2Ztv///+2Zv/bkP//tv//2wECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwUvICBOCCOeAGmiacm2K6q+M1vLrp3j8aQ0KorjAHgEDIhFQgARRQiBwKB5khAKpxAAOw==\" alt=\"image\"\u003e\u0026nbsp;19.6%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 424px;\"\u003e\n \u003cp\u003eLinUCB w/o Imagological findings\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 106px;\"\u003e\n \u003cp\u003e\u003cimg width=\"6\" height=\"15\" src=\"data:image/png;base64,R0lGODlhCQAXAHcAMSH+GlNvZnR3YXJlOiBNaWNyb3NvZnQgT2ZmaWNlACH5BAEAAAAALAAABQAJAA4AhAAAAAAAAAA6kABmtjoAADpmtjqQkDqQ22YAAGY6AGY6ZmaQkGa2/5Db/7ZmOtu2Ztv///+2Zv/bkP//tv//2wECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwUvICBOCCOeAGmiacm2K6q+M1vLrp3j8aQ0KorjAHgEDIhFQgARRQiBwKB5khAKpxAAOw==\" alt=\"image\"\u003e\u0026nbsp;19.6%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 424px;\"\u003e\n \u003cp\u003eBaseline\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 106px;\"\u003e\n \u003cp\u003e50%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eThis table presents a feature ablation experiment evaluating the effect of masking individual patient features on predicted survival rate using the LinUCB model. Missing clinical presentation, sex, or imageological information reduced model-predicted survival from 0.83 to 0.67, highlighting the need for complete, high-quality data input for optimal model performance.\u003c/p\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"bmc-infectious-diseases","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"infd","sideBox":"Learn more about [BMC Infectious Diseases](http://bmcinfectdis.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/infd","title":"BMC Infectious Diseases","twitterHandle":"#bmcinfectdis","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-6888468/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6888468/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eBackground\u003c/h2\u003e\u003cp\u003eCentral nervous system (CNS) aspergillosis is a rare but highly fatal infection, particularly among immunocompromised individuals. Timely diagnosis and optimal treatment selection are crucial for improving patient outcomes, yet clinical decision-making remains challenging.\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e\u003cp\u003eWe integrated clinical data from 64 published CNS aspergillosis cases (2014\u0026ndash;2024) and structured electronic medical records (EMRs) from 200 ICU patients. After preprocessing (one-hot encoding, Z-score standardization, BERT-based text feature extraction), a Gradient Boosting Classifier (GBC) was trained to predict 30-day survival. Additionally, a LinUCB-based adaptive treatment policy was developed to dynamically optimize therapy choices. Model performance was evaluated against logistic regression, random forest models, and baseline treatment policies.\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e\u003cp\u003eThe GBC model achieved 83% accuracy in predicting 30-day survival, outperforming logistic regression (72%) and random forests (78%). Key mortality predictors included older age, multiple CNS lesions, and delayed antifungal therapy. Feature ablation analysis confirmed the critical impact of clinical presentation, imaging findings, and treatment delay. The LinUCB adaptive policy demonstrated superior cumulative survival gain compared to random and ε-greedy strategies, achieving a stabilized survival probability of 0.81 by simulation step 300.\u003c/p\u003e\u003ch2\u003eConclusion\u003c/h2\u003e\u003cp\u003eIntegrating meta-analytic and EMR-derived data with machine learning models can accurately predict survival and inform adaptive treatment strategies in CNS aspergillosis. The proposed LinUCB-guided approach offers a promising framework for real-time, personalized decision-making in critically ill patients.\u003c/p\u003e","manuscriptTitle":"Deep Learning on Meta-Analytic Data for Therapeutic Decision-Making in Central Nervous System Aspergillosis","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-08-25 20:53:03","doi":"10.21203/rs.3.rs-6888468/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2025-10-14T19:33:56+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-08-21T19:30:32+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"75354765130979499531659486678652590491","date":"2025-08-11T12:05:54+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-08-06T02:10:57+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"115206144848568496383569745245939353802","date":"2025-08-06T00:59:36+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-08-04T09:23:24+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-07-30T09:22:29+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2025-07-11T05:50:13+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-07-10T13:40:18+00:00","index":"","fulltext":""},{"type":"submitted","content":"BMC Infectious Diseases","date":"2025-07-10T12:15:54+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"bmc-infectious-diseases","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"infd","sideBox":"Learn more about [BMC Infectious Diseases](http://bmcinfectdis.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/infd","title":"BMC Infectious Diseases","twitterHandle":"#bmcinfectdis","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"1ac69b98-9631-468a-ad4e-a47b02001754","owner":[],"postedDate":"August 25th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[],"tags":[],"updatedAt":"2026-01-19T16:47:48+00:00","versionOfRecord":{"articleIdentity":"rs-6888468","link":"https://doi.org/10.1186/s12879-026-12573-7","journal":{"identity":"bmc-infectious-diseases","isVorOnly":false,"title":"BMC Infectious Diseases"},"publishedOn":"2026-01-17 16:29:20","publishedOnDateReadable":"January 17th, 2026"},"versionCreatedAt":"2025-08-25 20:53:03","video":"","vorDoi":"10.1186/s12879-026-12573-7","vorDoiUrl":"https://doi.org/10.1186/s12879-026-12573-7","workflowStages":[]},"version":"v1","identity":"rs-6888468","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6888468","identity":"rs-6888468","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}