Predicting participant attrition in paediatric clinical trials using machine learning and deep learning models

Predicting participant attrition in paediatric clinical trials using machine learning and deep learning models | 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 Predicting participant attrition in paediatric clinical trials using machine learning and deep learning models Mailosi Innussa, Priscilla Maliwichi, Clement Nyirenda This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8970473/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Participant attrition in paediatric clinical trials remains a major challenge, leading to higher costs, longer trial durations, and compromised research outcomes. Existing retention and prediction methods have shown limited success. This study aimed to predict participant attrition in paediatric clinical trials using machine learning and deep learning models. Secondary data from a paediatric clinical trial in Malawi were used. Logistic regression, Random forest, Multi-layer perceptron, and One convolutional neural network models were trained and evaluated under four data settings: original, SMOTE- enhanced, GAN-augmented, and combined. Predictive performance was assessed using macro-averaged F1-score as the primary metric. Random forest achieved the highest F1- score of 0.514 on the SMOTE-enhanced dataset. Overall performance remained modest, and GAN augmentation did not consistently improve results. The findings suggest that data augmentation techniques provide limited benefits for attrition prediction in this context. Artificial Intelligence and Machine Learning Paediatric clinical trial Attrition prediction Machine learning Deep learning Data augmentation Figures Figure 1 Figure 2 Figure 3 1 Introduction Attrition in clinical trials refers to patient dropout or withdrawal. It is one of the significant challenges faced by researchers in clinical trials [ 1 ]. Attrition occurs when participants either voluntarily withdraw or are lost to follow-up by the clinical site and trial sponsor [ 2 ]. A study has identified that an 88% dropout rate is due to factors such as loss to follow-up, non-compliance with the protocol, and withdrawal of consent [ 3 ]. Other studies have found that the rates of attrition range from 3 to 63% [ 4 ]. It is generally considered that bias becomes a significant concern when the attrition rate surpasses 20% [ 5 ]. Attrition becomes more complicated in paediatric populations because trial involvement relies on both parental and child behaviours [ 6 ]. Attrition can cause the trial to last longer than anticipated, which may lead to additional financial costs. Additionally, attrition has the potential to heavily bias data, reduce statistical power, and undermine the overall statistical validity of an intervention. It may also make future study requirements more difficult [ 1 , 7 ]. High attrition rates can give the impression that the trial’s design and execution were poorly handled, which might lead to a loss of trust in clinical research [ 8 ]. Prior research has identified multiple factors associated with attrition, including lack of trust, socioeconomic barriers, participant inconvenience, real or perceived adverse events of intervention, migration from study sites, media influence, procedural complexity, and the volume of biological samples collected [ 8 – 10 ]. Trialists employ various methods to reduce attrition, such as building strong relation- ships between participants and site staff, offering incentives, enhancing communication, showing respect for participants, and ensuring effective informed consent. However, the success of these strategies often depends on the specific context [ 8 , 11 ]. Recently, machine learning and deep learning have been applied to predict attrition in clinical studies. Literature presents the potential of machine learning (ML) and deep learning (DL) in predicting attrition in clinical studies. However, significant gaps remain in the research, as much of the prior work has focused on adult populations [ 12 – 14 ], general populations [ 2 , 15 ], preterm infants [ 16 ], or weight management interventions [ 17 ]. These models may not be directly applicable to paediatric clinical trials, where attrition pat- terns are influenced by unique factors related to trial procedures, child participants, and caregiver involvement. In addition, attrition prediction datasets often suffer from severe class imbalance and limited sample sizes, which pose significant challenges to predictive performance [ 18 , 19 ]. To address these gaps, this study develops and evaluates machine learning models for predicting attrition in paediatric clinical trials. Data from completed clinical trials are collected and preprocessed, with particular emphasis on handling class imbalance using the Synthetic Minority Over-sampling Technique (SMOTE) [ 20 ] and investigating Generative Adversarial Network (GAN)-based data augmentation as a strategy for mitigating data scarcity [ 21 ]. The study then compared the predictive performance of Logistic regression(LR), Random forest (RF), Multilayer Perceptron (MLP), and One convolutional neural network (1DCNN) models, and identified and ranked features contributing to attrition. This approach allows assessment of model performance while providing insights into the key factors influencing dropout. This study makes several contributions to clinical trial research: Provides a comparative evaluation of LR, RF, MLP, and 1DCNN models in the context of paediatric attrition. Demonstrates the impact of data enhancement techniques, including SMOTE, GAN, and a combined approach, on predictive performance. Identifies and ranks the most influential features associated with dropout, offering actionable insights for designing trials that minimize attrition and improve participant retention. 2 Methods 2.1 Data Collection and Preparation The dataset comprised data from 2,168 participants aged six months enrolled in a paediatric clinical trial conducted in Malawi. Eighteen participants who exited the study due to death were excluded to ensure that the analysis focused specifically on attrition outcomes [ 2 ]. fourteen predictor were selected based on their relevance and prior evidence reported in the literature [ 6 , 9 , 10 , 22 – 25 ]. The predictor features included child age, child sex, residential area, maternal age, household size, number of household members under five years of age, father’s education level, mother’s education level, father’s occupation, mother’s occupation, family income source, monthly income category, income–expenditure deficit indicator, and adverse event count. The outcome variable was attrition. 2.1.1 Data preprocessing Missing values were identified and imputed using the K-Nearest Neighbors (KNN) imputation method to preserve dataset integrity. Outliers were detected through boxplot visualizations and addressed via Winsorizing to reduce their impact on the models [ 26 , 27 ]. Additionally, 18 participants who exited the study due to death were excluded from the dataset to focus the analysis on attrition outcomes [ 2 ]. Categorical variables were encoded using one-hot encoding, and all features were normalized to a common scale to ensure balanced model training [ 28 , 29 ]. 2.2 Exploratory Analysis and Data Enhancement To explore feature relationships, a correlation matrix was computed (Fig. 1 ). The matrix uses color gradients from blue to red, where grey indicates no correlation, red indicates strong positive correlation, and blue indicates negative correlation. Notable correlations observed include positive correlations between ’age mother’ with ’household number’, ’edu mother’ with ’edu father’, and ’household member5’ with ’household number’; negative correlations between ’occup mother’ with ’income source’ and ’edu mother’ with ’age mother’; and correlation between parental occupations: ’occup mother’ and ’occup father’. These relationships offer insights into underlying data patterns and inform feature selection. Feature importance was further assessed using SHAP values to quantify each feature’s contribution to prediction. Class imbalance was evident, with 1,750 participants completing the study and only 400 exiting early, representing an 18.6% attrition rate. Imbalanced classes can bias model performance toward the majority class. To mitigate this, Synthetic Minority Oversampling Technique (SMOTE) was applied to generate synthetic minority class samples, achieving balanced class distribution for improved classifier performance [20]. To further enhance generalization and reduce overfitting, a conditional tabular GAN was employed for synthetic data generation. The neural generator comprised fully connected layers with LeakyReLU activations and batch normalization, while the adversarial component was implemented using a LightGBM regressor (max depth = 2, learning rate = 0.02, n estimators = 500). Training parameters included ten-fold minor- ity oversampling gen x times = 10), outlier filtering (bot filter quantile = 0.0001 and top filter quantile = 0.999), and default library settings (batch size ≈ 500 with early stopping). This approach aimed to increase dataset diversity and model robustness [21, 30]. 2.3 Machine Learning Model Development and Evaluation Four machine learning algorithms, Logistic Regression, Random Forest, MLP, and 1DCNN were employed to classify participant attrition. Logistic Regression was included as a widely used baseline in attrition research. Models were trained using a stratified train-test split (70% training, 30% testing), with 10% of the training data reserved for validation. Eight features were selected from the initial 14 candidates in Data Collection using mean absolute SHAP values from the best-performing Random Forest model. These features age child, age mother, edu mother, resident, ae count, household num, deficit income, and child sex were retained as they contributed the majority of predictive signal (mean absolute SHAP > 0.02), with domain knowledge confirming their clinical relevance. All models were trained using these selected features across four experimental conditions: Original Dataset: Using the original preprocessed dataset without any balancing or augmentation SMOTE-Enhanced: Applying SMOTE to balance classes by generating synthetic minority samples GAN-Augmented: Using GAN-augmented data to increase dataset diversity Combined Enhancement: Combining SMOTE with GAN augmentation for balanced and diverse data Each model was trained and tested across ten independent runs under consistent conditions, with average performance metrics reported. Evaluation metrics included accuracy, precision, recall, F1-score, AUC-ROC, and confusion matrices [ 31 ]. The F1-score was used as the primary metric due to the dataset’s imbalance, as it provides a balanced measure of precision and recall, making it more reliable than accuracy or AUC for assessing the minority class. This metric is commonly applied in imbalanced classification tasks and related domains [ 31 – 33 ]. Statistical significance of performance differences was assessed using paired t-tests for within-model comparisons and one-way ANOVA for comparisons across models and configurations. 2.4 Ethical Consideration This study received ethical approval from the Malawi University of Science and Technology Research Committee (MUSTREC) under reference number P.01/2025/391. All ethical standards for research without direct contact with human or animal subjects were followed. 3 Results Our comparative analysis of four machine learning models across four data settings reveals significant variation in peadiatric attrition prediction performance as shown in Table 1 . The average F1-Scores across all algorithms and data setting ranged from 0.4487 to 0.514. Random forest achieved the highest F1-score (0.514 ± 0.01) with SMOTE-enhanced data. Logistic regression (0.4932 ± 0.40) and 1DCNN (0.5093 ± 1.21) achieved their best score on original dataset, while MLP achieved its best performance (0.5053 ± 0.76) with SMOTE enhancement. SMOTE augmentation resulted in notable performance improvements for Random forest (+ 0.0024) and MLP (+ 0.0566) but led to minor declines for Logistic regression (-0.0292) and 1DCNN (-0.0013). GAN augmentation negatively affected Random forest (-0.0375), Logistic regression (-0.0252) and 1DCNN (-0.0319) and did not result in any significant improvement for MLP. Combined augmentation slightly improved the performance of MLP (+ 0.0064) while adversely affecting the other models. Figure 2 presents the performance of all models across the different data settings. Table 1 F1-Scores Across Data Settings Model Original SMOTE GAN Combined Logistic regression 0.4932 0.464 0.4791 0.4613 Random forest 0.5116 0.514 0.4741 0.4551 MLP 0.4487 0.5053 0.4487 0.4551 1DCNN 0.5093 0.508 0.4832 0.4522 To provide a more comprehensive overview, Table 2 summarises the best-performing model for each data setting. On the original dataset, the Random forest model achieved an accuracy of 57.05%, with a precision of 0.5498, recall of 0.5819, and an AUC–ROC of 0.5848. Under the SMOTE-enhanced setting, Random forest again performed best, with marginal increases in precision (0.5537), recall (0.5883), and AUC–ROC (0.5874), while overall accuracy remained unchanged. In contrast, performance declined under GAN- based augmentation, where the 1DCNN achieved an accuracy of 55.66% alongside lower precision, recall, and AUC–ROC values. In the combined data setting, Logistic regression yielded the highest performance; however, accuracy (54.88%) and discrimination ability (AUC–ROC of 0.4870) remained comparatively low. Table 2 Best Model Performance by Data Setting Setting Model F1 Precision Recall AUC-ROC Accuracy% Original RF 0.5116 0.5498 0.5819 0.5848 57.05 SMOTE RF 0.5140 0.5537 0.5883 0.5874 57.05 GAN 1DCNN 0.4832 0.5174 0.5145 0.5296 55.66 Combined LR 0.4613 0.4926 0.4882 0.4870 54.88 The confusion matrix prediction performance evaluation results is presented in Fig. 3 . Under the original data setting, the Random forest model correctly classified 56.38% of non-attrition cases (true negatives) while misclassifying 43.62% as attrition (false positives), and achieved a recall of 60.00% for attrition cases, with 40.00% false negatives. With SMOTE enhancement, Random forest improved attrition recall to 61.67% (true positives), at the slight expense of a reduced true negative rate of 56.00% and an increase in false positives to 44.00%. Logistic regression achieved a true negative rate of 58.48%, with 42.52% false positives, and correctly identified 39.17% of attrition cases. The 1DCNN model exhibited a true negative rate of 51.24% and a false positive rate of 48.33%, while attrition recall remained at 51.67%. Overall, these confusion matrices indicate that Random forest with SMOTE enhancement provided stronger detection of the minority attrition class (recall of 61.67%). SHAP analysis on the best-performing model ranked features by their contribution to attrition predictions. The top influencers were residential area (mean SHAP value: 0.12), adverse event count (0.10), maternal age (0.09), child age (0.08), and maternal education level (0.07). Features like household size and income-expenditure deficit had moderate impact (0.05–0.06), while child sex was least influential (0.03). 4 Discussion The results of this study demonstrated the potential of machine learning and deep learning models for addressing participant attrition in paediatric clinical trials. Across all data settings, the Random forest model demonstrated comparatively stronger performance, particularly under SMOTE augmentation, where it achieved the highest macro-averaged F1-score of 0.514 and improved attrition recall of 61.67% (true positives), highlighting its robustness in handling class imbalance and detecting attrition cases. Although the improvement over the original dataset was marginal (+ 0.0024 in F1-score), this finding suggests that SMOTE can partially mitigate bias toward the majority class and enhance recall for the minority attrition class without substantial loss of precision. This is consistent with findings reported by [ 15 ]. The MLP model exhibited a larger performance gain under SMOTE augmentation (+ 0.0566), suggesting that its performance is more strongly influenced by class distribution than that of the other models. Comparatively, GAN augmentation and the combined approach yielded negative impacts on performance for most models, with declines in F1-scores ranging from − 0.0013 to -0.0375. The decline in performance with GAN augmentation is likely due to mode collapse, unstable training, and poor-quality synthetic samples on small, highly imbalanced tabular clinical data [ 34 – 36 ]. Unlike SMOTE, which reliably interpolates real minority instances, GANs struggle with scarce examples, often adding noise rather than benefit [ 35 ].These results suggest that more complex data augmentation techniques may not be well suited to the underlying structure and size of paediatric clinical trial data. The SHAP-based feature importance analysis provides valuable insights into attrition drivers, emphasizing sociodemographic and trial-related factors such as residential area, adverse event count, maternal age, child age, and maternal education. These align with existing literature highlighting socioeconomic barriers, participant inconvenience and caregiver characteristics as key contributors to dropout in paediatric trials [ 14 , 17 , 37 ]. Notably, residential area high influence (mean SHAP: 0.12) may reflect access to health facility challenges in low-resource settings including relocation and transportation issues [ 10 ]. Conversely, the lower impact of household size, income, and occupations contrasts with some studies [ 8 ], and suggesting that, in this context, individual caregiver attributes may play a more significant role than broader economic factors. Strengths of this study include its focus on attrition prediction in a paediatric clinical trial from a low-income setting, addressing a notable gap in the existing literature, which is largely dominated by adult or general populations [ 6 , 12 – 14 , 16 , 17 ]. The study demonstrates the substantial impact of data augmentation techniques on improving model performance. Additionally, it identifies and ranks the most influential factors contributing to attrition, providing actionable insights for this specific setting. Overall, the comparative evaluation of multiple models and augmentation strategies offers practical guidance for clinical trial researchers developing data-driven tools to enhance participant retention. Nonetheless, several limitations remain. The data, drawn from a low- income country, may not generalise to other contexts and no external validation was done. Furthermore, the highest F1-score achieved was only 0.514, with moderate recall (0.5883) and a high false negative rate (44%), limiting stand-alone clinical utility for early intervention. Unmeasured variables like trial complexity, relocation, or media influence known attrition factors [ 15 ] may have contributed to the models’ modest overall performance. These results have implications for clinical trial design, suggesting that integrating Random forest models with SMOTE could enable early identification of at-risk participants, allowing targeted interventions such as enhanced communication or incentives [ 11 ]. Future research should validate these models on multi-site, international datasets to improve robustness. Expanding to additional machine and deep learning algorithms may identify models that are better aligned with the characteristics of peadiatric clinical trial datasets. 5 Conclusion The results from the four experiments indicate the Random forest model demonstrates clear potential in predicting participant attrition in peadiatric clinical trials. The application of SMOTE alone effectively enhanced model performance by improving recall of at-risk participants, whereas GAN augmentation and combined approaches negatively impacted results for most models. Key predictors of attrition were identified and ranked, with residential area, adverse event count, maternal age, child age, and maternal education emerging as the most influential factors. While these insights provide a meaningful foundation for targeted retention strategies, further research is required to improve prediction accuracy and generalisability across diverse clinical and demographic settings. References Scout. Understanding Clinical Trial Patient Attrition (2023) Accessed: 2024-10-25. Available from: https://www.clinicalleader.com/doc/ understanding-clinical-trial-patient-attrition-0001 Hutchison E, Zhang Y, Nampally S, Neelufer IK, Malkov V, Weatherall J et al (2021) Modeling Clinical Trial Attrition Using Machine Intelligence: A driver analytics case study using 1,325 trials representing one million patients. medRxiv. :2021–2011 Desai M (2020) Recruitment and retention of participants in clinical studies: critical issues and challenges. Medknow Farris MS, Devoe DJ, Addington J (2020) Attrition rates in trials for adolescents and young adults at clinical high risk for psychosis: A systematic review and meta-analysis. Early Interv Psychiat 14(5):515–527 Marcellus L (2004) Are we missing anything? Pursuing research on attrition. Can J Nurs Res Gaunt DM, Papastavrou Brooks C, Pedder H, Crawley E, Horwood J, Metcalfe C (2023) Participant retention in paediatric randomised controlled trials published in six major journals 2015–2019: systematic review and meta-analysis. Trials 24(1):403 Hentzen S, Meirson T, Koehn K, Goodman AM, Chakraborty R, Prasad V et al (2022) Attrition and withdrawal amongst patients with multiple myeloma participating in randomized clinical trials: A systematic review. Blood 140(Supplement 1):7950–7951 Chaudhari N, Ravi R, Gogtay NJ, Thatte UM (2020) Recruitment and retention of the participants in clinical trials: challenges and solutions. Perspect Clin Res 11(2):64–69 Mohseni S, Tabatabaei-Malazy O, Peimani M, Ejtahed H, Khodaeian M, Nazeri E et al (2021) Withdrawal reasons of randomized controlled trials on type 2 diabetes: A systematic review. DARU J Pharm Sci 29:1–12. 10.1007/s40199-020-00380-7 Mistry A, Odwar B, Olewe F, Kurtis J, Moormann AM, Ong’echa JM (2022) Pedi- atric Participant Retention Rates in a Longitudinal Malaria Immunology Study. Am J Trop Med Hyg 106(6):1629–1636. 10.4269/ajtmh.21-1052 Poongothai S, Anjana RM, Aarthy R, Unnikrishnan R, Narayan KV, Ali MK et al (2023) Strategies for participant retention in long term clinical trials: A participant– centric approaches. Perspect Clin Res 14(1):3–9 Chan, Li SX, Chau SWH, Chan NY, Zhang J, Wing YK (2022) Prediction of Dropout in a Randomized Controlled Trial of Adjunctive Light Treatment in Patients with Non-Seasonal Depression and Evening Chronotype. Clocks Sleep 4(3):346–357 Jacobsen E, Ran X, Liu A, Chang CCH, Ganguli M (2021) Predictors of attrition in a longitudinal population-based study of aging. Int Psychogeriatr 33(8):767–778 Doring N, Hwang YI, Akpanekpo E, Gullotta M, Ton B, Knight L et al (2023) Predicting attrition of men with a history of violence from randomised clinical trials. Trials 24(1):740 Giesemann J, Delgadillo J, Schwartz B, Bennemann B, Lutz W (2023) Predicting dropout from psychological treatment using different machine learning algorithms, resampling methods, and sample sizes. Psychother Res 33(6):683–695 Teixeira R, Rodrigues C, Moreira C, Barros H, Camacho R (2022) Machine learning methods to predict attrition in a population-based cohort of very preterm infants. Sci Rep 12(1):10587 Fayyaz H, Phan TLT, Bunnell HT, Beheshti R (2022) Predicting attrition patterns from pediatric weight management programs. Machine Learning for Health. PMLR, pp 326–342 Elkin ME, Zhu X (2021) Predictive modeling of clinical trial terminations using feature engineering and embedding learning. Sci Rep 11(1):3446 Amboka P, Kurui D, Wamukoya M, Sindi JK, Vicente-Crespo M (2024) A landscape analysis of clinical trials and infant clinical trials in Kenya, Ethiopia, and Nigeria. Front Epidemiol 4:1417419 Elreedy D, Atiya AF (2019) A comprehensive analysis of synthetic minority oversam- pling technique (SMOTE) for handling class imbalance. Inf Sci 505:32–64 Itam DH, Martin EC, Horsfall IT (2024) Enhanced convolutional neural network method- ology for solid waste classification utilizing data augmentation techniques. Waste Manage Bull 2(4):184–193 DeMauro SB, Bellamy SL, Fernando M, Hoffmann J, Gratton T, Schmidt B et al (2019) Patient, Family, and Center-Based Factors Associated with Attrition in Neonatal Clinical Research: A Prospective Study. Neonatology 115(4):328–334. 10.1159/000494105 Robinson L, Adair P, Coffey M, Harris R, Burnside G (2016) Identifying the participant characteristics that predict recruitment and retention of participants to randomised controlled trials involving children: a systematic review. Trials 17:1–17 Katz KS, El-Mohandes A, Johnson DM, Jarrett M, Rose A, Cober M (2001) Retention of Low Income Mothers in a Parenting Intervention Study. J Community Health 26(3):203–218. 10.1023/A:1010373113060 Watson SE, Smith P, Snowden J, Vaughn V, Cottrell L, Madden CA et al (2022) Facilitators and barriers to pediatric clinical trial recruitment and retention in rural and community settings: A scoping review of the literature. Clin Transl Sci 15(4):838–853 Saar-Tsechansky M, Provost F (2007) Handling missing values when applying classifica- tion models. J Mach Learn Res Dash CSK, Behera AK, Dehuri S, Ghosh A (2023) An outliers detection and elimination framework in classification task of data mining. Decis Analytics J 6:100164 Brownlee J (2020) Imbalanced classification with Python: better metrics, balance skewed classes, cost-sensitive learning. Machine Learning Mastery Kotsiantis SB, Kanellopoulos D, Pintelas PE (2006) Data preprocessing for supervised learning. Int J Comput Sci 1(2):111–117 Zhao Z, Kunar A, Birke R, Van der Scheer H, Chen LY (2024) Ctab-gan+: Enhancing tabular data synthesis. Front big Data 6:1296508 Brownlee J (2016) Machine learning mastery with Python: understand your data, create accurate models, and work projects end-to-end. Machine Learning Mastery Weiss GM (2013) Foundations of imbalanced learning. Imbalanced learning: Foundations, algorithms, and applications. :13–41 Menardi G, Torelli N (2014) Training and assessing classification rules with imbalanced data. Data Min Knowl Disc 28(1):92–122 Xu L, Skoularidou M, Cuesta-Infante A, Veeramachaneni K (2019) Modeling tabular data using conditional gan. Adv Neural Inf Process Syst. ;32 Engelmann J, Lessmann S (2021) Conditional Wasserstein GAN-based oversampling of tabular data for imbalanced learning. Expert Syst Appl 174:114582 Figueira A, Vaz B (2022) Survey on synthetic data generation, evaluation methods and GANs. Mathematics 10(15):2733 Zhang Y, Nampally S, Hutchison E, Weatherall J, Khan F, Shameer K (2020) Predictive Modeling of Personalized Clinical Trial Attrition using Time-to-event Approaches. In: Machine Learning for Pharma and Healthcare Applications ECML PKDD 2020 Workshop; PharML Additional Declarations The authors declare no competing interests. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8970473","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":597107366,"identity":"25145c54-25fc-4417-98d9-1f7db634b0e7","order_by":0,"name":"Mailosi Innussa","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAu0lEQVRIiWNgGAWjYDACZijNDyISCkjRItkA0mJAim0GB8AkESp123mPPeb5UydvfH514ocHBgzy/GIH8GsxO8yXbszbdthw2423myWADjOcOTuBkBYeM2nehgOM226c3QDSkmBwmxgtQIfZb55xdvMPErSwMSdu4O/dRrwtknPbDifPuMG7zSLBQIIIv5w/Yybx5k+dbX//2c03f1TYyPNLE9CCABJglRLEKgcB/gOkqB4Fo2AUjIKRBABdaEERPd3s/QAAAABJRU5ErkJggg==","orcid":"https://orcid.org/0009-0005-7776-0555","institution":"Malawi university of science and technology","correspondingAuthor":true,"prefix":"","firstName":"Mailosi","middleName":"","lastName":"Innussa","suffix":""},{"id":597107367,"identity":"4fdef3a5-50e8-4742-84a7-b23afa096c6b","order_by":1,"name":"Priscilla Maliwichi","email":"","orcid":"https://orcid.org/0000-0001-5878-5355","institution":"Malawi university of science and technology","correspondingAuthor":false,"prefix":"","firstName":"Priscilla","middleName":"","lastName":"Maliwichi","suffix":""},{"id":597107368,"identity":"465b687f-8acf-470e-8504-039345ac92b4","order_by":2,"name":"Clement Nyirenda","email":"","orcid":"https://orcid.org/0000-0002-4181-0478","institution":"University of Western cape","correspondingAuthor":false,"prefix":"","firstName":"Clement","middleName":"","lastName":"Nyirenda","suffix":""}],"badges":[],"createdAt":"2026-02-25 18:14:01","currentVersionCode":1,"declarations":{"humanSubjects":true,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":true,"humanSubjectConsent":true,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-8970473/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8970473/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":103602243,"identity":"8542331a-2367-4fd8-9de7-e79e8ba23454","added_by":"auto","created_at":"2026-02-27 14:12:26","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":163020,"visible":true,"origin":"","legend":"\u003cp\u003eCorrelation Heat-map.\u003c/p\u003e","description":"","filename":"1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8970473/v1/d89962ae960028ef1cd8638a.jpg"},{"id":103602245,"identity":"fa14252e-4bfc-42ed-a46b-3485e1761a69","added_by":"auto","created_at":"2026-02-27 14:12:26","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":77338,"visible":true,"origin":"","legend":"\u003cp\u003eModel performance\u003c/p\u003e","description":"","filename":"2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8970473/v1/3b10abfdd26de701de67b3cd.jpg"},{"id":103602244,"identity":"7383a1d9-29fc-4b06-be5e-439c665b3d5c","added_by":"auto","created_at":"2026-02-27 14:12:26","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":81942,"visible":true,"origin":"","legend":"\u003cp\u003eConfusion matrix: Best performing model in each data setting\u003c/p\u003e","description":"","filename":"3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8970473/v1/ddb11adfb0ab8ae8d464bba8.jpg"},{"id":105033539,"identity":"3f65372f-d9ea-42ef-82c2-1856c5baf979","added_by":"auto","created_at":"2026-03-20 07:19:36","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":795826,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8970473/v1/ea066345-43f3-4698-b823-b53739f4fc28.pdf"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"\u003cp\u003ePredicting participant attrition in paediatric clinical trials using machine learning and deep learning models\u003c/p\u003e","fulltext":[{"header":"1 Introduction","content":"\u003cp\u003eAttrition in clinical trials refers to patient dropout or withdrawal. It is one of the significant challenges faced by researchers in clinical trials [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. Attrition occurs when participants either voluntarily withdraw or are lost to follow-up by the clinical site and trial sponsor [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. A study has identified that an 88% dropout rate is due to factors such as loss to follow-up, non-compliance with the protocol, and withdrawal of consent [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. Other studies have found that the rates of attrition range from 3 to 63% [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. It is generally considered that bias becomes a significant concern when the attrition rate surpasses 20% [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. Attrition becomes more complicated in paediatric populations because trial involvement relies on both parental and child behaviours [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e].\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eAttrition can cause the trial to last longer than anticipated, which may lead to additional financial costs. Additionally, attrition has the potential to heavily bias data, reduce statistical power, and undermine the overall statistical validity of an intervention. It may also make future study requirements more difficult [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e, \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. High attrition rates can give the impression that the trial\u0026rsquo;s design and execution were poorly handled, which might lead to a loss of trust in clinical research [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. Prior research has identified multiple factors associated with attrition, including lack of trust, socioeconomic barriers,\u003c/p\u003e\u003cp\u003eparticipant inconvenience, real or perceived adverse events of intervention, migration from study sites, media influence, procedural complexity, and the volume of biological samples collected [\u003cspan additionalcitationids=\"CR9\" citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eTrialists employ various methods to reduce attrition, such as building strong relation- ships between participants and site staff, offering incentives, enhancing communication, showing respect for participants, and ensuring effective informed consent. However, the success of these strategies often depends on the specific context [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e, \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. Recently, machine learning and deep learning have been applied to predict attrition in clinical studies. Literature presents the potential of machine learning (ML) and deep learning (DL) in predicting attrition in clinical studies. However, significant gaps remain in the research, as much of the prior work has focused on adult populations [\u003cspan additionalcitationids=\"CR13\" citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e], general populations [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e], preterm infants [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e], or weight management interventions [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]. These models may not be directly applicable to paediatric clinical trials, where attrition pat- terns are influenced by unique factors related to trial procedures, child participants, and caregiver involvement. In addition, attrition prediction datasets often suffer from severe class imbalance and limited sample sizes, which pose significant challenges to predictive performance [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e, \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]. To address these gaps, this study develops and evaluates machine learning models for predicting attrition in paediatric clinical trials. Data from completed clinical trials are collected and preprocessed, with particular emphasis on handling class imbalance using the Synthetic Minority Over-sampling Technique (SMOTE) [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e] and investigating Generative Adversarial Network (GAN)-based data augmentation as a strategy for mitigating data scarcity [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e]. The study then compared the predictive performance of Logistic regression(LR), Random forest (RF), Multilayer Perceptron (MLP), and One convolutional neural network (1DCNN) models, and identified and ranked features contributing to attrition. This approach allows assessment of model performance while providing insights into the key factors influencing dropout.\u003c/p\u003e\u003cp\u003eThis study makes several contributions to clinical trial research:\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eProvides a comparative evaluation of LR, RF, MLP, and 1DCNN models in the context of paediatric attrition.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eDemonstrates the impact of data enhancement techniques, including SMOTE, GAN, and a combined approach, on predictive performance.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eIdentifies and ranks the most influential features associated with dropout, offering actionable insights for designing trials that minimize attrition and improve participant retention.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e"},{"header":"2 Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1 Data Collection and Preparation\u003c/h2\u003e \u003cp\u003eThe dataset comprised data from 2,168 participants aged six months enrolled in a paediatric clinical trial conducted in Malawi. Eighteen participants who exited the study due to death were excluded to ensure that the analysis focused specifically on attrition outcomes [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. fourteen predictor were selected based on their relevance and prior evidence reported in the literature [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e, \u003cspan additionalcitationids=\"CR23 CR24\" citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e]. The predictor features included child age, child sex, residential area, maternal age, household size, number of household members under five years of age, father\u0026rsquo;s education level, mother\u0026rsquo;s education level, father\u0026rsquo;s occupation, mother\u0026rsquo;s occupation, family income source, monthly income category, income\u0026ndash;expenditure deficit indicator, and adverse event count. The outcome variable was attrition.\u003c/p\u003e \u003cdiv id=\"Sec4\" class=\"Section3\"\u003e \u003ch2\u003e2.1.1 Data preprocessing\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eMissing values were identified and imputed using the K-Nearest Neighbors (KNN) imputation method to preserve dataset integrity. Outliers were detected through boxplot visualizations and addressed via Winsorizing to reduce their impact on the models [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e]. Additionally, 18 participants who exited the study due to death were excluded from the dataset to focus the analysis on attrition outcomes [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. Categorical variables were encoded using one-hot encoding, and all features were normalized to a common scale to ensure balanced model training [\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e, \u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e].\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e2.2 Exploratory Analysis and Data Enhancement\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eTo explore feature relationships, a correlation matrix was computed (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). The matrix uses color gradients from blue to red, where grey indicates no correlation, red indicates strong positive correlation, and blue indicates negative correlation. Notable correlations observed include positive correlations between \u0026rsquo;age mother\u0026rsquo; with \u0026rsquo;household number\u0026rsquo;, \u0026rsquo;edu mother\u0026rsquo; with \u0026rsquo;edu father\u0026rsquo;, and \u0026rsquo;household member5\u0026rsquo; with \u0026rsquo;household number\u0026rsquo;; negative correlations between \u0026rsquo;occup mother\u0026rsquo; with \u0026rsquo;income source\u0026rsquo; and \u0026rsquo;edu mother\u0026rsquo; with \u0026rsquo;age mother\u0026rsquo;; and correlation between parental occupations: \u0026rsquo;occup mother\u0026rsquo; and \u0026rsquo;occup father\u0026rsquo;. These relationships offer insights into underlying data patterns and inform feature selection. Feature importance was further assessed using SHAP values to quantify each feature\u0026rsquo;s contribution to prediction.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eClass imbalance was evident, with 1,750 participants completing the study and only 400 exiting early, representing an 18.6% attrition rate. Imbalanced classes can bias\u003c/p\u003e \u003cp\u003emodel performance toward the majority class. To mitigate this, Synthetic Minority Oversampling Technique (SMOTE) was applied to generate synthetic minority class samples, achieving balanced class distribution for improved classifier performance [20]. To further enhance generalization and reduce overfitting, a conditional tabular GAN was employed for synthetic data generation. The neural generator comprised fully connected layers with LeakyReLU activations and batch normalization, while the adversarial component was implemented using a LightGBM regressor (max depth\u0026thinsp;=\u0026thinsp;2, learning rate\u0026thinsp;=\u0026thinsp;0.02, n estimators\u0026thinsp;=\u0026thinsp;500). Training parameters included ten-fold minor- ity oversampling gen x times\u0026thinsp;=\u0026thinsp;10), outlier filtering (bot filter quantile\u0026thinsp;=\u0026thinsp;0.0001 and top filter quantile\u0026thinsp;=\u0026thinsp;0.999), and default library settings (batch size\u0026thinsp;\u003cem\u003e\u0026asymp;\u003c/em\u003e\u0026thinsp;500 with early stopping). This approach aimed to increase dataset diversity and model robustness [21, 30].\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e2.3 Machine Learning Model Development and Evaluation\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003e Four machine learning algorithms, Logistic Regression, Random Forest, MLP, and 1DCNN were employed to classify participant attrition. Logistic Regression was included as a widely used baseline in attrition research. Models were trained using a stratified train-test split (70% training, 30% testing), with 10% of the training data reserved for validation. Eight features were selected from the initial 14 candidates in Data Collection using mean absolute SHAP values from the best-performing Random Forest model. These features age child, age mother, edu mother, resident, ae count, household num, deficit income, and child sex were retained as they contributed the majority of predictive signal (mean absolute SHAP\u0026thinsp;\u003cem\u003e\u0026gt;\u003c/em\u003e\u0026thinsp;0.02), with domain knowledge confirming their clinical relevance. All models were trained using these selected features across four experimental conditions:\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eOriginal Dataset: Using the original preprocessed dataset without any balancing or augmentation\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eSMOTE-Enhanced: Applying SMOTE to balance classes by generating synthetic minority samples\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eGAN-Augmented: Using GAN-augmented data to increase dataset diversity\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eCombined Enhancement: Combining SMOTE with GAN augmentation for balanced and diverse data\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eEach model was trained and tested across ten independent runs under consistent conditions, with average performance metrics reported. Evaluation metrics included accuracy, precision, recall, F1-score, AUC-ROC, and confusion matrices [\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e]. The F1-score was used as the primary metric due to the dataset\u0026rsquo;s imbalance, as it provides a balanced measure of precision and recall, making it more reliable than accuracy or AUC for assessing the minority class. This metric is commonly applied in imbalanced classification tasks and related domains [\u003cspan additionalcitationids=\"CR32\" citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e]. Statistical significance of performance differences was assessed using paired t-tests for within-model comparisons and one-way ANOVA for comparisons across models and configurations.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e2.4 Ethical Consideration\u003c/h2\u003e \u003cp\u003eThis study received ethical approval from the Malawi University of Science and Technology Research Committee (MUSTREC) under reference number P.01/2025/391. All ethical standards for research without direct contact with human or animal subjects were followed.\u003c/p\u003e \u003c/div\u003e"},{"header":"3 Results","content":"\u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eOur comparative analysis of four machine learning models across four data settings reveals significant variation in peadiatric attrition prediction performance as shown in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. The average F1-Scores across all algorithms and data setting ranged from 0.4487 to 0.514. Random forest achieved the highest F1-score (0.514\u0026thinsp;\u003cem\u003e\u0026plusmn;\u003c/em\u003e\u0026thinsp;0.01) with SMOTE-enhanced data. Logistic regression (0.4932\u0026thinsp;\u003cem\u003e\u0026plusmn;\u003c/em\u003e\u0026thinsp;0.40) and 1DCNN (0.5093\u0026thinsp;\u003cem\u003e\u0026plusmn;\u003c/em\u003e\u0026thinsp;1.21) achieved their best score on original dataset, while MLP achieved its best performance (0.5053\u0026thinsp;\u003cem\u003e\u0026plusmn;\u003c/em\u003e\u0026thinsp;0.76) with SMOTE enhancement. SMOTE augmentation resulted in notable performance improvements for Random forest (+\u0026thinsp;0.0024) and MLP (+\u0026thinsp;0.0566) but led to minor declines for Logistic regression (-0.0292) and 1DCNN (-0.0013). GAN augmentation negatively affected Random forest (-0.0375), Logistic regression (-0.0252) and 1DCNN (-0.0319) and did not result in any significant improvement for MLP. Combined augmentation slightly improved the performance of MLP (+\u0026thinsp;0.0064) while adversely affecting the other models. Figure\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e presents the performance of all models across the different data settings.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eF1-Scores Across Data Settings\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eModel\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eOriginal\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSMOTE\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eGAN\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eCombined\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLogistic regression\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.4932\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.464\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.4791\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e0.4613\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRandom forest\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003e0.5116\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e0.514\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.4741\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.4551\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMLP\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.4487\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.5053\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.4487\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.4551\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e1DCNN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.5093\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.508\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e0.4832\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.4522\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eTo provide a more comprehensive overview, Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e summarises the best-performing model for each data setting. On the original dataset, the Random forest model achieved an accuracy of 57.05%, with a precision of 0.5498, recall of 0.5819, and an AUC\u0026ndash;ROC of 0.5848. Under the SMOTE-enhanced setting, Random forest again performed best, with marginal increases in precision (0.5537), recall (0.5883), and AUC\u0026ndash;ROC (0.5874), while overall accuracy remained unchanged. In contrast, performance declined under GAN- based augmentation, where the 1DCNN achieved an accuracy of 55.66% alongside lower precision, recall, and AUC\u0026ndash;ROC values. In the combined data setting, Logistic regression yielded the highest performance; however, accuracy (54.88%) and discrimination ability (AUC\u0026ndash;ROC of 0.4870) remained comparatively low.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eBest Model Performance by Data Setting\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"7\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSetting\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eModel\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eF1\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePrecision\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eRecall\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eAUC-ROC\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eAccuracy%\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOriginal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRF\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.5116\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.5498\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.5819\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.5848\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e57.05\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSMOTE\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRF\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.5140\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.5537\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.5883\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.5874\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e57.05\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGAN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1DCNN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.4832\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.5174\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.5145\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.5296\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e55.66\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCombined\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eLR\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.4613\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.4926\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.4882\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.4870\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e54.88\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eThe confusion matrix prediction performance evaluation results is presented in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e. Under the original data setting, the Random forest model correctly classified 56.38% of non-attrition cases (true negatives) while misclassifying 43.62% as attrition (false positives), and achieved a recall of 60.00% for attrition cases, with 40.00% false negatives. With SMOTE enhancement, Random forest improved attrition recall to 61.67% (true positives), at the slight expense of a reduced true negative rate of 56.00% and an increase in false positives to 44.00%. Logistic regression achieved a true negative rate of 58.48%, with 42.52% false positives, and correctly identified 39.17% of attrition cases. The 1DCNN model exhibited a true negative rate of 51.24% and a false positive rate of 48.33%, while attrition recall remained at 51.67%. Overall, these confusion matrices indicate that Random forest with SMOTE enhancement provided stronger detection of the minority attrition class (recall of 61.67%).\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eSHAP analysis on the best-performing model ranked features by their contribution to attrition predictions. The top influencers were residential area (mean SHAP value: 0.12), adverse event count (0.10), maternal age (0.09), child age (0.08), and maternal education level (0.07). Features like household size and income-expenditure deficit had moderate impact (0.05\u0026ndash;0.06), while child sex was least influential (0.03).\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e"},{"header":"4 Discussion","content":"\u003cp\u003eThe results of this study demonstrated the potential of machine learning and deep learning models for addressing participant attrition in paediatric clinical trials. Across all data settings, the Random forest model demonstrated comparatively stronger performance, particularly under SMOTE augmentation, where it achieved the highest macro-averaged F1-score of 0.514 and improved attrition recall of 61.67% (true positives), highlighting its robustness in handling class imbalance and detecting attrition cases. Although the improvement over the original dataset was marginal (+\u0026thinsp;0.0024 in F1-score), this finding suggests that SMOTE can partially mitigate bias toward the majority class and enhance recall for the minority attrition class without substantial loss of precision. This is consistent with findings reported by [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]. The MLP model exhibited a larger performance gain under SMOTE augmentation (+\u0026thinsp;0.0566), suggesting that its performance is more strongly influenced by class distribution than that of the other models. Comparatively, GAN augmentation and the combined approach yielded negative impacts on performance for most models, with declines in F1-scores ranging from \u0026minus;\u0026thinsp;0.0013 to -0.0375. The decline in performance with GAN augmentation is likely due to mode collapse, unstable training, and poor-quality synthetic samples on small, highly imbalanced tabular clinical data [\u003cspan additionalcitationids=\"CR35\" citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e]. Unlike SMOTE, which reliably interpolates real minority instances, GANs struggle with scarce examples, often adding noise rather than benefit [\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e].These results suggest that more complex data augmentation techniques may not be well suited to the underlying structure and size of paediatric clinical trial data.\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eThe SHAP-based feature importance analysis provides valuable insights into attrition drivers, emphasizing sociodemographic and trial-related factors such as residential area, adverse event count, maternal age, child age, and maternal education. These align with existing literature highlighting socioeconomic barriers, participant inconvenience and caregiver characteristics as key contributors to dropout in paediatric trials [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e, \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e, \u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e]. Notably, residential area high influence (mean SHAP: 0.12) may reflect access to health facility challenges in low-resource settings including relocation and transportation issues [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]. Conversely, the lower impact of household size, income, and occupations contrasts with some studies [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e], and suggesting that, in this context, individual caregiver attributes may play a more significant role than broader economic factors.\u003c/p\u003e\u003cp\u003eStrengths of this study include its focus on attrition prediction in a paediatric clinical trial from a low-income setting, addressing a notable gap in the existing literature, which is largely dominated by adult or general populations [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e, \u003cspan additionalcitationids=\"CR13\" citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e, \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e, \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]. The study demonstrates the substantial impact of data augmentation techniques on improving model performance. Additionally, it identifies and ranks the most influential factors contributing to attrition, providing actionable insights for this specific setting. Overall, the comparative evaluation of multiple models and augmentation strategies offers practical guidance for clinical trial researchers developing data-driven tools to enhance participant retention. Nonetheless, several limitations remain. The data, drawn from a low- income country, may not generalise to other contexts and no external validation was done. Furthermore, the highest F1-score achieved was only 0.514, with moderate recall (0.5883) and a high false negative rate (44%), limiting stand-alone clinical utility for\u003c/p\u003e\u003cp\u003eearly intervention. Unmeasured variables like trial complexity, relocation, or media influence known attrition factors [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e] may have contributed to the models\u0026rsquo; modest overall performance. These results have implications for clinical trial design, suggesting that integrating Random forest models with SMOTE could enable early identification of at-risk participants, allowing targeted interventions such as enhanced communication or incentives [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. Future research should validate these models on multi-site, international datasets to improve robustness. Expanding to additional machine and deep learning algorithms may identify models that are better aligned with the characteristics of peadiatric clinical trial datasets.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e"},{"header":"5 Conclusion","content":"\u003cp\u003eThe results from the four experiments indicate the Random forest model demonstrates clear potential in predicting participant attrition in peadiatric clinical trials. The application of SMOTE alone effectively enhanced model performance by improving recall of at-risk participants, whereas GAN augmentation and combined approaches negatively impacted results for most models. Key predictors of attrition were identified and ranked, with residential area, adverse event count, maternal age, child age, and maternal education emerging as the most influential factors. While these insights provide a meaningful foundation for targeted retention strategies, further research is required to improve prediction accuracy and generalisability across diverse clinical and demographic settings.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eScout. Understanding Clinical Trial Patient Attrition (2023) Accessed: 2024-10-25. Available from: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.clinicalleader.com/doc/ understanding-clinical-trial-patient-attrition-0001\u003c/span\u003e\u003cspan address=\"https://www.clinicalleader.com/doc/ understanding-clinical-trial-patient-attrition-0001\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHutchison E, Zhang Y, Nampally S, Neelufer IK, Malkov V, Weatherall J et al (2021) Modeling Clinical Trial Attrition Using Machine Intelligence: A driver analytics case study using 1,325 trials representing one million patients. medRxiv. :2021\u0026ndash;2011\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDesai M (2020) Recruitment and retention of participants in clinical studies: critical issues and challenges. Medknow\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFarris MS, Devoe DJ, Addington J (2020) Attrition rates in trials for adolescents and young adults at clinical high risk for psychosis: A systematic review and meta-analysis. Early Interv Psychiat 14(5):515\u0026ndash;527\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMarcellus L (2004) Are we missing anything? Pursuing research on attrition. Can J Nurs Res\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGaunt DM, Papastavrou Brooks C, Pedder H, Crawley E, Horwood J, Metcalfe C (2023) Participant retention in paediatric randomised controlled trials published in six major journals 2015\u0026ndash;2019: systematic review and meta-analysis. Trials 24(1):403\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHentzen S, Meirson T, Koehn K, Goodman AM, Chakraborty R, Prasad V et al (2022) Attrition and withdrawal amongst patients with multiple myeloma participating in randomized clinical trials: A systematic review. Blood 140(Supplement 1):7950\u0026ndash;7951\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChaudhari N, Ravi R, Gogtay NJ, Thatte UM (2020) Recruitment and retention of the participants in clinical trials: challenges and solutions. Perspect Clin Res 11(2):64\u0026ndash;69\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMohseni S, Tabatabaei-Malazy O, Peimani M, Ejtahed H, Khodaeian M, Nazeri E et al (2021) Withdrawal reasons of randomized controlled trials on type 2 diabetes: A systematic review. DARU J Pharm Sci 29:1\u0026ndash;12. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1007/s40199-020-00380-7\u003c/span\u003e\u003cspan address=\"10.1007/s40199-020-00380-7\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMistry A, Odwar B, Olewe F, Kurtis J, Moormann AM, Ong\u0026rsquo;echa JM (2022) Pedi- atric Participant Retention Rates in a Longitudinal Malaria Immunology Study. Am J Trop Med Hyg 106(6):1629\u0026ndash;1636. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.4269/ajtmh.21-1052\u003c/span\u003e\u003cspan address=\"10.4269/ajtmh.21-1052\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePoongothai S, Anjana RM, Aarthy R, Unnikrishnan R, Narayan KV, Ali MK et al (2023) Strategies for participant retention in long term clinical trials: A participant\u0026ndash; centric approaches. Perspect Clin Res 14(1):3\u0026ndash;9\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChan, Li SX, Chau SWH, Chan NY, Zhang J, Wing YK (2022) Prediction of Dropout in a Randomized Controlled Trial of Adjunctive Light Treatment in Patients with Non-Seasonal Depression and Evening Chronotype. Clocks Sleep 4(3):346\u0026ndash;357\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJacobsen E, Ran X, Liu A, Chang CCH, Ganguli M (2021) Predictors of attrition in a longitudinal population-based study of aging. Int Psychogeriatr 33(8):767\u0026ndash;778\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDoring N, Hwang YI, Akpanekpo E, Gullotta M, Ton B, Knight L et al (2023) Predicting attrition of men with a history of violence from randomised clinical trials. Trials 24(1):740\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGiesemann J, Delgadillo J, Schwartz B, Bennemann B, Lutz W (2023) Predicting dropout from psychological treatment using different machine learning algorithms, resampling methods, and sample sizes. Psychother Res 33(6):683\u0026ndash;695\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTeixeira R, Rodrigues C, Moreira C, Barros H, Camacho R (2022) Machine learning methods to predict attrition in a population-based cohort of very preterm infants. Sci Rep 12(1):10587\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFayyaz H, Phan TLT, Bunnell HT, Beheshti R (2022) Predicting attrition patterns from pediatric weight management programs. Machine Learning for Health. PMLR, pp 326\u0026ndash;342\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eElkin ME, Zhu X (2021) Predictive modeling of clinical trial terminations using feature engineering and embedding learning. Sci Rep 11(1):3446\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAmboka P, Kurui D, Wamukoya M, Sindi JK, Vicente-Crespo M (2024) A landscape analysis of clinical trials and infant clinical trials in Kenya, Ethiopia, and Nigeria. Front Epidemiol 4:1417419\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eElreedy D, Atiya AF (2019) A comprehensive analysis of synthetic minority oversam- pling technique (SMOTE) for handling class imbalance. Inf Sci 505:32\u0026ndash;64\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eItam DH, Martin EC, Horsfall IT (2024) Enhanced convolutional neural network method- ology for solid waste classification utilizing data augmentation techniques. Waste Manage Bull 2(4):184\u0026ndash;193\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDeMauro SB, Bellamy SL, Fernando M, Hoffmann J, Gratton T, Schmidt B et al (2019) Patient, Family, and Center-Based Factors Associated with Attrition in Neonatal Clinical Research: A Prospective Study. Neonatology 115(4):328\u0026ndash;334. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1159/000494105\u003c/span\u003e\u003cspan address=\"10.1159/000494105\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRobinson L, Adair P, Coffey M, Harris R, Burnside G (2016) Identifying the participant characteristics that predict recruitment and retention of participants to randomised controlled trials involving children: a systematic review. Trials 17:1\u0026ndash;17\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKatz KS, El-Mohandes A, Johnson DM, Jarrett M, Rose A, Cober M (2001) Retention of Low Income Mothers in a Parenting Intervention Study. J Community Health 26(3):203\u0026ndash;218. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1023/A:1010373113060\u003c/span\u003e\u003cspan address=\"10.1023/A:1010373113060\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWatson SE, Smith P, Snowden J, Vaughn V, Cottrell L, Madden CA et al (2022) Facilitators and barriers to pediatric clinical trial recruitment and retention in rural and community settings: A scoping review of the literature. Clin Transl Sci 15(4):838\u0026ndash;853\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSaar-Tsechansky M, Provost F (2007) Handling missing values when applying classifica- tion models. J Mach Learn Res\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDash CSK, Behera AK, Dehuri S, Ghosh A (2023) An outliers detection and elimination framework in classification task of data mining. Decis Analytics J 6:100164\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBrownlee J (2020) Imbalanced classification with Python: better metrics, balance skewed classes, cost-sensitive learning. Machine Learning Mastery\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKotsiantis SB, Kanellopoulos D, Pintelas PE (2006) Data preprocessing for supervised learning. Int J Comput Sci 1(2):111\u0026ndash;117\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhao Z, Kunar A, Birke R, Van der Scheer H, Chen LY (2024) Ctab-gan+: Enhancing tabular data synthesis. Front big Data 6:1296508\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBrownlee J (2016) Machine learning mastery with Python: understand your data, create accurate models, and work projects end-to-end. Machine Learning Mastery\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWeiss GM (2013) Foundations of imbalanced learning. Imbalanced learning: Foundations, algorithms, and applications. :13\u0026ndash;41\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMenardi G, Torelli N (2014) Training and assessing classification rules with imbalanced data. Data Min Knowl Disc 28(1):92\u0026ndash;122\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eXu L, Skoularidou M, Cuesta-Infante A, Veeramachaneni K (2019) Modeling tabular data using conditional gan. Adv Neural Inf Process Syst. ;32\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eEngelmann J, Lessmann S (2021) Conditional Wasserstein GAN-based oversampling of tabular data for imbalanced learning. Expert Syst Appl 174:114582\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFigueira A, Vaz B (2022) Survey on synthetic data generation, evaluation methods and GANs. Mathematics 10(15):2733\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang Y, Nampally S, Hutchison E, Weatherall J, Khan F, Shameer K (2020) Predictive Modeling of Personalized Clinical Trial Attrition using Time-to-event Approaches. In: Machine Learning for Pharma and Healthcare Applications ECML PKDD 2020 Workshop; PharML\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"Malawi University of Science and Technology","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":true,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Paediatric clinical trial, Attrition prediction, Machine learning, Deep learning, Data augmentation","lastPublishedDoi":"10.21203/rs.3.rs-8970473/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8970473/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eParticipant attrition in paediatric clinical trials remains a major challenge, leading to higher costs, longer trial durations, and compromised research outcomes. Existing retention and prediction methods have shown limited success. This study aimed to predict participant attrition in paediatric clinical trials using machine learning and deep learning models. Secondary data from a paediatric clinical trial in Malawi were used. Logistic regression, Random forest, Multi-layer perceptron, and One convolutional neural network models were trained and evaluated under four data settings: original, SMOTE- enhanced, GAN-augmented, and combined. Predictive performance was assessed using macro-averaged F1-score as the primary metric. Random forest achieved the highest F1- score of 0.514 on the SMOTE-enhanced dataset. Overall performance remained modest, and GAN augmentation did not consistently improve results. The findings suggest that data augmentation techniques provide limited benefits for attrition prediction in this context.\u003c/p\u003e","manuscriptTitle":"Predicting participant attrition in paediatric clinical trials using machine learning and deep learning models","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-02-27 14:12:21","doi":"10.21203/rs.3.rs-8970473/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"97e856ce-4aa3-4743-8065-b5434c7cd6f5","owner":[],"postedDate":"February 27th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":63541296,"name":"Artificial Intelligence and Machine Learning"}],"tags":[],"updatedAt":"2026-03-11T18:31:21+00:00","versionOfRecord":[],"versionCreatedAt":"2026-02-27 14:12:21","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8970473","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8970473","identity":"rs-8970473","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}