Background: Patient non-adherence and drop-out increase the time and cost of clinical trials. A tool that predicts, at baseline, which patients are at risk of dropping out could enhance trial management through personalized engagement strategies. Understanding patient profiles and behaviors is essential for this. Our goal was to develop a machine learning-based model to predict drop-out likelihood in early-stage Type 1 Diabetes (T1D) clinical trials, through an evaluation of patient profile and behaviors.

Methods: Data from the IMPACT study and an associated substudy with 85 and 24 patients suffering from T1D were analyzed. The aim was to predict drop-out due to informed consent withdrawal or non-adherence, both often linked to lack of engagement. We modeled drop-out as a survival endpoint (time to drop-out) using a multivariate Cox model validated through Monte-Carlo cross validation. Baseline predictors of patient engagement, such as study site perception, belief in medicine, and health literacy, were collected via the Compl-AI questionnaire.       

Results: The model accurately identified early drop-outs, with a C-index of 0.82. The AUC of the ROC curve (0.80) further validated its strong ability to discriminate non-completers. For example, a sensitivity (True Positive Rate) of 93% can be achieved while maintaining a specificity (True Negative Rate) of 65%, making it an effective tool for distinguishing patients likely to drop out due to low engagement.

Conclusion: Informed consent withdrawal and non-adherence can be predicted at baseline, enabling early identification of patients at higher risk of drop-out. These patients could benefit from targeted interventions, improving retention rates and overall trial efficiency.