Recruitment inefficiencies pose significant challenges for the clinical research industry, resulting in longer timelines and underpowered studies. Despite technological advancements in patient recruitment, there is still potential to optimise models by considering behavioural, social vulnerability, and time-sensitive factors. This research aims to enhance recruitment by combining behavioural theory with artificial intelligence (AI), machine learning (ML), and survival analysis.

Patient pre-screening responses were mapped to Health Belief Model (HBM) constructs using Natural Language Processing (NLP), specifically Named Entity Recognition (NER) embedded within large language models (LLMs). These behavioural features were incorporated into an XGBoost classifier to predict patient progression to the site phone screening stage. The model achieved robust performance (AUC = 0.7398; Accuracy = 67.04%), with SHAP value analysis highlighting perceived barriers and overall social vulnerability as key predictors. Constructs such as perceived severity and benefits showed more limited influence, suggesting these features may yield greater value when proactively embedded into the design of pre-screening strategies.

Survival analysis revealed a significant decline in contact probability over time, with the probability dropping below 50% by Day 11. These results highlight the need for timely patient follow-up and effective, behaviourally informed referral strategies. This research demonstrates that incorporating survival-derived probabilities into a risk-scoring framework can enhance real-world patient engagement.

This research introduces a novel, behaviourally grounded pre-screening framework that enhances the precision and adaptability of AI/ML recruitment models, thereby contributing to both the academic literature and the development of AI/ML in clinical trial recruitment.