Knowledge, Innovation and Technology in Scientific Research Vol. 3

Background: Leave Against Medical Advice (LAMA), also termed Discharge Against Medical Advice (DAMA), is a clinically important event associated with interrupted treatment, avoidable care gaps, readmission and other adverse outcomes. Existing evidence indicates that LAMA is influenced by interacting clinical, socioeconomic, behavioural, communication and health-system factors. Although recent machine-learning studies demonstrate that LAMA/DAMA can be predicted from routinely collected electronic health record (EHR) and hospital data, prediction alone does not address the modifiable reasons that precipitate premature departure.

Objective: To develop and prospectively evaluate an explainable, patient-centred artificial intelligence (AI) early-warning and intervention system that links risk prediction to human-led preventive action and a structured continuity-of-care pathway.

Methods: We propose a prospective, pragmatic, mixed-methods, two-phase study. Phase I will develop and temporally evaluate candidate prediction models using routinely collected clinical, operational and patient-centred data. Candidate approaches will include regularised logistic regression, random forest and gradient-boosting models. Performance will be assessed using discrimination, calibration, decision-curve analysis, operational utility, temporal stability and subgroup fairness. Phase II will prospectively evaluate a human-in-the-loop intervention using a stepped-wedge or controlled before-and-after design, subject to institutional feasibility and ethical approval. Explainable model outputs will identify potentially modifiable contributors to elevated risk and trigger proportionate clinical, communication, social-support and care-coordination interventions. When a patient continues to choose LAMA, a structured Protect pathway will minimise avoidable harm through medication reconciliation, return precautions, follow-up planning and post-discharge contact.

Expected Results: The study is a research protocol and conceptual implementation framework; therefore, no empirical results are claimed. Expected outputs include a temporally evaluated prediction model, an actionable explanation layer, estimates of intervention effects on LAMA and continuity-of-care outcomes, subgroup fairness analyses and implementation measures.

Conclusion: The Predict–Prevent–Protect framework shifts the role of AI from retrospective risk labelling towards proactive, patient-centred risk reduction while preserving autonomy and clinician accountability. Prospective clinical evaluation is required to determine whether this closed-loop approach reduces avoidable LAMA and downstream harm without increasing inequity, alert burden or loss of trust.