An AI-enabled framework for pre-visit decision support using patient similarity analysis in melanoma.

Abstract

e21533 Background: Clinical decision-making in melanoma is challenged by heterogeneous disease biology, evolving treatment paradigms, and fragmented longitudinal data. Although EMRs contain real-world information, clinicians often lack tools that organize these data around clinically meaningful milestones to support pre-visit decision-making. We developed CODE-M, an AI–enabled framework designed to identify clinically similar patients (“patients like mine”) and surface outcome-relevant insights using real-world data. Methods: CODE-M was developed using data derived from EMRs at a large academic cancer center. Patient data were normalized and aligned to a clinically defined ordinal timeline spanning diagnosis, adjuvant therapy, recurrence, progression, systemic treatment, and outcomes. Similarity modeling incorporated demographics, disease stage, treatment history, longitudinal outcomes, and available molecular and genomic attributes, represented as contextual features within the timeline. AI-based similarity modeling generated retrospective cohorts reflecting shared clinical features of an index patient and supported iterative cohort generation for competing progression-free and overall survival analyses using automated Kaplan–Meier, Cox proportional hazards, and feature association models. A clinician-facing pre-visit summary presents cohort characteristics, outcomes, and relevant clinical trial options prior to patient encounters. Results: The framework reliably aligned heterogeneous longitudinal EMR data to a shared clinical timeline and generated reproducible, clinically coherent patient cohorts reflecting shared disease stage, treatment exposure, and melanoma care patterns. Iterative cohort generation enabled comparative analyses across competing cohorts, resulting in predictable shifts in observed progression-free and overall survival distributions. Pre-visit summaries and analytic outputs were reviewed by clinicians and demonstrated interpretability and relevance for visit preparation. The platform also identified clinical trials based on eligibility criteria, trial availability, and patient geographic proximity, integrating trial matching into the clinical decision workflow. Conclusions: An AI-enabled, ordinal timeline–based approach can structure real-world melanoma data to support pre-visit clinical context through patient similarity analysis and comparative outcome visualization. CODE-M summarizes observed outcomes among clinically similar patients and is intended to support contextual understanding of prior care experiences rather than generate treatment recommendations or patient-specific predictions. This framework supports clinical insight, hypothesis generation, and shared decision-making using longitudinal EMR-derived data, with applicability across oncology disease types.