Making LLM Predictions Interpretable: Fine-Tuning GPT-4o for Early Discontinuation of Cancer Medication

Medication discontinuation is a major challenge in oncology; predicting discontinuation that occurs before planned treatment completion enables earlier risk flagging and care-team awareness. We compared GPT-4o with traditional ML models to evaluate whether a general-purpose LLM can perform competitively on structured EHR prediction while also producing clinician-readable rationales that can be mapped to feature attributions, from 2,364 patient records at a large academic medical center. Among all evaluated approaches, fine-tuned GPT-4o achieved the highest F1 (0.867), exceeding the best traditional model (XGBoost, F1=0.825). SHAP (XGBoost) and a SHAP-like mimic attribution (GPT-4o) both prioritized age and BMI, enabling aligned interpretability comparisons across model types. These findings highlight the promise of LLMs for structured clinical prediction and enable direct interpretability comparisons with traditional ML.