MediLoRA: Clinical-risk-guided low-rank adaptation for parameter-efficient medical question answering

Résumé fourni par la source

Medical question-answering systems are commonly adapted to maximize aggregate accuracy even though the consequences of incorrect answers differ across clinical domains. MediLoRA is a physician-informed parameter-efficient fine-tuning framework in which domain criticality conditions a coupled adaptation policy comprising effective rank, target-module coverage, loss weighting, curriculum order, and tier-dependent dropout. Eight physicians rated ten specialties for safety impact, reverse-coded error reversibility, and knowledge complexity. The framework was evaluated with LLaMA-3-8B on 24,000 multiple-choice items, with a 2400-item held-out split containing 2000 MedMCQA questions and 400 clinician-validated synthetic stress-test cases. In the cross-model held-out comparison, MediLoRA achieved 84.5% exact-match accuracy (Wilson 95% CI, 83.0–85.9), compared with 77.5% for uniform LoRA, 64.2% for the frozen backbone, and 85.3% for a full fine-tuning reference. Five independent MediLoRA runs yielded 83.8% mean accuracy (SD, 1.2). For the three critical domains with available domain-level results, gains over uniform LoRA were +18.1 percentage points in pharmacology, +14.3 in emergency medicine, and +9.3 in cardiology. No predefined high-risk benchmark errors occurred for MediLoRA in the 400-case audit; the one-sided exact upper 95% bound was 0.75%. The maximal MediLoRA adapter configuration contained approximately 65.0 million trainable parameters (0.80% of the backbone), versus 6.8 million (0.08%) for uniform LoRA. Because the comparison was not parameter matched and multiple risk-conditioned components changed jointly, the results support the feasibility of the coupled policy rather than an isolated causal effect of rank allocation.