Consistency-Based Confidence in Multimodal Large Language Models on Radiology Cases: Comparison with Self-Report (Preprint)

Abstract

<sec> <title>BACKGROUND</title> Large language models (LLMs) require specialized methodologies to quantify model confidence for safe deployment in healthcare systems; however, there is a lack of established methods for confidence assessment. </sec> <sec> <title>OBJECTIVE</title> To evaluate output consistency as a confidence metrics for multimodal-LLMs interpreting radiology cases and compare with self-reported. </sec> <sec> <title>METHODS</title> From a total of 311 quizzes in the Korean Society of Ultrasound in Medicine digital platform, we selected 75 multiple-choice cases. Six multimodal-LLMs were evaluated, three reasoning-focused models (o1, Claude-3.7-Sonnet, Gemini-2.5-Pro) and three general models (GPT-4o, Claude-3.5-Sonnet, Gemini-1.5-Pro). Temperature was fixed at 1.0. Two confidence metrics were assessed: (i) self-reported by LLMs using prompts that elicited direct confidence percentages with answer, and (ii) consistency-based metrics derived from 20 repeated interpretations per case including relative entropy (R_H) calculated as 1 - H/log₂k (H = Shannon-entropy, k = number of repetitions) and majority vote proportion. Receiver Operating Characteristics (ROC) analysis for discrimination and Spearman correlation (r) between accuracy and each confidence metric was conducted. Additionally, model calibration was assessed using Expected Calibration Error (ECE). </sec> <sec> <title>RESULTS</title> Consistency-based metrics demonstrated significant correlation with diagnostic accuracy for Claude-3.7-Sonnet (percentage, r=0.314; R_H, r=0.310), Gemini-2.5-Pro (percentage, r=0.354; R_H, r=0.347), and GPT-4o (percentage, r=0.321; R_H, r=0.318). ROC analysis revealed that consistency-based metrics outperformed self-reported confidence in discriminative ability, with area under the curve values of 0.663 (95% CI: 0.545–0.768) for Claude-3.7-Sonnet, 0.694 (95% CI: 0.577–0.795) for Gemini-2.5-Pro, and 0.671 (95% CI: 0.533–0.775) for GPT-4o. For consistency-based metrics, Regular ECE (10-bin) ranged from 0.313–0.485, while optimal ECE ranged from 0.276–0.478 with varying bin configurations. </sec> <sec> <title>CONCLUSIONS</title> In multimodal-LLMs applied to radiology case, consistency-based metrics provide a more dependable indicator of diagnostic confidence than the self‑report. </sec>

Ähnliche Arbeiten