Warning: This is an author submission that is not peer-reviewed or edited. Preprints - unless they show as "accepted" - should not be relied on to guide clinical practice or health-related behavior and should not be reported in news media as established information.
Consistency-Based Confidence in Multimodal Large Language Models on Radiology Cases: Comparison with Self-Report
ABSTRACT
Background:
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.
Objective:
To evaluate output consistency as a confidence metrics for multimodal-LLMs interpreting radiology cases and compare with self-reported.
Methods:
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).
Results:
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.
Conclusions:
In multimodal-LLMs applied to radiology case, consistency-based metrics provide a more dependable indicator of diagnostic confidence than the self‑report.
Citation
Request queued. Please wait while the file is being generated. It may take some time.
Copyright
© The authors. All rights reserved. This is a privileged document currently under peer-review/community review (or an accepted/rejected manuscript). Authors have provided JMIR Publications with an exclusive license to publish this preprint on it's website for review and ahead-of-print citation purposes only. While the final peer-reviewed paper may be licensed under a cc-by license on publication, at this stage authors and publisher expressively prohibit redistribution of this draft paper other than for review purposes.