JMIR Publications

ABSTRACT

Background:

Rare genetic diseases often present with recognizable facial phenotypes. Whether multimodal large language models (MLLMs) can prioritize rare genetic disease diagnoses from facial images, and whether image enhancement or visual prompting can improve this process, remains unclear.

Objective:

To evaluate MLLM diagnostic accuracy for rare genetic diseases from facial images and assess whether super-resolution enhancement or red-circle visual prompting improves performance.

Methods:

This diagnostic accuracy study used 4,746 facial images from the GestaltMatcher Database, including 882 images in GMDB-Frequent from diseases with at least 6 images and 3,864 images in GMDB-Rare from diseases with fewer than 6 images. Each query sample was evaluated under four input conditions: original image, CodeFormer-enhanced super-resolution image, GPT-Image-2–generated red-circle visual-prompt image, and shuffled-patient red-circle control image. GPT-5.5 generated 30 ranked candidate diagnoses per input. Diagnostic performance was assessed using top-1, top-5, top-10, and top-30 accuracy. Red-circle prompt quality was rated by two physicians. Prediction-set overlap and disease-composition analyses were used to assess patient-specific image constraint and prediction bias.

Results:

Among 4746 images, physician evaluators rated 29.5% and 29.3% of red-circle annotations as good, with 80.6% exact agreement and a weighted Cohen κ of 0.712. In GMDB-Frequent, top-1 and top-30 accuracies were 4.20% and 29.25% for original images, 4.99% and 32.31% for super-resolution images, and 4.76% and 29.93% for red-circle visual-prompt images; pairwise differences were statistically nonsignificant. In GMDB-Rare, top-1 accuracy was below 0.5% and top-30 accuracy approximately 1% across the input conditions. Matched red-circle prompts outperformed shuffled controls. Prediction-overlap and disease-composition analyses showed weak patient-specific image constraint and tend to output common rare genetic diseases.

Conclusions:

GPT-5.5 showed limited ability to diagnose rare genetic diseases from facial images alone, with modest performance in GMDB-Frequent and very low performance in GMDB-Rare. Super-resolution enhancement and red-circle visual prompting did not consistently improve performance. Shuffled-patient controls and prediction-composition analyses showed outputs were weakly constrained by patient-specific facial features and tend to output common rare genetic diseases. Future systems should integrate facial images with structured clinical phenotypes, family history, and genetic evidence.