Enhancing the Accuracy of Human Phenotype Ontology Identification: Comparative Evaluation of Multimodal Large Language Models.

Background: Identifying Human Phenotype Ontology (HPO) terms is crucial for diagnosing and managing rare diseases. However, clinicians, especially junior physicians, often face challenges due to the complexity of describing patient phenotypes accurately. Traditional manual search methods using HPO databases are time-consuming and prone to errors. Objective: The aim of the study is to investigate whether the use of multimodal large language models (MLLMs) can improve the accuracy of junior physicians in identifying HPO terms from patient images related to rare diseases. Methods: In total, 20 junior physicians from 10 specialties participated. Each physician evaluated 27 patient images sourced from publicly available literature, with phenotypes relevant to rare diseases listed in the Chinese Rare Disease Catalogue. The study was divided into 2 groups: the manual search group relied on the Chinese Human Phenotype Ontology website, while the MLLM-assisted group used an electronic questionnaire that included HPO terms preidentified by ChatGPT-4o as prompts, followed by a search using the Chinese Human Phenotype Ontology. The primary outcome was the accuracy of HPO identification, defined as the proportion of correctly identified HPO terms compared to a standard set determined by an expert panel. Additionally, the accuracy of outputs from ChatGPT-4o and 2 open-source MLLMs (Llama3.2:11b and Llama3.2:90b) was evaluated using the same criteria, with hallucinations for each model documented separately. Furthermore, participating physicians completed an additional electronic questionnaire regarding their rare disease background to identify factors affecting their ability to accurately describe patient images using standardized HPO terms. Results: A total of 270 descriptions were evaluated per group. The MLLM-assisted group achieved a significantly higher accuracy rate of 67.4% (182/270) compared to 20.4% (55/270) in the manual group (relative risk 3.31, 95% CI 2.58-4.25; P<.001). The MLLM-assisted group demonstrated consistent performance across departments, whereas the manual group exhibited greater variability. Among standalone MLLMs, ChatGPT-4o achieved an accuracy of 48% (13/27), while the open-source models Llama3.2:11b and Llama3.2:90b achieved 15% (4/27) and 18% (5/27), respectively. However, MLLMs exhibited a high hallucination rate, frequently generating HPO terms with incorrect IDs or entirely fabricated content. Specifically, ChatGPT-4o, Llama3.2:11b, and Llama3.2:90b generated incorrect IDs in 57.3% (67/117), 98% (62/63), and 82% (46/56) of cases, respectively, and fabricated terms in 34.2% (40/117), 41% (26/63), and 32% (18/56) of cases, respectively. Additionally, a survey on the rare disease knowledge of junior physicians suggests that participation in rare disease and genetic disease training may enhance the performance of some physicians. Conclusions: The integration of MLLMs into clinical workflows significantly enhances the accuracy of HPO identification by junior physicians, offering promising potential to improve the diagnosis of rare diseases and standardize phenotype descriptions in medical research. However, the notable hallucination rate observed in MLLMs underscores the necessity for further refinement and rigorous validation before widespread adoption in clinical practice.