Evaluating the Multimodal Capabilities of Generative AI in Complex Clinical Diagnostics

In the rapidly evolving landscape of artificial intelligence (AI) in healthcare, the study explores the diagnostic capabilities of Generative Pre-trained Transformer 4 Vision (GPT-4V) in complex clinical scenarios involving both medical imaging and textual patient data. Conducted over a week in October 2023, the study employed 93 cases from the New England Journal of Medicines image challenges. These cases were categorized into four types based on the nature of the imaging data, ranging from radiological scans to pathological slides. GPT-4Vs diagnostic performance was evaluated using multimodal inputs (text and image), text-only, and image-only prompts. The results indicate that GPT-4Vs diagnostic accuracy was highest when provided with multimodal inputs, aligning with the confirmed diagnoses in 80.6% of cases. In contrast, text-only and image-only inputs yielded lower accuracies of 66.7% and 45.2%, respectively (after correcting for random guessing: multimodal: 70.5 %, text only: 54.3 %, image only: 29.3 %). No significant variation was observed in the models performance across different types of images or medical specialties. The study substantiates the utility of multimodal AI models like GPT-4V as potential aids in clinical diagnostics. However, the proprietary nature of the models training data and architecture warrants further investigation to uncover biases and limitations. Future research should aim to corroborate these findings with real-world clinical data while considering ethical and privacy concerns.