Med-PaLM M Overview

Evaluation of Med-PaLM Multimodal AI

Towards Generalist Biomedical AI was published in arXiv, by researchers from Google Research and Google DeepMind, evaluating the clinical capabilities of Med-PaLM Multimodal (Med-PaLM M), a generalist biomedical artificial intelligence (AI) system.

The practice of medicine involves diverse data modalities, including text records, imaging, genomics, etc. The potential of effectively utilizing this multimodal data, has prompted researchers to develop MultiMedBench, a new biomedical benchmark for developing AI systems.

Med-PaLM is evaluated for generalist biomedical AI systems. This large multimodal generative model can encode and interpret biomedical data, including clinical language, imaging, and genomics.

This paper claims that Med-PaLM M demonstrates performance that is competitive with, or exceeds, the current standards in MultiMedBench tasks.

The critical step now, despite the encouraging results, is the validation of Med-PaLM M in clinical practice. For example, An evaluation of model-generated and human chest X-ray reports was conducted in a retrospective study. Of 246 chest X-rays, clinicians expressed a preference for Med-PaLM M reports over those produced by radiologists in up to 40.50% of cases. Results such as this suggest a promising clinical utility for the model, however, extensive validation and development are required for effective deployment in a clinical setting.

I’ll be including an examination of Med-PaLM and similar technologies in my work-progress book, “A Clinician’s Guide to AI,” exclusively for my subscribers.