A new review paper takes stock of how large language models are being pointed at medical time series data, and the results are promising, not production-ready.
The paper synthesizes recent research on using LLMs to analyze electrocardiograms, electroencephalograms, photoplethysmography, and vital-sign recordings. It maps the model architectures, data resources, and processing pipelines researchers have tried, along with prompt-design strategies tuned for specific clinical scenarios. The authors then sort existing applications into buckets: diagnostic interpretation, automated report generation, longitudinal health monitoring, and synthetic physiological-signal generation. They also maintain a public GitHub list tracking new studies as they appear.
What's notable is what the review admits it can't yet deliver: a consistent way to evaluate these systems. Unlike LLMs applied to clinical text, where benchmarks are relatively mature, MedTS work is still comparing apples to EKGs - different datasets, different metrics, different definitions of success. That gap matters more here than in most LLM applications, because a hallucinated summary of a doctor's note is embarrassing; a hallucinated read of a cardiac rhythm is dangerous.
The paper does not report a working clinical deployment. It is a map of a fast-growing but scattered field, useful mainly to researchers deciding where to point their next experiment - not to anyone shopping for a bedside monitor.