Secure, Transactable, Agentic Health Data Sharing

Health data is most valuable where it is hardest to move. STARFISH treats that as a design problem rather than a legal one, building a framework in which hospitals and research institutions can run analyses together while each record stays where it was collected and the individual it describes keeps control over how it is used.

Starfish-FL is the federated learning and analysis component, open source under the Apache 2.0 license and published in ACM Transactions on Computing for Healthcare. Federated learning (FL) already lets institutions train models without pooling patient records, but orchestrating, monitoring, and interpreting a federated analysis has stayed out of reach for the biostatisticians and clinicians who need it most. Starfish-FL closes that gap with a multi-tier agent harness built around one safety invariant. The large language model is advisory and cannot independently trigger a consequential action. An Autonomous Experiment Agent runs end-to-end experiments through command line tools, an Embedded Router Agent handles adaptive aggregation, early stopping, and failure triage, and Controller Agent Hooks produce per-site summaries, outlier flags, and convergence signals for the people running each site.

Three hybrid safety policies named SafeStop, SafeAggregate, and SafeTriage gate every site exclusion, early stop, and recovery action behind deterministic preconditions, so a mistake by the language model cannot move the system into an unsafe state. Every agent feature is opt-in and degrades to standard federated learning when switched off. The platform ships 21 task implementations in both Python and R, from logistic regression and Cox proportional hazards to Kaplan-Meier curves, Poisson and negative binomial models, censored regression, and multiple imputation, each with built-in statistical diagnostics so researchers can stay in the language they already use.

The same approach has been applied to real trial data, reproducing the non-inferiority results of the AcT multicentre stroke trial through federated analysis without pooling patient records.