A new federated learning system teaches AI to spot early signs of neurodevelopmental risk in infant movement videos without moving patient data off-site.
Researchers built what they describe as the first federated learning framework for automated infant movement analysis, including General Movement Assessment, a clinical method that reads video-derived skeletal data for signs of neurodevelopmental disorders. They tested it on fidgety movement classification, a core part of that assessment, in a three-client setup where each site trains a local model and shares only model updates, not raw video or skeletal data. To gauge how confident each prediction is, the team added Monte Carlo Dropout, then built a new aggregation method called Uncertainty-Aware Federated Averaging that weighs each client's contribution by how uncertain its predictions are. In testing, the federated models beat versions trained independently at each site and came close to matching a model trained on centralized data.
Hospitals have wanted to pool infant movement data for years, but privacy rules and data-sharing agreements make that slow or impossible. The real contribution here is the uncertainty-weighting step, which could make federated results more trustworthy when some clinical sites have noisier or smaller datasets than others. This is a research proof-of-concept, not a deployable clinical tool.
The entire experiment ran on three simulated clients, not real multi-hospital data, so "first" claims like this one should be read as an opening move rather than a finished product, especially since the paper does not say how the method holds up with dozens of sites instead of three.