Foundation model extracts disease signals from 24-hour wrist movement data

The critical detail the summary glosses over: Sensori learns representations without any labeled health outcomes during training. It discovers disease markers and behavioral patterns purely from the structure of movement itself, then validates those learned features against held-out clinical labels. This inversion (representation first, validation second) is what enables scaling across unlabeled person-days.

This follows the same pattern as BEACON and the body-composition GNN work from late August. Both those papers showed how foundation models can extract human-relevant signals from raw sensor streams by fusing multiple data modalities or using specialized architectures to propagate information across structured graphs. Sensori applies that logic to temporal wearable data: instead of hand-crafted activity summaries, learn dense embeddings that capture what matters. The multi-cohort validation across geographies also echoes the infrastructure thinking in the tactile intelligence and MLOps papers, signaling that reproducibility across deployment contexts is becoming table stakes for sensor-based health models.

If Sensori's learned representations transfer to a fourth cohort (geographically or demographically distinct from the three used in training) without retraining, that confirms the model captured generalizable movement patterns rather than dataset artifacts. Conversely, if performance drops significantly on out-of-distribution populations, the claimed pathway to real-world deployment becomes much narrower.