A research framework is adapting heart-monitoring AI, built for hospital diagnostics, to measure how hard your brain is working — using only a wrist sensor.
CogAdapt takes ECG foundation models pre-trained on millions of clinical recordings and makes them usable on wearable devices that capture far less signal. The gap is significant: clinical setups use 12-lead electrode configurations; most wearables capture 3 leads at best. CogAdapt's LeadBridge component learns to translate the sparse wearable signal into a 12-lead-compatible format. A second component, ProFine, unfreezes the underlying model's layers gradually rather than all at once, which keeps the pre-trained knowledge intact while the model learns the new task. Tested on two public datasets under leave-one-subject-out cross-validation — the hardest standard, since the model sees no data from the test subject during training — CogAdapt hit macro-F1 scores of 0.626 and 0.768, beating from-scratch baselines by 11.2 and 16.1 percentage points respectively.
The labeled-data bottleneck has long kept cognitive load detection a lab curiosity rather than a shipping product. By borrowing representations from clinical pretraining, CogAdapt sidesteps the need to collect massive subject-specific datasets — the usual prerequisite for models that generalize across people. If the approach holds outside controlled experiments, it could push real-time cognitive monitoring into adaptive interfaces, driver safety systems, or workplace tools without requiring hospitals to get involved.
Still, a macro-F1 of 0.626 on one dataset is a long way from deployable reliability, and the jump from two research datasets to the noise and motion artifacts of daily life is the part these papers rarely address.