CogAdapt: Transferring Clinical ECG Foundation Models to Wearable Cognitive Load Assessment via Lead Adaptation

Researchers introduce CogAdapt, a framework that adapts clinical ECG foundation models to wearable cognitive load assessment by bridging the gap between hospital-grade 12-lead sensors and 3-lead wearable devices. The approach achieves strong cross-subject generalization on benchmark datasets, demonstrating the feasibility of transferring pre-trained medical models to consumer health applications.

CogAdapt addresses a fundamental challenge in healthcare AI: leveraging expensive clinical training data for practical wearable applications. Foundation models trained on millions of clinical ECG recordings contain rich physiological patterns, but these models assume specific sensor configurations unavailable in consumer devices. The researchers solve this mismatch through LeadBridge, a learnable adapter that mathematically reconstructs missing lead channels from limited wearable input, and ProFine, a training strategy that selectively unfreezes encoder layers to preserve learned knowledge while adapting to the cognitive load task.

The work builds on two converging trends: the success of foundation models across domains (vision, language, now medical) and the growing adoption of consumer wearables for health monitoring. Clinical ECG datasets contain millions of labeled examples, while cognitive load datasets remain scarce—a data asymmetry that foundation model adaptation directly exploits. By avoiding training from scratch, CogAdapt reduces labeled data requirements and improves generalization across subjects, a critical requirement for consumer deployment.

The reported performance metrics—macro-F1 scores of 0.626 and 0.768 on CLARE and CL-Drive datasets respectively under leave-one-subject-out validation—demonstrate meaningful improvement over baselines. This has implications for adaptive human-computer interfaces, driver monitoring systems, and personalized health applications. Companies developing wearable health platforms could adopt similar transfer learning approaches to improve model robustness with limited proprietary data.

Future research should explore whether this framework generalizes to other physiological signals and wearable form factors, and whether clinical foundation models pre-trained on diverse conditions transfer better than those specialized to single tasks.

• →CogAdapt bridges clinical and wearable ECG systems through LeadBridge, a learnable adapter that transforms 3-lead signals into 12-lead anatomically consistent representations.
• →Progressive fine-tuning prevents catastrophic forgetting while adapting foundation models from 1M+ clinical ECGs to cognitive load assessment with limited labeled data.
• →Cross-subject validation shows macro-F1 improvements to 0.768 over baselines, addressing a critical generalization challenge in wearable health AI.
• →The framework demonstrates practical transfer learning for consumer health devices, reducing dependency on large proprietary wearable datasets.
• →Results support broader trend of foundation models enabling efficient adaptation across healthcare domains and sensor modalities.