CogAdapt: Adapting Clinical ECG Foundation Models for Wearable Cognitive Load Assessment

Amir Mousavi; Erfan Nourbakhsh; Mohammad Sadegh Sirjani; Rocky Slavin; Mimi Xie; Leslie Neely; John Davis; John Quarles

CogAdapt: Adapting Clinical ECG Foundation Models for Wearable Cognitive Load Assessment

Amir Mousavi¹, Erfan Nourbakhsh¹, Mohammad Sadegh Sirjani¹, Rocky Slavin¹, Mimi Xie¹, Leslie Neely³, John Davis², John Quarles¹

¹Department of Computer Science, College of AI, Cyber and Computing, The University of Texas at San Antonio

²Department of Educational Psychology, College of Education and Human Development, The University of Texas at San Antonio

³Department of Neuroscience, Developmental and Regenerative Biology, College of Sciences, The University of Texas at San Antonio

IEEE-EMBS International Conference on Biomedical and Health Informatics (BHI 2026)

arXivCode

The Sensor Gap

Clinical ECG foundation models expect 12-lead hospital recordings. Wearable cognitive load datasets provide noisy 3-lead signals and a different task.

Clinical 12-lead ECG foundation model versus a wearable 3-lead cognitive load classifier — Figure 1. The challenge: pretrained foundation models expect 12-lead clinical ECG, but wearable cognitive load datasets provide only 3-lead recordings.

Abstract

Assessing cognitive load continuously and at low latency would help adaptive human-computer interaction, but it remains hard because labeled data are scarce and models generalize poorly across subjects. Recent ECG foundation models, pre-trained on millions of clinical diagnostic ECG recordings, yet they do not apply directly to wearable devices when the sensor configuration and the task both differ. We present CogAdapt, a framework that adapts a clinical ECG foundation model to wearable cognitive load assessment. CogAdapt has two parts. LeadBridge is a learnable adapter that maps 3-lead wearable signals to a 12-lead-compatible representation. ProFine is a progressive fine-tuning strategy that unfreezes encoder layers in stages while limiting representational drift in the pre-trained model. On two public datasets (CLARE and CL-Drive) under leave-one-subject-out cross-validation, CogAdapt reaches macro-F1 of 0.626 and 0.768, improving over from-scratch baselines by 11.2 and 16.1 percentage points. The results show that a clinical ECG pretraining can support subject-independent cognitive load assessment from wearable sensors.

CogAdapt Pipeline

Wearable preprocessing, LeadBridge 3→12 mapping, pretrained ECG-FM, and a binary load head.

ProFine Progressive Fine-Tuning

Three scenarios control how much of ECG-FM is updated: frozen, top layers, or full encoder with bucketed learning rates.

ProFine scenarios A, B, and C showing frozen versus trainable LeadBridge, ECG-FM, and classification head — Figure 3. Progressive fine-tuning scenarios. Scenario A freezes the encoder. Scenario B unfreezes top layers. Scenario C unfreezes all layers.

Main Results

Performance on CLARE and CL-Drive under K-fold and LOSO. Cells are mean ± std over folds. Best per column in bold.

Method	CLARE						CL-Drive
	K-Fold			LOSO			K-Fold			LOSO
	Acc ± std	F1 ± std	AUC	Acc ± std	F1 ± std	AUC	Acc ± std	F1 ± std	AUC	Acc ± std	F1 ± std	AUC
ECG-LightCNN	.774±.007	.726±.009	.832	.735±.136	.514±.044	.539	.814±.017	.808±.018	.887	.716±.114	.607±.093	.707
Transformer	.706±.012	.643±.011	.718	.640±.116	.496±.044	.506	.688±.014	.680±.013	.731	.617±.102	.541±.086	.604
HRV-RF	.756±.011	.688±.016	.795	.668±.167	.444±.065	.517	.768±.013	.758±.014	.835	.594±.137	.533±.123	.672
CogAdapt-A	.742±.012	.717±.010	.843	.681±.122	.527±.052	.598	.823±.018	.821±.018	.906	.665±.166	.578±.149	.705
CogAdapt-B	.762±.012	.736±.012	.857	.691±.127	.557±.085	.706	.833±.019	.831±.019	.919	.770±.073	.704±.099	.811
CogAdapt-C	.813±.033	.785±.034	.898	.736±.104	.626±.119	.799	.862±.025	.860±.025	.945	.831±.091	.768±.117	.889

Table I. Performance on CLARE and CL-Drive under K-fold and LOSO. Cells are mean ± std over folds (K-fold 10; LOSO 20 on CLARE, 21 on CL-Drive). AUC is AUROC.

LeadBridge Ablation

Frozen encoder (Scenario A). Only the 3→12 mapping and head are trained. LeadBridge wins on macro-F1 for both datasets.

3→12 Mapping	CLARE		CL-Drive
3→12 Mapping	F1	AUC	F1	AUC
Zero-padding	.380	.639	.519	.764
Random adapter	.409	.548	.516	.704
Dower transform	.407	.628	.508	.744
LeadBridge	.527	.598	.578	.705

Table II. LeadBridge ablation under the frozen ECG-FM encoder. Best F1 cells in bold.

PTB-XL Reconstruction

Held-out PTB-XL reconstruction for precordial leads V2–V6. LeadBridge leads on RMSE for V2–V5 and on correlation for most leads.

Lead	RMSE ↓ (µV)			Correlation ↑
Lead	Dower	Lin. Reg.	LeadBridge	Dower	Lin. Reg.	LeadBridge
V2	205.69	195.08	175.61	0.546	0.546	0.669
V3	290.39	194.52	189.77	0.569	0.570	0.613
V4	270.63	164.72	164.02	0.662	0.698	0.718
V5	215.06	143.45	140.76	0.679	0.694	0.717
V6	161.40	129.66	135.22	0.544	0.549	0.547

Table III. Held-out PTB-XL reconstruction for V2–V6. RMSE in µV.

Takeaways

Clinical ECG pretraining can transfer to wearable cognitive load when the sensor gap is closed with a learned lead adapter.
LeadBridge beats zero-padding, random adapters, and fixed Dower transforms on LOSO macro-F1 under a frozen encoder.
Progressive unfreezing (ProFine) improves monotonically from A → B → C; full adaptation yields the strongest subject-independent results.
Cross-dataset transfer still drops, so environment and label physiology remain open challenges.

BibTeX

@misc{mousavi2026cogadapt,
  author={Mousavi, Amir and Nourbakhsh, Erfan and Sirjani, Mohammad Sadegh and Slavin, Rocky and Xie, Mimi and Neely, Leslie and Davis, John and Quarles, John},
  title={CogAdapt: Adapting Clinical ECG Foundation Models for Wearable Cognitive Load Assessment},
  year={2026},
  eprint={2605.22774},
  archivePrefix={arXiv},
  primaryClass={cs.LG},
  url={https://arxiv.org/abs/2605.22774},
  note={Accepted at IEEE EMBS BHI 2026}
}