Med2ECG: Medical-Guided BCG-to-ECG Reconstruction for Diverse Populations

Lin Chen, Yandao Huang, Chenggao Li, Jun Chen, Shuxin Zhong, Minghui Qiu, Chunzhen Guo, Yi Wang, Qian Zhang, Kaishun Wu

SenSys, 2026

PDF Link GitHub

Minimal Med2ECG concept illustration showing sleep BCG transformed into ECG — Med2ECG turns passive sleep BCG signals into clinically useful ECG waveforms through medical-guided reconstruction.

A medical-guided system for reconstructing clinically useful ECG waveforms from passive, contactless BCG signals across healthy and clinical populations.

Med2ECG reconstructs ECG waveforms from contactless BCG signals by aligning latent cardiac event structures rather than treating the task as direct waveform regression. Its multi-scale feature extractor, shared-personalized experts, and medical-informed losses preserve morphology and diagnostic intervals across subjects, postures, and clinical settings.

0.927Pearson correlation on the self-collected clinical dataset

12.82%amplitude error on the clinical dataset

0.916Pearson correlation on the public dataset

5-20 msPR/QRS/QT/RR interval estimation range reported across datasets

Framework

From contactless BCG to diagnostic ECG

Med2ECG combines a multi-scale feature extractor, shared-personalized experts, and medical-informed objectives so the reconstruction follows latent cardiac event structure rather than only pointwise waveform similarity.

Relationship between BCG and ECG cardiac events

Physiology

BCG and ECG are aligned through cardiac event structure

The page foregrounds the paper's core reframing: BCG I/J/K events and ECG P/QRS/T waves are treated as related physiological sequences, making cross-modal reconstruction less brittle across people and postures.

Reconstruction

Sample traces show the reconstructed ECG against ground truth

The sample result makes the task concrete: Med2ECG starts from a passive BCG waveform and recovers ECG morphology useful for downstream cardiac interval estimation.

Clinical Dataset

Clinical evaluation covers healthy and patient data

On the self-collected in-hospital dataset, Med2ECG reports 0.927 PCC and 12.82% amplitude error while preserving clinically relevant timing metrics.

Public Dataset

Public data validates generalization

The public-dataset results summarize morphology and timing metrics against prior baselines under subject-independent evaluation.

Ablation

Each design choice is stress-tested

The ablation study isolates the impact of the expert architecture and medical-informed losses on reconstruction quality.

Med2ECG results comparing healthy and clinical individuals

Patient Study

Patient data is separated from healthy-subject results

The patient comparison highlights how Med2ECG behaves across diverse cardiovascular conditions rather than only controlled healthy recordings.

Med2ECG sample result for tachycardia patient data

Clinical Example

Clinical rhythm examples remain visible

A tachycardia case is included so readers can quickly inspect how reconstructed ECG traces look on patient data before opening the full paper.

Med2ECG results across different sleeping postures

Robustness

Sleeping posture is evaluated explicitly

The posture analysis reflects the practical deployment target: passive overnight sensing where users naturally shift position.

Med2ECG performance across different heart rates