EnECG: one ECG, five tasks, and a mixture of experts

Instead of training one huge ECG model, EnECG freezes five existing models, adds small LoRA-tuned layers, and lets a gate decide which model to trust for each ECG.

• ECG
• Foundation models
• Mixture of experts
• LoRA

Video resources

Related ECG learning videos, embedded for playback on this page.

Why I picked this

Most ECG papers solve one task. This one asks a more practical question: can one system handle several tasks without retraining giant models every time?

The problem

Two things get in the way:

• No single model extracts every feature needed for every ECG task.
• Full fine-tuning of big foundation models costs more GPU than most hospitals have.
• Fixed ensembles (plain averaging or voting) use the same weights for every patient.

How it works

1. Data. MIMIC-IV-ECG: 800,035 ECGs from 161,352 patients, 12 leads, 10 s at 500 Hz. Each run samples 10,000 patients, split 70/20/10, repeated three times.
2. Five tasks. RR interval and age (regression), sex and potassium abnormality (binary, only 3% abnormal), and 15-label arrhythmia detection.
3. Frozen experts. Each base model is frozen, gets a feed-forward output layer, and that layer is tuned with LoRA.
4. Gate. A Mixture of Experts looks at a subset of leads and outputs weights. The final prediction is the weighted sum of the expert logits.

Results

From the paper's Table 2. The RR gain is about 38% over TEMPO.

My reading notes

Biggest lesson: different models are good at different ECG tasks, so a learned gate beats a fixed average.

What's next

Only 10,000 of 161,352 patients were used per run. The authors plan adaptive ensembles and dynamic expert selection. I would test the full dataset.

Paper: Xu et al., EnECG, arXiv 2511.22935.

Research interpretation

EnECG treats ECG analysis as a family of related tasks rather than one label-prediction problem. Specialist models can capture different signal characteristics; a gating network learns how to weight their outputs for a given task or input. Low-Rank Adaptation (LoRA) is used as a parameter-efficient adaptation strategy, limiting how much of each model needs to be updated.

Why this matters

• Multi-task learning: related tasks may share useful representations, but labels and clinical endpoints differ.
• Mixture of experts: a gate can route information to specialist experts, although routing errors can propagate.
• Parameter efficiency: fewer trainable parameters can reduce adaptation cost; this does not automatically mean lower end-to-end inference cost.
• Fair comparison: compare against single models and simple ensembles using the same data splits and compute budget.

Questions for a careful reading

Check the datasets and task definitions, external validation, per-task metrics, expert ablations, calibration, computational cost, and whether patient-level splits prevent leakage. The preprint's results are research evidence, not proof of clinical readiness.

Primary sources

• Xu et al., EnECG: Efficient Ensemble Learning for ECG Multi-task Foundation Model (arXiv)
• Authors' code repository
• PTB-XL dataset on PhysioNet