MACHINE LEARNING · ECG SIMILARITY · TEAM RESEARCH
ECG patient matching across clinical and wearable devices
Can two electrocardiograms belong to the same unseen patient, even when taken at different times or with different devices? Our team investigated that question through contrastive learning, CNN/GRU representations and rigorous matching experiments.
ECG waveforms are affected by more than identity. A different patient's ECG may look familiar, while the same person's ECG may change across recordings.
Computer-generated traces, not patient data.
Our task was similarity scoring, not medical diagnosis. We used patient-level contrastive learning to teach models to rank same-person pairs above different-person pairs.
Prepare a standardized Lead I signal: clip length, missing-value handling, 250 Hz sampling, 0.5–40 Hz bandpass and normalization.
Process ECG A and ECG B through the same encoder weights to produce normalized 128-dimensional representations.
Compute cosine similarity for the two embeddings. Higher scores should correspond to more likely same-patient matches.
The final system did not stack a CNN and GRU into a single network. Four independently trained encoders each evaluated the same pair; their four cosine scores were averaged with equal weights.
Different model families can contribute complementary errors and representations, which motivated the score-level ensemble.
One cosine score from each model
One K2 CNN member starts with a 1D convolutional stem, then passes through residual convolution stages (32 → 64 → 128 → 256 channels). Attention pooling summarizes the time dimension; a projection head creates a unit-length 128-dimensional embedding. The Raw GRU member instead reads short waveform patches as a sequence before attention pooling.
Some of the hardest clinical-to-clinical examples were not simply cases separated by many months. They were pairs where the waveform morphology had changed significantly.
Heart rate, rhythm, physiological state, electrode placement and recording conditions may influence the trace. We investigated cases where a later recording from the same individual looked less similar than a different person's recording.
Observed in development data; we did not establish the physiological causes.
We grouped clinical positive pairs by heartbeat-shape change to investigate where errors concentrated. The most-changed group was consistently harder than more stable pairs.
Illustrative ordering, not plotted values.
We went beyond swapping in another network: we changed how many recordings each patient contributed, added time-stretch augmentation, and combined models that made different mistakes.
The later CNN training recipe sampled up to four clinical ECG recordings from a patient rather than relying on only two, selecting a same-person pair each epoch. This exposed the model to more within-person variation.
The training recipe also used ±20% time-stretch augmentation. Separately trained CNN members were averaged, then complemented by a Raw GRU in K2. More recordings and time stretch were changed together; their effects cannot be isolated from that particular comparison.
Earlier experiments used different patients, pair lists and evaluation protocols, so their AUROCs are not comparable with each other. Putting them on one chart would imply a fair comparison that did not exist. This page reports only the official competition scores.
The organizers scored two settings on their own test data: matching clinical ECGs to each other, and matching clinical ECGs to Apple Watch recordings.
2026 Wake Forest ECG-AI Build-a-Thon
Scored by the organizers on their own test data; the two settings are reported separately.
Every split was patient-disjoint, so a model could not score well by memorizing individuals.
Our one-time sealed test covered clinical pairs only. We checked Apple Watch performance on development patients but never ran an equally independent wearable holdout, so the official 0.930 was our first truly independent measure of cross-device performance. Next time, cross-device validation gets its own frozen test before submission.
Model experimentation: developed and evaluated beat-level CNN representations and ran CNN vs. RNN design comparisons.
Controlled evaluation: supported patient-disjoint benchmarks, comparable cohorts and cross-team model selection.
Failure analysis: traced the hardest same-patient matches to changes in waveform shape over time.
Final validation: helped coordinate the independent clinical-only holdout used to confirm K2 before submission.
The important next step is independent cross-device validation and better robustness to longitudinal ECG morphology change, not simply adding more model complexity.
The full record includes more experiments, ablations and model analysis than this page shows. Happy to walk through what can be shared.
Contact me about the research →A collaborative team project. Scores shown are the official competition results. ECG traces on this page are synthetic; no patient recordings, model weights or raw data are published.