Bridging the gap from clinical to home ECG: quantifying and overcoming accuracy loss in AI-enabled single-lead ECG models

Figures

• 

  Fig. 1. Study design and dataset workflow for training, fine-tuning, and testing of Lead-I ECG AI models. a Development of the base Lead-I model: 246,874 patients at Tri-Service General Hospital-Neihu General Hospital were randomized at a 70%/30% ratio into a training set (172,827 patients, 474,130 ECGs) and an internal test set (74,047 patients, 202,062 ECGs). An independent external test set comprised 65,338 patients (219,231 ECGs) from the Tingzhou Branch Hospital. Lead-I segments extracted from the 12-lead ECGs were used to train and validate the base model. b Fine-tuning on Apple Watch data: 8790 patients with single-lead Apple Watch recordings at the Neihu site were randomized 70%/30% into a training set (6113 patients, 7044 ECGs) and a test set (2677 patients, 3399 ECGs). ECG recordings judged as inconclusive or poor quality were excluded, resulting in 6618 ECGs for training and 3217 ECGs for testing. c Fine-tuning on QOCA ECG102D data: 27,983 patients with single-lead QOCA ECG102D recordings at the Neihu site were randomized 70%/30% into a training set (19,553 patients, 24,977 ECGs) and a test set (8430 patients, 10,942 ECGs). After excluding low-quality ECG recordings, 21,927…

• 

  Fig. 2. ECG acquisition interfaces, waveform presentation, and heart-rate agreement. a Single-lead ECG recording on an Apple Watch worn on the wrist. b QOCA ECG102D single-lead device held between two fingers to record Lead-I ECG. c For two representative patients (Patient 1, left column; Patient 2, right column), example 10-s Lead-I traces from three devices: Apple Watch (top row), QOCA ECG102D (middle row), and Philips TC-series 12-lead system (bottom row). Gray lines are raw, unfiltered signals; blue lines represent signals after band-pass filtering and baseline correction. d Bland–Altman plots comparing heart rates (beats per minute) reported by each device’s vendor algorithm against the Philips 12-lead ECG reference. Left: Apple Watch vs. 12-lead; right: QOCA ECG102D vs. 12-lead. Solid horizontal lines show the mean difference (bias), dashed lines show the limits of agreement (±1.96 SD), and each green dot represents one patient’s paired heart-rate difference plotted against the reference ECG value. We used the Philips 12-lead ECG as the clinical reference and therefore plotted differences against the reference value (rather than the mean of two methods) to directly visualize…

• 

  Fig. 3. Classification performance of base and fine-tuned Lead-I ECG AI models. AUCs (95% CI) for eight clinical and demographic classification tasks using Lead-I AI-ECG models before and after device-specific fine-tuning. Green bars show the base model trained on Lead-I segments from 12-lead ECGs on each test set. Orange bars show the same model after fine-tuning on Apple Watch Lead-I ECGs (tested on the Apple Watch cohort), and purple bars show performance after fine-tuning on QOCA ECG102D Lead-I ECGs (tested on the QOCA ECG102D cohort). Error bars denote 95% confidence intervals. Symbols indicate statistical comparisons: ns, not significant; *p < 0.05; **p < 0.01; ***p < 0.001.

• 

  Fig. 4. Continuous-variable prediction performance of base and fine-tuned Lead-I ECG AI models. a Continuous-variable prediction performance: Pearson’s correlation coefficients (r) with 95% CIs for seven continuous clinical variables predicted by Lead-I AI-ECG models before and after device-specific fine-tuning. b Continuous-variable prediction performance: Mean absolute error (MAE) with 95% CIs for seven continuous clinical variables predicted by Lead-I AI-ECG models before and after device-specific fine-tuning. Green bars show the base model trained on Lead-I segments from 12-lead ECGs on each test set. Orange bars show the same model after fine-tuning on Apple Watch Lead-I ECGs (tested on the Apple Watch cohort), and purple bars show performance after fine-tuning on QOCA ECG102D Lead-I ECGs (tested on the QOCA ECG102D cohort). Error bars denote 95% confidence intervals. Symbols indicate statistical comparisons: ns, not significant; *p < 0.05; **p < 0.01; ***p < 0.001.

• 

  Fig. 5. Fine-tuning sample size and performance change in home-based ECG models. a The change in discrimination (Delta AUC) for eight binary clinical tasks. b The change in continuous prediction accuracy (Delta Pearson’s r) for seven continuous variables under the same fine-tuning sample-size scheme. Left panel is for the Apple Watch cohort, right panel for the QOCA ECG102D cohort. The gray line represents the overall learning curve, and colored lines denote individual tasks (see legend). The horizontal dotted line at zero indicates parity with the base 12-lead model, and open circles with vertical dotted lines mark the approximate sample sizes at which the mean ΔAUC/Δr crosses zero. The number of training sizes is based on the number of ECGs rather than the number of patients. Since the number of patients varies for each task under different training sizes, specific patient numbers are presented in Figures S9 and S10.