The miniECG: Enabling interpretable detection of amplitude and intraventricular conduction ECG-abnormalities with a novel ECG device

Background The miniECG, a smartphone-sized, multi-lead device, offers a simple and fast alternative to the 12-lead ECG. We aimed to demonstrate the potential of the miniECG to record and detect amplitude and intraventricular conduction ECG abnormalities via interpretable models. Methods A miniECG was captured for patients undergoing a conventional 12-lead ECG in the University Medical Centre Utrecht. MiniECGs of patients with normal ECGs (controls), amplitude abnormalities (low QRS voltage (Microvoltage) and left ventricular hypertrophy (LVH)), conduction abnormalities (left or right bundle branch blocks (LBBB, RBBB), left anterior fascicular block (LAFB), bifascicular block (BfB)) were selected. Standard ECG-features were used as input for decision trees (DTs) for binary (normal vs abnormal) or multiclass classification, employing 10-fold stratified cross-validation. Results 1717 patients were included. For binary classification of conduction abnormalities, the decision tree showed an AUROC of 0.97±0.01 (NPV 0.95±0.03, sensitivity 0.95±0.02). For multiclass classification, AUROCs were 0.95±0.02 (BfB), 0.85±0.03 (LAFB), 0.87±0.05 (LBBB), 0.94±0.04 (RBBB) and 0.93±0.03 (controls). For binary classification of amplitude abnormalities, the AUROC was 0.76±0.04 (NPV 0.83± 0.05, sensitivity 0.84±0.06). For multiclass classification, the AUROCs were 0.80 ± 0.06 (LVH), 0.78±0.05 (microvoltage) and 0.76±0.05 (controls). Conclusion This is the first study to detect ECG abnormalities in amplitude and intraventricular conduction, with a four precordial electrodes set-up. DTs based on miniECG features offer an interpretable detection method with a performance that was comparable to other less interpretable models. Before clinical implementation, further research is necessary to optimize DT-structures and analyze abnormalities beyond the current study. Author Summary Electrocardiograms (ECGs) are vital for detecting abnormal electrical heart signals, but traditional 12-lead ECGs are often limited to in-hospital settings due to complexity. Our study introduces the miniECG, a smartphone-size device with four electrodes to be placed on the chest, that records eight leads. We evaluated its ability to detect amplitude and conduction cardiac abnormalities using interpretable machine learning models. In a hospital-based sample of over 1,700 patients, we trained decision trees (DTs) on standard ECG features to classify abnormalities. The models demonstrated accurate diagnostic potential, particularly for conduction abnormalities, while maintaining transparency and clinical interpretability. Further optimization is needed, but this tool could enable broader ECG acquisition and easier detection of cardiac abnormalities.