An Open-Source Analysis of Cardiomyopathy Using Machine Learning and Electrocardiograms

Figure 1. Two discharge notes from unique patients with corresponding ICD codes and final GPT-4.1 labels for their hospital visit. The patient on the (left) presents with clear non-ischemic DCM but received an incorrect HCM diagnosis code, which was later corrected by GPT-4.1. On the (right), the patient is indicated to have a history of HOCM and received an HOCM diagnosis. However, the echocardiogram findings from the hospital visit directly indicate no LVOT obstruction with no mention of Valsalva obstruction, so the patient should be labeled HNCM. This was correctly done by GPT-4.1, which gives full emphasis to the current echocardiogram.

Figure 2. Flow chart showing the study population.

Figure 3. Visual of all detected peaks, onsets, and offsets for a sample lead II in the study population using the “prominence” method. Heartbeats and ECG markers from the 10 s lead II are overlaid. Time (s) is relative to the location of R peaks. The example S-peak and R-offset markers are stacked at the same location. Note that there are no Q-onsets or S-offsets provided in the “prominence” method.

Figure 4. Kernel density estimate (KDE) plots of three ECG-derived features between obstructive and non-obstructive hypertrophic cardiomyopathy. One HOCM patient did not have a valid P-loop detected in the VCG feature extraction.

Figure 5. Kernel density estimate (KDE) plots of six ECG-derived features between non-ischemic DCM (DCM-NI), ischemic DCM (DCM-I), hypertrophic cardiomyopathy (HCM), and 500 electrically normal ECGs.

Figure 6. Representative vectorcardiograms for normal, HCM, DCM-I, and DCM-NI patients. The QRS loop is in blue, and the T-loop is in orange. The ventricular gradient (VG) vector is marked in green for the horizontal and sagittal projections; it was not featured for the frontal view, as the standard QRS and T axes from the frontal view were not significant. The length of the VG vector is the magnitude of the projection on the shown view. The HCM patient shows larger loops, with slightly larger VG azimuths in comparison to the normal patient. Both DCM patients show loops smaller in amplitude in all projections, accompanied by small amplitude VG vectors with far higher 3D elevation and higher azimuth angles within the horizontal and sagittal projections.

Figure 7. Pooled out-of-fold receiver operating characteristic (ROC) curves for each task and model type in the internal validation cohort. Each ROC curve plots sensitivity (true positive rate) against 1 −specificity (false positive rate). AUC is the area under the curve. The dashed diagonal line represents the performance of a random classifier (AUC = 0.5).