ISENet: a Deep Learning Model for Detecting Ischemic ST Changes in Long-Term ECG Monitoring

Comprehensive Summary

Lin et al. developed ISENet, a one-dimensional residual neural network designed to detect ischemic ST change events (ISE) from long-term ambulatory ECG recordings in the PhysioNet Long-Term ST Database (86 24-hour recordings). Rather than relying on handcrafted ST-segment features or predefined reference points, the model analyzed 10-second ECG segments directly after basic filtering and normalization. Using patient-independent tenfold cross-validation, ISENet exhibited 83.5% accuracy and an AUC of 0.89 in the primary task of distinguishing ISE from non-ISE events (including heart rate-related, axis shift, and conduction changes), outperforming VGG19 and ResNet50 models while using fewer parameters (0.3M vs. 25.6M and 2.6M respectively). In a mixed event setting (ISE + HRE vs. non-ISE + HRE), performance improved further (accuracy: 85.3%, AUC: 0.91). Separating ISE from morphologically similar HRE events remained more challenging, reflecting real-world waveform overlap.

Outcomes and Implications

Detecting transient ischemia in long-term ECG recordings is valuable but technically difficult since many non-ischemic ST changes mimic ischemia. ISENet shows that a low-complexity 1D model can achieve strong sensitivity while remaining computationally efficient enough for wearable or real-time applications. Its performance suggests potential use as a front-end screening tool to flag suspicious segments for clinician review rather than replace expert interpretation. Nonetheless, validation was limited to a single public dataset, and performance variability in certain subgroup comparisons highlights the need for external multicenter testing before clinical adoption.