[April 20, 2026, Taichung] A cross-institutional research team consisting of Hsien-Ju Ko from the Precision Health Research Center at Asia University, Professor Wen-Shyong Yu from the Department of Electrical Engineering at Tatung University, Yu-Chen Wang from the Department of Medical Laboratory Science and Biotechnology at Asia University, and Asia University President and IEEE Fellow Distinguished Professor Jeffrey J.P. Tsai, has published an innovative AI architecture, "Rot-IIR-SSM," in the top biomedical engineering journal *IEEE Transactions on Biomedical Engineering* (published online April 20, 2026; DOI: 10.1109/TBME.2026.3685682).

Research Breakthrough: While maintaining diagnostic accuracy comparable to mainstream deep learning models, this research is the first to provide ECG AI with mathematically provable stability, spectrum-level interpretability, and ultra-low-power streaming inference capable of running continuously on wearable devices. This opens a new pathway for the practical application of precision health, transitioning from hospital-based care to home-based monitoring.

Three Core Technical Features: (1) Extreme Lightweighting: Upon deployment, the model is converted into a biquad filter recursion with a state memory of only approximately 25 KB, making it ideal for 24-hour real-time monitoring on smartwatches or patch-type ECG devices; (2) Medical Interpretability: Through innovative pole-domain parameterization, the model primarily interprets based on the QRS complex morphology in the 10–20 Hz range, aligning with clinical intuition; (3) Provable Stability: BIBO stability is guaranteed by design, which is critical for the safety of medical devices.

Research Validation: In 12-lead ECG myocardial infarction detection (PTB-XL five-fold cross-validation, PTB-DB external testing), Rot-IIR-SSM achieved the highest external F1 score (0.8306) without post-processing, demonstrating the best cross-dataset generalization. Its internal F1 (0.8360) and AUROC (0.9251) were comparable to high-efficiency CNNs. Additionally, it achieved an F1 score of 0.8939 in supplementary experiments for atrial fibrillation.

The Precision Health Research Center at Asia University will continue to integrate big data, AI, and biomedicine to drive health management from a generalized approach toward precision health.

Figure 1. Rot-IIR-SSM model architecture

Figure 2. Pole Domain: Learned poles are all within the unit circle (BIBO stable), with frequency focused on the QRS band

Figure 3. Interpretability: The model primarily focuses on the QRS complex morphology between 10–20 Hz

Figure 4. ECG and spectral performance of myocardial infarction