Theoretical Frameworks
Complementary Feature Domain (CFD) Theory
CFD theory formalizes how time-, frequency-, and time-frequency ECG representations carry complementary rather than redundant diagnostic information. This theory reveals that these three domains capture different physiological patterns and that multimodal models perform best when they preserve and balance this complementarity.
Our work shows that fusion architectures that explicitly respect domain-specific information improve accuracy, robustness to noise, and data efficiency compared to simple feature stacking.
Physiologic Invariance Theory (PIT)
When should a model change its mind? PIT answers this fundamental question in trustworthy AI by defining robustness in terms of physiologic drift—the natural variability in rhythm, amplitude, and morphology that should not flip a model's prediction.
We develop metrics and training strategies that keep latent representations stable under benign perturbations while remaining sensitive to truly pathological changes. This framework ensures AI systems are robust in real clinical environments.
Multimodal ECG Systems
Hybrid deep learning systems that integrate time-series signals, frequency spectra, and time-frequency images from the same ECG. Using 1D-CNNs for temporal patterns, 2D-CNNs for spectral analysis, Transformers for long-range dependencies, and principled fusion strategies to combine insights from each modality.
These systems are designed to be deployable on resource-constrained devices and scalable to large clinical cohorts, making them practical for real-world healthcare environments.
Explainable AI Trustworthiness (EAT)
The EAT framework integrates saliency maps, stability analysis, and clinician-aligned explanations to rigorously evaluate whether model predictions are medically plausible. Rather than optimizing for black-box accuracy alone, EAT emphasizes interpretability as a first-class design objective.
We study how explanations change under noise and domain shifts, design visualization tools that highlight which parts of the ECG models truly rely on, and ensure clinicians can interrogate and refine model decisions in practice.