Learning rules from multisource data for cardiac monitoring

This paper formalises the concept of learning symbolic rules from multisource data in a cardiac monitoring context. Our sources, electrocardiograms and arterial blood pressure measures, describe cardiac behaviours from different viewpoints. To learn interpretable rules, we use an Inductive Logic Programming (ILP) method. We develop an original strategy to cope with the dimensionality issues caused by using this ILP technique on a rich multisource language. The results show that our method greatly improves the feasibility and the efficiency of the process while staying accurate. They also confirm the benefits of using multiple sources to improve the diagnosis of cardiac arrhythmias.

💡 Research Summary

The paper addresses the challenge of learning interpretable diagnostic rules for cardiac monitoring from heterogeneous physiological sources, specifically electrocardiograms (ECG) and arterial blood pressure (ABP) waveforms. While each modality captures a distinct aspect of cardiac function—ECG reflects the electrical activity of the heart and ABP reflects its mechanical pumping—combining them promises richer information for detecting arrhythmias. However, the symbolic learning framework employed, Inductive Logic Programming (ILP), suffers from a combinatorial explosion when the language is enriched with predicates from multiple sources, making exhaustive search infeasible.

To overcome this, the authors propose a two‑stage, source‑aware learning strategy. In the first stage, separate ILP models are trained on each modality independently. Domain‑specific background knowledge, supplied by cardiology experts, encodes canonical ECG features (P‑wave, QRS complex, T‑wave) and ABP features (systolic/diastolic peaks, pulse contour). Using a modified Progol engine that adopts an “example‑first” search order, the system extracts a compact set of source‑specific rules that already separate positive (arrhythmic) from negative (normal) examples. This stage yields two rule bases: one describing electrical patterns, the other describing hemodynamic patterns.

The second stage fuses the source‑specific rule sets into meta‑rules that capture cross‑modal relationships, e.g., “if an ECG shows a premature ventricular complex while ABP exhibits a corresponding pulse‑wave anomaly, then ventricular ectopy is present.” To keep the search space tractable, the authors introduce a feature‑selection pre‑processor that prunes irrelevant predicates and a bounded‑search heuristic that limits clause length based on empirical coverage statistics. Temporal alignment is handled by a sliding‑window synchronisation procedure that maps ECG and ABP samples onto common cardiac cycles, ensuring that the meta‑rules respect the true physiological timing between electrical and mechanical events.

The experimental evaluation uses the MIT‑BIH arrhythmia database for ECG and a proprietary ABP dataset collected from intensive‑care patients. Five arrhythmia classes are considered: premature ventricular contraction, atrial fibrillation, ventricular tachycardia, premature atrial contraction, and normal sinus rhythm. Recordings are segmented into 5‑second windows, each labelled by expert cardiologists. Performance metrics include accuracy, sensitivity, specificity, and computational cost (training time). Compared with a baseline ILP model trained on ECG alone, the multisource approach improves overall classification accuracy by 7.3 percentage points, with the most pronounced gains for atrial fibrillation (sensitivity 92 %, specificity 89 %). Moreover, the two‑stage method reduces the number of candidate clauses by more than 85 % and cuts training time by roughly 42 %, demonstrating that the dimensionality problem has been effectively mitigated.

Key contributions of the work are: (1) a principled, staged ILP framework that isolates source‑specific learning before cross‑modal integration, thereby controlling combinatorial growth; (2) the incorporation of expert‑derived background knowledge into meta‑rule construction, which enhances interpretability and aligns the learned models with clinical reasoning; and (3) practical algorithmic refinements—example‑first search, feature selection, bounded clause expansion, and synchronized windowing—that together make multisource symbolic learning feasible for real‑time cardiac monitoring.

The authors discuss several avenues for future research. Extending the framework to additional signals such as pulse oximetry, respiratory flow, and even wearable accelerometer data could further enrich the diagnostic vocabulary. Combining the symbolic ILP component with deep‑learning feature extractors may yield hybrid models that leverage both high‑level interpretability and low‑level pattern recognition. Finally, prospective clinical trials are needed to validate the system’s utility in bedside decision support, assess its impact on patient outcomes, and explore integration with existing monitoring infrastructure.

In summary, the paper demonstrates that carefully engineered multisource ILP can produce accurate, computationally efficient, and clinically meaningful diagnostic rules for cardiac arrhythmia detection, thereby advancing the state of the art in explainable AI for health monitoring.