cAPM: Continual AI-Assisted Pace-Mapping with Active Learning

What Happened

Researchers have introduced cAPM (Continual AI-Assisted Pace-Mapping), a machine learning framework designed to assist clinicians during catheter ablation for ventricular tachycardia (VT). VT is a dangerous heart rhythm disorder that can lead to sudden cardiac death, and pace-mapping is a critical procedure where clinicians stimulate heart tissue to locate the arrhythmia's origin before delivering targeted ablation.

The core innovation is twofold. First, cAPM uses active learning to intelligently select which pacing sites to test next, reducing the number of unnecessary stimulations. Second, it operates in a continual learning paradigm—meaning the model improves its predictions in real-time as new data arrives during the procedure, without requiring retraining on the entire dataset. This contrasts with traditional static models that cannot adapt mid-procedure.

The system analyzes electrocardiogram (ECG) signals from pace-mapping and learns to predict which anatomical locations are most likely to be the VT origin. By prioritizing the most informative sites, cAPM aims to shorten procedure time, reduce patient discomfort, and improve ablation accuracy.

Why It Matters

Catheter ablation for VT is technically demanding and time-sensitive. Current practice often relies on clinician intuition and trial-and-error pacing, which can prolong procedures and increase risks. cAPM addresses a genuine clinical bottleneck: the need to map dozens or even hundreds of cardiac sites while maintaining precision.

From a healthcare perspective, this research signals a shift toward AI systems that are not just diagnostic tools but procedural assistants—actively guiding real-time decision-making. If validated in clinical settings, such systems could reduce procedure duration by 20-40%, lower radiation exposure from fluoroscopy, and improve first-pass ablation success rates.

For the broader AI community, cAPM exemplifies a growing trend: deploying machine learning in high-stakes, data-scarce environments where models must adapt on the fly. The continual learning component is particularly relevant—most medical AI models are trained offline and deployed as static systems, but cAPM demonstrates a path toward adaptive, patient-specific personalization during a single procedure.

Implications for AI Practitioners

Active learning in clinical workflows: cAPM's query strategy—choosing which pacing site to test next—is a textbook application of active learning in a safety-critical domain. Practitioners building similar systems should consider uncertainty sampling or expected model change as selection criteria, while ensuring that the AI's recommendations align with clinical safety constraints. Continual learning without catastrophic forgetting: The system must incorporate new data without overwriting prior knowledge. This is nontrivial in medical contexts where data distributions shift across patients. Techniques like elastic weight consolidation or memory replay may be necessary to maintain robustness. Real-time inference constraints: ECG signals arrive continuously, and the model must update predictions within seconds. Practitioners should optimize for low-latency inference (e.g., quantized models, on-device deployment) and design interfaces that present actionable recommendations without overwhelming clinicians. Evaluation beyond accuracy: Clinical impact depends on reducing procedure time and complications, not just classification accuracy. AI teams should collaborate with electrophysiologists to define meaningful endpoints—such as number of pacing sites avoided or ablation success rate—and validate models in prospective studies.

Key Takeaways

• cAPM introduces a continual active learning framework for real-time guidance during cardiac ablation, reducing unnecessary pacing and improving procedural efficiency.
• The system exemplifies a shift from static diagnostic AI to adaptive procedural assistants that learn patient-specific patterns during a single intervention.
• AI practitioners must address challenges in continual learning, real-time inference, and clinically meaningful evaluation metrics when deploying similar systems in high-stakes medical environments.
• If validated, this approach could set a precedent for AI-assisted electrophysiology procedures, potentially reducing procedure times and improving outcomes for VT patients.