Accelerated LV Activation Mapping via Geometric Deep Learning for CRT Planning

Overview

This study introduces two geometric deep learning models, GNN and GINO, to rapidly predict left ventricular activation time maps, aiming to optimize cardiac resynchronization therapy (CRT) lead placement. The GINO model demonstrated superior accuracy on synthetic data, and both models performed comparably on real patient geometries, enabling robust identification of optimal pacing sites.

Background

Cardiac resynchronization therapy is a key treatment for patients with dyssynchronous heart failure but suffers from a high non-responder rate, partly due to suboptimal left ventricular lead placement. Current individualized planning methods are limited by anatomical variability and complexity. Computational modeling offers a promising avenue to personalize CRT by predicting cardiac activation patterns and guiding lead positioning. This study leverages geometric deep learning to accelerate and enhance the precision of such predictions.

Data Highlights

Key Findings

• Two geometric deep learning models, GNN and GINO, were developed to predict LV activation time maps in real time.
• GINO outperformed GNN on synthetic datasets with a lower prediction error (1.38% vs 2.44%).
• Both models showed comparable accuracy on real-world LV geometries (GINO 4.79%, GNN 4.07%).
• The models enabled robust recovery of subject-specific parameters even from noisy input data.
• An interactive web interface was created to facilitate clinical use and pacing site optimization.
• The approach supports personalized pre-procedural CRT planning to potentially improve response rates.

Clinical Implications

These geometric deep learning models provide a rapid and accurate method to predict LV activation patterns, which can guide optimal lead placement in CRT. By enabling personalized and data-driven pacing site selection, this approach may reduce non-responder rates and improve clinical outcomes. Integration into clinical workflows via user-friendly interfaces could facilitate adoption and enhance decision-making.

Conclusion

Geometric deep learning models represent a promising tool for accelerating and personalizing CRT planning by accurately predicting LV activation and optimizing pacing sites. Future work should focus on clinical validation and integration into decision-support systems.

References

1. Glikson et al. 2022 -- 2021 ESC guidelines on cardiac pacing and cardiac resynchronization therapy
2. Almeida et al. 2017 -- Cardiac resynchronization therapy in heart failure: do evidence-based guidelines follow the evidence?
3. McAlister et al. 2007 -- Cardiac resynchronization therapy for patients with left ventricular systolic dysfunction: a systematic review
4. Abraham et al. 2002 -- Cardiac resynchronization in chronic heart failure
5. Wijesuriya et al. 2023 -- Pacing interventions in non-responders to cardiac resynchronization therapy
6. Wouters et al. 2021 -- Optimizing lead placement for pacing in dyssynchronous heart failure: the patient in the lead