Learning to Predict Global Atrial Fibrillation Dynamics from Sparse Measurements

TL;DR

FibMap addresses the challenge of reconstructing global atrial fibrillation dynamics from sparse sequential contact mappings by formulating imputation mapping as a spatiotemporal graph neural network. It reconstructs whole-atria activity from as little as 10% observed surface, achieving a MAE of $0.0574 \pm 0.0005$ and markedly improved phase-singularity tracking ($\mathrm{TPR}=0.8924 \pm 0.0342$) compared with baselines. The method generalises to new patients via a fine-tuning procedure and demonstrates clinical relevance by matching fidelity to non-continuous clinical mappings (HD Grid) and capturing patient-specific dynamics through cross-modal validation against AcQMap ground truth. Additionally, FibMap reveals structured state- and embedding-space representations that support electrophenotyping and potential personalised ablation planning, highlighting its practical impact in improving AF care from sparse, routine measurements.

Abstract

Catheter ablation of Atrial Fibrillation (AF) consists of a one-size-fits-all treatment with limited success in persistent AF. This may be due to our inability to map the dynamics of AF with the limited resolution and coverage provided by sequential contact mapping catheters, preventing effective patient phenotyping for personalised, targeted ablation. Here we introduce FibMap, a graph recurrent neural network model that reconstructs global AF dynamics from sparse measurements. Trained and validated on 51 non-contact whole atria recordings, FibMap reconstructs whole atria dynamics from 10% surface coverage, achieving a 210% lower mean absolute error and an order of magnitude higher performance in tracking phase singularities compared to baseline methods. Clinical utility of FibMap is demonstrated on real-world contact mapping recordings, achieving reconstruction fidelity comparable to non-contact mapping. FibMap's state-spaces and patient-specific parameters offer insights for electrophenotyping AF. Integrating FibMap into clinical practice could enable personalised AF care and improve outcomes.

Figures (7)

• Figure 1: The inputs (A) and architecture (B) of FibMap. A) The atrium is discretised into nodes and edges via a triangulated mesh, and from this the graph adjacency matrix (describing the coupling between nodes) and the observed time series are derived. The patient-specific parameters (node and patient embeddings) are also provided as input to personalise FibMap’s imputation maps. B) FibMap is instantiated as a bidirectional graph recurrent neural network (GRNN), a state-space model that learns a representation for each node by propagating information forwards and backwards through space and time. Node and patient embeddings are provided to a multilayer perceptron (MLP) decoder, which personalises how the learnt representation is mapped to a signal value.
• Figure 2: Quantitative results of FibMap imputation mapping. A) Reconstruction loss as a function of the number of epochs for the fine-tuning of FibMap on the validation set, with a strong correlation of $r=0.97$ ($p<0.0001$) present between the loss curves of the observed patch and whole atria. B) Test set reconstruction performance of all models quantified using the mean absolute error (MAE) across all space and time. C) Test set performance for detecting phase singularities (PSs), quantified using the true positive rate of their detection.
• Figure 3: Qualitative results of FibMap imputation mapping. A) Snapshots of the imputation maps of FibMap and MF, versus the ground truth AcQMap phase maps for two different patients (rows). FibMap imputes the missing signals (grey regions) from sparse observations (10% of atria) of AcQMap recordings with a simulated sequential contact mapping multipolar catheter (surface area of 10%, no spatial overlap and 1 second dwell time). B) Node-level temporal reconstructions for two different patients (rows) for FibMap and MF. Both the maps and signal traces demonstrate the superior performance of FibMap.
• Figure 4: Validation of FibMap imputation maps from EnSite Precision HD Grid Mapping against non-contemporaneous ground truth AcQMap recordings. A) Sliding window cross-correlation analysis between AcQMap and FibMap phase signals, enabling comparison between non-contemporaneous recordings by measuring the similarity between AF patterns across different time windows. The correlation matrix below shows pairwise comparisons between all possible window combinations. For three patients (indexed i-iii), B) shows the kernel density estimates of pairwise cross-correlations computed between AcQMap and FibMap maps from the same patient (intra), different patients (inter), and spatiotemporally shuffled maps (random baseline), using 0.5-second sliding windows. Vertical dashed lines indicate the 99th percentile, with consistently higher values for intra-patient (0.19-0.22) versus inter-patient (0.16-0.17) correlations and shuffled baseline (0.02), demonstrating patient-specific pattern capture. C) Temporal robustness analysis showing the 99th percentile of cross-correlations against sliding window duration. Non-overlapping confidence intervals (computed via bootstrap sampling) between intra-patient and other distributions confirm that FibMap can capture patient-specific dynamics across different temporal scales. D) Representative snapshots of phase maps from highly correlated windows, comparing AcQMap recordings with FibMap maps derived from real HD Grid catheter measurements covering $<10$% of the atrial surface per time. The maps demonstrate the ability of FibMap to capture AF dynamics which are not directly visible in the original HD Grid measurements, including multiple wavefronts and rotational patterns.
• Figure 5: Results of FibMap sensitivity analysis. A) MAE of FibMap reconstructions as a function of catheter surface area, dwell time and entropy (left vs. right). B) Reconstruction MAE of FibMap and the predicted confidence intervals as a function of imputation horizon.