Projects/03
Accepted at NeurIPS 2026
Longitudinal AFib Prediction
SurviFormer: longitudinal ECG survival modelling with external validation
We developed SurviFormer to estimate future documented atrial fibrillation risk from serial ECGs. The model combines a causal transformer, relative time biases and multi-landmark survival training. Evaluated on CODE and externally on MIMIC-IV, the work was accepted at NeurIPS 2026.
Project overview
- Problem
- Estimate future atrial fibrillation risk from a patient’s ECG history.
- My contribution
- Designed and implemented the ECG pipeline, survival model, training and internal/external evaluation at Inria.
- Approach
- Causal transformer with relative time biases and multi-landmark training to estimate discrete-time hazards.
- Constraints
- Handle irregular visits and censored outcomes using ECGs acquired before the first documented AFib event.
- Result
- C-index 0.853 on CODE and 0.805 on external MIMIC-IV validation. Accepted at NeurIPS 2026.
My contribution
I designed and implemented SurviFormer end to end: the ECG data pipeline, model architecture, survival training, baseline comparisons, and internal and external evaluation. The research was conducted at Inria under the supervision of Maxime Sermesant, co-author of the paper.
Research focus
SurviFormer estimates the risk of a future first documented atrial fibrillation (AFib) event from a patient's ECG history. The task is longitudinal risk prediction from intermittent recordings, rather than detection of AFib already present in an ECG.
Technical approach
- Encode ECG recordings as learned representations and process each patient's history with a causally masked transformer.
- Represent irregular gaps between visits as pairwise relative attention biases.
- Predict a discrete-time hazard function and use multi-landmark training to learn from different points in the available history.
- Account for censored outcomes: patients without a documented AFib event still contribute follow-up information.
These choices connect physiological time-series representation learning with time-to-event modelling. Risk estimates can be updated when another ECG becomes available.
Internal and external evaluation
The public NeurIPS abstract reports training and evaluation on CODE, with approximately 140,000 patients and 517,000 ECGs, and external validation on MIMIC-IV.
| Metric | CODE internal test | MIMIC-IV external cohort |
|---|---|---|
| C-index | 0.853 | 0.805 |
| Mean time-dependent AUC | 0.867 | 0.848 |
| Integrated Brier score | 0.030 | 0.106 |
The evaluation includes static survival baselines using ECG embeddings and a dynamic competing-risk model. Reporting both discrimination and Brier scores makes the assessment broader than a single classification metric.
Scope and limitations
The target is a documented AFib event in the available records. Intermittent ECGs do not establish the precise biological onset of AFib or continuous AFib burden. External cohort evaluation provides evidence of transfer beyond CODE; it does not establish prospective clinical effectiveness or deployment readiness.
Publication and resources
I co-authored this work with Maxime Sermesant at Inria. The paper, "Survival Transformers for Longitudinal Data Analysis: Application to Atrial Fibrillation Risk from ECG," was accepted at NeurIPS 2026; proceedings publication is forthcoming.
The methods and results above are already public in the official NeurIPS abstract. The OpenReview paper page provides the research record.