AFNet AI Prediction Model (Paroxysmal AF Detection)
AFNet AI Prediction Model (Paroxysmal AF detection): explanation and context
This tool provides two complementary approaches: (1) a CHARGE-AF–style clinical proxy estimator that uses demographic and common clinical predictors to compute an approximate probability of paroxysmal AF detection on prolonged monitoring, and (2) an optional slot to combine an external ECG-AI model probability (if you have a deployed ECG model that returns a numeric probability). The clinical proxy implemented here is intentionally transparent and interpretable; it must be recalibrated to local populations before clinical use. Clinical risk models such as CHARGE-AF were developed to predict incident AF using simple clinical variables and remain widely used as a baseline comparator. Recent electrocardiogram-based deep learning models (end-to-end ECG AI and convolutional networks) have shown promising performance for predicting current or future AF from ECG traces, but these require model weights and careful external validation before clinical deployment. If you possess the original AFNet / AF detection model weights or an API endpoint that returns a calibrated probability, integrate that into the ECG-AI slot: send the raw ECG trace to your model, receive probability, then replace the ensemble logic here with the published model's exact formula or calibrated stacking.
References: Alonso A et al. CHARGE-AF risk model (J Am Heart Assoc. 2013) — CHARGE-AF is a widely cited clinical predictor for incident AF.
AFibNet / AFibNet-style convolutional ECG detection studies (implementation and example CNN architectures).
End-to-end ECG deep neural network risk prediction of AF (recent deep-learning work showing ECG-AI potential).
Note: I did not find a peer-reviewed model published under the exact name “AFNet” with publicly available coefficients; AFNET is also the name of the Atrial Fibrillation Network (research consortium) which runs AI/AF projects — if you have their published model file or reference, we can directly implement it and replace the proxy weights.