Atrial Fibrillation Risk Prediction using Deep Learning
Sun Young Park, Dustin Sargent, Marwan Sati, Rebecca Trunnell Hyman, Vadim Peretroukhine, Yury Vorontsov
Background
Atrial fibrillation (AF), which is associated with increased risk of stroke, heart failure and all-cause mortality, is expected to increase in prevalence in the coming years due to an aging population. AF is also correlated with risk of sudden cardiac death, which is one of the leading causes of death worldwide. Implanted cardioverter defibrillators (ICDs) can stop AF and other arrythmias and thus prevent adverse cardiac events. However, most applicable patients do not meet the clinical criteria for implantation of an ICD (left ventricular ejection fraction < 30-35%) prior to experiencing adverse events. AF is also intermittent and may not show up during clinical visits, thus often not being diagnosed until an event occurs. It is therefore important to develop prediction models to identify at-risk patients early and take preventative measures. In this paper, we present a novel machine learning model that predicts the likelihood of future occurrence of AF from a large set of electronic health record (EHR) data and echocardiography measurements. Conversely, previous work has focused on regression models for AF risk prediction using a small number of EHR features. We also analyze the features used by the model to determine which are most highly correlated with the AF condition.
Methods
Our models were trained on patient data from a database of 12607 patient encounter records with 232 recorded features including patient history and demographic information, study information and echo measurements. AF status was given by the echo baseline electrocardiogram (ECG) rhythm, which had 7 possible values: normal sinus, sinus bradycardia, sinus tachycardia, atrial flutter, atrial fibrillation, paced, or v-paced rhythm. The data contained 992 records with atrial fibrillation and 11615 with normal or other echo baseline ECG rhythm values. Patient age ranged from 16 to 64, 51.8% of the patients were female, and 74.4% were Caucasian. Tobacco use, diabetes, and hypertension were present in 0.7%, 0.6%, and 3.3% of the patients respectively. Positive cases were defined as patients who had atrial fibrillation recorded as their echo baseline ECG rhythm at any encounter. Negative cases were patients who had only normal echo baseline ECG rhythm at all encounters (AF vs normal), or patients who had any values other than atrial fibrillation at all encounters (AF vs non-AF).
We used a 1D convolutional neural network with residual blocks and fully connected layers for the AF classification task. We tested this model with three input variations: EHR data only, echo measurement data only, and combined EHR and echo measurement data. We also tested models that predicted current and future AF. For the combined input model, we first performed a regression analysis, and eliminated features whose correlation with the target value was not statistically significant. We then scaled the health record data and echocardiographic measurements to compatible ranges and combined them into single feature vectors for each patient encounter. Models were trained using 5-fold cross validation, and performance was evaluated with area under the ROC curve on a held-out test dataset.
Results
We first trained several models on single encounter current AF vs normal classification. The combined model achieved the best ROC-AUC score of 0.90 (95% CI: 0.88 – 0.92) on a held-out test set of 1547 normal encounter records and 214 AF records. The echo measurement-only model achieved 0.86 ROC-AUC on AF vs normal classification, and the EHR-only model achieved 0.62. This was likely due to a high percentage of missing values in the EHR data. We also observed that the combined model was able to distinguish AF from all other echo baseline ECG rhythm values, achieving 0.89 (95% CI: 0.86 – 0.91) ROC-AUC on AF vs non-AF classification on a held-out test set of 1738 non-AF records and 214 AF records. Finally, we trained a combined model on future AF vs non-AF classification using only the patient’s first encounter data as input, where patients were considered positive if they had any future AF diagnosis. The model achieved 0.81 (95% CI: 0.79 – 0.83) ROC-AUC for future AF prediction on a held-out test set of 791 negative (non-AF) patients and 110 positive patients.
We also performed independent training analyses with reduced feature sets to determine which features were most highly correlated with AF in our dataset. The most relevant features came from the echo measurements, and were primarily related to atrium and ventricle size, arterial stenosis, and valve function. Some examples from the top ten included left atrium size, right atrium size, mitral valve regurgitation, pulmonic valve stenosis, and left ventricle diastolic function.
Conclusions
We have presented a novel deep learning model for predicting atrial fibrillation from EHR and echo measurement data. The model accurately predicts current or future AF using data from a single encounter, and can distinguish patients who are at-risk of AF from those who may develop other abnormal heart rhythm conditions. This model can identify high risk patients who may not exhibit AF during visits with their physician, which may help to prevent future life-threatening cardiac events.

**Atrial Fibrillation Risk Prediction using Deep Learning**
Sun Young Park, Dustin Sargent, Marwan Sati, Rebecca Trunnell Hyman, Vadim Peretroukhine, Yury Vorontsov
**Background**
Atrial fibrillation (AF), which is associated with increased risk of stroke, heart failure and all-cause mortality, is expected to increase in prevalence in the coming years due to an aging population. AF is also correlated with risk of sudden cardiac death, which is one of the leading causes of death worldwide. Implanted cardioverter defibrillators (ICDs) can stop AF and other arrythmias and thus prevent adverse cardiac events. However, most applicable patients do not meet the clinical criteria for implantation of an ICD (left ventricular ejection fraction < 30-35%) prior to experiencing adverse events. AF is also intermittent and may not show up during clinical visits, thus often not being diagnosed until an event occurs. It is therefore important to develop prediction models to identify at-risk patients early and take preventative measures. In this paper, we present a novel machine learning model that predicts the likelihood of future occurrence of AF from a large set of electronic health record (EHR) data and echocardiography measurements. Conversely, previous work has focused on regression models for AF risk prediction using a small number of EHR features. We also analyze the features used by the model to determine which are most highly correlated with the AF condition.
**Methods**
Our models were trained on patient data from a database of 12607 patient encounter records with 232 recorded features including patient history and demographic information, study information and echo measurements. AF status was given by the echo baseline electrocardiogram (ECG) rhythm, which had 7 possible values: normal sinus, sinus bradycardia, sinus tachycardia, atrial flutter, atrial fibrillation, paced, or v-paced rhythm. The data contained 992 records with atrial fibrillation and 11615 with normal or other echo baseline ECG rhythm values. Patient age ranged from 16 to 64, 51.8% of the patients were female, and 74.4% were Caucasian. Tobacco use, diabetes, and hypertension were present in 0.7%, 0.6%, and 3.3% of the patients respectively. Positive cases were defined as patients who had atrial fibrillation recorded as their echo baseline ECG rhythm at any encounter. Negative cases were patients who had only normal echo baseline ECG rhythm at all encounters (AF vs normal), or patients who had any values other than atrial fibrillation at all encounters (AF vs non-AF).
We used a 1D convolutional neural network with residual blocks and fully connected layers for the AF classification task. We tested this model with three input variations: EHR data only, echo measurement data only, and combined EHR and echo measurement data. We also tested models that predicted current and future AF. For the combined input model, we first performed a regression analysis, and eliminated features whose correlation with the target value was not statistically significant. We then scaled the health record data and echocardiographic measurements to compatible ranges and combined them into single feature vectors for each patient encounter. Models were trained using 5-fold cross validation, and performance was evaluated with area under the ROC curve on a held-out test dataset.
**Results**
We first trained several models on single encounter current AF vs normal classification. The combined model achieved the best ROC-AUC score of 0.90 (95% CI: 0.88 – 0.92) on a held-out test set of 1547 normal encounter records and 214 AF records. The echo measurement-only model achieved 0.86 ROC-AUC on AF vs normal classification, and the EHR-only model achieved 0.62. This was likely due to a high percentage of missing values in the EHR data. We also observed that the combined model was able to distinguish AF from all other echo baseline ECG rhythm values, achieving 0.89 (95% CI: 0.86 – 0.91) ROC-AUC on AF vs non-AF classification on a held-out test set of 1738 non-AF records and 214 AF records. Finally, we trained a combined model on future AF vs non-AF classification using only the patient’s first encounter data as input, where patients were considered positive if they had any future AF diagnosis. The model achieved 0.81 (95% CI: 0.79 – 0.83) ROC-AUC for future AF prediction on a held-out test set of 791 negative (non-AF) patients and 110 positive patients.
We also performed independent training analyses with reduced feature sets to determine which features were most highly correlated with AF in our dataset. The most relevant features came from the echo measurements, and were primarily related to atrium and ventricle size, arterial stenosis, and valve function. Some examples from the top ten included left atrium size, right atrium size, mitral valve regurgitation, pulmonic valve stenosis, and left ventricle diastolic function.
**Conclusions**
We have presented a novel deep learning model for predicting atrial fibrillation from EHR and echo measurement data. The model accurately predicts current or future AF using data from a single encounter, and can distinguish patients who are at-risk of AF from those who may develop other abnormal heart rhythm conditions. This model can identify high risk patients who may not exhibit AF during visits with their physician, which may help to prevent future life-threatening cardiac events.