[Heart News] Artificial intelligence ECG and clinical risk factors forecast atrial fibrillation

Artificial Intelligence (AI) Analysis 12 Level ECG (ECGS) helps effectively evaluate the risk of trembling (AF). However, except for clinical risk factors, it is not clear whether artificial intelligence provides meaning and further improvement of accurate prediction of AF.

The study was conducted by 12 lead electrocardiograms for long-term primary health care patients (MGH) in Massachusetts General Hospital (MGH), which was inferred by convolutional neural network ("ECG-AI"). Subsequently, three Cox proportional risk models were constructed, each model included: a) ECG-AI 5-year room fuse, b) CHARGE-AF clinical risk score, and C) ECG-AI and Charge-AF rating ("CHAI"). The researchers evaluate model performance by calculating judgment (under the area of ??ROC curve, AUROC) and calibration in an internal test set and two external test sets (Brigham and Women’s Hospital and British Bioconbound). The model is re-calibrated in the British Biological Bank, and the risk of 2 years of room tribin is estimated. The most significant electrocardiogram characteristics of ECG-AI risk predictions were studied by significant mapping, and the correlation between ECG-AI and Charge-AF linear prediction factors were evaluated.

The training set includes 4,577 (age 55 ± 17 years old, 53% women, 2171 AF incidents), including 83,162 people (age 59 ± 13 years old, 56% women, 2424 AF events). CHARGE-AF (MGH 0.802, 95% Ci 0.767-0.752, 95% Ci 0.741-0.752, 95% Ci 0.741-0.763; UK Biobank 0.732, 95% Ci 0.704-0.759) and ECG-Ai (MGH 0.823, 95% Ci 0.790- 0.856; BWH 0.747, 95% Ci 0.736-0.759; UK BIOBANK 0.705, 95% CI 0.673-0.737) Auroc comparison. The CH-AI group Auroc has the highest: MGH 0.838, 95% Ci 0.807-0.869; BWH 0.777, 95% Ci 0.766-0.788; UK BIOBANK 0.746, 95% CI 0.716-0.776). ECG-AI (MGH 0.0212; BWH 0.0129; UK Biobank 0.0035) and CH-Ai (MGH 0.012; BWH 0.0108; UK BIOBANK 0.0001) Calibration error. In significant analysis, the effect of electrocardiogram P wave on artificial intelligence model predictions is the largest. ECG-AI and Charge-AF linear predictors (Pearson R MGH 0.61, BWH 0.66, UK BIOBANK 0.41).

Table 1 Baseline Information

This study developed a depth learning model ECG-AI, using 12 lead electrocardiogram data precisely predicts the time of the atrial fibrillation. ECG-Ai uses approximately 100,000 electrocardiogram training from more than 40,000 patients from the primary health queue. CH-AI is a model of ECG-AI and Charge-AF, compared with Charge-Af, in patients with more than 80000 clinical features, a large number of clinical features of more than 80000 cases, in multiple prognostic model indicators The effect is better. ECG-AI evaluates the risk of CHARGE-AF rating containing 11 components. In addition, the study found that ECG-AI and CHARGE-AF are highly correlated, indicating that ECG-AI can largely reflect the ECG performance of the determined atrial fibrillation clinical hazardous factor. The results show that the clinical risk signal based on deep learning provides similar predictive utility, ECG-AI and clinical risk factors complement each other, enhance risk prediction skills.

Table 2 Test Concentrated atrial fuse

The Attia team has developed a depth learning model that is 80% for the model of atrial fibrillation state classification of diagnosis room fibrillation and sinus rhythm. Raghunath et al. Subsequently developed a neural network that utilizes 12 lead-optic maps to predict atrial fibrillation event. In one year, the atrial fibrillation recognition capacity is good, in subset analysis, compared with Charge-Af, performance is better. The study made substantial supplements in the aforementioned work, introduced a deep learning model, explicitly incorporating the survival time, and conducted strict epidemiological surveys, including quantitative errors, calibration, re-classification and extensive external validation. However, due to the data is not applicable, the study cannot be directly compared to the aforementioned model.

The results show that when using a strict epidemiological index to assess, the depth learning model uses an electrocardiogram to assess the risk of atrial fibrillation is reliable and effective. Specifically, the study evaluated ECG-AI of the test set consisting of independent individuals from: a) the same institution with the training set, b) independent institutions in the same medical network, c) from atrial fibrillation risk Lower different region forward-looking research queues. With the previous research results, ECG-Ai is best in the closest population closest to the training set, and the difference in different samples is small, highlighting the importance of extensive external verification in assessing clinical utility. In the end, researchers suspect that the differences in identification ability may differ from different characteristics (eg, age, baseline atrial risks), and thus the specific electrocardiogram characteristics are different from the relevance between the risks of the long-term atrial fibrillation.

This study is of great significance to the relationship between depth study of ECG hazardous signals and traditional atrial fibrillation clinical risk factors. First, the clinical hazardous factors can be performed on an electrocardiogram in a perceived manner in the deep learning model. Specifically, the ECG-AI probability and the Charge-AF score have moderate correlations in each test. In addition, the characteristics map and median waveform analysis have found that ECG-Ai probability is largely influenced by atrial decoder and a replenishment period, and the atrial structure and function may be affected by chronic diseases such as age and hypertension.

Second, the depth learning model seems to extract a risk factor complementary to clinical hazardous factors. CH-AI is a combination of CHARGE-AF and ECG-AI, which always shows better AF identification, calibration, and re-classification. This shows that the predicted effect of the model combined with the clinical risk information and the artificial intelligent veneer risk is significantly improved compared to any of the methods alone. Similarly, atrial fibrillation is significantly higher based on the risk of ECG-AI and Charge-Af, based on the use of two models alone. In addition, the predictive differences in CH-AI and Charge-AF are more significant over time, which may be related to the time accumulation effect of clinical risk factors, while the authentication of ECG-Ai remains relatively unchanged, prompting an electrocardiogram-based atrial fibrillation Risk predictions can contribute the most in short-term predictions.

The design background is explained below. First, the researchers conduct ECG-Ai training for individuals who conduct at least one ECG clinical examination. At the same time, the patient baseline information includes each component of the charge-AF rating. These two requirements have introduced potential selection bias. However, the study included EHR samples accepted longitudinal primary care may reduce bias, and the research model continues to distinguish AF in a completely independent forward-looking queue study. Second, the research training set represents an individual of a single mechanism. Training for larger samples in multiple institutions helps to generate more accurate and more universal models. Third, due to the limited follow-up of British bio-bank, the evaluation time window is 2 years, while MGH and BWH are five years.

Fourth, ECG-AI is a black box model. However, compared with the previous AF prediction model, the study uses a significant mapping and median waveform analysis, and the biological reasonable electrocardiogram changes (for example, P-wave duration prolong) have the greatest impact on the AF risk assessment. Fifth, 5 years of predictive window may represent the risk of atrial fibrillation, but it cannot be intervened immediately. However, the study also has a consistent difference in a shorter time window. It is not yet excluded that ECG-AI recognizes the previously have unexposed atrial vibration. Sixth, ECG-AI and CH-AI need to recalibrate in Biobio Bioconbanks. It is necessary to recalibrate when replacing the prognostic model. It has been very accurately estimated by simple recalibration of the incidence of the atrial fibrillation of BBBs. Tip The initial calibration error can be completely attributed to the incidence of baseline atrial fibrillation in the British branch, approximately one-third of MGH and BWH. In addition, CH-Ai has a better calibration effect than Charge-Af, although Charge-AF has a similar recalibration in Bio Biological Bank.

Figure 1 Research overview

Figure 2 identification of atrial fibrillation

Figure 3 Correction of atrial fibrillation

Figure 4 In accordance with predicted risk stratified atrial fuse

Figure 5 ECG-AI depth learning model two visual behavior forms

In summary, in three independent test sets of more than 80,000 people, ECG-AI (a depth learning model using 12 lead electrocardiogram clearly predicts the time of AF occurrence time) is compared to Charge-AF clinical risk score containing 11 components. Provide a 5-year AF risk forecast similar to the effect. The CH-A model combines Charge-Af and ECG-AI to identify, calibrate, and re-categorize more accurately. Based on the depth learning ECG’s atrial risoting assessment has the potential for wide application, it can provide accurate and promotion of atrial fuser risk assessment within a few years of ECG inspection.


Khurshid Shaan, Friedman Samuel, Reeder ChristoPher, et al. Electrocardiogram-based deep Learning and Clinical Risk Fabrtors to Predict Atries, 2021, In Press.