Using artificial intelligence (AI) to examine echocardiograms (ECGs) can identify patients at risk of developing atrial fibrillation (AF), as well as accurately predict one-year all-cause mortality, data presented at the American Heart Association Scientific Sessions (AHA 2019; 16-18 November, Philadelphia, USA) has demonstrated.
Brandon Fornwalt (Geisinger, Danville, Pennsylvania, USA) outlined the one-year mortality findings on behalf of his fellow investigators. The researchers hypothesised that a deep neural network can predict an important future clinical event (one-year all-cause mortality) from ECG voltage-time traces. Using Geisinger electronic health records, they extracted 1,775,926 12-lead resting ECGs, collected from 397,840 patients over a period of 34 years, along with their age, sex, and survival status. From the ECGs, they extracted 15 voltage-time 250 Hz to 500 Hz traces, including three standard ‘long’ 10-second and 12 ‘short’ 2.5-second acquisitions. In addition, they extracted ECG measures in the form of 30 diagnostic patterns and nine standard measurements.
Fornwalt et al then trained a deep neural network to predict one-year mortality directly from the ECG traces. Using different variable inputs, they compared fivefold cross-validated model performance and performed a Cox proportional hazard survival analysis on the predicted groups. Once they had confirmed good predictive accuracy among the 297,548 ECGs that had been evaluated as ‘normal’ by the physician, the researchers performed a blinded survey of three cardiologists to determine whether they were capable of seeing features indicative of mortality risk.
They found that the model trained with the 15 traces alone yielded an average area under the curve (AUC) of 0.83. This improved to 0.85 after adding age and sex. The model was superior to a separate, non-linear model created from the 39 ECG measures (AUC=0.77 and 0.81 without and with age and sex, respectively, p<0.001). Even within the “normal” ECGs, the model performance remained high (AUC=0.84), and the hazard ratio was 6.6 (p<0.005) beyond one-year post-ECG. In a blinded survey, the patterns captured by the model were not visually apparent to cardiologists, even after they had been shown labelled true positives (from patients who were dead) and true negatives (from patients who were alive).
They conclude that deep learning can be a powerful tool for identifying patients with potential adverse outcome who may benefit from early interventions, even in cases that have been interpreted as normal by physicians.
And in a poster presentation, Sushravya M Raghunath (also Geisinger, Danville, Pennsylvania, USA) and colleagues outlined a deep learning model that predicts future atrial fibrillation directly from 12-lead resting ECG voltage-time traces, with a mean AUC of 0.75 ± 0.02. In the subset of ECGs interpreted by the physician as ‘normal’, the AUC was 0.72 ± 0.02.
Though the vast Geisinger database is a key strength of both studies, the findings should be tested at sites outside of Geisinger, the researchers noted.
Fornwalt and coresearcher Chris Haggerty (also Geisinger) commented to Cardiac Rhythm News on the significance of the findings, saying: “We believe this type of model can be incorporated into clinical workflows and help to inform the assessment of patient risk and prognosis. For example, in the setting of advanced chronic disease, an accurate assessment of mortality risk (which we envision to be the product of a physician interpreting this mortality risk prediction in the context of other clinical signals, not solely based on an algorithm output alone) can be used to help inform referrals for palliative care, or signal that more aggressive treatment may be warranted.”
And they added that the algorithm they have developed is transferable to other settings: “This model is actually fairly simplistic in terms of the data it needs to run, with one version requiring no patient data other than the voltages acquired from the ECG. This design means that it would be a lot simpler to implement in other centres than complex models that require deep integration with a system’s data infrastructure.”