Israeli scientists use machine learning for atrial fibrillation risk prediction.
====== AVISO IMPORTANTE! ======
Shany Biton and Sheina Gendelman, two M.Sc. students working under the supervision of Assistant Professor Joachim A. Behar, head of the Artificial Intelligence in Medicine laboratory (AIMLab.) in the Technion Faculty of Biomedical Engineering, wrote a machine learning algorithm capable of accurately predicting whether a patient will develop atrial fibrillation within 5 years. Conceptually, the researchers sought to find out whether a machine learning algorithm could capture patterns predictive of atrial fibrillation even though there was no atrial fibrillation diagnosed by a human cardiologist at the time.
Atrial fibrillation is an abnormal heart rhythm that is not immediately life-threatening but significantly increases patients’ risk of stroke and death. Warning patients that they are at risk of developing it can give them time to change their lifestyle and avoid or postpone the onset of the condition. It may also encourage regular follow-ups with the patient’s cardiologist, ensuring that if and when the condition develops, it will be identified quickly, and treatment will be started without delay. Known risk factors for atrial fibrillation include a sedentary lifestyle, obesity, smoking, genetic predisposition, and more.
Ms. Biton and Ms. Gendelman used more than one million 12-lead ECG recordings from more than 400,000 patients to train a deep neural network to recognize patients at risk of developing atrial fibrillation within 5 years. Then, they combined the deep neural network with clinical information about the patient, including some of the known risk factors. Both the ECG recordings and the patients’ electronic health records were provided by the Telehealth Network of Minas Gerais (TNMG), a public telehealth system assisting 811 of the 853 municipalities in the state of Minas Gerais, Brazil. The resulting machine learning model was able to correctly predict the development of atrial fibrillation risk in 60% of cases, while preserving a high specificity of 95%, meaning that only 5% of persons identified as being potentially at risk did not develop the condition.
“We do not seek to replace the human doctor – we don’t think that would be desirable,” said Prof. Behar of the results, “but we wish to put better decision support tools into the doctors’ hands. Computers are better equipped to process some forms of data. For example, examining an ECG recording today, a cardiologist would be looking for specific features which are known to be associated with a particular disease. Our model, on the other hand, can look for and identify patterns on its own, including patterns that might not be intelligible to the human eye.”
Doctors have progressed from taking a patient’s pulse manually, to using a statoscope, and then the ECG. Using machine learning to assist the analysis of ECG recordings could be the next step on that road.
Since ECG is a low-cost routine test, the machine learning model could easily be incorporated into clinical practice and improve healthcare management for many individuals. Access to more patients’ datasets would let the algorithm get progressively better as a risk prediction tool. The model could also be adapted to predict other cardiovascular conditions.
The study was conducted in collaboration with Antônio Ribeiro from the Uppsala University, Sweden and Gabriela Miana, Carla Moreira, Antonio Luiz Ribeiro from the Universidade Federal de Minas Gerais, Brazil.
The study was published in the European Heart Journal – Digital Health.