AI learned to predict failures in the heart with an accuracy of 73%

Achievements of modern medicine suggest a possible outcome for patients with heart disease, but this is a very long and laborious process - doctors have to analyze MRI scans, blood tests and other data without the help of computers, relying on their knowledge and experience. Scientists from Imperial College London came to the rescue of doctors who taught artificial intelligence to predict the risk of death in people with serious cardiovascular diseases faster and more accurately than other existing forecasting tools.

New software creates virtual 3D hearts for each patient that repeat every contraction of the organ. Artificial intelligence can quickly determine which functions of the heart indicate heart failure and death, using magnetic resonance imaging (MRI) data along with information about blood tests.
')

An example of a simulation for a patient with idiopathic pulmonary hypertension

Scientists have tested the work of technology in patients with pulmonary hypertension - a condition that leads to heart failure, if not treated properly. To prescribe treatment, doctors need to predict whether patients are in a high or low risk group, but modern methods do not allow it to do so with exceptional precision.

Pulmonary hypertension is characterized by high pressure in the blood vessels supplying the heart with oxygen — the pulmonary arteries, veins, or capillaries. The result is a load on the right side of the heart, which over time leads to progressive damage. If the patient does not go to the doctor on time, he risks losing the rest of his life with heart failure. In a patient with pulmonary hypertension, shortness of breath, fatigue, angina pectoris, fainting, coughing and other symptoms can be observed.

For the treatment of this disease, modern medicine uses drugs that allow blood to move more freely through the lungs, which helps patients at risk of living longer. In some cases, injections are administered directly into the blood vessels, and in particularly difficult situations, lung transplants are performed.

Until now, radiologists relied on time-consuming manual measurement of heart function to identify patients with the highest risk of deterioration. According to Dr. Declan O'Regan, the lead author of the study, for the first time they managed to teach a computer to interpret heart scans in order to determine exactly how long patients can live. Artificial intelligence can change the way people treat heart disease.

Using data from 256 patients, the software analyzed the moving MRI images of the heart of each patient and measured the movement of 30,000 different points in the structure of the organ during each heart beat. Then, the data were combined with medical records about the health of patients, which were maintained for eight years. The program has created a virtual 3D-heart of each person and automatically recognized which functions were the earliest harbingers of heart failure and death.


At the end of the observation, which lasted about four years, 36% of patients (93 of 256) died, and one of them underwent a lung transplantation. The computer correctly identified those who live more than a year, in 73% of cases. Given that the accuracy of the predictions of doctors in this situation was 60%.

Researchers are confident that the technology can be further used for patients with other types of cardiovascular diseases. Artificial intelligence is already helping research on cancer and brain diseases, but analyzing moving images of the heart is a more difficult operation.

Co-author of the study, Dr. Tim Dawes, who, together with his team, developed a learning algorithm, said: “The computer performs the analysis in seconds and simultaneously interprets data from images, blood tests and other studies without any human intervention. This can help doctors correctly diagnose disorders and promptly prescribe appropriate treatment or change the current one. ”

Scientists plan to test the software on patient data from other hospitals in order to decide whether there is a need for its wide distribution among doctors. Researchers also want to use this technology to predict other forms of heart failure, such as cardiomyopathy , to determine if a patient needs a pacemaker or other forms of treatment.

The ultimate goal is to develop software that can predict not only survival, but also which type of treatment is best suited for a particular patient.

Scientific work published in the journal Radiology
DOI: radiol.2016161315