An artificial intelligence system has accurately predicted when patients with heart conditions will die, according to new results published in the journal Radiology.
The study was conducted by a team of scientists at the London Institute of Medical Services, who trained the software to analyze blood tests and intricate 3D models of beating hearts in order to detect signs of failure. The AI was assigned 256 patients diagnosed with pulmonary hypertension, a type of high blood pressure which impacts the lungs and can cause dizziness, fainting, and shortness of breath.
By tracking the movement of 30,000 different points on a patient’s heart, it was able to construct an intricate 3D scan of the organ. Combining these models with patient health records going back eight years, the system could learn which abnormalities signaled a patient’s approaching death, making predictions about five years into the future.
The AI predicted with 80 percent accuracy which patients would die in the next year. The average doctor’s accuracy is about 60 percent.
Treatments for pulmonary hypertension include drugs, lung transplants, and targeted medicines, but the choice of treatment depends on the prognosis. As a result, an AI with such high accuracy can help physicians better treat patients.
“The AI really allows you to tailor the individual treatment,” Dr. Declan O’Regan, one on the researchers, told BBC News. “So it takes the results of dozens of different tests including imaging, to predict what’s going to happen to individual patients very accurately. So we can tailor getting absolutely the right intensive treatment to those who will benefit the most.”
AI software is becoming increasingly adept at diagnosing diseases. In July, Google announced success in diagnosing eye disease using machine learning software. A month earlier, researchers from Harvard Medical School (HMS) and Beth Israel Deaconess Medical Center (BIDMC) demonstrated a system that could detect breast cancer with 92 percent accuracy. When combined with the analysis of pathologists, that rate shot up to 99.5 percent.
“We expect the deep pathology network to continue to improve as it is trained on increasingly large and diverse pathology data sets,” Andrew Beck, Harvard professor and co-author of the study, told Digital Trends.
