Machine learning applications in medicine usually entail analysis and prediction. It’s easy to imagine multiple applications for AI in public health, resource allocation, disease trends, and cost containment. From a patient’s perspective, however, machine learning’s ability to analyze a vast array of variables only matters if it helps with diagnosis, care plans, and outcome prediction. We’ve written about studies of specific machine learning applications such as identifying suicidal adolescents, predicting pain from video data, and detecting sepsis.

Researchers at Turku PET Centre, Finland, presented a study that showed machine learning can outperform humans in successfully predicting death or heart attack. At the International Conference on Nuclear Cardiology and Cardiac CT (ICNC) in Lisbon, Portugal, Dr. Luis Eduardo Juarez-Orozco, the study author, advised caution in using machine learning algorithms in place of doctors but pointed out that the results of the study show advances far ahead of current medical practice.

The study analyzed six years’ of health data for 950 patients who reported chest pain and underwent the PET Centre’s standard protocol to detect coronary artery disease. During those six years, 24 people had heart attacks and 49 died from any cause.

Doctors typically use ten clinical variables to make treatment decisions. For this study, the research team gathered 75 variables from CT and PET scans along with the usual ten clinical variables, and then provided them to LogitBoost, a machine learning algorithm. LogitBoost repeatedly analyzed the data seeking the patterns that had the best success at predicting a heart attack or death. The conventional ten variables alone had unspecified “modest” results, according to Juarez-Orozco. With the additional CT and PET scan variables, the algorithm was more than 90% accurate.

Juarez-Orozco summarized the study results saying that while doctors collect a large amount of information about patients, machine learning that integrates that traditional information with additional data points can significantly improve individual risk prediction. Incorporating data from both sources results in personalized medical treatment with improved chances for positive patient outcomes.