Machine learning techniques show promise for supporting medical decisions

Feb. 28, 2018

Several studies being presented at the American College of Cardiology’s 67th Annual Scientific Session demonstrate how the computer science technique known as machine learning can be used to accurately predict clinical outcomes in patients with known or potential heart problems. Collectively, the findings suggest that machine learning may usher in a new era in digital healthcare tools capable of enhancing healthcare delivery by aiding routine processes and helping physicians assess patients’ risk.

While clinical scoring systems and algorithms have long been used in medical practice, there has been a marked uptick in the application of machine learning to improve such tools in recent years. In contrast to traditional algorithms that require all calculations to be pre-programmed, machine learning algorithms deduce the optimal set of calculations by looking for patterns in large collections of patient data.

The new studies presented at ACC.18 demonstrate how machine learning can be used to predict outcomes such as diagnosis, death, or hospital readmission; improve upon standard risk assessment tools; elucidate factors that contribute to disease progression; or to advance personalized medicine by predicting a patient’s response to treatment.

In one study, researchers used machine learning to predict which patients would ultimately be diagnosed with a heart attack after coming to a hospital emergency department for chest pain. Chest pain is one of the most common complaints among patients visiting the emergency department, yet only a minority of these patients are ultimately diagnosed with a heart attack. In a pilot test, the algorithm showed a 94% accuracy rate for predicting the ultimate diagnosis.

Daniel Lindholm, MD, PhD, postdoctoral research fellow at Uppsala University in Sweden and the study’s lead author, and his colleague developed and tested their algorithm using data from more than 8,200 emergency department visits in Stockholm between 2011 and 2013. In the first (training) phase, the algorithm used data from 5,800 patient visits to iteratively refine decision trees to identify a heart attack diagnosis based on factors such as blood test results, vital signs and the patient’s medical history—the same data that a physician would have upon first encountering a patient.

In the second (validation) phase, the algorithm uses the same types of patient data from a separate set of 2,400 visits, including blood test results, vital signs and medical history, but information about ultimate heart attack diagnosis was excluded to test how accurately the algorithm could predict which patients were ultimately diagnosed with a heart attack.

Overall, the algorithm was able to accurately predict a heart attack diagnosis 94% of the time in the validation data set. Ninety percent is generally considered a high rate of accuracy for models of this type. Researchers also ran the validation data through a standard clinical model (the hsTnT model, which incorporates only a patient’s age, sex and high-sensitivity troponin levels), which showed an accuracy of 88%. The results suggest the machine learning approach could offer a substantial improvement over current decision support tools.

To bring their algorithm closer to clinical application, Lindholm said the research team will first need to evaluate it in larger data sets that include more hospitals in diverse environments and then validate it in a prospective clinical study.

EurekAlert! has the full article