Application of machine learning methods to healthcare outcomes research

Value in Health, the official journal of ISPOR- the professional society for health economics and outcomes research, announced today the publication of a high-level overview of machine learning for healthcare outcomes researchers and decision makers. The report, "Machine Learning for Health Services Researchers," was published in the July 2019 issue of Value in Health.

Machine learning is a rapidly growing field that attempts to extract general concepts from large datasets, commonly in the form of an algorithm that predicts an outcome-;a task that has become increasingly difficult to accomplish by humans because data volume and complexity has increased beyond what was capable with traditional statistics and desktop computers.

Machine learning methods may be useful to health service researchers seeking to improve prediction of a healthcare outcome with large datasets available to train and refine an estimator algorithm. Machine learning methods can help generalizable data-driven estimators when many covariates are being selected among and when the outcome of interest may be produced by complex nonlinear relationships and interaction terms.

In this report, the authors introduce key concepts for understanding the application of machine learning methods to healthcare outcomes research. They first provide an overview of machine learning, then identify 5 steps to developing and applying a machine learning algorithm (commonly referred to as a predictive model or estimator): (1) data preparation, (2) estimator family selection, (3) estimator parameter learning, (4) estimator regularization, and (5) estimator evaluation.

The report goes on to compare 3 of the most common machine learning methods: (1) decision tree methods that can be useful for identifying how different subpopulations experience different risks for an outcome, (2) deep learning methods that can identify complex nonlinear patterns or interactions between variables predictive of an outcome, and (3) ensemble methods that can improve predictive performance by combining multiple machine learning methods. Finally, the authors demonstrate the application of common machine methods to a simulated insurance claims dataset.

While machine learning methods may be useful to health service researchers, they offer considerable challenges that are worth considering before engaging in a machine learning activity. Specifically, they may be difficult to interpret (particularly for deep learning), difficult to glean mechanistic understandings from, and may require substantial investment of time and resources for computation. Nevertheless, improvements in hardware and cloud computing technologies have made machine learning methods increasingly accessible to healthcare outcomes researchers and healthcare organizations. With this article, we aim to lower the barriers to implementing machine learning methods."

Patrick Doupe, PhD, Zalando SE, Berlin, Germany