Low-value health care – care that provides little health benefit in light of its costs and risks to patients – is a central concern for policy-makers. Diagnostic testing is a particularly important example: while the use of high cost diagnostic tests has skyrocketed, for many the frequency with which the tests that are performed identify new diagnoses or trigger effective interventions appears low, while some patients with likely high benefit go
untested. The challenge is to make an accurate prospective prediction of each patient’s likely benefit at the time the testing decision is made.
To this end, we draw on recent advances in machine learning algorithms to predict which Medicare patients are likely to benefit from specific diagnostic tests: stress testing for myocardial infarction; CT pulmonary angiography for pulmonary embolism; and spine MRI to detect treatable causes of back pain. These tests seek to detect treatable medical conditions with important health implications, but involve substantial costs, and often risks to patients. Building on early evidence that machine learning algorithms combined with massive datasets can make highly accurate predictions, we will predict test value using data available to the doctor at the time the test is ordered. We can then assess the extent to which doctors are (a) testing patients with predictably low benefit, or (b) failing to test patients with predictably high benefit. We can also use this approach to better understand whether well-documented variations in testing across regions and providers are
the result of provider behavior or underlying patient factors. We will then leverage this approach using electronic health record data. This will also allow us to assess the feasibility of translating our work into a decision support intervention that could be integrated into individual hospitals’ electronic health record systems.
With advances in medical technology and data science, there are new opportunities to better deploy health care resources to meet the health care needs of an aging population. By developing broadly applicable machine learning techniques for estimating risk in both Medicare claims and electronic health record data, this project will generate the tools and insights needed to help ensure that health care resources are devoted to patients likely to benefit the most from them.