LAUSR

We are well over a decade into the machine-learning revolution, but medicine, particularly laboratory medicine, has resisted. This is not for want of opportunity. Laboratory testing is the single highest volume medical activity, with 10 billion tests being performed in the US each year (1). Laboratory testing boasts the lowest error rates in medicine, at least in terms of analytical error, with well-defined reference ranges and regular quality control inspiring confidence in the results (2, 3). Results are almost all either numerical or categorical, perfect for machines. Testing not only informs diagnosis across medicine but also is inexpensive—at pennies on the healthcare dollar, vastly cheaper than the downstream care it guides—representing a financial, as well as medical, incentive for building a future based on systematic machine learning-based clinical decision support. In this issue of Clinical Chemistry , Wilkes and colleagues offer a glimpse of this future by using machine learning to predict clinical diagnoses using urine steroid profiles (4). Urine steroid profiles are used to diagnose endocrine conditions such as adrenal cancer and Cushing disease. A profile consists of the urine concentrations of steroid metabolites—aldosterone, cortisol, and many others—measured most often by GC-MS. To make a diagnosis, laboratorians or clinicians must integrate the information in these results and interpret it in the context of patient demographics such as age and sex. They do so by applying heuristics or rules. Sometimes the rules are explicit, something a person can describe. But often with human experts, decision-making becomes so automatic, with rules so complex …