Since the American Recovery and Reinvestment Act of 2009 incentivized the adoption and use of electronic health records (EHRs), EHRs have become ubiquitous in the health care industry. Recent federal reports show about 84% adoption in hospitals and about 86% adoption in office-based practices. Patient information that was once captured on paper is now being regularly recorded and stored in EHRs, creating new opportunities for analyzing and drawing insights from these increasingly rich data sets. While EHRs can support easier access to and sharing of information by individual providers and patients, larger efforts (e.g. Why is EHR data complex?

Imagine knowing, in real time, whether a patient will suffer a surgical infection as a surgeon closes up a wound. That's the kind of clinical situation that machine learning is enabling at the University of Iowa Hospitals & Clinics. To date, the health system's innovation with AI analytics has led to a 74 percent reduction in surgical site infection over a three-year period, which at scale is a $1.2 million cost savings – not including savings from value-based purchasing because of the reduced surgical site infection rate. Iowa's work with comes as more and more hospitals and tech vendors are undertaking innovative initiatives with machine learning and artificial intelligence. Johns Hopkins for instance, is using deep learning to improve how it handles pancreatic cancer and Amazon Web Services is harnessing machine learning to enable customers to better treat depression.

A large fraction of the electronic health records consists of clinical measurements collected over time, such as blood tests, which provide important information about the health status of a patient. These sequences of clinical measurements are naturally represented as time series, characterized by multiple variables and the presence of missing data, which complicate analysis. In this work, we propose a surgical site infection detection framework for patients undergoing colorectal cancer surgery that is completely unsupervised, hence alleviating the problem of getting access to labelled training data. The framework is based on powerful kernels for multivariate time series that account for missing data when computing similarities. Our approach show superior performance compared to baselines that have to resort to imputation techniques and performs comparable to a supervised classification baseline.