Medical Analytics in the Presence of Human-Error. An Exploration of EMR Data Quality using MIMIC-III. (Preprint)
BACKGROUND Public Electronic Medical Records (EMR) datasets are a goldmine for vendors and researchers seeking to develop analytics designed to assist caregivers in monitoring, diagnosis, and treatment of patients. Both complex machine-learning-based tools, which require copious amounts of data to train, and a simple trend graph presented in a patient-centered dashboard, are sensitive to noise. OBJECTIVE We aim to systematically explore data errors in MIMIC-III as a representative of secondary use datasets and the impact of these errors on downstream analytics. METHODS We discuss the unique challenge of accounting for the specific patient's medical condition and personal characteristics such as age, weight, gender, and others, in identifying data errors when only a few measurements of each patient are available. To do so, we examine the prevalence and manifestations of errors in one of the most popular public medical research databases - MIMIC-III. We then evaluate how these errors impact visual analytics, score-based sepsis analytics SOFA and qSOFA, and a machine-learning-based sepsis predictor. RESULTS We find a variety of error patterns in MIMIC-III and highlight effective methods to find them. All analytics are found to be sensitive to sporadic error. Visual analytics are severely impacted, limiting their usefulness in the presence of error. qSOFA and SOFA suffer a score change of +1 (of 3) and +2.3-4 (of 15). The sepsis predictor suffers from a 0.01-0.3 score change compared to a median score of 0.08. CONCLUSIONS The use of statistical methods to detect data errors is limited to high-throughput scenarios and large data aggregations. There is a dearth of medical guidelines and error-detection practices to support rule-based systems, required to keep analytics safe and trustworthy in low-volume scenarios. Analytics developers should test their software’s sensitivity to error on public datasets. The medical informatics community should improve support for medical data-quality endeavors by creating guidelines for plausible values and analytics robustness to error and collecting real-world dirty datasets which contain errors as they appear in normal EMR use.