Most "close calls" of medication errors are not analyzed because most organizations manually review mistakes that lead to patient harm. Researchers at Columbia, Md.-based MedStar Health used machine learning to fix this lack of insight.
Two researchers from MedStar and one at Georgetown University School of Medicine in Washington, D.C., evaluated 3,861 patient safety events from a 10-hospital system with logistic regression, elastic net and software library XGBoost.
They found XGBoost's semi-automated machine performed best, and there are multiple opportunities to increase patient safety with this "mining" technique, according to their study, which Nature published Oct. 26.
Natural language processing and machine learning algorithms can help investigate "near misses," which often aren't examined since most manual review time is allocated to when patients are harmed. Implementing this strategy can also save time, "reduce inappropriate classifications and the labor-intensive recoding of reports," and be applied to non-medication reports, the researchers concluded.