»AI based triage - predicting late and early mortality after emergency department visit.«
2019-06-03, 16:30–16:55, Hall 2 (PyData)
In this talk I will demonstrate how AI outperforms traditional triage measures for predicting early and late mortality after emergency department (ED) visit, using EMR data of ~56K ED visits of adults patients over a period of 5 years.
Emergency departments (ED) are becoming more often overwhelmed increasing the poor outcomes of ED overcrowding. Triage scores aim to optimize the waiting time and prioritize the resource usage according to the severity of the medical condition. However, the most widely used triage scores, relies heavily on provider judgment which can lead to inaccuracy and misclassification. Artificial intelligence (AI) algorithms offer advantages for creating predictive clinical applications because of flexibility in handling large datasets from electronic medical records (EMR), and are becoming better at prediction tasks, often outperforming current clinical scoring systems.
In this talk I will present a collaborative study between Intuit’s AI-ML researchers and Shiba hospital innovation center. In this study we used EMR data to predict mortality in the at early triage and ED (emergency department) visit level. The study included 559,353 ED visits of adults patients over five years (2012-2017). Variables included: demographics, admission date, arrival mode, referral code, chief complaint, previous ED visits, previous hospitalizations, background diagnoses, regular drugs, vital Signs and ESI score. Using XGboost, Catboost and deep-leaning we yielded an AUC of 0.97 for early mortality, an AUC of 0.93 for short term mortality, and an AUC of 0.93 for long term mortality (90 days from admission). Single variable analysis shows that the two variables with the highest AUC were: age and arrival mode for early mortality; age and main cause for short term mortality; and are age and number of drugs for long term mortality. The highest information gain variables for early mortality (calculated using XGBoost with a 1000 trees) were SBP, HR, days to most recent previous ED visit and fever. The results outperform e