Predicting risk for trauma patients using static and dynamic information from the MIMIC III database

• Published in: PloS one Vol. 17; no. 1; p. e0262523
• Main Authors: Tsiklidis, Evan J, Sinno, Talid, Diamond, Scott L
• Format: Journal Article
• Language: English
• Published: United States Public Library of Science 19.01.2022; Public Library of Science (PLoS)
• Subjects: Adult; Age; Aged; Algorithms; Blood pressure; Clinical outcomes; Computer and Information Sciences; Critically ill; Data Management - methods; Databases, Factual; Datasets; Electronic Health Records; Electronic medical records; Electronic records; Engineering; Female; Gender; Health aspects; Health risks; Heart rate; Hospital Mortality - trends; Hospital patients; Hospitalization - statistics & numerical data; Hospitalization - trends; Humans; Injury Severity Score; Intensive care; Intensive Care Units - statistics & numerical data; Machine Learning; Male; Medical records; Medicine and Health Sciences; Middle Aged; Mortality; Ordinary differential equations; Patients; Physical Sciences; Prediction models; Probability; Prognosis; Research and Analysis Methods; Retrospective Studies; Risk Assessment - methods; Risk Factors; Risk groups; Risk management; ROC Curve; Standard deviation; Survival; Training; Trauma; Variables; Windows (intervals); Wounds and injuries; United States; United States > US
• ISSN: 1932-6203; 1932-6203
• DOI: 10.1371/journal.pone.0262523

Risk quantification algorithms in the ICU can provide (1) an early alert to the clinician that a patient is at extreme risk and (2) help manage limited resources efficiently or remotely. With electronic health records, large data sets allow the training of predictive models to quantify patient risk. A gradient boosting classifier was trained to predict high-risk and low-risk trauma patients, where patients were labeled high-risk if they expired within the next 10 hours or within the last 10% of their ICU stay duration. The MIMIC-III database was filtered to extract 5,400 trauma patient records (526 non-survivors) each of which contained 5 static variables (age, gender, etc.) and 28 dynamic variables (e.g., vital signs and metabolic panel). Training data was also extracted from the dynamic variables using a 3-hour moving time window whereby each window was treated as a unique patient-time fragment. We extracted the mean, standard deviation, and skew from each of these 3-hour fragments and included them as inputs for training. Additionally, a survival metric upon admission was calculated for each patient using a previously developed National Trauma Data Bank (NTDB)-trained gradient booster model. The final model was able to distinguish between high-risk and low-risk patients to an AUROC of 92.9%, defined as the area under the receiver operator characteristic curve. Importantly, the dynamic survival probability plots for patients who die appear considerably different from those who survive, an example of reducing the high dimensionality of the patient record to a single trauma trajectory.