A deep learning solution to recommend laboratory reduction strategies in ICU

• Published in: International journal of medical informatics (Shannon, Ireland) Vol. 144; p. 104282
• Main Authors: Yu, Lishan, Li, Linda, Bernstam, Elmer, Jiang, Xiaoqian
• Format: Journal Article
• Language: English
• Published: Ireland Elsevier B.V 01.12.2020
• Subjects: Deep Learning; Dynamic decision-making problem; Humans; Intensive Care Units; Laboratories; Laboratory test reduction; Machine Learning; ROC Curve; Deep learning; Dynamic decision-making problem; Laboratory test reduction
• ISSN: 1386-5056; 1872-8243
• DOI: 10.1016/j.ijmedinf.2020.104282

•Achieving > 98% accuracy on the abnormality prediction on a nearly 20% recommended reduction of lab tests.•A joint consideration of the optimal reduction strategy with long-term prediction (rather than short-term prediction and greedy reduction strategy).•Online adjustable model to accommodate expert decisions and change recommendations in a dynamic manner. To build a machine-learning model that predicts laboratory test results and provides a promising lab test reduction strategy, using spatial-temporal correlations. We developed a global prediction model to treat laboratory testing as a series of decisions by considering contextual information over time and across modalities. We validated our method using a critical care database (MIMIC III), which includes 4,570,709 observations of 12 standard laboratory tests, among 38,773 critical care patients. Our deep-learning model made real-time laboratory reduction recommendations and predicted the properties of lab tests, including values, normal/abnormal (whether labs were within the normal range) and transition (normal to abnormal or abnormal to normal from the latest lab test). We reported area under the receiver operating characteristic curve (AUC) for predicting normal/abnormal, evaluated accuracy and absolute bias on prediction vs. observation against lab test reduction proportion. We compared our model against baseline models and analyzed the impact of variations on the recommended reduction strategy. Our best model offered a 20.26 % reduction in the number of laboratory tests. By applying the recommended reduction policy on the hold-out dataset (7755 patients), our model predicted normality/abnormality of laboratory tests with a 98.27 % accuracy (AUC, 0.9885; sensitivity, 97.84 %; specificity, 98.80 %; PPV, 99.01 %; NPV, 97.39 %) on 20.26 % reduced lab tests, and recommended 98.10 % of transitions to be checked. Our model performed better than the greedy models, and the recommended reduction strategy was robust. Strong spatial and temporal correlations between laboratory tests can be used to optimize policies for reducing laboratory tests throughout the hospital course. Our method allows for iterative predictions and provides a superior solution for the dynamic decision-making laboratory reduction problem. This work demonstrates a machine-learning model that assists physicians in determining which laboratory tests may be omitted.