Welcome to the new version of CaltechAUTHORS. Login is currently restricted to library staff. If you notice any issues, please email coda@library.caltech.edu
Published October 30, 2018 | Supplemental Material + Submitted
Report Open

Preoperative predictions of in-hospital mortality using electronic medical record data

Abstract

Background: Predicting preoperative in-hospital mortality using readily-available electronic medical record (EMR) data can aid clinicians in accurately and rapidly determining surgical risk. While previous work has shown that the American Society of Anesthesiologists (ASA) Physical Status Classification is a useful, though subjective, feature for predicting surgical outcomes, obtaining this classification requires a clinician to review the patient's medical records. Our goal here is to create an improved risk score using electronic medical records and demonstrate its utility in predicting in-hospital mortality without requiring clinician-derived ASA scores. Methods: Data from 49,513 surgical patients were used to train logistic regression, random forest, and gradient boosted tree classifiers for predicting in-hospital mortality. The features used are readily available before surgery from EMR databases. A gradient boosted tree regression model was trained to impute the ASA Physical Status Classification, and this new, imputed score was included as an additional feature to preoperatively predict in-hospital post-surgical mortality. The preoperative risk prediction was then used as an input feature to a deep neural network (DNN), along with intraoperative features, to predict postoperative in-hospital mortality risk. Performance was measured using the area under the receiver operating characteristic (ROC) curve (AUC). Results: We found that the random forest classifier (AUC 0.921, 95%CI 0.908-0.934) outperforms logistic regression (AUC 0.871, 95%CI 0.841-0.900) and gradient boosted trees (AUC 0.897, 95%CI 0.881-0.912) in predicting in-hospital post-surgical mortality. Using logistic regression, the ASA Physical Status Classification score alone had an AUC of 0.865 (95%CI 0.848-0.882). Adding preoperative features to the ASA Physical Status Classification improved the random forest AUC to 0.929 (95%CI 0.915-0.943). Using only automatically obtained preoperative features with no clinician intervention, we found that the random forest model achieved an AUC of 0.921 (95%CI 0.908-0.934). Integrating the preoperative risk prediction into the DNN for postoperative risk prediction results in an AUC of 0.924 (95%CI 0.905-0.941), and with both a preoperative and postoperative risk score for each patient, we were able to show that the mortality risk changes over time. Conclusions: Features easily extracted from EMR data can be used to preoperatively predict the risk of in-hospital post-surgical mortality in a fully automated fashion, with accuracy comparable to models trained on features that require clinical expertise. This preoperative risk score can then be compared to the postoperative risk score to show that the risk changes, and therefore should be monitored longitudinally over time.

Additional Information

The copyright holder for this preprint is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a CC-BY-NC-ND 4.0 International license. bioRxiv preprint first posted online May. 25, 2018. L.O.L. was financially supported by the National Institute of Mental Health under award number K99MH116115. S.S. was supported in part by NIH grants R00GM111744, R35GM125055, NSF Grant III-1705121, an Alfred P. Sloan Research Fellowship, and a gift from the Okawa Foundation. R.B. was funded by a UCLA QCB Collaboratory Postdoctoral Fellowship directed by Matteo Pellegrini. U.M. was financially supported by the Bial Foundation. B.J. was supported by the National Science Foundation Graduate Research Fellowship Program under Grant No. DGE-1650604. M.C. is co-owner of US patent serial no. 61/432,081 for a closed-loop fluid administration system based on the dynamic predictors of fluid responsiveness which has been licensed to Edwards Lifesciences. M.C. is a consultant for Edwards Lifesciences (Irvine, CA), Medtronic (Boulder, CO), Masimo Corp. (Irvine, CA). M.C. has received research support from Edwards Lifesciences through his Department and NIH R01 GM117622 - Machine learning of physiological variables to predict diagnose and treat cardiorespiratory instability and NIH R01 NR013912 - Predicting Patient Instability Non invasively for Nursing Care-Two (PPINNC-2). No funding bodies had any role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

Attached Files

Submitted - 329813.1.full.pdf

Supplemental Material - 329813-1.pdf

Files

329813.1.full.pdf
Files (4.7 MB)
Name Size Download all
md5:ffc58bdc35a10d1b92ccb99e65aaa4cf
938.4 kB Preview Download
md5:fbedba3271c3552ceb695c20f65df656
3.7 MB Preview Download

Additional details

Created:
August 19, 2023
Modified:
October 18, 2023