Integration of Feature Vectors from Raw Laboratory, Medication and Procedure Names Improves the Precision and Recall of Models to Predict Postoperative Mortality and Acute Kidney Injury

Introduction: Manuscripts that have successfully used machine learning (ML) to predict a variety of perioperative outcomes often use only a limited number of features selected by a clinician. We hypothesized that techniques leveraging a broad set of features for patient laboratory results, medications, and the surgical procedure name would improve performance as compared to a more limited set of features chosen by clinicians. Methods Feature vectors for laboratory results included 702 features total derived from 39 laboratory tests, medications consisted of a binary flag for 126 commonly used medications, procedure name used the Word2Vec package for create a vector of length 100. Nine models were trained: Baseline Features, one for each of the three types of data Baseline+Each data type (, all features, and then all features with feature reduction algorithm. Results Across both outcomes the models that contained all features (model 8) (Mortality ROC-AUC 94.42, PR-AUC 31.0; AKI ROC-AUC 92.47, PR-AUC 76.73) was superior to models with only subsets of features Conclusion Featurization techniques leveraging a broad away of clinical data can improve performance of perioperative prediction models.