Medication Recommendation for Critical Care Patients Using Patient Similarity in Clinical Records

Abstract

This paper presents an effective clinical decision support technique for recommending medications for critical care patients by combining various clinical data of the patients. Specifically, our goal is to infer necessity of medications for intensive care unit (ICU) patients by looking into similar patients' medication. The patient similarity is inferred utilizing their first 24-hour clinical data. Modern ICUs are equipped with numerous monitoring devices, including devices that monitor vital signs such as heart rate, blood pressure, Oxygen saturation and so on. Patients also undergo different pathological (i.e., laboratory) tests, medications and nursing. In this work, we utilize a combination of several of these data for inferring the necessity of prescribing or administering certain medication. We encounter several challenges associated with this task, including heterogeneous sources of features such as lab test results, prescription, fluid intake and fluid output, missing values, class imbalance and high dimensionality. We address some of these challenges by feature selection, missing value imputation and class balancing. We investigate the effectiveness of using patient similarity network and fusion of these networks for the recommendation task. We also study the effectiveness of ensemble classification models, where each model is trained from a specific set of features. When recommending certain medication for a new patient, we first find the most similar patients of the query patient based on the first 24-hour clinical data and recommend or decline based on the medication data of those patients. The proposed technique has been evaluated on a real ICU patients' database and exhibits AUROC score of up to 0.83 for certain prescription drugs.