An Ensemble Method for Data Imputation

Healthcare analytics is transforming the healthcare industry, finding novel and useful patterns in patient data such as electronic health records (EHRs), to provide patients with improved care and service. Researchers train machine learning (ML) algorithms to discover new knowledge by mining patients' clinical data to provide better care such as accurate diagnoses and personalized therapy. However, the quality of clinical patient data may inhibit the discovery process. The type and frequency of collected data varies based on a patient' s clinical condition and administrative requirements. Patients can have different diagnostic tests and treatments at different times, even with the same symptoms. Therefore in EHRs, many aspects of a patient' s clinical condition could be unmeasured at different timestamps. Missing measurements may be clinically important, but cannot be used by ML algorithms. To utilize all clinical data and achieve optimal performance of ML algorithms, we address the missing data issue by imputing missing time series values. We will describe our imputation methods and apply them to 13 common clinical laboratory test results obtained from a set of 8267 inpatients to evaluate their performance using normalized root-mean-squared deviation (nRMSD) [1].