A Novel Approach for Handling Missing Data in Multivariate Time Series Clustering: Case Study on Predicting Delayed Graft Function

Single value imputation and Multiple Imputation Chained Equation (MICE) are widely used for handling missing data cross-sectionally. Although several predictive and generative deep learning based methods address missing multivariate time series (MTS) data under Missing at Random /Missing Completely at Random types of missingness, their performance banks on the interpretation of missingness. Missing data is common in electronic health record (EHR) data and typically falls under Missing Not at Random, for which few suitable imputation methods exist. We propose the Phasing and List-Splitting (PALS) technique, a novel approach to handling irregularly spaced missing data, designed for applications associated with MTS Clustering, where the objective is to group data points based on their observable trajectories. With PALS, the priority is to minimize the amount of imputation while maintaining temporal alignment. We compared PALS with existing imputation methods and evaluated its performance based on imputation accuracy and on the quality of clustering obtained from a dataset of patient trajectories with delayed graft function following kidney transplants.