All Times EDT
Keywords: Clustering, Long Short-Term Memory, Auto-encoder, Trajectory analysis
Background: In precision medicine, it is crucial to identify subgroups of similar patterns from electronic health records data and further develop specific treatment strategies. Clustering can reveal the association and distinction between subjects by maximizing the internal similarity inside each cluster and maximizing the external similarity between different clusters. However, clustering longitudinal data is challenging due to the time dependency between repeated measurements. Methods: This retrospective, observational study utilized data for RYGB patients up to seven years post-surgery. We proposed Long Short-Term Memory (LSTM) auto-encoder based clustering method to identify group-based weight trajectories of Roux-en-Y gastric bypass (RYGB) patients. Results: Data from 2,918 RYGB patients from a comprehensive medical center between January 2004-November 2016 were included. The experimental results showed the superiority of the proposed LSTM in missing data imputation in longitudinal data. Based on LSTM representations, clustering results using K-Means and Gaussian Mixture Models confirmed the existed heterogeneity in the weight change trajectories following RYGB surgery. Conclusions: LSTM is promising in longitudinal clustering with missingness to identify heterogeneous subgroups of patient behaviors. Those findings can help inform patient-provider discussions surrounding postoperative expectations after RYGB and guide clinical care.