|
Activity Number:
|
288
- SLDS CSpeed 5
|
|
Type:
|
Contributed
|
|
Date/Time:
|
Wednesday, August 11, 2021 : 1:30 PM to 3:20 PM
|
|
Sponsor:
|
Section on Statistical Learning and Data Science
|
|
Abstract #318011
|
|
|
Title:
|
Longitudinal Cluster Analysis Using Segmented-LSTM with Applications of Weight Loss Trajectories Following RYGB Surgery
|
|
Author(s):
|
Yirui Hu* and Simo Wu and Kunpeng Liu and Michelle R. Lent and Peter N. Benotti and Anthony T. Petrick and G. Craig Wood and Christopher D. Still and H. Lester Kirchner
|
|
Companies:
|
Geisinger and Facebook and University of Central Florida and Philadelphia College of Osteopathic Medicine and Geisinger and Geisinger and Geisinger and Geisinger and Geisinger
|
|
Keywords:
|
Recurrent Neural Network (RNN);
Long Short-Term Memory (LSTM);
Roux-en-Y gastric bypass (RYGB) bariatric surgery ;
Longitudinal Cluster Analysis ;
LSTM autoencoder
|
|
Abstract:
|
Background: Identifying subgroups of bariatric surgery patients with similar post-surgery weight loss may be helpful for treatment targeting. Recurrent neural networks are powerful to learn the sequential pattern in longitudinal data. Methods: A novel segmented-LSTM autoencoder based clustering architecture was proposed to leverage the phased longitudinal data and identify group-based weight trajectories for patients after RYGB surgery. Silhouette scores were estimated to determine the optimal number of clusters. Results: 2,918 RYGB patients between 2004-2016 were included. Phased longitudinal patterns were observed, where the majority of weight loss occurred within 1st year post-surgery. We compared the clustering results from K-Shape algorithm, Gaussian Mixture Model based on LSTM and segmented-LSTM representations. Silhouette scores suggested that four clusters achieved best quality of clusters across algorithms. Specifically, the Silhouette score was 0.22 from K-Shape algorithm; 0.35 from LSTM; 0.42 from segmented-LSTM. Conclusion: Segmented-LSTM successfully learns the heterogeneous sequential patterns of the phased longitudinal data to inform the variability in treatment.
|
Authors who are presenting talks have a * after their name.
