Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Collaborative learning of common latent representations in routinely collected multivariate ICU physiological signals

Hollan Haule, Ian Piper, Patricia Jones, Tsz-Yan Milly Lo, Javier Escudero

TL;DR

The study tackles ICU phenotyping from routinely collected multivariate time series by learning common latent representations across patients using a collaborative-filtering framework coupled with LSTM embeddings. It introduces a loss that aligns cross-patient latent states via a similarity target and employs collaborative inference by least squares to predict IH-related states, evaluated on a pediatric TBI dataset. Results show strong IH detection performance (IH AUC up to $0.889$ and AP up to $0.725$) and more compact, discriminative latent spaces compared to autoencoder baselines, though cross-channel attention may detract in this setup. The approach demonstrates the feasibility of deriving clinically meaningful, shared representations from high-dimensional ICU signals, with potential to inform patient subgroups and personalized care.

Abstract

In Intensive Care Units (ICU), the abundance of multivariate time series presents an opportunity for machine learning (ML) to enhance patient phenotyping. In contrast to previous research focused on electronic health records (EHR), here we propose an ML approach for phenotyping using routinely collected physiological time series data. Our new algorithm integrates Long Short-Term Memory (LSTM) networks with collaborative filtering concepts to identify common physiological states across patients. Tested on real-world ICU clinical data for intracranial hypertension (IH) detection in patients with brain injury, our method achieved an area under the curve (AUC) of 0.889 and average precision (AP) of 0.725. Moreover, our algorithm outperforms autoencoders in learning more structured latent representations of the physiological signals. These findings highlight the promise of our methodology for patient phenotyping, leveraging routinely collected multivariate time series to improve clinical care practices.

Collaborative learning of common latent representations in routinely collected multivariate ICU physiological signals

TL;DR

The study tackles ICU phenotyping from routinely collected multivariate time series by learning common latent representations across patients using a collaborative-filtering framework coupled with LSTM embeddings. It introduces a loss that aligns cross-patient latent states via a similarity target and employs collaborative inference by least squares to predict IH-related states, evaluated on a pediatric TBI dataset. Results show strong IH detection performance (IH AUC up to and AP up to ) and more compact, discriminative latent spaces compared to autoencoder baselines, though cross-channel attention may detract in this setup. The approach demonstrates the feasibility of deriving clinically meaningful, shared representations from high-dimensional ICU signals, with potential to inform patient subgroups and personalized care.

Abstract

In Intensive Care Units (ICU), the abundance of multivariate time series presents an opportunity for machine learning (ML) to enhance patient phenotyping. In contrast to previous research focused on electronic health records (EHR), here we propose an ML approach for phenotyping using routinely collected physiological time series data. Our new algorithm integrates Long Short-Term Memory (LSTM) networks with collaborative filtering concepts to identify common physiological states across patients. Tested on real-world ICU clinical data for intracranial hypertension (IH) detection in patients with brain injury, our method achieved an area under the curve (AUC) of 0.889 and average precision (AP) of 0.725. Moreover, our algorithm outperforms autoencoders in learning more structured latent representations of the physiological signals. These findings highlight the promise of our methodology for patient phenotyping, leveraging routinely collected multivariate time series to improve clinical care practices.
Paper Structure (15 sections, 9 equations, 2 figures, 1 table, 2 algorithms)

This paper contains 15 sections, 9 equations, 2 figures, 1 table, 2 algorithms.

Table of Contents

  1. Introduction
  2. Methodology
  3. Core principle and loss function formulation
  4. Training algorithm
  5. Collaborative Inference by Least Squares
  6. LSTM Model for Extracting Embeddings
  7. Self-Attention and Cross-channel attention
  8. Dataset description
  9. Preprocessing
  10. Experiments
  11. Results and discussion
  12. Classification performance
  13. Cross-channel attention ablation
  14. Visualising embeddings with t-SNE
  15. Conclusion

Figures (2)

  • Figure 1: Distribution of correlation values between the presence of patients in the training set and average AP. This shows that the presence of some patients in the training set strongly affects performance.
  • Figure 2: Comparative t-SNE 2D projections of patients in a test set.