Temporal Supervised Contrastive Learning for Modeling Patient Risk Progression.

Area of Science:

• Machine Learning
• Medical Informatics
• Time Series Analysis

Background:

• Predicting patient outcomes over time is crucial in healthcare.
• Existing methods struggle with the dynamic nature of patient data.
• Clinical data presents unique challenges for standard machine learning techniques.

Purpose of the Study:

• To develop a novel supervised contrastive learning framework for patient time series.
• To learn embedding representations that capture temporal dependencies and outcome probabilities.
• To improve the accuracy of predicting patient outcomes and tracking disease progression.

Main Methods:

• Proposed a supervised contrastive learning framework for patient time series.
• Learned embedding representations where similar outcomes are close in space.
• Employed a nearest neighbor pairing mechanism in raw feature space as an alternative to data augmentation.
• Validated the approach on predicting mortality in septic patients (MIMIC-III) and cognitive impairment progression (ADNI).

Main Results:

• Outperformed state-of-the-art baselines in predicting mortality and cognitive impairment progression.
• Successfully recovered synthetic dataset embedding structures, a feat not achieved by baselines.
• Demonstrated the critical role of the nearest neighbor pairing mechanism through ablation studies.

Conclusions:

• The proposed framework effectively learns patient time series embeddings for outcome prediction.
• Nearest neighbor pairing is a vital component for handling clinical tabular data in contrastive learning.
• This approach offers a promising direction for dynamic patient outcome prediction and disease tracking.