Contrastive Transfer Learning for Prediction of Adverse Events in Hospitalized Patients

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Predictive Analytics

Background:

• Hospitalized patients face risks of adverse events.
• The Deterioration Index (DI) is a composite score reflecting patient condition.
• Early prediction of adverse events is crucial for timely intervention.

Purpose of the Study:

• To develop and validate a contrastive transfer learning (CL) method for early prediction of adverse events.
• To utilize Deterioration Index (DI) scores as a primary data source.
• To compare the proposed method against supervised deep learning models.

Main Methods:

• An unsupervised contrastive learning (CL) model with a classifier was developed.
• The model was pretrained on large-scale time series data and fine-tuned with DI scores.
• Performance was evaluated against supervised deep learning models for time series classification.

Main Results:

• The proposed unsupervised contrastive transfer learning model outperformed supervised deep learning solutions.
• Pretraining on extensive time series data and fine-tuning with DI scores improved prediction accuracy.
• A significant relationship was observed between longitudinal DI scores and patient outcomes.

Conclusions:

• Unsupervised contrastive transfer learning using DI scores can effectively predict and prevent adverse outcomes.
• The developed algorithm can function as an early warning system in hospitals.
• This technology has the potential to mitigate adverse events and improve patient care.