Unsupervised learning approach for predicting sepsis onset in ICU patients.

Area of Science:

• Critical Care Medicine
• Machine Learning
• Biomedical Informatics

Background:

• Sepsis is a life-threatening condition with poor outcomes if not diagnosed early.
• Current sepsis prediction relies on supervised models needing extensive labeled data.
• Septic shock onset requires timely prediction for improved patient management in ICUs.

Purpose of the Study:

• To develop and evaluate fully unsupervised learning approaches for predicting septic shock onset.
• To leverage Recurrent Autoencoders for multivariate time-series representation learning.
• To apply anomaly detection on learned representations for early septic shock detection.

Main Methods:

• Utilized Recurrent Autoencoders (including Variational Autoencoder - VAE) for unsupervised representation learning from patient time-series data.
• Implemented clustering-based anomaly detection algorithms on the learned feature space.
• Compared unsupervised VAE with Gaussian Mixture Models against a supervised LSTM network.

Main Results:

• The unsupervised VAE with Gaussian Mixture Models achieved an AUC of 0.82 and F1-score of 0.65.
• Performance was competitive with a supervised LSTM network (AUC 0.80, F1-score 0.66).
• Demonstrated the efficacy of unsupervised learning for septic shock prediction.

Conclusions:

• Unsupervised learning offers a viable alternative for septic shock onset prediction, reducing reliance on labeled data.
• The proposed framework can enhance current infection progression monitoring in the ICU.
• This approach holds promise for earlier and more effective intervention in septic patients.