A customizable deep learning model for nosocomial risk prediction from critical care notes with indirect supervision

Area of Science:

• Clinical informatics
• Artificial intelligence in healthcare
• Predictive modeling

Background:

• Longitudinal risk prediction for hospitalized patients is crucial for quality care.
• Current methods often rely on structured data, potentially missing insights from clinical notes.

Purpose of the Study:

• To develop a generalizable model leveraging clinical notes for early prediction of healthcare-associated diseases.
• To predict the risk of hospital-acquired acute kidney injury, pressure injury, or anemia 24-96 hours in advance.

Main Methods:

• Developed a reCurrent Additive Network for Temporal RIsk Prediction (CANTRIP).
• Utilized the MIMIC III critical care database, extracting positive and negative cohorts for specific diseases.
• Employed indirect supervision using retrospectively determined event dates from structured and unstructured data to train CANTRIP on clinical notes.

Main Results:

• CANTRIP, using only text, achieved 74%-87% area under the curve and 77%-85% specificity.
• Outperformed baseline shallow models across all metrics.
• Bidirectional long short-term memory models showed higher sensitivity but lower specificity and precision.

Conclusions:

• Clinical text alone can be effectively used to predict nosocomial diseases.
• CANTRIP generalizes across diseases without disease-specific feature extraction, matching or exceeding existing models.
• The CANTRIP model offers a competitive alternative to traditional structured features for risk prediction.