Unfolding Physiological State: Mortality Modelling in Intensive Care Units.

Area of Science:

• Clinical Informatics
• Natural Language Processing
• Machine Learning

Background:

• Accurate patient outcome prediction is vital for effective clinical care.
• Electronic healthcare records (EHRs) contain vast amounts of data, offering potential for improved predictions.
• Identifying factors influencing patient trajectories can enhance healthcare efficiency and quality.

Purpose of the Study:

• To investigate the utility of latent variable models, specifically Latent Dirichlet Allocation (LDA), for extracting features from free-text hospital notes.
• To evaluate the predictive power of these extracted features for patient mortality across different timeframes.
• To compare the performance of latent topic features against structured EHR data and their combinations.

Main Methods:

• Applied Latent Dirichlet Allocation (LDA) to decompose free-text hospital notes into latent topic features.
• Developed prediction models for patient mortality in three regimes: baseline, dynamic (time-varying), and retrospective.
• Evaluated model performance using Area Under the Curve (AUC) for in-hospital, 30-day, and 1-year post-discharge mortality.

Main Results:

• Latent topic features effectively predicted patient mortality across all evaluated timelines.
• Features predictive of in-hospital mortality differed significantly from those predicting post-discharge mortality.
• Combined latent topic and structured features generally outperformed either feature type alone.
• Dynamic models achieved AUCs of 0.85 (in-hospital), 0.80 (30-day), and 0.77 (1-year).
• Retrospective models achieved AUCs of 0.96 (in-hospital), 0.82 (30-day), and 0.81 (1-year).

Conclusions:

• Latent topic features derived from clinical notes are valuable predictors of patient mortality.
• The combination of latent topic and structured data offers superior predictive performance.
• Dynamic models utilizing these features can support ongoing severity stratification and resource allocation systems.