Combining patient visual timelines with deep learning to predict mortality.

Area of Science:

• Artificial Intelligence
• Clinical Informatics
• Biomedical Data Science

Background:

• Electronic health records contain valuable clinical data, primarily in non-image formats.
• Deep learning models excel at image recognition but require data transformation for clinical applications.
• Visualizing longitudinal patient data can bridge the gap between clinical data and advanced predictive modeling.

Purpose of the Study:

• To develop a framework for transforming longitudinal patient data into visual timelines.
• To apply deep learning models to these visual timelines for predicting in-hospital mortality.
• To enhance clinician interpretability of predictive models.

Main Methods:

• Retrospective analysis of adult patient admissions from 2008-2016.
• Creation of 2D visual timelines using clinical variables and time as dimensions.
• Utilized a convolutional neural network with a recurrent layer for mortality prediction.
• Model derivation on 70% of the cohort and independent validation on 30%.

Main Results:

• The deep learning model achieved an AUC of 0.91 for predicting in-hospital mortality.
• Outperformed the Modified Early Warning Score (AUC: 0.76) and SOFA score (AUC: 0.57).
• Class-activation heatmaps provided insights into prediction drivers, aiding interpretability.

Conclusions:

• A novel framework successfully converts longitudinal patient data into visual timelines for deep learning.
• The deep learning model demonstrates superior accuracy in predicting in-hospital mortality compared to established scores.
• This approach offers a promising method for predicting clinical outcomes and improving interpretability.