Two-step interpretable modeling of ICU-AIs.

Area of Science:

• Biostatistics
• Medical Informatics
• Machine Learning in Healthcare

Background:

• Accurate prediction of patient outcomes is crucial in intensive care units (ICUs).
• Traditional survival models often lack interpretability or struggle with high-resolution longitudinal data.
• Artificial neural networks offer predictive power but often function as 'black boxes'.

Purpose of the Study:

• To develop a novel methodology integrating high-resolution longitudinal data with survival models for improved predictive power and interpretability.
• To move beyond 'black box' models by proposing a parsimonious and robust semi-parametric approach.
• To enhance dynamic prediction capabilities for clinical events, specifically healthcare-associated infections in ICUs.

Main Methods:

• A parsimonious and robust semi-parametric approach, specifically a landmarking competing risks model.
• Integration of routinely collected low-resolution data with predictive features from a convolutional neural network (CNN).
• CNN trained on high-resolution time-dependent information, with saliency maps used for model explanation.

Main Results:

• The proposed methodology successfully integrates high-resolution data into interpretable survival models.
• The combined approach demonstrated enhanced predictive power compared to traditional methods.
• Saliency maps provided insights into the contribution of CNN-extracted features to predictive accuracy.

Conclusions:

• The novel methodology offers a powerful tool for dynamic prediction in clinical settings, balancing accuracy and interpretability.
• This approach can improve the prediction of healthcare-associated infections in ICU patients.
• The use of saliency maps facilitates understanding and trust in complex predictive models.