Predicting malnutrition from longitudinal patient trajectories with deep learning.

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Clinical Nutrition

Background:

• Malnutrition is a prevalent condition often diagnosed late, leading to adverse clinical, functional, and economic consequences.
• Improved screening and prediction methods are essential for timely intervention and better patient outcomes.

Purpose of the Study:

• To assess the predictability of malnutrition using longitudinal patient records.
• To evaluate the external generalizability of predictive models across different states.

Main Methods:

• Developed and validated predictive models using longitudinal patient data from emergency department and hospital admission databases in California, Florida, and New York (2015-2018).
• Employed long short-term memory (LSTM) recurrent neural networks for longitudinal data and gradient-boosted tree/logistic regression for cross-sectional data.
• Utilized Area Under the Receiver-Operating Characteristic (AUROC) and Precision-Recall (AUPRC) curves to evaluate prediction accuracy for first malnutrition diagnoses.

Main Results:

• The longitudinal LSTM model demonstrated high and comparable predictive performance across California, Florida, and New York (AUROC range: 0.854-0.869).
• Malnutrition diagnoses ranged from 4.0% to 6.2% across the studied states.
• Deep learning models achieved superior predictive accuracy with lower data requirements compared to existing instruments.

Conclusions:

• Deep learning models can reliably predict malnutrition from existing longitudinal patient records.
• This predictive approach facilitates early nutritional intervention through automated screening at the point of care.
• The findings support the integration of AI-driven tools for proactive malnutrition management in healthcare settings.