Predicting post-stroke pneumonia using deep neural network approaches.

Area of Science:

• Neurology
• Medical Informatics
• Artificial Intelligence

Background:

• Pneumonia is a frequent and serious complication following stroke.
• Early and accurate prediction of post-stroke pneumonia is crucial for improved patient outcomes and reduced hospital stays.
• Traditional risk scores often rely on basic statistical methods, necessitating more advanced approaches.

Purpose of the Study:

• To investigate the efficacy of advanced machine learning algorithms, particularly deep learning, for predicting pneumonia in stroke patients.
• To compare the performance of deep learning models against traditional methods and existing risk scores.
• To leverage electronic health record (EHR) data for developing robust predictive models.

Main Methods:

• Utilized EHR data from 13,930 acute ischemic stroke patients (2007-2017) for model development and validation.
• Implemented and compared various machine learning techniques, including logistic regression, SVMs, XGBoost, MLPs, and attention-augmented GRUs.
• Developed prediction models for pneumonia occurring within 7 days (stroke-associated pneumonia, SAP) and 14 days post-stroke.

Main Results:

• The attention-augmented GRU model demonstrated superior performance, achieving an AUC of 0.928 for 7-day and 0.905 for 14-day pneumonia prediction.
• This deep learning model outperformed other machine learning methods and previous risk score models.
• At 90% sensitivity, the GRU model achieved high specificity (0.85 for 7-day, 0.82 for 14-day) and NPV (0.99 for both windows).

Conclusions:

• Deep learning-based predictive models are effective and feasible for managing stroke patients.
• The attention-augmented GRU model offers optimal performance for predicting post-stroke pneumonia.
• These advanced models can significantly enhance clinical decision-making and preventative care for stroke patients.