An Artificial Neural Network-Based Pediatric Mortality Risk Score: Development and Performance Evaluation Using Data...

Area of Science:

• Pediatric Intensive Care
• Machine Learning in Medicine
• Biostatistics

Background:

• Quantifying pediatric intensive care unit (PICU) illness severity aids timely intervention.
• Existing risk models like PIM-2 and PRISM-III use routine admission data.
• Artificial neural networks (ANNs) show potential for improved performance over regression models in healthcare.

Purpose of the Study:

• To compare the performance of ANNs against logistic regression models for pediatric mortality risk estimation in the PICU.
• To evaluate ANN models using data comparable to PIM-2 and PRISM-III.
• To assess the utility of ANNs in predicting mortality for critically ill children.

Main Methods:

• Utilized a dataset of 102,945 patients from North American PICUs with infection-related diagnoses.
• Developed a 2-layer ANN model and a logistic regression model using PIM-2 and PRISM-III features.
• Preprocessed data by imputing missing values and normalizing variables.
• Compared model performance using Area Under the Receiver Operating Characteristic Curve (AUROC) and Area Under the Precision-Recall Curve (AUPRC).

Main Results:

• The best-performing ANN (normalized data) achieved an AUROC of 0.871 and AUPRC of 0.372.
• This ANN outperformed PIM-2 (AUROC 0.805, AUPRC 0.234) and PRISM-III (AUROC 0.844, AUPRC 0.348).
• The ANN also showed superior performance compared to the logistic regression model (AUROC 0.862, AUPRC 0.329).

Conclusions:

• A simple 2-layer ANN model demonstrated slightly better performance than benchmark models (PIM-2, PRISM-III) and logistic regression.
• The observed performance gains may not yet translate to clinically significant improvements.
• Further investigation into advanced ANN architectures and data imputation techniques is warranted.