Prediction of perioperative transfusions using an artificial neural network

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine
• Surgical Outcomes Research

Background:

• Accurate prediction of operative transfusions is critical for efficient resource allocation.
• Identifying patients at risk of postoperative adverse events is essential for patient safety.
• This study focuses on predicting perioperative transfusions for all inpatient operations.

Purpose of the Study:

• To evaluate the efficacy of artificial neural networks (ANNs) in predicting operative transfusions.
• To compare ANN model performance against traditional logistic regression models.
• To develop a robust model for transfusion prediction across diverse surgical procedures.

Main Methods:

• Utilized a large dataset of over 1.6 million surgical cases from the NSQIP-PUF database (2014 for development, 2015 for validation).
• Developed and compared various artificial neural network (ANN) models against logistic regression models.
• Employed four distinct variable sets for model development and analysis.

Main Results:

• The best-performing ANN models significantly outperformed logistic regression models (p < .001) in predicting perioperative transfusions.
• ANN models achieved high performance metrics, including 70-80% specificity and 62-75% sensitivity.
• ANNs successfully predicted over 75% of patients requiring transfusion and 70% of those not requiring transfusion.

Conclusions:

• Artificial neural networks (ANNs) demonstrate superior efficacy in predicting the need for operative transfusions.
• A single ANN model can effectively predict transfusions across a wide spectrum of surgical procedures.
• The findings support the use of ANNs for improved patient risk stratification and resource management in surgery.