Development and Deployment of Web Application Using Machine Learning for Predicting Intraoperative Transfusions in Neurosurgical Operations

Abstract

Background and Aim: Preoperative blood product preparation is a common practice in neurosurgical patients. However, over-requesting of blood is common and leads to the wastage of blood bank resources. Machine learning (ML) is currently one of the novel computational data analysis methods for assisting neurosurgeons in their decision-making process. The objective of the present study was to use machine learning to predict intraoperative packed red cell transfusion. Additionally, a secondary objective focused on estimating the effectiveness of blood utilization in neurosurgical operations. Methods and Materials/Patients: This was a retrospective cohort study of 3,021 patients who had previously undergone neurosurgical operations. Data from the total cohort were randomly divided into a training dataset (N=2115) and a testing dataset (N=906). The supervised ML models of various algorithms were trained and tested with test data using both classification and regression algorithms. Results: Almost all neurosurgical conditions had a cross-match to transfusion ratio of more than 2.5. Support vector machine (SVM) with linear kernel, SVM radial kernel, and random forest (RF) classification had a performance with good AUC of 0.83,0.82, and 0.82, respectively, while RF regression had the lowest root mean squared error and mean absolute error. Conclusion: In almost all neurosurgical surgeries, preoperative overpreparation of blood products was detected. The ML algorithm was proposed as a high-performance method for optimizing blood preparation and intraoperative consumption. Furthermore, ML has the potential to be incorporated into clinical practice as a calculator for the optimal cross-match to transfusion ratio.