Authors of a paper published in Health Data Science, a Science Partner Journal, developed an artificial intelligence (AI) model based on machine learning to predict blood transfusion needs for non-traumatic intensive care unit (ICU) patients. Specifically, the researchers at Emory University developed their AI model to predict blood transfusions using a dataset that included 72,072 total patients with 18,314 transfused individuals and 53,758 that were not transfused from 2016 to 2020 at Emory University Hospital. They noted that the transfused cohort was adult patients 18 years of age and older who were not suffering from massive bleeding but had received a blood product. The non-transfused group featured individuals 18 and older who did not receive a transfusion of any blood product.
The paper explained that in addition to their AI model to predict blood transfusions, five distinct other models or algorithms were developed and used to, “predict the probability of necessity for blood transfusions 24 hours in advance during ICU stays.” The researchers described those models as, “logistic regression (LR), random forest (RF), feedforward neural networks (FNNs), support vector machines (SVMs), and XGBoost (XGB). [The AI model was created to] improve the predictive performance of the blood transfusion receipt.” They noted that, “[o]ur primary performance metric was the area under the receiver operating characteristic curve (AUROC) [as models] with a higher AUROC potentially lead to more efficient models in the prediction of blood transfusion by maintaining the balance between specificity and sensitivity metrics.”
The researchers discovered that, “the [AI model] consistently outperform[ed] other models across various scenarios.” They explained that the AI model, “[w]hen evaluated on unseen data from the year 2018 and trained on data from other years, [it] achieves an impressive performance, boasting an AUROC of 0.97, an accuracy rate of 0.93, and an F1 score of 0.89. The main contribution of the AI model can be seen in its ability to maintain high precision while improving recall. That is, it [can] identify a high proportion of the true positive cases it predicts as such, ensuring that the predictions it makes are highly reliable.”
The authors concluded that, “[t]he developed [AI] model demonstrated superior performance across various training scenarios, with a full year’s data utilized for evaluation. The ability to analyze the underlying reasons behind the [AI] model’s decision-making using its base models and patient features offers better communication with healthcare providers and builds trust. By enabling healthcare providers to predict transfusion recipients, [the AI] model can allow for proactive management of patients at risk, potentially improving recovery rates and reducing complications due to delayed transfusions. Additionally, improved predictive capabilities can streamline hospital operations, from optimizing blood supply management to planning staffing and procedural logistics more efficiently.” The researchers also acknowledged that, “our [AI] model needs to be cross-validated with other hospitals for more generalization. Hence, future endeavors will aim to validate extensively and integrate these models into clinical workflows and assess their effectiveness on a broader scale, with the ultimate goal of refining and personalizing care in critical settings.”
Citation: Rafiei, A., Moore, R., and Choudhary, T. et al. “Robust meta-model for predicting the likelihood of receiving blood transfusion in non-traumatic intensive care unit patients.” Health Data Sciences. 2024.