Development and Validation of an Explainable Machine Learning Model for Prediction of Massive Transfusion in Upper...

Area of Science:

• Medical research
• Clinical informatics
• Machine learning applications in healthcare

Background:

• Upper gastrointestinal bleeding (UGIB) is a critical condition with high mortality, particularly when massive transfusion (MT) is needed.
• Existing scoring systems for MT prediction in UGIB have moderate accuracy and do not fully capture complex variable interactions.

Purpose of the Study:

• To develop and validate an explainable machine learning (ML) model for predicting the need for MT in patients with UGIB.
• To improve the precision and interpretability of MT risk assessment in UGIB compared to traditional methods.

Main Methods:

• A retrospective study involving 700 UGIB patients, with data split into training, testing, and external validation cohorts.
• Feature selection using Boruta and LASSO regression identified 8 key clinical variables from an initial 18.
• Seven ML algorithms were evaluated, with the optimal model further assessed for discrimination, calibration, and clinical utility using SHapley Additive exPlanations (SHAP) for interpretability.

Main Results:

• The Random Forest (RF) model demonstrated superior performance with high AUC values across training (0.862), testing (0.823), and external validation (0.807) cohorts.
• Calibration plots confirmed strong agreement between predicted and observed probabilities, and decision curve analysis showed clinical utility.
• SHAP analysis identified key predictors such as impaired mental status, liver cirrhosis, and INR, aligning with clinical expertise.

Conclusions:

• The developed ML model shows significant promise for accurately identifying UGIB patients at high risk of requiring MT.
• The model's interpretability enhances its potential to assist clinicians in optimizing blood management strategies.
• Further prospective validation is recommended to confirm the model's clinical utility in diverse healthcare settings.