Related Experiment Videos
A Clustering-Based Machine Learning Approach for Mortality Prediction in Gastrointestinal Bleeding: Development and
Laith Alomari1, Zaid Al-Fakhouri2, Jaber Jaradat3
1Department of Medicine, Jefferson Einstein Philadelphia Hospital, Philadelphia, Pennsylvania.
Background And Aims:
Gastrointestinal bleeding (GIB) is a life-threatening emergency with considerable morbidity and mortality. Traditional risk scores like AIMS65 and Glasgow-Blatchford Score (GBS) are limited in capturing nonlinear clinical interactions. We developed and externally validated a machine learning model to predict 30-day mortality in GIB patients.
Methods:
We retrospectively analyzed 5453 emergency department patients with GIB from the Medical Information Mart for Intensive Care IV-Emergency Department database for model development, with external validation using 7166 patients from Jefferson Health. Sixteen clinical and laboratory variables were selected based on a literature review and clinical relevance. The development cohort was divided into training (80%) and internal validation (20%) sets. Survivors were partitioned into 24 clusters using K-means, with separate random forest models trained on each cluster, combined with all deceased cases. Performance was evaluated using the area under the receiver-operating characteristic curve, sensitivity, and specificity on the external validation set, then benchmarked against AIMS65 and the GBS.
Results:
The model achieved an area under the receiver-operating characteristic curve of 0.884 (95% confidence interval: 0.863-0.905) on internal validation and 0.882 (95% confidence interval: 0.863-0.900) on external validation, significantly outperforming AIMS65 (0.737) and GBS (0.768) (P < .001). At the optimal threshold, the model achieved 87.9% sensitivity and 74.3% specificity on the external validation cohort. At maximum sensitivity thresholds, the model maintained higher specificity (54.4%) than AIMS65 (29.7%) and GBS (17.0%) (P < .001). Clustering identified distinct phenotypes with mortality ranging from 0.6% to 15.3%. SHapley Additive exPlanations analysis identified age, albumin, hemodynamic parameters, and presenting hemoglobin and platelet count as key predictors.
Conclusion:
Our model provides superior risk stratification for 30-day mortality in GIB compared to conventional scores, with validated generalizability and potential for integration into electronic health record systems.
