A gradient boosting tree model for multi-department venous thromboembolism risk assessment with imbalanced data

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Clinical Decision Support

Background:

• Venous thromboembolism (VTE) is a leading cause of vascular mortality, necessitating accurate in-hospital risk assessment.
• Traditional VTE risk assessment tools struggle with large-scale application due to reliance on expert-designed, population-specific rules.
• Electronic Health Record (EHR) data enables data-driven, machine learning approaches for superior VTE risk prediction.

Purpose of the Study:

• To develop and evaluate a novel machine learning model for VTE risk assessment using multi-department hospital data.
• To address the challenges of data heterogeneity and distribution across different hospital departments.
• To improve the accuracy and applicability of VTE risk prediction models in diverse clinical settings.

Main Methods:

• Utilized a gradient boosting tree model for VTE risk assessment.
• Framed the multi-department prediction task as a multi-task learning problem.
• Introduced a task-aware tree-based method (TSGB) to handle departmental data heterogeneity.
• Proposed two variants of TSGB to address performance decline in imbalanced VTE data volumes.

Main Results:

• The proposed TSGB method demonstrated improved overall cross-department Area Under the Curve (AUC) performance.
• The model variants effectively mitigated performance degradation associated with imbalanced VTE data.
• Experimental results showed enhanced prediction performance across individual departments compared to baseline methods.

Conclusions:

• The task-aware multi-task learning approach significantly improves VTE risk prediction accuracy in heterogeneous hospital data.
• The developed TSGB variants offer a robust solution for VTE risk assessment, outperforming existing methods.
• This data-driven strategy enhances clinical intervention timeliness and patient safety by providing more reliable VTE risk stratification.