Optimizing platelet transfusion through a personalized deep learning risk assessment system for demand management.

Area of Science:

• Biomedical Informatics
• Machine Learning in Healthcare
• Hematology

Background:

• Platelet demand management (PDM) is complex and resource-intensive for hospitals.
• Current PDM relies on inpatient numbers and institutional standards, often leading to shortages.
• A patient-specific predictive approach is needed to optimize platelet utilization.

Purpose of the Study:

• To develop and validate a deep learning-based model for forecasting individual patient platelet transfusions.
• To support Platelet Demand Management (PDM) by providing 24-hour risk assessments.
• To mitigate platelet shortages and improve patient care through predictive analytics.

Main Methods:

• Developed a deep learning model using retrospective electronic health record data from 34,809 patients (2017-2022).
• Utilized static and time-dependent features including demographics, diagnoses, procedures, blood counts, and transfusion history.
• Implemented an expanding window approach for training and live prediction with a 30-day input and 24-hour forecast.

Main Results:

• The hematology/oncology model achieved high performance (AUC-PR: 0.84, ROC-AUC: 0.98).
• A multispecialty model showed good performance (AUC-PR: 0.73).
• The cardiothoracic surgery model had lower performance (AUC-PR: 0.42), potentially due to intraoperative bleeding.

Conclusions:

• This is the first deep learning predictor for individualized 24-hour platelet transfusion risk assessment.
• The decision-support system can aid in early detection of platelet demand.
• Implementation may prevent critical transfusion shortages and enhance patient care.