Platelet inventory management is challenging due to their short shelf-life and fluctuating demand. Researchers in Germany created a risk assessment system for platelet inventory using artificial intelligence (A.I.). Based on retrospective electronic health records from 2017-2022 at one hospital, over 34,800 patients (median age, 64 years; 42% female) with hospital stays greater than one day and minimum platelet counts <150 X 103 platelets/µL were used to train and test a deep learning risk-assessment model to forecast platelet transfusions for patients within the following 24 hours. Features in the model included demographics, diagnoses, lab results, past and present medical procedures, previous transfusions, blood counts, medications, and hospital duration. The model was best at predicting the platelet demands of hematology-oncology patients (area under the precision-recall curve, 0.84; area under the receiver operating curve, 0.98); the model had the lowest performance for cardiothoracic surgery patients, likely because of unexpected bleeding. While artificial intelligence models have the power to help manage platelet inventory, they have relatively low sensitivity and must be optimized for unique patient populations.
Reference: