Predicting Individual Patient Platelet Demand in a Large Tertiary Care Hospital Using Machine Learning

<b><i>Introduction:</i></b> An increasing shortage of donor blood is expected, considering the demographic change in Germany. Due to short shelf life and varying daily fluctuations consumption, storage platelet concentrates (PCs) becomes challenging. This emphasizes need for reliable prediction needed PCs bank inventories. Therefore, objective this study was evaluate multimodal data from multiple source systems within a hospital predict number transfusions 3 days on per-patient level. <b><i>Methods:</i></b> Data were collected 25,190 (42% female 58% male) patients between 2017 2021. For each patient, received PCs, count tests, drugs causing thrombocytopenia, acute diseases, procedures, age, gender, period patient’s stay collected. Two models trained samples using sliding window 7 as input day target. The model predicts whether patient will be transfused future. with an excessive hyperparameter search patient-level repeated 5-fold cross-validation optimize average macro F2-score. <b><i>Results:</i></b> tested 5,022 unique patients. best-performing has specificity 0.99, sensitivity 0.37, area under precision-recall curve score 0.45, MCC 0.43, F1-score 0.43. However, does not generalize well cases when transfusion recognized. <b><i>Conclusion:</i></b> A AI-based forecast could improve logistics management reduce product waste. In study, we build first individual demand. To best our knowledge, are introduce approach. Our units While underperforms, performs reliably. may clinical use pretest potential needing next days. As needs improved, further studies should deep learning wider characterization methodological multimodal, multisource Furthermore, hospital-wide consumption derived predictions.