Architecture of the LSTM–based multi-input single output model. The first input type is a tensor representing the generated time series, with 7 features for each time series element, including platelet count, day of the week (represented by cosine and sine), and Boolean variables indicating measurements, PC given, holiday, and weekend. The second input type is location features uniquely identifying a patient's last institute and ward. The third input type is categorical features generated from past procedures, diagnosed conditions, and given medications within the last 30 days. The sample size for each input type is batch size 60 × 7, batch size ×2, and batch size ×16, respectively. The institute encoding feature is disregarded when training department-specific models. The model's final output, generated through the sigmoid activation function in the last dense layer, represents the likelihood that a patient will need a platelet transfusion.