Data sampling from raw patient samples to trainable samples. Before using the patient data samples retrieved from the data mining process for training, each sample had to pass another preprocessing step. First, metadata, such as department, specific ward, and sex, had to be encoded, and age was normalized. Next, the received PCs per time window were encoded. Then, categorical features of the time-distributed data were transformed into a one-hot encoding (see supplemental Figures 1-4 for a complete list of conditions, medications, and procedures). Finally, the observations were normalized. This was done for every sample before the data were batched.