Fig. 4. Development of a deep learning model for calculation of an SCC-Score.

a Time series of vital signs and physical activity recorded by a medical wearable. b Clinical documentation, such as patient charts or laboratory results, that were reviewed for identifying SCC events. c According to the clinical documentation, the hours without evidence of SCC were annotated as regular hours, the remaining hours were regarded as non-regular. d regular hours for each individual patient were randomly split into two datasets: 90% for training and 10% for testing and generating a null-distribution. For cross-validation, the splitting was repeated ten times. For training the deep learning model, the regular hours were presented to a deep neural network as part of a self-supervised contrastive learning objective. An SCC-Score based on the similarity between a test hour and the closest regular hour from the training set was calculated. e A null-distribution of SCC-Scores from regular hours in the test set was established. f For a given hour, a statistical test under the null-distribution was applied to detect SCC, with a significance level selected by clinical requirements.