Fig. 6: Processing steps from wearable features to CDI outputs.

The source minute features (3 of 17 shown above) are windowed using a moving 24 h window with a 1 h step size. Each windowed set of minute features are processed to produce a feature vector, and timestamped to the end of the 24 h window. Each feature vector is independently passed to the CDI model to produce a decompensation probability at a cadence of once per hour. The positive detection window is defined as up to 14 days before the time of hospitalization and is evaluated at corresponding CDI decisions.