Fig. 1.

Adaptive metrics. The diagram shows the model-fitting process using 4 wk of training data. We trained a model using the 4 most recent weeks with complete outcome data, including inputs from w − 6 to w − 3 and outputs from w − 3 to w. We then used this model, with input data from w, to estimate the probability of “high” or “very high” future mortality at w + 3 and specified a binary prediction based on whether this probability exceeded the user’s cutoff. (When a single indicator is used as the only input, this process is equivalent to identifying the optimal threshold for the indicator over the training period, accounting for user preferences.)