Fig. 4: Model performance of the integrated system for ventilator resource planning (RMS) in an ICU with an effective capacity of 42 beds.
From: RMS: a ML-based system for ICU respiratory monitoring and resource planning
a Observed ventilator usage pattern in the HiRID-II dataset, in terms of days on which a particular number of ventilators is used. b Performance of RMS compared with the baseline, in terms of mean absolute error (MAE) for predicting the maximum number of used ventilators at a fixed horizon in the future. c Example of a typical ICU setting shown for a duration of one day, annotated with ground-truth, RMS, and baseline predictions. In the rug plot, relevant better predictions are marked for the offsets 1–4. d RMS compared with the baseline, for different absolute differences of predictions (1–4) in the rows, for a prediction horizon of 8–16 h in the future. The table entries denote the proportion of time points in which either RMS or the baseline is better by at least the difference of the row. e Performance of RMS for different current ventilator ICU usage scenarios, ranging from low usage, to high usage, compared with the baseline, for a prediction horizon of 8–16 h in the future. The number of time points falling into each bin is denoted in parentheses. f Performance of RMS, by hour of the day when the prediction is performed, compared with the baseline, for a prediction horizon of 8–16 h in the future.
