Prediction of Patient Deterioration Using Machine Learning

Sensitivity for recognition of a safety composite

The sensitivity (true positives divided by condition positives) for detection of a safety composite (overnight visit, extra unplanned visit, transfer back to the hospital, death during admission, delirium, loss of consciousness, or other major event).

Time frame: From admission to discharge, on average 5 days

Specificity for recognition of a safety composite

The specificity (true negatives divided by condition negatives) for detection of a safety composite (overnight visit, extra unplanned visit, transfer back to the hospital, death during admission, delirium, loss of consciousness, or other major event).

Time frame: From admission to discharge, on average 5 days

Positive predictive value for recognition of a safety composite

The positive predictive value (true positives divided by the sum of true positives plus false positives) for detection of a safety composite (overnight visit, extra unplanned visit, transfer back to the hospital, death during admission, delirium, loss of consciousness, or other major event).

Time frame: From admission to discharge, on average 5 days

Negative predictive value for recognition of a safety composite

The negative predictive value (true negatives divided by the sum of true negatives plus false negatives) for detection of a safety composite (overnight visit, extra unplanned visit, transfer back to the hospital, death during admission, delirium, loss of consciousness, or other major event).

Time frame: From admission to discharge, on average 5 days

Rate of alarms with clinical utility

We will use general estimating equations (GEE) with three outcomes per patient (the number of clinically important alarms for BioVitals, NEWS2, and traditional vital signs); the GEE will account for the clustering between the three outcomes on a patient. The GEE will use a negative binomial marginal model with a log-link for the number of alarms with clinical utility and an offset for log length-of stay (in hours); with this model, we model the rate per hour of number of alarms with clinical utility with BI, NEWS2, and traditional vital signs. The main covariate in the negative binomial model will be a three-level covariate for method: BI vs NEWS2 vs traditional vital signs, and the exponential of the effect of this covariate will be a pair-wise rate ratio for BI vs NEWS2 vs traditional vital signs.

Time frame: From admission to discharge, on average 5 days