The value of progressively accrued information during initial post-admission hours in forecasting future cardiorespiratory instability

Abstract

Introduction: Enabling clinicians to prospectively identify patients who will later become unstable would enable targeting resources to patients most in need as well as potential application of preventive care.

Objectives: To determine the incremental contribution of information
progressively available within the first 4 hours of (SDU) admission to improve models forecasting later development of cardiorespiratory instability (CRI), including a novel CRI relative risk score.

Methods: Continuous noninvasive vital sign (VS) monitoring data (heart
rate [HR], respiratory rate [RR; bioimpedance], oscillometric blood pressure [BP], peripheral oximetry [SpO2]) were collected from 1971 stepdown unit (SDU) patients, and CRI episodes defined as VS deviation beyond stability thresholds. Patients with any CRI (cases, n = 918) and those never displaying CRI (controls, n = 1053) were identified. We computed a minute-by-minute integrated CRI risk score based on the method described in [1], using features computed from VS data streams during trailing 15 minute rolling windows and a trained random forest machine learning model. We then computed for each patient a mean risk score aggregated from the risk scores during first 4 hours of SDU stay. Next we built a logistic regression model to forecast whether or not there will be a CRI event in the future. To mimic the temporal availability of data following patient admission, we first entered demographics available at patient admission (age, gender, Charlson Comorbidity Index score) into the model, and then the initial VS (5-minute average of continuous VS data accrued from minutes 10 to 15 after admission), and finally the relative risk score derived in the first 4 hours. We assessed the predictive contribution of information from these 3 progressively accrued categories
(demographics, initial VS, 4-hr risk score) by the Area Under Receiver Operating Curve (AUC) in a 10-fold cross validation experiment setup.

Results: The risk score derived from admission demographics alone
yielded an AUC of 58 ± 0.002 % to forecast future CRI. Adding the initial VS improved the AUC to 64 ± 0.003 %, and with further adding the 4-hr risk score the AUC became 67 ± 0.002 %.

Conclusions: A predictive model which incorporates patient data as
it becomes available, including a risk score derived within the first 4 hours, progressively improves the models ability to forecast future CRI development. Such forecasting information could enable clinicians to identify those patients who will become unstable in future very soon after admission in order to triage patients needing closer surveillance and potentially apply preemptive interventions.