Patients in intensive care units (ICU) face a high risk of subacute illnesses, such as hemorrhage, respiratory failure, and sepsis. Our early works have produced and validated pre-trained predictive models of these subacute illnesses using physiological data from one institution. Because these subacute illnesses may share physiological characteristics that precede in-hospital cardiac arrest (IHCA), we therefore aim to test the pre-trained models to predict IHCA using data from a different institution to leverage the prior work to improve IHCA prediction.
Six pre-trained models, as provided in the CoMET® risk prediction software (NKDHS, Irvine, CA), were selected. The models were originally trained for three subacute illnesses with data from either medical (MICU) or surgical (SICU) units at an academic medical center in Virginia. The models employ features derived from time series of physiological vital signs and waveforms that characterize specific dynamic physiological patterns associated with decompensation. The external dataset, collected from a different health institution in California, included 28.5 bed-years of physiological time series of 2808 patients, with 173 IHCA events. We used the area under the receiver characteristics curve (AUC) as the performance metric.
Results showed the hemorrhage model for MICU had an AUC of 68.2% (95%CI: 64.4%~71.5%) for predicting IHCA, while the performance was 70.2% (95%CI: 65.1%~74.8%) for the SICU model. Intubation models exhibited AUCs of 67.5% (95%CI: 63.3%~72.3%) and 70.2% (95%CI: 65.8%~74.1%), and sepsis models had AUCs of 67.3% (95%CI: 62.38%~71.5%) and 63.9% (95%CI: 60.4%~67.2%).
Despite a slight performance drop (performance for the intended subacute illnesses: 61%~77%), the pre-trained models effectively predicted IHCA. This variation may stem from the different degrees of feature sharing among three subacute conditions for IHCA. Our study highlights the potential of incorporating pre-trained models for predicting additional related clinical outcomes and encourages further investigation to optimize their use in diverse clinical situations.