Machine Learning-Based Models for Prediction of Critical Illness at Community, Paramedic, and Hospital Stages

Table 3

Comparison of models for prehospital assessment to predict critical illness.

Authors	Data source	Predictors	Model	AUC (95% CI)
Lee et al. [21]	Single medical center	Age, sex, chief complaints, vital signs, and comorbidities	Neural network	0.801 (0.796–0.805)
Kang et al. [11]	NEDIS	Age, sex, chief complaints, vital signs, and symptom onset to arrival time	FFNN	0.867 (0.864–0.871)
			ESI	0.839 (0.831–0.846)
			NEWS	0.741 (0.734–0.748)
Shirakawa et al. [22]	Single medical center	Age, sex, chief complaints, and vital signs	LR	0.805 (0.782–0.827)
			RF	0.813 (0.786–0.834)
			GBM	0.818 (0.792–0.839)
Raita et al. [12]	NHAMCS	Age, sex, chief complaints, vital signs, comorbidities, and mode of arrivals	LR	0.84 (0.83–0.85)
			RF	0.85 (0.84–0.87)
			GBM	0.85 (0.83–0.86)
			DNN	0.86 (0.85–0.87)
Spangler et al. [13]	Uppsala ambulance service	Age, sex, chief complaints, vital signs, and operational characteristics of calls	GBM	0.79 (0.78–0.80)
			NEWS	0.76 (0.75–0.78)

AUC: area under the curve; CI: confidence interval; FFNN: feedforward neural network; ESI: emergency severity index; NEWS: national early warning score; LR: logistic regression; RF: random forest; GBM: gradient boosting machine; DNN: deep neural network; NEDIS: national emergency department information system; NHAMCS: national hospital and ambulatory medical care survey; and CHARS: comprehensive hospital abstract reporting system.