Transferability and interpretability of the sepsis prediction models in the intensive care unit

Table 2 The results of models trained on different datasets on the HDRJH test set

Model	Transfer learning	Training set	Accuracy	AUC
LightGBM	N	MIMIC-III	0.74	0.86
LightGBM	N	HDRJH	0.81	0.92
LightGBM	Y	MIMIC-III + HDRJH	0.83	0.93
MLP	N	MIMIC-III	0.76	0.86
MLP	N	HDRJH	0.84	0.92
MLP	Y	MIMIC-III + HDRJH	0.78	0.93

These models predict sepsis in 4-h preceding and the complete results can be found in Additional file 1: Tables S5 and S6. MIMIC-III + HDRJH means training on MIMIC-III first and then tuning on the training set of HDRJH by transfer learning techniques

ISSN: 1472-6947