Table 1.

Selected examples of the use of interpretable machine learning approaches to NICU data

Article	Study population	Data set(s)	Predicted variable	Machine learning algorithm applied	Interpretability technique(s)
Overweg et al. [40]	ICU and TBI	CENTER-TBI, MIMIC-III	ICU/NICU mortality	BNN	HorseshoeBNN—a novel approach proposed by the authors; the horseshoe prior has been added to induce sparsity in the first layer of the BNN, enabling feature selection
Caicedo-Torres and Gutierrez [41]	ICU	MIMIC-III	ICU mortality	Multiscale deep convolutional neural network (ConvNet)	DeepLIFT, visualizations
Thorsen-Meyer et al. [42]	ICU	5 Danish medical and surgical ICUs	All-cause 90-day mortality	Recurrent neural network with LSTM architecture	SHAP
Wang et al. [43]	ICU patients diagnosed with cardiovascular disease	MIMIC-III	Survival	LSTM network	Counterfactual explanations
Fong et al. [44]	ICU	eICU collaborative research database and 5 ICUs in Hong Kong	Hospital mortality	XGBoost	SHAP
Che et al. [45]	Pediatric ICU patients with acute lung injury	Pediatric ICU at Children’s Hospital Los Angeles	Mortality, ventilator-free days	Interpretable mimic learning (using gradient boosting trees)	Partial dependence plots, feature importance, intrinsic interpretability of tree structure
Shickel et al. [46]	ICU	UFHealth, MIMIC-III	In-hospital mortality	RNN with GRU	Modified GRU-RNN network with final self-attention mechanism (to identify feature importance)
Farzanah et al. [47]	TBI	ProTECT III	Functional outcome – GOSE at 6 months	XGBoost	SHAP
Gao et al. [48]	TBI	NICU at Cambridge University Hospitals, Cambridge	Mortality 6 months post brain injury	Decision tree	Intrinsic interpretability of model
Thoral et al. [49]	ICU	AmsterdamUMCdb	ICU readmission and/or death, both within 7 days of ICU discharge	XGBoost	SHAP