Table 3.

Comparison between STL and MTL in stroke prediction. MAND-LR, MAND-MLP, MAND-LSTM, and MAND-MHSA denote the MAND architecture integrated with logistic regression, multilayer perceptron (MLP), LSTM, and multi-head self-attention as ICD feature extraction modules, respectively. FM and DCN represent CTR-based approaches. BAC: balanced accuracy; FPR: false positive rate; FNR: false negative rate.

Backbone model	STL/MTL	Log loss	AUC	BAC	Precision	Recall	F1 score	FPR	FNR
MAND-LR18	STL	0.2120	0.8585	0.6023	0.7794	0.2102	0.3311	0.0056	0.7898
	MTL	0.2134	0.8586	0.6009	0.7857	0.2071	0.3279	0.0053	0.7929
MAND-MLP18	STL	0.2042	0.8626	0.6169	0.8939	0.2364	0.3739	0.0026	0.7636
	MTL	0.2198	0.8467	0.5694	0.8637	0.1409	0.2422	0.0021	0.8591
MAND-LSTM18	STL	0.1974	0.8700	0.6460	0.8339	0.2977	0.4387	0.0057	0.7023
	MTL	0.2050	0.8625	0.6286	0.8240	0.2626	0.3983	0.0054	0.7374
MAND-MHSA18	STL	0.2020	0.8703	0.6445	0.7540	0.2982	0.4274	0.0092	0.7018
	MTL	0.2076	0.8643	0.6392	0.7367	0.2882	0.4143	0.0098	0.7118
FM19	STL	0.2619	0.8330	0.6226	0.5708	0.2642	0.3612	0.0190	0.7358
	MTL	0.2288	0.8351	0.6216	0.6718	0.2551	0.3698	0.0119	0.7449
DCN22	STL	0.2013	0.8649	0.6336	0.8600	0.2715	0.4127	0.0043	0.7285
	MTL	0.2040	0.8627	0.6288	0.8509	0.2620	0.4003	0.0044	0.7380

Bold font indicates the better performance values between STL and MTL.