Table 3.

Models for poor outcome (paradigm II).

Scenario	GLM	Lasso	ElasticNet	Catboost	MLP	SVMC	Naive bayes
II A	0.67	0.7	0.64	0.7	0.71	0.59	0.69
II A+B	0.65*	0.7	0.62*	0.7	0.69	0.57	0.65*
II A+B+C	0.68+	0.71	0.65+	0.73*+	0.7	0.65*+	0.66*

The prediction variable sets (Table 1) were used to predict poor outcome (mRS dichotomized as 5 and 6 vs. 0–4). In contrast to the favorable outcome paradigm, the addition of thrombectomy-associated variables (set C) did not lead to relevant improvements in the performance of machine learning models. Only the Catboost model profited slightly in a clinically relevant AUC range. Highest AUC results for each variable set are marked in bold. Confidence intervals for all models are included in the Supplementary Material. Statistically significant difference in model performance between variable sets are marked with * and + to signal difference from A and from A+B, respectively. Significance was determined by a value of p lower than 0.05, resulting from the Wilcoxon signed-rank test.