Table 2.

Benchmark results as median RMSE_bl.NLI and (2.5%-ile, 97.5%-ile) on held-out data for EMSCI and Sygen (external validation). The best-performing approach (lowest RMSE_bl.NLI, if tied on median lower value for 97.5%-ile better) is highlighted. Columns labeled All times include all instances. Columns labeled 4 ± 1 weeks include instances with an initial assessment between 21 and 35 DAI

	EMSCI		Sygen
	All times	4 ± 1 weeks	All times	4 ± 1 weeks
Linear regression	0.63 (0.00, 2.22)	0.64 (0.02, 2.15)	0.70 (0.00, 2.32)	0.68 (0.08, 2.15)
Random forest	0.61 (0.01, 2.17)	0.60 (0.01, 2.08)	0.64 (0.03, 2.37)	0.61 (0.03, 1.89)
XGBoost	0.61 (0.01, 2.12)	0.61 (0.01, 2.08)	0.70 (0.01, 2.33)	0.68 (0.02, 2.00)
CNN	0.55 (0.00, 2.50)	0.55 (0.00, 2.57)	0.57 (0.00, 2.85)	0.60 (0.00, 2.90)
Transformer	0.63 (0.00, 2.58)	0.63 (0.00, 2.57)	0.65 (0.00, 3.00)	0.71 (0.00, 3.27)
GNN	0.58 (0.00, 2.60)	0.60 (0.00, 2.47)	0.68 (0.00, 2.61)	0.63 (0.00, 2.57)

Abbreviations: RMSE_bl.NLI root mean squared error (RMSE) below neurological level of injury (bl.NLI), CNN convolutional neural network, GNN graph neural network