Predicting cognitive scores from wearable-based digital physiological features using machine learning: data from a clinical trial in mild cognitive impairment

Table 3 Performance of the best models predicting NTB scores

Cross-validation method	Cognitive measure	Machine Learning algorithm	Number of used features	Feature selection p-value	r	r gain	rho	rho gain	MAE
Subject-based	Global Cognition	ElasticNet	11	0.001	0.54	0.02	0.50	− 0.06	0.59
	Executive Function	ElasticNet	13	0.001	0.69	0.15	0.70	0.46	0.46
	Processing Speed	ElasticNet	14	0.001	0.47	0.08	0.48	0.27	0.67
	Memory Immediate	ElasticNet	4	0.001	0.44	− 0.03	0.44	0.01	1.24
	Memory Delayed	ElasticNet	6	0.001	0.48	0.22	0.61	0.33	0.97
Interval-based	Global Cognition	Random Forest	7	0.0001	0.92	0.25	0.91	0.25	0.16
	Executive Function	Random Forest	9	0.00001	0.89	0.22	0.87	0.33	0.15
	Processing Speed	XGBoost	11	0.0001	0.85	0.33	0.82	0.31	0.21
	Memory Immediate	Random Forest	3	0.0001	0.92	0.29	0.87	0.28	0.33
	Memory Delayed	XGBoost	3	0.0001	0.86	0.38	0.86	0.37	0.35

ISSN: 1741-7015