TABLE 2.

Feature set extracted from the DEEP backend metrics and selected in the final model.

Game	Features from the tablet	Derived features	Total	Feature set used for ML modeling#	Selected in final model
Single tap	3	6	9	4	–
Alternate tap	4	12	16	8	–
Popping Balloons	6	30	36	11	–
Grow your garden	30	90	120	40	–
Hidden objects	37	119	156	86	–
Odd one out	74	166	240	105	–
Matching shapes	27	59	86	13	1
Jigsaw puzzles	32	66	98	24	3
Location recall	49	131	180	111	–
Across games	0	30	30	10	4
Interaction terms$	–	–	–	83	22
Principle components*	–	–	–	26	–
Mas-o-menos	–	–	–	1	1
Total	262	709	971	522	31

^#Highly correlated features (Pearson’s r > 0.9) were dropped to avoid multicollinearity during modeling, resulting in 522 features being used for training the machine learning models from the initial set of 971 features. ^$Interaction terms were generated by computing products and ratios of a subset of the features selected from the initial exploratory analysis (see Supplementary Table S2). *26 principle components explained 70% of the variance in the dataset.