Table 2.

Predicting intervention outcomes using data collected pre-intervention and a k-nearest neighbor model

	Machine learning: k-nearest neighbor model
Feature sets	All participants			Unexpected non-responder
	Outcome prediction for			Outcome prediction for
	ARAT	BBT	NHPT	ARAT	BBT	NHPT
	Balanced accuracy (%)			Correct (yes/no)
1	55	63	43	y	y	y
2	49	28	54	n	n	y
3	52	37	43	y	y	y
4	43	53	40	n	n	y
5	77	66	71	y	y	y
6	80	63	64	n	y	n
1, 2	83	80	43	y	y	y
1, 3	60	25	50	y	y	y
1, 4	55	46	63	y	y	y
1, 5	69	55	64	y	y	y
1, 6	93	59	33	n	y	n
1, 2, 3	71	35	50	y	y	y
1, 4, 6	68	42	48	n	y	n
1, 5, 6	85	60	68	n	y	y
1, 4, 5, 6	64	69	61	n	y	n
1, 2, 3, 4, 5, 6	68	59	64	n	y	y

Multiple machine learning models were trained using different feature sets (independent variables, 1–6). The training label indicated whether a considerable change across intervention was observed in a specific conventional score (dependent variable; ARAT, BBT, or NHPT). The models were evaluated in a leave-one-out cross-validation and specifically tested for one individual with strong activity limitations who did not show improvements across neurorehabilitation (referred to as unexpected non-responder). Feature set nomenclature: (1) patient master data (ms type, chronicity, age, sex); (2) intervention group; (3) disability (EDSS, disability group); (4) conventional scales of body functions (motricity index, static fatigue index, monofilament index, symbol digit modality test, Fahn’s tremor rating scale); (5) digital health metrics of sensorimotor impairments (ten VPIT metrics); (6) Conv. scale of activity (ARAT, NHPT, BBT). The best performing (accuracy and unexpected non-responder) models relying on the least amount of features are highlighted in bold for each conventional scale. ARAT: Action Research Arm Test. BBT: Box and Block Test. NHPT: Nine Hole Peg Test. VPIT: Virtual Peg Insertion Test