Table 2.

Comparison of HC vs. PwMS subgroup classification results between various models for each task subset, $\mathcal{T}$. Results are presented as: (1) the posterior overall subject-wise outcome for one cross-validation (CV) run as well as (2) the 2MWT test-wise median and interquartile range (IQR) across that CV in brackets. The best performing model for each $\mathcal{T}$ are highlighted in bold. Acc: Accuracy; $\kappa$, Cohen’s Kappa statistic; MF${}_1$, Macro-F1 score.

$f(·)$	Acc.	$\kappa$	MF${}_1$
	HC vs. PwMSmild
Features + SVM${}^1$	0.671 (0.576, 0.544–0.696)	0.212 (0.153, 0.088–0.393)	0.605 (0.575, 0.527–0.694)
DCNN (end-to-end)${}^2$	0.658 (0.601, 0.517–0.641)	0.226 (0.082, 0.037–0.194)	0.613 (0.541, 0.494–0.588)
DCNN (UCI HAR$\to$FL)${}^3$	0.776 (0.741, 0.688–0.767)	0.510 (0.435, 0.346–0.481)	0.754 (0.716, 0.662–0.737)
DCNN (WISDM$\to$FL)${}^3$	0.763 (0.733, 0.698–0.761)	0.486 (0.479, 0.343–0.490)	0.741 (0.727, 0.667–0.743)
PwMSmild vs. PwMSmod
Features + SVM${}^1$	0.849 (0.783, 0.706–0.858)	0.627 (0.566, 0.412–0.708)	0.813 (0.778, 0.692–0.853)
DCNN (end-to-end)${}^2$	0.822 (0.682, 0.617–0.763)	0.583 (0.356, 0.166–0.444)	0.791 (0.675, 0.562–0.721)
DCNN (UCI HAR$\to$ FL)${}^3$	0.904 (0.849, 0.839–0.873)	0.776 (0.675, 0.650–0.707)	0.888 (0.837, 0.823–0.852)
DCNN (WISDM$\to$FL)${}^3$	0.918 (0.869, 0.833–0.935)	0.810 (0.690, 0.630–0.844)	0.905 (0.845, 0.812–0.922)
HC vs. PwMSmod
Features + SVM ${}^1$	0.800 (0.773, 0.737–0.881)	0.595 (0.546, 0.474–0.763)	0.796 (0.772, 0.737–0.881)
DCNN (end-to-end)${}^2$	0.822 (0.734, 0.663–0.831)	0.641 (0.462, 0.292–0.657)	0.820 (0.730, 0.618–0.828)
DCNN (UCI HAR$\to$FL)${}^3$	0.889 (0.873, 0.730–0.929)	0.777 (0.743, 0.446–0.847)	0.889 (0.870, 0.723–0.924)
DCNN (WISDM$\to$FL)${}^3$	0.911 (0.886, 0.766–0.911)	0.821 (0.772, 0.520–0.820)	0.911 (0.886, 0.760–0.910)
HC vs. PwMSmild vs. PwMSmod
Features + SVM${}^1$	0.629 (0.551, 0.510–0.577)	0.368 (0.093, 0.020–0.103)	0.580 (0.510, 0.495–0.540)
DCNN (end-to-end)${}^2$	0.608 (0.503, 0.488–0.516)	0.274 (0.106, 0.081–0.130)	0.523 (0.446, 0.402–0.483)
DCNN (UCI HAR$\to$FL)${}^3$	0.814 (0.703, 0.700–0.744)	0.673 (0.331, 0.325–0.423)	0.796 (0.672, 0.664–0.720)
DCNN (WISDM$\to$FL)${}^{13}$	0.763 (0.690, 0.677–0.737)	0.571 (0.303, 0.274–0.407)	0.725 (0.671, 0.644–0.699)

${}^1$ “features + SVM” refers to classification performed using features and a SVM with the pipeline described in³;

${}^2$ “end-to-end”, refers to a model trained and validated end-to-end exclusively on ${\mathcal{D}}_T$ data;

${}^3$ “$\to$” denotes the source HAR dataset ${\mathcal{D}}_S$ used and transferred to FL ${\mathcal{D}}_S$ and ${\mathcal{T}}_S$. See Fig. 6 for a more detailed description of the TL approach used in this study.