31 participants were 100% compliant in wearing the sensors at all times for 7 consecutive days. The remaining 8 participants had an average non-wear duration of 72.3 ± 14.5 minutes (average compliance of about 95%). Figure 1 illustrates the correlations between sensor-derived physical activity metrics measured by the PAMSys pendant with the related clinical outcomes of FRDA (i.e., FA-ADL, FA-ADL LL, mFARS, and mFARS Section E), disease duration and GAA and FXN. The physical activity metrics were reported under three categories: 1) postures, 2) locomotion, and 3) postural transitions. Most locomotion-related metrics were negatively correlated with disease severity – See Figure 1 and Supplementary Table S1. The correlation coefficients ranged from -0.384 to 0.763 (p<0.05) indicating moderate to high effect size for clinical measures and biological biomarkers (i.e., GAA and FXN). Median step count per walking bout demonstrated the highest correlations (-0.627 to -0.763, p<0.05) with the clinical scores of mFARS, mFARS Section E, FA-ADL, and FA-ADL LL scores) but insignificant correlations of 0.05 and -0.24 with GAA and FXN, respectively.
Furthermore, Figure 2 illustrates the associations between sensor-derived GDM metrics of upper extremity with the related clinical outcomes of FRDA (i.e., FA-ADL, FA-ADL UL, mFARS, and mFARS Section B, Average 9 HPT), disease duration and GAA and FXN. The GDM metrics were reported under three categories: 1) GDM counts, 2) velocity features of GDMs, and 3) acceleration features of GDMs. Multiple velocity and acceleration features of GDMs were significantly correlated with the clinical scores in the range from -0.34 to -0.54 (p < 0.05), as well GAA from 0.35 to 0.6 (p < 0.05) and FXN from 0.38 to 0.58 (p < 0.05) – See Figure 2 and Supplementary Table S2.
We developed various machine learning models to leverage demographics, disease duration, and GAA, as well as both sensor-derived physical activity and GDM metrics – See Table 2. Model 1 achieved the following coefficient of determination (explained variance, R2) of 0.15 for FA-ADL, 0.22 for FA-ADL UL, 0.33 for FA-ADL LL, 0.47 for mFARS, 0.51 for mFARS Section B, and 0.15 for mFARS Section E. Additionally, Model 1 yielded R2 values of 0.39 and 0.55 for the biological biomarkers (i.e., GAA and FXN) of disease severity. The Model 2 provided R2 of 0.32 for FA-ADL, 0.56 for FA-ADL LL, 0.40 for mFARS, and 0.69 for mFARS Section E. Moreover, prediction for GAA and FXN using Model 2 yielded R2 of 0.03 and 0.42, respectively. Compared Model 1, Model 3 provided higher R2 values of 0.39 for FA-ADL, 0.65 for FA-ADL LL, 0.69 for mFARS, and 0.72 for mFARS Section E. Moreover, prediction for GAA and FXN using Model 3 yielded R2 of 0.65 and 0.58, respectively. For Model 4, R2 values of 0.28, 0.50, 0.70, and 0.60 were observed for FA-ADL, FA-ADL UL, mFARS, and mFARS Section B, respectively. Additionally, for biological outcomes, Model 4 accounted for 0.74 of the variance in GAA and 0.74 in FXN. Lastly, Model 5 yielded 0.36, 0.61, 0.65, 0.73, 0.61, and 0.85 of the variance in FA-ADL, FA-ADL UL, FA-ADL LL, mFARS, mFARS Section B, and mFARS Section E, respectively. Additionally, for biological outcomes, Model 5 accounted for an R2 of 0.77 in GAA and 0.81 in FXN.