Ambulatory disability is one of the most common, bothersome and limiting symptoms for people living with multiple sclerosis (MS) and greatly decreasing quality of life (1, 2). Walking capacity is measured in the clinic using a variety of validated outcomes (e.g. Timed 25-Foot Walk [T25FW] test), however, measurement and evaluation of walking performance (i.e. what they actually do in daily life) may be more important to the patient and reflect actual function (3, 4).

Efforts by several groups focused on remote (real-world) monitoring of ambulatory function mostly using average daily step count (STEPS), obtained from research-based and commercially available accelerometers (5–12). In the Fitbit remote monitoring in MS (FITriMS) study, daily step counts were collected continuously for over 1-year (5, 6). The STEPS averaged over the first 30 days correlated with disability (Expanded Disability Status Score [EDSS]), clinic-based metrics (T25FW, Timed-Up and Go Test [TUG], 2-min walk test [2MWT]) and patient reported outcomes (i.e. 12-item MS Walking Scale [MSWS-12]) (5) Longitudinal analysis over 1 year demonstrated a change in STEPS over time, even when conventional measures remained stable (6). These findings suggest remote physical activity monitoring provides additional sensitivity when capturing change in performance in people with MS.

Physical activity (activity) is quantified in different ways. The STEPS summarizes the total number of steps taken during an allotted epoch (usually 1 day) but does not reflect how different ambulatory behavior results in unique or distinctive step distributions, nor does it provide information or understanding of variability and intensity of the activity. Minute-by-minute (M-M) step count data can provide more granular information on the intensity, duration and frequency of ambulatory behavior. The aims for this analysis were to: determine the best probabilistic model using M-M step data to characterize activity distribution in people with MS with a range of ambulatory disability, evaluate the statistical validity of this new outcome, and compare with STEPS and conventional disability correlates at baseline.

From the 104 patients recruited into FITriMS, 10 did not have M-M data - due to sporadic syncing resulting in only daily step count data rather than M-M which requires weekly syncing. Of 94 participants used for this analysis, 63.5% carried a relapsing MS diagnosis (the remaining had progressive forms of MS) and more than two-thirds (76.3%) were women. The mean (SD) age was 55.5 years (13.7), median disease duration was 19.6 (IQR: 16) years, and median EDSS 4.0 (IQR: 3.5). All participant characteristics are summarized in Table 1 and EDSS distribution in Figure 1.

The GMM was found to be the best fit for individual subject data (see Supplemental Table 1 for full comparison of distributions). GMM fits a greater variability in activity distribution and provides more flexibility in generalizable representation of MS activity. For example, Figure 2A shows a participant who spend most of their time in low levels of activity and Figure 2B depicts the GMM fit for a participant who perform some higher activity over the day (100–130 steps per min).

All Gaussian parameters (μ, σ, and π) except for π3 (since π1, π2 and π3 are perfectly collinear [π1 + π2 + π3 = 1]) correlated with clinical metrics (EDSS, TUG, MSWS12, 2MWT, T25FW). Using individual participant data, GMM was fit to M-M step data.

We propose that the 3 Gaussians correspond to 3 activity levels for each patient (low activity, medium activity, high activity – as ordered by their μ [means]): each μ represents the average step count we would expect from each step activity level, as characterized by its corresponding Gaussian; each σ represents the variability we would expect for each activity level; and each π represents the proportion of activities we would expect from each activity region.

Linear regression was used to show moderate to high correlation between GMM acquired over the study's baseline first 4 weeks of monitoring, and both clinic-based and patient-reported outcome measures. The results using GMM were consistently stronger than results obtained using only STEPS. Adding STEPS to GMM (GMM + STEPS) consistently outperforms either measure its own. Adjusting for sex, age and disease duration improved all models (Table 2).

Table 3A shows the centers for each Gaussian for different EDSS groupings. EDSS groups with lower levels of disability had consistently higher Gaussian centers. Further, Gaussians at higher activity levels had higher variances than those at lower levels.

These results provide preliminary evidence for the use of a GMM probabilistic model to characterize activity distribution using granular M-M step count in people with MS. This model performs better (stronger correlations and adjusted R²) than previous methods using crude STEPS.

The GMM model was able to generalize over a range of activity profiles. More specifically, it was able to capture the step distributions of those in the cohort where a significant percentage of steps come from high activity levels. Previous work from our group illustrated the wide variability in activity levels in people with MS, within and between all ambulatory disability levels (5). Therefore, the ability to generalize M-M modeling to highly variable distributions using the 3 compartment GMM has clinical appeal. In addition, the GMM dovetails well with existing literature regarding activity levels classified into three levels: low, moderate/moderate to vigorous, and high physical activity (21–23).

The GMM representation of activity outperforms other statistical models and performs better than STEPS as compared to conventional disability correlates (5). The GMM may provide an interpretable framework to better understand the association between different levels of activity and clinical metrics. It also allows further analysis of walking performance and behavior by taking step distribution and proportion of time at each intensity into account. GMM and STEPS are complimentary; STEPS provides a mean, whereas GMM presents information regarding intensity, variance, and proportional step distribution. The model including STEPS and GMM generated high correlation with the conventional outcomes, suggesting that the overall mean (STEPS) is a useful metric in combination with the more descriptive GMM.

This analysis has important limitations. Although this cohort was well-phenotyped, larger studies in more heterogeneous populations are needed to provide additional evidence of generalizability and replicability. Due to the limited availability of M-M, longitudinal (>7 days) datasets in people with MS, we were not able to perform a replication analysis. In addition, analysis of larger datasets collected in randomly recruited cohorts (rather than block enrolled) will be required. Data processed with the same granularity (minute-by-minute steps) from healthy age-matched controls would provide a better understanding about the proportions of time spent in each activity level.

A GMM based model is also relatively inflexible when approximating activity distributions that are not Gaussian in nature. A possible solution to overcome this representational limitation is to use an autoencoder (24), a type of neural network, to compress the distribution into a flexible lower dimensional representation with greater generalizability.

Without access to a platform that automatically pulls the M-M data, retrieving these detailed metrics would be more burdensome than simply downloading daily step count from the Fitbit.com website. Although we were fortunate to be able to use an in-house platform (Eureka: https://info.eurekaplatform.org/), there are fee-based companies that offer such services. In addition, the M-M data and GMM model provide improved correlations with conventional measures and provide insight into how a patient spends their time in different activity levels. For instance, a larger “μ3” represents greater average step count; a smaller “σ1” denotes less variability; and “π2” denotes greater proportion of activities in the moderate range. These granular information combined with the level of disability (Tables 3A–C), provide a potential avenue for predictive algorithms and later, individualized rehabilitation plans.

Moderate-to-Vigorous activity equates with our π2, and has frequently been cited as the benchmark for determining optimal physical activity in MS and the general population. People needing double support to ambulate (EDSS = 6.5) tended to spend a larger proportion their activity in π₁ (corresponding to low levels of activity or sedentarism). On the other hand, people with lower disability (EDSS = 0.0–3.5) presented with a greater proportion in π₂ and π₃ (corresponding with more moderate and higher activity levels). Therefore, it may be more clinically useful to evaluate π1, and π3 (the time spent in low levels and very high levels of activity) when assessing an individual patient's activity level and subsequent risk factors or rehabilitation needs. For example, understanding fluctuations in activity over the day to better personalize when to focus rehabilitation interventions and on what (intensity and duration of activity). Research awareness already shifted toward investigating the effect of sedentary time on health and disease (25–27). Higher physical activity has been associated with beneficial changes in the brain and spinal cord, as well with decreased levels of disability (28, 29). However, this area of investigation is in its infancy and the underlying mechanism of action is not yet understood. Methods presented in this paper could build on and support these lines of investigation – with aims at promoting greater overall wellness, and potentially delay disease progression in people with MS (29–32).

This model will be used as a framework to predict disease progression over the longer term (>2 years) and to develop further descriptive metrics for activity. How time spent at various activity levels is associated with MS disability over time – i.e., temporal validation of the Gaussian parameters for prediction on disability progression – remains to be determined. Models including associations of fall-risk prediction in people with MS would also be highly clinically valuable.

Results from this analysis favor a 3 compartment GMM as the best probabilistic model to characterize dynamic ambulatory activity in people with MS with a wide range of disability. As compared to STEPS as a sole outcome, this method demonstrated stronger associations with conventional clinic-based and patient-reported outcomes. To unearth the full potential of this method, additional longitudinal exploration of predictive value is required.

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

The studies involving human participants were reviewed and approved by UCSF Human Research Protection Program, Box 1288 490 Illinois Street, Floor 6 San Francisco, CA 94143. The patients/participants provided their written informed consent to participate in this study.

VB and MW: design and conceptualized study, analyzed the data, and drafted the manuscript for intellectual content. ZX and AA: analyzed the data and drafted the manuscript for intellectual content. RB, MP, GM, JO, and BC: interpreted the data and revised the manuscript for intellectual content. JG: design and conceptualized study, interpreted the data, and revised the manuscript for intellectual content. RH: design and conceptualized study, analyzed and interpreted the data, and revised the manuscript for intellectual content. All authors contributed to the article and approved the submitted version.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.