Характеристика походки у людей с множественным склерозом с использованием цифровых данных с смартфонов: предлагаемая основа.

Background

Walking disability is common in people with multiple sclerosis (PwMS), owing to lower limb motor function deficits caused by neurological impairment
1
and/or other mobility difficulties caused by multiple sclerosis (MS) symptoms, such as fatigue and visual and cognitive impairment. 2 – 4 The impact of MS on walking, or gait, can manifest in various ways, including slower walking speed, instability and asymmetry, among others.
5
These impairments have a considerable negative impact on mobility, activities of daily living and quality of life. 1 , 6 In turn, this may then place an additional burden on healthcare resources. As such, measurement of gait characteristics is an important aspect of patient management in MS; not only may such measurements be used to monitor disease activity over time
7
but they may also enable assessment of interventions to preserve walking ability in PwMS.

The standard assessment of mobility in MS is the Kurtzke Expanded Disability Status Scale (EDSS), which is a clinician-reported outcome based on patient assessment, including distance walked by the patient and their reliance on assistive devices.
8
However, this broad assessment can fail to detect more granular changes in mobility, especially in early disease. 9 , 10 Some gait metrics may be acquired more accurately using clinical gait tests, for example, the 6-minute walk test (6MWT) and the timed 25-foot walk (T25FW) for walking speed, and the Berg Balance Scale
11
and the Timed Up and Go test
12
for balance and postural stability. However, similar to the EDSS, these tests are limited in their ability to capture subtle changes in gait. 13 – 15 In addition, they require a clinic visit, as well as space and time to perform the test, which is not always readily available, and the tests are not intended to differentiate between specific gait presentations.
16

Mobility monitoring using digital instruments allows the collection of quantitative information and a thorough characterization of gait patterns. 5 , 17 – 23 While some methods of digital mobility monitoring (e.g. electronic treadmill, video capture) may still require a clinic visit to perform the assessment, 5 , 17 – 24 wearable sensors may be used to capture gait data in unsupervised settings. Although wearables such as smartwatches, motion trackers and pressure insoles are ideal for continuous unintrusive monitoring, smartphones are more effective tools for implementing remote standardized assessments because instructions can be provided to the patient via apps while collecting the data of interest. Indeed, this approach has already been explored in a variety of neurodegenerative diseases, including Parkinson’s disease,
25
spinal muscular atrophy 26 , 27 and MS, 28 , 29 leading to highly reproducible results from remote gait assessments.

Smartphones typically contain accelerometers, gyroscopes and magnetometers, which allow measurement of various signals during a movement, including linear accelerations, angular velocities and amplitude of magnetic field.
30
The outcome measures captured from these signals – digital measures (DMs) – can then be used to reliably characterize gait. 28 , 31 – 33 In MS, studies have shown that DMs derived from a smartphone at the waist level during prescribed walking tests correlate with standard clinical measures used to quantify functional impairment and overall disability. 28 , 29 , 32 Data captured from smartphones during ‘free-walking’ (i.e. real-life walking with no instruction in terms of how to walk or where to wear the monitoring device), conversely, may be affected by confounding contextual factors (e.g. crowded spaces, changes in phone positioning), requiring extensive quality control.
34
The accuracy of single inertial measurement unit (IMU)-based gait measures appears to be improved when the sensors are located close to the centre of mass,
35
which is also generally well accepted by users.
36

Despite the clear potential clinical value of gait DMs as a monitoring tool in PwMS, 16 , 37 , 38 there is still a lack of agreement on which gait properties to measure, and how to use the resulting data to interpret gait impairment, and detect improvement or deterioration. This is partly due to a lack of normative data and limited evidence on the value of DMs compared with standard clinical measures. To better enable such evidence to be generated, it would be valuable to increase awareness of the essential steps associated with data collection, analysis and reporting, as well as to have a common framework for DMs. Previous efforts aimed to provide consensus in digital mobility monitoring by agreeing on definitions relevant to real-world gait, such as purposeful walking and the process of turning,
39
and by summarizing the greatest challenges and barriers to standard use of digital health technology in clinical trials. 40 , 41 However, the literature has not yet addressed the problem of providing a thorough contextual framework to measure and interpret specific alterations in gait patterns relevant to the neurological impairment of PwMS.

With the goal of fostering the adoption of smartphones as a standard digital health tool to measure gait in PwMS, this paper aims to describe a novel framework for capturing and interpreting key aspects of impaired gait. The aims were to provide standardized definitions of relevant gait domains and to identify corresponding DMs that can be extracted from the signals recorded with a smartphone worn at the waist level and to illustrate how they can be specifically used to quantify MS gait patterns.

Overview of methodological approach

The initial stage of the process was to identify the most common gait patterns associated with the presence of MS. Subsequently, a robust systematic literature review (SLR) was conducted on articles published in the last 10 years that measured one or more gait domains to determine how clinical manifestations of MS affect different domains of gait. These domains were pre-selected for being well-established within the scientific community and were reinforced by the findings from the SLR.

The outputs arising from this process were threefold: (1) a proposal for standardized terminology, (2) an evaluation of the impact of gait domains in disease and (3) the identification of a set of DMs directly relevant to gait domains with established acceptance within the scientific community.

Gait patterns in PwMS

Typical neurological impairments that affect gait in PwMS include paresis (weakness), spastic paraparesis, cerebellar, sensory and vestibular ataxia, extrapyramidal disorder and frontal lesions. 3 , 4 , 42 These can lead to a wide range of gait patterns, which can be distinctive in terms of the nature of the specific impairment. For example, gait ataxia, which is generally associated with instability, may lead to a veering gait and/or wobbly stance and sometimes a slow, stomping gait, with high lifting of feet. 3 , 43 People with spastic paraparesis may appear to drag each leg during walking, while those with frontal gait disorders are likely to have a short step length.
3
Full descriptions of these gait patterns are provided in Table 1 .

Table 1. Gait patterns associated with neurological impairment in multiple sclerosis. Neurological impairment Description/characteristics of impairment Resulting gait pattern Cerebellar ataxia Loss of coordination and poor dynamic balance as a result of cerebellar pathology. Gait shows signs of postural instability, poor control of limb movement, a broad-based and wobbly stance, and increased variability of leg and trunk movement and step length. 3 , 4 , 43 Additionally, an increase in double limb support time has been noted in adult patients with cerebellar ataxia, relative to healthy controls. Vestibular ataxia Loss of coordination and poor dynamic balance resulting from disturbances in the vestibular system. A ‘veering gait’ pattern resulting from the inability to walk in a straight line. Bilateral vestibular issues can lead to general instability while walking. 3 , 44 Sensory ataxia Loss of coordination and poor dynamic balance resulting from disturbances in proprioceptive function. Gait is slower and more cautious compared with cerebellar ataxia. The feet are sometimes lifted high, and gait may have a stomping quality. Visual deficits can lead to worsening of sensory ataxia because patients often rely on vision to compensate for loss of proprioception. 3 , 44 Spastic paraparesis Lower limb weakness and spasticity. Gait can appear stiff or insecure, and patients often appear to drag each leg forward when walking.
3
Patients with spastic paraparesis show reduced dorsiflexion of ankle and foot, as well as reduced knee flexion range of motion during the swing phase of the gait cycle, leading to gait asymmetry and reduced walking speed. Common patterns of spastic paraparesis can include dragging of each leg during walking,
44
as well as a common gait deviation known as ‘foot drop’. Foot drop is characterized by the inability to achieve adequate dorsiflexion and obtain a sufficient distance from the ground during the swing phase of gait. Extrapyramidal disorder Spontaneous motor activity.
42
Extrapyramidal symptoms, including gait disturbances, can occur in up to 8% of PwMS, particularly in the case of basal ganglia involvement in more advanced cases. Extrapyramidal gait discloses a typical pattern characterized by diminished or absent arm swing, forward-bent torso, stiff legs bent at the knees and hips, short or shuffling steps, turning ‘en bloc’, hesitation in starting to walk, and shuffling or freezing when encountering obstacles. While walking, the torso tends to advance ahead of the legs, impelling the patient to increase the cadence of shorts steps with increased risks of fall or collision. This sequence is known as ‘festination’.
44
Frontal lesions Frontal or parietal lesions.
3
Frontal gait disturbances are uncommon in PwMS but are recognized in advanced stages or cases with significant lesion burden. The frontal gait pattern is linked to lesions in the frontal lobes or their connections, especially in secondary progressive MS. This pattern is characterized by wide base stance, short step length and irregular walking pattern. 3 , 44 , 45
Patients with frontal gait disorders present as having difficulties standing up, inadequately adapting posture when changing position, and can have difficulties achieving a stable position. The arms may be extended laterally and arm swing may be reduced. Trunk posture may be stooped, upright or hyperextended. Inhibition of gait initiation is common. Walking is laboured and feet will shuffle or can appear to be glued to the ground (‘magnetic feet’). 3 , 44 , 45 MS: multiple sclerosis; PwMS: people with MS.

Gait patterns associated with neurological impairment in multiple sclerosis.

MS: multiple sclerosis; PwMS: people with MS.

From smartphone signals to gait DMs

Gait DMs can be captured from IMU signals after a series of pre-processing steps that are needed to mitigate the effects of noisy IMU measurements and transform the signal to match the expected input of the gait analysis algorithms ( Figure 1 ). When using a smartphone, the most critical pre-processing steps (described in Supplementary Text 1 ) entail reducing the variability of sampling rates between different devices and calibrating the sensors to correct for inaccuracies due to their electronics or configuration (e.g. bias and misalignment). To isolate the signal content of interest, noise is filtered, and the gravity signal is removed. Finally, for certain DMs to be calculated, the signals may need to be transformed into a frame aligned with the walking direction, compensating for variations in the phone positioning on the body.
34
All these steps can critically affect the calculation of the outcomes and should be reported to allow for data comparison. After pre-processing, the next step entails detecting walking bouts (i.e. segment of the signal containing continuous walking between pauses and sharp turns) on which the gait DMs are calculated. It is important to highlight that during unsupervised testing, the reliability of DMs may be affected by the quality of the walking bout from which they are derived (e.g. whether the participants were walking along a rectilinear path or not) but also by the adherence to the measurement instructions (e.g. the measuring smartphone might be moved during the measurement to answer a phone call). Some of these issues can be automatically detected and addressed with adequate signal-processing techniques. 46 , 47

Example data processing flow from walking bout to digital measures.

Example illustration of walking trajectory (top panel), IMU signals (middle panel) and digital measures derived during part of an active gait test from a lumbar mounted sensor (bottom panel). The individual started and ended at a standing still position and walked through a trajectory defined by three rectilinear walking bouts split by two approximately orthogonal turns indicated on the accelerometer (m/s 2 ) and gyroscope (rad/s) subplots. The first walking part of the recording (around 20–100 s) was performed at a higher speed than the second one (100–180 s), demonstrated by the differences in step power and step time features. Note that step regularity is less affected by such changes in walking speed because it expresses a different gait domain. In the example, step regularity is computed for the entire bout. IMU: inertial measurement unit.

Commonly adopted gait DMs range from classic spatiotemporal metrics (e.g. step/stride length, cadence, gait speed) to accelerometry-derived measures in the time and frequency domains associated with the overall quality of gait and its efficiency in terms of spent energy. 16 , 18 , 19 , 21 , 22 , 48 , 49 According to their intrinsic correlation to the specific aspects of impaired gait that they intend to measure, these DMs are typically not reported in isolation but grouped into domains used to build disease-specific conceptual gait models. In recent years, such models have been proposed in MS, 50 , 51 where domains include pace, rhythm, variability, asymmetry and forward and lateral dynamic balance. These have then been expanded to include other aspects of interest related to complexity, smoothness, endurance and fatigability when using data from longer observation, such as the 6MWT. 52 – 54 Using this approach, it has been suggested that the variability and rhythm domains explain most of the total variance in gait performance in PwMS, with variability potentially being a clinical predictor of gait impairment and falls. 2 , 7 , 55 , 56 This is further supported by the link between gait variability and neurological clinical presentations, such as spasticity, muscle weakness, impaired proprioception and balance,
57
motor control function,
58
a higher energetic cost of walking and fatigue.
56
The domains used in previous models, as above, were used as a basis for the SLR described in the next section.

SLR

Overview

An SLR was performed to determine the acceptance of the various gait DMs and of the associated gait domains when these are obtained from devices located at waist level (i.e. close to the centre of mass). For each domain, relevant electronic search strings were used to identify articles published between 2011 and 2021 that reported on studies using digital solutions in PwMS and containing the name of the relevant gait domain ( Supplementary Tables 1a–1c ). The SLR process, including date ranges and databases used, is described in Supplementary Text 2 . Predefined eligibility criteria for the SLR are presented in Supplementary Table 2 .

Selection of DMs

Based on the findings of the SLR, DMs of interest were selected for consideration in the proposed framework using the following inclusion criteria:

Selected DMs were required to have been described in at least one publication retrieved via the searches (except for classic spatiotemporal gait parameters, which are usually not clearly linked to a single domain, so were identified via targeted, ad hoc literature searches).

The publication describing the DM was required to contain an application in PwMS.

The digital health tool used to collect the signals from which the DMs were derived had to include a single IMU placed close to the centre of body mass.

Summary of evidence

In total, 1908 abstracts identified in the searches were screened against eligibility criteria ( Supplementary Table 2 ). Following screening and full-text review, a few commonly used definitions of gait emerged. Specifically, 200 articles were relevant for endurance, speed, dynamic balance or symmetry of gait, 35 for complexity, smoothness or variability of gait, and 33 for cadence of gait ( Supplementary Figures 1a–c ).

Proposed DM framework

On the basis of findings from the SLR, we have proposed a two-part framework. Part 1 of the framework provides standardized definitions for describing key gait domains in MS ( Table 2 ), and part 2 aims to identify specific DMs associated with each gait domain.

Table 2. Proposed framework part 1: gait domains and their definitions. Gait domain Definition Pace Speed and vigour of a walking bout Rhythm Gait temporal parameters Stability Ability to maintain control of the body centre of mass during walking, without excessive or variable trunk motion Symmetry Absence of lateral differences in the gait pattern Variability Fluctuations in major spatiotemporal parameters of gait across multiple strides, associated with changes in the peripheral realization of the gait pattern Smoothness Absence of intermittency of movements during gait, where intermittency is characterized by alternately accelerating and decelerating movement Complexity Presence of irregular and non-repeatable patterns within a walking period Fatigability Intra-task decline in performance over time

Proposed framework part 1: gait domains and their definitions.

For part 1, the framework addresses pace, rhythm and stability domains that are commonly and non-quantitatively assessed during a neurological examination.
44
Performance tests such as the T25FW, 2-minute walk test (2MWT) or 6MWT are used primarily to assess the pace domain, with a time- or distance-based outcome. 15 , 59 , 60 Conversely, smoothness, complexity and fatigability domains are more difficult to quantify with usual clinical examination and tests.

The findings from the literature highlighted that there is little experimental evidence on how MS affects these domains. The expected effect of MS gait impairments on selected gait domains is shown in Supplementary Table 3 , together with the level of evidence that was found in the literature from each study/publication. Supplementary Table 3 shows that all types of impairment similarly affect the gait domains, suggesting that a further level of detail is needed to capture specific neurological evolution. This is expected to be provided by DMs, as described in part 2 of our proposed framework. For each gait domain, specific DMs were identified from the SLR, and the framework describes the DMs that can be used to detect impairment, explains how they are calculated and interpreted, and includes key caveats relevant to measurement and interpretation ( Table 3 ). It is intended that multiple, complementary DMs in our framework are used to capture a fuller understanding of gait patterns, as opposed to a single DM. Note that the list has been limited to those DMs that are expected to be reliably extracted with a single IMU worn at the waist level and, as such, do not include a comprehensive set of classical spatiotemporal parameters (see Supplementary Figure 2 for an illustration of typical parameters measured during a gait cycle).

Table 3. Proposed framework part 2: recommended digital measures,
a
by gait domain. Relevant gait domain DM Description Unit Interpretation Caveats Pace Speed Ratio of distance walked to walking time. m/s Reduced speed is one of the major signs of reduced walking ability and is a common feature in ageing or in the presence of a disease. In PwMS, walking speed has been reported to decrease by 0.12 m/s for every 1-point increase in EDSS and to be 18% lower in recent fallers compared with non-fallers.
61
Highly dependent on contextual confounding factors (e.g. crowded environment, indoor/outdoor, straight/curvilinear walking) and on device location and wearing modality (e.g. loose belt). If GPS-based, only feasible when walking outdoors. Definitions about whether calculated at step, stride or walking bout level should be clear. Intensity This can be calculated as the RMS of the magnitude of the acceleration signal over a step. m/s Lower intensity is associated with reduced walking speed and energy. In PwMS, the RMS is lower in patients with higher EDSS values
50
and it was recently shown to be strongly correlated with the risk of falls.
62
Highly dependent on changes in walking speed due to contextual confounding factors, on device location, on wearing modality and on accurate detection of individual step durations. Power Ability to generate walking energy. It can be calculated as the integral of the mean-centred acceleration magnitude signal over a step.
63
m 2 /s
3
Lower power is associated with reduced walking ability. Montalban et al.
64
found a reduced step power measured during 2MWT at baseline (i.e. average measures over the first week) in PwMS when compared with healthy controls (mean (SD): 3.7 (2.2) vs 4.8 (2.6)) Montalban et al.
28
also showed a significant correlation of step power with T25FW time ( r = −0.31, p < 0.01) and with EDSS ( r = 0.28, p < 0.05). Highly dependent on changes in walking speed due to contextual factors, on device location, on wearing modality and on accurate detection of individual step durations. Rhythm Cadence Number of steps per unit of time. Steps/min Increasing/decreasing cadence is a strategy to increase/reduce speed and is one of the safety strategies adopted by PwMS, in whom it decreases by −4.4 steps per minute for every 1-point increase in EDSS.
61
Usually reliable to estimate with a single sensor. Can be significantly affected by contextual confounding factors. Stride frequency Number of strides per unit of time. Strides/min Similar to step frequency, lower values are associated with a higher likelihood of disability in PwMS.
65
Step/stride time Time between the beginning and end of a step/stride. s Longer step/stride times are associated with slower gait and reduced gait ability. A recent literature meta-analysis showed that step duration increases by + 0.04 s with every 1-point increase in EDSS.
61
If calculated as times between subsequent initial contacts, these DMs depend on the accurate detection of these gait events. Also, these events might not always be present in pathological walking (e.g. shuffling gait contact might happen as flat feet). Stability Lyapunov exponent Measure of dynamic balance control, intended as the ability to compensate for small perturbations in the walking pattern.
52
a.u. The Lyapunov exponent measures the amount of divergence from the initial walking pattern as the walking bout proceeds. Dynamic balance impairment results in a locally unstable system with a high Lyapunov exponent, corresponding to the slope of the divergence over time. Unlike healthy people, PwMS become less stable over time, as shown by larger values of the Lyapunov exponent. However, dynamic stability differences between PwMS and healthy people are only apparent after about 3 minutes of walking.
52
The walking bout should be long enough to compute this DM.
66
Trunk sway Amount of angular motion and speed in the ML and AP directions. 67 , 68 Degree or degrees/s Higher trunk sway has been reported in PwMS compared with controls,
69
indicating lower balance and stability. Ataxia would affect the trunk sway but this effect would probably only be measurable for people with high impairment. The amount of sway in the ML and AP directions might be difficult to discriminate. RMS ratio RMS of the acceleration components acc x (x = AP, ML or V) normalized by the total RMS vector.
70
a.u. The RMS ratio values exhibit common proportions for healthy people if they walk at their own preferred speed. However, in pathological gait, the RMS ratios deviate from such common proportions. RMS ratios significantly deteriorate in PwMS, and stability along AP and ML directions tends to decrease as disability levels increase.
50
Positioning of the sensor is critical and realignment with respect to gravity (see Supplementary Text 1 ) is essential when comparing data from different sessions. Symmetry Side-to-side symmetry Equivalence of ipsi- and contra-lateral values of step duration and step length. 38 , 71 s or m A difference between values measured for the two limbs can provide information about unilateral impairments and their effect on mobility. PwMS exhibit a lower side-to-side symmetry (i.e. larger asymmetry between the right and left legs) in both step duration and step length than healthy people. However, the reduced side-to-side symmetry in step duration and step length is only statistically significant among PwMS with a Modified Fatigue Impact Scale score of ⩾ 38.
72
Labelling of right and left steps using only one sensor at the waist level might be inaccurate. Step regularity Degree of similarity between subsequent steps,
73
calculated by an unbiased autocorrelation procedure applied to the acceleration signal. The autocorrelation coefficient at the first dominant period (Ad1) is an expression of the regularity of the acceleration signal between neighbouring steps. It can be calculated for individual components or for the resultant signal. a.u. Step regularity close to 1 indicates high symmetry between subsequent steps. A low value of Ad1 is the result of a generally low regularity between steps, the other a systematic asymmetry between left and right steps. In PwMS, step regularity has already been proven to be reduced in patients with mild impairment (EDSS < 3.5), with further reduction at higher EDSS scores.
50
The walking bout should be long enough to compute this DM. Realignment with respect to gravity is essential if the DM is computed along each component.
b
Swing time ratio Absolute of the log ratio of the mean left and right swing time.
38
a.u. Expressed as a percentage; if the log ratio is zero, there is perfect symmetry. The further from zero, the higher asymmetry is expected. In PwMS, the swing time ratio increases with higher levels of disability, indicating reduced symmetry in patients with severe impairment when compared with those with mild and moderate impairment. 21 , 38 , 50 Validated for straight line walking (turns removed). It requires accurate detection of swing duration, which is highly complex with a single sensor approach. Harmonic ratio Used to quantify step-to-step symmetry within a stride, which is reduced in the presence of ageing and neurological diseases. It may also carry information about more functional aspects, such as balance or stability, with higher ratios representing a more stable walking pattern.
74
a.u. Stride segmentation is needed to focus on symmetry, because without it the harmonic ratio would be affected by gait variability, given that frequencies lower than the stride would also be modelled. 75 , 76
In PwMS, the harmonic ratio in the direction of motion is already reduced in the early stages of the disease. 37 , 48 A frequency domain transformation is applied to the signal to obtain its harmonics. The harmonic ratio is defined as the ratio of the sum of even harmonics ( n = 2, 4, 6, . . .) and the sum of odd harmonics ( n = 1, 3, 5, . . .). For the ML movement, owing to its monophasic nature, the inverse ratio is used (i.e. the ratio of odd to even harmonics).
74
Variability Variability in stride/step/swing duration Can be calculated as combined within-person SD, which directly reflects step-to-step variability, or CV, 38 , 50 which is more associated with the asymmetry between the left and right steps.
50
SD (s) CV (%) Low value of SD or CV indicates low variability, reflecting the underlying neural control of gait. In PwMS, this could be associated also to spasticity with ataxic movements. In PwMS, increases in gait variability occur early on in the disease process and apparently worsen as disability increases.
7
This measure should be calculated by using long bouts. Galna et al.
77
suggested that 30 or more steps are preferable to have a reliable measure. At the same time, when looking at long walking bouts, fatigue may be associated with gait variability in PwMS, 7 , 57 potentially acting as a confounding factor. If calculated using swing duration, it requires accurate detection of swing duration, which is highly complex with a single sensor approach. Stride regularity Degree of variability between subsequent steps,
73
calculated by an unbiased autocorrelation procedure applied to the acceleration signal. The autocorrelation coefficient at the second dominant period (Ad2) is an expression of the regularity of the acceleration signal between neighbouring strides. 37 , 78 – 80 It can be calculated for individual components or for the resultant signal. a.u. Stride regularity close to 1 indicates low variability between subsequent strides. In PwMS, stride regularity has already been proven to be reduced in patients with mild impairment (EDSS < 1.5),
37
with further reduction at higher EDSS scores.
50
Orientation of the phone with respect to movement direction (i.e. AP/ML alignment) is critical for reliable estimation. This can be overcome by calculating the feature using the acceleration magnitude, which is independent of smartphone orientation (i.e.
a c c A P 2 + a c c M L 2 + a c c V 2
). Smoothness RMS jerk The RMS of the rate of change of acceleration, 50 , 81 representing a measure of how much the acceleration varies over time. m/s
3
Jerky movements will have increased variation due to abrupt stops and starts. Given its dependence on speed, jerk-based measures in PwMS have also been reported as part of the pace domain.
50
Some standard measures of jerk are sensitive to movement amplitude, duration and speed.
82

Measuring smoothness with the jerk requires identification and exclusion of the initial contacts,
b
because these would artificially increase the jerk values without lack of smoothness to the overall gait, e.g. due to rigidity. Spectral arc length Spectral arc length is defined as the negative arc length of the amplitude and frequency-normalized Fourier magnitude spectrum of the speed profile. 83 , 84 a.u. It is used to quantify the intermittency of gait. Arc length is the length of a curve: spectra with many peaks, corresponding to jerky movements, will have a smaller spectral arc length. Some literature describes it as a measure of the complexity of the spectrum, not to be confused with gait complexity. In PwMS, smoothness, as measured by spectral arc length, is lower than in healthy controls.
85
The algorithm requires two threshold values to determine the spectral arc length: (1) a frequency upper threshold that should bound the ‘normal and abnormal’ movement range and (2) a lower bound threshold that controls the trade-off between SPARC sensitivity and noise contamination.
84
These thresholds need tuning for both specific diseases and instruments. Complexity Sample entropy (SampEn) SampEn is derived by comparing subsamples of the signal recorded during one stride and calculating the conditional probability of the difference between two subsamples remaining within the same tolerance bound at the next sample time point. a.u. It measures the entropy, i.e. the amount of information in the signal, resulting from the amount of regularity and predictability in the acceleration signal. 86 , 87
It will generally increase with increasing complexity;
86
in PwMS, it has been shown that patients walk with a lower SampEn (i.e. decreased complexity of gait acceleration) than healthy individuals.
80
SampEn is independent of the time series length only when the total number of data points is large enough.
86
For smaller time series, errors grow very fast because the total number of data points is reduced. SampEn may also increase for other reasons, e.g. through the introduction of randomness. Informally, one can think of this as the amount of ‘surprise’ at observing a value at time point t
+
1 , given what we know about the signal up to time t . SampEn could be affected by walking speed, by the type of task protocol and participant instructions, and by the context (e.g. crowded environment, turning).
86
Fatigability Frequency and/or regularity decrease over a walking period Can be quantified using subsequent windows during a prolonged observation (e.g. six 1-minute windows in a 6MWT) and comparing the mean values of frequency or regularity over subsequent windows. a.u. In PwMS, step frequency and symmetry markedly decline over the course of a 6MWT 21 , 88 , 89 and this effect seems stronger in a group with moderate impairment than in a group with low impairment.
21

In addition to frequency and regularity, increased gait variability might be associated with fatigue.
57
In groups with low disability, fatigability might not be observable for short-duration active tests such as the 6MWT. 2MWT: 2-minute walk test; 6MWT: 6-minute walk test; Ad1: autocorrelation coefficient at the first dominant period; Ad2: autocorrelation coefficient at the second dominant period; a.u.: arbitrary units; AP: antero-posterior; CV: coefficient of variation; DM: digital measure; EDSS: Expanded Disability Status Scale; GPS: global positioning system; ML: medio-lateral; PwMS: people with multiple sclerosis; RMS: root mean square; SampEn: sample entropy; SD: standard deviation; SPARC: spectral arc length measure; T25FW: timed 25-foot walk; V: vertical. a DMs recommended in this table are all intended to be captured from smartphone sensor data while the smartphone is worn at waist level during a specific walking protocol. b See Supplementary Figure 2 for an illustration of the gait cycle and potential parameters measured during a gait cycle, demonstrating the contact points where realignment is needed.

Proposed framework part 2: recommended digital measures,
a
by gait domain.

2MWT: 2-minute walk test; 6MWT: 6-minute walk test; Ad1: autocorrelation coefficient at the first dominant period; Ad2: autocorrelation coefficient at the second dominant period; a.u.: arbitrary units; AP: antero-posterior; CV: coefficient of variation; DM: digital measure; EDSS: Expanded Disability Status Scale; GPS: global positioning system; ML: medio-lateral; PwMS: people with multiple sclerosis; RMS: root mean square; SampEn: sample entropy; SD: standard deviation; SPARC: spectral arc length measure; T25FW: timed 25-foot walk; V: vertical.

DMs recommended in this table are all intended to be captured from smartphone sensor data while the smartphone is worn at waist level during a specific walking protocol.

See Supplementary Figure 2 for an illustration of the gait cycle and potential parameters measured during a gait cycle, demonstrating the contact points where realignment is needed.

Active gait testing typically involves replicating commonly adopted clinical tests (e.g. 6MWT in MS). When deployed in unsupervised remote conditions, specific recommendations emerging from the literature are associated with avoiding excessive changes in walking speed by choosing terrains and paths such that the individual can walk comfortably at a desired speed. In addition, to ensure that gait is evaluated only under the ideal and steady-state condition, the minimum duration or the minimum number of steps within such walking bouts must be taken into consideration. The definition of a walking bout, encompassing both its duration and number of steps, can be adjusted according to the DMs of interest. For example, symmetry and regularity can be reliably estimated through accelerometry-based autocorrelation analyses with a minimum of four or eight steps in a steady-state condition and a minimum of 30 or 40 steps when incorporating transitional phases of gait (e.g. initiation and termination of gait).
90

Discussion

This paper proposes a framework, developed via expert collaboration and supported by literature, which is intended to serve as a reference for those interested in using gait DMs extracted from sensor data of a smartphone device worn at the waist level during active walking tests. The framework was designed with a specific focus on assessing gait impairments in PwMS and will hopefully encourage standardization and collaboration in the field, which is needed to enable validation of DMs, consistency in clinical interpretation and comparison of data across studies. The resulting framework provides literature-based definitions for relevant gait domains as well as a set of DMs intended to capture MS-related impairment across these domains. The framework aims to promote the adoption of standard gait measures obtained via smartphone-based sensors in PwMS and drive researcher consensus on gait domain digital measurement and interpretation. Although this work is focused on measuring gait impairment in MS, similar frameworks could be developed for other neurological diseases affecting gait.

Using proposed standardized terminology, the findings from the SLR showed that different neurological impairments similarly affect the different gait domains, and as such these should be investigated simultaneously to gather a clear understanding of interactions between new and pre-existing impairments (e.g. slowly evolving lesions in the frontal region might not significantly reduce the walking speed of a patient already experiencing severe cerebellar ataxia).

The next part of establishing the framework was to identify a set of DMs directly relevant to gait domains with validity and with established acceptance within the scientific community. The use of complementary DMs, rather than focusing on a single metric, is essential to identify and monitor key clinical patterns in PwMS and to quantify various control mechanisms simultaneously utilized during gait to interact with the environment safely.

As a prerequisite for DM calculation and interpretation, we also highlighted the critical signal processing steps required to treat the signals measured by the smartphones located at waist level. This provides the necessary background and context for standardization of gait data capturing, processing and reporting. Although a wide range of solutions for quantitative assessment of gait are being explored in this field, the choice of narrowing the focus to a smartphone worn at the waist level position (rather than a multi-sensor approach for joint angle kinematic analysis, for instance) and on data captured during walking tests (as opposed to free-walking) was motivated by choosing a widely available technology and by the possibility of reducing confounding contextual factors during remote monitoring. While these might limit the number of DMs that can be reliably calculated, the integration of information coming from multiple domains is expected to meet a substantial unmet clinical need. In fact, this work defined the relationship between gait domains and MS-related gait impairment, based on findings leveraging on the proposed DMs. These relationships are consistent with results from a recent review, which showed that measurements from key gait domains can successfully discriminate between PwMS and healthy controls, and are also promising candidates in terms of convergent validity, responsiveness and ecological validity.
40
Various considerations should be noted in the context of the proposed DMs, some of which we have noted within the framework itself. For example, the main approach for measuring gait complexity is to use the sample entropy of the acceleration signal,
86
even though this is an imperfect proxy for our concept of gait complexity. We also note the potential for bias in DMs as a result of data quality issues, which might arise from differences in device hardware or their malfunctioning or from a lack of concordance with test instructions. For example, the patient may not walk consistently at the instructed speed, or they might move the belt that holds the smartphone device in a fixed location.

Several other considerations related to using smartphones for digital measurement are worth highlighting. Privacy and data security are important considerations and can be potential barriers to using mobile health technologies. 91 , 92 Safeguards such as data encryption, use of firewalls, device configuration to allow patient-level control over data usage and regulatory oversight can all enhance data security and privacy.
93
In addition, differences in smartphone device hardware, configuration or operating system may contribute to variability in sensor data, challenges that can be mitigated with pre-processing protocols ( Supplementary Text 1 ). Finally, using smartphones for digital measurement in free-living environments requires patients to have access to a smartphone and literacy in basic technology. Social determinants of health may impact digital health literacy and access to smartphones, which is an important consideration for integrating these technologies into research and healthcare settings.
94

Although this study has focused on smartphone-based gait measures for the longitudinal monitoring of PwMS, our approach could be extended to include other neurological diseases. This may, in turn, support the generation of further disease-specific evidence to overcome the limitations associated with the paucity of evidence linking specific measures and gait domains to individual neurological syndromes.

In conclusion, the proposed framework is expected to foster rapid advances in terms of realizing the value of digital mobility monitoring as a necessary complementary tool to standard methods of measuring disease activity in PwMS, such as the EDSS. Our framework provides recommendations for measuring complex gait patterns using widely available technology, thus encouraging standardization of outcome measurement in MS. Further development of DMs may address some of the limitations identified in our framework, and recommendations may need to be regularly updated owing to the quickly evolving landscape of digital monitoring. However, this work sets the groundwork for improving the quality of evidence in MS and, ultimately, improving patient care.

Supplemental Material

sj-docx-1-msj-10.1177_13524585251316242 – Supplemental material for Characterizing gait in people with multiple sclerosis using digital data from smartphone sensors: A proposed framework

Supplemental material, sj-docx-1-msj-10.1177_13524585251316242 for Characterizing gait in people with multiple sclerosis using digital data from smartphone sensors: A proposed framework by Angelos Karatsidis, Lorenza Angelini, Matthew Scaramozza, Emmanuel Bartholome, Susanne P Clinch, Changyu Shen, Michael Lindemann, Claudia Mazzà, Alf Scotland, Johan van Beek, Shibeshih Belachew and Licinio Craveiro in Multiple Sclerosis Journal