Analytical and clinical validation of step counting method in people living with amyotrophic lateral sclerosis

1. Background

Accelerometer-based step counts have emerged as a valuable indicator of physical functioning in both clinical and epidemiological research, offering a scalable, low-burden approach to monitoring mobility across diverse populations. Gait speed has been referred to as the “sixth vital sign” due to its strong association with functional status, fall risk, hospitalization, and mortality (1–3). In people living with amyotrophic lateral sclerosis (ALS), a progressive neurodegenerative condition characterized by motor neuron loss, functional decline, and ultimately respiratory failure, capturing ambulatory function is particularly critical. Gait impairment is often among the earliest manifestations of ALS (4–6), and progressively worsens, leading to decreased mobility, increased fall risk, and compounding health complications (7–9).

Despite its clinical relevance, mobility assessment in ALS has traditionally relied on infrequent and coarse-grained outcome measures such as the ALS Functional Rating Scale-Revised (ALSFRS-R), which may fail to detect subtle yet meaningful changes in walking behavior (10–13). As therapeutics evolve, there is an urgent need for sensitive, responsive, and ecologically valid outcome measures capable of detecting functional improvements and disease progression (14–16).

Digital health technologies (DHTs), including wearable accelerometers and smartphones, enable continuous, high-frequency assessment of motor function in both clinical and free-living environments (17–20). However, existing step counting algorithms are often developed for healthy populations and may perform poorly in individuals with slower, asymmetric, or variable gait patterns typical to ALS (21,22). Furthermore, the clinical validity of step counts as a digital measure of disease progression is ALS remains unexplored, despite successful use of similar endpoints in other neuromuscular diseases (23).

To address these gaps, we propose a simple and robust step counting algorithm that leverages ankle-worn accelerometer data. The algorithm builds on fundamental features of walking, i.e., periodicity, intensity, and cadence, and is designed to accommodate the heterogeneity, speed fluctuations, and asymmetries characteristic to ALS gait. Because our purpose was to develop an algorithm that quantifies gait in individuals that may use assistive devices, we relied solely on ankle-worn sensors. While wrist-worn accelerometers are commonly used in other populations (24–26), their application in ALS is limited due to interruptions in typical arm swings, that can lead to substantial errors when participants begin to use assistive devices.

We validate the method using several publicly available datasets of healthy adults as well as clinical and free-living data from individuals with ALS, capturing a range of walking speeds, assistive device use, and disease severity. Furthermore, we test the accuracy of the method to distinguish walking from other common daily activities (27). This multi-context validation framework enables analytical and clinical evaluation of algorithmic performance and supports the feasibility of ankle-worn step counting as a sensitive, interpretable, and scalable digital endpoint for mobility in ALS.

2. Methods

2.2. Algorithm description

The proposed step counting method leverages the observation that walking consists of repetitive, oscillatory movements characterized by distinct periodicity and amplitude, particularly when captured by worn accelerometers. These periodic features arise from biomechanical events such as heel strikes, which produce transient, high-amplitude changes in acceleration along the axis aligned with the shin (Figure 1A). To discriminate walking from other daily activities and to detect individual steps in heterogeneous free-living conditions, we apply a continuous wavelet transform (CWT) to the raw accelerometer signal. This transform decomposes the signal into a time-frequency representation, allowing for the detection of localized periodic events while being robust to signal non-stationarity (Figure 1B). By focusing on the signal’s vertical component, i.e., where heel strikes are reflected as sharp spikes, our method provides a robust framework for detecting gait in individuals with varying movement patterns and gait abnormalities.

• Download figure
• Open in new tab

Figure 1.

Framework for detecting heel strikes using ankle-worn accelerometers. Signal from vertical axis (A) is decomposed using a continuous wavelet transform which uncovers sharp spikes corresponding to individual heel strikes (B). Local peaks in frequency-wise summation of wavelet coefficients help identify heel strikes locations within the signal s(C). Identified locations correspond to heel strikes from one foot (D); assuming another foot making the same number of steps, total step counts are calculated as number of heel strikes multiplied by 2.

In our method, the CWT is implemented using the generalized Morse wavelet, defined as parameterized with a symmetry constant γ = 3, ensuring zero skewness, and a time-bandwidth product P² = 10, which balances time and frequency localization (Supplementary Figure 1). This configuration allows the wavelet coefficients to be particularly sensitive to abrupt changes in signal amplitude, such as those produced during heel strikes.

In our approach, wavelet coefficients (product of CWT) are computed across a physiologically relevant frequency range (up to 5Hz), and their values are summed at each time point to enhance the signal-to-noise ratio of periodic gait-related events (Figure 1C). We identify local peaks in the resulting time series of summed coefficients and classify them as heel strikes if they exceed a fixed threshold set at 0.1 (Figure 1D).

To prevent spurious peaks (e.g., from isolated stomps or noise) from being misclassified as steps, we apply additional temporal time constraints. Specifically, our algorithm enforces that timestamps t_i, t_i+1, …, t_i+k of consecutive heel strikes from the same leg must be separated by time intervals Δ_i= t_i+1 − t_i with physiologically plausible Δ_min= 0.85 seconds and Δ_max= 2.5 seconds, corresponding to a very fast (about 2.35 steps per second) and a very slow walking speed (0.8 steps per second), respectively. Furthermore, the difference between consecutive between-strike intervals δ_i = |Δ_i+1 − Δ_i| should not exceed 0.5 seconds to further exclude implausible or artifact-driven detections.

In the analytical validation, we focus on two gait-related measures: walking activity time and step counts. Walking activity is identified for all nonsingular heel strikes if ∀_j∈ {i, …, i + k − 1}, t_i+1 − t_i ≤ 3 seconds. Steps were included in the step count outcome measure if they were classified as walking activity. Assuming another foot making the same number of steps, total step counts were computed as number of detected heel strikes multiplied by 2. Walking activity recognition rate was investigated to determine potential activities of daily living that might inflate step counts in free-living observations.

3. Results

The aggregated dataset used to validate the method comprised of data from 452 individuals, including 138 healthy individuals and 314 individuals diagnosed with ALS. Data from 93 individuals were used to determine the accuracy to recognize walking activity, data from 57 (45 healthy, 12 ALS) individuals were used for analytical validation of step counting, while data from 302 individuals were used to clinical validation in ALS.

Altogether we analyzed data from 80,989 hours of activity, including over 21 hours for analytical validation of walking recognition (8.75 h of walking activities, 12.65 h of non-walking activities), 15.57 hours for analytical validation of step counting, and over 80,952 hours (about 3373 days) for clinical validation in ALS.

3.1. Analytical validation

3.1.1 Walking recognition accuracy

Walking recognition was highly accurate for a wide range of walking activities, such as normal walking (e.g., in SPADES the mean sensitivity was 0.995, CI = [0.997, 0.999]), ascending stairs (0.947 [0.908, 0.987]), descending stairs (0.915 [0.847, 0.983]), or walking on a treadmill at 1 mph (0.983 [0.978, 0.987]) (Figure 2A). Our method also accurately recognized non-walking activities, including repetitive activities, such as cycling (e.g., in DaLiAc mean specificity was equal 1 [1, 1]) or running (0.981 [0.951, 1]) (Figure 2B). Our results showed that car driving data collected in IUWCD, often represented as high-frequency vibrations (34), was entirely classified as non-walking (1 [1, 1]) (Figure 2C). Decreased specificity was observed for sweeping the floor (in DaLiAc 0.851 [0.815, 0.887]) and vacuuming (0.877 [0.814, 0.940]). However, it is unknown to the authors whether these activities in fact consisted of entirely non-walking movements. Finally, systematically decreased specificity was reported for jumping (in SPADES 0.669 [0.545, 0.792], in DaLiAc 0.852 [0.770, 0.934]).

• Download figure
• Open in new tab

Figure 2.

Sensitivity (blue) and specificity (red) estimated for walking recognition in populations of ambulatory participants investigated in SPADES (A), DaLiAc (B), and IUWCD (C) studies.

3.1.2. Step counting accuracy

Altogether, according to the ground truth, participants included in PedEval performed a mean (SD) of 1045.73 (85.81) steps during continuous, 681.91 (88.37) steps during semi-continuous, and 180.81 (16.65) steps during discontinuous walking activities, respectively. The corresponding estimations calculated using our method were 1048.15 (84.89) steps for continuous, 660.67 (93.45) steps for semi-continuous, and 190.29 (18.29) steps for discontinuous walking activities.

Using Bland-Altman analysis, we determined that the mean bias (LoA) was equal to 0.42 (−11.60, 12.44) or 0.04 % for continuous walking (Figure 3A), −4.33 (−61.81, 53.15) or 0.64 % for semi-continuous walking (Figure 3B), and 1.36 (−44.45, 47.17) or 0.76 % for discontinuous walking (Figure 3C), indicating excellent reliability of the algorithm on a population level. The mean (SD) ALSFRS-R total score of individuals participating in PFALS dataset was equal to 34.9 (4.1), mean response to Q7 was equal to 3.2 (0.7), Q8 2.5 (0.7), and Q9 1.5 (1.4).

• Download figure
• Open in new tab

Figure 3.

Bland-Altman plots on step counting accuracy in three walking scenarios investigated in PedEval dataset (continuous walk A, semi-continuous walk B, and discontinuous walk C) and one walking scenario investigated in PFALS dataset (D). In each scenario, ground truth (GT) was assessed by a visual observer. Black dots correspond to observations from healthy individuals, grey dots correspond to observations from individuals diagnosed with ALS.

In PFALS, ALS participants and healthy controls performed a similar number of steps, i.e., 1269 and 1249, respectively. In this controlled experiment, our method accurately counted steps regardless of participants’ health conditions. In healthy participants, mean bias (LoA) was equal to −0.75 (−5.06, 3.56) or −1.75 %, while in ALS participants mean bias was equal to 1.40 (−4.28, 7.08) or 0.4 %. Cumulatively, mean bias (LoA) was equal to 0.44 (54.01, 5.90) or 0.56 %, suggesting accurate step counting performance across the two populations (Figure 3D).

3.2. Clinical validation

Clinical validation data were collected between September 2014 and January 2023 on 305 participants diagnosed with ALS. There were 251 participants who met the data inclusions criteria (at least 3 days with a minimum wear-time of 16 hours). The mean (SD) data collection period was equal to 13.44 (8.44) days, while step counts were estimated for a mean (SD) of 5.13 (1.89) days. At baseline, study participants reported a mean (SD) ALSFRS-R total score equal to 40.21 (5.58), while their mean (SD) responses to Q7, Q8, and Q9 were equal to 3.31 (0.80), 2.96 (0.88), and 2.76 (1.20), respectively.

Using step counts, we observed statistically significant differences between participants reporting normal gross motor functioning and those reporting impairments in this domain (Figure 4). For example, participants reporting “normal” gait (Q8) performed a mean of 3984 [3537, 4432] step counts per day, significantly more compared to those reporting “early ambulation difficulties” (3197 [2837, 3557], p = 0.009), and “walking with assistance” (1283 [1063, 1503], p = <0.001). Importantly, the algorithm estimated almost no steps for those reporting “nonambulatory functional movement” (81 [12, 150], p = <0.001), indicating that it is robust to non-walking activities of daily living.

• Download figure
• Open in new tab

Figure 4.

Mean and 95% CI of step counts stratified by responses to ALSFRS-R questions on gross motor functioning subscale (Q7—Q9) along with p-values of two sample t-test (i.e., step counts observed in participants reporting normal functioning and those reporting decreased functioning). Values displayed next to the mean step counts indicate group size reporting a given functional status.

4. Discussion

This study demonstrates the validity and applicability of a step-counting method for quantifying ambulatory activity in individuals with ALS. By leveraging a continuous wavelet transform to detect periodic peaks corresponding to heel strikes, our algorithm offers a robust and interpretable approach for step detection using accelerometers placed on a foot or ankle, body location which appears favorable to capture gait events in individuals using assistive devices (Supplementary Figure 2 panels B—D). We determined the utility of the method following principles of the V3 framework, which underscores the need for analytical and clinical validation of sensor-based digital health technologies in populations with changing health status (35,36).

Our analytical validation was conducted in two ways, (1) focusing on the accuracy of the method to identify periods of walking activity and (2) counting steps within these periods. The walking recognition accuracy was verified using three independent, publicly available datasets. It was determined to be highly sensitive to walking activities and highly specific to most non-walking activities defined by the respective dataset authors. Notably, a marginally decreased specificity was observed for selected household activities, such as vacuuming or sweeping the floor. These activities, however, might include a small number of steps and a certain level of misclassification is warranted. Importantly, the method remained insensitive to activities that might present in the data as repetitive spikes, such as cycling or driving a car.

The proposed approach also demonstrated accurate estimation of step counts both in controlled and semi-controlled assessments. In each investigated scenario the mean bias was close to zero, while wider levels of agreement were calculated for semi-controlled tasks. This variation may be attributed to complex structure of movement patterns, such as hesitations, pauses, or weight-bearing steps that can produce signal artifacts resembling step-like events. Additionally, the absence of a standardized protocol for specific transitions, such as turning around, navigating obstacles, or climbing stairs, may have introduced variability in recorded signals that contributed to reported misclassification. These factors highlight inherent challenges of step counting in naturalistic settings where movements often deviate from steady, linear gait patterns.

For clinical validation, we evaluated whether the method provides clinically meaningful insights into mobility patterns in people living with ALS, a population characterized by progressive motor decline and high interindividual variability in gait. Our results show that the method can differentiate step counts between individuals with normal gait, early-stage ambulatory difficulties, and those requiring assistance. This distinction aligns with known clinical trajectories in ALS and demonstrates the method’s ability to detect functionally and clinically relevant changes in walking behavior, which is a critical feature for a potential digital outcome measure used in a therapeutic evaluation or clinical trial. According to our knowledge, this is the first study to assess step count levels in a large-scale free-living observation involving individuals with ALS. In our previous study (5), we enrolled twenty people living with ALS whose baseline daily step counts were estimated as 1871 [1137, 2605] steps, substantially lower than the step counts reported in this study. The lower step counts in the earlier cohort may reflect participants in a more advanced disease stage, as indicated by a substantially lower average ALSFRS-R total score (mean 31.40) compared to the group reported here (mean 40.21).

Our findings align with prior work which emphasizes that ankle-worn devices are particularly suited for step detection in populations with altered gait patterns, including those with neurodegenerative diseases and requiring assistance. It is worth highlighting that our method detects heel strikes using data collected with a low sampling frequency (10 Hz), in contrast to many commercial algorithms that require higher sampling frequencies to achieve comparable accuracy (22,37). This property is especially important in free-living observations of individuals with ALS, where slower gait speeds, the use of assistive devices (21), or variable walking surfaces (27) may introduce artefacts that hinder approaches relying on time-domain signal analyses, and where employing more sophisticated devices may not be feasible. The clear distinction in step counts observed between participants with early ambulatory impairment and those requiring assistance (3197 [2837, 3557] vs. 1283 [1063, 1503]) underscores the method’s sensitivity to clinically relevant differences in mobility status, reinforcing its utility as a candidate digital endpoint for ALS studies.

Several limitations should also be acknowledged. First, the method assumes a basic level of periodicity and intensity in the accelerometer signal, which may lead to errors in contexts involving sporadic or irregular stepping, such as turning, standing, or shuffling. This challenge is compounded by the lack of consensus on how to define “walking” in real-world settings, as it remains unclear whether a single step, two steps, or a continuous sequence of steps should be classified as walking (27). Second, due to the secondary nature of this analysis and the absence of a standardized turning protocol in the original datasets (e.g., in PFALS), we were unable to assess the algorithm’s performance specifically during transitions such as straight walks versus turning. Third, while the algorithm provided meaningful indications across the included ALS datasets, the free-living observations did not include ground truth step counts, limiting our ability to assess its accuracy under naturalistic conditions. Future research should explore these limitations but also consider the integration of additional sensor modalities (e.g., gyroscopes, magnetometers) to capture a broader range of gait-related metrics, including stride length, cadence variability, and measures of balance.

5. Conclusions

The reliance of our method on the fundamental properties of the gait signal, recorded from the ankle, rather than the wrist, supports its applicability across diverse measurement conditions. By validating the algorithm across multiple independent datasets, including those collected in controlled, semi-controlled, and free-living settings, we demonstrated its reproducibility and potential for use in real-world studies. Importantly, the algorithm accurately counts steps and walking activity in healthy participants and those with ALS who have gait disturbances – even those requiring assistance. This overcomes a major existing shortcoming of gait quantification in ALS trials – the floor effect driven by the inability to accurately count steps as gait deteriorates. With this method, researchers can more assuredly employ accelerometry in clinical trials. We have made our algorithm open-source to enhance transparency and support adoption by researchers in ALS and other populations. The method can also be adapted for other devices, or research questions, as needed. Together, these attributes position the method as a promising foundation for future development and application of digital outcome measures on mobility in ALS and other diseases characterized by deterioration of gait.

7. Declarations