Dear Sirs,

Ataxia-telangiectasia (A-T) is a neurodegenerative disease characterized by onset of ataxia in early childhood with impairments in gait, balance, and coordination [https://doi.org/10.1002/mds.27319 " data-track="click" data-track-action="reference anchor" data-track-label="link" data-test="citation-ref" aria-label="Reference 1">1]. As drug development efforts in A-T accelerate, it is increasingly important to develop objective motor assessments that can detect early disease features to reduce delays in diagnosis and support early therapies, and to monitor progression in support of clinical trials [https://doi.org/10.1002/mds.27319 " data-track="click" data-track-action="reference anchor" data-track-label="link" data-test="citation-ref" aria-label="Reference 1">1].

Current clinical rating scales to assess ataxia severity, such as the Scale for the Assessment and Rating of Ataxia (SARA), Brief Ataxia Rating Scale (BARS), and International Cooperative Ataxia Rating Scale (ICARS), are largely based on visual evaluation of motor signs as patients perform predefined motor tasks [2]. These scales, unfortunately, cannot effectively support frequent assessment and monitoring of disease progression, particularly in individuals living in underserved, rural areas and/or with limited mobility. Because these scales can be confounded by immature motor patterns present in pediatric populations [https://doi.org/10.1111/dmcn.12369 " data-track="click" data-track-action="reference anchor" data-track-label="link" data-test="citation-ref" aria-label="Reference 3">3, https://doi.org/10.1111/dmcn.13507 " data-track="click" data-track-action="reference anchor" data-track-label="link" data-test="citation-ref" aria-label="Reference 4">4], they typically require the presence of a pediatric neurologist, a rare human resource.

This study presents a novel approach for precise and objective assessment of ataxia severity in children with A-T using two ankle-worn inertial sensors during a simple gait task. Gait is a critical motor function containing important kinematic information related to ataxia, and gait-based analyses have successfully estimated ataxia severity in adults [https://doi.org/10.1212/WNL.0000000000010176 " data-track="click" data-track-action="reference anchor" data-track-label="link" data-test="citation-ref">5,https://doi.org/10.1109/TBME.2022.3142504 " data-track="click" data-track-action="reference anchor" data-track-label="link" data-test="citation-ref">6,https://doi.org/10.1002/mds.28740 " data-track="click" data-track-action="reference anchor" data-track-label="link" data-test="citation-ref" aria-label="Reference 7">7]. While most previous studies employed gait parameters or time/frequency-domain features to assess ataxic gait, our prior study in adults with ataxia [https://doi.org/10.1109/TBME.2022.3142504 " data-track="click" data-track-action="reference anchor" data-track-label="link" data-test="citation-ref" aria-label="Reference 6">6] decomposed ambulatory lower-limb movement into a series of scaled, one-dimensional submovements called movement elements (MEs) [https://doi.org/10.1038/s41598-018-29470-y " data-track="click" data-track-action="reference anchor" data-track-label="link" data-test="citation-ref" aria-label="Reference 8">8]. MEs are obtained by first removing the signal component corresponding to walking velocity from the ankle velocity time-series. Subsequently, the resulting ankle velocity time-series is segmented with respect to the body’s reference frame at its zero-crossings along each anatomical axis. MEs extracted from voluntary movements of neurologically healthy individuals show smooth, bell-shaped velocity profiles [https://doi.org/10.1038/s41598-018-29470-y " data-track="click" data-track-action="reference anchor" data-track-label="link" data-test="citation-ref" aria-label="Reference 8">8]. In contrast, MEs generated from lower-limb movements of adults with ataxia are shorter, slower, and more variable as severity worsens [https://doi.org/10.1109/TBME.2022.3142504 " data-track="click" data-track-action="reference anchor" data-track-label="link" data-test="citation-ref" aria-label="Reference 6">6]. Prior studies demonstrated that ME-based features could effectively represent motor impairments in ataxias [https://doi.org/10.1109/TBME.2022.3142504 " data-track="click" data-track-action="reference anchor" data-track-label="link" data-test="citation-ref" aria-label="Reference 6">6, https://doi.org/10.1007/s12311-021-01247-6 " data-track="click" data-track-action="reference anchor" data-track-label="link" data-test="citation-ref" aria-label="Reference 9">9, https://doi.org/10.1007/s12311-022-01385-5 " data-track="click" data-track-action="reference anchor" data-track-label="link" data-test="citation-ref" aria-label="Reference 10">10]. Hence, we hypothesized that ME-based features could objectively capture segmented and uncoordinated movements in children with various severity levels of A-T, including younger children with early disease and older children requiring substantial assistance to ambulate, with minimal effect from age-dependent, immature motor characteristics.

Twenty children with A-T (age: 3–15) and thirteen neurologically intact, healthy children (age: 2–16) participated in the study (Table 1). The two groups showed no significant differences in age (Welch’s http://www.w3.org/1998/Math/MathML"><mi>t</mi></math>" role="presentation">𝑡-test: http://www.w3.org/1998/Math/MathML"><mi>t</mi><mo>=</mo><mn>0.15</mn><mo>,</mo><mi>p</mi><mo>=</mo><mn>0.705</mn></math>" role="presentation">𝑡=0.15,𝑝=0.705). Inclusion criteria stipulated that participants were (1) genetically confirmed to have A-T or were neurologically healthy, (2) able to perform the gait task either with or without assistance, and (3) younger than 18 years old. Motor impairment severity in each child with A-T was assessed by a neurologist using the half-point version of BARS [https://doi.org/10.1002/mds.22681 " data-track="click" data-track-action="reference anchor" data-track-label="link" data-test="citation-ref" aria-label="Reference 11">11], which has been suggested for the assessment of individuals with A-T [2]. Total BARS is computed by summing sub-scores from five motor sub-tasks and ranges from 0 (normal) to 30 (most severe). For fourteen A-T participants that did not perform one or two sub-tasks, linear regression models were used to approximate total BARS based on available sub-scores. The coefficients of determination of these linear models were http://www.w3.org/1998/Math/MathML"><mn>0.93</mn><mo>&#x00B1;</mo><mn>0.07</mn></math>" role="presentation">0.93±0.07, which supports excellent accuracy. Healthy children were assumed to have a total BARS of 0. Eight children with A-T and five healthy children had two data points, and one child with A-T had three data points from participation in multiple data collection sessions separated by http://www.w3.org/1998/Math/MathML"><mn>413.9</mn><mo>&#x00B1;</mo><mn>113.0</mn></math>" role="presentation">413.9±113.0 days (Range: 236–736).

Table 1 Participant demographics and summary of walk-and-turn task performance for A-T and healthy groups

Participants were instrumented with a nine-axis inertial measurement unit (Opal, APDM Wearable Technologies) on each ankle and walked along a straight, approximately 5-m long path at their preferred speed, made a 180-degree turn, and returned to the starting position. The use of a relatively short 5-m walking distance was determined to represent ecological-valid gait patterns based on prior findings that gait in free-living settings is limited to only a few strides, lasting less than 10 s [https://doi.org/10.1088/1361-6579/38/1/n1 " data-track="click" data-track-action="reference anchor" data-track-label="link" data-test="citation-ref" aria-label="Reference 12">12]. Most participants repeated this motor task twice or three times continuously (see Table 1). Ten A-T participants who could not walk independently received assistance from a clinician during the task. Three children held a clinician’s hand, and six received two-hand assistance. One child received two-hand assistance for the first visit and held a hand for the second visit. The data from assisted walking were included in the analysis because we hypothesized that ME-based features, which focus on analyzing kinematic characteristics of voluntary, limb-specific movements apart from the assistance provided to the upper body, could capture inherent information related to disease severity. Data processing methods are described in the Supplementary Methods and our prior paper [https://doi.org/10.1109/TBME.2022.3142504 " data-track="click" data-track-action="reference anchor" data-track-label="link" data-test="citation-ref" aria-label="Reference 6">6]. The study protocol was approved by the Institutional Review Board at Massachusetts General Hospital, and all participants provided written informed assent and/or consent.

Regression models were trained to estimate total BARS and the BARS gait sub-score using ME-based features (see Supplementary Materials and Table S3). Total BARS estimates had strong agreement with clinician scores, with a root mean square error (RMSE) of 3.69 BARS points and a coefficient of determination R2 = 0.82 (Pearson’s http://www.w3.org/1998/Math/MathML"><mi>r</mi><mo>=</mo><mn>0.91</mn><mo>,</mo><mi>p</mi><mo>&lt;</mo><mn>0.001</mn><mo stretchy="false">)</mo></math>" role="presentation">𝑟=0.91,𝑝<0.001) across all participants (Fig. 1a). When considering only A-T participants, the estimations had an RMSE of 4.05 with R2 = 0.45 (Pearson’s http://www.w3.org/1998/Math/MathML"><mi>r</mi><mo>=</mo><mn>0.76</mn><mo>,</mo><mi>p</mi><mo>&lt;</mo><mn>0.001</mn><mo stretchy="false">)</mo></math>" role="presentation">𝑟=0.76,𝑝<0.001). Gait severity estimation also showed comparable agreement with clinician scores with an RMSE of 1.09 BARS points and R2 = 0.80 (Pearson’s http://www.w3.org/1998/Math/MathML"><mi>r</mi><mo>=</mo><mn>0.89</mn><mo>,</mo><mi>p</mi><mo>&lt;</mo><mn>0.001</mn><mo stretchy="false">)</mo></math>" role="presentation">𝑟=0.89,𝑝<0.001) across all participants (Fig. 1b). Considering only A-T participants, the gait BARS estimates had an RMSE of 1.25 with R2 = 0.51 (Pearson’s http://www.w3.org/1998/Math/MathML"><mi>r</mi><mo>=</mo><mn>0.76</mn><mo>,</mo><mi>p</mi><mo>&lt;</mo><mn>0.001</mn><mo stretchy="false">)</mo></math>" role="presentation">𝑟=0.76,𝑝<0.001). The longitudinal changes were analyzed for nine children with A-T who completed multiple data collection sessions to evaluate agreement between clinician-scored and estimated BARS (Fig. 2). The changes in the clinician-evaluated vs. estimated BARS showed significant correlations for total BARS (Pearson’s http://www.w3.org/1998/Math/MathML"><mi>r</mi><mo>=</mo><mn>0.68</mn><mo>,</mo><mi>p</mi><mo>=</mo><mn>0.032</mn></math>" role="presentation">𝑟=0.68,𝑝=0.032; Fig. 2a) and gait sub-score (Pearson’s http://www.w3.org/1998/Math/MathML"><mi>r</mi><mo>=</mo><mn>0.78</mn><mo>,</mo><mi>p</mi><mo>=</mo><mn>0.008</mn></math>" role="presentation">𝑟=0.78,𝑝=0.008; Fig. 2b), indicating a strong relationship in capturing severity changes over time.

Fig. 1
figure 1

BARS estimation results and relationships between age and ataxia severity. A-T participants walking without assistance and healthy participants are denoted as blue and green dots, respectively. A-T participants walking with assistance are denoted as blue triangles. The black solid line indicates perfect regression (http://www.w3.org/1998/Math/MathML"><mi>y</mi><mo>=</mo><mi>x</mi></math>" role="presentation">𝑦=𝑥). The black dotted line indicates a least-squares regression line of estimated BARS scores. The blue and green solid lines indicate a least-squares regression line of A-T and healthy groups’ data points, respectively. a Clinician-scored total BARS vs. total BARS estimated by the proposed model. b Clinician-scored gait sub-scores vs. estimated gait sub-score. c Age vs. clinician-scored total BARS for A-T participants. d Age vs. total BARS estimated by the proposed model. A-T participants were color-coded with clinician-scored total BARS. Darker colors indicate higher total BARS evaluated by a clinician. http://www.w3.org/1998/Math/MathML"><msup><mrow class="MJX-TeXAtom-ORD"><mi>R</mi></mrow><mrow class="MJX-TeXAtom-ORD"><mn>2</mn></mrow></msup></math>" role="presentation">𝑅2: coefficient of determination; RMSE: root mean square error

Fig. 2
figure 2

Longitudinal differences in the clinician-scored and estimated BARS. The black solid line indicates perfect correlation (y = x). The black dotted line indicates a least-squares regression line of BARS differences. a Clinician-scored total BARS differences vs. estimated total BARS differences. b Clinician-scored gait sub-score differences vs. estimated gait sub-score differences

Correlations between age and both clinician-scored BARS and sensor-based severity estimates were analyzed to investigate the influence of age. Total BARS scored by clinicians showed a significant correlation with A-T participants’ ages (Pearson’s http://www.w3.org/1998/Math/MathML"><mi>r</mi><mo>=</mo><mn>0.63</mn><mo>,</mo><mi>p</mi><mo>&lt;</mo><mn>0.001</mn></math>" role="presentation">𝑟=0.63,𝑝<0.001; Fig. 1c), which reflects population-level disease progression. The estimated total BARS also showed a significant correlation with A-T participants’ ages (Pearson’s http://www.w3.org/1998/Math/MathML"><mi>r</mi><mo>=</mo><mn>0.46</mn><mo>,</mo><mi>p</mi><mo>=</mo><mn>0.011</mn></math>" role="presentation">𝑟=0.46,𝑝=0.011; Fig. 1d). While estimated BARS in healthy participants showed a negative trend with age, there was not significant correlation (Pearson’s http://www.w3.org/1998/Math/MathML"><mi>r</mi><mo>=</mo><mo>&#x2212;</mo><mn>0.32</mn><mo>,</mo><mi>p</mi><mo>=</mo><mn>0.190</mn></math>" role="presentation">𝑟=−0.32,𝑝=0.190). The difference in estimated BARS between A-T and healthy participants increased as age increases. Specifically, the difference in estimated total BARS score between A-T and healthy in the older age group (9–16 years old) showed a larger effect size (Cohen’s http://www.w3.org/1998/Math/MathML"><mi>d</mi><mo>=</mo><mn>3.08</mn></math>" role="presentation">𝑑=3.08) than in the younger age group (2–8 years old, Cohen’s http://www.w3.org/1998/Math/MathML"><mi>d</mi><mo>=</mo><mn>2.69</mn></math>" role="presentation">𝑑=2.69) (see Supplementary Table S4). Nonetheless, strong separation was observed between A-T and healthy participants regardless of age, with the separation increasing as expected for older individuals.

ME feature-based regression models were compared with benchmark models trained on conventional gait parameter-based features. The detailed comparison methods and benchmark results from all combinations of feature selection and machine learning algorithms are described in the Supplementary Results. The gait parameter-based model that exhibited the best performance showed only moderate agreement with the clinician-scored total BARS, as evidenced by a root mean square error (RMSE) of 5.08 BARS points and R2 = 0.65 (Pearson’s http://www.w3.org/1998/Math/MathML"><mi>r</mi><mo>=</mo><mn>0.81</mn><mo>,</mo><mi>p</mi><mo>&lt;</mo><mn>0.001</mn></math>" role="presentation">𝑟=0.81,𝑝<0.001; Fig. 3a). In summary, the gait sub-score estimates had an RMSE of 1.38 with R2 = 0.68 (Pearson’s http://www.w3.org/1998/Math/MathML"><mi>r</mi><mo>=</mo><mn>0.82</mn><mo>,</mo><mi>p</mi><mo>&lt;</mo><mn>0.001</mn></math>" role="presentation">𝑟=0.82,𝑝<0.001; Fig. 3b). Hence, ME feature-based model showed stronger agreement with clinician-scored BARS compared to the model trained on conventional gait parameter-based features.

Fig. 3
figure 3

BARS estimation results of the model trained on conventional gait parameter-based features. A-T participants walking without assistance and healthy participants are denoted as blue and green dots, respectively. A-T participants walking with assistance are denoted as blue triangles. The black solid line indicates perfect regression. The black dotted line indicates a least-squares regression line of estimated BARS scores. The blue and green solid lines indicate a least-squares regression line of A-T and healthy groups’ data points, respectively. a Clinician-scored total BARS vs. total BARS estimated by the benchmark model trained on conventional gait parameters. b Clinician-scored gait sub-scores vs. estimated gait sub-score

This study demonstrates a novel method to objectively capture overall ataxia severity in children with A-T based on ankle inertial data and a simple gait task. We envision that this sensor-based estimation model could facilitate frequent, objective, and accurate remote assessments for individuals with A-T and other neurological movement disorders across a wide range of severity levels.

We identified four ME features that were most frequently selected by the estimation model, all of which showed significant correlations with total BARS (see Supplementary Table S5). Three of these features were extracted from the anteroposterior axis and showed that lower-limb movements in more severe A-T participants were decomposed into smaller and more variable MEs. These features represented uncoordinated and segmented limb-specific movements during walking. The feature from the mediolateral axis showed increased variability in the speeds of MEs for more severe A-T participants, capturing the instability and imbalance of lower-limb movements in the mediolateral direction during walking. These findings are consistent with previous studies reporting that increased gait variability and slower walking speed are distinctive kinematic characteristics of ataxic gait [https://doi.org/10.1212/WNL.0000000000010176 " data-track="click" data-track-action="reference anchor" data-track-label="link" data-test="citation-ref" aria-label="Reference 5">5, https://doi.org/10.1002/mds.28740 " data-track="click" data-track-action="reference anchor" data-track-label="link" data-test="citation-ref" aria-label="Reference 7">7].

There are limited studies focused on estimating motor impairment severity in pediatric ataxias using gait analysis [https://doi.org/10.1016/j.cmpb.2020.105705 " data-track="click" data-track-action="reference anchor" data-track-label="link" data-test="citation-ref" aria-label="Reference 13">13, https://doi.org/10.1007/s12311-021-01348-2 " data-track="click" data-track-action="reference anchor" data-track-label="link" data-test="citation-ref" aria-label="Reference 14">14]. Milne et al. analyzed gait parameters using instrumented walkways and showed reduced walking velocity and increased base of support variability in children with Friedreich Ataxia [https://doi.org/10.1007/s12311-021-01348-2 " data-track="click" data-track-action="reference anchor" data-track-label="link" data-test="citation-ref" aria-label="Reference 14">14]. However, this work used instrumented walkways requiring specialized infrastructure. In contrast, the proposed method employed wearable sensors and a sub-movement-based analysis approach that may support continuous and remote monitoring of ataxia severity in real-life environments.

Our ME-based method showed potential to ataxia motor impairments with minimal influence of age-dependent, immature developmental motor characteristics. Prior studies investigated the influence of age and motor development on several ataxia rating scales [https://doi.org/10.1111/dmcn.12369 " data-track="click" data-track-action="reference anchor" data-track-label="link" data-test="citation-ref" aria-label="Reference 3">3, https://doi.org/10.1111/dmcn.13507 " data-track="click" data-track-action="reference anchor" data-track-label="link" data-test="citation-ref" aria-label="Reference 4">4]. Brandsma et al. assessed different age groups of healthy children using the BARS, SARA, and ICARS and reported that these scales showed significant age-dependency up to 10–12.5 years old [https://doi.org/10.1111/dmcn.12369 " data-track="click" data-track-action="reference anchor" data-track-label="link" data-test="citation-ref" aria-label="Reference 3">3]. Thus, detecting early motor impairments and measuring disease-related change in the setting of motor development is challenging using these scales, including BARS. On the other hand, although the estimates from proposed model showed a negative trend similar to these scales, they did not show significant correlation with healthy individuals’ ages (http://www.w3.org/1998/Math/MathML"><mi>r</mi><mo>=</mo><mo>&#x2212;</mo><mn>0.32</mn><mo>,</mo><mi>p</mi><mo>=</mo><mn>0.190</mn></math>" role="presentation">𝑟=−0.32,𝑝=0.190) In addition, the model showed significant differences in estimated total BARS between healthy and A-T participants regardless of age. However, the larger sample size of healthy participants is needed to demonstrate the age-dependency of ME features more clearly.

This study has some limitations. First, the sample size was relatively small as A-T is a rare disease. Second, the A-T and healthy groups were not sex-matched. However, age is a more critical factor than sex in gait development [https://doi.org/10.1002/mds.27319 " data-track="click" data-track-action="reference anchor" data-track-label="link" data-test="citation-ref" aria-label="Reference 1">1]. Third, the data were collected from a single center and controlled setting. More data from multi-institutions and real-world settings are necessary to demonstrate the extensibility of the proposed approach. Furthermore, the use of a relatively short walking distance may hinder accurate measures of conventional gait parameters and a fair comparison between our approach and the model trained on gait parameters. Finally, more longitudinal data will be necessary for additional investigation of this approach's sensitivity to disease progression.