Scholarlink

Background <p>Spinocerebellar ataxia type 3 (SCA3) and the cerebellar subtype of multiple system atrophy (MSA-C) both manifest with cerebellar ataxia and postural instability in the early stages, leading to considerable clinical overlap and frequent misdiagnosis or delayed intervention. Current diagnostic tools offer limited support when genetic confirmation is unavailable or neuroimaging findings are inconclusive. Clinical rating scales, while commonly used, rely heavily on subjective judgment and lack sensitivity to dynamic monitoring. Therefore, this study aims to develop an interpretable machine learning model based on gait and postural features acquired through wearable sensors, in order to assist in distinguishing SCA3 from MSA-C and to explore their potential as digital biomarkers for clinical application.</p> Methods <p>This study included 74 individuals diagnosed with SCA3, 43 individuals with multiple system atrophy of MSA-C, and 45 age-matched healthy controls (HC). All participants were in the mild to moderate stages of disease. Gait and postural control data were collected using a wearable smart insole system, and 98 digital features were extracted. Classification models were developed using XGBoost, LightGBM, random forest, and logistic regression. To enhance model interpretability, SHapley Additive exPlanations (SHAP) were applied to quantify the contribution of each feature and to identify the most informative digital biomarkers.</p> Result <p>Among all models, LightGBM achieved the highest accuracy (95.91%) and ROC-AUC (0.9962). SHAP analysis identified Cadence CV and center of pressure (COP) features under eyes-closed (EC) conditions as top contributors across all models, supporting their potential as digital biomarkers. Gait and postural features provided complementary information. Even at mild to moderate disease stages, individuals with SCA3 and those with MSA-C showed distinguishable motor patterns. Specifically, individuals with SCA3 exhibited greater medio-lateral (ML) postural instability under visual deprivation, while those with MSA-C showed increased variability in gait rhythm, indicating distinct movement phenotypes.</p> Conclusion <p>The integration of gait and postural stability features with interpretable machine learning models enables effective differentiation between SCA3 and MSA-C. It also facilitates the identification of clinically relevant digital biomarkers, supporting early diagnosis and personalized rehabilitation planning in cerebellar ataxia.</p> <p><i>Clinical trial number</i> NTC04010214.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Interpretable machine learning for differentiating SCA3 and MSA-C using gait and postural features from wearable sensors

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年12月20日

查看全文
  • Yuanyuan Xiao,
  • Kailiang Luo,
  • Yue Zhang,
  • Wanli Zhang,
  • QiKui Sun,
  • Bingwei He,
  • ShiRui Gan,
  • Xinyuan Chen

摘要

Background

Spinocerebellar ataxia type 3 (SCA3) and the cerebellar subtype of multiple system atrophy (MSA-C) both manifest with cerebellar ataxia and postural instability in the early stages, leading to considerable clinical overlap and frequent misdiagnosis or delayed intervention. Current diagnostic tools offer limited support when genetic confirmation is unavailable or neuroimaging findings are inconclusive. Clinical rating scales, while commonly used, rely heavily on subjective judgment and lack sensitivity to dynamic monitoring. Therefore, this study aims to develop an interpretable machine learning model based on gait and postural features acquired through wearable sensors, in order to assist in distinguishing SCA3 from MSA-C and to explore their potential as digital biomarkers for clinical application.

Methods

This study included 74 individuals diagnosed with SCA3, 43 individuals with multiple system atrophy of MSA-C, and 45 age-matched healthy controls (HC). All participants were in the mild to moderate stages of disease. Gait and postural control data were collected using a wearable smart insole system, and 98 digital features were extracted. Classification models were developed using XGBoost, LightGBM, random forest, and logistic regression. To enhance model interpretability, SHapley Additive exPlanations (SHAP) were applied to quantify the contribution of each feature and to identify the most informative digital biomarkers.

Result

Among all models, LightGBM achieved the highest accuracy (95.91%) and ROC-AUC (0.9962). SHAP analysis identified Cadence CV and center of pressure (COP) features under eyes-closed (EC) conditions as top contributors across all models, supporting their potential as digital biomarkers. Gait and postural features provided complementary information. Even at mild to moderate disease stages, individuals with SCA3 and those with MSA-C showed distinguishable motor patterns. Specifically, individuals with SCA3 exhibited greater medio-lateral (ML) postural instability under visual deprivation, while those with MSA-C showed increased variability in gait rhythm, indicating distinct movement phenotypes.

Conclusion

The integration of gait and postural stability features with interpretable machine learning models enables effective differentiation between SCA3 and MSA-C. It also facilitates the identification of clinically relevant digital biomarkers, supporting early diagnosis and personalized rehabilitation planning in cerebellar ataxia.

Clinical trial number NTC04010214.