Wearable sensors and machine learning fusion-based fall risk prediction in covert cerebral small vessel disease

Yuanyuan Zhou ^1,2 Dingwen Zhang ³ Yingxiao Ji ⁴ Shuohan Bu ⁵ Xinzhu Hu ¹ Congying Zhao ⁴ Zhou Lv ⁶ Litao Li ^4,7*

• ¹ Department of Neurology, Hebei Medical University, Shijiazhuang, China
• ² Department of Neurology, Baoding NO.1 Central Hospital, Baoding, China
• ³ China National Clinical Research Center for Neurological Diseases, Beijing, China
• ⁴ Department of Neurology, Hebei General Hospital, Shijiazhuang, Hebei Province, China
• ⁵ Department of Neurology, Hebei North University, Zhangjiakou, Hebei Province, China
• ⁶ Department of Neurology, Handan Central Hospital, Handan, China
• ⁷ Hebei Provincial Key Laboratory of Cerebral Networks and Cognitive Disorders, Hebei General Hospital, Shijiazhuang, China

Background: Fall risk prediction is crucial for preventing falls in patients with cerebral small vessel disease (CSVD), especially for those with gait disturbances. However, research in this area is limited, particularly in the early, asymptomatic phase. Wearable sensors offer an objective method for gait assessment. This study integrating wearable sensors and machine learning, aimed to predict fall risk in patients with covert CSVD. Methods: We employed soft robotic exoskeleton (SRE) to acquire gait characteristics and surface electromyography (sEMG) system to collect sEMG features, constructing three datasets: gait-only, sEMG-only, and their combination. Using Support Vector Machine (SVM), Random Forest (RF) , Gradient Boosting Decision Tree (GBDT), and Neural Network (NN) algorithms, we developed twelve predictive models. Furthermore, we integrated the selected baseline data and imaging markers with the three original datasets to create three new integrated datasets, and constructed another twelve optimized predictive models using the same methods. A total of 117 participants were enrolled in the study. Results: Of the 28 features, ANOVA identified 10 significant indicators. The Gait & sEMG integration dataset, analyzed using the SVM algorithm, demonstrated superior performance compared to other models. This model exhibited an area under the curve (AUC) of 0.986, along with a sensitivity of 0.909 and a specificity of0.923, reflecting its robust discriminatory capability. Conclusions: This study highlights the essential role of gait characteristics, electromyographic features, baseline data, and imaging markers in predicting fall risk. It also successfully developed an SVM-based model integrating these features. This model offers a valuable tool for early detection of fall risk in CSVD patients, potentially enhancing clinical decision-making and prognosis.