Leveraging Machine Learning Techniques using Inertial Sensor‐based Static Balance Data for Early Detection of Mild Cognitive Impairment

TL;DR

This study uses inertial sensors and machine learning to detect early signs of mild cognitive impairment through balance data.

Contribution

The study introduces a novel approach combining inertial sensor data and machine learning for early MCI detection.

Findings

Key balance features like mean-distance and frequency-domain measures were identified as significant for MCI detection.

ANN and SVM models achieved high accuracy in classifying MCI under eyes-open conditions.

Eyes-open balance data proved highly distinctive for early dementia detection.

Abstract

Mild Cognitive Impairment (MCI), an early stage of dementia, is often difficult to diagnose due to its subtle and transitional nature. Research has shown that balance impairments can serve as early indicators of cognitive decline, highlighting the potential of wearable sensor technology for detecting MCI at an early stage.

In this study, balance data was collected from 60 participants (30 cognitively normal, 30 with MCI) at the National Research Center for Dementia, South Korea. Shimmer‐3 inertial sensors were placed on the lower back, left and right thigh, left and right legs. Data acquisition involved four conditions: eyes open (EO), eyes closed (EC), right‐leg lift (RL), and left‐leg lift (LL). A total of 76 features were extracted from each sensor, comprising 43 time‐domain and 33 frequency‐domain measures. In the first step, features from all sensors were combined. Extensive feature selection was performed using a combination of different Wrapper methods, namely ANOVA (Analysis of Variance), Fisher's Score and Mutual Information, followed by dimensionality‐reduction through Principal Component Analysis (PCA). The transformed dataset was subsequently utilized for classification, including Support Vector Machine (SVM) and Random Forest. Additionally, the extracted features were utilized to perform classification through an Artificial Neural Network (ANN) model. These models were evaluated using Leave‐One‐Subject‐Out (LOSO) cross‐validation technique.

Key time‐domain features, such as mean‐distance, root‐mean‐square‐of‐acceleration, range and summed‐axis‐acceleration, alongside frequency‐domain features like median, peak and mean frequency, were identified as significant balance features. In eyes‐open condition, SVM achieved classification accuracy of 84%, whereas 83% validation accuracy and 86% test accuracy were obtained by ANN. Therefore, balance characteristics in eyes‐open condition were proven to be highly distinctive, providing promising opportunities for early dementia detection.

The proposed balance biomarkers derived from inertial sensors demonstrate the potential for early detection of MCI, paving the way for non‐invasive screening solutions to facilitate timely intervention in dementia care.

Figures