PaperlayerView exact revision on arXiv

Source-linked AI summary

Integrating Triaxial IMU Sensors and Ensemble Learning for Effective Parkinson Disease Severity Classification

Rehan Khan, Muhammad Junaid Asif, Rana Fayyaz Ahmad

arXiv:2608.28602v1cs.AIcs.CV

TL;DR

Parkinson’s disease severity is difficult to assess continuously because conventional clinical evaluations are periodic, partial, and potentially subjective. The paper analyzes triaxial wearable-IMU motion and tremor signals with multiple machine-learning classifiers, finding LightGBM performs best at around 97% across the reported metrics. The results support the proposed approach as an accurate predictive system for PD severity classification.

  • Problem

    Conventional Parkinson’s disease evaluations are periodic and may provide incomplete, subjective monitoring, motivating data-driven approaches for timely diagnosis and ongoing assessment.

  • Method

    The framework processes triaxial accelerometer and gyroscope signals from wearable IMUs and compares Logistic Regression, KNN, SVM, Decision Tree, XGBoost, and LightGBM.

  • Results

    LightGBM achieved the best overall performance, with Accuracy, Precision, Recall, and F1-score of around 97%.

  • Takeaways & Limitations

    The study concludes that gradient boosting models, particularly LightGBM, are promising for interpreting complex wearable-IMU tremor and movement patterns in PD monitoring and clinical decision support.

Abstract

from arXiv · show

Parkinson disease PD is a progressive neurodegenerative disease that can have a significant impact on motor performance resulting in the appearance of symptoms such as tremors rigidity postural instabilities and bradykinesia. Timely clinical treatment disease management and quality life of the patients are closely linked to early and appropriate identification of PD. Over the past few years the growth of wearable sensor technology and artificial intelligence AI have made it possible to create noninvasive and data driven disease detection methods. This paper proposes a comparative system using artificial intelligence to detect Parkinsons disease by analyzing the motion and tremor data captured by an inertial measurement unit IMU. The data comprises the signals of the acceleration and gyroscope sensors measuring movement in three directions X Y and Z. The signs and symptoms provide helpful information about subtle motor deficits associated with PD. Several classification models like Support Vector Machine SVM Logistic Regression LR KNearest Neighbors KNN Decision Tree DT Extreme Gradient Boosting XGBoost and Light Gradient Boosting Machine LightGBM were used to compare their effectiveness. The Logistic Regression model had a performance around 75 percent in all evaluation metrics and KNearest Neighbours KNN around 90 percent. The support vector machine SVM performed almost 94 percent whereas the performance of classifiers such as Decision Tree and XGBoost was close to 96 percent and overall classification efficacy respectively. LightGBM model performs consistently at the best rank among all of the evaluated methods having Accuracy, Precision, Recall and F1score of around 97 percent. The results show that the proposed machine learning approach offers an accurate and effective predictive capability in the classification of PD severity.

I. INTRODUCTION

The paper motivates wearable IMU and machine-learning methods for Parkinson’s disease severity classification, addressing limits in periodic, subjective clinical assessment and incomplete continuous monitoring. It proposes a triaxial sensor framework and compares multiple classifiers for data-driven disease monitoring and clinical decision support.

  • Parkinson’s disease progressively impairs motor functions, producing tremors, bradykinesia, rigidity, and balance problems that affect daily activities and quality of life.
  • Timely diagnosis and regular monitoring matter because symptoms change over time and can influence treatment management and disease progression assessment.
  • Conventional examinations provide only periodic, partial assessments and may miss tremors while introducing inconsistency through subjective clinical judgment.
  • Wearable IMUs combine accelerometers and gyroscopes to capture movement-related information, while machine learning can analyze motion and tremor patterns for subtle abnormalities.
  • The proposed framework processes triaxial accelerometer and gyrometer signals along X, Y, and Z axes to classify Parkinson’s disease severity.
  • The study comparatively evaluates Logistic Regression, KNN, SVM, Decision Tree, XGBoost, and LightGBM for automated severity classification and wearable-based clinical support.

II. RELATED WORK

Related work has combined wearable technologies, motion analysis, and machine learning to support earlier Parkinson’s diagnosis, symptom assessment, and ongoing monitoring.

  • Prior Parkinson’s disease studies use wearable technologies, motion analysis, and machine-learning techniques for earlier diagnosis, movement and tremor assessment, and ongoing patient monitoring.

A. IMU-Based Detection of Gait Abnormalities in Parkinson’s

Related work has used multimodal wearable sensing and inertial data to analyze gait abnormalities and Parkinson’s motor symptoms. Gyroscope-derived features and combined accelerometer–gyroscope sensing provide useful information for distinguishing motor states and gait impairments.

  • A multisource gait-monitoring system combined IMUs, force-sensitive sensors, and piezoelectric sensors to collect plantar pressure, postural angles, and movement dynamics.
  • The multisource sensor data supported gait-abnormality analysis and Parkinson’s disease monitoring and diagnosis.
  • Gyroscope features were more effective than accelerometer features for distinguishing On and Off medication states, while combining both modalities improved gait-abnormality detection and monitoring.
  • A review of 24 studies from 2010–2020 found wearable inertial sensors, including accelerometers, gyroscopes, and magnetometers, were used mainly in home environments to capture natural movements over longer periods.

C. Movement Disorder Detection Using Wearable Inertial Sensors

Related studies have used wearable inertial sensors and machine learning to detect Parkinsonian movement disorders and monitor symptoms. These approaches commonly process accelerometer and gyroscope data into features evaluated with standard classification metrics.

  • Wearable inertial sensors have been used to detect tremors, dyskinesia, and bradykinesia, with possible support for real-time monitoring through smartwatches and fitness bands.
  • One kinetic-tremor study placed IMUs on the wrists of 30 patients performing spiral and line-drawing tasks, recording three-axis accelerometer, gyroscope, and magnetometer signals at 100 Hz.
  • A smartwatch and fitness-band system used triaxial accelerometer data from 28 Parkinson’s patients in a two-stage approach to detect abnormal movements and classify tremor, dyskinesia, or bradykinesia.
  • A Kaggle dataset study preprocessed accelerometer and gyroscope data, segmented signals into fixed windows, extracted motion features, and evaluated machine-learning models.
  • The dataset study assessed model effectiveness using accuracy, precision, recall, and F1-score.

A. Dataset Details

The dataset combines accelerometer and gyroscope measurements across spatial axes with labels for three Parkinson’s disease severity classes. The framework uses these motion features directly or with limited preprocessing, then trains and evaluates multiple classifiers.

  • Dataset Details: The dataset contains digital accelerometer and gyroscope motion values recorded across the x, y, and z spatial axes.Accelerometers measure linear acceleration, while gyroscopes measure angular speed and turning.
  • Dataset Details: Each sample is labeled as mild, moderate, or severe to indicate Parkinson’s disease progression.The dataset is balanced, with equal numbers of samples in labels 1, 2, and 3.
  • Dataset Details: The proposed approach analyzes motion and tremor signals from wearable triaxial IMU sensors across specified acceleration and angular-velocity axes.The sensors measure angular velocity in X, Y, and Z and linear acceleration in X and Y.
  • Dataset Details: The raw sensor feature matrix is used without heavy preprocessing or handcrafted signal transformations, preserving motion characteristics while lowering computational complexity.The matrix contains accelerometer and gyroscope readings.
  • Dataset Details: The feature matrix is classified with Logistic Regression, KNN, SVM, Decision Tree, XGBoost, and LightGBM models.The models learn discriminative patterns from triaxial motion signals to assign one of three severity stages.
  • Dataset Details: Training and testing splits support generalization assessment, using accuracy, precision, recall, F1-score, and confusion matrices for comparison.The methodology also includes data acquisition, exploratory analysis, training, evaluation, deployment, and clinical support.

C. Evaluation Metrics

The framework evaluates Parkinson’s disease severity classifiers with standard performance measures and confusion-matrix analysis. These measures assess correctness, positive-prediction accuracy, sensitivity, and the balance between precision and recall.

  • Evaluation Metrics: Accuracy measures the percentage of classifications that are correct out of all classifications.
  • Evaluation Metrics: Precision measures the proportion of observed positive cases among the cases predicted as positive.
  • Evaluation Metrics: Recall or sensitivity measures the model’s ability to identify all true positives in the dataset.
  • Evaluation Metrics: F1-score balances classification precision and recall for evaluating robustness.

A. Performance with Logistic Regression

Logistic Regression provided a moderate and balanced baseline for classifying Parkinson’s disease severity from triaxial IMU sensor data. Its performance was consistent across the evaluated severity classes.

  • Performance with Logistic Regression: Confusion matrices and graphical visualizations were used to compare the model’s classification capabilities and predictive performance.
  • Performance with Logistic Regression: 75% macro-averaged precision, recall, and F1-score were also obtained, indicating similar performance across severity classes.
  • Performance with Logistic Regression: Logistic Regression extracted most of the relevant tremor patterns and confirmed a significant proportion of actual positive cases through recall.

B. Performance with KNN

K-Nearest Neighbors performed substantially better than Logistic Regression for Parkinson’s disease severity classification. Its metrics were balanced across the severity classes and reflected local relationships in the triaxial signals.

  • Performance with KNN: KNN significantly outperformed Logistic Regression in the reported severity-classification comparison.
  • Performance with KNN: Around 90% accuracy, precision, recall, and F1-score were achieved by K-Nearest Neighbors.
  • Performance with KNN: KNN’s balanced metric values suggest similar classification performance across all severity classes.
  • Performance with KNN: The model captured local feature relationships and similarities in the triaxial motion signals.

C. Performance with Support Vector Machines (SVM)

SVM achieved approximately 94% across key classification metrics, while Decision Tree and XGBoost reached about 96% with strong performance on triaxial IMU signals.

  • C. Performance with Support Vector Machines (SVM): Approximately 94% accuracy, precision, and recall were achieved by SVM for each severity class.The paper attributes this performance to optimal decision boundaries for complex tremor-related motion patterns.
  • C. Performance with Support Vector Machines (SVM): About 96% Accuracy, Precision, Recall, and F1-score were obtained by the Decision Tree classifier.The model learned discriminating movement patterns and captured nonlinear feature interactions while retaining interpretability.
  • C. Performance with Support Vector Machines (SVM): About 96% Accuracy, Precision, Recall, and F1-score were achieved by XGBoost with consistent classification performance.The results are associated with ensemble accelerating strategies learning complex tremor attributes from triaxial IMU signals.

F. Performance with LightGBM

LightGBM outperformed the other evaluated classifiers for Parkinson disease severity classification using IMU sensor data, achieving approximately 97% across all reported metrics.

  • F. Performance with LightGBM: Around 97% Accuracy, Precision, Recall, and F1-score were achieved by LightGBM.The paper reports balanced high metrics for classifying Parkinson disease severity from accelerometer and gyroscope measurements.
  • F. Performance with LightGBM: Figure 4 compares Accuracy, Precision, Recall, and F1-Score across six implemented classification models.The models are shown in panels for Logistic Regression, K-NN, SVM, Decision Tree, XGBoost, and LightGBM.
  • F. Performance with LightGBM: LightGBM achieved the highest classification accuracy at 97% among the evaluated approaches.Figure 5 compares Logistic Regression, K-NN, SVM, Decision Tree, XGBoost, and LightGBM.

V. CONCLUSIONS

The comparative evaluation shows substantial performance differences across classifiers, with ensemble methods outperforming introductory approaches. LightGBM achieved the strongest overall results, reaching around 97% across the reported metrics.

  • About 75% Accuracy, Precision, Recall, and F1-score made Logistic Regression the baseline and lowest-performing classifier.Its linear form was not sufficiently able to represent complex nonlinear tremor and movement patterns.
  • About 90% across the evaluation measures showed that K-Nearest Neighbors captured useful similarity relationships in triaxial motion signals.
  • Around 94% Accuracy, Precision, Recall, and F1-score indicated that SVM produced accurate, balanced predictions for all severity classes.Its decision boundaries modeled complex motion patterns in accelerometer and gyroscope signals.
  • Around 96% across the evaluation parameters was achieved by Decision Tree and XGBoost through modeling nonlinear feature relationships.
  • Around 97% Accuracy, Precision, Recall, and F1-score made LightGBM the best overall and most reliable classifier in this case study.The comparative analysis also found ensemble methods, particularly LightGBM and XGBoost, stronger than conventional methods.
  • Stable Accuracy, Precision, Recall, and F1-score values evidenced consistent model performance under macro averaging.
Loading 2608.28602v1…