TL;DR
This paper introduces OA-Pain-Sense, a novel machine learning framework that uses Inertial Measurement Unit (IMU) data from a single accelerometer to predict hip and knee osteoarthritis (OA) pain. This passive, objective approach offers a promising alternative to subjective self-report questionnaires, potentially improving clinical decision-making and patient management.
What Problem Does This Paper Address?
Osteoarthritis (OA) is a debilitating joint disease characterized by pain, which significantly impacts patients' quality of life and incurs substantial healthcare costs. Accurately assessing OA pain is crucial for effective treatment, but current methods rely heavily on self-report questionnaires. These methods are prone to bias due to psychological factors, sociocultural influences, and difficulties older adults may have in accurately reporting their pain. This subjectivity can lead to erroneous clinical decisions. The paper addresses the need for an objective, passive, and reliable method to assess OA pain, particularly for hip and knee OA (HKOA), which are the most prevalent types.
Methods
The core of the OA-Pain-Sense framework lies in its use of Spatio-Temporal Gait Parameters (STGPs) extracted from the anterior-posterior (AP) accelerometer signal of a lumbar-mounted IMU. The methodology involves several key steps:
- Data Collection: IMU data was collected from participants walking for 2 minutes.
- Data Preprocessing: Raw accelerometer signals were filtered (Butterworth low-pass filter) to reduce noise. Gait events, specifically Heel Strike (HS) and Toe-Off (TO), were identified to segment gait cycles.
- STGP Extraction: A comprehensive set of 21 STGP features were computed, including spatial parameters (e.g., step length, stride length) and temporal parameters (e.g., step time, stride time), along with measures of time variability and fluctuation magnitude.
- Data Rescaling: To mitigate inter-subject variability, four data rescaling methods were explored:
- No normalization (raw STGP)
- Dataset-level normalization (min-max scaling)
- Dataset-level standardization (z-score scaling)
- Subject-level normalization (dimensionless scaling based on height)
- Machine Learning Classification: Six different binary classification models (Logistic Regression, Support Vector Machine, K-Nearest Neighbors, Decision Tree, Random Forest, and XGBoost) were trained and tested to distinguish between patients with moderate-to-severe OA pain and those without pain.
- Hyperparameter Optimization: Grid search was employed to tune the hyperparameters of the ML models to optimize their performance.
- Feature Ranking: Random Forest was used to rank the importance of STGP features, exploring the impact of using a reduced set of features on classification accuracy.
- Model Evaluation: Performance was evaluated using accuracy and Area Under the Curve (AUC) metrics, with 5-fold stratified cross-validation to ensure robustness.
Datasets
The study utilized data collected from:
- 53 patients with unilateral end-stage hip (n=26) or knee (n=25) OA.
- 27 age-matched healthy controls (HC).
The OA severity was assessed using the Kellgren-Lawrence (KL) radiographic grading scale. Patients' overall HOOS (Hip disability and Osteoarthritis Outcome Score) or KOOS (Knee Injury and Osteoarthritis Outcome Score) scores were used as the target variables for pain prediction.
Key Insights and Ideas
- Single-Axis IMU Efficacy: The study demonstrates that a single-axis accelerometer (AP direction) from a lumbar-mounted IMU is sufficient for extracting meaningful gait features to predict OA pain. This simplifies hardware requirements and aligns with the ubiquity of accelerometers in smartphones.
- Dimensionless Normalization's Impact: Subject-level dimensionless normalization, which accounts for anthropometric variations like height, significantly improved the performance of scale-variant ML models (LR, SVM, KNN) by reducing data dispersion. This highlights the importance of accounting for individual physical characteristics in gait analysis.
- Feature Importance of Gait Velocity: Gait velocity emerged as the most important feature for predicting OA pain, aligning with existing knowledge that slower walking speed is associated with OA. This suggests that simple, readily available gait metrics can be highly predictive.
- Rescaling is Crucial: Data rescaling, particularly dataset-level methods, was essential for optimizing the performance of scale-variant ML models like LR, SVM, and KNN, whereas tree-based models showed less sensitivity. This underscores the need for careful data preprocessing tailored to the chosen ML algorithms.
Key Results
- High Classification Accuracy: The OA-Pain-Sense framework achieved impressive classification accuracies. The Decision Tree classifier, using STGP-Norm data, reached an average accuracy of 86.79% for distinguishing Hip OA (HOA) from healthy controls (HC). The Support Vector Machine classifier, using STGP-Normz data, achieved 83.57% for distinguishing Knee OA (KOA) from HC.
- Rescaling Benefits: Dataset-level rescaling methods significantly improved the performance of SVM and KNN models. Dimensionless normalization specifically led to noticeable improvements in LR, SVM, and KNN accuracies.
- Feature Ranking Insights: Reducing the number of features generally decreased performance for most models, but gait velocity alone showed comparable performance to using all features for some models (RF, XGB), highlighting its predictive power.
Limitations and Future Work
- Study Population: The study population consisted exclusively of elderly subjects with unilateral end-stage HKOA. This raises questions about the generalizability of the findings to younger individuals or those with different stages or types of OA.
- Dataset Size: The datasets used for classification were relatively small, which may have influenced the observed performance improvements from hyperparameter tuning.
- Gender Matching: Participants were not gender-matched, which could have introduced bias.
- Self-Report Bias: The reliance on patient self-reports for pain levels inherently introduces bias.
- Gold Standard for Gait Events: The lack of a definitive gold standard for validating the peak detection algorithm used to identify gait events is a limitation.
Future Work: The authors plan to:
- Collect data from patients at different stages of HKOA and across a wider age range to improve generalizability.
- Incorporate features from other physical activities (e.g., sit-to-stand, stair ascent/descent) alongside gait features to enhance model robustness.
- Explore emerging deep learning models to potentially improve robustness and reliability.
- Conduct a comprehensive analysis of HKOA gait abnormalities and compare findings to normative STGP values.
- Validate OA-Pain-Sense in a live deployment setting.
Summary updated