TL;DR
This systematic review analyzes 46 studies using machine learning regression techniques to estimate biomechanical time-series (joint/segment kinematics and kinetics) from wearable sensors. The authors highlight that incorporating domain knowledge and using nonparametric regression often leads to better performance, especially with limited data, but significant challenges remain in practical implementation, including the need for subject-specific models and validation on impaired populations.
What Problem Does This Paper Address?
The paper addresses the challenge of accurately estimating biomechanical time-series data (like joint angles, velocities, and forces) from wearable sensors for applications in healthcare and rehabilitation. While wearable sensors offer the promise of continuous, remote patient monitoring, directly measuring these biomechanical quantities often requires complex physical models that necessitate numerous sensors, making them impractical for daily use. Regression-based machine learning techniques offer a potential solution by allowing estimation with fewer sensors, but a comprehensive understanding of the current landscape of these techniques, their performance, and their limitations for real-world deployment is lacking.
Methods
The authors conducted a systematic review of 46 peer-reviewed articles published between 1995 and 2019. They searched PubMed and IEEE Xplore databases using terms related to regression, biomechanical time-series, and wearable sensors. Inclusion criteria focused on studies using non-physical regression to estimate biomechanical time-series from wearable sensor data, with objective validation of testing error. The reviewed papers were then analyzed and categorized based on sensor types, biomechanical variables estimated, tasks performed, input features used, and the machine learning models employed. A classification system was developed to group prediction equations into three main categories: polynomial mixtures (Pn), neural networks (NN), and nonparametric regression (NP).
Datasets
The paper does not introduce new datasets. Instead, it reviews existing studies that utilized various datasets for training and validation. The datasets were derived from participants (mostly unimpaired males) performing specific tasks (e.g., isometric tasks, walking, activities of daily living). A significant limitation noted is that most studies trained and validated their models on small, subject-specific datasets, with only a small percentage validating on impaired populations.
Key Insights and Ideas
- Domain Knowledge is King (Often): Incorporating prior knowledge about biomechanics, physiology, or system dynamics into the regression models (e.g., through feature engineering or model structure) consistently led to improved estimation performance. This suggests that purely data-driven approaches might not be sufficient.
- Nonparametric Regression Shines with Small Data: When models were trained on limited datasets, nonparametric regression techniques often outperformed other methods. This is crucial for practical applications where collecting massive datasets can be challenging.
- The "Black Box" Problem Persists: None of the reviewed prediction equations were open-sourced. This lack of transparency and accessibility hinders broader adoption, validation, and further development by the research community.
- Impaired Populations are Underrepresented: The vast majority of studies validated their models on unimpaired individuals. This is a major hurdle for clinical translation, as performance often degrades when applied to populations with actual impairments.
- Subject-Specific Models Limit Scalability: The prevalence of subject-specific models (80% of studies) means that each user would require individual in-lab calibration, significantly limiting the scalability of these solutions for remote patient monitoring.
Key Results
- Popularity Contest: Neural networks (NN) were the most popular modeling approach (72% of studies), followed by polynomial mixtures (30%) and nonparametric regression (15%).
- Sensor Landscape: Surface electromyography (sEMG) sensors were the most common (32 studies), followed by inertial sensors (9 studies).
- Focus on Kinematics: Joint kinematics was the most frequently estimated biomechanical variable (23 studies), followed by joint kinetics (16 studies).
- Sagittal Plane Dominance: Most studies focused on biomechanics in the sagittal plane (87%), with less attention paid to frontal and transverse planes.
- Performance Varies: While specific quantitative results are hard to compare directly due to differing validation methods, the review highlights that studies incorporating domain knowledge and nonparametric methods often reported superior performance, especially with smaller datasets.
Limitations and Future Work
The authors identify several key limitations and suggest future directions:
- Lack of Open-Source Models: The absence of open-source algorithms prevents wider adoption and validation. Future work should prioritize developing and sharing open-source solutions.
- Limited Validation on Impaired Populations: Models need to be rigorously validated on individuals with clinical impairments to assess their true utility.
- Subject-Specific Models: The reliance on subject-specific models hinders scalability. Research into subject-general models or effective real-time calibration methods is crucial.
- Hardware Constraints: Practical limitations like battery life and memory capacity of wearable sensors need to be addressed. Future work should consider hardware specifications for remote, 24-hour monitoring.
- Focus on Muscle Forces and Joint Contact Forces: There's a notable lack of studies estimating muscle forces and joint contact forces, which are clinically important. Future research should explore these areas.
- Hybrid Approaches: Combining physics-based models with machine learning techniques (hybrid estimation) holds significant promise for improving accuracy and robustness.
- Frontal and Transverse Plane Mechanics: More research is needed to accurately estimate biomechanics in these planes, which are relevant for conditions like knee osteoarthritis.
Summary updated