Evaluating the Influence of Sensor Configuration and Hyperparameter Optimization on Wearable-Based Knee Moment Estimation During Running

Accurate assessment of training status in team sports is crucial for optimising performance and reducing injury risk. This pilot study investigates the feasibility of using machine learning (ML) models to estimate oxygen uptake (VO2) with wearable sensors during team sports activities. Six healthy male team sports athletes participated in the study. Data were collected using inertial measurement units (IMU), heart rate monitors, and breathing rate sensors during incremental fitness tests. The performance of different ML models, including multiple linear regression (MLR), XGBoost, and deep learning models (LSTM, CNN, MLP), was compared using raw and engineered features from IMU data. Results indicate that while LSTM models with raw IMU data provided the most accurate predictions (RMSE: 4.976, MAE: 3.698 [Formula: see text]), MLR models remained competitive, especially with engineered features. Multi-sensor configurations, particularly those including sensors on the torso and limbs, enhanced prediction accuracy. The findings demonstrate the potential of ML models to monitor VO2 noninvasively in real-time, offering valuable insights into the internal physiological demand during team sports activities.

Fall-risk assessment of aged workers using wearable inertial measurement units based on machine learning

Due to the recent technological advances in inertial measurement units (IMUs), many applications for the measurement of human motion using multiple body-worn IMUs have been developed. In these applications, each IMU has to be attached to a predefined body segment. A technique to identify the body segment on which each IMU is mounted allows users to attach inertial sensors to arbitrary body segments, which avoids having to remeasure due to incorrect attachment of the sensors. We address this IMU-to-segment assignment problem and propose a novel end-to-end learning model that incorporates a global feature generation module and an attention-based mechanism. The former extracts the feature representing the motion of all attached IMUs, and the latter enables the model to learn the dependency relationships between the IMUs. The proposed model thus identifies the IMU placement based on the features from global motion and relevant IMUs. We quantitatively evaluated the proposed method using synthetic and real public datasets with three sensor configurations, including a full-body configuration mounting 15 sensors. The results demonstrated that our approach significantly outperformed the conventional and baseline methods for all datasets and sensor configurations.

Gait rehabilitation commonly relies on bodyweight unloading mechanisms, such as overhead mechanical support and underwater buoyancy. Lightweight and wireless inertial measurement unit (IMU) sensors provide a cost-effective tool for quantifying body segment motions without the need for video recordings or ground reaction force measures. Identifying the instant when the foot contacts and leaves the ground from IMU data can be challenging, often requiring scrupulous parameter selection and researcher supervision. We aimed to assess the use of machine learning methods for gait event detection based on features from foot segment rotational velocity using foot-worn IMU sensors during bodyweight-supported treadmill walking on land and underwater. Twelve healthy subjects completed on-land treadmill walking with overhead mechanical bodyweight support, and three subjects completed underwater treadmill walking. We placed IMU sensors on the foot and recorded motion capture and ground reaction force data on land and recorded IMU sensor data from wireless foot pressure insoles underwater. To detect gait events based on IMU data features, we used random forest machine learning classification. We achieved high gait event detection accuracy (95–96%) during on-land bodyweight-supported treadmill walking across a range of gait speeds and bodyweight support levels. Due to biomechanical changes during underwater treadmill walking compared to on land, accurate underwater gait event detection required specific underwater training data. Using single-axis IMU data and machine learning classification, we were able to effectively identify gait events during bodyweight-supported treadmill walking on land and underwater. Robust and automated gait event detection methods can enable advances in gait rehabilitation.

BACKGROUND Wearable sensor technologies, such as inertial measurement units (IMUs), smartwatches, and multi-sensor systems, have emerged as valuable tools in clinical and real-world health monitoring. These devices allow continuous, non-invasive tracking of gait, mobility, and functional health across a variety of populations. However, significant challenges remain, including variability in sensor placement, data processing methodologies, and insufficient validation in real-world settings. OBJECTIVE This systematic review aims to evaluate recent literature on the clinical and research applications of wearable sensors. Specifically, it investigates how these technologies are used to assess mobility, predict disease risk, and support rehabilitation. It also identifies limitations and proposes future research directions. METHODS The review was conducted according to PRISMA guidelines. A comprehensive search of PubMed, Scopus, and Web of Science databases was performed for studies published in the past ten years. Inclusion criteria focused on studies using wearable sensors in clinical or real-world environments. A total of 30 eligible studies were identified for qualitative synthesis. Data extracted included study design, population characteristics, sensor type and placement, machine learning algorithms, and clinical outcomes. RESULTS Among the reviewed studies, observational designs were the most common (43.3%), followed by experimental studies (26.7%) and randomized controlled trials (10%). IMU-based sensors were used in 66.7% of studies, with wrist-worn devices being the most common placement (43.3%). Machine learning techniques were frequently applied, with random forest (20%) and deep learning (16.7%) models predominating. Clinical applications spanned Parkinson’s disease, stroke, multiple sclerosis, and frailty, with several studies reporting high predictive accuracy for fall risk and mobility decline (AUROC up to 0.919, p < 0.05). CONCLUSIONS Wearable sensors demonstrate strong potential for enhancing mobility monitoring, disease risk assessment, and rehabilitation tracking in both clinical and real-world settings. However, challenges remain in standardizing sensor protocols and data analysis. Future research should focus on large-scale, longitudinal studies, harmonized machine learning pipelines, and integration with cloud-based health systems to improve scalability and clinical translation.

Background: Adherence to scheduled radiation therapy (RT) is a key determinant of cancer treatment quality and outcomes. For this study, we developed an interpretable AI model to identify 1) patients at risk for multiple unplanned RT interruptions and 2) modifiable factors contributing to an elevated risk of RT interruption. Methods: We retrospectively analyzed clinical, socioeconomic, demographic, and behavioral data from 2,525 RT patients treated at the University of Tennessee Medical Center (UTMC) in Knoxville. The study cohort was dichotomized into patients with 0-1 unplanned RT interruptions (Class 0; n≈2000) and those missing >2 sessions (Class 1; n≈500). The dataset was partitioned into training, validation, and test sets (70:15:15 ratio), with class imbalance addressed in the training set by synthetic data generation via Tabular Variational Autoencoder. Twenty-seven candidate features were initially evaluated for multicollinearity using correlation matrices, heatmap visualization, and Variance Inflation Factor analysis. We applied feature selection methods (correlation-based techniques and causality-based approaches) to limit further modeling to the most predictive 15 core features. We compared XGBoost and Neural Networks-based classifiers, with each model undergoing hyperparameter optimization using Bayesian optimization methods. SHapley Additive exPlanations (SHAP) analysis was used to identify influential predictors. Results: The final optimized XGBoost model provided an overall accuracy of 82% and AUC-ROC of 63% on the independent test set. All tested models yielded similar performance, confirming the consistent predictive value of our selected features despite class imbalance. SHAP analyses identified dominant predictive contributions from treatment factors (prescribed radiation dose per session), patient resources (insurance coverage, marital status, social vulnerability indices), and travel distance to the radiotherapy facility. Supplementary causal analysis employing total causal effect methods further corroborated the direct influence of all these features. Conclusions: Our results suggest that causal inference and explainable AI modeling can provide useful interrogative strategies to identify modifiable predictors of RT adherence. Further refinement of predictive decision-support tools may lead to automated approaches to match high-risk patients with personalized interventions (e.g. community-based care navigation and/or patient psychosocial support) in real-world clinical settings to overcome social barriers to RT access. Citation Format: Rezaur Rashid, Soheil Hashtarkhani, Parnian K. Rahimabad, Brianna M. White, Fekede A. Kumsa, Lokesh Chinthala, Janet A. Zink, Christopher L. Brett, Robert L. Davis, David L. Schwartz, Arash Shaban-Nejad. Machine Learning and Causal Inference-Based Predictive Risk Modeling of Unplanned Radiation Treatment Interruption [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Artificial Intelligence and Machine Learning; 2025 Jul 10-12; Montreal, QC, Canada. Philadelphia (PA): AACR; Clin Cancer Res 2025;31(13_Suppl):Abstract nr A061.

We compared the ability of seven machine learning algorithms to use wearable inertial measurement unit (IMU) data to identify the severe knee loading cycles known to induce microdamage associated with anterior cruciate ligament rupture. Sixteen cadaveric knee specimens, dissected free of skin and muscle, were mounted in a rig simulating standardized jump landings. One IMU was located above and the other below the knee, the applied three-dimensional action and reaction loads were measured via six-axis load cells, and the three-dimensional knee kinematics were also recorded by a laboratory motion capture system. Machine learning algorithms were used to predict the knee moments and the tibial and femur vertical forces; 13 knees were utilized for training each model, while three were used for testing its accuracy (i.e., normalized root-mean-square error) and reliability (Bland-Altman limits of agreement). The results showed the models predicted force and knee moment values with acceptable levels of error and, although several models exhibited some form of bias, acceptable reliability. Further research will be needed to determine whether these types of models can be modified to attenuate the inevitable in vivo soft tissue motion artifact associated with highly dynamic activities like jump landings.

Athlete performance and monitoring is an area of interest across many sporting codes, for indoor sports however, the primary methods of athlete monitoring are manually recording statistics or video analysis.This paper contains two studies, the first to validate an automatic jump detection method which was developed using inertial sensors and the second to determine the accuracy at which the time of flight (ToF) can be detected using this method.To evaluate the automatic jump detection method both male (n=7) and female (n=5) volleyball players performed a total of 1201 jumps manually identified on video during a typical training session. Of these 1201 jumps, the method correctly identified 1144 (95%) jumps in the inertial data with only 57 (5%) being false negatives and 54 (4%) false positives.The ToF was then found to be underestimated with a mean error of -0.015s ± 0.058s when compared to the ToF obtained from a force plate.Overall the system provided a means to quickly and easily track the number of jumps being performed by each player and the approximate ToF of each jump when compared to existing methods.

This paper presents the design of an experimental platform for rapid prototyping, testing, evaluation, and validation of different sensor types and configurations for Inertial Measurement Unit (IMU) applications. Two example IMU designs are rapidly prototyped and tested on this experimental platform. The first design was based on a triad of gyroscopes, three dual-axis accelerometers and a triad of magnetometers to provide information on the Earth’s magnetic field for the purpose of bounding the attitude drift. The second system was based on twelve accelerometers and again a triad of magnetometers. An Unscented Kalman Filtering technique is used to fuse the IMU and magnetometer data in both cases for an Inertial Navigation System (INS). Test results are presented comparing the two IMU designs to each other and to ground truth with minimal magnetic interference and with introduction of deliberate magnetic interference. The results indicate that the two systems have excellent performance for high-speed applications, while at lower speeds the first design outperforms the second design. It is also shown that external magnetic disturbance degrades the performance of both designs.

In most real world rehabilitation training, patients are trained to regain motion capabilities with the aid of functional/epidural electrical stimulation (FES/EES), under the support of gravity-assist systems to prevent falls. However, the lack of motion analysis dataset designed specifically for rehabilitation-related applications largely limits the conduct of pilot research. We provide an open access dataset, consisting of multimodal data collected via 16 electromyography (EMG) sensors, 6 inertial measurement unit (IMU) sensors, and 230 insole pressure sensors (IPS) per foot, together with a 26-sensor motion capture system, under different MOVEments and POstures for Rehabilitation Training (MovePort). Data were collected under diverse experimental paradigms. Twenty four participants first imitated multiple normal and abnormal body postures including (1) normal standing still, (2) leaning forward, (3) leaning back, and (4) half-squat, which in practical applications, can be detected as feedback to tune the parameters of FES/EES and gravity-assist systems to keep patients in a target body posture. Data under imitated abnormal gaits, e.g., (1) with legs raised higher under excessive electrical stimulation, and (2) with dragging legs under insufficient stimulation, were also collected. Data under normal gaits with low, medium and high speeds are also included. Pathological gait data from a subject with spastic paraplegia further increases the clinical value of our dataset. We also provide source codes to perform both intra- and inter-participant motion analyses of our dataset. We expect our dataset can provide a unique platform to promote collaboration among neurorehabilitation engineers.

Anxiety disorders affect hundreds of millions of people worldwide, yet objective and continuous assessment remains limited in clinical practice. To our knowledge, this is the first modality-specific, translational synthesis focusing on wearable ECG and PPG for anxiety detection. Wearable electrocardiography (ECG) and photoplethysmography (PPG), combined with data-driven analytics, have emerged as promising tools for anxiety monitoring, but translation into routine care has been slow. Here, we present a PRISMA-guided systematic review of 38 studies (2015-2025) investigating wearable ECG- and PPG-based anxiety detection. We analyze anxiety induction paradigms, sensor configurations, signal acquisition strategies, and analytical approaches, including statistical, machine learning, and hybrid methods. While autonomic markers derived from ECG and PPG consistently reflect anxiety-related physiological changes, substantial heterogeneity in study design, limited population diversity, and laboratory-centric validation constrain clinical generalizability. Critically, most studies lack evaluation in real-world settings and do not demonstrate clinical utility or impact on patient outcomes. We identify key translational barriers and propose a digital medicine roadmap emphasizing standardized protocols, robust validation across diverse populations, workflow integration, and outcome-driven evaluation to enable clinically actionable, real-world anxiety monitoring.

Machine learning is a promising approach to evaluate human movement based on wearable sensor data. A representative dataset for training data-driven models is crucial to ensure that the model generalizes well to unseen data. However, the acquisition of sufficient data is time-consuming and often infeasible. We present a method to create realistic inertial sensor data with corresponding biomechanical variables by 2D walking and running simulations. We augmented a measured inertial sensor dataset with simulated data for the training of convolutional neural networks to estimate sagittal plane joint angles, joint moments, and ground reaction forces (GRFs) of walking and running. When adding simulated data, the root mean square error (RMSE) of the test set of hip, knee, and ankle joint angles decreased up to 17%, 27% and 23%, the RMSE of knee and ankle joint moments up to 6% and the RMSE of anterior-posterior and vertical GRF up to 2 and 6%. Simulation-aided estimation of joint moments and GRFs was limited by inaccuracies of the biomechanical model. Improving the physics-based model and domain adaptation learning may further increase the benefit of simulated data. Future work can exploit biomechanical simulations to connect different data sources in order to create representative datasets of human movement. In conclusion, machine learning can benefit from available domain knowledge on biomechanical simulations to supplement cumbersome data collections.

Given the high rates of both primary and secondary anterior cruciate ligament (ACL) injuries in multidirectional field sports, there is a need to develop easily accessible methods for practitioners to monitor ACL injury risk. Field-based methods to assess knee variables associated with ACL injury are of particular interest to practitioners for monitoring injury risk in applied sports settings. Knee variables or proxy measures derived from wearable inertial measurement units (IMUs) may thus provide a powerful tool for efficient injury risk management. Therefore, the aim of this study was to identify whether there were correlations between laboratory-derived knee variables (knee range of motion (RoM), change in knee moment, and knee stiffness) and metrics derived from IMUs (angular velocities and accelerations) placed on the tibia and thigh, across a range of movements performed in practitioner assessments used to monitor ACL injury risk. Ground reaction forces, three-dimensional kinematics, and triaxial IMU data were recorded from nineteen healthy male participants performing bilateral and unilateral drop jumps, and a 90° cutting task. Spearman’s correlations were used to examine the correlations between knee variables and IMU-derived metrics. A significant strong positive correlation was observed between knee RoM and the area under the tibia angular velocity curve in all movements. Significant strong correlations were also observed in the unilateral drop jump between knee RoM, change in knee moment, and knee stiffness, and the area under the tibia acceleration curve (rs = 0.776, rs = −0.712, and rs = −0.765, respectively). A significant moderate correlation was observed between both knee RoM and knee stiffness, and the area under the thigh angular velocity curve (rs = 0.682 and rs = −0.641, respectively). The findings from this study suggest that it may be feasible to use IMU-derived angular velocities and acceleration measurements as proxy measures of knee variables in movements included in practitioner assessments used to monitor ACL injury risk.

This study examined whether an inertial measurement unit (IMU), in combination with machine learning, could accurately predict two indirect measures of bowling intensity through ball release speed (BRS) and perceived intensity zone (PIZ). One IMU was attached to the thoracic back of 44 fast bowlers. Each participant bowled 36 deliveries at two different PIZ zones (Zone 1 = 24 deliveries at 70% to 85% of maximum perceived bowling effort; Zone 2 = 12 deliveries at 100% of maximum perceived bowling effort) in a random order. IMU data (sampling rate = 250 Hz) were downsampled to 125 Hz, 50 Hz, and 25 Hz to determine if model accuracy was affected by the sampling frequency. Data were analysed using four machine learning models. A two-way repeated-measures ANOVA was used to compare the mean absolute error (MAE) and accuracy scores (separately) across the four models and four sampling frequencies. Gradient boosting models were shown to be the most consistent at measuring BRS (MAE = 3.61 km/h) and PIZ (F-score = 88%) across all sampling frequencies. This method could be used to measure BRS and PIZ which may contribute to a better understanding of overall bowling load which may help to reduce injuries.

Inertial Measurement Unit (IMU) sensors are the essential components in navigation systems. Micro-Electro-Mechanical System (MEMS) sensors are a less affordable and lightweight IMU. MEMS suffer from larger stochastic errors that accumulate over time, eventually causing drifts in navigation. The integration of Global Navigation Satellite System (GNSS) measurements with IMU data solves the issue of navigation drift and provides accurate navigation. However, this integration falls short in during GNSS outage scenarios, primarily due to persistent IMU errors. This thesis introduces an effective strategy to reduce navigation drifts by rectifying IMU errors using Machine Learning (ML) algorithms such as Light Gradient Boosting Machine (LightGBM) and Categorical Boosting (CatBoost). In contrast to many other methodologies reliant on high-end and expensive IMUs for denoising low-cost MEMS IMUs, this thesis proposes Inverse Kinematics (IK). The proposed method undergoes testing in Loosely coupled and Tightly coupled schemes under varying GNSS outage durations.