Objective: In Parkinson's disease (PD), gait-related digital mobility outcomes (DMOs) show promise for monitoring mobility decline, but convergent validity remains limited. To improve convergent validity, demonstrating convergence with motor severity scales and PD-specific neural mechanisms underlying mobility has been proposed. However, severity scales capture both PD-specific and non-specific factors. Requiring mechanistic evidence may favor DMOs that converge with underlying mechanisms, while those converging only with severity scales may be overlooked despite capturing broader mobility dimensions. Here, we asked whether PD-specific neural mechanisms underlying mobility enhanced the convergence of DMOs with motor severity scales - so that integrating mechanistic evidence in validation could improve (and not impede) convergent validity. Method: Principal component analysis was applied to task-based functional neuroimaging data to identify a measure associated with PD motor network dysfunction. An optimization problem was then formulated in which deep learning examined the convergence of a signal-based DMO with motor severity in laboratory and real-world contexts, and Tracing Gradient Descent assessed the influence of the identified measure on DMO-severity convergence. Results: Greater PD motor network dysfunction was associated with reduced Attractor Complexity Index (ACI) values (rho=-0.54). Strong DMO-severity convergence was found across contexts (rho=|0.81-0.82|). Reduced ACI (i.e., greater PD motor network dysfunction) markedly enhanced DMO-severity convergence across contexts (rrb=|0.63-0.29|). Conclusion: PD-specific neural mechanisms underlying mobility enhanced the convergence of a DMO with severity scales. Significance: Integrating mechanistic validation into DMO validation could improve convergent (and construct) validity, a prerequisite for regulatory approval and adoption in clinical trials and practice.

AIM:To identify the barriers and facilitators that shape how clinical gait analysis (CGA) informs family decision-making in different contexts. METHOD:A qualitative descriptive study design was chosen to explore the current clinical landscape. Semistructured interviews were conducted with 15 healthcare professionals and eight caregivers (n = 23) from three pediatric rehabilitation centers in Quebec, Canada, that provide CGA to children with cerebral palsy. Data were analyzed thematically with an inductive approach using NVivo software. RESULTS:Key barriers include a lack of structure, limited financial and human resources, and restricted access to CGA. The primary facilitators are healthcare professionals and caregivers who aim to enhance the care process by more effectively integrating CGA. Other common topics include the benefits of care management with CGA and the need to clarify the expertise required to ensure the optimal functioning and clinical application of a gait laboratory. INTERPRETATION:Participants emphasized the importance of addressing disparities, securing institutional support, and promoting intercenter collaboration to achieve sustainable improvements in the care of ambulatory children with cerebral palsy.

In ice hockey, automatic activity detection using wearable sensors and machine learning could provide objective feedback to support coaches and players during performance evaluation. The primary objective was to assess the predictive ability of a deep learning model to recognize common ice hockey stick striking actions (passing, shooting) from inertial measurement unit sensors. This study implemented a fully connected convolutional neural network model to classify seven ice hockey-related technical actions (wrist shot, slap shot, backhand shot, one-timers, pass, other, and rest) using acceleration data via two setups: an all-sensor configuration (17 sensors) and a hands-only sensor configuration (2 sensors) in 43 elite players. Data were split into 80/20 train/test sets, with a five-fold cross-validation applied to the training data. The train/test split was repeated 10 times with different random splits to assess stability of results. The model achieved high classification accuracy, with the all-sensor model reaching an average F1 score of 95.0 ± 3.0% and the hands-only model achieving 93.5 ± 1.6%. These findings support the use of convolutional neural networks for automatic shooting action classification in ice hockey and highlight the feasibility of using minimal sensor configurations, such as sensor-integrated gloves, for real-world applications. This approach could further enhance training practices by providing objective performance metrics and allowing coaches to deliver data-driven feedback to players.

BackgroundViral Respiratory Tract Infections (VRTIs) are a major public health threat. Early detection and preventive measures are key to controlling their spread. Current machine learning approaches often depend on symptom onset, costly equipment, trained personnel, and slow results. This study aims to evaluate whether a machine learning algorithm using physiological data from wearable biosensors during a constant-rate stair-stepping task (3-min test, 2-min recovery) can predict inflammation levels, and to identify the most predictive indicators of VRTI.Methods55 Healthy participants (27 males and 28 females) aged 18–59 years, were recruited and inoculated with a live influenza vaccine to induce an immune response, assessed via changes in circulating inflammatory biomarkers. Physiological markers, including breathing rate and heart rate, during a series of clinically controlled stair tests, were monitored by a wearable biosensor. These data were collected to develop a prediction model using gradient-boosting machine learning algorithms combined with hyperparameter tuning and a leave-one-subject-out method to train the models.ResultsThe study developed a predictive model that accurately estimates inflammation levels in individuals. Features from heart rate variability (HRV) showed the greatest potential, with 70% sensitivity and 77% specificity, and physiological markers from controlled stair tests correlated with VRTI-related inflammatory responses.ImpactThe prediction model linked to stair-stepping tests offers clinicians and the public a tool for self-monitoring and early intervention. Using machine learning and physiological markers, especially HRV features, it can help guide timely treatments and reduce the impact of future outbreaks.

(1) Background: Navigating surfaces during walking can alter gait patterns. This study aims to develop tools for automatic walking condition classification using inertial measurement unit (IMU) and foot pressure sensors. We compared sensor modalities (IMUs on lower-limbs, IMUs on feet, IMUs on the pelvis, pressure insoles, and IMUs on the feet or pelvis combined with pressure insoles) and evaluated whether gait cycle segmentation improves performance compared to a sliding window. (2) Methods: Twenty participants performed flat, stairs up, stairs down, slope up, and slope down walking trials while fitted with IMUs and pressure insoles. Machine learning (ML; Extreme Gradient Boosting) and deep learning (DL; Convolutional Neural Network + Long Short-Term Memory) models were trained to classify these conditions. (3) Results: Overall, a DL model using lower-limb IMUs processed with gait segmentation performed the best (F1=0.89). Models trained with IMUs outperformed those trained on pressure insoles (p<0.01). Combining sensor modalities and gait segmentation improved performance for ML models (p<0.01). The best minimal model was a DL model trained on IMU pelvis + pressure insole data using sliding window segmentation (F1=0.83). (4) Conclusions: IMUs provide the most discriminative features for automatic walking condition classification. Combining sensor modalities may be helpful for some model architectures. DL models perform well without gait segmentation, making them independent of gait event identification algorithms.

BACKGROUND:Presymptomatic or asymptomatic immune system signals and subclinical physiological changes might provide a more objective measure of early viral upper respiratory tract infections (VRTIs) compared with symptom-based detection. We aimed to use multimodal wearable sensors, host-response biomarkers, and machine learning to predict systemic inflammation following controlled exposure to a live attenuated influenza vaccine, without relying on symptoms. METHODS:WE SENSE study is a single-centre (McGill University Health Center, Montreal, QC, Canada), prospective controlled trial that recruited healthy adults aged 18-59 years who had not received or were not planning to receive the seasonal influenza vaccine or any other vaccine during the study period. We excluded participants with any infectious symptoms within 7 days before screening. We collected physiological and activity data (eg, heart rate, breathing rate, and acceleration) through continuous monitoring with a smart ring (Oura ring Gen 2, Oura Oy, Finland), smart watch (Biobeat watch, Biobeat Technologies, Israel), and smart shirt (Astroskin-Hexoskin shirt, Hexoskin, Canada) along with high temporal resolution systemic inflammatory biomarker mapping over 12 days (7 days before inoculation and 5 days after). We frequently tested participants both before and after inoculation via PCR for respiratory pathogens, and monitored them via apps for symptoms and free-text annotations. Machine learning algorithms predicting systemic inflammatory surges were trained (35 participants), validated (ten participants), and tested (ten participants) using gradient-boosting techniques. FINDINGS:Between Dec 10, 2021, and Feb, 28, 2022, we enrolled 56 participants, of whom 55 had available data; all 55 participants continuously wore the Oura ring, 54 participants wore the Astroskin-Hexoskin shirt, and 50 wore the Biobeat watch. 27 (49%) participants were female and 28 (51%) were male; 31 (56%) participants were White, eight (15%) were Asian, four (7%) were Black, two (4%) were Latino or Hispanic, and ten (18%) did not disclose. We used model 2, which included handpicked features from the Oura ring night-time data, as the candidate model because it was built on the lowest number of features (more practical). This model predicted inflammatory surges with receiver operating characteristic area under the curve (ROC-AUC) of 0·73 (95% CI 0·71-0·74) for real-time prediction and 0·89 (0·87-0·90) for a 24-h tolerance prediction window (24h-tol) using night-time data from the Oura ring. Incorporating both night-time and daytime data from the Astroskin-Hexoskin shirt yielded ROC-AUC values of 0·73 (0·71-0·75) for real-time and 0·91 (0·90-0·92) for 24h-tol along with improved precision (ie, specificity [0·83, 0·79-0·87] and F1 score [0·65, 0·58-0·71]). The model based on symptoms alone had lower performance, with ROC-AUC values of 0·66 (0·63-0·68) for real-time and 0·79 (0·77-0·82) for 24h-tol. INTERPRETATION:Systemic inflammatory biomarkers coupled with physiological data from wearable biosensors provided rich and objective data from which to train machine learning algorithms to predict systemic inflammation from a low-grade influenza challenge. This approach outperformed symptom-based detection and has the potential to improve detection of VRTIs such as influenza and decrease time to detection, even among asymptomatic people. FUNDING:The Canadian Institutes of Health Research.

In Parkinson’s disease (PD), the development of accurate wearable biomarkers for real-world monitoring is a priority. Developers tend to prioritize agreement with clinical features (e.g., neurological tests). However, wearable biomarkers should also reflect the pathogenic processes underlying these clinical features. This critical aspect is often overlooked in validation studies, raising doubts about construct validity and limiting adoption of these biomarkers. Here, we propose a solution to address this gap. We examined whether a previously validated wearable biomarker, derived from a deep learning model trained on raw accelerometer signals during walking to estimate motor symptom severity scores, can also reflect the pathogenic processes associated with motor dysfunction in people with PD (PwP). The model was reproduced and evaluated in-the-wild, before being deployed on a subset of PwP for whom neuroimaging data were also available. Neuroimaging data were analyzed to extract the brain activity pattern associated with predicted motor symptoms severity scores. The topographic similarity between the extracted pattern and two established brain patterns (one underlying motor symptoms in PD and one not) was assessed. The model accurately estimated ground-truth motor severity scores (mean absolute error = 5.20). Despite not being explicitly trained for this purpose, the model was also able to capture pathogenic mechanisms specifically linked to motor dysfunction in PD (dice similarity = 0.653). These findings represent an initial step toward linking wearable biomarkers not only to clinical features, but also to underlying mechanistic representations. This supports the wider adoption of wearable biomarkers in clinical practice and trials.

Background/Objective: Understanding the training effect in high-level running is important for performance optimization and injury prevention. This includes awareness of how different running surface types (e.g., hard versus soft) may modify biomechanics. Recent studies have demonstrated that deep learning algorithms, such as convolutional neural networks (CNNs), can accurately classify human activity collected via body-worn sensors. To date, no study has assessed optimal signal type, sensor location, and model architecture to classify running surfaces. This study aimed to determine which combination of signal type, sensor location, and CNN architecture would yield the highest accuracy in classifying grass and asphalt surfaces using inertial measurement unit (IMU) sensors. Methods: Running data were collected from forty participants (27.4 years + 7.8 SD, 10.5 ± 7.3 SD years of running) with a full-body IMU system (head, sternum, pelvis, upper legs, lower legs, feet, and arms) on grass and asphalt outdoor surfaces. Performance (accuracy) for signal type (acceleration and angular velocity), sensor configuration (full body, lower body, pelvis, and feet), and CNN model architecture was tested for this specific task. Moreover, the effect of preprocessing steps (separating into running cycles and amplitude normalization) and two different data splitting protocols (leave-n-subject-out and subject-dependent split) was evaluated. Results: In general, acceleration signals improved classification results compared to angular velocity (3.8%). Moreover, the foot sensor configuration had the best performance-to-number of sensor ratio (95.5% accuracy). Finally, separating trials into gait cycles and not normalizing the raw signals improved accuracy by approximately 28%. Conclusion: This analysis sheds light on the important parameters to consider when developing machine learning classifiers in the human activity recognition field. A surface classification tool could provide useful quantitative feedback to athletes and coaches in terms of running technique effort on varied terrain surfaces, improve training personalization, prevent injuries, and improve performance.

BACKGROUND:The influence of gait speed on lower-extremity coordination while walking in people with Parkinson's disease (pwPD) is poorly understood. This study sought to investigate the relationship between gait speed and hip-knee coordination and coordination variability in older adults and pwPD. METHODS:A total of 27 pwPD and 21 healthy older adults were recruited. Participants walked in a straight line at slow, preferred, and fast walking speeds. Gait data were collected using inertial measurement units, and the kinematics of the hip and knee were calculated. Coordination and coordination variability at the hip-knee joint pair were determined using continuous relative phase. A repeated measures two-way ANCOVA tested the impact of gait speed on coordination and coordination variability, while group differences were evaluated using statistical parametric mapping (SPM). RESULTS:Neither the healthy older adults nor the pwPD adjusted their hip-knee coordination in response to changes in gait speed. pwPD also displayed a trend towards restricted hip and knee joint excursion compared to older adults, which may further limit their ability to adapt gait strategies. CONCLUSIONS:These findings suggest that interventions addressing both joint excursion and motor adaptability may be important for improving gait function in individuals with Parkinson's disease. Real-world applicability can be found in the potential of wearable sensors to become a valuable tool in routine clinical practice for both diagnosis and ongoing management. TRIAL REGISTRATION:The study is registered in the German Clinical Trials Register (DRKS00022998).

For children with cerebral palsy (CP), walking on uneven surfaces (US) is a challenging task essential for their engagement in their daily lives. This study aims to compare spatiotemporal parameters of multiple domains of walking (pace, rhythm, stability, variability) in children with spastic CP between gait on an uneven surface (US) and an even surface (ES) and assess differences against their typically developing (TD) peers. A total of 34 children (17CP/17TD) walked at a self-selected speed on an US and an ES. Gait speed, stride length, stride time, walk ratio, cadence, double and single support time, and stride width were calculated. For each parameter, stride-to-stride variability was calculated using the coefficient of variation. A 2-way ANOVA (group, surface) was conducted on each parameter. Stride width, and variability of gait speed, cadence, and walk ratio presented a group × surface interaction (p ≤ 0.042). Post-hoc tests revealed a greater stride width, and variability of gait speed, and walk ratio in the CP, compared to the TD group (p ≤ 0.005) only on an US, and on both surfaces for cadence variability (p = 0.017). Gait analysis on an US reveals gait changes in children with CP, highlighting the importance of using more ecological approaches for gait assessment.

The Oxford Foot Model (OFM) is a widely-used multi-segment foot model for the evaluation of foot motion. To date, custom code based on the original scientific publications have failed to reproduce results available through the Vicon plug-in (ViconOFM). This highlights a lack of transparency, affecting the accessibility and understanding of the model. Therefore, the aims of this study are to (1) replicate ViconOFM using Python for open-source distribution (openOFM v1.0) and (2) reproduce the original scientific description of the OFM in a second version (openOFM v1.1), highlighting differences between both versions. A dataset comprising one healthy adult and a set of five patients with heterogeneous foot pathologies was used for analyses. Evaluation was conducted using the normalised root mean square error (NRMSE) between the inter-segment angles and arch heights of both implementations. The openOFM v1.1 was developed based on the original OFM publications. The average NRMSE between ViconOFM and openOFM v1.0, using both healthy and pathological gait, was of 0.0012. Based on our openOFM v1.1 implementation, differences between ViconOFM and the original OFM description from the literature are due to an integrated smoothing and gap filling function and changes in segment definitions. The negligible differences between ViconOFM and openOFM v1.0 in healthy and pathological gait supports the concurrent validity of openOFM. Providing users with both openOFM versions enables informed use of either model and allows further investigation into the implications of these differences. The open-source nature of the project promotes further development.

Objective: To develop a deep learning framework to predict lower-limb joint kinematics from inertial measurement unit (IMU) data across multiple gait tasks (walking, jogging, and running) and evaluate the impact of dynamic time warping (DTW) on reducing prediction errors. Patients and Methods: Data were collected from 18 participants fitted with IMUs and an optical motion capture system between May 25, 2023, and May 30, 2023. A long short-term memory autoencoder supervised regression model was developed. The model consisted of multiple long short-term memory and convolution layers. Acceleration and gyroscope data from the IMUs in 3 axes and their magnitude for the proximal and distal sensors of each joint (hip, knee, and ankle) were inputs to the model. Optical motion capture kinematics were considered ground truth and used as an output to train the prediction model. Results: The deep learning models achieved a root-mean-square error of less than 6° for hip, knee, and ankle joint sagittal plane angles, with the ankle showing the lowest error (5.1°). Task-specific models reported enhanced performance during certain gait phases, such as knee flexion during running. The application of DTW significantly reduced root-mean-square error across all tasks by at least 3° to 4°. External validation of independent data confirmed the model’s generalizability. Conclusion: Our findings underscore the potential of IMU-based deep learning models for joint kinematic predictions, offering a practical solution for remote and continuous biomechanical assessments in health care and sports science.

(1) Background: Human walking involves adapting to diverse terrains, influencing gait biomechanics. This study examined how seven outdoor surfaces—flat–even, banked-right/-left, cobblestone, grass, sloped-down, and sloped-up—affect nonlinear gait dynamics in 30 healthy adults (14 females and 15 males). (2) Methods: Trunk and shank accelerations were analyzed for movement predictability (sample entropy, SE), smoothness (log dimensionless jerk, LDLJ), symmetry (step/stride regularity), and stability (short-/long-term Lyapunov exponents, LyEs, LyEl). (3) Results: Surface type significantly influenced all gait metrics, regardless of sex. Banked-right and sloped-down walking reduced SE, indicating less predictable movements. All surfaces except flat–even increased LDLJ, suggesting reduced smoothness. Cobblestone and sloped-down surfaces impaired step symmetry, while banked surfaces enhanced stride symmetry. LyEs decreased on cobblestones (lower variability), while sloped-up increased it. LyEl rose on all surfaces except cobblestones, indicating a more chaotic gait. No significant sex differences were found, though males showed a non-significant trend toward lower LyEs. Notably, sex–surface interactions emerged for SE and stride symmetry on banked-right surfaces, with females showing decreased SE and increased symmetry. (4) Conclusions: These findings underscore the importance of terrain and sex in gait dynamics research.

The remote monitoring of walking behavior, such as step counts, using wearables is increasingly common in research and clinical settings. However, contextless step counts have limited value for assessing orthopedic patients recovering from trauma or surgery. Instead, step analysis on specific surfaces, like stairs or slopes, is more relevant for evaluating rehabilitation progress and tailoring therapeutic decision making. Algorithms that classify walking surfaces from inertial measurement unit (IMU) signals have been developed, but the lack of standardized and practical methods for IMU data collection and analysis has hindered the creation of robust, generalizable models. This study investigates whether simplifying IMU-based gait monitoring through deep learning (DL) models—via sensor reduction (fewer sensors/signals) or surface class grouping—can maintain or improve classification performance. Data were sourced from an open-science multi-modal Gait Database (Losing et. al, 2022), comprising 20 subjects (5 female, 15 male; 18–69 years) walking on different urban surfaces wearing lycra suits embedded with 17 IMU sensors. Our baseline DL model was limited to lower-body sensors (pelvis, thighs, shanks & feet) which we deemed a feasible, but still encompassing setup. The original surface classes (n=5) included flat walking, stairs up/down, and slopes up/down. Using a previously validated Bi-Directional CNN-LSTM model with batch normalization (Vinco et. al, 2024), we tested variations which included: (a) acceleration signals only, (b) a single pelvis sensor, and (c) surface class reduction to two groups: i) by type (flat, slope, stairs) and ii) by elevation (flat, up, down). Simplifications aimed to maximize usability by enhancing patient compliance (single pelvis sensor) or battery life (acceleration-only signals). Using the baseline sensor arrangement and both the acceleration and gyroscope signals, the model achieved 84.01% accuracy (F1: 0.77–0.98). Using only acceleration, lower-body sensors reached 65.75% accuracy (F1: 0.53–0.96), and the pelvis sensor alone scored 67.08% (F1: 0.63–0.75). With surface type grouping, lower-body sensors achieved 86.86% accuracy (F1: 0.74–0.94); pelvis sensor, 65.80%. Grouping by elevation improved lower-body sensor accuracy to 89.69% (F1: 0.83–0.92) and pelvis sensor to 85.27% (F1: 0.69–0.94). High classification accuracy (>85%) was achieved, even with a single pelvis sensor and grouped classes. While sensor reduction decreased performance, grouping by elevation produced comparable results. These findings suggest practical, patient-compliant solutions for gait monitoring with potential clinical applications. Including gyroscope data appears effective towards model performance especially when many surface types need to be classified.

Purpose:The modified Star Excursion Balance Test (mSEBT) is a clinical tool for dynamic balance assessment. While researchers have examined individual joint kinematic predictors of dynamic balance performance, limited data exist on body and joint sway during the test execution. Further investigation of kinematic predictors' influence on dynamic balance is needed to improve assessment methodologies. This study aimed to examine the relationship between the center of mass and lower limb kinematics as predictors of mSEBT performance. Patients and Methods:Twenty-seven participants with no history of lower limb joint instability were recruited for this study. The inertial sensors were positioned on the non-dominant leg: trunk, thigh, shank, and foot. The participants completed the mSEBT barefoot following standardized practice trials with three test trials per direction. The reach distance and lower limb kinematic data were recorded. Spearman rank's correlation and stepwise multiple regression analyses identified key predictors of dynamic balance performance. Results:Ankle dorsiflexion was a strong predictor of normalized reach distance in the anterior direction (r² = 0.34, p < 0.001). Body center of mass displacement was the strongest predictor of posteromedial and posterolateral reach (r² = 0.55, p < 0.001; r2 = 0.57, p = < 0.001, respectively). The combined influence of the body center of mass and hip flexion accounted for 65% of the variance in the posterior reach assessments. Conclusion:This study highlights the key biomechanical factors that influence dynamic balance, focusing on the interaction between joint mobility and segmental control. Ankle dorsiflexion is critical for anterior balance, whereas hip flexion and body center of mass displacement are essential for posterior balance.

Coughing behavior is associated with conditions such as sleep apnea, asthma, and chronic obstructive pulmonary disorder and can severely affect quality of life in those affected. In this context, coughing quantification is often important, but routinely performed via questionnaires. This approach is dependent on patient compliance or recall, which may affect validity and be especially difficult for nocturnal coughs. Manual review of audio recordings is potentially more accurate, but raises privacy concerns due to the collection and review of sensitive audio-data by a human annotator. Today, machine learning approaches are increasingly used to quantify coughs; however, algorithms often rely on microphone recordings, resulting in the same privacy issues, especially if data are sent to a remote server for analysis. The aims of this study are to determine if (1) a suite of sensors, excluding microphone recordings, can accurately detect coughs unobtrusively and (2) what the relative importance of each sensor-type on model performance may be. Data from 44 healthy young adult participants performing on-demand coughs and other tasks (breathing, talking, throat clearing, laughing, sniffing) in supine and sitting conditions were collected for this observational, cross-sectional study using a multi-sensor smart-garment device. Synchronized video was used to annotate tasks. Three-dimension acceleration, respiration (inductance plethysmography), and electrical activity (electrocardiography) signals were extracted into 1 s strips and binarized into coughs and non-coughs. Data were split into train and test sets using an inter-subject 80:20 split, ensuring that data from a particular participant are found in a single set. This procedure was repeated 10 times with different random inter-subject splits to assess the variability of results. Statistical and frequency-based features were computed and used as inputs to a Random Forest Classifier to predict classes (cough vs not-cough). Model hyperparameters were tuned to maximize F1-score using five-fold cross validation of the training set. Final model performance was assessed using F1-score, precision, and recall (sensitivity) on the test sets with mean (standard deviation) reported. Single sensor models based on acceleration, respiration, or electrocardiography revealed F1 scores of 92.6 (1.2)%, 88.9 (3.2)%, and 77.5 (3.4)%, respectively. Overall, the dual (acceleration, respiration) sensor model achieved the highest performance (F1-score 93.0 (1.1)%, precision 84.2 (4.2)%, and recall 95.5 (1.6)%). The multi-modal wearable device was able to distinguish coughs from other respiratory maneuvers, with acceleration and respiration sensors providing the most valuable information. Future studies could implement this approach for remote monitoring of coughs in patients suffering from coughing symptoms.

This study evaluates the ability of body segment kinematic data to identify skating tasks in ice hockey using machine learning models and compares the performance of models trained on different body segments. We employed XGBoost, Support Vector Machine and Random Forest models to classify four primary ice-hockey skating tasks: forward skating start and strides, skating stop & go, and skating into a wrist shot. Trunk, pelvis, thigh, shank, and foot segment centre of mass linear accelerations were derived from retro-reflective markers and used as inputs for feature engineering. The models were trained and evaluated using a 10-fold cross-validation stratified by participant. Overall, the machine learning models demonstrated strong performance, with mean accuracy scores ranging from 86.5% to 98.9%. The pelvis yielded the best overall performance, followed by the trunk and foot, whereas the thigh segment generally exhibited lower accuracies across models. These results indicate that prediction performance depends on the body segment kinematic data used as input. This study highlights the potential of body segment kinematic data for automated identification of ice hockey skating tasks, providing insights into sports analytics and player performance assessment.

As advancements in inertial measurement units (IMUs) for motion analysis progress, the inability to directly apply decades of research-based optical motion capture (OMC) methodologies presents a significant challenge. This study aims to bridge this gap by proposing an innovative deep learning approach to predict marker positions from IMU data, allowing traditional OMC-based calculations to estimate joint kinematics. Eighteen participants walked on a treadmill with seven IMUs and retroreflective markers. Trials were divided into normalized gait cycles (101 frames), and an autoencoder network with a custom Biomech loss function was used to predict 16 marker positions from IMU data. The model was validated using the leave-one-subject-out method and assessed using root mean squared error (RMSE). Joint angles in the sagittal plane were calculated using OMC methods, and RMSE was computed with and without alignment using dynamic time warping (DTW). The models were also tested on external datasets. Marker predictions achieved RMSE values of 2–4 cm, enabling joint angle predictions with 4–7° RMSE without alignment and 2–4° RMSE after DTW for sagittal plane joint angles (ankle, knee, hip). Validation using separate and open-source datasets confirmed the model’s generalizability, with similar RMSE values across datasets (4–7° RMSE without DTW and 2–4° with DTW). This study demonstrates the feasibility of applying conventional biomechanical models to IMUs, enabling accurate movement analysis and visualization outside controlled environments. This approach to predicting marker positions helps to bridge the gap between IMUs and OMC systems, enabling decades of research-based biomechanical methodologies to be applied to IMU data.