Feature Learning Networks for Floor Sensor-based Gait Recognition.

Deep learning (DL) has become a powerful tool in many image classification applications but often requires large training sets to achieve high accuracy. For applications where the available data are limited, this can become a severely limiting factor in model performance. To address this limitation, feature learning network approaches that integrate traditional feature extraction methods with DL frameworks have been proposed. In this study, the performances of traditional methods: discrete wavelet transform (DWT), discrete cosine transform (DCT), independent component analysis (ICA), and principal component analysis (PCA); and their corresponding feature networks based on a convolutional neural network (CNN) framework: ScatNet (wavelet scattering network), DCTNet, ICANet, and PCANet, were investigated for use in pressure-based footstep recognition when the limited sample size is available for person authentication. The results show that the feature learning networks (90.6% accuracy) achieved significantly better performance on average than the conventional feature extraction methods (79.7% accuracy) (p < 0.05). Among the different feature networks, PCANet provided the best verification performance, with an accuracy of 92.2%. Feature learning networks are simple and effective approaches that can be a promising solution for applications like floor-based gait recognition in a security access scenario (such as workspace environment and border control) when small amounts of data are available for training models to differentiate between a larger group of users.

Infrared cavity ring-down spectroscopy for detecting non-small cell lung cancer in exhaled breath.

Early diagnosis of lung cancer greatly improves the likelihood of survival and remission, but limitations in existing technologies like low-dose computed tomography have prevented the implementation of widespread screening programs. Breath-based solutions that seek disease biomarkers in exhaled volatile organic compound (VOC) profiles show promise as affordable, accessible and non-invasive alternatives to traditional imaging. In this pilot work, we present a lung cancer detection framework using cavity ring-down spectroscopy (CRDS), an effective and practical laser absorption spectroscopy technique that has the ability to advance breath screening into clinical reality. The main aims of this work were to (1) test the utility of infrared CRDS breath profiles for discriminating non-small cell lung cancer (NSCLC) patients from controls, (2) compare models with VOCs as predictors to those with patterns from the CRDS spectra (breathprints) as predictors, and (3) present a robust approach for identifying relevant disease biomarkers. First, based on a proposed learning curve technique that estimated the limits of a model's performance at multiple sample sizes (10-158), the CRDS-based models developed in this work were found to achieve classification performance comparable or superior to like mass spectroscopy and sensor-based systems. Second, using 158 collected samples (62 NSCLC subjects and 96 controls), the accuracy range for the VOC-based model was 65.19%-85.44% (51.61%-66.13% sensitivity and 73.96%-97.92% specificity), depending on the employed cross-validation technique. The model based on breathprint predictors generally performed better, with accuracy ranging from 71.52%-86.08% (58.06%-82.26% sensitivity and 80.21%-88.54% specificity). Lastly, using a protocol based on consensus feature selection, three VOCs (isopropanol, dimethyl sulfide, and butyric acid) and two breathprint features (from a local binary pattern transformation of the spectra) were identified as possible NSCLC biomarkers. This research demonstrates the potential of infrared CRDS breath profiles and the developed early-stage classification techniques for lung cancer biomarker detection and screening.

Verification-Based Design of a Robust EMG Wake Word.

Surface electromyography (sEMG) signals are now commonly used in continuous myoelectric control of prostheses. More recently, researchers have considered EMG-based gesture recognition systems for human computer interaction research. These systems instead focus on recognizing discrete gestures (like a finger snap). The majority of works, however, have focused on improving multi-class performance, with little consideration for false activations from "other" classes. Consequently, they lack the robustness needed for real-world applications which generally require a single motion class such as a mouse click or a wake word. Furthermore, many works have borrowed the windowed classification schemes from continuous control, and thus fail to leverage the temporal structure of the gesture. In this paper, we propose a verification-based approach to creating a robust EMG wake word using one-class classifiers (Support Vector Data Description, One Class-Support Vector Machine, Dynamic Time Warping (DTW) & Hidden Markov Models). The area under the ROC curve (AUC) is used as a feature optimization objective as it provides a better representation of the verification performance. Equal error rate (EER) and AUC are then used as evaluation metrics. The results are computed using both window-based and temporal classifiers on a dataset consisting of five different gestures, with a best EER of 0.04 and AUC of 0.98, recorded using a DTW scheme. These results demonstrate a design framework that may benefit the development of more robust solutions for EMG-based wake words or input commands for a variety of interactive applications.

Data Pre-Processing of Infrared Spectral Breathprints for Lung Cancer Detection.

Though breath analysis shows promise as a noninvasive and cost-effective approach to lung cancer screening, biomarkers in exhaled breath samples can be overwhelmed by irrelevant internal and environmental volatile organic compounds (VOCs). These extraneous VOCs can obscure the disease signature in a spectral breathprint, hindering the performance of pattern recognition models. In this work, pre-processing pipelines consisting of missing value replacement, detrending, and normalization techniques were evaluated to reduce these effects and enhance the features of interest in infrared cavity ring-down spectra. The best performing pipeline consisted of moving average detrending, linear interpolation for missing values, and vector normalization. This model achieved an average accuracy of 73.04% across five types of classifiers, exhibiting an 8.36% improvement compared to a baseline model (p < 0.05). A linear support vector machine classifier yielded the best performance (79.75% accuracy, 67.74% sensitivity, 87.50% specificity). This work can serve to guide pre-processing in future lung cancer breath research and, more broadly, in infrared laser absorption spectroscopy in general.

Machine Learning Model Validation for Early Stage Studies with Small Sample Sizes.

In early stage biomedical studies, small datasets are common due to the high cost and difficulty of sample collection with human subjects. This complicates the validation of machine learning models, which are best suited for large datasets. In this work, we examined feature selection techniques, validation frameworks, and learning curve fitting for small simulated datasets with known underlying discriminability, with the aim of identifying a protocol for estimating and interpreting early stage model performance and for planning future studies. Of a variety of examined validation configurations, a nested cross-validation framework provided the most accurate reflection of the selected features' discriminability, but the relevant features were often not properly identified during the feature selection stage for datasets with small sample sizes. Ultimately, we recommend that: (1) filter-based feature selection methods should be used to minimize overfitting to noise-based features, (2) statistical exploration should be conducted on datasets as a whole to estimate the level of discriminability and the feasibility of the classification problems, and (3) learning curves should be employed using nested cross-validation performance estimates for forecasting accuracy at larger sample sizes and estimating the required number of samples to converge towards best performance. This work should serve as a guideline for researchers incorporating machine learning in small-scale pilot studies.

Nanoparticle surface-enhanced Raman spectroscopy as a noninvasive, label-free tool to monitor hematological malignancy.

Aim: Monitoring minimal residual disease remains a challenge to the effective medical management of hematological malignancies; yet surface-enhanced Raman spectroscopy (SERS) has emerged as a potential clinical tool to do so. Materials & methods: We developed a cell-free, label-free SERS approach using gold nanoparticles (nanoSERS) to classify hematological malignancies referenced against two control cohorts: healthy and noncancer cardiovascular disease. A predictive model was built using machine-learning algorithms to incorporate disease burden scores for patients under standard treatment upon. Results: Linear- and quadratic-discriminant analysis distinguished three cohorts with 69.8 and 71.4% accuracies, respectively. A predictive nanoSERS model correlated (MSE = 1.6) with established clinical parameters. Conclusion: This study offers a proof-of-concept for the noninvasive monitoring of disease progression, highlighting the potential to incorporate nanoSERS into translational medicine.

Cancer patient quality of life is achieved by reassurance from informed doctors using the best clinical tools. Confirming the earliest detection or absence of disease ensures treatment is timely and recovery optimal. Here we show the potential for a new tool to be developed to reassure patients and inform doctors. We examined the 'chemical fingerprints' (Raman spectroscopic profiling) of patient's blood, enhanced by gold nanoparticles with a double-referenced machine learning algorithm. Teaching a machine to learn as it works ensures it is improving how it finds clinically important features in the chemical fingerprint. This helps patients live more confidently with cancer or in cancer recovery. Eventually, once fully trained and translated into a real-world hospital application, this could improve patient outcomes and quality of life.

Deep Cross-User Models Reduce the Training Burden in Myoelectric Control.

The effort, focus, and time to collect data and train EMG pattern recognition systems is one of the largest barriers to their widespread adoption in commercial applications. In addition to multiple repetitions of motions, including exemplars of confounding factors during the training protocol has been shown to be critical for robust machine learning models. This added training burden is prohibitive for most regular use cases, so cross-user models have been proposed that could leverage inter-repetition variability supplied by other users. Existing cross-user models have not yet achieved performance levels sufficient for commercialization and require users to closely adhere to a training protocol that is impractical without expert guidance. In this work, we extend a previously reported adaptive domain adversarial neural network (ADANN) to a cross-subject framework that requires very little training data from the end-user. We compare its performance to single-repetition within-user training and the previous state-of-the-art cross-subject technique, canonical correlation analysis (CCA). ADANN significantly outperformed CCA for both intact-limb (86.8-96.2%) and amputee (64.1-84.2%) populations. Moreover, the ADANN adaptation computation time was substantially lower than the time otherwise devoted to conducting a full within-subject training protocol. This study shows that cross-user models, enabled by deep-learned adaptations, may be a viable option for improved generalized pattern recognition-based myoelectric control.

A Transferable Adaptive Domain Adversarial Neural Network for Virtual Reality Augmented EMG-Based Gesture Recognition.

Within the field of electromyography-based (EMG) gesture recognition, disparities exist between the offline accuracy reported in the literature and the real-time usability of a classifier. This gap mainly stems from two factors: 1) The absence of a controller, making the data collected dissimilar to actual control. 2) The difficulty of including the four main dynamic factors (gesture intensity, limb position, electrode shift, and transient changes in the signal), as including their permutations drastically increases the amount of data to be recorded. Contrarily, online datasets are limited to the exact EMG-based controller used to record them, necessitating the recording of a new dataset for each control method or variant to be tested. Consequently, this paper proposes a new type of dataset to serve as an intermediate between offline and online datasets, by recording the data using a real-time experimental protocol. The protocol, performed in virtual reality, includes the four main dynamic factors and uses an EMG-independent controller to guide movements. This EMG-independent feedback ensures that the user is in-the-loop during recording, while enabling the resulting dynamic dataset to be used as an EMG-based benchmark. The dataset is comprised of 20 able-bodied participants completing three to four sessions over a period of 14 to 21 days. The ability of the dynamic dataset to serve as a benchmark is leveraged to evaluate the impact of different recalibration techniques for long-term (across-day) gesture recognition, including a novel algorithm, named TADANN. TADANN consistently and significantly ( [Formula: see text]) outperforms using fine-tuning as the recalibration technique.

Assessment of EMG Benchmark Data for Gesture Recognition Using the NinaPro Database.

In recent years, many electromyography (EMG) benchmark databases have been made publicly available to the myoelectric control research community. Many small laboratories that lack the instrumentation, access, and experience needed to collect quality EMG data have used these benchmark datasets to explore and propose new signal processing and pattern recognition algorithms. It is widely accepted that noise contamination can affect the performance of myoelectric control systems, and so useful datasets should maintain good signal quality to ensure accurate results for proposed EMG-based gesture recognition systems. Despite the availability and adoption of benchmarks datasets, however, the quality of the EMG signals in these benchmarks has not yet been examined. In this study, the signal quality of the Non-Invasive Adaptive Prosthetics (NinaPro) dataset, the most widely known publicly available benchmark database to date, was comprehensively investigated with the goals of: 1) reporting the level of noise contamination in each NinaPro sub-dataset, 2) proposing signal quality criteria for assessing EMG datasets, 3) analyzing the effect of signal quality on classification performance, and 4) examining the quality of the data labels.

Wearable EMG-Based Gesture Recognition Systems During Activities of Daily Living: An Exploratory Study.

Recent advancements in wearable technologies have increased the potential for practical gesture recognition systems using electromyogram (EMG) signals. However, despite the high classification accuracies reported in many studies (> 90%), there is a gap between academic results and industrial success. This is in part because state-of-the-art EMG-based gesture recognition systems are commonly evaluated in highly-controlled laboratory environments, where users are assumed to be resting and performing one of a closed set of target gestures. In real world conditions, however, a variety of non-target gestures are performed during activities of daily living (ADLs), resulting in many false positive activations. In this study, the effect of ADLs on the performance of EMG-based gesture recognition using a wearable EMG device was investigated. EMG data for 14 hand and finger gestures, as well as continuous activity during uncontrolled ADLs (>10 hours in total) were collected and analyzed. Results showed that (1) the cluster separability of 14 different gestures during ADLs was 171 times worse than during rest; (2) the probability distributions of EMG features extracted from different ADLs were significantly different (p <; 0.05). (3) of the 14 target gestures, a right angle gesture (extension of the thumb and index finger) was least often inadvertently activated during ADLs. These results suggest that ADLs and other non-trained gestures must be taken into consideration when designing EMG-based gesture recognition systems.

Fractal Analysis of Human Gait Variability via Stride Interval Time Series.

Fractal analysis of stride interval time series is a useful tool in human gait research which could be used as a marker for gait adaptability, gait disorder, and fall risk among patients with movement disorders. This study is designed to systematically and comprehensively investigate two practical aspects of fractal analysis which significantly affect the outcome: the series length and the parameters used in the algorithm. The Hurst exponent, scaling exponent, and/or fractal dimension are computed from both simulated and experimental data using three fractal methods, namely detrended fluctuation analysis, box-counting dimension, and Higuchi's fractal dimension. The advantages and drawbacks of each method are discussed, in terms of biases and variability. The results demonstrate that a careful selection of fractal analysis methods and their parameters is required, which is dependent on the aim of study (either analyzing differences between experimental groups or estimating an accurate determination of fractal features). A set of guidelines for the selection of the fractal methods and the length of stride interval time series is provided, along with the optimal parameters for a robust implementation for each method.

Interpreting Deep Learning Features for Myoelectric Control: A Comparison With Handcrafted Features.

Existing research on myoelectric control systems primarily focuses on extracting discriminative characteristics of the electromyographic (EMG) signal by designing handcrafted features. Recently, however, deep learning techniques have been applied to the challenging task of EMG-based gesture recognition. The adoption of these techniques slowly shifts the focus from feature engineering to feature learning. Nevertheless, the black-box nature of deep learning makes it hard to understand the type of information learned by the network and how it relates to handcrafted features. Additionally, due to the high variability in EMG recordings between participants, deep features tend to generalize poorly across subjects using standard training methods. Consequently, this work introduces a new multi-domain learning algorithm, named ADANN (Adaptive Domain Adversarial Neural Network), which significantly enhances (p = 0.00004) inter-subject classification accuracy by an average of 19.40% compared to standard training. Using ADANN-generated features, this work provides the first topological data analysis of EMG-based gesture recognition for the characterization of the information encoded within a deep network, using handcrafted features as landmarks. This analysis reveals that handcrafted features and the learned features (in the earlier layers) both try to discriminate between all gestures, but do not encode the same information to do so. In the later layers, the learned features are inclined to instead adopt a one-vs.-all strategy for a given class. Furthermore, by using convolutional network visualization techniques, it is revealed that learned features actually tend to ignore the most activated channel during contraction, which is in stark contrast with the prevalence of handcrafted features designed to capture amplitude information. Overall, this work paves the way for hybrid feature sets by providing a clear guideline of complementary information encoded within learned and handcrafted features.

Current Trends and Confounding Factors in Myoelectric Control: Limb Position and Contraction Intensity.

This manuscript presents a hybrid study of a comprehensive review and a systematic(research) analysis. Myoelectric control is the cornerstone ofmany assistive technologies used in clinicalpractice, such as prosthetics and orthoses, and human-computer interaction, such as virtual reality control.Although the classification accuracy of such devices exceeds 90% in a controlled laboratory setting,myoelectric devices still face challenges in robustness to variability of daily living conditions.The intrinsic physiological mechanisms limiting practical implementations of myoelectric deviceswere explored: the limb position effect and the contraction intensity effect. The degradationof electromyography (EMG) pattern recognition in the presence of these factors was demonstratedon six datasets, where classification performance was 13% and 20% lower than the controlledsetting for the limb position and contraction intensity effect, respectively. The experimental designsof limb position and contraction intensity literature were surveyed. Current state-of-the-art trainingstrategies and robust algorithms for both effects were compiled and presented. Recommendationsfor future limb position effect studies include: the collection protocol providing exemplars of at least 6positions (four limb positions and three forearm orientations), three-dimensional space experimentaldesigns, transfer learning approaches, and multi-modal sensor configurations. Recommendationsfor future contraction intensity effect studies include: the collection of dynamic contractions, nonlinearcomplexity features, and proportional control.

Topology highlights mesoscopic functional equivalence between imagery and perception: The case of hypnotizability.

The functional equivalence (FE) between imagery and perception or motion has been proposed on the basis of neuroimaging evidence of large spatially overlapping activations between real and imagined sensori-motor conditions. However, similar local activation patterns do not imply the same mesoscopic integration of brain regions, which can be described by tools from Topological Data Analysis (TDA). On the basis of behavioral findings, stronger FE has been hypothesized in the individuals with high scores of hypnotizability scores (highs) with respect to low hypnotizable participants (lows) who differ between each other in the proneness to modify memory, perception and behavior according to specific imaginative suggestions. Here we present the first EEG evidence of stronger FE in highs. In fact, persistent homology shows that the highs EEG topological asset during real and imagined sensory conditions is significantly more similar than the lows. As a corollary finding, persistent homology shows lower restructuring of the EEG asset in highs than in lows during both sensory and imagery tasks with respect to basal conditions. Present findings support the view that greater embodiment of mental images may be responsible for the highs greater proneness to respond to sensori-motor suggestions and to report involuntariness in action. In addition, findings indicate hypnotizability-related sensory and cognitive information processing and suggest that the psycho-physiological trait of hypnotizability may modulate more than one aspect of the everyday life.

Classification of higher- and lower-mileage runners based on running kinematics.

BACKGROUND: Running-related overuse injuries can result from the combination of extrinsic (e.g., running mileage) and intrinsic risk factors (e.g., biomechanics and gender), but the relationship between these factors is not fully understood. Therefore, the first purpose of this study was to determine whether we could classify higher- and lower-mileage runners according to differences in lower extremity kinematics during the stance and swing phases of running gait. The second purpose was to subgroup the runners by gender and determine whether we could classify higher- and lower-mileage runners in male and female subgroups. METHODS: Participants were allocated to the "higher-mileage" group (≥32 km/week; n = 41 (30 females)) or to the "lower-mileage" group (≤25 km; n = 40 (29 females)). Three-dimensional kinematic data were collected during 60 s of treadmill running at a self-selected speed (2.61 ± 0.23 m/s). A support vector machine classifier identified kinematic differences between higher- and lower-mileage groups based on principal component scores. RESULTS: Higher- and lower-mileage runners (both genders) could be separated with 92.59% classification accuracy. When subgrouping by gender, higher- and lower-mileage female runners could be separated with 89.83% classification accuracy, and higher- and lower-mileage male runners could be separated with 100% classification accuracy. CONCLUSION: These results demonstrate there are distinct kinematic differences between subgroups related to both mileage and gender, and that these factors need to be considered in future research.

Feature Extraction and Selection for Pain Recognition Using Peripheral Physiological Signals.

In pattern recognition, the selection of appropriate features is paramount to both the performance and the robustness of the system. Over-reliance on machine learning-based feature selection methods can, therefore, be problematic; especially when conducted using small snapshots of data. The results of these methods, if adopted without proper interpretation, can lead to sub-optimal system design or worse, the abandonment of otherwise viable and important features. In this work, a deep exploration of pain-based emotion classification was conducted to better understand differences in the results of the related literature. In total, 155 different time domain and frequency domain features were explored, derived from electromyogram (EMG), skin conductance levels (SCL), and electrocardiogram (ECG) readings taken from the 85 subjects in response to heat-induced pain. To address the inconsistency in the optimal feature sets found in related works, an exhaustive and interpretable feature selection protocol was followed to obtain a generalizable feature set. Associations between features were then visualized using a topologically-informed chart, called Mapper, of this physiological feature space, including synthesis and comparison of results from previous literature. This topological feature chart was able to identify key sources of information that led to the formation of five main functional feature groups: signal amplitude and power, frequency information, nonlinear complexity, unique, and connecting. These functional groupings were used to extract further insight into observable autonomic responses to pain through a complementary statistical interaction analysis. From this chart, it was observed that EMG and SCL derived features could functionally replace those obtained from ECG. These insights motivate future work on novel sensing modalities, feature design, deep learning approaches, and dimensionality reduction techniques.

An Investigation of Temporally Inspired Time Domain Features for Electromyographic Pattern Recognition.

This paper presents a novel set of temporally inspiredtime domain features for electromyographic (EMG) pattern recognition. The proposed methods employ simple time series measures derived from peak detection, and could better reflect EMG activity over time. Multiple EMG datasets consisting of 68 able-bodied and transradial amputee subjects performing a large variety of hand, wrist, fingers, and grasping movements are used to evaluate the performance of the proposed features and to design robust EMG feature sets. The results show that the average classification accuracy of two novel features, the mean prominence of local peaks and valleys, outperform several commonly used time domain features, autoregressive coefficients, histogram, and zero crossing, by 8{\%, 11{\%, and 17{\%, respectively. The proposed features are also shown to provide additional information as part of a robust feature set when compared to the common Hudgins' timedomain feature set, as selected by sequential forward selection and through empirical feature set design.

Feature Extraction and Selection for Myoelectric Control Based on Wearable EMG Sensors.

Specialized myoelectric sensors have been used in prosthetics for decades, but, with recent advancements in wearable sensors, wireless communication and embedded technologies, wearable electromyographic (EMG) armbands are now commercially available for the general public. Due to physical, processing, and cost constraints, however, these armbands typically sample EMG signals at a lower frequency (e.g., 200 Hz for the Myo armband) than their clinical counterparts. It remains unclear whether existing EMG feature extraction methods, which largely evolved based on EMG signals sampled at 1000 Hz or above, are still effective for use with these emerging lower-bandwidth systems. In this study, the effects of sampling rate (low: 200 Hz vs. high: 1000 Hz) on the classification of hand and finger movements were evaluated for twenty-six different individual features and eight sets of multiple features using a variety of datasets comprised of both able-bodied and amputee subjects. The results show that, on average, classification accuracies drop significantly ( p.

Analysis of Big Data in Gait Biomechanics: Current Trends and Future Directions.

The increasing amount of data in biomechanics research has greatly increased the importance of developing advanced multivariate analysis and machine learning techniques, which are better able to handle "big data". Consequently, advances in data science methods will expand the knowledge for testing new hypotheses about biomechanical risk factors associated with walking and running gait-related musculoskeletal injury. This paper begins with a brief introduction to an automated three-dimensional (3D) biomechanical gait data collection system: 3D GAIT, followed by how the studies in the field of gait biomechanics fit the quantities in the 5 V's definition of big data: volume, velocity, variety, veracity, and value. Next, we provide a review of recent research and development in multivariate and machine learning methods-based gait analysis that can be applied to big data analytics. These modern biomechanical gait analysis methods include several main modules such as initial input features, dimensionality reduction (feature selection and extraction), and learning algorithms (classification and clustering). Finally, a promising big data exploration tool called "topological data analysis" and directions for future research are outlined and discussed.

Runners with patellofemoral pain demonstrate sub-groups of pelvic acceleration profiles using hierarchical cluster analysis: an exploratory cross-sectional study.

BACKGROUND: Previous studies have suggested that distinct and homogenous sub-groups of gait patterns exist among runners with patellofemoral pain (PFP), based on gait analysis. However, acquisition of 3D kinematic data using optical systems is time consuming and prone to marker placement errors. In contrast, axial segment acceleration data can represent an overall running pattern, being easy to acquire and not influenced by marker placement error. Therefore, the purpose of this study was to determine if pelvic acceleration patterns during running could be used to classify PFP patients into homogeneous sub-groups. A secondary purpose was to analyze lower limb kinematic data to investigate the practical implications of clustering these subjects based on 3D pelvic acceleration data. METHODS: A hierarchical cluster analysis was used to determine sub-groups of similar running profiles among 110 PFP subjects, separately for males (n = 44) and females (n = 66), using pelvic acceleration data (reduced with principal component analysis) during treadmill running acquired with optical motion capture system. In a secondary analysis, peak joint angles were compared between clusters (α = 0.05) to provide clinical context and deeper understanding of variables that separated clusters. RESULTS: The results reveal two distinct running gait sub-groups (C1 and C2) for female subjects and no sub-groups were identified for males. Two pelvic acceleration components were different between clusters (PC1 and PC5; p < 0.001). While females in C1 presented similar acceleration patterns to males, C2 presented greater vertical and anterior peak accelerations. All females presented higher and delayed mediolateral acceleration peaks than males. Males presented greater ankle eversion (p < 0.001), lower knee abduction (p = 0.007) and hip adduction (p = 0.002) than all females, and lower hip internal rotation than C1 (p = 0.007). CONCLUSIONS: Two distinct and homogeneous kinematic PFP sub-groups were identified for female subjects, but not for males. The results suggest that differences in running gait patterns between clusters occur mainly due to sex-related factors, but there are subtle differences among female subjects. This study shows the potential use of pelvic acceleration patterns, which can be acquired with accessible wearable technology (i.e. accelerometers).