Anterior cruciate ligament (ACL) injuries are common among athletes. Despite a successful return to sport (RTS) for most of the injured athletes, a significant proportion do not return to competitive levels, and thus RTS post ACL reconstruction still represents a challenge for clinicians. Wearable sensors, owing to their small size and low cost, can represent an opportunity for the management of athletes on-the-field after RTS by providing guidance to associated clinicians. In particular, this study aims to investigate the ability of a set of inertial sensors worn on the lower-limbs by rugby players involved in a change-of-direction (COD) activity to differentiate between healthy and post-ACL groups via the use of machine learning. Twelve male participants (six healthy and six post-ACL athletes who were deemed to have successfully returned to competitive rugby and tested in the 5–10 year period following the injury) were recruited for the study. Time- and frequency-domain features were extracted from the raw inertial data collected. Several machine learning models were tested, such as k-nearest neighbors, naïve Bayes, support vector machine, gradient boosting tree, multi-layer perceptron, and stacking. Feature selection was implemented in the learning model, and leave-one-subject-out cross-validation (LOSO-CV) was adopted to estimate training and test errors. Results obtained show that it is possible to correctly discriminate between healthy and post-ACL injury subjects with an accuracy of 73.07% (multi-layer perceptron) and sensitivity of 81.8% (gradient boosting). The results of this study demonstrate the feasibility of using body-worn motion sensors and machine learning approaches for the identification of post-ACL gait patterns in athletes performing sport tasks on-the-field even a number of years after the injury occurred.

Every year over 75 000 firefighters are injured and 159 die in the line of duty. Some of these accidents could be averted if first response team leaders had better information about the situation on the ground. The SAFESENS project is developing a novel monitoring system for first responders designed to provide response team leaders with timely and reliable information about their firefighters' status during operations, based on data from wireless inertial measurement units. In this paper we investigate if Gradient Boosted Trees (GBT) could be used for recognising 17 activities, selected in consultation with first responders, from inertial data. By arranging these into more general groups we generate three additional classification problems which are used for comparing GBT with k-Nearest Neighbours (kNN) and Support Vector Machines (SVM). The results show that GBT outperforms both kNN and SVM for three of these four problems with a mean absolute error of less than 7%, which is distributed more evenly across the target activities than that from either kNN or SVM.

The distinction between subject-dependent and subject-independent performance is ubiquitous in the human activity recognition (HAR) literature. We assess whether HAR models really do achieve better subject-dependent performance than subject-independent performance, whether a model trained with data from many users achieves better subject-independent performance than one trained with data from a single person, and whether one trained with data from a single specific target user performs better for that user than one trained with data from many. To those ends, we compare four popular machine learning algorithms’ subject-dependent and subject-independent performances across eight datasets using three different personalisation–generalisation approaches, which we term person-independent models (PIMs), person-specific models (PSMs), and ensembles of PSMs (EPSMs). We further consider three different ways to construct such an ensemble: unweighted, $\kappa$ -weighted, and baseline-feature-weighted. Our analysis shows that PSMs outperform PIMs by 43.5% in terms of their subject-dependent performances, whereas PIMs outperform PSMs by 55.9% and $\kappa$ -weighted EPSMs—the best-performing EPSM type—by 16.4% in terms of the subject-independent performance.

Human activity recognition (HAR) has become an increasingly popular application of machine learning across a range of domains. Typically the HAR task that a machine learning algorithm is trained for requires separating multiple activities such as walking, running, sitting, and falling from each other. Despite a large body of work on multi-class HAR, and the well-known fact that the performance on a multi-class problem can be significantly affected by how it is decomposed into a set of binary problems, there has been little research into how the choice of multi-class decomposition method affects the performance of HAR systems. This paper presents the first empirical comparison of multi-class decomposition methods in a HAR context by estimating the performance of five machine learning algorithms when used in their multi-class formulation, with four popular multi-class decomposition methods, five expert hierarchies—nested dichotomies constructed from domain knowledge—or an ensemble of expert hierarchies on a 17-class HAR data-set which consists of features extracted from tri-axial accelerometer and gyroscope signals. We further compare performance on two binary classification problems, each based on the topmost dichotomy of an expert hierarchy. The results show that expert hierarchies can indeed compete with one-vs-all, both on the original multi-class problem and on a more general binary classification problem, such as that induced by an expert hierarchy’s topmost dichotomy. Finally, we show that an ensemble of expert hierarchies performs better than one-vs-all and comparably to one-vs-one, despite being of lower time and space complexity, on the multi-class problem, and outperforms all other multi-class decomposition methods on the two dichotomous problems.

Human activity recognition (HAR) has a wide range of applications, such as monitoring ambulatory patients' recovery, workers for harmful movement patterns, or elderly populations for falls. These systems often operate in an environment where battery lifespan, power consumption, and hence computational complexity, are of prime concern. This work explores three methods for reducing the dimensionality of a HAR problem in the context of an emergency first responders monitoring system. We empirically estimate the accuracy of k-Nearest Neighbours, Support Vector Machines, and Gradient Boosted Trees when using different combinations of (A)ccelerometer, (G)yroscope and (P)ressure sensors. We then apply Principal Component Analysis for dimensionality reduction, and the Kruskal-Wallis test for feature selection. Our results show that the best combination is that which includes all three sensors (MAE: 3.6%), followed by the A/G (MAE: 3.7%), and the A/P combination (MAE 4.3%): the same as that when using the accelerometer alone. Moreover, our results show that the Kruskal-Wallis test can be used to discard up to 50% of the features, and yet improve the performance of classification algorithms.