The use of wearable technology and machine learning to prevent musculoskeletal injury in military training

The following thesis investigates approaches and analyses for quantifying the arduous, injuriousinjurious, and repetitive nature of military training. This type of training has been historically reported as a contributor to musculoskeletal injury over the past 30 years. As a result, this has placed a significant strain on recruitment, rehabilitation, and retainment in military populations worldwide. The thesis is written from a mathematics and sports engineering background. Specifically, making use of machine learning, deep learning, physical modelling, and optimization to best understand the lower limb demands of military training for new recruits using data derived from wearable technology. The tools described within could potentially be applied for injury prevention, thus reducing the present strain on the Ministry of Defence’s rehabilitation, financial and training resources.

Chapter 1 – A general overview of the project and relevant literature is given. The history of military training is reviewed, referring to key events in time that have affected training methods and injury causation. A detailed summary of attempts to reduce the incidence of injury to reach current government and ministry of defence targets is also included, highlighting the current gap in research and conspicuous need for a more detailed appraisal of the current variety of injurious training programmes administered around the world. A brief introduction into the theory and application of machine learning and wearable technology are also provided to give context and help with the understanding of later Chapters.

Chapter 2 – A review into the nature of training over the past 30 years is provided in the form of a systematic review and meta-analysis. This review set out to report musculoskeletal and stress fracture injury incidence for a variety of initial military training programmes delivered worldwide. Furthermore, to analyse factors associated with the abovementioned injuries and report any preventive measures, interventions and predictive models that have been taken or created to reduce the onset of musculoskeletal injury in military training. In total, 81 studies were reviewed that included initial military training from around the world. Meta analyses showed the average period prevalence for musculoskeletal injuries and stress fractures specifically was equal to 32.3% and 6.2% respectively. Age and sex of recruits were consistently reported as significant risk factors for musculoskeletal injury and stress fracture. Pooled meta-analysis of studies that reported age as a risk factor showed those aged over 25 had a relative risk of 1.40 (1-10-1.79) and 2.15 (1.72-2.67) for musculoskeletal injury and stress fracture respectively compared to recruits aged 18 and under. Similarly, females had a relative risk of 2.07 (1.29-3.32) and 2.45 (1.97-3.04) for musculoskeletal injuries and stress fractures respectively compared to male recruits. A wide variety of other risk factors have also been identified. Several studies have successfully presented predictions of musculoskeletal and stress fracture cases using logistic regression modelling. The review gives a detailed insight into period prevalence of musculoskeletal injury and stress fractures in military training along with consistent and contradictory risk factors. However, the review clearly identifies a current gap in the literature that warrants further investigation. This gap included the rich objective measurement of both the intensity and type of lower limb loading throughout the duration of a training programme.

As a result, the development of a series of methods to quantify lower limb loading derived from an accelerometer (an affordable and feasible device) placed on the lower leg are fully described and evaluated in Chapters 3 and 4. In Chapter 3, several machine learning predictors and a novel non-linear auto-regressive neural network are tested to predict peak vertical ground reaction forces of several movements often seen during military training. Predictions were extensively tested against data collected from a gold standard fixed force plate. Analysis through different data preparation techniques highlighted that a Gaussian process regression technique is the best predictor of peak vertical ground reaction force. Results of this method showed. Resulting in root mean square error for all predictions of 0.0003 body weights and limits of agreement equal to 0.0006 body weights of force.

Where Chapter 3 provides a prediction of the peak vertical force for a variety of movements, Chapter 4 provides a classification of each type of load. Several machine learning and deep learning classifiers were again developed and tested to recognise a combination of activities that would normally be performed in a military environment. As in Chapter 3, this Chapter provided an optimised model which turned out to be a simple two-layer feed-forward neural network that achieved a classification accuracy of 96.95% for all activities. The largest inaccuracies in developed models were shown between marching and weighted marching movements.

Chapter 5 – The feasibility and acceptability of a selection of desired wearable and nearable technologies were tested in a small sample (n=20) for two weeks of training in this Chapter. All participants completed the study and on average recorded 322.4, 1032.3 and 510.3 minutes of data for heart rate monitors, accelerometers, and sleep ballistocardiography respectively. Descriptive measures for the abovementioned measures from Chapters 3 and 4 along with other accelerometer summary statistics were collected. No significant differences were observed between sleep variables, but sedentary periods calculated from accelerometer data and time spent in HR zones 1,2,5 and 6 were significantly different between weeks 1 and 2 of training. In conclusion, this study determined that an accelerometer worn on the lower leg, a chest worn heart rate monitor and an under-mattress sleep sensor are all feasible and acceptable tools to measure lower limb loading and recovery (sleep) in a military environment. Furthermore, the observed differences between accelerometer and HR data over this short period highlights that these technologies could be useful in highlighting changes in the demands of training over time and a potential tool for identifying increased risk to injury onset.

Chapter 6 – In this chapter, the findings from previous studies were applied towards the full objective measurement of lower limb loading for the duration of a phase one training program (Chapter 6) is provided. Quantification of the lower limb demands, and activity type is were provided predicted for each second, minute, hour, and week of the training programme. Between weeks, significant differences were observed between measures of predicted ground reaction force, acceleration vector magnitude and physical activity categories (sedentary, light, moderate and vigorous). However, compliance was noticeably low (45.37%) compared to previous tests where compliance was excellent (100%). Nevertheless, the methods presented provide a first of its kind representation of the lower limb demands of a military training programme along with a detailed methodology that can be repeated and improved to assess injury occurrence and individual characteristics of recruits throughout training.

Chapter 7 - In conclusion, the introduction and feasibility of a series of novel lower limb loading measures has been demonstrated within this thesis. It is recommended that the measurement of lower limb force and activity type through the use of accelerometer data is followed up with a larger cohort to test the methodology displayed in Chapter 6. If successful, this data could be applied in future cohorts to reduce injuries for military recruits during their initial training.