A novel method to assist clinical management decisions following mTBI using wearable sensors and machine learning

Introduction: The consequences of traumatic brain injury (TBI) significantly affect athletic populations, military personnel, and the general public around the world. Mild TBI (mTBI) accounts for 95% of all TBIs, up to 50% of patients with these “mild” injuries will struggle to perform tasks that would typically be trivial for weeks/months post-injury1. Some patients present with functional impairments due to a predominantly physiological cause, while others suffer due to a primarily vestibular issue. Best practice is to prescribe an individualised treatment plan based on the underlying source of symptomology2, but this process is currently highly reliant on the availability of an adequately trained clinician. Laboratory studies using electrocardiography (ECG) and force plates have identified differences between mTBI patients and healthy controls3,4, but no research has investigated whether physiological and biomechanical data can differentiate between subgroups of mTBI patients requiring different treatment modalities. The purpose of this exploratory study was to evaluate whether a convolutional neural network could accurately classify mTBI patients who present with different symptom profiles using features within physiological and biomechanical data collected during an exertional protocol. Methods: A cross-sectional design enabled evaluation of a deep learning model trained with ECG and accelerometry (ACC) data to classify patients identified to have predominantly physiological (n = 12) or vestibular symptomology (n = 5) by an expert clinician. Time-series ECG and ACC data were acquired using a Zephyr BioHarness (Medtronic, Minneapolis, USA) during a validated exertional protocol on a treadmill performed in a real-world clinical environment. Three temporal slices of the ECG and ACC signals were defined at different stages of the protocol to standardise data inputs into the model. To be representative of clinical decision making, a k-folds approach was adopted to implement leave-one-out-cross-validation to evaluate the performance of a CNN trained on data from n-1 participants to correctly classify the left-out case. Results: Using data acquired during the third minute of exertional testing the CNN correctly classified 82% of the cases (3 of 5 vestibular cases and 11 of 12 physiological cases) using ECG and ACC data. The same classification accuracy was observed using only ACC data from the third minute. Discussion/conclusions: Our results provide proof of concept that the application of machine learning techniques to wearable sensor data collected during exertional testing may improve recovery outcomes post-mTBI through more accurate prescription of individualised treatment plans, particularly in the absence of an expert clinician. References 1Theadom A, Parag V, Dowell T, et al. Persistent problems 1 year after mild traumatic brain injury: a longitudinal population study in New Zealand. Br J Gen Pract 2016; 66(642):e16-e23. https://doi.org/10.3399/bjgp16X683161 2Ellis MJ, Leddy JJ, Willer B. Physiological, vestibulo-ocular and cervicogenic post-concussion disorders: an evidence-based classification system with directions for treatment. Brain Inj 2015; 29(2):238-248. https://doi.org/10.3109/02699052.2014.965207 3Gall B, Parkhouse W, Goodman D. Heart rate variability of recently concussed athletes at rest and exercise. Med Sci Sports Exerc 2004; 36(8):1269-1274. https://doi.org/10.1249/01.MSS.0000135787.73757.4D 4Doherty C, Zhao L, Ryan J, et al. Quantification of postural control deficits in patients with recent concussion: an inertial-sensor based approach. Clin Biomech 2017; 42:79-84. https://doi.org/10.1016/j.clinbiomech.2017.01.007