Wearable Sensors for Brain Injury Prediction and Frailty Assessment

Abstract

Advances in medical technology, accessibility to health care, healthier lifestyle, diet, and hygiene have allowed us to increase life expectancy from 46 years in 1950 to almost 74 in 2019. Worldwide, the population is becoming older, and new aging-related challenges have appeared, increasing the burden on a debilitated-in-workforce healthcare system. Wearable sensors and sensor-enhanced health information systems are regarded as one way to face the implications of the aforementioned demographic change. Indeed, continuously monitoring health time series would bring detailed aspects that could potentially enable early diagnosis and an informed prognosis of any disease or injury. This dissertation presents some efforts on developing methods for injury prediction and frailty assessment from real-time health signals obtained from wearable sensors. This work was divided into three consecutive studies.In the first study, we studied an existing data set of 537 head impact kinematics in sports settings, consisting of 6 degrees of freedom measurements. We presented a principal component analysis-based method for building an accurate low-rank approximation of head impact kinematics. We found that only a few modes were sufficient for the accurate reconstruction of the entire data set. Then, we studied these representative modes in the frequency domain, which were primarily low frequency under 40Hz. We compared our approximation against existing impact parametrization methods and showed significantly better performance in injury prediction using a range of kinematic and brain tissue deformation-based metrics. In all cases, our representation reproduced similar results in injury prediction to the ground truth. Finally, we developed a data-driven emulator capable of generating new kinematic data sets of any size, being particularly useful for training machine learning algorithms that have been investigated for approximating brain deformation in real-time for early brain injury diagnosis. In the second study, we focused on frailty syndrome, which is associated with a lack of physiological reserve and consequent adverse outcomes (treatment complications and death). Previous research has shown associations between heart rate (HR) dynamics (HR changes during physical activity) and frailty. We analyzed the effect of frailty on the interconnection between the motor and cardiac systems during a localized upper-extremity function (UEF) test. Fifty-six older adults aged 65 or older were recruited to perform a rapid elbow flexion-extension test for 20 seconds while we monitored motor performance and heart rate, utilizing wearable gyroscopes and electrocardiogram devices, respectively. Frailty was assessed using the gold-standard Fried phenotype. We quantified the interconnection between motor (angular displacement) and cardiac (HR) performance using convergent cross-mapping (CCM). A significantly weaker interconnection was observed among pre-frail and frail participants compared to non-frail individuals. Findings suggested a strong association between cardiac-motor interconnection and frailty. In the final study, we continued the development of an integral frailty assessment tool, now specifically for heart diseases. Aortic stenosis (AS) is the most commonly acquired valvar disease and is associated with an increased risk of frailty. The current work aimed to assess differences in motor and ANS performance as symptoms of frailty between community-dwelling older adults with and without AS. Older adults, 55 years and older, with and without AS were recruited. Frailty was assessed using the Fried phenotype. Participants performed an upper-extremity function (UEF) physical task – 20 seconds of rapid elbow flexion of the right arm. Arm motion was measured by gyroscopes and heart rate (HR) was measured using an electrocardiogram (ECG) sensor attached to the left side of the chest. Outcomes included UEF motor score (a validated score from 0: “not frail” to 1: “extremely frail” based on slowness, weakness, exhaustion, and flexibility), HR percentage increase due to physical activity and decrease due to rest, and ANS performance (scored from 0: “poor” to 1: “excellent”). ANS performance was measured using convergent cross mapping (CCM) representing the interconnection between HR and motor data. ANOVA models were used with the Fried frailty, AS condition, age, BMI, and sex as independent variables and UEF outcomes as dependent variables. There was a significant difference in UEF motor scores between older adults with and without AS and between the frailty groups. CCM parameters showed significant differences between the frailty groups; however, they were not found to be significantly different between the AS groups. No significant interaction was observed between frailty and AS condition. Our findings suggested that ANS measures may be highly associated with frailty regardless of AS condition. Combining motor and HR dynamics parameters in a multimodal model may provide a promising tool for frailty assessment. The results presented in this work show ways in which data acquired from wearable sensors could be translated into injury prediction and frailty assessment. These are topics that seem to be deeply important and responsive to future demographic changes in healthcare systems, regarding frailty and aging.