AI SenDOC Project

Advancements in healthcare have resulted in longer life expectancy. This ageing population, accompanied by an increasing number of the elderly becoming physically inactive, is placing an additional burden upon healthcare systems and directing research towards early detection or prevention of future medical issues. Remote digital health technology presents a potential solution to relieve strain on healthcare systems, reducing long-term care costs whilst improving patient outcomes and quality-of-life for older adults.

A critical factor that sets the SENDoc project apart from previous remote monitoring research is that it prioritized feasibility, usability, and acceptability of remote monitoring technology deployed in real-world conditions. The project developed several AI-based technologies to provide older adults and clinicians with novel insights/predictions about health conditions. These predictions can enable more timely and personalised interventions and ultimately prevent health conditions deteriorating. Three types of health predictions were developed: 1) Health Status Prediction, 2) Fall Risk Prediction, and 3) Mortality Prediction.

To increase the likelihood that any developed technology could be deployed in real-world conditions, solutions needed to be usable and feasible from a cost and technology deployment point-of-view as well as being acceptable by older adults. All solutions were therefore based on a single wearable sensor. This reduces the cost while also limiting invasiveness. As a consequence, the quantity and quality of data recorded about the user was also reduced. With the sensor being deployed in uncontrolled real-world conditions, sensor data were also likely to include significant noise levels. Thus, a critical technical challenge was developing AI techniques that could accurately extract insights from these limited and noisy sensor signals. Descriptions of the three individual elements of the project are described below:

Health Status Prediction:                                                                                                                                                    This work aims to predict the subjects’ health status as defined by the standardized SF-36 questionnaire. SF -36 is a commonly used tool in healthcare to measure health status. Self-reported questionnaires have a number of disadvantages due to their subjective nature and due to their inability to continuously measure health status over time. The developed technology continuously measures health status allowing healthcare providers to evaluate potential changes in health status over time and enabling more personalised and timely interventions.The SENDoc project harnessed the power of smartphones and AI to act as a remote health-status sensor. The developed technology uses GPS, accelerometer and microphone sensors embedded in a smartphone to measure human behaviour and generate behaviour profiles which can be used to make objective predictions related to health status. Motion sensors are utilised to measure physical activity, location sensors are utilised to measure travel behaviour, and sound sensors are used to measure voice activity related behaviour. Sensor fusion, through a genetic algorithm, was performed to find complementary and co-operative features. Using a behaviour measurement composed of motion, sound, and location data, results showed that a Support Vector Machine (SVM) can predict ten different health metrics with an error that did not exceed a clinical error benchmark. Evaluations of the technology, using 186 participants, showed that utilising all three sensors produced the overall best prediction performance with an average mean absolute error and correlation of 10.9% and 0.771, respectively. Taking the minimum clinically important difference into account, error rates were below the suggested benchmark of half a standard deviation and therefore deemed accurate enough to detect clinically significant changes in health status.

Fall Risk Prediction:                                                                                                                                                         Globally, falls are one of the costliest population health issues in terms of reduced quality-of-life and economic burden on healthcare service. Screening of older adults for fall risks can allow for earlier interventions and ultimately lead to better outcomes and reduced public health spending. The SENDoc project developed a novel solution to limitations in existing fall screening techniques by utilizing a hip-based accelerometer worn at-home. The technology used AI techniques to extract novel fall risk features from periods of free-living ambulatory activity. Analysis of the techniques were conducted and compared with existing screening methods using functional tests and lab- based gait analysis. The technology was evaluated using data obtained from 1705 older adults. Data consisted of 1 week of hip-worn accelerometer data, gait measurements, and performance metrics for three functional tests. Retrospective and prospective fall data were also recorded based on the incidence of falls occurring 12 months before and after the study commencing respectively. ML techniques, using novel accelerometer-based measures, performed best when predicting falls compared to more traditional lab-based measures. Prospective falls had a sensitivity and specificity of 0.61 and 0.66, respectively, while retrospective falls had 0.61 and 0.68, respectively.

Mortality Prediction:                                                                                                                                                          The application of proper prognostic indices in clinical decisions regarding mortality prediction has assumed a significant importance for personalized risk management. This requires identifying patients who are at high or low risk of death to help ensure effective healthcare services are given to at risk patients. The project developed ML models for all-cause mortality prediction in a cohort of healthy older adults. The models are based on features based on signals from a single wearable sensor device in free-living conditions as well as anthropometric variables, physical and lab examinations, questionnaires, and lifestyles collected from 2291 recruited participants.

A novel ensemble model for was developed. Results (AUC-ROC: 0.88) were in line with previous investigations. Accelerometer-related metrics, with demographics, may predict all-cause mortality in older adults with sufficient performance (AUC-ROC: 0.763) and cancer-related mortality with even higher performance (AUC-ROC: 0.857). Results showed that ML models could provide comparable results to standard epidemiological models while being completely data-driven and disease-agnostic, and prove that wearables may have a role in mortality prediction modelling in older adults.

Overall Impact:                                                                                                                                                                     After testing the developed technologies in real-world uncontrolled conditions, the main conclusion was that wearable sensor systems and the developed AI, were able to make predictions about health risk with accuracy levels that either matched or exceeded traditional infrequent and costly measures. The implications of this could have major impacts in the way health is managed. Rather that fitness trackers just providing feedback on activity metrics, inexpensive and readily available wearable sensors could be deployed to give feedback on health risk.