(C) PLOS One
This story was originally published by PLOS One and is unaltered.
 . . . . . . . . . .



Population analysis of mortality risk: Predictive models from passive monitors using motion sensors for 100,000 UK Biobank participants [1]

['Haowen Zhou', 'Department Of Statistics', 'University Of Illinois At Urbana-Champaign', 'Champaign', 'Illinois', 'United States Of America', 'Carl R. Woese Institute For Genomic Biology', 'Urbana', 'Ruoqing Zhu', 'Anita Ung']

Date: 2022-11 

Abstract Many studies have utilized physical activity for predicting mortality risk, using measures such as participant walk tests and self-reported walking pace. The rise of passive monitors to measure participant activity without requiring specific actions opens the possibility for population level analysis. We have developed novel technology for this predictive health monitoring, using limited sensor inputs. In previous studies, we validated these models in clinical experiments with carried smartphones, using only their embedded accelerometers as motion sensors. Using smartphones as passive monitors for population measurement is critically important for health equity, since they are already ubiquitous in high-income countries and increasingly common in low-income countries. Our current study simulates smartphone data by extracting walking window inputs from wrist worn sensors. To analyze a population at national scale, we studied 100,000 participants in the UK Biobank who wore activity monitors with motion sensors for 1 week. This national cohort is demographically representative of the UK population, and this dataset represents the largest such available sensor record. We characterized participant motion during normal activities, including daily living equivalent of timed walk tests. We then compute walking intensity from sensor data, as input to survival analysis. Simulating passive smartphone monitoring, we validated predictive models using only sensors and demographics. This resulted in C-index of 0.76 for 1-year risk decreasing to 0.73 for 5-year. A minimum set of sensor features achieves C-index of 0.72 for 5-year risk, which is similar accuracy to other studies using methods not achievable with smartphone sensors. The smallest minimum model uses average acceleration, which has predictive value independent of demographics of age and sex, similar to physical measures of gait speed. Our results show passive measures with motion sensors can achieve similar accuracy to active measures of gait speed and walk pace, which utilize physical walk tests and self-reported questionnaires.

Author summary Healthcare infrastructure implementation could benefit tremendously from national scale screening with passive monitors. Large scale population data could delineate health risks without intruding into daily living. Digital health offers potential solutions if sensor devices of adequate accuracy for predictive models could be widely deployed. The only such current devices are cheap phones, smartphone devices with embedded accelerometers. This limits measures to motion sensor data collected when the phones are carried during normal activities. So measuring walking intensity is possible, but the total activity measure that is possible with 24-hour wearable devices is not. Our study simulates the use of smartphone sensors to predict mortality risk in the largest national cohort with sensor records, the demographically representative UK Biobank. Mortality is the most definitive outcome, with accurate death records for five years available for the 100,000 participants who wore sensor devices. We analyzed this dataset to extract walking sessions during daily living, then used characteristic motions of these walking sessions to predict mortality risk. The accuracy achieved was similar to activity monitors measuring total activity and even similar to physical measures such as gait speed during observed walks. Our scalable methods offer a feasible pathway towards national screening for health risk.

Citation: Zhou H, Zhu R, Ung A, Schatz B (2022) Population analysis of mortality risk: Predictive models from passive monitors using motion sensors for 100,000 UK Biobank participants. PLOS Digit Health 1(10): e0000045. https://doi.org/10.1371/journal.pdig.0000045 Editor: Yuan Lai, Tsinghua University, CHINA Received: April 15, 2022; Accepted: September 12, 2022; Published: October 20, 2022 Copyright: © 2022 Zhou et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Data Availability: The datasets are produced by the UK Biobank, a national research resource in the United Kingdom. As a public resource, all this data is available to researchers worldwide. All such requests must be formally approved, see application process at https://www.ukbiobank.ac.uk/enable-your-research/apply-for-access. Funding: Principal Investigator Bruce Schatz. [Beckman Award 2019] RB19125, Predicting Mortality from Wearable Devices. University of Illinois at Urbana-Champaign, Campus Research Board, https://crb.research.illinois.edu/past-awards The funder had no role in the study design, data collection and analysis, decision to publish, or in the preparation of the manuscript. Competing interests: The authors have declared that no competing interests exist.

Introduction The association of physical activity with mortality risk is well established. National cohort studies based on self-report have shown intensity to be correlated with survival, as persons who engage in more moderate-to-vigorous activity and less sedentary activity have lower mortality rates [1]. These studies focus upon amount of activity at given intensity level. These findings have been confirmed in large meta-analyses using objective physical activity, in which body worn sensors record total activity and statistical models predict mortality risk using accelerometers [2]. Cohort meta-analyses also show sensor features improve model performance beyond traditional risk factors [3], e.g. smoking and alcohol, independent of demographics, e.g. age and sex. In addition to the quantity of intensity, there are also implications for the quality of intensity. Physical measurements focus upon walking as a moderate activity, intermediate between vigorous and sedentary activity. Large cohort studies show gait speed is correlated with mortality risk [4], with timed walking over short distances such as 6 seconds for 4 meters. National cohort studies based on self-report reveal walking pace as a unique characteristic beyond traditional demographic risk factors which mediate cardiovascular mortality risk [5]. The 6 minute walk test [6]—where persons walk steadily in hospital corridor, a standard evaluation for cardiopulmonary disease—has been shown in large meta-analysis studies to be a strong independent predictor of mortality from heart failure [7]. Our current study focuses on the largest available national cohort, the UK Biobank [8], where 103,683 participants wore wrist devices with accelerometer sensors for 1 week [9]. In keeping with our previous accelerometer based analysis of the physical activity level of a national cohort [10], the US Women’s Health Initiative, we use raw sensor data during labelled walking sessions to identify characteristic motions for predictive models. This is the first population analysis of walking intensity with mobile sensors, accepting only input types which can be accurately gathered via personal smartphones. There are four primary methods for measuring physical activity, which all achieve roughly the same accuracy for predictive models of mortality risk. Two methods require active individual participation, such as answering a questionnaire concerning health status (self-report) or walking a fixed distance under observation (gait-speed). These have proven feasible within limited cohort studies, but are problematic to scale for population level assessment, due to logistic difficulty of getting large numbers of people to perform the required tasks on a routine basis. Two measures are passive, collecting data through devices such as activity monitors worn on the body: total amount of physical activity performed during the day and intensity of physical activity such as walking pace over a limited period. These sensor-based methods have the major advantage that they can measure physical activity during daily living, without requiring persons to change their normal activity other than wearing the devices. However, such digital health approaches have had limited success due to health equity issues relating to access to wearable sensors. Results based on wearable activity monitors come largely from recruited cohorts, such as the nationally representative sample we analyze in this study, rather than from actual community populations. For population measurement in health systems to be routinely available, the measurement devices must be already widely deployed and familiar to the public [11], mandating the choice of mobile phones at present. In the United States for example, the Pew Research Center estimates 97% of the population own cell phones, with 83% possessing smart phones containing motion sensors [12], while only 21% of the population wear sensors such as smart watches or fitness devices [13]. Thus scalable methods for predictive models using mobile phones would have great impact if data analytic limitations can be overcome. Mobile phones are often carried while walking, so they could easily capture walking sessions. Conversely, they are rarely carried all day, so would not be effective at collecting the total amount of physical activity achieved during a day, unlike wearable devices. The smartphone penetration rate in the United Kingdom, where our dataset was gathered, has increased every year over the past decade, reaching an overall ownership of 92% in 2021. For older adults in the statistical survey [14], less than half of all respondents over the age of 55 owned such a device in 2016, but this total rose to 83% in 2021. Soon adequate devices will be everywhere, with even the cheapest flip phones incorporating motion sensors. Furthermore, inexpensive smartphones are already widespread worldwide, even in the poorest countries [15]. The global smartphone penetration rate is estimated to have reached over 78% in 2020. This is based on 6.4 billion smartphone subscriptions in a global population of 7.8 billion. The global smartphone penetration rate in the general population has great regional variation. In North America and Europe, the smartphone adoption rates are roughly 82% and 78% respectively, whereas in Sub-Saharan Africa, the same rate currently stands at 48 percent. While there is a 30% difference in adoption rates between the highest and lowest ranked regions, note that even in low-income regions half the population already has smart phones with motion sensors. Cheap phones could have major impact in addressing health equity if proper models can be developed to utilize the limitations of the data provided by their sensors when they are carried. Our study uses the sensor dataset from the largest current national cohort, the UK Biobank, the largest sensor dataset currently available. Although this data was gathered from activity monitors, our sensor models use only the inputs that would be feasible to gather using inexpensive, currently available, phones. This is possible because of our extensive clinical experiments with cheap phones, developing highly accurate predictive models for health status for cardiopulmonary patients [16]. In addition, the 100K participants included in this paper are demographically similar to the overall 500K UK Biobank participants, who match the characteristics of the national population [17], thus providing significant generality to the model results.

Discussion Measuring physical activity via walking intensity has become a standard practice for certain clinical settings, where gait speed can be quantified with a short walk. Detailed meta-analyses showed that gait speed is a predictor independent of age/sex [4], with a pooled C-index close to 0.72 model accuracy for 5-year mortality risk. Other metrics like Objective Physical Activity (OPA) look at the “quantity” of physical activity, such as total amount of moderate-to-vigorous physical activity, requiring sensor devices to be worn all day. For example, the concurrent study [26] of the same UK Biobank sensor dataset developed a model where the highest predictor of mortality was relative amplitude (RA), the ratio of the most active 10 hours of average acceleration to the least active 5 hours. The C-index studied was 0.72, with RA plus age/sex for 5-year risk, based upon 600 minutes per day of sensor records. A walk test measures “quality” (intensity) rather than “quantity” (duration). Our previous work showed accelerometer sensors in carried smartphones can digitally model physical distance [27] and oxygen saturation [28] during a Six Minute Walk Test (6MWT). We also showed that the pulmonary models similarly worked with smartphones carried during daily living [16]. The logistical advantage of using 6 minutes of walking intensity is two orders of magnitude less frequent sensor input, using ENMO for quality instead of RA for quantity. Measuring intensity/quality makes it possible to effectively utilize smart phones instead of wearable sensors for predictive models. Our Min Model with only sensor features holds at the same C-index of 0.72 for 5-year mortality risk. For continuous features only without categorical features, our Max Model with all the sensor features yields 0.73 C-index for 5-year risk. This model has greater accuracy in earlier years yielding 0.76 for 1-year risk. We note model accuracy varies by local sites, as shown in S2 Fig, with 0.77 for 5-year risk at the Scottish sites of Glasgow and Edinburgh, where the original mortality study using self reports also did best [23]. There are significant limitations to our current research. The most obvious is that the UK Biobank dataset was generated by wrist-worn motion-sensors. The sensors themselves are equivalent to those contained in smartphones, but the wearing patterns may not be, so results may differ when large datasets generated by personal smartphones become available. The walking patterns of large populations chosen for health equity may also differ, since low-income lifestyles differ from high-income lifestyles even when the demographics of age and sex are the same, as in the UK Biobank dataset. The methodology of what is considered to be walking sessions might be thus affected, since 6 minutes of steady walking was chosen to mimic walk tests for hospital patients with cardiopulmonary diseases. Our models computed all-cause mortality of patients aged 45–79 for the 5 years past when sensors were recorded. Utilizing walking intensity implies that higher predictive accuracy might be achieved for older patients only, especially those who ultimately die from cardiopulmonary diseases where characteristic motions are more discriminating than with other causes of mortality. We have planned large population trials with only cardiopulmonary patients carrying their personal smartphones, to investigate whether such walk test cohorts produce more accurate predictive models. We hope our research makes clear that large trials employing passive monitors with diverse populations using cheap phones are now technically feasible and socially desirable. In terms of future directions, we are involved in planning the physical activity study for the US Precision Medicine Initiative (All of Us Research Program), especially the use of phones for health monitoring. This historic longitudinal cohort is planned to have more than 1M participants and is already over 50% enrollment. Participants are being recruited to be representative of the US national population, which is considerably more diverse than that of the UK. For example, the ethnicity “white” covers 94% of all UK Biobank participants, so “race” is weakly correlated with mortality risk in our current analysis. Race/ethnicity is more easily stratified with the US Precision Medicine Initiative population. All consenting participants would be longitudinally measured on their personal smartphones, directly utilizing smartphone sensors, with both a larger sample size and a longer time horizon than our current mortality analysis. Our previous work showed accelerometer motion sensors in cheap smart phones can capture predictive model input for walking intensity analysis equivalent to expensive medical devices. This is particularly important for health equity purposes, given populations at highest health risk are often the least resourced—so persons most likely to have cheap phones rather than wearable devices would benefit most from easy assessment. Phone apps could record six minutes of consecutive walking during daily living, then compute predictive models for risk stratification via population analysis [11]. To test this strategy, we have planned large trials with minority populations using personal smartphones, within the US Hispanic Community Health Study. Our results from high-income countries may be directly applicable to low-income countries. Major cohort studies using self-reported status have shown cardiovascular health is strongly correlated with physical activity, largely independent of the socioeconomic level of the participants country [29]. Healthy longevity can be facilitated globally for all adults possessing cheap phones, using the minimum model to assess gait status, computed on their phones for the maximum privacy. Implementing effective healthcare infrastructure requires continued research into screening populations with ubiquitous sensors [30].

Acknowledgments We thank Qian Cheng of Salesforce AI Research for advice on predictive models of wearable sensors.

[END]
---
[1] Url: https://journals.plos.org/digitalhealth/article?id=10.1371/journal.pdig.0000045

Published and (C) by PLOS One
Content appears here under this condition or license: Creative Commons - Attribution BY 4.0.

via Magical.Fish Gopher News Feeds:
gopher://magical.fish/1/feeds/news/plosone/