Use of Smart Technology Tools for Supporting Public Health Surveillance: From Development of a Mobile Health Platform to Application in Stress Prediction
MetadataShow full item record
BACKGROUND Traditional public health data collection methods are typically based on self-reported data and may be subject to limitations such as biases, delays between collection and reporting, costs, and logistics. These may affect the quality of collected information and the ability of public health agencies to monitor and improve the health of populations. An alternative may be the use of personal, off-the-shelf smart devices (e.g., smartphones and smartwatches) as additional data collection tools. These devices can collect passive, continuous, real-time and objective health-related data, mitigating some of the limitations of self-reported information. The novel data types can then be used to further study and predict a condition in a population through advanced analytics. In this context, this thesis’ goal is to investigate new ways to support public health through the use of consumer-level smart technologies as complementary survey, monitoring and analyses tools, with a focus on perceived stress. To this end, a mobile health platform (MHP) that collects data from devices connected to Apple Health was developed and tested in a pilot study collecting self-reported and objective stress-related information, and a number of Machine Learning (ML) models were developed based on these data to monitor and predict the stress levels of participants. METHODS The mobile platform was created for iOS using the XCode software, allowing users to self-report their stress levels based on the stress subscale of the Depression, Anxiety and Stress Scale (DASS-21) as well as a single-item LIKERT-based scale. The platform also collects objective data from sensors that integrate with Apple Health, one of the most popular mobile health data repositories. A pilot study with 45 participants was conducted that uses the platform to collects stress self-reports and variables associated with stress from Apple Health, including heart rate, heart rate variability, ECG, sleep, blood pressure, weight, temperature, and steps. To this end, participants were given an iPhone with the platform installed as well as an Apple Watch, Withings Sleep, Withings Thermos, Withings BPM Connect, Withings Body+, and an Empatica E4 (the only device that does not connect to Apple Health but included due to its wide use in research). Participants were instructed to take device measurements and self-report stress levels 6 times per day for 14 days. Several experiments were conducted involving the development of ML models to predict stress based on the data, using Random Forests and Support Vector Machines. In each experiment, different subsets of the data from the full sample of 45 participants were used. 3 approaches to model development were followed: a) creating generalized models with all data; b) a hybrid approach using 80% of participants to train and 20% to test the model c) creating individualized user-specific models for each participant. In addition, statistical analyses of the data – specifically Spearman correlation and repeated measures ANOVA – were conducted. RESULTS Statistical analyses did not find significant differences between groups and only weak significant correlations. Among the Machine Learning models, the approach of using generalized models performed well, with f1-macro scores above 60% for several of the samples and features investigated. User-specific models also showed promise, with 82% achieving accuracies higher than 60% (the bottom limit of the state-of-the-art). While the hybrid approach had lower f1-macro scores, suggesting the models could not predict the two classes well, the accuracy of several of these models was in line with the state-of-the-art. Apple Watch sleep features, as well as weight, blood pressure, and temperature, were shown to be important in building the models. DISCUSSION AND CONCLUSION ML-based models built with data collected from the MHP in real-life conditions were able to predict stress with results often in line with state-of-the- art, showing that smart technology data can be a promising tool to support public health surveillance. In particular, the approaches of creating models for each participant or one generalized model were successful, although more validation is needed in future studies (e.g., with more purposeful sampling) for increased generalizability and validity on the use of these technologies in the real-world. The hybrid approach had good accuracy but lower f1-scores, indicating results could potentially be improved (e.g., possibly with less missing or noisy data, collected in more controlled conditions). For feature selection, important features included sleep data as well as weight, blood pressure and temperature from mobile and wearable devices. In summary, this study indicates that a platform such as the MHP, collecting data from smart technologies, could potentially be a novel tool to complement population-level public health data collection. The predictive stress modelling might be used to monitor stress levels in a population and provide personalized interventions. Although more validation may be needed, this work represents a step in this direction.
Cite this version of the work
Pedro Elkind Velmovitsky (2023). Use of Smart Technology Tools for Supporting Public Health Surveillance: From Development of a Mobile Health Platform to Application in Stress Prediction. UWSpace. http://hdl.handle.net/10012/19669