Predicting Depression From Smartphone Behavioral Markers Using Machine Learning Methods, Hyperparameter Optimization, and Feature Importance Analysis: Exploratory Study.

BACKGROUND: Depression is a prevalent mental health challenge. Current depression assessment methods using self-reported and clinician-administered questionnaires have limitations. Instrumenting smartphones to passively and continuously collect moment-by-moment data sets to quantify human behaviors has the potential to augment current depression assessment methods for early diagnosis, scalable, and longitudinal monitoring of depression. OBJECTIVE: The objective of this study was to investigate the feasibility of predicting depression with human behaviors quantified from smartphone data sets, and to identify behaviors that can influence depression. METHODS: Smartphone data sets and self-reported 8-item Patient Health Questionnaire (PHQ-8) depression assessments were collected from 629 participants in an exploratory longitudinal study over an average of 22.1 days (SD 17.90; range 8-86). We quantified 22 regularity, entropy, and SD behavioral markers from the smartphone data. We explored the relationship between the behavioral features and depression using correlation and bivariate linear mixed models (LMMs). We leveraged 5 supervised machine learning (ML) algorithms with hyperparameter optimization, nested cross-validation, and imbalanced data handling to predict depression. Finally, with the permutation importance method, we identified influential behavioral markers in predicting depression. RESULTS: Of the 629 participants from at least 56 countries, 69 (10.97%) were females, 546 (86.8%) were males, and 14 (2.2%) were nonbinary. Participants' age distribution is as follows: 73/629 (11.6%) were aged between 18 and 24, 204/629 (32.4%) were aged between 25 and 34, 156/629 (24.8%) were aged between 35 and 44, 166/629 (26.4%) were aged between 45 and 64, and 30/629 (4.8%) were aged 65 years and over. Of the 1374 PHQ-8 assessments, 1143 (83.19%) responses were nondepressed scores (PHQ-8 score <10), while 231 (16.81%) were depressed scores (PHQ-8 score ≥10), as identified based on PHQ-8 cut-off. A significant positive Pearson correlation was found between screen status-normalized entropy and depression (r=0.14, P<.001). LMM demonstrates an intraclass correlation of 0.7584 and a significant positive association between screen status-normalized entropy and depression (ß=.48, P=.03). The best ML algorithms achieved the following metrics: precision, 85.55%-92.51%; recall, 92.19%-95.56%; F1, 88.73%-94.00%; area under the curve receiver operating characteristic, 94.69%-99.06%; Cohen κ, 86.61%-92.90%; and accuracy, 96.44%-98.14%. Including age group and gender as predictors improved the ML performances. Screen and internet connectivity features were the most influential in predicting depression. CONCLUSIONS: Our findings demonstrate that behavioral markers indicative of depression can be unobtrusively identified from smartphone sensors' data. Traditional assessment of depression can be augmented with behavioral markers from smartphones for depression diagnosis and monitoring.

EDISON: An Edge-Native Method and Architecture for Distributed Interpolation.

Spatio-temporal interpolation provides estimates of observations in unobserved locations and time slots. In smart cities, interpolation helps to provide a fine-grained contextual and situational understanding of the urban environment, in terms of both short-term (e.g., weather, air quality, traffic) or long term (e.g., crime, demographics) spatio-temporal phenomena. Various initiatives improve spatio-temporal interpolation results by including additional data sources such as vehicle-fitted sensors, mobile phones, or micro weather stations of, for example, smart homes. However, the underlying computing paradigm in such initiatives is predominantly centralized, with all data collected and analyzed in the cloud. This solution is not scalable, as when the spatial and temporal density of sensor data grows, the required transmission bandwidth and computational capacity become unfeasible. To address the scaling problem, we propose EDISON: algorithms for distributed learning and inference, and an edge-native architecture for distributing spatio-temporal interpolation models, their computations, and the observed data vertically and horizontally between device, edge and cloud layers. We demonstrate EDISON functionality in a controlled, simulated spatio-temporal setup with 1 M artificial data points. While the main motivation of EDISON is the distribution of the heavy computations, the results show that EDISON also provides an improvement over alternative approaches, reaching at best a 10% smaller RMSE than a global interpolation and 6% smaller RMSE than a baseline distributed approach.

Internet of Things for Smart Spaces: A University Campus Case Study.

Advances in technology and data analysis provide rich opportunities for developing intelligent environments assisting their inhabitants, so-called smart environments or smart spaces. Enhanced with technology, sensors, user interfaces, and various applications, such smart spaces are capable of recognizing users and situations they are in, react accordingly, e.g., by providing certain services or changes to the environment itself. Therefore, smart space solutions are gradually coming to different application domains, each with corresponding specific characteristics. In this article, we discuss our experiences and explore the challenges of a long-term real-world Internet of Things (IoT) deployment at a University campus. We demonstrate the technical implementation and data quality issues. We conduct several studies, from data analysis to interaction with space, utilizing the developed infrastructure, and we also share our actions to open the data for education purposes and discuss their outcomes. With this article, we aim to share our experience and provide real-world lessons learned when building an open, multipurpose, publicly used smart space at a University campus.

Understanding Smartwatch Battery Utilization in the Wild.

Smartwatch battery limitations are one of the biggest hurdles to their acceptability in the consumer market. To our knowledge, despite promising studies analyzing smartwatch battery data, there has been little research that has analyzed the battery usage of a diverse set of smartwatches in a real-world setting. To address this challenge, this paper utilizes a smartwatch dataset collected from 832 real-world users, including different smartwatch brands and geographic locations. First, we employ clustering to identify common patterns of smartwatch battery utilization; second, we introduce a transparent low-parameter convolutional neural network model, which allows us to identify the latent patterns of smartwatch battery utilization. Our model converts the battery consumption rate into a binary classification problem; i.e., low and high consumption. Our model has 85.3% accuracy in predicting high battery discharge events, outperforming other machine learning algorithms that have been used in state-of-the-art research. Besides this, it can be used to extract information from filters of our deep learning model, based on learned filters of the feature extractor, which is impossible for other models. Third, we introduce an indexing method that includes a longitudinal study to quantify smartwatch battery quality changes over time. Our novel findings can assist device manufacturers, vendors and application developers, as well as end-users, to improve smartwatch battery utilization.

Mobile road weather sensor calibration by sensor fusion and linear mixed models.

Mobile, vehicle-installed road weather sensors are becoming ubiquitous. While mobile sensors are often capable of making observations on a high frequency, their reliability and accuracy may vary. Large-scale road weather observation and forecasting are still mostly based on stationary road weather stations (RWS). Though expensive, sparsely located and making observations on a relatively low frequency, RWS' reliability and accuracy are well-known and accommodated for in the road weather forecasting models. Statistical analysis revealed that road weather conditions indeed have a great effect on how the observations of mobile and stationary road weather temperature sensors differ from each other. Consequently, we calibrated the observations of mobile sensors with a linear mixed model. The mixed model was fitted fusing ca. 20 000 pairs of mobile and RWS observations of the same location at the same time, following a rendezvous model of sensor calibration. The calibration nearly halved the MSE between the observations of the mobile and the RWS sensor types. Computationally very light, the calibration can be embedded directly in the sensors.