Kenan Li

Advances in measurement technology are producing increasingly time-resolved environmental exposure data. We aim to gain new insights into exposures and their potential health impacts by moving beyond simple summary statistics (e.g., means, maxima) to characterize more detailed features of high-frequency time series data. This study proposes a novel variant of the Self-Organizing Map (SOM) algorithm called Dynamic Time Warping Self-Organizing Map (DTW-SOM) for unsupervised pattern discovery in time series. This algorithm uses DTW, a similarity measure that optimally aligns interior patterns of sequential data, both as the similarity measure and training guide of the neural network. We applied DTW-SOM to a panel study monitoring indoor and outdoor residential temperature and particulate matter air pollution (PM2.5) for 10 patients with asthma from 7 households near Salt Lake City, UT; the patients were followed for up to 373 days each. Compared to previous SOM algorithms using timestamp alignment on time series data, the DTW-SOM algorithm produced fewer quantization errors and more detailed diurnal patterns. DTW-SOM identified the expected typical diurnal patterns in outdoor temperature which varied by season, as well diurnal patterns in PM2.5 which may be related to daily asthma outcomes. In summary, DTW-SOM is an innovative feature engineering method that can be applied to highly time-resolved environmental exposures assessed by sensors to identify typical diurnal (or hourly or monthly) patterns and provide new insights into the health effects of environmental exposures.

Background: Time-resolved quantification of physical activity can contribute to both personalized medicine and epidemiological research studies, for example, managing and identifying triggers of asthma exacerbations. A growing number of reportedly accurate machine learning algorithms for human activity recognition (HAR) have been developed using data from wearable devices (eg, smartwatch and smartphone). However, many HAR algorithms depend on fixed-size sampling windows that may poorly adapt to real-world conditions in which activity bouts are of unequal duration. A small sliding window can produce noisy predictions under stable conditions, whereas a large sliding window may miss brief bursts of intense activity. Objective: We aimed to create an HAR framework adapted to variable duration activity bouts by (1) detecting the change points of activity bouts in a multivariate time series and (2) predicting activity for each homogeneous window defined by these change points. Methods: We applied standard fixed-width sliding windows (4-6 different sizes) or greedy Gaussian segmentation (GGS) to identify break points in filtered triaxial accelerometer and gyroscope data. After standard feature engineering, we applied an Xgboost model to predict physical activity within each window and then converted windowed predictions to instantaneous predictions to facilitate comparison across segmentation methods. We applied these methods in 2 datasets: the human activity recognition using smartphones (HARuS) dataset where a total of 30 adults performed activities of approximately equal duration (approximately 20 seconds each) while wearing a waist-worn smartphone, and the Biomedical REAl-Time Health Evaluation for Pediatric Asthma (BREATHE) dataset where a total of 14 children performed 6 activities for approximately 10 min each while wearing a smartwatch. To mimic a real-world scenario, we generated artificial unequal activity bout durations in the BREATHE data by randomly subdividing each activity bout into 10 segments and randomly concatenating the 60 activity bouts. Each dataset was divided into similar to 90% training and similar to 10% holdout testing. Results: In the HARuS data, GGS produced the least noisy predictions of 6 physical activities and had the second highest accuracy rate of 91.06% (the highest accuracy rate was 91.79% for the sliding window of size 0.8 second). In the BREATHE data, GGS again produced the least noisy predictions and had the highest accuracy rate of 79.4% of predictions for 6 physical activities. Conclusions: In a scenario with variable duration activity bouts, GGS multivariate segmentation produced smart-sized windows with more stable predictions and a higher accuracy rate than traditional fixed-size sliding window approaches. Overall, accuracy was good in both datasets but, as expected, it was slightly lower in the more real-world study using wrist-worn smartwatches in children (BREATHE) than in the more tightly controlled study using waist-worn smartphones in adults (HARuS). We implemented GGS in an offline setting, but it could be adapted for real-time prediction with streaming data.

This study developed a spatial dynamic model to examine the coupled natural-human responses in the form of changes in population and associated developed land area in the Lower Mississippi River Basin region. The goal was to identify key socioeconomic factors (utility) and environmental factors (hazard damages, elevation, and subsidence rate) that affected population changes, as well as to examine how population changes affected the local utility and the local environment reciprocally. We first applied areal interpolation techniques with the volume-preserving property to transform all the data at Year 2000 into a unified 3km by 3km cellular space. We then built an Elastic Net model to extract 12 variables from a set of 33 for the spatial dynamic model. Afterward, we calibrated the neighborhood effects with a genetic algorithm and use the spatial dynamic model to simulate population and developed land area in 2010. Furthermore, we took a Monte Carlo approach for analyzing the uncertainty of the model outcome. Our accuracy assessment shows that the model on average slightly overpredicts the number of population and the developed land percentage at 2010, as indicated by the low values of mean absolute deviation (MAD) due to quantity. On the other hand, the MADs due to allocation are larger than the MADs due to quantity, with most outliers found in the New Orleans region where population and urban development declined significantly during 2000-2010 after Hurricane Katrina. The proposed model sheds light on the complex relationships between coastal hazards and human responses and provides useful insights to strategic development for coastal sustainability.

This study develops a linear regression model to select local, low-collinear explanatory variables. This model combines two well-known models: geographically weighted regression (GWR) and elastic net (EN). The GWR model posits that the regression coefficients vary as a function of location and focuses on solving the problem of explaining the relationships under the spatially nonstationary condition, which a global model cannot solve. GWR cannot fulfill the task of variable selection, however, which is problematic when there are many explanatory variables with nonnegligible multicollinearity. On the other hand, the EN model is a member of the regulated regression family. EN can trim the number of explanatory variables and select the most important ones by adding penalty terms in its cost function, and it has been proven to be robust under the high-multicollinearity condition. The EN model is a global model, however, and does not consider the spatial nonstationarity. To overcome these deficiencies, we proposed the geographically weighted elastic net (GWEN) model. GWEN uses the kernel weights derived from GWR and applies EN locally to select variables for each geographical location. The result is a set of locally selected, low-collinear explanatory variables with spatially varying coefficients. We demonstrated the GWEN method on a data set relating population changes to a set of social, economic, and environmental variables in the Lower Mississippi River Basin. The results show that GWEN has the advantages of both the high prediction accuracy of GWR and the low multicollinearity among explanatory variables of EN.