Logo image
Back
Integrating sensor technology and machine learning to target dietary lapses
Thesis   Open access

Integrating sensor technology and machine learning to target dietary lapses

Rebecca Jane Crochiere
Master of Science (M.S.), Drexel University
Apr 2020
DOI:
https://doi.org/10.17918/00000994
pdf
Crochiere_Rebecca_20201.88 MBDownloadView
PDF Open Access Open Access (License Unspecified)

Abstract

Reducing diets Weight loss
Over 160 million American adults have overweight/obesity-a medical issue that increases risk of serious health consequences and reduces quality of life. Yet, self-help and gold-standard behavioral weight loss treatments produce less weight loss than is recommended to lower health risks. Suboptimal weight loss outcomes are largely attributable to individuals' inability to consistently adhere to a prescribed diet, i.e., dietary lapse. Research has identified specific factors that increase risk of dietary lapse, making interventions that provide support based on the presence of these risk factors ripe for development. However, to date, these momentary risk factors have been almost exclusively measured via ecological momentary assessment (EMA), which has several shortcomings that reduce its feasibility, acceptability, and efficacy, including participant burden, self-report bias (and thus limited predictive accuracy), deterioration of compliance over time, and restricted data collection periods (i.e., only collecting data at times during which participants complete surveys). Passive sensing systems, which are capable of objective, automatic, quasi-continuous measurement of real-time factors that influence dietary lapses, are not subject to these same limitations and thus may offer discrete advantages over EMA. However, no study to date has evaluated whether passively sensed variables can predict dietary lapse. The primary aims of this study were: 1) to determine if certain well-validated sensors and sensor features contribute to the prediction of dietary lapse and 2) to use predictive sensor features to build and evaluate outcomes of (accuracy, sensitivity, and specificity) a machine learning algorithm designed to classify dietary lapse cases. Results showed 74 features from 4 sensor variables (sleep, PA, location, time) contributed to the predictive accuracy of the model and 10 were found to be stable predictors. In addition, the optimal LASSO supervised binary classification group model predicted lapse with 67% accuracy, i.e., 73% sensitivity and 66% specificity, which was statistically significantly greater than random chance predictions. Of the individual models examined, 7 out of 14 (50%) met minimum model adequacy criteria ([greater than or equal to]70% sensitivity and [greater than or equal to]50% specificity), with 10 out of 14 (71.4%) meeting less stringent criteria ([greater than or equal to]60% sensitivity and [greater than or equal to]50% specificity). This project showed that many sensor features contribute to predictive models of lapse and that sensor-measured risk factors and machine learning can be used to predict lapse with accuracy. These findings lay the groundwork for future research that may utilize sensors to predict when dietary lapses will occur to inform the delivery of a just-in-time intervention (e.g., a JITAI).

Metrics

56 File views/ downloads
52 Record Views

Details

Logo image