Logo image
Back
Application of machine learning to predict loss of control eating and compensatory behavior episodes during cognitive behavioral therapy for bulimia nervosa
Dissertation   Open access

Application of machine learning to predict loss of control eating and compensatory behavior episodes during cognitive behavioral therapy for bulimia nervosa

Christina Robinson Felonis
Doctor of Philosophy (Ph.D.), Drexel University
Jun 2024
DOI:
https://doi.org/10.17918/00010709
pdf
Felonis_Christina_20241.09 MBDownloadView
PDF Open Access Open Access (License Unspecified)

Abstract

Cognitive therapy Compensatory behavior Loss of control eating Predict Eating Disorders Machine Learning
Just-in-time adaptative interventions (JITAIs) present a promising solution to improving outcomes from cognitive behavioral therapy (CBT) for bulimia nervosa (BN) by prompting individuals to practice skills in their exact moments of need via their smartphones. JITAIs have been utilized in the field of overweight and obesity to predict and prevent dietary lapses, however, only one study has tested a JITAI as an adjunct to CBT for BN and this utilized a theory-driven algorithm. Alternative approaches like machine learning (ML) may be more effective as they develop computer algorithms that predict behaviors in real-time with incoming group data that are continuously updated with individual data to learn different individuals' patterns over time. As part of a larger scale JITAI study, the current study used existing baseline and electronic self-monitoring data from 55 individuals with BN over 16 weeks of outpatient CBT to: (1) develop two group-level ML models for predicting loss of control eating (LOCE) and compensatory behavior (CB) episodes over the course of CBT that achieve acceptable classification accuracy, (2) compare the performance of the group-level models within-subjects, between-subjects, and when combined with individual-level data, (3) examine the classification accuracy of the group-level models in the remaining data after 4 and 8 weeks, (4) evaluate the importance of each feature in the group-level models, and (5) evaluate if certain subgroups of participants respond poorly in the group-level models. Results showed that several ML algorithms can successfully build group-level models for predicting LOCE and CB episodes, with generalized linear model performing best (LOCE 90% accuracy, 91% sensitivity, 90% specificity; CB 94% accuracy, 96% sensitivity, 94% specificity). The models generalized somewhat sufficiently within- and between-subjects (LOCE 30-50% of participants; CB 50-60% of participants), suggesting that they lack the ability to capture individual differences in predicting eating disorder (ED) behaviors. Importantly, models utilizing both group- and individual-level data performed best (LOCE and CB 75% of participants), indicating that JITAI systems incorporating both would perform best. The group-level models performed less well although still acceptably when created after 4 weeks (LOCE 92% accuracy, 86% sensitivity, 92% specificity; CB 92% accuracy, 63% sensitivity, 93% specificity) and 8 weeks (LOCE 95% accuracy, 80% sensitivity, 96% specificity; CB 98% accuracy, 72% sensitivity, 99% specificity). Prioritizing tracking of ED behaviors and urges over a 2-day window, eating episodes during certain times of day, meal types, mood, and session frequency may be most effective in predicting and preventing ED behaviors. The current study was the first in the ED field to support the development of ML models for predicting LOCE and CB episodes during CBT.

Metrics

174 File views/ downloads
40 Record Views

Details

Logo image