Abstract：
OBJECTIVE: Binge eating and self-induced vomiting are common, transdiagnostic eating disorder (ED) symptoms. Efforts to understand these behaviors in research and clinical settings have historically relied on self-report measures, which may be biased and have limited ecological validity. It may be possible to passively detect binge eating and vomiting using data collected by continuous glucose monitors (CGMs; minimally invasive sensors that measure blood glucose levels), as these behaviors yield characteristic glucose responses.
METHODS: This study developed machine learning classification algorithms to classify binge eating and vomiting among 22 adults with binge-spectrum EDs using CGM data. Participants wore Dexcom G6 CGMs and reported eating episodes and disordered eating symptoms using ecological momentary assessment for 2 weeks. Group-level random forest models were generated to distinguish binge eating from typical eating episodes and to classify instances of vomiting.
RESULTS: The binge eating model had accuracy of 0.88 (95% CI: 0.83, 0.92), sensitivity of 0.56, and specificity of 0.90. The vomiting model demonstrated accuracy of 0.79 (95% CI: 0.62, 0.91), sensitivity of 0.88, and specificity of 0.71.
CONCLUSIONS: Results suggest that CGM may be a promising avenue for passively classifying binge eating and vomiting, with implications for innovative research and clinical applications.
METHODS: This study developed machine learning classification algorithms to classify binge eating and vomiting among 22 adults with binge-spectrum EDs using CGM data. Participants wore Dexcom G6 CGMs and reported eating episodes and disordered eating symptoms using ecological momentary assessment for 2 weeks. Group-level random forest models were generated to distinguish binge eating from typical eating episodes and to classify instances of vomiting.
RESULTS: The binge eating model had accuracy of 0.88 (95% CI: 0.83, 0.92), sensitivity of 0.56, and specificity of 0.90. The vomiting model demonstrated accuracy of 0.79 (95% CI: 0.62, 0.91), sensitivity of 0.88, and specificity of 0.71.
CONCLUSIONS: Results suggest that CGM may be a promising avenue for passively classifying binge eating and vomiting, with implications for innovative research and clinical applications.
摘要：
目的：暴饮暴食和自我诱发的呕吐是常见的，经诊断的进食障碍（ED）症状。在研究和临床环境中理解这些行为的努力历来依赖于自我报告措施，这可能是有偏见的，生态有效性有限。使用连续葡萄糖监测仪（CGM；测量血糖水平的微创传感器）收集的数据，可以被动地检测暴饮暴食和呕吐，因为这些行为产生了特征性的葡萄糖反应。
方法：这项研究开发了机器学习分类算法，使用CGM数据对22名暴饮暴食和呕吐进行分类。参与者穿着DexcomG6CGMs，并使用2周的生态瞬时评估报告了饮食发作和饮食紊乱症状。生成组级别的随机森林模型以区分暴饮暴食与典型的进食发作，并对呕吐的情况进行分类。
结果：暴食模型的准确性为0.88（95％CI：0.83，0.92），敏感性为0.56，特异性为0.90。呕吐模型的准确性为0.79（95％CI：0.62，0.91），敏感性为0.88，特异性为0.71。
结论：结果表明，CGM可能是被动分类暴饮暴食和呕吐的有希望的途径，对创新研究和临床应用具有重要意义。
方法：这项研究开发了机器学习分类算法，使用CGM数据对22名暴饮暴食和呕吐进行分类。参与者穿着DexcomG6CGMs，并使用2周的生态瞬时评估报告了饮食发作和饮食紊乱症状。生成组级别的随机森林模型以区分暴饮暴食与典型的进食发作，并对呕吐的情况进行分类。
结果：暴食模型的准确性为0.88（95％CI：0.83，0.92），敏感性为0.56，特异性为0.90。呕吐模型的准确性为0.79（95％CI：0.62，0.91），敏感性为0.88，特异性为0.71。
结论：结果表明，CGM可能是被动分类暴饮暴食和呕吐的有希望的途径，对创新研究和临床应用具有重要意义。
