Differentiation between atypical anorexia nervosa and anorexia nervosa using machine learning.

Area of Science:

• Eating disorder diagnosis
• Machine learning in healthcare
• Clinical psychology

Background:

• Anorexia nervosa (AN) and atypical anorexia nervosa (AAN) are often differentiated by body mass index (BMI), despite exhibiting similar clinical features.
• Prior research suggests minimal differences between AN and AAN beyond BMI criteria.

Purpose of the Study:

• To evaluate the efficacy of machine learning (ML) algorithms in distinguishing between AN and AAN using a comprehensive feature set.
• To determine if BMI is essential for accurate ML-based classification of AN versus AAN, or if other features suffice.

Main Methods:

• Utilized logistic regression, decision tree, and random forest ML models on an aggregated dataset (N=448) of individuals with AN and AAN.
• Trained models on two distinct datasets: one excluding BMI and another including all demographic, eating disorder, and comorbid features along with BMI.

Main Results:

• ML models achieved acceptable performance (mean accuracy 74.98%, mean AUC 74.75%) when BMI was included as a feature.
• Classification accuracy significantly diminished (mean accuracy 59.37%, mean AUC 59.98%) when BMI was excluded from the models.

Conclusions:

• Body mass index (BMI) is a critical feature for machine learning algorithms to accurately differentiate between anorexia nervosa and atypical anorexia nervosa.
• The inclusion of other demographic and clinical characteristics did not substantially improve classification accuracy when BMI was absent.
• These findings suggest a need to reconsider the diagnostic differentiation between AN and AAN, as BMI appears to be the primary distinguishing factor.