Redefining severity in anorexia nervosa: a multivariable prognostic model with incremental predictive validity beyond DSM-5 BMI-based classification

The DSM-5 severity specifier for anorexia nervosa (AN), which relies primarily on body mass index (BMI) thresholds, has shown limited prognostic utility for predicting treatment outcomes. The present study aimed to develop a multivariable prognostic severity model for AN based on the risk of achieving full remission following treatment. A clinical sample of N = 398 patients with AN treated in a day patient program was used to develop a prognostic severity algorithm (366 women and 32 men, with age between 16 and 68 years). Candidate predictors included BMI-related indicators, illness duration, eating disorder (ED) symptom severity, previous treatment history, non-suicidal self-injury (NSSI) and lifetime substance use. An elastic-net logistic regression (ENLR) model was fitted and internally validated using nested cross-validation. Its performance was subsequently compared with that of a multilayer perceptron (MLP) neural network. The ENLR model identified current BMI (at pre-treatment) as the strongest predictor, followed by treatment history and NSSI. Compared with the DSM-5 BMI-based severity classification (AUC = 0.56), the ENLR model showed substantially superior prognostic discriminative performance (AUC = 0.68), and adequate calibration (Brier score = 0.21), with no evidence of systematic over- or underestimation of risk. Results were largely replicated by the MLP model, which yielded comparable discrimination and only marginally improved probabilistic accuracy. The ENLR-derived severity groups exhibited a markedly steeper gradient in remission rates (43.4% to 12.0%) than the DSM-5 severity categories (34.2% to 17.4%), resulting in significantly better prognostic stratification. The ENLR showed moderate predictive performance across AN subtypes and age groups, with slightly higher discrimination among patients aged 20–24 years and those with binge–purge symptomatology. A multivariable prognostic severity model integrating anthropometric, clinical, and treatment-history indicators substantially outperformed the DSM-5 BMI-based severity criterion in predicting treatment outcomes in AN. These findings support the adoption of multidimensional severity frameworks and provide a clinically interpretable algorithm for prognostic stratification and personalized treatment planning in AN. Anorexia nervosa is a serious eating disorder, and predicting which patients are most likely to recover following treatment remains a major clinical challenge. Current severity ratings rely primarily on body weight-related measures, but these may provide limited information about an individual’s likelihood of recovery. In this study, we developed a machine learning-based prognostic tool to empirically derive a predictive model of full remission using a small set of variables routinely collected in clinical practice. Data were obtained from 398 patients with anorexia nervosa (men and women, aged 16 to 68 years). The model integrated several clinical factors, including body mass index, treatment history, non-suicidal self-injury, eating disorder severity, and substance use. The new tool predicted treatment outcomes substantially more accurately than the current weight-based severity classification (measured at pre-treatment). Current body mass index was the most important predictor, followed by previous treatment history and non-suicidal self-injury. The tool showed slightly better accuracy in young adults aged 20–24 years and in patients with binge-purge symptoms. These findings suggest that considering multiple aspects of a patient’s condition may provide a more useful and personalized assessment of illness severity and recovery potential than relying on body weight alone.