The study found that people with unhealthy obesity have lower blood oxytocin and higher leptin, and that lower oxytocin levels are linked to more severe disordered‑eating behaviors. Adding other health measures (like BMI, waist size, and leptin) makes it much better at spotting those at risk than looking at oxytocin alone.

Obesity is heterogeneous across metabolic and behavioral dimensions. Oxytocin, a hypothalamic neuropeptide, and leptin, an adiposity signal, have been implicated in appetite and reward, yet their relationships with disordered eating across metabolic obesity phenotypes remain unclear. We examined these associations and evaluated the predictive value of oxytocin alone versus multivariable models. In a cross-sectional cohort of 99 adults, we assessed anthropometry, biochemistry, oxytocin and leptin, and three validated questionnaires (EDE-Q, DEBQ, EBA-O). Participants were classified into four metabolic obesity phenotypes. Group differences used Kruskal-Wallis with Dunn's correction; associations used Spearman correlation and OLS with HC3 robust SEs. Predictive modeling used logistic regression with restricted cubic splines for oxytocin and an elastic-net multivariable model (oxytocin spline + leptin, BMI, waist circumference, HSI, VAI, and a PCA-derived EDE-Q component). Performance was estimated via leakage-free nested cross-validation (outer 5-fold, inner 5-fold) using out-of-fold (OOF) ROC AUC, Brier score, bootstrap CIs, calibration, and decision-curve analysis. Oxytocin was lower and leptin higher in metabolically unhealthy obesity (both p<0.01). Oxytocin correlated inversely with disordered-eating severity, while leptin correlated positively. The oxytocin-only spline model achieved OOF AUC 0.87 (95% CI 0.76-0.95; Brier 0.10). The combined elastic-net model achieved OOF AUC 0.97 (95% CI 0.90-1.00; Brier 0.05) and provided significantly better discrimination than oxytocin alone (ΔAUC 0.11, 95% CI 0.01-0.22; p=0.02). Using Youden's index on OOF predictions, the oxytocin-only model's optimal operating probability (0.69) mapped to ~90.5 pg/mL (95% CI 74.8-103.3), yielding sensitivity of 0.94 (0.87-0.99) and specificity of 0.83 (0.70-0.95). Decision-curve analysis showed higher net benefit for multivariable models across clinically relevant thresholds. Lower oxytocin is associated with greater disordered-eating severity, but oxytocin is most informative when integrated with metabolic and behavioral markers. A multivariable model substantially improved discrimination and net benefit over oxytocin alone. The ~90.5 pg/mL value is an exploratory operating point rather than a clinical cutoff; external validation and prospective evaluation are needed before translation to practice.