Background/Aims: To empirically classify phenotypes of eating disorders (ED) using latent class analysis (LCA), and to validate this classification based on clinical outcomes. Methods: LCA was applied to 968 inpatients. The resultant classes were validated by clinical outcomes including mortality. Results: A 5-class solution showed the best fit. The symptoms of latent class 1 (LC1; 26% of the sample) resembled anorexia nervosa (AN), bingeing-purging (AN-B/P) subtype; those of LC2 (23%) resembled bulimia nervosa; those of LC3 (11%) were close to AN-B/P without weight and body concerns; those of LC4 resembled restrictive anorexia nervosa (RAN) without weight and body concerns, and those of LC5 RAN. A history of hospitalization for ED was significantly more frequent for LC3 and LC4. The lowest BMI at admission were presented in LC4. LC1 showed the highest level of psychological disturbances and LC4 the lowest. LC3 and LC4 differed from LC1 and LC5 by higher percentages of treatment dropout (64.9 vs. 57.2 and 55.7 vs. 47.5%, respectively; overall p = 0.001). Survival rates tended to be different between the LC (p = 0.09). Conclusions: Subgroups of AN patients with low weight and body concerns seem more severe at hospitalization and more difficult to manage, with a higher rate of treatment dropout than the ‘typical’ AN patients.