The purpose of this study was to develop and validate the Nursing Severity Index, a new method used to measure the admission severity of illness of hospital patients using nursing diagnoses, which categorize biologic, functional, cognitive, and psychosocial abnormalities. This retrospective cohort study with independent development and testing phases was conducted at a U.S. academic medical center. In the development phase, data regarding 14,183 adult medical-surgical patients admitted to the medical center in 1985 and 1986 was used. In the testing phase, data regarding 7,302 patients admitted in 1987 and 1988 was used. Primary nurses prospectively recorded the presence or absence of 61 nursing diagnoses on admission. Demographic and clinical data were obtained from hospital data bases. In the development phase, the number of admission nursing diagnoses was highly related (P < 0.001) to in-hospital mortality. Using multiple logistic regression, 34 nursing diagnoses were identified as independent predictors of mortality; the Nursing Severity Index equals the number of these 34 diagnoses. In the testing phase of 7,302 patients, the Nursing Severity Index was related (P < 0.001) to mortality rates, which were 0.5%, 1%, 2%, 6%, 13%, 22%, and 31% in seven hierarchical strata defined by the Index. The Index was as accurate in predicting mortality as MedisGroups (receiver-operating-characteristic curve areas, 0.814 +/- 0.016 vs. 0.845 +/- 0.015, respectively, P = 0.12). Furthermore, the Nursing Severity Index and MedisGroups together (receiver operating characteristic curve area 0.880 +/- 0.014), were more accurate (P < 0.01) than either measure alone. The Nursing Severity Index assesses multiple dimensions of illness, can be easily measured during routine patient care, accurately predicts the risk of in-hospital death, and has similar prognostic accuracy as MedisGroups. Its usefulness in outcomes assessment, quality assurance, and case management merits further study.