Predicting inpatient costs with admitting clinical data.

Hospital cost-containment programs should themselves be cost-effective, targeting high-cost physicians (which requires adjusting for case mix) and patients (which requires early identification). In this study, clinical data available within 24 hours of admission from an electronic medical record system were used to develop statistical models to predict hospital costs. In this retrospective analysis of clinical data and diagnosis-related groups (DRGs), study subjects were 2,355 patients admitted for at least 1 day to the medicine service at an urban teaching hospital with sophisticated electronic medical records. Of these 2,355 patients, 1,663 (71%) had one of the 41 most common DRGs. Predictive models were derived on a random subset of two thirds of the patients and were validated on the remaining third. The following patient data were obtained: admission and prior diagnostic test results, diagnoses, vital signs; demographic data; prior inpatient and outpatient visits; tests and treatments ordered within 24 hours of admission (discretionary data); DRGs; and total inpatient costs (estimated from charges). Diagnosis-related groups explained 24% of the variance in total costs in the derivation patient set and 16% in the validation set. When only nondiscretionary data were used, the models retained only clinical laboratory results and prior diagnoses, explaining 20% of the derivation set variance in total costs and 16% in the validation set. Adding DRGs increased the variance explained in the derivation set to 34%, but decreased to 24% in the validation set. Adding discretionary data substantially increased the explained variance in the derivation and validation patient sets. The models' median predicted costs underestimated true costs by 10% to 13%, with the lowest error in the models using all types of variables. Clinical data gathered during routine clinical care can be used to adjust for case mix and identify high-cost patients early in their hospital stays, when they could be targeted by cost-containment interventions.