Objective: The cost-effectiveness of heart failure (HF) disease management depends on avoiding future high costs. Prospectively identifying HF patients who are likely to incur high costs would be beneficial. Methods: We used a 100% sample of 1,363,977 Medicare beneficiaries hospitalized with a primary diagnosis of HF (ICD-9-CM codes 428.x, 402.x1, 404.x1, 404.x3) between 2001 and 2004. The earliest HF hospitalization for each beneficiary was considered the index. We summed Medicare payments for rehospitalizations in the year following the index hospitalization, adjusted costs to 2001 dollars, and created a binary variable, with patients in the 4th quartile (>$16,500) defined as “high cost.” Comorbidities and risks were obtained from the index claim and from inpatient claims in the prior year. Logistic regression was used to predict high cost status in a 75% random derivation sample; the model was validated in the remaining 25%. We evaluated the calibration and discrimination of the model in both samples and refit the model on the entire sample. Results: Average Medicare payments in the year following index hospitalization were $38,300 (SD $29,146) among high cost patients and $4272 (SD $4857) among patients in the lower 3 quartiles. Inpatient cost in the prior year was the strongest predictor of inpatient cost in the subsequent year (OR 2.31, 95% CI: 2.27-2.35 for prior year inpatient costs >$16,500 vs. no inpatient costs in the prior year.) In both the derivation and validation cohorts, 11% of patients in the lowest decile and 45% of patients in the highest decile were high cost. The model was well-calibrated. The c-statistic was 0.65 for both the derivation and validation cohorts. Conclusion: There is limited ability to predict high cost HF patients using claims data alone. Future studies should assess the value of incorporating clinical variables.