We defined hospitalists as physicians in general internal medicine who had at least five evaluation-and-management billings in a given year and generated at least 90% of their total evaluation-and-management billings in the year from services to hospital inpatients (Fig. 1 ). Since our source of data is a 5% sample, these five evaluation-and-management billings represent 100 or more charges to Medicare patients. Using inpatient evaluation-and-management billing codes (Current Procedural Terminology [CPT] codes 99221–99223, 99231–99233, and 99251–99255) and outpatient evaluation-and-management billing codes (CPT codes 99201–99205, 99211–99215, and 99241–99245), we calculated the percentage of each physician’s evaluation-and-management claims that were generated from services provided to hospitalized patients.
We analyzed the effect of different cutoff points according to the percentage of evaluation-and-management charges generated from care provided to hospitalized patients (≥80% vs. ≥90%) and according to the minimum number of evaluation-and-management charges in a given year in the 5% sample of Medicare data (≥5 vs. ≥10) in the algorithm to identify hospitalists. We tested the algorithm in a validation set of 57 hospitalists and 172 physicians in traditional non–hospital-based general internal medicine (hereafter referred to as nonhospitalists) employed in 2006 at seven hospitals. These hospitals were located in California (University of California, Los Angeles, Medical Center), Michigan (Wayne State University Detroit Medical Center), Virginia (Hospital Corporation of America [HCA]–affiliated hospitals in Richmond, including Henrico Doctors’ Hospital, John Randolph Medical Center, HCA Retreat Hospital, Johnston Willis Hospital, and Chippenham Hospital), Wisconsin (Sinai Samaritan Medical Center), and Texas (University of Texas Medical Branch, Clear Lake Regional Medical Center, and University of Texas Health Science Center at San Antonio). The algorithm requiring a minimum of 5 evaluation-and-management charges per physician in a given year and the algorithm requiring 10 or more such charges, with both requiring that 90% or more of the charges represent the care of hospitalized patients, had a sensitivity of 84.2% and 71.9%, a specificity of 96.5% and 97.1%, and a positive predictive value of 88.9% and 89.1%, respectively. The algorithm requiring 5 or more evaluation-and-management charges and the algorithm requiring 10 or more such charges, with both requiring that 80% or more of the charges represent the care of hospital inpatients, had a sensitivity of 87.7% and 73.7%, a specificity of 93.0% and 94.2%, and a positive predictive value of 80.6% and 80.8%, respectively. The sensitivities of the four algorithms were very similar (91.1%, 88.9%, 91.1%, and 88.9%, respectively) when applied to 45 hospitalists in two hospitalist groups serving community hospitals in the Houston and Austin metropolitan areas. We selected the algorithm requiring at least five evaluation-and-management charges with at least 90% of such charges generated from the care of hospital inpatients because the sensitivity (84.2%) and positive predictive value (88.9%) were acceptable.
We also evaluated the specificity of the algorithm by assessing whether hospitalists in general internal medicine identified by the algorithm submitted claims for procedures that are not usually performed by general internists; these procedures included colonoscopy, upper endoscopy, liver biopsy, hemodialysis, peritoneal dialysis, kidney biopsy, bronchoscopy, and cardiac catheterization. In 1995, the proportion of physicians identified as hospitalists who billed for one or more of these procedures was 14.9%; this percentage decreased to 2.3% in 2006. In some analyses, we also calculated the percentage of physicians in other specialties for whom more than 90% of evaluation-and-management billing codes were generated from services provided to hospitalized patients.
We analyzed the effect of different cutoff points according to the percentage of evaluation-and-management charges generated from care provided to hospitalized patients (≥80% vs. ≥90%) and according to the minimum number of evaluation-and-management charges in a given year in the 5% sample of Medicare data (≥5 vs. ≥10) in the algorithm to identify hospitalists. We tested the algorithm in a validation set of 57 hospitalists and 172 physicians in traditional non–hospital-based general internal medicine (hereafter referred to as nonhospitalists) employed in 2006 at seven hospitals. These hospitals were located in California (University of California, Los Angeles, Medical Center), Michigan (Wayne State University Detroit Medical Center), Virginia (Hospital Corporation of America [HCA]–affiliated hospitals in Richmond, including Henrico Doctors’ Hospital, John Randolph Medical Center, HCA Retreat Hospital, Johnston Willis Hospital, and Chippenham Hospital), Wisconsin (Sinai Samaritan Medical Center), and Texas (University of Texas Medical Branch, Clear Lake Regional Medical Center, and University of Texas Health Science Center at San Antonio). The algorithm requiring a minimum of 5 evaluation-and-management charges per physician in a given year and the algorithm requiring 10 or more such charges, with both requiring that 90% or more of the charges represent the care of hospitalized patients, had a sensitivity of 84.2% and 71.9%, a specificity of 96.5% and 97.1%, and a positive predictive value of 88.9% and 89.1%, respectively. The algorithm requiring 5 or more evaluation-and-management charges and the algorithm requiring 10 or more such charges, with both requiring that 80% or more of the charges represent the care of hospital inpatients, had a sensitivity of 87.7% and 73.7%, a specificity of 93.0% and 94.2%, and a positive predictive value of 80.6% and 80.8%, respectively. The sensitivities of the four algorithms were very similar (91.1%, 88.9%, 91.1%, and 88.9%, respectively) when applied to 45 hospitalists in two hospitalist groups serving community hospitals in the Houston and Austin metropolitan areas. We selected the algorithm requiring at least five evaluation-and-management charges with at least 90% of such charges generated from the care of hospital inpatients because the sensitivity (84.2%) and positive predictive value (88.9%) were acceptable.
We also evaluated the specificity of the algorithm by assessing whether hospitalists in general internal medicine identified by the algorithm submitted claims for procedures that are not usually performed by general internists; these procedures included colonoscopy, upper endoscopy, liver biopsy, hemodialysis, peritoneal dialysis, kidney biopsy, bronchoscopy, and cardiac catheterization. In 1995, the proportion of physicians identified as hospitalists who billed for one or more of these procedures was 14.9%; this percentage decreased to 2.3% in 2006. In some analyses, we also calculated the percentage of physicians in other specialties for whom more than 90% of evaluation-and-management billing codes were generated from services provided to hospitalized patients.