Clinical and Nonclinical Predictors of the Cost of Coronary Bypass Surgery:  Potential Effects on Health Care Delivery and Reimbursement FREE

THE RESTRUCTURING of American health care has been driven primarily by economic concerns.¹ Because of intense competition and price controls, providers are being reimbursed less to furnish the same or better service than before.^2- 3 Formerly, reimbursement was correlated with the cost of providing care, a paradigm that accounted for individual patient characteristics. Now, many methods of reimbursement involve single prices—there is no variation in reimbursement based on individual patient characteristics. As competition increases and reimbursement continues to decline, providers may not be able to cover the cost of care for many inherently high-cost patients. This situation may lead to the denial of care to a patient (based on cost) who might receive medical benefit from an expensive therapy.^4- 5

To optimize pricing in such a competitive environment, both payers and providers must understand the drivers of cost and how to influence them. In the case of coronary artery bypass surgery (CAB), investigators have identified a series of preoperative variables that are predictive of cost in individual patients,^6- 8 but these studies have focused almost solely on traditional clinical variables. Patients who undergo CAB are quite varied from a clinical point of view, but even more so from a demographic perspective.^9- 10 Because our health system serves a broad range of patients—both clinically and demographically—it is important to examine the role of both types of variables as contributors to the cost of care. With a view toward probing patient-specific predictors of the cost of CAB, we tested the hypothesis that nonclinical variables will add significant, incremental cost information to traditional clinical variables alone.

Included in this study were 882 consecutive patients who underwent routine CAB without catheterization on the same admission at Cedars-Sinai Medical Center, Los Angeles, Calif, from March 14, 1990, to March 22, 1995. Diagnosis-related group (DRG) 107 (group associated with CAB and no catheterization) was chosen because it is a common procedure for which payers and providers negotiate specific contracts for large numbers of patients. Also, DRG 107 is not confounded by variable preoperative lengths of stay or other related therapies as would be the case with DRG 106 (CAB with catheterization on the same admission). Furthermore, because the DRG classification system is widely used, these data should have applicability in centers that do not have sophisticated clinical information systems.

All data were obtained from our cardiothoracic surgery database as previously described.¹¹ Briefly, demographic and nonclinical data are obtained by an electronic link between the hospital administrative database and the cardiothoracic surgery clinical database. Preoperative history, physical examination results, and laboratory information are prospectively obtained by trained clinical personnel and by chart review, when necessary. Intraoperative data collection is performed by operating room personnel who interview the surgeon during the procedure. Financial data are obtained by an electronic link between the hospital financial system and the clinical database.

A series of 21 clinical and 34 nonclinical variables were selected for investigation. All variables used are listed in Table 1, and below we define those that are not self-explanatory.

Patient type refers to our grouping of all patients into 1 of 3 medical-care categories. A private patient is one who is admitted to the hospital under the care of a private physician. A clinic patient receives routine care in the Cedars-Sinai internal medicine resident's clinic. A teaching patient is one who has no private physician, is not followed up by our clinic, and often has had little contact with the medical system before surgery. Both clinic and teaching patients are cared for by residents and full-time hospital staff.

We used a formal socioeconomic-status scale that is based on a ZIP code algorithm developed by Claritas Inc (Los Angeles, Calif) in which each ZIP code is given a geounit quality score between 0 and 100. The scores are based on the 1990 US Census data and are derived from a weighting of household income, educational level, occupation, and home value. The 7 commonly used socioeconomic classes are parsed by the quality score as follows: bottom, 1 to 19; lower, 20 to 29; lower-middle, 30 to 39; middle, 40 to 59; upper-middle, 60 to 69; upper, 70 to 79; and top, 80 to 100. The mean score for the US population is 50. Each score was divided by 10 so that a unit change in socioeconomic status approximates a change of 1 socioeconomic class.

Primary insurer is categorized into private insurance (a combination of indemnity insurance, health maintenance organization, and preferred provider organization), Medicare, MediCal (the California equivalent of Medicaid), cash, and other. Living arrangements were defined as living with a spouse (or other support person) or alone. Activity level was classified into active, moderately active, sedentary, and restricted based on the patients' own assessments of their physical activity during the interview.

We defined 3 primary symptoms present at admission: angina, shortness of breath, and other (a combined variable of syncope, near syncope, fatigue, and sudden cardiac death). The primary symptom is the main symptom that leads to the procedure. For example, a patient may have angina with congestive heart failure, but the angina may be minimal. In this case, the primary symptom might be shortness of breath, but angina will also be present (see below).

If the patient has angina (with or without other symptoms), we include type of angina (typical, atypical, or nonanginal chest pain), frequency of angina (many times per day, once per day, a few times per week, or weekly or less), Canadian Cardiovascular Society class, and stability of symptoms during the last 6 weeks (stable, improved, progressive, or unstable). If the patient has congestive heart failure, we include the frequency of the symptoms (many times per day, once per day, a few times per week, or weekly or less), the New York Heart Association class, and the stability of the symptoms during the last 6 weeks (improved, stable, or worsening).

The Cedars-Sinai cost-accounting system tracks all patient-specific costs by a bottom-up approach (direct measurement of costs, not estimates of costs from charges) and measures direct and indirect costs of hospitalization. Because indirect costs are not directly related to patient care, and often vary significantly between hospitals, we focused this study solely on direct costs. The financial system of the hospital defines direct costs as those related to providing direct patient care, including salaries, supplies, repair and maintenance of equipment, dues and subscriptions, outside training, and transfer expenses. Outpatient costs and professional fees were not included in this analysis. Because of the assumptions of multiple linear regression (see below), the dependent variable used in the statistical analysis was the natural logarithm of direct cost.

All statistical analysis was performed by a commercially available software package (Statistica, StatSoft Inc, Tulsa, Okla). Where applicable, all data are presented as mean ± SD. For explanatory purposes,¹² multiple linear regression with 1 dependent variable (natural logarithm of direct cost) and 55 independent variables was performed, and variables with a P value of .05 or less were determined to contribute significantly to direct cost. Missing values were replaced with variable means. Regression results are expressed by the common raw regression coefficient (B), its SE, the P value, and the standardized regression coefficient (β), which is a calculated coefficient assuming all independent variables were normalized to a mean of 0 and an SD of 1. Using β allows the comparison of the relative magnitude of contribution of the independent variables to the dependent variable prediction.

Because the dependent variable (direct cost) was logarithmically transformed (to satisfy the normal distribution requirements for linear regression), the results of the regression cannot be interpreted directly into dollar amounts. Accordingly, the regression coefficient B represents the logarithm of a ratio of costs based on an increment of the variable in question. For example, in the case of a categorical variable such as hypertension, a B of 0.05 would be transformed to its exponent (1.05), and this exponent would indicate that the presence of hypertension increases cost by 5%. Similarly, given a B for age (per decade) of 0.06, its exponent (1.06) indicates a 6% increase in cost for each increment of 1 decade.

The mean age was 66.7 ± 9.6 years with 84% male patients. The average number of distal anastomoses was 3.2 ± 1.0, the surgical mortality (in-hospital and 30-day mortality) was 0.8% (7/882 patients), and the average length of stay was 9.6 ± 5.7 days. Table 1 presents greater detail of the population's characteristics. The mean direct cost in this population was $13183 ± $6448 with a range of $3458 to $75349. The geometric mean of direct cost was $11614.

Fifteen of the 55 independent variables significantly contributed to the direct cost estimates (Table 2); 11 were traditional clinical variables, and 4 were nonclinical. Based on the magnitude of the exponent of B, the variables with the largest cost impact (>10% increase in cost) include 4 of the 11 traditional clinical variables (other symptoms, congestive heart failure–New York Heart Association, peripheral vascular disease, and preoperative dialysis) and 3 of the 4 nonclinical variables (teaching patient, surviving spouse, and restricted activity).

Teaching patients had an increase of 25% in their direct costs, while a patient with restricted physical activity preoperatively had an increase in cost of 61%. A surviving spouse had an increased cost of 12.4%, and an elevated socioeconomic status actually decreased cost by 2.6% per socioeconomic class.

Of the traditional clinical variables, preoperative dialysis increased the cost of CAB by 279.3—the largest magnitude of any of the variables. A presentation of other symptoms (no chest pain or shortness of breath) increased cost by 12.9%. Each increment in the congestive heart failure–New York Heart Association variable increased cost by 11.3%. The presence of peripheral vascular disease increased direct costs by 10.7%.

Using the β coefficient to compare the relative contribution of normalized variables, the strongest predictors of direct cost were age, socioeconomic status, angina class, congestive heart failure class, and preoperative dialysis. The presence of hypercholesterolemia, a high socioeconomic status or prior percutaneous transluminal coronary angioplasty tended to lower costs (a negative coefficient). Two of the nonclinical variables (teaching patient and restricted activity) were stronger predictors of cost than most of the traditional clinical variables.

Market forces are moderating the cost of health care.^13- 14 Competition among providers is being encouraged by payers to improve service and lower cost. As reimbursement continues to decline, profitability and survival may be affected unless providers return to economic basics—understanding and controlling costs. Classically, the costs incurred in the production of a tangible product are usually fixed and, therefore, highly predictable—a defined number of components along with a given number of working hours. But in providing a service, such as medical care, the associated costs are highly variable and dependent on the region of the country, the hospital, the individual physician, and individual patient characteristics.^9- 10 The cost of providing health care is inextricably bound to a highly variable patient population (in clinical and demographic terms) and to the inherent variability in the way that providers care for those patients (differing approaches to patient care).

In this study on the determinants of hospital provider cost, we have demonstrated that a series of preoperative clinical and nonclinical characteristics significantly contribute to the cost of surgery. Our results are similar to previously published reports^6- 8 based on estimates of cost from charge data. The major finding of our study, and the difference from prior studies, is that preoperative nonclinical variables can have a major influence on the cost of CAB. Teaching patients, surviving spouses, high socioeconomic status, and restricted activity had significant effects on cost, and 3 of these had greater effects than many of the clinical variables. We believe these data expand our understanding of CAB procedural costs and will aid providers in developing local contracting strategies, will improve the models used in physician economic credentialing, and could provide a basis for reevaluating national reimbursement strategies.

The melding of traditional clinical databases with accurate financial data provides a powerful tool to understand and control the hospital costs of CAB and to thereby optimize contracting strategies in a competitive marketplace. Accurate knowledge of individual patient costs, linked to patient clinical characteristics, would allow the development of market and patient-specific pricing. Assume that a provider is negotiating a carve-out contract (a contract in which a provider will perform a specific service, usually at a reduced price) for CAB with a payer. Decisions regarding proposed pricing for contracts such as these are often based on hospital-specific estimates of average costs for all patients undergoing CAB. If we know the characteristics of the population of potential patients, we can use statistical models to determine the typical costs of providing that service. As a specific example, based on our data, we would estimate that a Medicare contract for CAB (assuming an average age of 75 years) should cost 13% more than a commercial contract (assuming an average age of 55 years) based on age characteristics alone. More detailed knowledge of the population's characteristics would allow further refinement of that estimate, since the combination of clinical and nonclinical data can improve the accuracy of these cost estimates.

Another use of these cost models would include pricing estimates for individual cash-paying patients. Many hospitals care for patients who pay for services in cash, and these patients will often contact a variety of providers to obtain the best price for any particular procedure. Instead of quoting a single price for a given procedure, the price might be decreased for a potentially lower-cost patient and raised proportionately for one of potentially higher cost. By knowing the clinical and nonclinical characteristics that affect cost, a more accurate cost estimate can be computed, thus optimizing individual patient pricing.

It has become common to compare physicians based on their clinical outcomes of care, but now physicians are being compared based on their financial outcomes. One complaint regarding physician comparisons has been that raw data do not accurately reflect an individual physician's outcomes since the raw data do not take into account the characteristics of the population cared for. Raw data can be adjusted to account for variability in patient characteristics—an example is risk-adjusted mortality.¹⁵ Similarly, if physicians are to be compared based on their economic outcomes,^16- 19 then these comparisons should be risk adjusted to account for patient variability and referral bias. For example, a single physician in a group may be referred most of the potentially high-cost patients, but this would not be accounted for by observing raw cost data. The use of models similar to what we have described here (including clinical and nonclinical variables) would make economic credentialing a more credible method for addressing physician cost variability.

The variability in procedural reimbursement and provider cost structure has placed a tremendous burden on our health system. Many providers have adapted well and been profitable, while others have had to downsize, consolidate, or close.²⁰ Some providers are clearly burdened with higher-cost patients compared with others, potentially impacting their competitiveness.^21- 25 These adverse conditions may be exacerbated in large urban centers, especially in inner-city, university teaching programs, because of population demographics and because competition among providers is greatest in large urban areas where health insurance companies have the largest number of insured lives.

A potential solution to the large variability in reimbursement would be to adjust reimbursements based on individual patient characteristics (both clinical and nonclinical). Objective cost data, such as we have described here, could demonstrate the fact that some providers do care for higher-cost patients, and thus they deserve a differential based on these higher costs. The present DRG-based reimbursement system is an example of the lack of adjustment based on individual patient characteristics—the only patient-specific classifier is the DRG itself. The only adjustments (other than outliers) are based on the health system's characteristics (the type of hospital, the region of the country, and local wage rates²⁶), not the individual patient. A reimbursement system with continuously declining payments (eg, Medicare) that ignores individual patient characteristics will eventually be unable to cover all the costs of providing care for high-cost patients. The implications for restricted access based on economics are of concern to all involved in health care delivery.

Furthermore, data regarding cost variability as a function of clinical and sociodemographics add an important dimension to the national debate about the funding of health care, specifically Medicare. In our approach to health care cost containment, we must be attentive to the fact that patients differ widely and a single reimbursement cannot apply to all. Budgeted reimbursement and competition have reduced costs for Medicare and "commercial" (<65 years old) populations, but our data indicate that, if Medicare reimbursement approaches that of a commercial population, the provider will not be able to cover all costs. Thus, a provider may choose not to perform a beneficial procedure because it is not profitable.^27- 29 Policy makers must, therefore, be informed regarding the inherently higher procedural costs of specific populations (Medicare included) when they rethink reimbursement strategies.

Two factors strengthen the results of our study. First, as opposed to other investigations, we have used a bottom-up approach to the calculation of cost. Some investigators do not have a mechanism for measuring costs directly and use a percentage of charges as a costing basis. Investigators who have access to directly measured expenses (where supplies, person-hours, and other costs are captured by the accounting system) use the bottom-up approach. Because of the poor correlation between costs and charges, true costs (those measured directly and not calculated as a percentage of charges) should be used whenever possible.³⁰ Second, our results are consistent with previously published reports about the contributions of specific clinical variables to the cost of CAB in which cost-to-charge ratios were used instead of direct measurement of cost.^6- 8

As with any investigation, there are many significant limitations to our study. First, the data were derived from our observational cardiothoracic surgical database, which, while providing prospective data, is nonrandomized. Also, our hospital, located in a high-cost metropolitan area, is a university-affiliated, tertiary-care center in which costs have been traditionally higher, and, thus, these data may not directly apply to other hospitals. Furthermore, our patients tend to have a higher socioeconomic status, compared with the national average. A large proportion of the patients are fully insured, have excellent family support, know how to access the health system, and can afford to purchase health care aid in the general marketplace.

Also, during this period, our low surgical mortality rates (0.8%) could have skewed the cost data. Higher death rates could increase costs for patients who die in the hospital late after surgery or lower costs if they die early in the postoperative course. Furthermore, we have not included the potentially large costs of complications,³¹ though we are presently investigating this issue.

In a competitive health care market, providers are increasingly being forced to focus on financial outcomes as well as the quality of clinical outcomes. Because of the large variation in individual patient costs, providers must understand which individual patient characteristics drive costs. In a select cohort of patients undergoing CAB, we have demonstrated that in addition to previously described clinical variables that affect the cost of CAB, a series of nonclinical variables add important incremental information to these cost estimates. This is an important step in defining the drivers of cost for CAB and other expensive therapies. Data such as these could improve our ability to optimize contract pricing, improve the process of economic credentialing, and provide a basis for rethinking national budgeting and contracting strategies. The restructuring of the financial relationship between payer and provider is key to assuring optimal pricing strategies for high-cost therapies and maintaining the excellence and viability of our health system—the combination of patient, payer, and provider.

Accepted for publication July 9, 1997.

These data were presented in part at the meeting of the American College of Cardiology, March 22, 1995, New Orleans, La.

Reprints: Timothy A. Denton, MD, Cedars-Sinai Medical Center, 8700 Beverly Blvd, Room 6215, Los Angeles, CA 90048 (e-mail: denton@csmc.edu).