The term "medical necessity" generally refers to whether a health service should be used in a particular instance or for a specific person. Making this decision is distinct from defining insurance benefits, a process that determines whether a service is covered under a health care contract. A service may be covered but not necessary (for example, not all patients with coronary artery disease need coronary bypass surgery) or may be necessary but not covered (for example, by contractual exclusion). In some cases, such as in deciding when to cover a new technology or drug, the two concepts overlap [7]. Although this issue is important, we focus on the manner in which "medical necessity" is used by providers to decide whether to offer or deny a covered and established service to a particular patient.

In deciding whether a service is medically necessary, insurers and other health care providers have traditionally focused on simple, deterministic decision rules based on the idea that without the service, harm will come to a patient and that with the service, a potentially beneficial outcome will result. In other words, they apply a rule that allows for one of two outcomes: Yes, a service is necessary; or no, a service is not necessary. However, setting such a threshold is illusory because patients, providers, insurers, courts, and society have different values and objectives [8]. These differences would mean little if language pertaining to necessity were not found in almost every health insurance contract, both private and public [3], and if decisions involving medical necessity did not affect health care for thousands of patients in the United States.

The failure of the concept of medical necessity to address such diverse issues and differing priorities has led some to advocate cost-effectiveness as a basis for clinical choices [9]. However, methods used to define and assess what is and is not cost-effective vary [10]. Because we are concerned with the manner in which cost-effectiveness will be used to offer or deny patient services, we refer to the definition given by Eddy [9] in a recent article on coverage criteria and benefit language. He states that "an intervention is considered cost-effective if there is no other available intervention that offers a clinically appropriate benefit at a lower cost."

We describe a simple framework that shows why medical necessity and cost-effectiveness do not provide deterministic thresholds for clinical behavior. We first use Bayesian 2 x 2 tables and a receiver-operating characteristic curve to show that medical necessity involves a tradeoff between trying to extend a service to all patients who might benefit from it and trying to deny a service to those who will not benefit. We then use this framework and an optimization Equation based on maximizing the number of lives saved to show why patient, societal, and insurer perspectives on cost-effectiveness may differ. Our goal is not to downplay the need to further refine how to provide health care but rather to discuss the implications inherent in using medical necessity or cost-effectiveness to deny or provide services. Although these issues may be familiar to economists and decision analysts, we seek to clarify and present them in a way that is familiar to practicing physicians and policymakers.

Defining Medical Necessity

The accuracy of any definition of medical necessity should be assessed by comparing the definition to "true" medical necessity for the population to which the definition is applied. This assessment is problematic because there are few guaranteed outcomes in medical science. Instead, we must use imperfect estimates of potential benefit and harm and of diagnostic and therapeutic accuracy. We rarely know in advance who will benefit from a service and who will not; thus, we can rarely identify those for whom a service is truly "necessary." A decision rule based on medical necessity will therefore intuitively involve a tradeoff: We either treat some patients to benefit a few or we do not treat some patients because only a few would benefit. For example, a decision rule stating that a magnetic resonance imaging scan is "not necessary" in the assessment of patients with uncomplicated headaches might lead to lack of detection of only one brain tumor in, say, 20 000 patients. But for the one person who actually had a brain tumor, the test was, in retrospect, "necessary," albeit at the expense of many scans that were not.

If we cannot know in advance who should receive a service, how can we compare our estimate of necessity (whether a service should be provided) with "actual" necessity (whether our decision to provide or deny a service was correct or incorrect)? As Park and Brook [11] showed in a discussion about appropriateness, we can divide patients into two groups: those whose cases were classified as "necessary" and those whose cases were classified as "not necessary." Hypothetically, these classifications could then be evaluated as to whether the case was actually necessary or not actually necessary. This can be summarized by a Bayesian 2 x 2 table (Table 1). We can now calculate the sensitivity and specificity of any given definition of medical necessity. Errors are either "actually necessary but defined not necessary" (false-negative errors) or "not actually necessary but defined necessary" (false-positive errors). The sensitivity is the likelihood that a necessary case is correctly identified: Equation 1

(1)

Similarly, specificity is the probability that a "not necessary" case is appropriately classified: Equation 2

(2)

By changing the criteria for what is necessary, we can change the sensitivity and specificity of a definition as it affects a population. This technique is used to assess the accuracy of diagnostic tests, for which we vary sensitivity and specificity by using different cut-offs for positive and negative results. For example, variably defining a positive exercise tolerance test result by the degree of ST-segment depression [12] identifies different sets of patients within a population as eligible for the next procedure, angiography. The lesson for medical necessity is similar when different criteria are applied to a population. In other words, we can increase the sensitivity of a medical necessity decision rule only at the expense of decreasing its specificity.

An example helps to illustrate this tradeoff. Let us address the issue of how large an asymptomatic abdominal aortic aneurysm should be to qualify for surgical repair, a topic that has been extensively debated [13, 14]. Using estimates of rupture rates for aneurysms of different sizes, we can determine the sensitivity and specificity for different decision rules on when to intervene surgically. The expected rates, approximated from aggregated data, are as follows: 0.25% for aneurysms 3.5 to 3.9 cm in diameter; 3% for those 4.0 to 4.9 cm in diameter; 9% for those 5.0 to 5.9 cm in diameter; and 24% for those greater than 6 cm in diameter [13-15]. For simplicity, we assume that a patient population of 40 000 persons with aneurysms of different sizes, all larger than 3.5 cm, is evenly distributed and that aneurysms smaller than 3.5 cm in diameter have a rupture rate of 0%. We also assume that the expected number of ruptures actually occurs in our population. We define "actually necessary" as operating on an aneurysm that would have ruptured and "not actually necessary" as operating on an aneurysm that would not have ruptured.

We can now calculate the sensitivity and specificity of surgical interventions on aneurysms of different sizes (Table 2). If we operate on all aneurysms that are 4 cm in diameter or larger, our sensitivity is 94% and our specificity is 27%. Operating only on aneurysms that are 5 cm in diameter or larger yields a sensitivity of 86% and a specificity of 54%; limiting intervention to aneurysms that are 6 cm in diameter or larger allows a sensitivity of 62% and a specificity of 79%.

The tradeoff between sensitivity and specificity can be seen on a receiver-operating characteristic curve (Figure 1) on which the true-positive rate (sensitivity) is mapped against the corresponding false-positive rate (1 minus the specificity). The curve therefore represents a continuum of decision rules. It would obviously vary if we used different underlying assumptions about the incidence and rupture rates of aneurysms, but the overall conclusion would be the same. In other words, because we cannot prospectively identify the patients in a population who will benefit from surgery (for example, the patients in whom the asymptomatic abdominal aneurysm would have ruptured without intervention), we operate on many persons who might not ultimately have benefited from surgery. Putting this in human terms means that in choosing the 4-cm necessity rule, we potentially save 3600 persons from a ruptured aneurysm but operate on an extra 26 400 persons with "not necessary" cases. In contrast, choosing the more stringent 6-cm rule leads to only 7600 "not necessary" operations but allows 1200 aneurysms to rupture.

View larger version (11K): [in this window] [in a new window]   Figure 1. Receiver-operating characteristic curve for surgery done on patients with 4-cm, 5-cm, and 6-cm abdominal aortic aneurysms.

The model is incomplete because patient factors and issues such as perioperative mortality and financial and human costs have been left out. Nonetheless, choosing a predetermined point of intervention affects how risks and potential benefits are acted on and interpreted by both individual patients and targeted populations [8]. Because perfect predictive value for any decision is impossible to attain, any intervention point must balance the false-positive rate with the true-positive rate in a given population. Thus, any simple deterministic rule invoking necessity will have to choose a particular level of sensitivity and specificity; as we discuss in the following section, each choice results in a different set of human and financial costs.

From Necessity to Cost-Effectiveness

Given the problems with defining medical necessity and applying the definition, some authors have reasonably questioned whether necessity is a helpful construct in deciding who should receive a service [3, 6, 16]. Cost-effectiveness has increasingly been advocated as a basis on which to decide whether or not to provide services [9]. Let us consider this issue further. Using our model, we can, from a societal viewpoint, maximize the number of lives saved by determining how large an aneurysm should be to qualify for surgery. The optimal threshold for intervention balances the number of lives saved by surgery (number of ruptures prevented) against the number of deaths caused by surgery (perioperative mortality rate) [17]. We first calculate the percentage of expected deaths prevented by our intervention. To do this, let us assume an overall mortality rate of 80% for a ruptured aneurysm. Thus, overall, 3080 patients would die of a ruptured aneurysm if surgery is not done (0.80 x 3850). We can now determine the number of lives saved for each decision rule under varying perioperative mortality rates (we have optimistically chosen 2%, 4%, and 6%). As can be seen in Table 3, the optimal decision rule for maximizing the number of lives saved (2275) would involve surgery for aneurysms 4 cm in diameter or larger.

Unfortunately, the choice of any decision rule has financial implications. The dilemma for providers and payers in a world where health care services are open to competitive bidding is to balance financial costs and outcomes. This is especially difficult to do when costs and outcomes are not constant across hospitals or communities. Let us now consider the following: Hospital network A charges $40 000 per aneurysmal repair and has a perioperative mortality rate of 2%; hospital network B charges $30 000 and has a perioperative rate of 4%; and hospital network C charges $20 000 and has a perioperative mortality rate of 6%. Table 3 summarizes the number of lives saved, the total cost of all interventions, and the cost per life saved for each potential intervention point.

The optimal societal threshold occurs when surgery is done on persons with aneurysms 4 cm in diameter, given the 2% perioperative mortality rate within hospital network A. Given unlimited resources, society would prefer this choice if the only objective were to maximize the number of lives saved. That decision, under our stated conditions, is the most costly overall and the least "cost-effective" in terms of cost per life saved. On the other hand, the most cost-effective choice is to treat the largest aneurysms within hospital network C (with its 6% perioperative mortality rate). Thus, optimizing outcomes in terms of lives saved and optimizing cost-effectiveness in terms of cost per life saved do not necessarily go hand in hand.

Further complicating the issue is that individual patients may have an entirely different perspective when weighing relative risks and benefits [8]. Patients having surgery will preferentially pay a higher cost if the potential for a positive outcome from surgery (that is, life rather than death) is improved. Patients will preferentially choose the hospital that has the lowest perioperative mortality rate (and worst cost-effectiveness), even if their choice does not mesh with society's or the payer's interests. Ultimately, the needs of individual patients must be balanced against financial constraints, but the choice is not always clear. For instance, within the category of 5-cm aneurysms, hospital networks A, B, and C differ little in the total number of lives saved but differ greatly in the cost per life saved.

Policy Implications

We have used a model based on concepts familiar to most physicians to show that decision rules based on medical necessity and cost-effectiveness are not truly deterministic. In the case of medical necessity, there is a tradeoff between the sensitivity and the specificity of a decision rule as applied to a population. For cost-effectiveness, the tradeoff is more complicated and includes balancing outcomes (however specified), relative costs, societal values, and the patients' risks and preferences. Although the example of repair of asymptomatic abdominal aortic aneurysms should not be construed as a formal analysis, it shows a straightforward application of the principles of our model and shows that decision rules have unavoidable tradeoffs. This, combined with the uncertainty inherent in health care and the fact that decision analyses are very sensitive to assumptions, argues against using medical necessity or cost-effectiveness as a basis for a deterministic rule.

If we cannot use rules based on medical necessity or cost-effectiveness as deterministic thresholds, the problem of how to anchor clinical decisions still remains. We believe that rules based on medical necessity and cost-effectiveness assist in defining relative but not absolute thresholds for decision making. Decisions rules must be able to incorporate patients' individualized risks, benefits, and preferences [18] and must allow for reasonable differences in physicians' judgments and beliefs about available treatments. Variance will naturally occur and should be expected. Therefore, decision rules based on medical necessity and cost-effectiveness should be viewed as tools to help patients, providers, and payers make better-informed decisions about patient care.

Finally, in considering any proposed decision rule, the objectives being maximized must be made explicit and individual utilities and societal values should, whenever possible, be considered. Our model provides another way to visualize how changing decision rules might affect a population (that is, if we choose rule X rather than rule Y, how many more persons will derive a specified outcome, how many will be harmed, and how much will it all cost?) [3, 19]. In the current climate of managed care and cost-cutting, this type of analysis ties rules on providing care to the leading edge of medical decision making and, in so doing, makes the system of care as a whole more coherent.

Dr. Model: National Economic Research Associates, 555 S. Flower Street, Los Angeles, CA 90071.

Dr. Kahan: European American Center, Langbergstraat 6, 2628 CE Delft, the Netherlands.

Dr. Jacobson: Department of Health Management and Policy, University of Michigan School of Public Health, 109 Observatory, Ann Arbor, MI 48109-2029.

Dr. Peabody: RAND, PO Box 2138, 1700 Main Street, Santa Monica, CA 90407-2138.