Disease magnitude approach for estimating disease specific cost in health insurance claims AtoZ OKAMOTO,MD.MPH., Yasuyoshi SEKITA,Ph.D Kindai Medical School, OSAKA, JAPAN [atoz@med.kindai.ac.jp] Tohoku University Graduate School of Economics & Management, MIYAGI, JAPAN [Problem/Question] Health insurance claims commonly contain multiple diagnoses and it is often difficult to break down the aggregate cost into different disease categories. The author proposes an approach to break the aggregate cost based on the magnitudes derived from the primary/secondary relationships between two diseases when they coexist. [Modelling] Consider the following five insurance claims containing multiple diagnoses (disease A, B, C, D) and aggregate charges: diagnoses charges ----------------------------------------------------- claim 1: A, B, D 10,000 yen claim 2: B, D 150,000 yen claim 3: A, B 50,000 yen [1] claim 4: B, C, D 30,000 yen claim 5: A, B, C, D 40,000 yen total 280,000 yen Total of 280,000 yen was spent for treating five patients for four diseases namely A, B, C and D. Although it is obvious how much spent for each patient, it is less obvious how much spent for each disease. It may be possible to break down the total cost into disease categories for medication or examination because each drug and examination has specific indication. For example one can safely assume that the drug cost for H2 blocker was spent for peptic ulcer and EKG cost was for cardiac disease. However it is not practical to do the same for more generic type of medical services such as physicians' evaluation or nursing fee. Also breaking down the aggregate cost into disease categories requires medical knowledge and therefore not fit for computer processing. First, let us simply add up the number of diseases. Then we get: A x 3 B x 5 C x 2 [2] D x 4 total of 14 diseases for total cost of 280,000 yen The simplest way of dividing the total of 280,000 yen will be distributing the total cost in proporation to the number of diseases, i.e. giving an equal weight to each disease and assume that each disease cost 20,000 yen (280,000 yen divided by 14). Then we get: cost for A: 60,000 yen cost for B: 100,000 yen cost for C: 40,000 yen [3] cost for D: 80,000 yen total 280,000 yen However this method is obviously too simplistic because some diseases cost more than others to treat. Let us hypothetically assume that the "price" of each disease, or cost of cases whose primary diagnosis is each disease as follows: "price" for A: 30,000 yen "price" for B: 10,000 yen [4] "price" for C: 40,000 yen "price" for D: 50,000 yen By applying [4] to [2], we will get: cost for A: 30,000 X 3 = 90,000 yen cost for B: 10,000 X 5 = 50,000 yen cost for C: 40,000 X 2 = 80,000 yen [5] cost for D: 50.000 X 4 = 200,000 yen total 420,000 yen By adjusting the calculated total cost [420,000 yen] to the original total cost [ 280,000 yen] through multiplying with 2/3, we get the breakdown of the total cost weighted by the "price" of diseases. cost for A: 90,000 X 2/3 = 60,000 yen cost for B: 50,000 X 2/3 = 33,333 yen cost for C: 80,000 X 2/3 = 53,333 yen [6] cost for D: 200.000 X 2/3 = 133,333 yen total 280,000 yen [Disease magnitude approach] Still, this method is too crude because we assumed that all diagnosis contained in the claims are primary diagnoses, hence giving all of them the "price" while actually not all of them are primary diagnoses. Some diseases are more important than others in treatment. How can we know if disease A is more likely to be primary diagnoses than disease B and what likelihood it is? Fortunately we have the answer. Every three years Japan's Ministry of Health, Labor and Welfare (MHLW) conducts a nation wide, cross-sectional survey on sampled in and outpatients as to the primary and secondary (if any) dianoses to estimate the number of patients in each disease classification. This survey is called The Patient Survey, the latest of which was conducted in October 1996. The number of patients with disease A is defined as follows. A + Ax + Xa Whereas A is the number of patients who has only disease A and no secondary diagnoses, Ax is the number of patients who has disease A as a primary diagnosis and a secondary diagnosis of other disease X, Xa is the number of patients who has any disease X as primary a diagnosis but also has disease A as a secondary diagnosis. Then the likelihood of disease A to be a primary diagnosis is expressed as follows: A + Ax ------------- [7] A + Ax + Xa The likelihood to be a primary diagnosis varies from disease to disease ranging from as low as 6.8% (hypotension) to as high as 98.8% (schizophrenia) for inpatient according to the 96 Patient Survey. This variance makes one cautious in interpreting the list of dianoses in the claims. When one sees a diagnosis of schizophrenia in a claim, one can 98.8% certain that the diagnosis in the claim is treated as a primary dignosis. But if one sees a diagnosis of hypotension in a claim, one can assume that the diganosis is primary only with 6.8% likelihood. In other words, when one faces with 100 claims with diagnoses of schizophrenia, one can assume that 98.8 claims treat schzophrenia as primary diagnoses. However, for 100 claims with diagnoses of hypotension, one can assume that only 6.8 of them are primarily intended for treatment of the disease and. One never knows which of the 100 claims are primary diagnoses, but one does not need to know it just to evaluate how much the 100 claims are collectively worth: 100 claims with schizophrenia are worth 98.8 times the "price" of schizophrenia while 100 claims with hypotension are worth 6.8 times the "price" of hypotension. Let us hypothetically assume that the likelihood of each disease is as follows: likelihood of A to be primary diagnosis: 60% likelihood of B to be primary diagnosis: 50% [8] likelihood of C to be primary diagnosis: 80% likelihood of D to be primary diagnosis: 70% The "magnitude" of each disease will be expressed as a product of "price"[4] and likelihood[8] of the disease: "magnitude" of A: 30,000 yen X 60% =18,000 yen "magnitude" of B: 10,000 yen X 50% = 5,000 yen [9] "magnitude" of C: 40,000 yen X 80% =32,000 yen "magnitude" of D: 50,000 yen X 70% =35,000 yen Although disease B is most prevalent (every claim contains this disease), its magnitude in terms of health care cost is lowest among four diseases, and therefore should be given "lighter" weight in breaking down the total cost into disease specific cost. For example, claim 2 in [1] costs 150,000 yen for treatment of disease B and D. Given the heavier "magnitude" of disease D over disease B, one should assume that majority of 150,000 must have been spent for disease D. By distributing the 150,000 yen to disease B and D in proportion to their "magnitude (5,000 yen for B and 35,000 for D)", we estimate that the claim 2 spent 18,750 yen for disease B and 131,250 yen for disease D. The breakdown of the five claims and aggregate estimate of disease specific cost is given as follows: claim 1: A(3,103), B(862), D(6,035) total 10,000 yen claim 2: B(18,750), D(131,250) total 150,000 yen claim 3: A(39,130),B(10,870) total 50,000 yen claim 4: B(2,083), C(13,333),D(14,584) total 30,000 yen [10] claim 5: A(8,000), B(2,222), C(14,222),D(15,556) total 40,000 yen aggregate total A50,233, B34,787, C27,555, D167,425 total 280,000 yen Note that the disease specific cost in each claim varies within the total of each individual claim. Disease D accounts for nearly 60% of the total cost of five claims not only because it has a heavy magnitude but also because it occurred in an expensive claim. Disease B has a light "magnitude" but because of its high prevalence accounts for a relatively large share of the cost. This original method, Proporational Disease Magnitude (PDM) method, was first published in 1996 and a computer program is now available. [New approach using "paired" likelihood] The method described so far assumes that the likelihood of a disease to be a primary diagnosis and consequently the magnitude is constant regardless of other diseases present in a claim (comorbidity). For example, the likelihood of disease B to be a primary diagnosis is assumed to be 50% in both claim 2 and 3 where comorbidity is D and A respectively. However this is not a precise way to measure the magnitude of disease B because the likelihood of disease B to be a primary diagnosis may well vary depending on which disease is a comorbidiy. Disease B may have 80% likelihood to become a primary diagnosis when it coexists with disease A and may have 30% likelihood to become a primary diagnosis when it coexists with disease D. Likelihood to become a primary diagnosis is inherently relative: it is affected by other diseases coexisting in a claim. The Patient Survey gives information on the relative likelihood of which disease is likely to be a primary diagnosis when two diseases coexist. Let us call this "paired" likelihood because it refers to the relative likelihood for one disease to be a primary diagnosis over the other in a pair of diseases. For example, hypertension (HT) and diabetes (DM) commonly coexist in a patient. According to the 96 Patient Survey, there were estimated to be 59,500 outpatients visiting hospitals or clinics on a survey date for treatment of both diseases. Of those, 30,300 patients were diagnosed as having HT as a primary diagnosis and DM as a secondary diagnosis, and the rest 29,200 patients were diagnosed as vice versa. In other words, HT has 50.9% likelihood and DM has 49.1% likelihood when these two diseases coexist in a claim. HT also commonly coexists with peptic ulcer (PU). There were estimated to be 16,700 such patients. However HT was diagnosed as a primary diagnosis in 67% of those patients while in only 33% the primary diagnosis was PU. Therefore the likelihood of HT to be a primary diagnosis, and therefore its magnitude, is not the same depending on whether the comorbidity is DM or PU. Paired likelihood of selected pairs of diseases are expressed as combination of two diseases from disease categories. If there are 57 disease categories, there will be 1,596 pairs. But many of such potential pairs have few cases and may be omitted. According to the 96 Patient Survey, there are 995 pairs with meaningful sample size. If the likelihood of disease A to be a primary diagnosis over disease B is expressed as Ab, and Ba vice versa, Ab+Ba=1. [Application to multiple diagnoses] Many claims contain three or more diagnoses, for example, HT, DM and PU. For such "multiple" diagnoses, one can deal with it by treating such cases as if a "tournament", i.e. calculating the average likelihood of one disease against all other diseases. This appears as if calculating the probability of one team winning against all other teams. There are two ways for calculating the average likelihood: arithmetic mean and geometric mean. In the case of HT, DM and PU, HT has 50.9% likelihood against DM and 67% against PU. The average likelihood of HT against both DM and PU is calculated by either arithmetic mean: (50.9%+67%)/2 = 59% or geometric mean: SQRT(50.9% X 67%) = 58.4% By the same token, the average likelihood of DM and PU are calculated as: DM: arithmetic mean, 56.8%; geometric mean 56.2% GU: arithmetic mean, 34.3%; geometric mean 34.3% According to the 98 Social Insurance Claims Survey, per claim cost for these diseases were, 14,527 yen (HT), 19,437 yen (DM) and 16,284 yen (PU). If we take geometric mean, the magnitude of each disease will be: HT: 14,527 yen X 58.4% = 8,484 yen DM: 19,437 yen X 56.2% = 10,924 yen [11] PU: 16,284 yen X 34.3% = 5,585 yen For a claim containing these three diagnoses and the total cost of 100,000 yen, each disease is expected to be spent: HT: 33,946 yen DM: 43,708 yen [12] PU: 22,346 yen [Discussions] The logic of disease magnitude approach for estimating disease specific cost of a population of health insurance claims has the following advantages over traditional method of choosing a primary diagnoses arbitrarily and attribute all the cost contained in the claim to the selected one diagnoses: First, it is an objective method and is not affected by human judgement, and therefore reproduceable. This is appropriate for comparison between different populations either separated by time or geography. Second, it is suitable for computerized claims processing. This will eliminate time and cost of human work. Third, since it will consider all diagnoses listed in the claims, it will be able to estimate the cost for uncommon diseases. On the other side, this method has some limitations. It requires that all diagnoses be input to computer and the complexity of the logic requires a sophisticated program to hand a large number of claims. And for logical reasons, this method is not intended to estimate a disease specific cost in individual claim level because this is intended chiefly to break down the aggregate cost of a population of relatively large size. Also its validity has yet to be established. In order to demonstrate its validity, matching and linkage the claims information with medical records may be necessary. However for the sake of inter-population and inter-time comparison of the morbidity structure of insured populations, this method has enough utility without such verification. This method will be a useful tool for insurers to analyse the morbidity structure of their insured population for health planning and actuarial development. [caption]Matrix of "Paired" Likelihood to be Primary Diagnosis between Two Diseases (Ab and Ba are located symmetrically across the black cells. Ab+Ba=1)