Table of Contents Author Guidelines Submit a Manuscript
Computational and Mathematical Methods in Medicine
Volume 2018, Article ID 5091096, 6 pages
https://doi.org/10.1155/2018/5091096
Research Article

Illness-Death Model in Chronic Disease Epidemiology: Characteristics of a Related, Differential Equation and an Inverse Problem

1Hiller Research Unit for Rheumatology, University Hospital Duesseldorf, Duesseldorf, Germany
2Institute for Biometry and Epidemiology, German Diabetes Center, Duesseldorf, Germany

Correspondence should be addressed to Ralph Brinks; ed.frodlesseud-inu.zdd@sknirb.hplar

Received 30 January 2018; Revised 7 August 2018; Accepted 16 August 2018; Published 12 September 2018

Academic Editor: Andrzej Kloczkowski

Copyright © 2018 Ralph Brinks. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

Abstract

Chronic diseases impose a huge burden for mankind. Recently, a mathematical relation between the incidence and prevalence of a chronic disease in terms of a differential equation has been described. In this article, we study the characteristics of this differential equation. Furthermore, we prove the ill-posedness of a related inverse problem arising in chronic disease epidemiology. An example application for the inverse problem about type 1 diabetes in German women aged up to 35 years is given.

1. Introduction

Chronic diseases impose an enormous burden for mankind. It has been estimated that 71% of the 56 million global deaths in 2015 were attributable to noncommunicable diseases with an upward trend in the past decades [1]. Leading causes of death in 2015 were ischemic heart disease and cerebrovascular disease. Both are irreversible, i.e., chronic, conditions. Modelling of chronic conditions is often accomplished by compartment models like the illness-death model shown in Figure 1.

Figure 1: Illness-death model of a chronic disease. The rates , and describe the transitions between the states.

The illness-death model goes back at least until the 1950s [2] and consists of the three states: nondiseased, diseased, and dead [3, 4].

The prevalence of a chronic disease can be related to the transition rates in the illness-death model by a scalar partial differential equation [5]. Using the method of characteristics [6], this partial differential equation can be reduced to an ordinary differential equation (ODE), which pictures the temporal change of the prevalence along the characteristic lines. Thus, one time variable and a scalar ODE are sufficient to describe the change of the prevalence as a function of the transition rates in the illness-death model. Until now, a rigorous mathematical treatment of the ODE is missing.

The ODE has an important epidemiological application. Given the mortality rates ( in Figure 1) and the age-specific prevalence, the ODE may be used to derive the incidence rate (). Estimating the incidence rate of a disease is an important epidemiological problem [3, 7, 8]. This task can be interpreted as an inverse problem, which is often examined with respect to ill- or well-posedness. A well-posed inverse problem in the sense of Hadamard means that a solution for the inverse problem exists and that the solution is unique and continuous [9]. In this article, the ill-posedness of the inverse problem is proven, and an example application from the field of diabetes is given.

This article is organized as follows: Section 2 reviews the derivation of the ODE. Then, some properties of the ODE and its solution are examined. In Section 3, the inverse problem is introduced and the ill-posedness of the inverse problem is proven. To demonstrate the importance and applicability of the theory, an example for the inverse problem is given. Finally, in Section 4, the results and their consequences are discussed.

2. Derivation and Properties of the ODE

A popular framework for studying relations between prevalence and incidence of a chronic disease is the illness-death model shown in Figure 1. People in the population under consideration can contract the disease at the incidence rate , and they can die either with the disease at the age-specific mortality rate or without the disease at the mortality rate . The numbers of individuals in the nondiseased state and in the diseased state are denoted by (susceptibles) and (cases). Both numbers, and , are assumed to be sufficiently large, such that they can be considered as continuously differentiable functions. As described in Introduction, one time variable, the age , is sufficient to describe the temporal evolution of the population in the illness-death model.

We additionally assume that the population is closed, i.e., there is no migration. Furthermore, the age-specific functions are nonnegative and differentiable in for some with . Henceforth, is considered as age when all members (diseased and nondiseased) of the population are deceased. Then, the system of ODEs given by (1) and (2) describes the change rates of the numbers and of nondiseased and diseased individuals, respectively:

The system (1) and (2) is linear and of first order. Due to the simple structure of the ODEs for a given age-specific incidence and mortality rates and , the analytical solution of the corresponding initial value problem with initial conditions with can be calculated easily:

Obviously, from , it follows that , and for all .

Sometimes, it is important to consider the number of persons aged , who are alive, i.e., . For , we have the following initial value problem:with initial condition . From Equation (5), we find , where . This yields

For chronic diseases, it is reasonable to assume . Then, from Equation (6), we see that for all . We may conclude

Theorem 1. For nonnegative and differentiable functions , with for all , the system of ODEs given by equations (1) and (2) with initial conditions has a unique solution with for all .

From for all , we can infer

Definition 1. With the assumptions of Theorem 1, the function withis well defined for all . The function in Equation (7) is called the age-specific prevalence.

Next, we show that, with the assumptions of Theorem 1, the age-specific prevalence is epidemiologically meaningful.

Theorem 2. With the assumptions of Theorem 1, the age-specific prevalence defined by Equation (7) is bounded with

Proof. Let . If , then it is . For , we have with . This proves . From Equation (4), we see that , which implies .

Interestingly, the two-dimensional system (1) and (2) can be reduced to a scalar ODE.

Theorem 3. Let be nonnegative and differentiable functions with for all . Let be the unique solution of the system (1) and (2) with initial conditions . Then, is differentiable and the unique solution of the ODE iswith initial condition .

Proof. Essentially, this follows from the quotient rule applied to and inserting Equations (1) and (2).

Example 1. Figure 2 shows the slope field of an exemplary ODE (9) with the incidence rate chosen to be . The mortality rates and are assumed to be of Gompertz type:with the coefficients , as shown in Table 1.

Figure 2: Slope field and solution of an associated initial value problem (red) in Example 1.
Table 1: Coefficients of Gompertz mortality rates and .

The slope field in Figure 2 shows the solution of an associated initial value problem with initial condition (red line). The existence of a (local) maximum of the age-specific prevalence (here at an age of about 80 years) is typical for many chronic diseases, e.g., dementia [10], diabetes [11], or rheumatic diseases [12].

The ODE (9) is of Riccati type [13]. In epidemiological contexts, the mortality of the nondiseased people is usually unknown. Frequently, for a population under consideration, the overall mortality rate (general mortality) is known from vital statistics. The mortality rate is a convex combination of the mortality rate of the nondiseased people and the mortality rate of the diseased people:where is the relative risk, .

Apart from the incidence rate , two pieces of information about the mortality are necessary to solve the ODE (9). For instance, and are sufficient to determine the right-hand side of Equation (9). Depending on the type of information about the mortality (, or ) from the epidemiological context, the ODE (9) changes its type, which is important when solving the ODE. The possible combinations are shown in Table 2. In case the ODE is linear, an easy analytical solution exists. If the ODE is of Riccati or Abelian type [13], a general analytical solution does not exist. An extensive monograph about Riccati equations is in [14].

Table 2: Types of the ODE (9) depending on the given mortality.

The fractions and in the last two rows of Table 2 are well-known quantities in epidemiology. These are the exposition attributable fraction (EAF) and the population attributable fraction (PAF), respectively [15].

Remark 1. For the special case that for all —this case is called nondifferential mortality—the solution of Equation (9) with initial condition is

Starting from the system of ODEs (1) and (2), we have deduced the scalar ODE (9), which is well defined and epidemiologically meaningful. At the end of this section, we prove that the reverse way—from ODE (9) to the system (1) and (2)—is also possible. Before we can prove this, we need a slightly modified version of ODE (5).

Lemma 1. With the general mortality m defined in Equation (11), the ODE (5) can be reformulated intowith initial condition .

Then, from calculus, we can deduce the following correspondence between the system (1) and (2) and the two ODEs (9) and (13):

Theorem 4. (correspondence theorem). Let be nonnegative differentiable functions.(a)If are solutions of (1) and (2) with , then and are solutions of (9) and (13) with and . The general mortality in Equation (13) is defined by (11).(b)If are solutions of (9) and (13) with and initial conditions and , then and are solutions of (1) and (2) with initial conditions and .

Proof. Part (a) follows largely from Theorem 3 and the fact that . For part (b), we have to apply the product rule to and . An easy calculation yields the results about the initial conditions.

Remark 2. In [14], it has been shown that any solution of a Riccati ODE corresponds to a solution of a two-dimensional system of linear ODEs with . It can be shown that the functions and have the roles of and , respectively. Here, Theorem 4 has been tailored to the epidemiological context.

3. The Inverse Problem

A key application for the ODE (9) is the derivation of the age-specific incidence rate from the age-specific prevalence if the mortalities (or any equivalent information in the first column of Table 2) are known. In epidemiology, incidences rates are typically surveyed in follow-up studies, which may be lengthy and costly. However, Equation (9) allows a new way of estimating the incidence . Besides mortality information, the age-specific prevalence () has to be known, which can be obtained from relatively cheap cross-sectional studies. An example about type 1 diabetes is shown below.

In such an application with given mortalities, we conclude from an effect (the prevalence) the underlying cause (the incidence), which can be interpreted as an inverse problem [16]. The inverse problem is opposed to the direct problem of inferring from the incidence (i.e., the cause) the prevalence (the effect) by ODE (9).

3.1. Ill-Posedness of the Inverse Problem

We show that the inverse problem is ill-posed in the sense of Hadamard [9]. Let the mortalities and be continuous and nonnegative (in this section, denotes the set of all k-times continuously differentiable real-valued functions, ). For , define the operator , such that and is the solution of (9). To show that the inverse problem is ill-posed, we prove that is discontinuous. It is sufficient to show this for the special case of nondifferential mortality (). Let , be equipped with the norm . Choose , and define . Then, it is and

For and , the term is unbounded as , which implies that is discontinuous and the inverse problem is ill-posed.

3.2. Example: Incidence of Type 1 Diabetes

In this section, the inverse problem is solved in the context of type 1 diabetes. Type 1 diabetes is a chronic condition that mostly arises in the early decades of life. Figure 3 shows the age-specific prevalence () of type 1 diabetes in German women aged up to 35 years in the year 2010. Diagnoses stem from health insurance claims of 65 million people. Details about the data collection are given in [11]. The age-specific prevalence is steeply increasing from birth to the age of about 15 years. At age 20, a first plateau is reached, and then a second increase up to age 30 can be observed. At age 35 a second plateau is reached.

Figure 3: Age-specific prevalence of type 1 diabetes in German women aged 35 and below in 2010.

To estimate the age-specific incidence (), we use the following equation:where is the general mortality for the year 2010 and is the relative mortality risk. The general mortality is publicly available from the Federal Statistical Office of Germany. The population-wide relative mortality is unknown for Germany. Thus, we consider two extreme scenarios: and , and one realistic scenario: . Choosing the two extreme scenarios is done because the true (but unknown) age-specific incidence rate will be located between those rates derived from the two extreme scenarios (sandwich principle). The scenario assumes that women with type 1 diabetes undergo half the mortality rate of women without diabetes, which means that type 1 diabetes is a protective factor against death. The second scenario considers type 1 diabetes to be a strong risk factor for death and that women with type 1 diabetes undergo a fivefold mortality rate than those women without type 1 diabetes. Both scenarios are unrealistic, and the true relative mortality risk is certainly somewhere between these two extreme scenarios. Empirical data from countries with comparable health care systems indicate that is approximately 3 [17, 18]. In case of patients with severe late complications, reaches a value of about 4 [19].

Figure 4 shows the age-specific incidence rate of type 1 diabetes in German women in 2010. The different scenarios of the relative mortality () are indicated by different line types. There is no visual difference between all of the scenarios until age 20. For age 20+, the scenarios and are virtually indistinguishable. The scenario leads to a slightly elevated incidence rate compared to the scenario . At age 35, the difference is 0.43 per 100000.

Figure 4: Age-specific incidence of type 1 diabetes in German women in 2010. The different scenarios in the relative mortality are indicated by different line types.

All three scenarios unveil that the age-specific incidence of type 1 diabetes decreases from birth to the age of 20 years. Then, a second peak of the age-specific incidence rate occurs at the age of about 28 years.

4. Discussion

By extending the framework in [4] for studying the relation between prevalence and incidence, it had been found that prevalence, incidence, and mortality are linked by a one-dimensional ODE [20]. In this article, it has been shown that the solutions of this ODE are epidemiologically meaningful. Depending on the type of mortality information available, the ODE changes its type, which has implications about existence of general analytical solutions. In many epidemiologically relevant cases, an analytical solution does not exist, and numerical treatment has to be used instead.

An important application of the ODE is the derivation of age-specific incidence rates from the age distribution of the prevalence. This article shows that this question can be interpreted as an ill-posed inverse problem. The proof of the ill-posedness shows that an additive high-frequency distortion () of the prevalence may lead to an unbounded inaccuracy in the derived incidence. However, high-frequency distortions might be unlikely in real chronic diseases. Hence, the consequences in practical epidemiology are presumably small. In addition, it has recently been shown that, for the task of estimating incidence rates from prevalence data, methods based on ODEs, like the one presented in this article, are superior compared to other methods [21].

An application has been given from the context of type 1 diabetes in young women. The solution of the inverse problem showed a peak of the age-specific incidence rate at the age of about 28 years. Data about the incidence of type 1 diabetes in age 20+ are extremely scarce. Thus, the presented findings may gain insights into the epidemiology of type 1 diabetes.

In the discussed ODE model, several assumptions have been made. The ODE is valid only if incidence and mortality rates are independent of calendar time. In demography and epidemiology, this assumption is usually called time homogeneity. Due to changes in medical progress, hygiene, nutrition, and lifestyle, mortality does undergo secular trends. Thus, it is appropriate to formulate Equation (9) in terms of a partial differential equation (PDE) [22]. However, the associated PDE can be reduced to an ODE again by the method of characteristics [6]. By this, the presented results of this article remain valid in the case when time homogeneity does not hold.

Moreover, in real diseases, mortality of the diseased persons depends on the duration of the disease. An example is diabetes, where the relative mortality over the diabetes duration is U-shaped [23]. In this case, the theory of differential equations still can be applied if the mortality rate of the diseased is modified slightly [5].

Furthermore, the ODE (9) holds true only if the considered population is closed. In a recent article, we described the necessary changes to deal with immigration or emigration [10]. Thus, the results described here show a possible path of generalization to populations with migration.

A last note refers to the term “chronic disease.” In this article, chronic means irreversible, i.e., there is no way back from the diseased state to the nondiseased state. However, most of the results presented here remain true, if there is remission back to the state nondiseased. Then, the fundamental ODE (9) has an additional term that depends on the remission rate [20, 22].

Data Availability

Data presented in the example application have been obtained from the Deutsches Institut für Medizinische Dokumentation und Information (DIMDI). As a result of Germany’s strict regulations on data protection, data are only available in an anonymous and aggregated form (§5 Data Transparency Regulation). Eligible research institutes according to §303e Section 1 German Code of Social Law, Book V, can obtain the aggregated data from the DIMDI after application and positive approval. The author and the affiliated institutes are not allowed to provide data access (§8 of Terms of Use).

Disclosure

A preliminary analysis of these data was presented as a poster on the European Congress of Epidemiology in Maastricht, Netherlands, 2015. Neither the Hiller Foundation nor the German Diabetes Center had any role in the study design, data collection, analysis and interpretation, decision to publish, or preparation of the manuscript.

Conflicts of Interest

The author declares that there are no conflicts of interest with respect to any aspect of this work.

Acknowledgments

The author thanks the Deutsches Institut für Medizinische Dokumentation und Information (DIMDI) for providing the data in the application of the inverse problem. The Hiller Research Unit for Rheumatology is funded by the Hiller Foundation, Erkrath, Germany. The German Diabetes Center is funded by the German Federal Ministry of Health and the Ministry of Innovation, Science and Research of the State of North Rhine-Westphalia.

References

  1. GBD 2015 Mortality and Causes of Death Collaborators, “Global, regional, and national life expectancy, all-cause mortality, and cause-specific mortality for 249 causes of death, 1980‒2015: a systematic analysis for the Global Burden of Disease Study 2015,” The Lancet, vol. 388, no. 10053, pp. 1459–1544, 2016. View at Google Scholar
  2. E. Fix and J. Neyman, “A simple stochastic model of recovery, relapse, death and loss of patients,” Human Biology, vol. 3, no. 3, pp. 205–241, 1951. View at Google Scholar
  3. N. Keiding, “Age-specific incidence and prevalence: a statistical perspective,” Journal of the Royal Statistical Society. Series A, vol. 154, no. 3, pp. 371–412, 1991. View at Publisher · View at Google Scholar
  4. C. J. Murray and A. D. Lopez, “Quantifying disability: data, methods and results,” Bulletin of the World Health Organization, vol. 72, no. 3, pp. 481–494, 1994. View at Google Scholar
  5. R. Brinks and S. Landwehr, “A new relation between prevalence and incidence of a chronic disease,” Mathematical Medicine and Biology, vol. 32, no. 4, pp. 425–435, 2015. View at Publisher · View at Google Scholar · View at Scopus
  6. E. C. Zachmanoglou and D. W. Thoe, Introduction to Partial Differential Equations with Applications, Williams & Wilkins, Baltimore, MD, USA, 1976.
  7. N. Hens, M. Aerts, C. Faes et al., “Seventy-five years of estimating the force of infection from current status data,” Epidemiology and Infection, vol. 138, no. 6, pp. 802–812, 2010. View at Publisher · View at Google Scholar · View at Scopus
  8. H. Muench, “Derivation of rates from summation data by the catalytic curve,” Journal of the American Statistical Association, vol. 29, no. 185, pp. 25–38, 1934. View at Publisher · View at Google Scholar · View at Scopus
  9. J. Hadamard, Lectures on the Cauchy Problem in Linear Partial Differential Equations, Yale University Press, New York, NY, USA, 1923.
  10. R. Brinks and S. Landwehr, “Age- and time-dependent model of the prevalence of non-communicable diseases and application to dementia in Germany,” Theoretical Population Biology, vol. 92, pp. 62–68, 2014. View at Publisher · View at Google Scholar · View at Scopus
  11. T. Tamayo, R. Brinks, A. Hoyer, O. Kuss, and W. Rathmann, “The prevalence and incidence of diabetes in Germany: an analysis of statutory health insurance data on 65 million individuals from the years 2009 and 2010,” Deutsches Aerzteblatt Online, vol. 113, no. 11, pp. 177–182, 2016. View at Google Scholar
  12. R. Brinks, R. Fischer-Betz, O. Sander, J. Richter, G. Chehab, and M. Schneider, “Age-specific prevalence of diagnosed systemic lupus erythematosus in Germany 2002 and projection to 2030,” Lupus, vol. 23, no. 13, pp. 1407–1411, 2014. View at Publisher · View at Google Scholar · View at Scopus
  13. E. Kamke, Differentialgleichungen, Teubner Verlag, Stuttgart, Germany, 1983.
  14. W. T. Reid, Riccati Differential Equations, Academic Press, New York, NY, USA, 1972.
  15. C. H. Hennekens and J. E. Buring, Epidemiology in Medicine, Lippincot Williams & Wilkins, Philadelphia, PA, USA, 1987.
  16. A. Tarantola, Inverse Problem Theory, SIAM, Philadelphia, PA, USA, 2005.
  17. D. R. Wasag, J. W. Gregory, C. Dayan, and J. N. Harvey, “Excess all-cause mortality before age 30 in childhood onset type 1 diabetes: data from the Brecon Group Cohort in Wales,” Archives of Disease in Childhood, vol. 103, no. 1, pp. 44–48, 2018. View at Publisher · View at Google Scholar · View at Scopus
  18. D. Petrie, T. W. C. Lung, A. Rawshani et al., “Recent trends in life expectancy for people with type 1 diabetes in Sweden,” Diabetologia, vol. 59, no. 6, pp. 1167–1176, 2016. View at Publisher · View at Google Scholar · View at Scopus
  19. T. Heller, C. Kloos, T. Lehmann et al., “Mortality and its causes in a german cohort with diabetes mellitus type 1 after 20 years of follow-up: The JEVIN Trial,” Experimental and Clinical Endocrinology & Diabetes, vol. 126, no. 6, pp. 387–393, 2017. View at Publisher · View at Google Scholar · View at Scopus
  20. R. Brinks, S. Landwehr, A. Icks, M. Koch, and G. Giani, “Deriving age-specific incidence from prevalence with an ordinary differential equation,” Statistics in Medicine, vol. 32, no. 12, pp. 2070–2078, 2013. View at Publisher · View at Google Scholar · View at Scopus
  21. S. Landwehr and R. Brinks, “A comparative study of prevalence-based incidence estimation techniques with application to dementia data in Germany,” Statistics in Medicine, vol. 35, no. 5, pp. 768–781, 2016. View at Publisher · View at Google Scholar · View at Scopus
  22. R. Brinks and S. Landwehr, “Change rates and prevalence of a dichotomous variable: simulations and applications,” PLoS One, vol. 10, no. 3, Article ID e0118955, 2015. View at Publisher · View at Google Scholar · View at Scopus
  23. B. Carstensen, J. K. Kristensen, P. Ottosen, and K. Borch-Johnsen, “The Danish National Diabetes Register: trends in incidence, prevalence and mortality,” Diabetologia, vol. 51, no. 12, pp. 2187–2196, 2008. View at Publisher · View at Google Scholar · View at Scopus