This article has Open Peer Review reports available.

# Development and demonstration of a state model for the estimation of incidence of partly undetected chronic diseases

- Ralph Brinks
^{1}Email author, - Barbara H. Bardenheier
^{2}, - Annika Hoyer
^{1}, - Ji Lin
^{2}, - Sandra Landwehr
^{3}and - Edward W. Gregg
^{2}

**15**:98

https://doi.org/10.1186/s12874-015-0094-y

© Brinks et al. 2015

**Received: **9 August 2015

**Accepted: **6 November 2015

**Published: **11 November 2015

## Abstract

### Background

Estimation of incidence of the state of undiagnosed chronic disease provides a crucial missing link for the monitoring of chronic disease epidemics and determining the degree to which changes in prevalence are affected or biased by detection.

### Methods

We developed a four-part compartment model for undiagnosed cases of irreversible chronic diseases with a preclinical state that precedes the diagnosis. Applicability of the model is tested in a simulation study of a hypothetical chronic disease and using diabetes data from the *Health and Retirement Study (HRS)*.

### Results

A two dimensional system of partial differential equations forms the basis for estimating incidence of the undiagnosed and diagnosed disease states from the prevalence of the associated states. In the simulation study we reach very good agreement between the estimates and the true values. Application to the HRS data demonstrates practical relevance of the methods.

### Discussion

We have demonstrated the applicability of the modeling framework in a simulation study and in the analysis of the *Health and Retirement Study*. The model provides insight into the epidemiology of undiagnosed chronic diseases.

### Keywords

Compartment model Incidence Prevalence Diabetes Chronic disease Undiagnosed disease Case finding Screening Health and Retirement Study## Background

Most major causes of chronic morbidity and mortality, including diabetes, cancer, osteoporosis, cardiovascular disease, and dementia, pass through undiagnosed stages, at which clinically defined and recognized thresholds for a particular disease have been met, but diagnosis has not occurred due to either lack of awareness, symptoms, or access to care [1–3]. In the case of diabetes, population surveys have shown that 24 % to 75 % of prevalent cases across different countries and settings have not been diagnosed and the diagnosis lag has been estimated as ranging from three to seven years [4, 5]. With regard to dementia, it is estimated that more than a half of all patients are undiagnosed [6].

High proportions or long durations of undiagnosed chronic disease have several important clinical and epidemiological ramifications. First, the period prior to diagnosis may be a missed opportunity to implement effective preventive interventions in clinical settings [7, 8]. Second, the undiagnosed state creates problems for the accurate monitoring of population health and response to public health interventions [9]. In the United States, for example, trends in diabetes incidence at a national level are assessed using self-reports of diagnosed cases [10]; this means that the degree to which recent diabetes trends have been influenced by shifting awareness or detection of existing cases, as opposed to the rate of occurrence of new cases of disease, is unclear.

## Methods

Building upon previously published state models, for this study we develop a model including an *Undiagnosed* state (Fig. 1). The population of interest is partitioned into the four states *Normal* (i.e., healthy with respect to the chronic disease under consideration), *Undiagnosed, Diagnosed* (i.e., without and with a physician’s diagnosis), and *Dead*. The transition rates between the states are denoted as in the figure. The model described here is able to cope with secular trends, (i.e., involves calendar time *t*) and the different ages *a* of the subjects in the population, and thus these models are called *age-structured* [13].

The proportion of the living population in the states *Normal, Undiagnosed,* and *Diagnosed* are determined by their initial values and the rates *λ*
_{
ℓ
},*μ*
_{
k
}, *ℓ*=0,1,*k*=0,1,2. Let *N*
_{0},*N*
_{1}, and *N*
_{2} denote the numbers of persons in the respective state *Normal, Undiagnosed,* and *Diagnosed*. In addition, we set *N*(*t, a*):=*N*
_{0}(*t, a*)+*N*
_{1}(*t, a*)+*N*
_{2}(*t, a*). For (*t, a*) with *N*(*t, a*)>0 define the prevalences \(p_{k}(\textit {t, a}) := \frac {N_{k}(\textit {t, a})}{N(\textit {t, a})}, ~k = 0, 1, 2.\) For example, *N*
_{1}(*t, a*) denotes the number of persons aged *a* at time *t* with the disease, but without a diagnosis.

After deriving the governing equations for the state model in Fig. 1, we study an example of how the prevalences *p*
_{
k
}, *k*=0,1,2, evolve if the rates *λ*
_{
ℓ
}, *ℓ*=0,1, and *μ*
_{
k
}, *k*=0,1,2, are known. As we know the rates (i.e., the “causes”) and want to calculate the prevalences (i.e., the “effects”) we call this problem the *forward problem*.

Then, we examine whether the rates *λ*
_{
ℓ
}, *ℓ*=0,1, can be estimated if the prevalences *p*
_{
k
} and the mortality rates *μ*
_{
k
}, *k*=0,1,2, are known. We call this problem the *inverse problem*. The inverse problem is important in epidemiology, in which surveying the prevalences *p*
_{
k
} is much easier than surveying the transition rates *λ*
_{
ℓ
}. For surveying prevalences, cross-sectional studies suffice, whereas examining rates requires lengthy follow-up studies. We propose two approaches to solve the inverse problem.

After this, we describe and validate the methods in a simulation study and apply it to U.S. nationally representative data from the *Health and Retirement Study* (HRS). The HRS is a nationally representative longitudinal biannual survey of individuals 50 years of age and older in the United States. The survey is sponsored by the National Institute on Ageing and performed by the Institute for Social Research at the University of Michigan. The Health Sciences Institutional Review Board at the University of Michigan approved the HRS study design. The data used for this analysis contain no unique personal identifiers and are publicly available (after application). Permission to use the HRS data was obtained from the University of Michigan (Survey Research Center, 426 Thompson Street, Ann Arbor, MI 48104).

All calculations for this work have been performed with the statistical software R (The R Foundation for Statistical Computing). The scripts for usage in R are provided as an additional zip-file.

## Results

### The governing equations

*N*

_{0}(

*t, a*),

*N*

_{1}(

*t, a*) and

*N*

_{2}(

*t,a*) of healthy, undiagnosed, and diagnosed persons in terms of partial differential equations (PDEs), which can be derived from the disease model in Fig. 1. For the healthy persons, we get the following initial value problem of Cauchy type:

Here *S*
_{0}(*t*) is the number of healthy newborns at calendar time *t*. Note that, in this work, we just consider diseases contracted after birth. The notation *∂*
_{
x
} denotes the partial derivative with respect to *x*, *x*∈{*t, a*}.

*d*is also possible [12], hereinafter it is assumed that none of the rates depend on

*d*. Then, the numbers

*N*

_{1}and

*N*

_{2}of diseased persons without and with diagnosis, respectively, are described similarly:

### Prevalence, incidence and mortality

*p*

_{ k }instead of the absolute numbers

*N*

_{ k },

*k*=0,1,2. We expressed Eqs. (2) and (3) in terms of prevalences

*p*

_{1}and

*p*

_{2}. The prevalence

*p*

_{0}can be substituted by using the equation

*p*

_{0}=1−

*p*

_{1}−

*p*

_{2}. In addition, often the mortality

*μ*

_{0}is unknown and the overall mortality (general mortality)

*μ*is given. The overall mortality

*μ*in the population may be written as

Together with the initial conditions *p*
_{1}(*t*,0)=*p*
_{2}(*t*,0)=0 for all *t*, the system (4) - (5) completely describes the dynamics of the disease in the considered population. Note that the system (4) - (5) does not explicitly depend on the mortality of the healthy subjects *μ*
_{0}, which is typically unknown. The remaining rates are either accessible by (specially designed) epidemiological studies (*λ*
_{0},*λ*
_{1},*μ*
_{1},*μ*
_{2}) or by official vital statistics (*μ*).

### Relation to the conventional illness-death model

*Undiagnosed*and

*Diagnosed*to be pooled. If we define the prevalence

*p*as the pooled prevalence

*p*=

*p*

_{1}+

*p*

_{2}, the system (4) - (5) can be used to derive the following equation:

This equation has been proven in [14] for the conventional illness-death model. Thus, the system (4) - (5) is consistent with the conventional illness-death model if we pool the *Undiagnosed* and *Diagnosed* states together.

### Detection ratio

*λ*

_{0}and

*λ*

_{1}for the model in Fig. 1, we can calculate a measure that we call the age-specific

*detection ratio*

*DR*. For (

*t, a*) with

*λ*

_{0}(

*t, a*)>0 define

The detection ratio is a rate ratio. In the context of survival analysis such a ratio is called a *hazard ratio* [16]. For a point in time *t*, it describes the (instantaneous) probability of detecting an undiagnosed person of age *a* in relation to the (instantaneous) probability of a healthy subject aged *a* entering the *Undiagnosed* state.

*D*

*R*(

*t, a*) implies that

*p*

_{1}(

*t, a*) increases. More precisely: For (

*t, a*) let be

*λ*

_{0}(

*t, a*)>0 and

*p*

_{1}(

*t, a*)>0, then for small time intervals

*δ*>0, a detection ratio

*D*

*R*(

*t,a*) with

implies *p*
_{1}(*t*+*δ*,*a*+*δ*)>*p*
_{1}(*t, a*). Vice versa, a high detection ratio *D*
*R*(*t, a*)>*Δ*(*t,a*), implies *p*
_{1}(*t*+*δ*,*a*+*δ*)<*p*
_{1}(*t, a*). This follows from (*∂*
_{
t
}+*∂*
_{
a
})*p*
_{1}=0 for *D*
*R*=*Δ*, see Eq. (4).

### Simulation: forward problem

*t*

^{⋆}. At

*t*

^{⋆}the disease is detected and no longer unknown. This could happen by the discovery of a new pathogen or a novel diagnostic technique or by increased awareness, attention, or access to care. Henceforth, physicians start to look for the disease. As a consequence, after

*t*

^{⋆}the prevalence

*p*

_{1}of undetected cases decreases, whereas the prevalence

*p*

_{2}of detected cases increases. As an example, the general mortality

*μ*is chosen as the (approximated) general mortality of the German male population from 1900 to 2010. For the approximation of the mortality, we make the following approach:

with *β*
_{0}(*t*)=−7.078−0.02592 *t* and *β*
_{1}(*t*)=0.06401+2.455 10^{−4}
*t*. Calendar time *t* is counted in years since January 1st, 1900.

For simplicity, the mortality rates *μ*
_{
ℓ
}, *ℓ*=1,2, are assumed to be proportional to *μ*:*μ*
_{1}=3.5 *μ* and *μ*
_{2}=2.5 *μ*. The factor for *μ*
_{1} is chosen to be larger than the one for *μ*
_{2}, because in contrast to persons in the *Diagnosed* state, persons in the *Undiagnosed* state cannot be treated for the disease. The magnitude of the factors is motivated by dementia [17].

*λ*

_{0}(Table 1) is the 1.5-fold of the age-specific incidence rate of dementia in German males [18]. Dementia serves as a demonstration for an important chronic disease. However, as we are mixing data from different sources in different populations, the example is hypothetical and inferences about the disease itself should be drawn very carefully.

Age-specific incidence rates in the simulation

Age | Incidence | Incidence |
---|---|---|

(years) | (per 100 person-years) | (per 100 person-years) |

≤62.5 | 0 | 0 |

67.5 | 0.45 | 3.3 |

72.5 | 1.05 | 7.8 |

77.5 | 2.55 | 18.8 |

82.5 | 4.50 | 33.3 |

87.5 | 7.80 | 57.6 |

92.5 | 11.40 | 84.2 |

97.5 | 14.85 | 109.7 |

≥100 | 16.80 | 124.1 |

For year *t*=75, the rates *λ*
_{1} are also shown in Table 1. We assume a secular trend in *λ*
_{1}, mimicking increasing awareness of the hypothetical disease. In the simulation, *λ*
_{1} increases by 1 % per year for all ages *a*.

*t*=75 in both prevalences

*p*

_{1}and

*p*

_{2}is clearly visible in the upper right corner of the figures.

*t*=70 and

*t*=80 are additionally shown in Fig. 4. At

*t*=70, there are no diagnosed cases (the hypothetical disease is not detected yet). The prevalence of the undiagnosed cases (

*p*

_{1}) peaks at about 16 %, at the age of 91 years. Ten years later, the disease has been detected and the medical community is making diagnoses. Hence, the prevalence of the undiagnosed disease has decreased substantially - to less than 7 %. Especially in the higher age groups (≥85), physicians are aware of and detect a high proportion of cases and the prevalence of diagnosed cases (

*p*

_{2}) has increased.

*a*. It depends only on the calendar time

*t*. The time course of

*DR*is shown in Fig. 5. Before year 75, the detection ratio is 0. Later, the physicians start to diagnose the hypothetical disease at increasing rates.

*p*(=

*p*

_{1}+

*p*

_{2}) in year

*t*=70 differs substantially from the one at

*t*=80 (Fig. 6), which is an effect of the lowered mortality for those diseased persons whose condition has been detected. As the mortality

*μ*

_{2}is considerably lower than

*μ*

_{1}, the overall survival of the diseased persons is improved after

*t*=75 and the overall prevalence increases.

### Inverse problem

An important epidemiological application is the calculation (of some) of the rates in the model, if the prevalences *p*
_{
k
}, *k*=1,2, are known. A typical situation might be that the mortality rates are recorded in death registries (or other vital statistics) and two cross-sectional surveys are conducted to obtain the age-specific prevalences *p*
_{
k
}, *k*=1,2, at two points in time, *t*
_{1} and *t*
_{2}. The *inverse problem* is about whether the underlying rates *λ*
_{
ℓ
}, *ℓ*=0,1, can be *reconstructed* from the mortality and the prevalences. In the next two subsections we will present two ways for solving the inverse problem.

#### Direct solution of the inverse problem

*λ*

_{1}. For

*p*

_{1}>0 it holds:

*λ*

_{1}, Eq. (4) can be solved for

*λ*

_{0}. For 1−

*p*

_{1}−

*p*

_{2}>0 it is:

This is the *direct solution* of the inverse problem.

*p*

_{ k },

*k*=1,2, at two points in time

*t*

_{ j },

*j*=1,2, and the mortality rates

*μ*,

*μ*

_{1}, and

*μ*

_{2}are given at some time

*t*

^{′}with

*t*

_{1}<

*t*

^{′}<

*t*

_{2}. Then we can approximate

The symbol \(\doteq \) means that the partial derivative is approximated by its linearisation. Terms of quadratic or higher order in (*t*
_{2}−*t*
_{1}) are neglected.

*p*

_{1}and

*p*

_{2}in

*t*

_{1}=99 and

*t*

_{2}=101 by solving the forward problem, and then apply Eqs. (8) and (9) by using the approximation in Eq. (10) for

*t*

^{′}=100, we obtain the incidences

*λ*

_{0}and

*λ*

_{1}as shown Fig. 7. For comparison, the true incidences are shown as blue dashed lines. From visual inspection, the reconstructed incidences do not differ from the true incidences. Indeed, the maximum relative error in the age range

*a*=70,71,…,100 is 0.90 % for

*λ*

_{0}and 1.26 % for

*λ*

_{1}. The median relative errors are 0.13 % and 0.14 %.

#### Least squares solution

*p*

_{ k },

*k*=1,2, at two points in time

*t*

_{ j },

*j*=1,2, and the mortality rates

*μ*,

*μ*

_{1}, and

*μ*

_{2}at some time

*t*

^{′}with

*t*

_{1}<

*t*

^{′}<

*t*

_{2}. Typically,

*p*

_{ k },

*k*=1,2, are subject to sampling uncertainty. Let

*σ*

_{ k }denote the standard error of

*p*

_{ k }. For a moment let us assume that we know

*p*

_{ k }at

*t*

^{′}, and that we have a “guess” \(\lambda _{\ell }^{(g)}(t').\) Then, we can use the system (4) - (5) to approximate

*p*

_{ k }at

*t*

_{2}by

in which *h*
_{2}=*t*
_{2}−*t*
^{′}. The values of the partial derivatives (*∂*
_{
t
}+*∂*
_{
a
})*p*
_{
k
} are calculated by the right-hand sides of the associated Eqs. (4) and (5), respectively.

*p*

_{ k }at

*t*

_{1}:

in which *h*
_{1}=*t*
^{′}−*t*
_{1}.

*p*

_{ k }(

*t*

_{ j },

*a*). Define the sum of standardized squared error \(X^{2}(\lambda _{\ell }^{(g)})\) as

Hence, *λ*
_{
ℓ
} is the weighted least squares solution, which minimizes the squared deviation between the estimated and measured *p*
_{
k
} in *t*
_{1} and *t*
_{2}. Underlying the minimisation approach is the idea that the error \(p_{k}(t_{j}, a) - p_{k}(t_{j}, a | \lambda _{\ell }^{(g)})\) is approximately normally distributed with mean 0 and standard deviation *σ*
_{
k
}(*t*
_{
j
},*a*) [20].

*p*

_{ k }at

*t*

^{′}, which is not the case if we just have data from two cross-sections at

*t*

_{1}and

*t*

_{2}. In this case, we can estimate

*p*

_{ k }(

*t*

^{′},

*a*) by

*σ*

_{ k }=1. For solving the (constraint) minimisation problem, we use the R package nloptr [21]. The result is shown in Fig. 8. The reconstructed incidences visually do not differ from the true incidences. The maximum relative error in the age range

*a*=70,71,…,100 is 1.7 % for

*λ*

_{0}and 3.2 % for

*λ*

_{1}. The median relative errors are 0.29 % and 0.67 %.

Compared to the direct solution the median and maximum relative error increases, which is a consequence of the approximations (11) and (12). However, the least squares approach allows the inclusion of an error model and an estimation of the resulting uncertainty in the *λ*
_{
ℓ
} as shown in the next section.

### Example from the Health and Retirement Study

Estimates of diabetes prevalence and mortality were based on data from the 2006, 2008, and 2010 waves of the HRS.

Prevalence of undiagnosed diabetes was calculated from the random half sample of those participants aged 50 to 95 years, selected for the biomarkers blood tests in 2006 and 2008 [22] who had a valid HbA1c result (*n*=6300 and *n*=6115, respectively). Respondents who had a baseline HbA1c of >6.5 % and did not report a diabetes diagnosis were defined as having undiagnosed diabetes. Respondents in the 2006 sample (*n*=243) with undiagnosed diabetes were followed to 2008 to assess the risk of dying; similarly those in the 2008 sample (*n*=284) were followed to 2010. Due to the relatively low number of persons who died (*n*=19 and *n*=16), the mortality data of 2006 and 2008 have been pooled.

Diagnosed diabetes was identified if the respondent reported they had been told by a doctor that they had diabetes or high blood sugar [23]. Prevalence of diagnosed diabetes was based on 17,860 persons aged 50 to 95 years sampled in 2006 and 16,777 persons sampled in 2008. Respondents in the 2006 sample with diagnosed diabetes (*n*=3714) were followed to 2008 to assess mortality. During that time 408 subjects died. Similarly those in the 2008 sample (*n*=3768) were followed to 2010, with 503 death cases. For consistency reasons, we pooled the mortality data of both samples as the death cases in undiagnosed diabetes.

We used the survey sample in 2008 of participants aged 50 to 95 years who were alive, or who had died and had a proxy interview conducted by a family member or friend (*n*=17,970), to assess mortality (*n*=1173 died during the period 2008–10). To obtain mortality risk in the general population we ran a logistic model with death as the dependent variable and age and sex as the independent variables. Using the regression estimates (e.g., converting the odds to probabilities), we obtained the risk of mortality for each age, and then averaged every two years of age (i.e., 50–51, 52–53, …, 94–95).

*p*

_{1}, left part of Fig. 9) ranges from 3–4 %. Two years later this prevalence is about 5–6 %. Similarly, the prevalence of diagnosed diabetes (

*p*

_{2}) has increased for all age groups (right part of Fig. 9).

*X*

^{2}in Eq. (13) was augmented by the summand

in which *p*
_{
m
}(*a*) is the observed age-specific mortality risk with standard deviation *σ*
_{
m
}. The modeled mortality risk *p*
_{
m
}(*a*|*R*
^{(g)}) is assumed be to proportional to the mortality risk *π*(*a*) of the general population, with *R*
^{(g)} being the proportionality factor: *p*
_{
m
}(*a*|*R*
^{(g)})=*R*
^{(g)}
*π*(*a*).

To obtain estimates of the standard error of the *λ*
_{
ℓ
}, *ℓ*=0,1, we use a probabilistic sensitivity analysis [24]: 10,000 samples from the distributions of the input values are drawn and the associated least squares estimates for *λ*
_{
ℓ
}, *ℓ*=0,1, are calculated. This leads to an empirical estimate for the distribution of *λ*
_{
ℓ
}.

*λ*

_{ ℓ },

*ℓ*=0,1, for men and women. From the empirical standard deviations of the estimates it can be seen that the uncertainty in the estimates is rather high compared to the empirical mean. This is a result of the uncertainty in the input data, especially in the mortality of the persons with undiagnosed diabetes.

Age-specific incidence rates for men in HRS

Age | Incidence | Incidence | ||
---|---|---|---|---|

(years) | (per 100 person-years) | (per 100 person-years) | ||

Mean | SD | Mean | SD | |

54 | 2.05 | 1.12 | 54.18 | 30.11 |

62 | 2.22 | 1.25 | 24.62 | 16.18 |

70 | 3.13 | 1.21 | 35.50 | 16.48 |

78 | 1.87 | 1.40 | 30.55 | 18.45 |

86 | 0.25 | 0.64 | 36.66 | 28.70 |

Age-specific incidence rates for women in HRS

Age | Incidence | Incidence | ||
---|---|---|---|---|

(years) | (per 100 person-years) | (per 100 person-years) | ||

Mean | SD | Mean | SD | |

54 | 1.41 | 0.88 | 25.39 | 16.00 |

62 | 2.77 | 0.96 | 42.51 | 19.20 |

70 | 2.60 | 1.02 | 19.07 | 12.07 |

78 | 1.53 | 1.11 | 21.28 | 12.97 |

86 | 1.46 | 1.32 | 7.85 | 9.13 |

*DR*are estimated after performing a log-transformation. It holds that log(

*D*

*R*)= log

*λ*

_{1}− log

*λ*

_{0}. Thus, the variance of log(

*D*

*R*) can be estimated by the variances of log

*λ*

_{ ℓ },

*ℓ*=0,1, and the covariance. The corresponding estimates are shown in Table 4. We confine ourselves to reporting the estimates without interpreting them, because the uncertainty in estimated rates is too high to allow valid conclusions from these ratios.

Logarithms of the age- and sex-specific detection ratios in HRS

Age (years) | Detection ratio | |||
---|---|---|---|---|

Men | Women | |||

Mean | SD | Mean | SD | |

54 | 3.29 | 0.81 | 2.88 | 0.92 |

62 | 2.36 | 0.94 | 2.67 | 0.58 |

70 | 2.39 | 0.63 | 1.89 | 0.80 |

78 | 2.82 | 0.96 | 2.65 | 0.97 |

86 | 4.19 | 1.27 | 1.73 | 1.26 |

## Discussion

In this work, we used a state model to derive relations between incidence and mortality rates and the prevalence of undiagnosed and diagnosed persons in a chronic disease. The result is a two dimensional system of partial differential equations (PDEs) that forms a basis for estimating the incidence of the undiagnosed and diagnosed disease states from the prevalence of the associated states. In a simulation study and data from the *Health and Retirement Study* (HRS) we were able to show the accuracy and demonstrate the practical applicability of the method.

This method has several potential applications. First, it provides an approach to estimate the combined incidence for diabetes and similar conditions for which a large proportion of cases are undiagnosed and there is a lag in the identification of cases due to lack of symptoms, awareness, or health care access. Second, the methods also provide a way to determine the degree to which trends in incidence are biased by changing levels of detection by examining the ratio of diagnosed to undiagnosed incidence.

In an example, we have demonstrated the applicability of the modeling framework for a hypothetical chronic disease that has been discovered at a specific point in time, and has been diagnosed and treated thereafter. Apart from the hypothetical example, the analysis of the HRS data has proven applicability to real world data. Unfortunately, the uncertainty in the input data from HRS leads to relatively high uncertainty in the estimated incidence rates (Tables 2 and 3). As the HRS study has not been powered to accurately estimate the mortality of the study participants, the high uncertainty is a consequence of the study design. A more general analysis of how uncertainties and errors in the input data propagate into the results of the estimation, are subject of future work.

Our state model is an extension of the well-known illness-death model [15, 25], which has one additional state, *Undiagnosed*, which represents the subjects having contracted the disease but who are as yet undiagnosed. Using PDEs in the context of state models is not new [14, 26] and neither is taking into account undiagnosed diabetes [27]. However, the combination of both approaches is novel, and although our examples only considered non-communicable diseases, the model is potentially also applicable to some incurable infectious diseases, such as Hepatitis C or HIV, that have an asymptomatic preclinical phase [28].

The system of PDEs essentially has three advantages compared to other modelling techniques. First, the discretization errors using models with discrete time increments can be avoided. An example of these errors and the enormous impact they may have is demonstrated in [14], [Sect. 5]. Thus, the approach used here is more accurate than using discrete time models. The second advantage of PDEs lies in the fact that these equations are very well understood from the mathematical point of view. With very few assumptions on the smoothness of the right-hand side of the PDE, the existence and uniqueness of the solution is guaranteed [19]. Furthermore, there are a variety of freely available numerical routines to calculate the solution of PDEs. The third advantage is their flexibility: the new method may be applied to other chronic diseases as well, such as chronic kidney disease, osteoporosis, and cardiovascular disease. For each of these conditions, there exist modeling approaches including undiagnosed cases, which are important in health-economic models and screening [29, 30].

A drawback of the method is the restriction to *irreversible* chronic diseases. As shown in Fig. 1, it is assumed that there is no possibility to return to the *Normal* state. In case of diabetes, we know that this assumption is false, as bariatric surgery leads to remission in a large proportion of cases [31]. Even in the case of a modest intervention, like that received by the control condition of the *Look AHEAD Study*, 2 % had remission in the first year [32]. Thus, a small percentage of the incident cases may return to the *Normal* state [33, 34]. For diabetes, however, these cases are rare and have little impact on the population level that we are interested in. Other chronic diseases, like dementia do not have the possibility of remission at all. Thus, we consider the proposed methods useful for exploring how awareness for a disease and diagnostic possibilities may have an impact on the incidence of the disease. The full potential of the method is likely to become clear when time trends of the detection ratio of a chronic disease are studied.

In summary, we have developed a four-part compartment model with differential equations to estimate undiagnosed and diagnosed disease incidence and detection ratios for chronic diseases with common undiagnosed states. Future studies should validate our model using prospective, population-based studies, and surveillance systems.

## Conclusion

Based on the four-state compartment model we derived relations between the prevalences and the transition rates in terms of a system of partial differential equations. The partial differential equations provide insight into the epidemiology of undiagnosed chronic diseases. The applicability of the modeling framework has been demonstrated in a simulation study and in the analysis of the *Health and Retirement Study*.

## Declarations

### Acknowledgements

The authors have not received any funding with respect to any aspect of this work.

**Open Access** This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.

## Authors’ Affiliations

## References

- Beagley J, Guariguata L, Weil C, Motala AA. Global estimates of undiagnosed diabetes in adults. Diabetes Res Clin Pract. 2014; 103(2):150–60.View ArticlePubMedGoogle Scholar
- Lao C, Brown C, Obertová Z, Edlin R, Rouse P, Hodgson F, et al. The costs of identifying undiagnosed prostate cancer in asymptomatic men in New Zealand general practice. Family practice. 2013; 30(6):641–47.View ArticlePubMedGoogle Scholar
- Hodgson N, Gitlin LN, Winter L, Czekanski K. Undiagnosed illness and neuropsychiatric behaviors in community-residing older adults with dementia. Alzheimer Dis Assoc Disord. 2011; 25(2):109.PubMed CentralView ArticlePubMedGoogle Scholar
- Harris M, Klein R, Welborn T, Knuiman M. Onset of NIDDM occurs at least 4–7 yr before clinical diagnosis. Diabetes Care. 1992; 15(7):815–9.View ArticlePubMedGoogle Scholar
- Porta M, Curletto G, Cipullo D, de la Longrais RR, Trento M, Passera P, et al. Estimating the delay between onset and diagnosis of type 2 diabetes from the time course of retinopathy prevalence. Diabetes Care. 2014; 37(6):1668–74.View ArticlePubMedGoogle Scholar
- Solomon PR, Murphy CA. Should we screen for Alzheimer’s disease? A review of the evidence for and against screening Alzheimer’s disease in primary care practice. Geriatrics. 2005; 60(11):26–31.PubMedGoogle Scholar
- Kahn R, Alperin P, Eddy D, Borch-Johnsen K, Buse J, Feigelman J, et al. Age at initiation and frequency of screening to detect type 2 diabetes: a cost-effectiveness analysis. Lancet. 2010; 375:1365–74.View ArticlePubMedGoogle Scholar
- Black J, Sharp S, Wareham N, Sandbaek A, Rutten G, Lauritzen T, et al. Does early intensive multifactorial therapy reduce modelled cardiovascular risk in individuals with screen-detected diabetes? Results from the ADDITION-Europe cluster randomized trial. Diabet Med. 2014; 31(6):647–56.PubMed CentralView ArticlePubMedGoogle Scholar
- CDC: National Diabetes Statistics Report: estimates of diabetes and its burden in the United States. Atlanta GA: US Department of Health and Human Services; 2014. www.cdc.gov/diabetes/pubs/statsreport14/national-diabetes-report-web.pdf. Accessed 10 Nov 2015.
- Geiss LS, Wang J, Cheng YJ, Thompson TJ, Barker L, Li Y, et al.Prevalence and incidence trends for diagnosed diabetes among adults aged 20 to 79 years, United States, 1980-2012. JAMA. 2014; 312(12):1218–26.View ArticlePubMedGoogle Scholar
- Brinks R, Landwehr S, Icks A, Koch M, Giani G. Deriving age-specific incidence from prevalence with an ordinary differential equation. Stat Med. 2013; 32(12):2070–78.View ArticlePubMedGoogle Scholar
- Brinks R, Landwehr S. A new relation between prevalence and incidence of a chronic disease. Math Med Biol. 2015. doi:10.1093/imammb/dqu024.
- Schenzle D. An age-structured model of pre-and post-vaccination measles transmission. Math Med Biol. 1984; 1(2):169–91.View ArticleGoogle Scholar
- Brinks R, Landwehr S. Age-and time-dependent model of the prevalence of non-communicable diseases and application to dementia in Germany. Theor Popul Biol. 2014; 92:62–8.View ArticlePubMedGoogle Scholar
- Kalbfleisch J, Prentice R. The Statistical Analysis of Failure Time Data, 2nd edn. Hoboken: John Wiley & Sons; 2002.View ArticleGoogle Scholar
- Woodward M. Epidemiology: Study Design and Data Analysis. Texts in statistical science. Boca Raton: Chapman & Hall/CRC; 2005.Google Scholar
- Rait G, Walters K, Bottomley C, Petersen I, Iliffe S, Nazareth I. Survival of people with clinical diagnosis of dementia in primary care: cohort study. BMJ. 2010; 341:3584. doi:10.1136/bmj.c3584.View ArticleGoogle Scholar
- Ziegler U, Doblhammer G. Prävalenz und Inzidenz von Demenz in Deutschland–Eine Studie auf Basis von Daten der gesetzlichen Krankenversicherungen von 2002. Gesundheitswesen. 2009; 71(5):281–90.View ArticlePubMedGoogle Scholar
- Polyanin AD, Zaitsev VF, Moussiaux A. Handbook of First-Order Partial Differential Equations. Boca Raton: CRC Press; 2001.Google Scholar
- Draper NR, Smith H. Applied Regression Analysis. Wiley Series in Probability and Statistics. Hoboken: Wiley; 2014.Google Scholar
- Ypma J, Borchers HW, Eddelbuettel D. R Package nloptr. 2014. http://cran.r-project.org/web/packages/nloptr/nloptr.pdf. Accessed 10 Nov 2015.
- Crimmins E, Faul J, Kim J, Guyer H, Langa K, Ofstedal M, et al. Documentation of biomarkers in the 2006 and 2008 Health and Retirement Study. Ann Arbor, MI: Survey Research Center University of Michigan; 2013. hrsonline.isr.umich.edu/sitedocs/userg/Biomarker2006and2008.pdf. Accessed 10 Nov 2015.Google Scholar
- Clair PS, Bugliari D, Campbell N, Chien S, Hayden O, Hurd M, et al.RAND HRS data documentation - Version L. Technical report, Labor & Population Program, RAND Center for the Study of Aging. 2011.Google Scholar
- Oakley JE, O’Hagan A. Probabilistic sensitivity analysis of complex models: a Bayesian approach. J R Stat Soc Ser B (Stat Methodol). 2004; 66(3):751–69.View ArticleGoogle Scholar
- Keiding N. Age-specific incidence and prevalence: a statistical perspective. J R Stat Soc A. 1991; 154:371–412.View ArticleGoogle Scholar
- Brunet RC, Struchiner CJ. A non-parametric method for the reconstruction of age-and time-dependent incidence from the prevalence data of irreversible diseases with differential mortality. Theor Popul Biol. 1999; 56(1):76–90.View ArticlePubMedGoogle Scholar
- Boyle JP, Thompson TJ, Gregg EW, Barker LE, Williamson DF. Projection of the year 2050 burden of diabetes in the US adult population: dynamic modeling of incidence, mortality, and prediabetes prevalence. Popul Health Metrics. 2010; 8(1):29.View ArticleGoogle Scholar
- Hamers F, Phillips A. Diagnosed and undiagnosed HIV-infected populations in Europe. HIV Med. 2008; 9(s2):6–12.View ArticlePubMedGoogle Scholar
- Najafzadeh M, Marra CA, Lynd LD, Sadatsafavi M, FitzGerald JM, McManus B, Sin D. Future impact of various interventions on the burden of COPD in Canada: a dynamic population model. PloS one. 2012; 7(10):46746.View ArticleGoogle Scholar
- Manns B, Hemmelgarn B, Tonelli M, Au F, Chiasson TC, Dong J, et al.Population based screening for chronic kidney disease: cost effectiveness study. BMJ. 2010; 341:5869. doi:10.1136/bmj.c5869.View ArticleGoogle Scholar
- Raffaelli M, Sessa L, Mingrone G, Bellantone R. Assessing the obese diabetic patient for bariatric surgery: which candidate do i choose?Diabetes, metabolic syndrome and obesity: targets and therapy. 2015; 8:255.View ArticleGoogle Scholar
- Gregg E, Chen H, Wagenknecht L, Clark J, Delahanty L, Bantle J, et al.Association of an intensive lifestyle intervention with remission of type 2 diabetes. JAMA. 2012; 308(23):2489–96.View ArticlePubMedGoogle Scholar
- Pournaras DJ, Osborne A, Hawkins SC, Vincent RP, Mahon D, Ewings P, et al.Remission of type 2 diabetes after gastric bypass and banding: mechanisms and 2 year outcomes. Ann Surg. 2010; 252(6):966–71.View ArticlePubMedGoogle Scholar
- Ardestani A, Rhoads D, Tavakkoli A. Insulin cessation and diabetes remission after bariatric surgery in insulin-treated type 2 diabetic adults. Diabetes care. 2015; 38(4):659–64.PubMedGoogle Scholar