Study setting
In Finland, patients with T2D are typically treated and monitored in primary health care, where over 300 municipalities (local authorities) are responsible for the provision of these services to their residents. Health centres commonly have General Practitioner (GP)-run inpatient units, largely for chronic and long-term care patients. In addition, these Finnish municipalities form 20 hospital districts to provide hospital care as secondary and tertiary settings. Secondary care (including specialised outpatient care, inpatient care, and day surgery) is mainly provided by hospitals and tertiary care is delivered in five university hospitals. Patients need a referral to access specialist care, except for emergency cases. The management and follow-up of T2D patient population with micro- and macrovascular complications takes place in hospital setting in Finland.
The present study was conducted in Kuopio University Hospital (KUH), which provides the secondary health care services for around 250 000 citizens in its catchment area. In addition, it provides highly specialised tertiary health care services for around 810 000 inhabitants in Eastern and Central Finland, which is around 15% of all Finnish citizens (N = 5.5 million).
Study population and data linking
In the present study, people with diabetes were identified from the comprehensive national FinDM diabetes database [5]. The national FinDM database includes all Finnish patients with diabetes between years 1964–2017 and their linked health care data from several national registers. For the purposes of the present study, the T2D population having a home municipality code (any time during 2012–2017) within the catchment area of KUH (the Hospital District of Northern Savo) or any treatment at KUH during 2012–2017 were identified from the FinDM database following its definition for T2D [5]. For these patients, detailed data on patient-level events (of which only minor part were directly available from the FinDM database) including inpatient days, surgical procedures, ED visits, physician and nurse outpatient visits, laboratory examinations, diagnoses, medications (utilized during hospitalizations) as well as other accountable services and related detailed hospital claim data were gathered from KUH’s information systems for the period 2012–2018 by using unique national personal identity codes.
From this population, prevalent cases were defined as patients with T2D living in KUH’s catchment area at the final day of December 2011, whereas incident cases were defined as patients with newly diagnosed T2D between years 2012–2017. People living outside the KUH’s catchment area were excluded from the analyses. There were totally 35 292 individuals to whom the linkage to the KUH data was conducted. Of these, 27 255 were finally included to the actual study cohort (Fig. 1).
Next, the complications of these patients were screened during the years 1996–2017 using the same national quality register definitions as in FinDM [5], i.e., by identifying the given ICD-10 diagnoses or operation codes from the hospital admissions for the cerebrovascular, neurological, foot, nephropathy, eye, and cardiovascular complications (Supplement Table 3). The first diagnosis was selected as an index event of complication and all repeated diagnoses of the same complication class within a one year since the index event were considered to belong to the same “episode”, i.e. the diagnosis was considered as a new complication event if it occurred at least one-year since the previous complication event. In other words, the complication “episode” corresponds to the one-year period since the index event fulfilling the criteria of a complication.
Look back-period of ≥ 15 years allowed to further classify the episodes to prediabetic, incident or recurrent episodes depending on the timing of the episode, and diabetic complication episodes with a start date at the KUH during the years 2012–2016 were included in the main analyses to guarantee complete data for the one-year follow-up period (Fig. 1). Comorbid diseases were detected for each episode using disease classes based on the Charlson comorbidity classes [6] by screening for the hospital diagnoses within two years preceding the episode. Due to small number of cases, classes of HIV/AIDS and hemiplegia/paraplegia were dropped from the analyses, and liver diseases as well as malignancies being originally categories with two classes were reduced to any chronic liver disease and any malignancy. In addition, diabetes without complications was dropped as all patients had T2D. Diabetes with complications was kept in the analyses as its definition was based on the recent type of T2D diagnosis and not on the actual diagnoses of complication disorders.
Resource use and unit costs
Resource use and costs were considered from the health care payer perspective excluding patients’ out-of-pocket costs, direct non-health care costs such as travelling, and productivity loss costs due to T2D-related complications. Thus, health care resource use and costs considered in this study included all inpatient days, surgical procedures, emergency department visits, physician and nurse outpatient visits, laboratory examinations, diagnoses, medications (utilized during hospitalizations) as well as other accountable services related the treatment and monitoring of patients with T2D complications.
To evaluate costs at the complication level from the payer’s perspective, the costs taking place at more general resource related aggregate-level (e.g. salaries of the staff, purchases of equipment and drugs needed for the treatment, maintenance costs of operating rooms and other required facilities etc.) must be assigned to actual outputs, i.e. to the treatment of patients at the hospital. The standard way is to assign the costs to the observed components of the given treatment using suitable weights determined using e.g. diagnosis related group (DRG) or per diem costs. In the current study, the unit and total costs were obtained from the detailed patient-specific municipality billing data linked to the study dataset from KUH’s financial administration system. The KUH’s financial administration data provides an exceptionally detailed classification of the components of the treatment that allows comprehensive “reconstruction” of the actual treatment costs. Those costs are linked to the primary admission (inpatient period or outpatient visit) and the calculated cost is the price that is accounted from the payer of the treatment (a municipality of the patient).
As part of the treatment during the follow-up period have been given in the other (local) hospitals within the hospital district, costs for other hospital care than given in KUH were calculated using the CRHC data containing reported costs for part of the service providers, and for the rest of the costs of in- and outpatient admissions observed in the CRHC data were approximated by weighting the available treatment data (DRG-group or specialty-specific treatment days) using the published unit prices of health care [7].
Finally, we had component specific cost data for each admission (costs time stamped to accounting/discharge date). All costs were deflated to the price level of 2019 using a price index for health care. We used those to derive the cost outcomes of which the one-year costs after the complication event was the primary one. One-year costs are widely used in health economics and capture the burden of costs associated with diabetic complications in a way that also allows comparisons to any other cost-estimates of one-year hospital costs. As the KUH is a public non-profit university hospital with accurate costing system operating in the Finnish welfare state providing universal access to care, it is likely that the costs mapped to the components reflect the “real” costs of the treatment in a generalizable way – at least in the sense that costs within the hospital are on the same scale so that the relative differences are reliably reflecting the real differences of costs within hospital.
Statistical analysis
The statistical analysis unit was a complication event and there were six main classes of complications. For each complication event, three observation time periods were used: 1) the first year since the onset of complication, 2) the year preceding the complication, and 3) the second year since the onset of complication. The later time period was expected to reflect the ongoing long-term costs of complications, such as rehabilitation and monitoring, but also subsequent events of the same type [8, 9]. The complication events were further divided to first (incident) ones and to recurrent ones so that also the potential effect of repeated complications could be detected. Costs were stratified by detailed hospital resource categories. Cross tabulations, means and standard deviations as well as visualisations using the bar charts and radar plots were used for the descriptive analyses.
In addition, multivariable gamma regression models with log link functions taking an account to the positive skewness of cost data were applied to study the effects of patient characteristics on one-year costs and to generate covariate-adjusted mean estimates of one-year costs. For the incident complications, studied variables included sex, age, duration of T2D, and several comorbidities. In addition, death during the period of one year after the complication was included as a variable as that is known to modify the costs of treatment [10]. For the recurrent complications two additional variables were used: the number of earlier episodes of the same type and an indicator for exceptionally high (> 50 000€) one-year cost of the previous complication episode. These high costs represent a static construction of a variable that would adequately capture 10% of highest costs by using the 90% percentile as a cut point. This cut point approximately corresponded to the limit of 2 SD giving a further justification for the cut point. Regression models were estimated for each incident and recurrent complication type independently (i.e., 12 models in total). Sandwich estimator R package [11, 12] was used to consider the potential effect of clustering, due to the possible multiple recurrent complications of the same type on a single patient. In addition, average marginal effects of regression coefficients were estimated using Margins R package [13].
To dynamically estimate cost distributions for the patients with varying background characteristics, a Cholesky decomposition approach based on regression coefficients and covariance matrices of the models was used to simulate a large amount of possible predictions of the model for the populations with wanted background characteristics. In addition, freely accessible website application was developed for the further analysis of complication costs and their modifiers (https://uef-phoru.shinyapps.io/T2DCost/). The developed website application fits, in real time, customised regression models based on user-selected subset of predictors to predict the hospital treatment costs of different T2D-related complications.
Software package R version 4.0.2 was used for data processing and statistical analysis. Dynamic simulations utilize the Shiny application to run interactive R code.
