Effectiveness of integrated care model for type 2 diabetes: A population-based study in Reggio Emilia (Italy)

Aims To compare the effectiveness of integrated care with that of the diabetes clinic care model in terms of mortality and hospitalisation of type 2 diabetes patients with low risk of complications. Methods Out of 27234 people with type 2 diabetes residing in the province of Reggio Emilia on 31/12/2011, 3071 were included in this cohort study as eligible for integrated care (i.e., low risk of complications) and cared for with the same care model for at least two years. These patients were followed up from 2012 to 2016, for all-cause and diabetes-related mortality and hospital admissions. We performed a Poisson regression model, using the proportion of eligible patients included in the integrated care model for each general practitioner as an instrumental variable. Results 1700 patients were cared for by integrated care and 1371 by diabetes clinics. Mortality rate ratios were 0.83 (95%CI 0.60–1.13) and 0.95 (95%CI 0.54–1.70) for all-cause and cardiovascular mortality, respectively, and incidence rate ratios were 0.90 (95%CI 0.76–1.06) and 0.91 (95%CI 0.69–1.20) for all-cause and cardiovascular disease hospitalisation, respectively. Conclusion For low risk patients with type 2 diabetes, the integrated care model involving both general practitioner and diabetes clinic professionals showed similar mortality and hospitalisation as a model with higher use of specialized care in an exclusively diabetes clinic setting.


Introduction
The prevalence of type 2 diabetes (T2D) is increasing worldwide [1].In Italy, it has reached 6% of the adult population [2,3].T2D increases the risk of many severe diseases and of death [4].Control of glycaemia and early identification of possible complications of diabetes are effective ways of reducing the burden of disease [5].In Italy, normal treatment of T2D requires a referral by a general practitioner (GP) to a diabetes clinic (DC) for diagnostic confirmation, therapy, prevention and early diagnosis of complications through close patient follow up by a team of diabetologists, nurses and dieticians, and for scheduling of regular check-ups.This model of care is quite intensive and resource-consuming, and many diabetes clinics cannot take care of all the patients in their areas.Since 2006, a new model of care called integrated care (IC) was introduced, a shared care model involving both GPs and DC professionals [6].The IC model targets low risk T2D patients, i.e., those with good glycaemic control who do not use rapid-acting insulin and are not suffering from mild or severe complications.The care plan envisages visit every two years at the DC, as well as quarterly check-ups performed by a GP.Participation in IC is voluntary for both patients and physicians.The IC model is consistent with the WHO recommendations regarding the management of chronic diseases [7].
The effectiveness of a shared care health service has been debated in a recent review, which found that shared care probably has limited or no effect on physical clinical outcomes, except for a tendency towards improved blood pressure, and little or no difference in hospital admission rates [8].The review included studies focusing on diabetes [9][10][11][12][13][14][15][16][17].
Although randomized trials are the gold standard for comparing treatments or interventions, comparative effectiveness research has emphasised the use of observational studies to investigate treatment or intervention choices within larger and more representative populations [18].However, strong biases can affect the results of observational studies where the allocation to one treatment or another is based on the physician's assessment of prognostic factors [19,20].
In the comparison between the DC and IC models for T2D, two limitations can lead to biases in the results.First, the eligibility criteria for the new model of care (i.e., IC) define a very low risk group of patients compared to those in the DC model.This selection bias could be partially overcome by limiting the comparison only to those T2D patients who meet the IC criteria.The second bias regards confounding by indication bias (or treatment selection bias); physician and patient preferences and opportunities as well as patient frailty, lack of motivation or low self-efficacy, and/ or social or economic deprivation could be related to both treatment selection and outcome.As a result, residual confounding is expected when traditional methods of adjustment (regression or propensity score analysis) are used to compare the treatment effects [21].One way to overcome this is to use the instrumental variable approach [22].Similar to randomisation, an instrumental (or endogenous) variable is related to the treatment selection but is not directly related to the outcome.Its occurrence creates a natural experiment, i.e., a situation that allows for random or seemingly random assignment and can overcome the effect of unmeasured confounders.
In the province of Reggio Emilia (northern Italy), the IC model has been implemented gradually; while GP participation in some districts has increased rapidly, very few GPs in other districts initially agreed to participate.Furthermore, during the study period, each GP had a different level of engagement with IC.Therefore, the percentage of eligible patients who were cared for by the IC model varied from 0 to 100%, depending on the GP.Thus, the gradual implementation over time and across GPs has created a natural experiment that can be used to examine the comparative effectiveness of alternative care models.Moreover, the Reggio Emilia diabetes register (REDR) [2] permits conducting observational studies on a cohort of prevalent T2D patients to investigate differences in outcome measures.
This study aims to compare the effectiveness of IC with that of the DC care model in terms of mortality and hospitalisation rates of T2D patients with low risk of complications over a five-year follow-up period.The hypothesis is that IC is not inferior to the DC care model in controlling diabetes-related morbidity, while it should lead to lower resource consumption and be less time consuming for patients.

Study population and study design
The province of Reggio Emilia is situated in northern Italy and has approximatively 530,000 inhabitants, 300 general practitioners, 6 outpatient diabetes clinics and 1 diabetes inpatient unit in the main hospital.
This cohort study included people residing in the province of Reggio Emilia on December 31, 2011, diagnosed with T2D, under the age of 85 and cared for by the same care model for at least two years.Accordingly, three groups were created: 1) cared for by diabetes outpatient clinics (DC); 2) integrated care (IC); 3) Other-group (neither DC nor IC, maybe only cared for by own GP but also voluntary opt-outs who turn to private care or neglected patients).
Patients were followed up from 1 January 2012 to 31 December 2016, for all-cause and diabetes-related mortality and hospital admissions.
T2D status was retrieved from the Reggio Emilia diabetes register (REDR).
REDR is a validated register created by the deterministic linkage of six routinely collected data sources through a definite algorithm able to ascertain cases and to distinguish the type of diabetes and model of care [2].The date of inclusion in the register is the date when a person first meets one of the following inclusion criteria: (1) Disease-specific exemption database: exemption from co-payments due to diabetes; (2) Hospital discharge database: hospitalisation with diabetes diagnosis in whichever position by ICD-9 (International Classification of Diseases Clinical Modification, 9th Edition) codes 250.xx, 357.2x, 362.0x, 366.41, or 648.0x, excluding MDC14; (3) Biochemistry laboratory database: two glycated haemoglobin (HbA1c) tests > = 6.5% (48 mmol/mol) or one HbA1c> = 6.5% and (48 mmol/mol) test followed by a fasting blood sugar test> = 126 mg/dl; (4) Drug prescription databases: either at least two redeemed prescriptions for anti-diabetic drugs in a pharmacy or one prescription directly distributed by the hospital; (5) Diabetes outpatient clinics database: diagnosis by a diabetologist; (6) Mortality registry: cause of death by ICD-10 (International Classification of Diseases, 10th Edition) codes E10 -E14.Women with gestational diabetes or women receiving treatment for polycystic ovarian syndrome or obesity are excluded from the register.Cases initially notified to the registry through record linkage are retained if they are clinically confirmed by a diabetologist or another physician.
Exposure definition was assessed in the 2009-2011 period (see S1 Fig for study timeline chart).Patients were retained in the study and assigned to IC or DC care if they remained in the same care model for at least two years before the start of follow up.Shifts from one care model to the other after the start of follow up did not modify the exposure allocation, because the worsening of the disease would imply shifting from IC to DC.Thus, most of the investigated outcomes tended to occur in DC even if the worsening began during IC.
T2D patients were classified as eligible for IC according to regional guidelines criteria [6].The criteria were assessed in the period before the start of follow up: 1) no hospital admissions for diabetes-related complications in the previous two years, i.e., from 2009 to 2010 (see S1 Table for list of first diagnosis codes considered); 2) no mild or severe renal complications (i.e, those with the last glomerular filtrate before start of follow up> = 60 ml/min/1.73m 2 , considering 2010-2011 measurements); 3) HbA1c on target (last value before start of follow up < = 7% (53 mmol/mol) if aged less than 75 or < = 8% (64 mmol/mol) if aged equal to or more than 75, considering 2010-2011 measurements); 4) no use of rapid-acting insulin in 2011.Other covariates were assessed in 2011.

Statistical analysis
We compared patients' baseline characteristics stratified by current care model using the χ2 or Student t test, as appropriate.After classifying GPs according to the proportion of their patients eligible for IC model and those who were actually included in IC, we attributed this variable to each eligible patient, and we divided eligible patient population into quartiles.Eligible patients' baseline characteristics are reported according to quartile distribution to show the comparability of the population included in the 1 st quartile (i.e., assigned to GPs with lower adoption of IC) and the 4 th quartile (i.e., assigned to GPs with higher adoption of IC).
To measure the current differences in the care models, we calculated crude and standardised rates (by sex and age) and 95% confidence intervals (95%CI) for mortality and hospital admission outcomes, the latter considering all discharges occurring for the same patient in the follow-up period.
The conceptual non-inferiority hypothesis of IC compared to the DC care model was assessed by calculating and comparing RRs and related 95%CI through multivariate models in order to reduce the effect of a potential bias.First, we used a Poisson regression model, calculating mortality and hospitalisation rate ratios (RRs) and 95%CI only among IC-eligible patients cared for by DC or IC models.Persons-time was calculated as the difference between the start of the follow-up period (January 1, 2012) and the date of the event (death or first hospitalisation), moving out of the province or end of follow up (December 31, 2016), whichever came first.The DC care model was used as a reference and the covariates were sex, citizenship, age, years since diagnosis, the last two as continuous variables.Clinical conditions at baseline (HbA1c, triglycerides, BMI, GFR) were not included in the multivariate analysis because the values were balanced between groups.The only possible confounder was LDL, but the high proportion of missing values made it impossible to use to adjust the models.In this analysis we excluded the Other-group (only the number of events was reported) because it was not a care model foreseen by the guideline, but a residual group, the consequence of the incomplete application of either of the other two care models.Relative risks were not calculated for diabetes-related causes for which the number of expected events in at least one of the two care models was less than five.
Secondly, we performed a Poisson regression model, using our instrumental variables (i.e., the proportion of GP-eligible patients actually included in the IC model) as a main predictor.This variable was attributed individually to all eligible patients, and the model took into account that observations within the same GP's patient list were not independent (STATA command Poisson, cluster option).The other covariates were the same as the above multivariate model.
The analyses were performed using the STATA statistical package Version 13.0

Ethical approval
This is an observational study and the data were collected retrospectively.

Results
On December 31, 2011, the REDR included 27,234 prevalent patients with T2D (5.3% out of the total resident population, Fig 1  1).The percentage of women was greater the IC model, and the percentage of foreigners was smaller.As expected, the percentage of insulin users among IC patients was very low.
Looking at the characteristics of patients included in the Other-group, the most important result coming from the data was the low serum test rates compared to the other two groups, mainly from the HbA1c test, considered a primary test to check glucose control in T2D.BMI was not available for Other-group patients because the information was retrieved from the database of diabetes clinics.Distribution by care model and district of residence was very different, reflecting the different levels of implementation achieved by the IC model in 2009-11.Among the study population, one in four was unclassifiable with respect to IC eligibility criteria, there were 3,812 patients eligible for IC (21.8% of the total study population).Only slightly more than one third of the patients in IC were eligible for this care model, about one half were not eligible and the rest were unclassifiable.A median of 58 T2D patients were in the care of each GP and the interquartile range (IQR) was 47-70, while the median number of eligible patients was 12 and IQR 9-17, and the median number of unclassifiable patients was 13 and IQR 9-17.
When we compare the quartile with the lowest proportion of eligible patients actually included in IC and the highest quartile (Table 2), we go from <20% to >65% of eligible patients cared for in the IC model, and 78% of eligible patients actually included in IC are in the highest quartile.Distributions across quartiles of the proportion of IC-eligible patients actually included in IC by GPs was balanced in terms of age, time since diagnosis, HbA1c, BMI and triglycerides means, while the percentage of females decreased from the lowest to highest quartile (48.5% to 41.1%), as did the percentage of foreigners (3.1% to 2.3%).On the contrary, the number of available serum tests increased, as did the percentage of patients without glucose-lowering medication (the latter from 23.5% to 40.9%).Since activation of the IC model also depended on how the various health districts implemented organisational and logistical actions, there were significant differences in the distribution of quartiles between districts.All-cause mortality standardised rates for patients in the DC care model were double those in IC (Table 3); the all-cause mortality standardised rate for those in the Other-group was between the other two (DC: 44.0, 95%CI 42.0-45.9;IC: 24.5, 95%CI 22.5-26.5;Other-group: 30.4,95%CI 27.8-33.0,per 1000 person-year).In the comparison between the DC and IC models, the pattern was similar for all investigated causes except for AMI, where the gap increased slightly (DC: 2.2, 95%CI 1.8-2.7;IC: 0.9, 95%CI 0.5-1.3;per 1000 person-year), and for cerebrovascular diseases, where the gap decreased slightly (DC: 3.6, 95%CI 3.0-4.1;IC: 2.3, 95%CI 1.7-3.0;per 1000 person-year).The comparison between patients in the DC care model and Other-group was more erratic, with the lowest gap observed for cerebrovascular diseases (DC: 3.6, 95%CI 3.0-4.1;Other-group: 2.6, 95%CI 1.9-3.4;per 1000 person-year).
Restricting the analysis to IC-eligible patients actually placed in the DC or IC model (Table 4), all-cause mortality rate ratios still showed a reduced risk of 40% for IC patients compared to the DC model (adjusted IRR: 0.62, 95%CI 0.51-0.76).The same, even more pronounced, pattern was observed for all the investigated groups of causes.In the case of all-cause hospital admissions, the risk was almost 25% lower for IC patients than for DC patients (adjusted IRR: 0.74, 95%CI 0.66-0.82),and the gap was particularly wide for renal complications (adjusted IRR: 0.31, 95%CI 0.15-0.65).
When we adopted the instrumental variable approach using the proportion of IC-eligible patients actually included in the IC model by GPs as a main predictor (Table 5), the estimated rate ratios for all-cause mortality was 0.83 (95%CI 0.60-1.13)and 0.95 (95%CI 0.54-1.70)for cardiovascular diseases, under the hypothesis of a linear association between the outcomes and the instrumental variable and comparing a GP with 0% of eligible patients in IC vs. a GP with 100%.For diabetes mortality, there was still a decreased risk, largely compatible with random fluctuations.Similarly, the rate ratios for all-cause hospitalisations of a GP with 0% of eligible patients in IC vs. a GP with 100% were 0.90 (95%CI 0.76-1.06)and 0.91 (95%CI 0.69-1.20),respectively, for cardiovascular diseases.Regarding renal complications, the risk was reduced for GPs with high proportion of patients in IC, but the difference was compatible with random fluctuations.

Discussion
Our population-based cohort study using an instrumental variable approach showed that the IC model probably is not inferior to the DC care model for the main investigated clinical  outcomes.It showed negligible effects on mortality and hospital admissions.The progressive steps of the analysis (i.e., from current patients' distribution by care model to instrumental variable approach) allowed us to overcome both the selection and indication bias.Results from models adjusted by the main clinical characteristics, and also restricting the population to only those eligible, still showed a large difference in mortality and hospitalisation in favour of the IC model, but when an analysis similar to randomisation was adopted the differences almost disappeared.How to manage chronic patients is one of the main issues of this century.Many health systems in Europe and in the US have begun to involve primary care services more actively to reduce the workload of specialist clinics, to reduce patients' physical and personal distance to increase access to care, and finally, to favour proactive medicine interventions to prevent diabetes complications.The IC model implemented in Reggio Emilia was inspired by these principles.Nevertheless, the choice of a non-specialist setting for a disease as complex as diabetes could be not justified even in the presence of all these possible advantages if it resulted in less safe and less effective care.
The relevance of the question is confirmed by a recent Cochrane systematic review, conducted to determine the effectiveness of shared care health service interventions designed to improve the management of chronic disease across the primary/specialty care interface [24].Our results are consistent with the conclusions of this review.Specifically, among the main results, the systematic review highlights that the studies probably showed little or no difference in hospital admissions (with evidence of moderate certainty).As far as diabetes is concerned, one included RCT [10] found no significant differences in unscheduled admissions, one [25] found no significant differences in the number of patients admitted to hospital for a diabetesrelated reason, and number of deaths.A third RCT [26] included in the review was not taken into consideration because it compared all-cause mortality between "specialty care" (i.e., consultation care provided by an endocrinologist or a general internist in concert with the patient's primary care doctor) and "primary care" (cared for by family doctors alone).A further RCT aimed at evaluating the quality of a shared care programme for patients with T2D; providing evidence on the advantages and disadvantages compared with a programme in a specialised outpatient clinic is ongoing and the results are still unpublished [27].
Our starting hypothesis is that IC is not inferior to the DC care model in controlling diabetes-related morbidity and mortality, even though it should consume fewer resources and be less time consuming for the patient.
Indeed, an analysis of the current patients' distribution by care model, mortality and hospital admission showed that rates were higher for patients in DC care model compared to their IC counterparts.The guidelines recommend caring for low-risk profile T2D patients in the IC model.Therefore, by definition, patients included in the IC are those at lower risk of complications, hospitalisation and of death than the other T2D patients [6].Although different factors such as physician and patient preferences and medical or social patient frailty can influence the decision to care for T2D patients in the IC model, this was an expected result and is indirect evidence of adherence to guidelines.
Even patients included in the Other-group showed lower rates compared to their DC counterparts.This latter group is heterogeneous, and includes neglected patients belonging to difficult-to-reach populations, subjects being treated in other structured care models not specifically dedicated to diabetes (patients institutionalised or cared for at home for other chronic diseases), subjects who have opted out of any structured model because they prefer to be treated only by their GP or because they have a private diabetologist, and finally, newly diagnosed patients who have yet to choose a care model.
The results from previous observational studies on this topic are consistent with our analyses based on real-world data.Compared to the most similar studies [28,29], we found a lower proportion of patients who are not in a DC or IC model and higher mortality rates for each model of care.Possibly, the introduction of the DC and IC model in our province since 2006 has enhanced the accuracy in assigning patients to one of the two care models.Differences in standardised mortality rates can be due to the population used for standardising (the general population for the Turin study, while we used the population with diabetes) and inclusion criteria of the study population (our study is registry-based, while the other two studies defined a cohort based on administrative data).Both previous studies focused on comparing patients under structural care (IC and DC) and patients cared for by GPs alone, and showed higher mortality [28,29] and hospitalisation [29] for the latter group.Nevertheless, authors of the Baldo study [28] found lower mortality risks for IC compared to DC.
In our study, we chose to restrict the analysis to eligible patients to minimise selection bias induced by the presence of strict criteria to define IC-eligible patients.If the guidelines were applied in a uniform and complete manner and no other factors influenced the choice of assigning a patient to one care model rather than another, one would expect to find no eligible patients in the DC care model and all patients in the IC model defined as eligible.In fact, 21.8% of the subjects in our study population were IC-eligible, with the percentage varying across care models: 14.5% for DC, 37.1% for IC, and 21.6% for GP.Even when we restrict the comparison to IC-eligible patients, those actually cared for in IC experienced lower mortality and hospitalization risks than those cared for in the DC; for mortality, the reduction was consistent for all-causes, cardiovascular, cerebrovascular and diabetes, and for hospitalisation, for all-causes, cardiovascular, cerebrovascular, and renal complications.The difference is plausibly due to unregistered prognostic factors that were at the basis of the physician's decision to include or not include the patient in a low-risk model of care.
Finally, we adopted an instrumental variable analysis approach to minimise indication bias, i.e., to take into account latent factors such as physician and patient preferences and opportunity, as well as unregistered prognostic factors.We used the proportion of IC-eligible patients that were actually included by GPs in the IC model as an instrumental variable.We assumed that the case mix for each GP would be an almost random sample of the population.Indeed, we showed that mean age, time since diagnosis, HbA1c, mean BMI and triglycerides were balanced among each GP's eligible patients, which supports the validity of our instrumental variable.Only the availability of a serum test was not balanced, but this variable is clearly linked to the care model and to the GP's attitudes (i.e., ability to keep up to date, initiative, aptitude for teamwork), thus it is not surprising that it varies between GPs.Fortunately, not all the variability between GPs is due to each GP's characteristics: the activation of the IC model differed according to other organisational, logistical and timing factors (i.e., information technology facilities, contracts with GPs, differences in protocols of T2D management among diabetes clinics), as is clear if we look at the differences in proportion to IC-eligible patients actually included in the IC model between districts.

Strengths and limitation
The main limitation of this method is that we are using an ecological variable instead of an individual one, thus losing statistical power (the 95%CI gets larger in Table 5).We are also misclassifying patients, since our instrumental variable is a continuous variable; except for those GPs who have 0% or 100% of eligible patients included in the IC model, we are considering IC-eligible patients that are actually cared for by one definite care model as cared for by % of IC.The loss of power is quantified by the precision of our estimates and we can say that it is acceptable for all-cause hospitalisation and for all-cause mortality.For specific cause of death and hospitalisation, the estimates are very imprecise.The loss of discriminating power due to the use of an ecological variable depends on the variance in the proportion of patients in the IC models between GPs; the lowest quartile has less than 20% of IC-eligible patients in the IC model and the highest quartile more than 65% of IC-eligible patients are in the IC model.
Moreover, although our study included almost everyone with T2D living in the Reggio Emilia province, it is still underpowered to perform a formal non-inferiority test, which would assess acceptable differences in the considered major outcomes.This cohort study is population based and includes the results of implementing the IC model as it took place in all the resident population of IC-eligible patients in the province of Reggio Emilia.Furthermore, for our analysis, we considered only patients with T2D treated by the same care model for at least two years before the start of follow up, excluding those who shifted from one model to another.This choice was necessary because by definition, any worsening of the disease could imply a shift from IC to DC.Thus, most of the study outcomes occurred when the patient was in the DC model, even if resulting from the care received during the period when the patient was in the IC model.On the other hand, this choice introduced further misclassification of the exposure, as in any intention-to-treat analysis, driving our results toward a null hypothesis.

Conclusions
In low-risk patients with T2D, the integrated care model involving both GPs and diabetes clinic professionals showed similar mortality and hospitalisation risks as a model with higher use of specialized care, with exclusive patient management by the diabetes clinic.This was a necessary, but not sufficient, condition to prove the effectiveness of the integrated care model.Further research is needed to assess whether this model, aimed at reducing the workload of specialist care, increasing accessibility and facilitating proactive and initiative medicine through the involvement of primary care physicians, is actually less resource-consuming and more acceptable for patients.
The comparison between two care models with observational studies is affected by strong biases that systematically favour the one that should be prescribed to less complicated and less severe cases.

Fig 1 .
Fig 1. Flowchart showing the study population selection and the steps of analysis.S1 = step1 -mortality and hospitalisation rates by care model, no relative risk has been computed.S2 = step2 -to overcome selection bias, comparison was limited to patients eligible for the IC model, through mortality and hospitalisation risk rates, IC vs DC (Other-group vs DC analysis was omitted because the former was not a care model foreseen by the guideline).S3 = step3 -to overcome indication bias, we compared mortality and hospitalisation risk using the proportion of eligible patients actually in the IC model by GP (100% IC vs. 0% IC).DC = diabetes clinic; IC = integrated care.https://doi.org/10.1371/journal.pone.0194784.g001

Table 4 .
Mortality and hospitalisation rate ratios by care model.Mortality and hospitalisation rate ratios, with 95% confidence interval (95%CI), for all and diabetesrelated causes for patients in DC and IC model.Analyses restricted to IC-eligible patients.DC = diabetes clinic; IC = integrated care; DC was used as reference.

Table 1 .
Baseline characteristics.Patient's baseline characteristics by care model.DC = diabetes clinic; IC = integrated care; variables were collected on 31/12/2011; only patients with T2D and aged <85 placed in the same care model for at least 2 years are included.

Table 2 . Baseline characteristics by GPs' use of integrated care model.
GPs are classified according to the proportion of their IC-eligible patients who are included in the IC model.The eligible population is divided into quartiles according to this GP characteristic.The Table presents the covariates by quartiles of proportion of the GP's eligible patients included in IC model (1 st quartile lowest proportion of eligible patients included in IC model, 4 th quartile highest proportion).

Table 5 . Mortality and hospitalisation rate ratios by GPs' use of integrated care model.
Mortality and hospitalisation rate ratios, with 95% confidence interval (95% CI), for all and diabetes-related causes, comparing the proportion of each GP's eligible patients included in IC model.Analyses restricted to patients eligible for IC (N = 3812).