The effect of continuity of care on medical costs in patients with chronic shoulder pain

Unnecessary surgery could be prevented through continuity of care (COC). The present study aimed to investigate the relationships between COC, surgery and cost associated with chronic shoulder pain. We used the Health Insurance Review and Assessment Service national patient sample (HIRA-NPS) in 2017. A total of 1717 patients were included. Bice–Boxerman Continuity of Care Index was used as the indicator for measuring the COC. Occurrence of surgery, associated costs, and direct medical costs were analysed. Logistic regression, a two-part model with recycled predictions and generalized linear model with gamma distribution were used. The majority of patients were 40–65 years old (high COC: 68.4%; low COC: 64.4%). The odds ratio (OR) for surgery was 0.41 in the high-COC group compared to the low COC group (95% CI, 0.20 to 0.84). Direct medical cost was 14.09% (95% CI, 8.12% to 19.66%) and 58.00% lower in surgery cost (95% CI, 57.95 to 58.05) in the high-COC group. Interaction with COC and shoulder impingement syndrome was significant lower in direct medical cost (15.05% [95% CI, 1.81% to 26.51%]). High COC was associated with low medical cost in patients diagnosed with chronic shoulder pain.

Study population. The study population consisted of patients who developed chronic shoulder pain after suffering an acute shoulder injury. To select patients with acute shoulder injury, 31,621 patients whose initial episode of shoulder pain started as dislocation and sprain of joints and ligaments of shoulder girdle (S43) was selected. The period of progression into a chronic disease was defined as three months after the initial onset 20 . Accordingly, if shoulder lesion (M75) was claimed as main diagnosis after the initial claim of S43, the patients was considered progressed to chronic shoulder pain (n = 2648). Patients who underwent shoulder surgery during acute phase (i.e. within three months after initial claim of S43) (n = 114) were excluded. Patients aged < 20 years (n = 27) were excluded. Patients diagnosed as red flags such as cancer (C, D) or serious injury (T, V) (n = 484) were excluded. If the number of visits is too low during the overall period of the episodes, the measurement of COC become unstable 21 . Accordingly, previous studies usually included only patients with outpatient medical visits of four or more times during total episodes 14,22,23 . The present study applied the same criterion and excluded patients with outpatient medical visits of three times or fewer during total episodes (n = 306). As a result, 1717 patients were included in the final analysis (Fig. 1).
Continuity of care. The present study used the Bice-Boxerman COC index 24 as the index for measuring COC. COC ranges between 0 and 1, with a higher value indicating higher COC. COC, which is a frequently used index to measure continuity 25 , accounts for variance of service providers that a single patient visited during the total episodes and concentration of visits for a specific service provider (Eq. 1). In the process of calculating COC, utilisation of outpatient care for main diagnosis of shoulder pain (S43, M75) was used. High-and low-COC groups were divided based on median values. To our best knowledge, there is no absolute standard for differentiating high and low COC, and as a result, previous studies applied various standards according to the research model, including the median 26 , tertiles 27 , quartiles 15 , and continuous 13 . Due to the relatively fewer events as compared to previous studies, we used the median as the standard for differentiating high and low COC in the main study. Other standard was used for sensitivity analysis. Outcome. The dependent variables included occurrence of surgery, conservative treatments, surgery associated cost, and total medical cost. Occurrence of surgery was defined as possible surgery due to failed care of chronic shoulder surgery.
Surgical intervention must be chosen when there is no improvement or exacerbation of symptoms during rehabilitation 28 . Typical diseases of this type include those that involve the rotator cuff, such as shoulder impingement syndrome and rotator cuff tear. Surgical intervention includes acromioplasty and suturing if rotator cuff tear is involved 28 . Also, Partial or total arthroplasty is needed if open reduction is not possible due to inadequate fixation of the humeral head or humeral head wear due to serious injury or deterioration 28 . Accordingly, among the possible surgical interventions after an acute sprain, acromioplasty, rotator cuff repair (N0935, N0936, N0937, N0938), and shoulder arthroplasty (N2071, N2711) are defined as events 29 . Conservative treatments patients received during the episodes includes physical therapy, X-ray, injection, acupuncture, nerve block, non-steroidal anti-inflammatory drugs, paracetamol, muscle relaxants, corticosteroids and opioids. The details of conservative treatments are described in supplementary table S1.
Regarding medical costs, total claimed medical costs from the national health insurance claims data were analysed. Medical costs were divided into direct and indirect medical costs. Direct medical costs include costs associated with outpatient care, inpatient care, ED use, medication, diagnosis, treatment, medical devices, and hospital stay. Indirect medical costs include loss of productivity due to a disease and associated costs 30 . Only direct medical costs associated with medical utilisation were included as medical costs 31 while total medical costs Covariates. Gender, age, and insurance type were included as sociodemographic characteristics. Age groups were divided based on 40 and 65 years according to the definition of life transition periods in Korea. To adjust for severity, adhesive capsulitis of the shoulder (ACS; ICD-10: M750) and shoulder impingement syndrome (SIS; ICD-10: M754), which have the highest prevalence among chronic shoulder diseases in Korea, were used as comorbidities. The Charlson Comorbidity Index (CCI) was also used 32 . The CCI is most widely used index and also used for musculoskeletal disease 33 . The CCI was calculated with presences of 17 comorbidities. Presence of comorbidity was examined based on whether those disease codes were diagnosed before the end of the episodes. Moreover, the total number of medical visits for shoulder pain during the episodes was adjusted as previous studies 13 , assuming that more severe patients would be more likely to use medical services. It was considered that the characteristics of the main attending medical institution would affect continuity and patient behaviour, and thus, the institution that the patient had visited the most during the total episode was defined as the primary provider and accounted for whether the institution is a clinic or hospital 11,34,35 . A total of 405 patients had two www.nature.com/scientificreports/ or more medical institutions most frequently visited with the same number of visits (e.g., total of 15 visits with 5 visits each to institution A, B and C). In this case, one institution was selected randomly.
Statistical analysis. The difference in basic characteristics between groups was examined with independent t-test and chi-square test. To analyse the effects of COC on surgery, a multivariate logistic regression analysis was performed with covariates controlled. However, a quasi-complete problem occurred in the logistic regression analysis due to the relatively small number of events (n = 41). Since this made it impossible to estimate the maximum likelihood, we performed Firth's logistic regression analysis 36 . Two types of costs were analysed. First, the effects of COC on medical costs during the total episodes were analysed. The distribution of medical costs was right-skewed, and as a result, heteroscedasticity existed in the residuals. To resolve this, generalised linear model (GLM) with gamma distribution and log link function were used 37 . The difference in cost was expressed as percentage reduction due to high COC. For this, the low-COC was referenced, coefficients were exponentiated, multiplied by 100, and subtracted from 100 to analyse the effects of COC on surgery cost, a two-part model was employed first 38 . The two-part model was used because most patients did not undergo surgery, and therefore, their surgery cost was 0. As the first part, multivariate logistic regression for analysing the effects of COC on surgery was performed. With estimated model, the probability of undergoing surgery was predicted for each patients. As the second part, GLM with gamma distribution was performed only on patients who underwent surgery. Based on the estimated model, the surgery cost of the entire study population was predicted. Subsequently, the predicted probability and surgery costs in each part were multiplied. The second part was analysed only if the effects of COC was statistically significant (p < 0.05) in the first part, because if the effect of COC on surgery was not significant, the analysis on the effects of COC on surgery costs would be meaningless.
Next, recycled predictions were performed to analyse the marginal effects of COC on surgery cost 39 . With two-part model estimated previously, the costs of surgery were predicted with two scenario, all patients with low COC scenario and high COC scenario. The other covariates were fixed. Then, differences in surgery costs between each scenario were calculated. The mean and the 95% confidence interval was estimated with 1000 bootstrapping.
However, using data from patients who underwent surgery to predict the cost of other patients may lead to biased results. If generalised linear model is adopted under such situations, over-fitting may occur due to the number of patients being too low. Following the discussion by Kogure 40 , we performed a Lasso regression analysis 41 . With this, L1 penalty is applied when estimating the coefficient by ordinary least squares as a tradeoff between bias and variance. Kogure 40 demonstrated that the accuracy of prediction could be increased by regularisation through a penalised model in the two-part model. In sensitivity analysis, a model including basic demographic characteristics and severity, ridge regression 42 , and elastic net 43 were performed, and the results were compared. Lasso, ridge regression, and elastic net were performed with y log normalised, and Duan's smearing factor was added in the process of deriving the exponential.
For the sensitivity analysis, we first used the tertiles as the standard for differentiating high, middle and low COC. Due to the very small of number of the events, two-part model was not performed. Second, we assumed situations where the act of insurance claim may have been inaccurate. According to Park 44 , the accuracy of insurance claims in Korea is approximately 80%. We assumed that the act of insurance claims most likely to be inaccurate was from patients who developed chronic shoulder pain (M75) as acute pain (S43). Accordingly, the same process was carried out after randomly including 20% patients among those who claimed S43 even three months after the onset of acute shoulder injury. Moreover, subgroup analyses were performed based on the presence of chronic diseases including frozen shoulder and shoulder impingement syndrome, and the interaction effect between COC and each disease was derived. When analysing the medical costs, the number of visits was considered to be important. Accordingly, the covariates were included in the model hierarchically, and we observed changes in the coefficient of COC each time a covariate was added. SAS 9.4 and R studio version 1.1.463 was used for statistical analysis.

Results
The basic patient demographics are shown in Table 1. The patients were divided into high-and low-COC groups based on the median COC value of 0.50 and were then compared. The overall distribution of COC is shown in Fig. 2. With respect to basic demographics, their distribution was similar between the high-and low-COC groups. However, the low-COC group had a higher prevalence of shoulder impingement syndrome (high COC: 25.5%; low COC: 33.4%), while the high-COC group had a higher total number of medical visits (high COC: mean = 16.1 ± 16.1, median = 11; low COC: mean = 12.1 ± 10.9, median = 9). Table 2 describes the type of shoulder surgery and associated costs for the high-and low-COC groups. There were 41 surgery cases, most of which were to repair a rotator cuff tear (n = 40). The number of surgeries was higher in the low-COC group (high COC: n = 9; low COC: n = 32), with a noteworthy difference in incidence between the high COC (11.25 per 1000 population) and low COC (34.89 per 1000 population) groups. With respect to total medical cost, the median was higher and mean was lower in the high-COC group than in the low-COC group (high COC: mean = $364.43 ± 401.36, median = $236.24; low COC: mean = $396.52 ± 602.08, median = $220.99). A similar tendency was found for surgery cost. The distributions of total medical cost and surgery cost are shown in Figs. 3 and 4. The details of conservative treatments each group received during the episodes were described in Supplementary Table 1. In brief, low-COC group had more nerve block, and medications such as corticosteroids and opioids.
In the analysis of the effects of COC on surgery and medical costs (Table 3), odds ratio (OR) for surgery was lower by 0.41-fold in the high-COC group than in the low COC group (95% CI 0. 20 Table S3). The tendency in the analysis accounting for inaccurate claims in data was similar to that of the main analysis (Supplementary Table S4). When the covariates were hierarchically included in the model, the largest decrease in AIC and BIC was found when the total number of medical visits was included (decrease in AIC: 635.9957; decrease in BIC: 630.5473) (Supplementary Table S5).      11,13,14,[16][17][18] , in which high COC showed an association with low medical cost. They were also similar to other studies on ED visits 16,18 in that exacerbation of symptoms was prevented. However, most of these preceding studies involved metabolic disorders, while we were able to find very few articles on musculoskeletal disorders and COC 15 . This article 15 reported that higher COC resulted in lower hospitalisation and treatment costs among patients with osteoarthritis of the knees. The present study demonstrated that medical cost and risk and cost of surgery for severe cases decreased in patients with chronic shoulder pain. Subgroup analyses showed that COC effects on the direct medical costs was greater in patients with shoulder impingement syndrome (15.05%; 95% CI, 1.81% to 26.51%). According to a previous study 45 patients with shoulder impingement syndrome had higher costs associated with procedure, surgery, and hospitalisation than patients with frozen shoulder. It is believed that COC prevented such medical uses, which made the medical cost reduced more significantly than other diseases. However, the high-COC group had a higher number of medical visits and a higher median value for direct medical costs. Such results suggest the need for developing a policy that could reduce the number of medical visits while maintaining a high level of COC.
In caring for chronic shoulder disease after an injury, COC may provide high quality care in conservative treatment. Most cases of shoulder pain are primarily treated by conservative treatment. For example, rotator cuff tear or impingement syndrome with mild severity could be improved with at least three months of conservative treatment, such as rehabilitation therapy including muscle strengthening exercises 28 . During the process of taking part in a repetitive rehabilitation program, forming a rapport between the patient and physician could be helpful. According to the previous study, COC could be helpful in forming such rapport 46 .
The present study had some limitations. Firstly, the number of events was too small. In the two-part model prediction, the number of surgery cases was low at only 41. To compensate for this limitation, penalised regression and multiple models were used. The results showed similar tendencies, but because of the small number of events, a validation set for LASSO regression could not be created. Second, due to the limitation of the raw data (e.g., claims data), only a cross-sectional study with one-year observational period of 2017 could be conducted. Due to the short follow-up period, the analysis might be restricted to the effect of COC in the early phase of chronic shoulder pain. In future studies, it would be necessary to establish a long-term cohort to investigate the prognosis of chronic diseases. Third, in the analysis with tertiles standard, the dose-response was unclear in the OR for surgery. But the total cases of surgery was very small and dividing the groups into tertiles could dramatically lower the power of the logistic regression. The further analysis should be conducted with more samples. Fourth, although sensitivity analysis was performed, accuracy of claims data might be problematic 47 . However, considering that previous studies reported that the accuracy of diagnosis in insurance claims data is 80%, the sensitivity analysis showed a similar tendency with the main results. Lastly, because only insurance claims data for the shoulder were analysed, we could not predict all costs associated with the shoulder. Generally, total medical costs are categorised as direct, indirect, and non-medical costs 48 . The present study estimated only the direct medical costs covered by insurance. Non-medical costs and indirect cost should be considered in the further studies.
Despite the limitations, the present study has potential importance in that it used health insurance claims data most representative of the general population of Korea to investigate samples of patients with shoulder disease. We confirmed robustness through various sensitivity analyses, and the heterogeneity of outcome according to patient characteristics was analysed by subgroup analyses. Moreover, to our best knowledge, the present study is the first study that examines the relationship between COC and shoulder disease. Therefore, the findings in the present study could provide meaningful and useful information to clinicians, policy makers, and patients who use medical services for shoulder pain.

Data availability
The HIRA-NPS is provided by the Health Insurance Service & Assessment Service in Korea. To protect privacy, access to the data is available only for certified researchers in South Korea. The study protocol was approved by the Institutional Review Board of Jaseng Hospital of Korean Medicine (JASENG 2019-07-005) and followed relevant guidelines. The requirement of informed consent from the study population was waived by the same IRB.