Data source and study population
Since 2011, the Catalan Health Department surveillance system (CHSS) collects detailed information on healthcare usage for the entire population of Catalonia (North-Eastern Spain, 7.5 million inhabitants) . It includes information from hospitalization, primary care visits, emergency department visits, skilled nursing facilities, palliative care and the mental health services, information on pharmacy prescription and expenditure, and a registry on the billing record also encompassing outpatient visits to specialists, home hospitalization, medical transportation (urgent and non-urgent), ambulatory rehabilitation, respiratory therapies and dialysis.
The registry has an automated data validation system that checks the consistency of the data and identifies potential errors. Moreover, as this information is used for provider payment purposes, external audits are performed periodically to ensure the quality and reliability of the data. The CHSS is also used to elaborate, on a six-month basis, the regional population-based health risk assessment tool, known as GMA (Adjusted Morbidity Groups), which generates the health risk strata pyramid of the general population of Catalonia [18, 19].
For the purposes of the current study, all adult residents (≥18 years) in Catalonia on 31st December 2014 were included in the analysis. This yielded a final study population of 6,102,595 cases. The research was undertaken under the umbrella of the Nextcare project , approved by the Ethical Committee for Human Research at Hospital Clínic de Barcelona (HCB/2018/0805). We used retrospective de-identified data from administrative databases and, therefore, the need for informed consent was waived.
The study compared the predictive power of four different measures assessing multimorbidity: the Charlson index , number of chronic diseases, Clinical Risk Groups (CRG) , and GMA [18, 19, 23].
The Charlson index was included because it is the most broadly used parameter to assess multimorbidity . This index was initially developed in hospitalized patients to estimate mortality prognosis based on age and the fixed weights of 20 specific disorders . The current study used the 2007 updated version of the Charlson index  adapted to primary care, further refined in 2014 .
The number of chronic diseases was based on the Clinical Classifications Software (CCS)  and the Chronic Condition Indicator (CCI)  elaborated by the Healthcare Cost and Utilization Project (HCUP) of the Agency for Healthcare Research and Quality (AHRQ). The CCS aggregates all diagnosis codes into 262 mutually exclusive, clinically homogeneous categories; whereas the CCI allows to determine if a diagnosis is a chronic condition. The combination of CCS and CCI provides the number of chronic conditions for a given subject.
The study also included information provided by the Clinical Risk Groups (CRG), elaborated to predict total annual health costs for large patient groups . CRG consist of mutually exclusive risk groups estimating past and future use of healthcare resources. It is of note that calculation of CRG required information on diagnosis across the Catalan health system during 2014, as well as data on pharmacological prescriptions during the same period. Only estimation of future use of resources was considered in the current study.
Finally, we assessed the role of the morbidity grouper developed in Catalonia (GMA) [18, 19]. GMA classifies the population into 31 mutually exclusive categories based on both multimorbidity and levels of patient complexity (see detailed information on the GMA’s algorithm and validation in Additional file 1: Figures S1–S3).
The outcome variables considered in the current study were: (A) Frequent attenders in primary care, defined by ≥12 visits to the primary care team irrespective of the professional (physician, nurse, physiotherapist, etc.) and the type of visit (primary care unit, home, remote) during the year 2015; (B) Patients receiving home care support either by the primary care team, emergency services or teams specialized in geriatric and palliative care during 2015; (C) Patients receiving social support visits defined as patients that performed visits to the community-based social care worker during 2015; and, (D) Patients receiving polypharmacy, defined by prescription of more than eight drugs during the year 2015. All the outcome variables were treated as dichotomous.
A sensitivity analysis carried out to determine the cut-off points for A) and D) showed similar results for percentiles 95 and 85. The later (P85) was used in the study. The analysis of patients receiving social support was limited to the primary care centres that had one social worker assigned to the staff (n = 4,776,005) due to the fact that in some geographical areas social support is directly linked to city council services and information was not available for the current analysis.
The current study consisted of a prospective analysis of existing registry information at 31 December 2014 to calculate the four multimorbidity measures (Charlson index, number of chronic diseases, CRG and GMA) and the events occurring during the entire 2015 for the four outcomes variables (A to D) described above. Results are expressed as mean values, standard deviations and 95% confidence intervals.
Logistic regressions were carried using each outcome variable as the dependent variable. For each model, the following covariates were considered: (i) age, (ii) sex, and (ii) socioeconomic level, as well as all first order interactions among those covariates. Moreover, the individual contributions of the three multimorbidity measures to the performance of the resulting predictive models was assessed, with a log-likelihood ratio test, through its inclusion as a covariate in the regression analyses. Accordingly, the model with age, sex and socioeconomic status was the baseline model. Age was analysed as a categorical variable grouped in 5-year intervals except for the two extreme periods, 18–19 years and > 94 years. Socioeconomic level was calculated as average income of all the residents living in the primary care area and expressed as a categorical variable using five levels . It is of note that multimorbidity measures were included in the predictive modelling as categorical variables to allow for possible non-linearity in the relationship between multimorbidity and the relevant outcome variable.
To evaluate the performance of the resulting predictive models, we calculated the Akaike information criterion (AIC) , the deviance-based R-square (R2) and the area under the receiver operating characteristic (AUROC) curve .
Statistical analyses were performed using SPSS software, version 18.0. All statistical tests and confidence intervals were constructed with a type I error (alpha) level of 5%, and p-values < 0.05, were considered statistically significant.