A high number of diseases and consultations: a warning signal for GPs when following up a multimorbid patient

Background: the European General Practitioners Research Network (EGPRN) designed and validated a comprehensive denition of multimorbidity using a systematic literature review and qualitative research throughout Europe. Detecting risk factors for decompensation would be an interesting challenge for family physicians (FPs) in the management of multimorbid patients. The purpose of the survey was to assess which items belonging to the EGPRN multimorbidity denition could help to identify patients at risk of decompensation in a cohort pilot study over a 24-month follow-up among primary care outpatients. Method : 131 patients meeting the multimorbidity denition were included using two inclusion periods between 2014 and 2015. Over a 24-month follow-up, the « decompensation » or « nothing to report » status was collected. A logistic regression, following a Cox model, was then performed to identify risk factors for decompensation. Results : After 24 months of follow-up, 120 patients were analyzed. 3 clusters were identied. 44 patients, representing 36.6 % of the population, were still alive and had not been hospitalized for a period exceeding 6 days. Two variables were signicantly linked to decompensation: the number of visits to the FP per year (HR 1.06, IC 95 %, 1,03-1,10, p-value <0,001) and the total number of diseases (HR 1,12, IC 95 %, 1,013-1,33, p-value = 0,039). Conclusion: FPs should be aware that a high number of consultations and a high total number of diseases are linked to severe outcomes such as death or unplanned hospitalization. A large-scale cohort in primary care seems feasible to conrm these results.

in investigating the complexity of patient conditions was an exciting project that could improve GPs' ability to detect a patient's frailty and to prevent severe outcomes such as hospitalization or death.
The concept of multimorbidity had stirred up some interest in the research agenda of the European General Practitioners Research Network (EGPRN) (14).
A research team which included 9 European countries, all involved in the European General Practitioners Research Network (EGPRN), aimed to clarify the concept of multimorbidity (15) . In 2012, the EGPRN presented a comprehensive de nition of multimorbidity in family medicine and long-term care.(16) It was the rst attempt to de ne multimorbidity which bene ted from the FP's pragmatic point of view (17) . According to the EGPRN, multimorbidity is de ned as « any combination of chronic disease with at least one other disease (acute or chronic) or psychosocial factor (associated or not) or somatic risk factor, any bio-psychosocial factor, any somatic risk factor, the social network, the burden of diseases, the health care consumption and the patient's coping strategies may function as modi ers (of the effects of Multimorbidity). Multimorbidity modi es the health outcomes and leads to increased disability or a decreased quality of life or frailty. » Thirteen themes have been outlined and then translated into ten European languages with the intention of inducing and standardizing further collaborative studies (18) .
Nevertheless, that exhaustive de nition covers too large a part of the population which renders it ineffective as it reduces the opportunity to pinpoint a patient at risk of severe outcomes. Since preventing acute hospitalization or death is a major concern of FM, (19) the EGPRN considered how to determine which variables within the concept would be effective in preventing those two events (20).
As FPs are very familiar with their patient's health status, (21) there could be some factors that might be neglected but which could help to avoid severe outcomes if they were noticed in time (22).
If the EGPRN could succeed in highlighting those variables, they could be integrated within the FPs' medical software and be a powerful tool in the care of multimorbid patients (23) .
Consequently, a feasibility pilot cohort study was started in 2014 which included patients who met the EGPRN multimorbidity de nition (24) . Data were analysed every three months.
The main objective of this study was to assess which criteria, within the EGPRN concept of multimorbidity, could identify multimorbid patients at risk of decompensation in a primary care cohort in France over a 24-month follow-up.

Ethic statement
The study was approved by the ethics committee of the « Université de Bretagne Occidentale » Faculty of Medicine, Brest. The participants had to sign a written informed consent to participate in the study. The Family Physicians involved provided a verbal consent.
The questionnaire is fully available (Appendix 1) as it was given to the FPs (in the French language).
From the rst pilot study, some irrelevant variables were deleted for different reasons: -Chronic condition redundancy with chronic disease or psychological risk factor) -Cost of care (impossible to estimate given the time and resources dedicated to the study) -Disability (disability / impairment, qualitiy of life and health outcomes (consequences rather than the characteristics of the multimorbidity) -frailty (absence of consensual de nition, criterion assessed by study, methodologically impossible to assess at the beginning of the study) -physiology (too broad a notion, impossible to evaluate) -disease and assessment (present in the theoretical de nition but missing from the coding book and not found in the transcripts) -demography and aging (redundant with sociodemographic characteristics) Somatic risks were evaluated as cardiovascular risk factor, risk of falling factors (calculated with the CETAF score) (25) , an assessment of hygiene, nutrition and physical activity at the discretion of the FP. This questionnaire was accepted by the scienti c committee of the research team and tested with FPs and medical students.
Thanks to the comments made during the rst pilot study, a revised questionnaire was used for those patients included from September to November 2015. The question order was actually rearranged to ease and shorten the time needed to complete it. No question could be asked before the previous one had been answered in order to avoid omitting data. The FP was only asked question number 40 if he/she had answered « yes » to organized or individual screening in question number 39 to avoid errors in completion identifed in previous studies.
Data were saved using Microsoft Excel and computed by the online survey software EVA-LANDGO ®. 24 months after inclusion, FPs were contacted by email or phone (INSERER APPENDIX 5) to collect patient status information. Two types of status were de ned after a consensus had been reached in a peer group gathering on multimorbidity made up of physicians, residential students and researchers in family practice: « decompensation » (D) and « Nothing to Report » (NTR). Decompensation was understood as the occurrence of hospitalization of at least 7 days duration, or death, during the 24 months of follow-up, as the mean duration for hospitalization in the European Union is 6.7 days (26) .
Groups labelled « frail » or « not frail » in the feasibility study were changed at the six-month follow-up as confusion between « frail » and « frailty » might occur and this would include a de nition which was not consensual (27) .

Data analysis
Data cleaning was performed to harmonize data for analysis. The ICD 10 was used to standardize the mentioned chronic diseases. 102 chronic diseases were reported.
Missing data were spotted during descriptive statistics analysis. They were replaced by the median value to be incorporated into the statistical analysis.
Each modi cation and the reasons for it were compiled in the « dictionary » available on demand from the corresponding author.
A description of the population was the rst step. Both types of status « decompensation » and « nothing to Report » were compared using a bi-dimensional analysis for each variable.
Quantitative variables were compared using a Fischer's exact test or a chi-2 test with an alpha level set at 5 %. Qualitative variables were compared using a Student's test when it followed a normal distribution to compare the means of the two groups from a normal population, and a Shapiro-Wilk's test when it did not follow a normal distribution, to compare the medians of the two groups.
Patients with the same characteristics. regardless of status « Decompensation » or « Nothing to Report », were grouped together using a multidimensional analysis. Non-discriminating and non-descriptive variables were removed. Then, a clustering represented in a dendrogram and a multiple correspondence analysis (MCA) were performedto identify discriminating variables for each group and the resulting information was combined using the technique of hierarchical clustering on principal components (HCPC).
The second step comprised a statistical analysis of the follow-up. A logistic regression was used for the six and nine-month follow-up as the dependent variable was binary (D versus NTR), regardless of followup time. A Cox model was chosen to complete the analysis from the twelve-month study, making it possible to apply different durations of follow-up time to support each patient.The aim was to nd the best subgroup of variables for predicting and explaining the patient's status at 24 months.
At rst, the overall survival of the two populations for each variable were compared using a nonparametric estimator by Kaplan-Meier, with a Wilcoxon test of alpha-risk at 5%. Then, the team estimated the survival function using semi-parametric models. A Hazard Ratio (HR) was rst obtained by a univariate analysis using Cox's Model. Then, an adjusted HR was obtained with a multivariate analysis using Cox's model, representing the association between a variable and the decompensation risk factor.

Results
Sample participants 137 patients were included by 31 FPs. Out of the 137 patients, 6 were excluded for failing to complete questionnaires or for duplicate questionnaires. The status at 24 months was collected for all 131 patients. 11 were lost in follow-up because of a change of FP or because the FP ceased working. ( Figure  1) Data cleaning and recoding Some variables were removed from data analysis because they were: not discriminating: divorce, use of pharmacological treatment, neglect of the patient, patient victim of iatrogeny, nursing home, lack of entourage of no use: identi cation number, inclusion date, date of birth (expressed as age) Irrelevant for the objective (referring to FPs feelings): variables related to quality of care, detection of multimorbidity and doctor self-assessment, detailed and/or complex medical history.
The recoding data work was transcribed in a dictionary (available on demand from the corresponding author).

Characteristics of the patients included
A cluster dendrogram was summarized from a hierarchical ascending classi cation (FIGURE 2). Inertia gain determined the number of clusters and, thanks to the clustering quality index, three groups were retained. The MCA factor map projected those three groups in two dimensions.
A comparison of the proportion within the group (PwG), and within the study population (PwP), was carried out for each variable for the purpose of characterizing all three groups. (Figure 3 The analysis highlighted several characteristics for the patients belonging to the D group: they were more likely to suffer from postural instability (73% vs 49%, p-value = 0.018) They were more likely to be single or widowed (52% vs 30%, p-value = 0.028) Equipment at home was more readily available (39% vs 13%, p-value = 0.003) More human help was available at home (52% vs 26%, p-value = 0,008) Their medical history was detailed and more complex (93% vs 68%, p-value = 0.004).
Among the quantitative variables, six were signi cant with an alpha-risk at 5%: the patients in the D-group were older (80 years old vs 69 yo, p-value <0.001) They had more diseases, taking into account both chronic and acute conditions (7 vs 6, p-value = 0,016) They visited their FPs more often (12 vs 4, p-value = 0.010) They took more medication per day (8 vs 7, p-value = 0.003)

Survival analysis
The overall survival for each variable and between the D and NTR groups was compared using the nonparametric estimation of Kaplan-Meier.
The probability of no decompensation at 24 months was 63.3 % (95% IC, (55.3%-72.6%). Two variables had a signi cant protective effect: excess weight (Log-rank test, p-value = 0.038) and not being single or widowed (Log-rank test, p-value = 0.015). On the other hand, four were signi cantly linked to decompensation: detailed and complex medical history (Log-rank test, p-value = 0.003), human help at home (Log-rank test, p-value: 0.002), equipment to help at home (Log-rank test, p-value <. 0.001) and multiple complaints patient (Log-rank test, p-value = 0.03). (Figure 4) Twenty-ve variables appeared to be linked to the risk of decompensation according to a uni-variate analysis by Cox regression. They are referred in TABLE 3. Sixteen of them were statistically signi cant with an alpha risk at 5% (in bold in

Main results
The purpose of this survey was to assess which FP criteria in the EGPRN de nition of multi morbidity were the most accurate for identifying patients at risk of decompensation. Over 24 months of follow-up, this study highlighted 2 variables associated with decompensation: « number of visits to FPs » and « total number of diseases ». Those two variables are sub themes of « health care consumption » in the de nition of multimorbidity, according to the EGPRN.

Strengths and limitations of the study
Selections bias. FPs who selected multimorbid patients were aware of the study's aim. They may have selected patients with a high risk of decompensation although this bias was minimized by the exclusion criteria « estimated survival less than 3 months ». For the most part, FP recruiters were clinical teachers. It is a well-known fact that clinical teaching FPs undertake work which differs from that of non-teaching FPs (28) (29). Therefore, results with a more general population might be different.
Information bias. To avoid missing data, the questionnaire was modi ed after the rst inclusion period: moving to the next question was impossible without answering the previous one. As this move has effectively avoided omitting data and increased the response rate to some questions, it may have signi cant impact on the results of the statistical analysis. Some data were missing and others were inconsistent. The origin of that issue may be due to some unclear and/or laborious questions and to the length of the questionnaire. Missing data were removed in the statistical analysis to reduce information bias.
Exhaustion after 24 months of follow-up and lack of time to nd the answer to some questions may have caused some errors or omitted data.
The 102 chronic diseases reported by the FPs during the study were clustered in a single category, and others were moved into the acute disease or risk factor category. To limit information bias, the clustering was decided by the scienti c committee using the ICD 10.
Although the CETAF score was not validated for the under 65-year-olds, the team assumed that it would not be high for people under 65 years old and would have no impact on the statistical analysis result. Therefore, the CETAF score was calculated for every patient (30) .
Data transcription from the Evalongo software which was used to complete each questionnaire as an Excel le, in order to ease the analysis, may result in some transcription errors.

Confounding bias
With regard to the questionnaire, themes and sub themes of the English EGPRN de nition of multi morbidity had to be translated into French, and some errors of translation could have occured. Between the two inclusion periods, some quantitative concepts were transformed into qualitative variables, which may have led to errors of transformation.
Given the small number of patients, there was a large number of variables that might have hampered the analysis. With the intention of reducing those di culties, the decision was made to reduce the number of variables, following expert advice. The variables removed were those which were redundant or not statistically relevant according to the peer group.
Lastly, and despite the fact that it has the virtue of being objective, clinical and valid in literature, the judgment criterion for « decompensation » de ned as hospitalization for at least seven days, or death, could be a subject for discussion. However, this choice avoided confounding bias.

Key points
At the 24-month follow-up, the number of visits to the FPs and total number of diseases were the most useful variables of the EGPRN de nition of multimorbidity to predict decompensation. Contrary to the follow-up at 6 months, family problems were not found to be signi cant.
Every study before that cohort has failed to assess the meaning or the intensity of the relationship between multimorbidity and health outcomes (31) (32). These two variables will help to clarify the concept of multimorbidity when the subject in question is speci cally the decompensation outcome. In addition, it will ease the burden for FPs in their clinical practice as they work to identify patients who are at risk of decompensation, and to prevent this outcome.
The number of FP consultation per year was also found to be a risk factor for decompensation in the previous studies at 6, 9, 12, 15 and 18 months (24) .
Earlier studies had found an association between multimorbidity and the number of FP visits (18) (33) (34) .
The follow up at 6 months found that « age », « number of visits to FPs » and « family problems » were linked to the risk of decompensation. As « age » is a non-modi able factor, only « family problems » and « number of visits to FPs » could help to prevent decompensation. At 24 months, « age » was signi cantly linked to decompensation in the univariate analysis but the expert group didn't integrate it into the multivariate analysis.
As regards « family problems », belonging to the psychosocial risk factor theme, the univariate analysis at 24 months did not nd a link with a risk of decompensation, contrary to the previous studies.
These differences may be explained by the fact that more patients were included than previously which can lead to some changes in the characteristics of the population, as could the fact that eleven were lost to follow-up. Finally, the amendments made to the questionnaire may have interfered with the results of the statistical analysis.

Implications for practice, teaching and future research
In everyday practice, FPs should keep in mind that a multimorbid patient who frequently visits them is at risk of decompensation. This point is easy to deal with and costs nothing to monitor in primary care.
Trainees in family medicine should be aware of this risk factor. This study is included in an EGPRN project and is destined to be reproduced as a large-scale European study. It is apparent that a clear understanding of the concept of multimorbidity and of the risk factors for decompensation should have a major impact on managing multimorbid patients and on health system expenditure.
It could be interesting that some variables originated from patients, even though this contribution may be considered subjective.
A new, simpler and less demanding questionnaire, should be proposed in further studies in order to avoid errors or data omission due to recruiters being short of time.
As the total number of diseases was not found to be a risk factor at the previous follow-up, a future study should be run in order to establish a standardized disease classi cation that could help to identify multimorbidity and its risks of decompensation.

Conclusions
Once again, the number of FP consultation is found to be a risk factor for decompensation among multimorbid patients in this 24 months of follow-up study. For the rst time in this cohort, the total number of diseases was shown to be a risk factor for decompensation. Meanwhile family problems are no longer considered a risk factor. A large-scale study should complete and con rm these outcomes which would, in turn, facilitate research and clinical practice on the concept of Multimorbidity, an key topic of the EGPRN.    Overall survival curve of the cohort

Supplementary Files
This is a list of supplementary les associated with this preprint. Click to download.