We conducted a prospective multicentre trial at 19 primary care practices in the county of Östergötland in Sweden as previously described [26].
We rated the trial design as pragmatic according to all ten domains of the PRECIS-tool (Pragmatic Explanatory Continuum Indicator Summary) [27] as we studied the effects of implementation of a new work-mode without extra measurements or assessments in the intervention or control groups. As this work-mode was already decided by the county, participants were not asked for consent, in accordance with the decision of the ethical board.
We decided to start the collection of healthcare data (the follow-up period) when 90% of all available participants in the intervention group that accepted to participate had received the CGA (the run-in period). Our original plan was a run-in period of 3 months, but we had to prolong this period to 9 months because the start-up process required more time than expected. The staff information and training started in January 2017, recruitment of patients for CGA in April 2017 and the follow-up period of 24 months started in January 2018.
Setting
We were unable to randomise the practices, therefore we decided to compare the volunteering intervention practices with care-as-usual at matched practices in the county. We identified practices with similar locations, size and populations over 75 years of age and selected them as controls.
The practices were situated in both rural and densely populated areas with listed populations from 6000 to 21,000 inhabitants. Together, the 19 practices covered 40% of the total population aged 75 years and older across the county. Prior to the intervention, there was no tradition of performing CGA in primary care. Primary care in Sweden is financed by the county council and organised in primary care practices with a typical population of 6000–20,000 persons and an average workforce of 3–10 physicians, 6–15 nurses and administrative staff. Nurses are an important part of the workforce and have experience of working independently with both acute illness and chronic diseases, which makes them capable of performing a significant part of the CGA. The intervention practices did not receive any extra staff linked to the intervention.
Staff training and monitoring
The practices in the intervention received an introductory visit of 2 h during the run-in period where researchers introduced the CGA tool, the concept of frailty and the other features of the intervention to the nurses and physicians involved. During the whole study period, we visited the practices every 6 months giving advice and answering questions. Between these visits, the nurses at the intervention practices had access to a supervisor via email or telephone for additional support. In addition, we organised network meetings every 6 months where nurses and physicians at the intervention practices shared experiences, discussed CGA and related topics. There was no interaction with the control practices.
Participants
We identified 1604 individuals aged 75 years or older at the participating practices in March 2017. We used a recently developed and validated prediction model that contains 38 variables identified with multivariable logistic regression [23]. Age and healthcare use are the principal predictors, together with diagnoses from inpatient care and outpatient visits. Data from the preceding 12 months of electronic medical records was extracted to calculate a risk score for unplanned hospital admission. Our participants constituted the 11% with the highest risk score of the aged population. There were no exclusion criteria; all the selected individuals were included. We distributed a list of participants to the intervention practices at the end of March 2017.
Sample size
We hypothesised a reduction in hospital care days of 20% as a result of the intervention based on an earlier study where a reduction was found for individuals at high risk [28]. In a pilot study, we found an incidence of hospitalisations over 60% in the predicted target population. A calculation based on this, a power of 0.8 and a significance level of 0.05 led to a minimum number of participants of 270 in each group. We estimated a dropout rate of 30% giving a number of 380 participants in each group. Then we doubled the number of participants taking into account that the heterogeneity of the practices and the participants would lower the likelihood of detecting a meaningful intervention effect.
Intervention
The intervention comprised two main components. The first component involved presentation of the list of participants to the intervention practices. The list included the risk score and the number of hospitalizations, visits to emergency room and to any physician during the preceding 12 months for each participant. In this way, the practices were made aware of a group of older adults at high risk for hospitalisation.
The second component was the comprehensive geriatric assessment (CGA) performed by nurses at the intervention practices together with the listed/responsible physician of the participant. The CGA was carried out during the run-in period between April and December 2017. The primary care nurses contacted the participants in the intervention group by telephone and offered them a health evaluation/CGA and follow-up at the practice. Around 10% of the participants had greater reduction of mobility and received the CGA at home. We constructed an instrument for the CGA in this intervention, the Primary care Assessment Tool for Elderly (PASTEL) [25]. This 4-page form contains two parts. The first part is performed by a nurse and contains self-rating of health and about 20 items covering different perspectives of health and frailty including for example social network, vision, hearing, falls, incontinence, weight loss and psychological problems. This is followed by a medication review and physical measures. It ends with questions to the patient regarding the main concerns about their health and their needs in the future.
The second part is a template for a team meeting where the responsible physician and the nurse that performed the initial assessment together grade frailty with Clinical frailty scale [6] and review a check-list for further investigation and supportive actions, based on the assessment, medical records and personal knowledge. Typically, the assessment by the nurse lasted 1 h and the team meeting 15–30 min. In some cases, the participant was not known by the physician, and a visit for further medical assessment was planned. The participants did not participate in the team meeting, but was contacted by the nurse and involved in the care planning afterwards.
After the CGA, the participants were given care according to their individual needs. This included referral to occupational therapy, physiotherapy or other specialist services when needed. We encouraged the teams to provide continuity for the participants to both physician and nurse and to facilitate accessibility to the practices for the participants. The intervention had no standard treatment or action that was offered to all participants except the CGA described above. We instructed the teams to follow their clinical judgement and to individualise treatment and follow-up intervals. Control practices performed care as usual. We did not gather information about what specific actions the participants received.
Outcomes
Primary outcome was number of hospital care days.
Secondary outcomes were number of hospital care episodes and number of outpatient visits, health care costs and mortality. In analyses, outpatient visits were also subdivided into visits in emergency care, primary care and other outpatient care.
Using the unique 10-digit Swedish personal identity numbers, we linked patient data to the Care Data Warehouse in Östergötland (CDW) and Cost Per Patient database (CPP). The data warehouse includes all healthcare contacts (inpatient, outpatient and primary care) for both private care and public care, and the cost database includes costs for all contacts within public care. For an adequate assessment of cost data, all costs collected from CPP were adjusted to the price level of 2019 with an increase of historical costs with 3% per year. We converted costs to euros (one € = 10 SEK). Costs were missing in 11% of all contacts, mainly because of care given by private primary care providers. For these contacts, healthcare costs were imputed with the average cost calculated per contact type for existing data. Contact types were inpatient care, physician visit, and visit with a professional other than a physician. Less than 0.03% of contacts with missing costs were related to inpatient care. We also collected the number of co-morbidities from CDW as defined by the Royal College of Surgeons Charlson Score [29].
Statistical analysis
We compared baseline characteristics; risk score, age and gender, between the intervention group and the control group. All outcomes were analysed according to intention to treat. Continuous data were analysed with t-test and categorical data were analysed with χ2-test. We used generalised linear mixed models to compare the intervention group and control group regarding all outcomes, controlling for baseline covariates. We expected that two patients in the same primary care centre were likely to have more similar rates of the different outcomes than two patients in randomly different primary care centres. Therefore, our models included primary care centre as a random intercept effect. Risk score, age and gender were used as fixed covariates. For healthcare use data, we selected the distribution for our regression models based on an assessment of Pearson residuals. Since we found evidence of over-dispersion in healthcare use data, we preferred a negative binomial model with a log-link over a Poisson model. For cost data, we used a normal distribution with an identity link.
For healthcare use data, we calculated event rates for the two groups from actual data. From these event rates, absolute risk reductions for the intervention group were calculated. The relative risk reductions (RRRs) for the intervention group were analysed by the multivariable mixed count data models. For cost data, mean costs and mean differences between the groups were calculated from actual data. Adjusted mean differences were analysed by the multivariable mixed cost models.
We analysed mortality (1-overall survival) by the Kaplan-Meier method with any cause of death as an event and patients were censored by the end of follow-up 31 December 2019. Hazard ratio (HR) with 95% CIs between the intervention group and control group were estimated by multivariable Cox regression. The Cox regression was adjusted for the baseline covariates, risk score, age and gender.
For the primary outcome, the level of significance was set at 0.05 (two-tailed). Despite multiple secondary outcomes, we decided to keep the level of significance of 0.05 in order to detect meaningful differences. Therefore, these analyses should be interpreted with care. All analyses, except for Kaplan-Meier, were performed using SPSS version 25.0 (IBM Corp., Armonk, NY). The Kaplan-Meier plot was produced in R version 3.5.2 (R Core Team, Vienna, Austria).