The TRIAGE-trial
The TRIAGE-trial was set up to determine the effectiveness and safety of a nurse-led triage system that assigns low-risk patients from an ED to the GP. A single-centre cluster randomised trial was performed with weekends and bank holidays (hereinafter called weekends) serving as units of randomisation and patients as units of analysis. The trial ran from 01/03/2019 to 30/12/2019. During intervention weekends, patients were assigned to a particular care setting but they had the possibility to refuse: low-risk patients were considered as candidates for primary care and were assigned to the GPC, while patients in need of more urgent or advanced care were assigned to the ED. Control weekends are not of interest in this study, as the advice was not communicated to patients and they all remained at the ED. The trial was executed in the ED of the Belgian general hospital ‘AZ Monica’ and the adjacent GPC ‘Antwerpen Oost’. The surrounding area has citizens from a variety of ethnicities and consists of both middle income and socially deprived neighbourhoods. The Belgian healthcare system is mainly organised as a fee-for-service system and is characterised by free choice and open access for patients to all medical services.
The triaging of patients was done using a locally developed extension to the MTS (eMTS). The eMTS contains the entire MTS version 3.6, one of the main triage systems used worldwide [22] The system is a tool for prioritisation in the ED, but previous studies have also used it to relocate patients. They have illustrated that the system presents an acceptable validity [13, 15, 16] The MTS is a five-level triage system and consists of 53 presentational flowcharts. Each flowchart consists of discriminators, eventually leading to an urgency category ranging from level one (immediate care necessary) to level five (non-urgent). In the adapted version, 44 flowcharts were extended with GP risk discriminators whenever the urgency category was four or five. If such discriminator was present, patients were assigned to the ED [12].
Outcome measures
This study is a secondary analysis of the TRIAGE-trial. The predefined primary outcome is the proportion of patients that were assigned to the GPC but refused. They were treated at the ED, despite the advice to go to the GPC. The secondary outcomes of this article (not predefined) are the determinants of non-compliance and the impact on the costs.
Data collection
The following patient characteristics were collected and used in this study: age; sex; patient lives nearby (within the four communities covered by the GPC); and socio-economic status (whether patients receive an increased reimbursement or not, which is predominantly determined by an upper bound on household income). Information on the patients’ race, education, primary language, or previous experience with the ED/GPC was unavailable. The eMTS flowchart (53 flowcharts combined into 15 categories), urgency level, type of admission to the ED (walk-in or ambulance) offered information on the patients’ presentation. Other confounders, such as their baseline health or patient distress were, however, not collected. The time period (day, evening, or night), subjective crowding at the ED (quiet, normal, or busy), and anonymous ID of the triaging nurse were also registered. All 22 nurses who performed a triage were numbered. The data from the ED and GPC were linked through their pseudonymised national insurance number using iCAREdata, which is a database for medical records during OOH care [23, 24].
After the trial, the patient-level costs of treatment at the ED and GPC were received from the billing department of AZ Monica and the GPC respectively. Both settings make use of a fee-for-service system. The data consisted of the (pseudo)nomenclature codes of all medical services provided to the patients, as captured on the invoice. The codes were grouped to reflect different cost categories: consultation fees, medical imaging, clinical biology, technical procedures, medication, hospital lump sums, and non-refundable items. Data on medical imaging or clinical laboratory tests ordered by the GP were not available at the patient level. The category non-refundable items consists of various articles at the request of the patient (e.g., a toothbrush) or necessary for their medical care (e.g., crutches). Medication costs only include medicines given to the patient during a consultation and not the prescriptions given to them. The various cost categories (except consultation fees) give insight into the treatment people received, as prices for medical services are similar for both the GPC and the ED. Consultation fees are predetermined. In Belgium, ED physicians and GPs receive different consultation fees, depending on the medical specialty of the physician and on the arrival time of the patient. For instance, under the current remuneration scheme, consultations during the night are more expensive at the GPC than at the ED, while the opposite occurs during daytime. The data also show the proportion of the invoice paid by the patient and by the national health insurance. The division is predetermined as well and depends on whether the consultation is with or without referral and on the socio-economic status of the patient. Consultations at the ED without referral require a higher share of co-payment from the patient [25] Due to anonymity, data were matched with the medical records from above on the basis of sex, birth year, postal code, and time. For nine patients (1.2%) no invoice could be matched. Ten (1.3%) patients were hospitalized. They were excluded from the financial analysis, as only their ambulant costs were available.
Statistical methodology
The determinants of non-compliance were first considered using a bivariate analysis. The proportions of patient characteristics, patient status, eMTS components, and variables related to the time of admission were compared between refusers and non-refusers. Bivariate logistic regressions were used to calculate odds ratios. The data were analysed using JMP pro® version 14. Those variables found significant at an alpha of 0.10 were considered significant and incorporated in the multivariate analysis. A significance level of 0.10 was used since the smaller dataset and consequently larger standard errors were unlikely to produce more significant results.
A similar bivariate analysis was performed on the costs. The mean costs of compliers and non-compliers were compared using a T-test for unequal variances. A two-sided F-test for equal variance indicated this was most appropriate. A distinction was made between the fraction of the invoice paid by the national insurance and the fraction paid by the patient, as well as between the period of the day.
The multivariate analysis consisted of a chi-square automatic interaction detection (CHAID) decision tree [26, 27] This methodology is commonly used for building prediction algorithms for a target variable and can deal with large, complicated datasets in an efficient manner, without imposing a complicated parametric structure. This method classifies the population into branch-like segments that construct an inverted tree with a root node, internal nodes, and leaf nodes [28] For this article, a decision tree based on Bonferroni-Holm corrected chi-squared tests was constructed with as target variable the likelihood of refusing the advice to visit the GPC. The independent variables were all patient characteristics, subjective crowding, period of the day, flowchart category and nurse ID. A 10-fold cross validation was used to evaluate the model. The CHAID-analysis was performed using IBM SPSS® version 27.