The DACOS was a nationwide case series study with a pre-test and post-test design. We collected de-identified data extracted from electronic and/or paper-based medical records at eight clinics, including private and public hospital clinics. We identified patients with obesity who had a body mass index (BMI) of 30 kg/m2 or higher, referred for clinical obesity services, and followed for at least 6 months and up to 24 months, from their initial review between 2015 and 2020. There were no specific exclusion criteria for patients identified as eligible. To date, data from 273 patients across eight clinics in Australia have been extracted. For this study, we selected records for patients aged 16 years and over for whom weight loss could be estimated at 6 and 9 or 12 months after first clinic visit.

The data extracted from patients for this study included demographic information, anthropometric measurements, medical history, comorbidities, procedures, and current medications as part of routine clinical practice. The data also included information on services provided, such as the use of continuous positive airway pressure devices, the use of diet replacement products, the use of pharmacotherapies for weight management, and whether bariatric surgery was received.

The primary outcome was percentage total weight loss (%TWL) from baseline to 6 months and 9 or 12 months. Data were recorded at baseline and then at 3-month intervals to 24 months. Although few patients visited the clinic at such regularly spaced times, we sought to extract data from the first patient visit within one month of the target date. Patients’ current weight was missing whenever no visit was made during that time period or a visit was made but patient weight was not recorded. Where weight was missing for one or more visits, but available from both an earlier and later time point, we used linear interpolation to impute the missing weights. When the patient’s final weight measurement was at 9 months, we included their weight at 9 months in the 12-month analysis. Summary statistics are also provided for %TWL, change in BMI, and the proportion of patients achieving weight loss targets (defined as at least 5% of total body weight for non-surgical patients and at least 15% of total body weight for those who received surgical interventions).

We considered measurements which would typically be available to general practitioners when discussing referral to weight management clinics with their patients. These included demographics (age, sex), body mass index, smoking and alcohol status, mental and physical health history, number of medications, and current use of a continuous positive airway pressure (CPAP) machine. To keep the scoring system as simple as possible while incorporating the most important information, predictors were expressed as categorical variables with just two or three categories.

We aimed to develop a tool which could aid decision planning around referral to an obesity care clinic using information readily accessible to patients and their general practitioners. Predictors were selected according to statistical evidence, effect size, and clinical judgment. For simplicity, we only used whole number weightings.

The most important test of the decision aid was whether it could accurately predict future weight loss. We compared the %TWL predicted by the tool with actual %TWL recorded by the individuals in our data set and report the proportion of predictions that were accurate within ±5%. We also defined a target weight loss of 5% for non-surgical patients and 15% for those undergoing bariatric surgery and checked how well the tool could predict the achievement of these targets. Accuracy is reported in terms of sensitivity and specificity.

To ensure compatibility across clinics, we developed standard definitions of variables (data dictionary) and converted all incoming data into a single common format. We checked the distribution of each variable using frequency counts (for categorical variables) or histograms (for numeric variables) and checked the relationships between pairs of variables using cross-tabulations, side-by-side boxplots, and scatterplots as appropriate. We referred any questions to the data provider (clinical lead at each site) for verification.

As the participants were clustered by clinic, participants within the same clinic may be more similar and have more similar results than participants from different clinics. We visually inspected for outlier clinics using side-by-side boxplots and tested for outlier clinics by fitting clinic as a fixed effect in the regression models. As bariatric surgery is a strong predictor of weight loss and its availability differed between clinics, the confounding effect of bariatric surgery was adjusted for by including it in the model.

Even after adjustment for the use of bariatric surgery, variations in %TWL were observed between clinics. Therefore, all regression models include clinic as a random effect, allowing appropriate adjustment for clinic-based clustering when reporting evidence of association between the predictor variables and the outcome. We modeled clinic as a random effect rather than a fixed effect as fixed effects would compromise the confidentiality of participating clinics and undermine the generalizability of the resulting tool.

For each potential predictor, a mixed effect linear regression model of %TWL adjusted for the use of bariatric surgery (fixed effect) and clinic (random effect), was fit. These models provide estimates of both clinical effect size (the regression coefficient and 95% confidence intervals) and statistical evidence of association (as p-values). Pseudo R2 [15] statistics are also reported.

Review of the standardized residuals and Cook’s distances for each individual in the data set confirmed the presence of high and low outliers and some influential observations. Weight loss can be highly variable between individuals and challenging to accurately model. One outlier, for example, was an individual who lost 1/3 of their body weight from baseline without surgery, while another individual managed to gain a little weight despite bariatric surgery. Although outliers, there is no reason to doubt the validity of these data. We used 95% winsorization to limit the influence of extreme observations during model fitting, but all model testing was conducted on the original observations.

Candidate predictors which displayed a clinical meaningful effect (±2% weight loss), any statistical evidence of association (p<0.20) or which were deemed to be clinically meaningful predictors were combined in a single multivariate mixed model that provides the basis for the decision aid tool.

To produce the decision aid tool, we rounded the regression coefficients of the full model to the closest integer. The success of the prediction is reported as the proportion of predictions falling within 5 percentage points of actual %TWL and the sensitivity and specificity of the prediction in meeting weight loss targets.

As there was no specific hypothesis testing, choice of sample size was somewhat subjective. Clearly the larger the sample size, the better we would be able to specify the predictive model. A minimum sample target of 30 participants (each with complete data) from each of eight sites would have at least 80% power to detect a 0.25 standard deviation decrease in mean %TWL (from baseline to follow-up) assuming the intra-class correlation is 0.05 or less. (GLIMMPSE online calculator https://glimmpse.samplesizeshop.org).

As an external validation, we compared the %TWL predicted by our DACOS tool with the %TWL predictions from the American College of Surgeons Metabolic and Bariatric Surgery Accreditation and Quality Improvement Program (MBSAQIP) surgical risk/benefit calculator (Supplementary 2).