Our initial sample included 187 patients with ARFID seeking virtual-FBT treatment from September 2020 to May 2023, who received at least 20 weeks of treatment. Seventy-six percent of these patients required weight restoration; those who did not require weight restoration were excluded from the analysis (n = 44). The analytical sample was further reduced because weight data was not available at week 0 (i.e., first week of treatment) (n = 2) or week 20 (n = 10), or because EBW was not available (n = 15).

Our final analytical sample was N = 130. Patients ranged from 5 to 29 years old (M = 14.3, SD = 4.1) and were primarily white (n = 91, 70.0%) and cisgender girls/women (n = 75, 57.7%). Average percent EBW at admission was 84.6% (SD = 7.4%). Patients, or caregivers for minors, gave informed consent for treatment data to be analyzed and disseminated for research purposes. The analysis of patient treatment outcomes was also reviewed by Western Institutional Review Board (WIRB, Puyallup, WA), an independent ethics committee. WIRB determined the evaluation of our patient outcomes does not meet the definition of human subjects research.

Patients were enrolled in a virtual ED treatment program that uses an enhanced FBT approach. The enhanced approach consists of the conventional treatment team (e.g., family therapist, registered dietitian, and medical provider), and also includes a family mentor and peer mentor. A family mentor is a caregiver who has previous experience of caring for a loved one undergoing ED treatment, and a peer mentor is an individual who has recovered from an ED. Both the family mentor and peer mentor serve as additional support for the family and patient throughout treatment. Sessions are conducted via a HIPAA compliant telehealth platform with caregivers, patients, or both. Further details on the treatment approach and effectiveness are described in detail elsewhere [29, 30].

Weight was measured at home, by a family member who received in-depth training from the treatment team on weight monitoring. Patient's weight was checked two times a week with minimal clothing, after voiding, and prior to food or beverage consumption. Family members received automated prompts to enter weight into the electronic health record. EBW was determined using an individualized approach and was set by the patient's registered dietitian [31]. Dietitians calculated EBW by using each patient’s age-adjusted body mass index (BMI) and growth charts to determine where patient percentile BMI was trending before onset of the eating disorder. In addition, dietary intake, eating behaviors, physical activity patterns, medical data (vitals, blood work) and menstruation history (for those who menstruate) were used to establish EBW.

We defined weight restoration based on previous literature, such that a patient was considered weight restored if they reached 95% of their EBW by 20 weeks of treatment [16, 21, 24], a timeframe that generally aligns with how end of treatment is defined in FBT clinical trials. [16, 21,22,23,24].

First, to describe our sample, we used a combination of t-tests and linear regressions. Weight gain over time was assessed using two multilevel linear models, one to fit weight in pounds, the other fit weight as a proportion of EBW. Both models were identical except for the outcome variable, and included the log treatment week as a term. The models also included random intercepts and slopes on log treatment week.

We used ROC analysis in each week of treatment using weight gain as the predictor value to predict weight restoration at 20 weeks of treatment. The weight gain cutpoint chosen for each week was the one that maximized the sum of sensitivity and specificity. We report the area under the curve (AUC), its confidence intervals, and the sensitivity and specificity. We performed the analysis using weight change in pounds as well as the percent weight change from admission as the predictor and achieved similar results.

In ROC analysis, there is a binary outcome variable (in this study, weight restoration by 20 weeks), and predictor value (weight gain in a particular week) that can be used to distinguish between patients who reach weight restoration and those who do not. The cutpoint is a number that is meant to cleanly separate individuals into these two categories based on the predictor value. If a good cutpoint is found, it can be used in a clinical setting to determine whether progress early on in treatment is likely to end in weight restoration for a particular patient. For example, a weight gain cutpoint determined by the ROC analysis of 6.2 pounds in week 5 means that any patient who gains more than 6.2 pounds by week 5 is predicted to be weight restored at the end of treatment, and any patient who gains less than that is predicted not to. In this example, the cutpoint that maximizes the sum of the sensitivity (true positive rate) and specificity (true negative rate) is chosen. The ROC curve is the plot of the sensitivity against the specificity for each possible cutpoint value. The AUC characterizes how well the predictor value is able to categorize patients, regardless of the specific cutpoint. Because the ROC curve is built from a finite sample of data and possible cutpoints, the AUC we calculate is an estimate of the true AUC, and comes with uncertainty, measured by confidence intervals.

The cutpoint that maximizes the sum of sensitivity and specificity is the most common method used in the ED literature, although other methods exist. For example, if maximizing accuracy (the percentage correctly classified) was more important for the analysis, it would result in a cutpoint of 3.7 pounds in week 5. Maximizing the F1 score (a measure of accuracy based on precision and recall) would result in a cutpoint of 1 pound in week 5. The cutpoint R package contains 14 possible ways of estimating the cutpoint, and the possible cutpoints for the ROC analysis for week 5 ranged from 1 to 7.7 pounds (M = 4.9, SD = 2.1).

Lastly, we used logistic regression in each week of treatment with a set of coefficients (the patients’ %EBW at admission, the weight gain by that week, and the interaction between the first two variables) that would predict whether a patient would be weight restored at 20 weeks of treatment. Similar to the ROC analysis, we report the AUC, its confidence intervals, and the sensitivity and specificity. These models outperformed their ROC counterparts. We report the estimated probabilities of a patient reaching 95% EBW as a function of the patient’s %EBW at admission and their weight gain in a particular week (see Table 2).

All analyses were conducted with R version 4.3.0 [32]. Fitting was done using base R (lm and glm), lme4 version 1.1-33 [33], and lmerTest packages version 3.1-3 [34]. ROC analysis was performed using pROC version 1.18.0 [35] and cutpointr version 1.1.2 [36]. Model coefficients are presented with their corresponding standard errors. The targets package version 0.14.2 [37] was used for project management.