Introduction
Existing epidemiological data support conflicting views (Keski-Rahkonen, Reference Keski-Rahkonen2021; Wilfley, Wilson, & Agras, Reference Wilfley, Wilson and Agras2003) of the natural course of binge-eating disorder (BED), the most prevalent eating disorder in the United States (Hudson, Hiripi, Pope, & Kessler, Reference Hudson, Hiripi, Pope and Kessler2007; Taylor, Caldwell, Baser, Faison, & Jackson, Reference Taylor, Caldwell, Baser, Faison and Jackson2007; Udo & Grilo, Reference Udo and Grilo2018). Community-based studies using retrospective assessments, including three nationally representative samples, suggest a protracted course for BED, with mean durations ranging from 7 to 16 years (Hudson et al., Reference Hudson, Hiripi, Pope and Kessler2007; Pope et al., Reference Pope, Lalonde, Pindyck, Walsh, Bulik, Crow and Hudson2006; Taylor et al., Reference Taylor, Caldwell, Baser, Faison and Jackson2007; Udo & Grilo, Reference Udo and Grilo2018). By contrast, several small prospective, community-based studies potentially suggest a more transient course (Brown, Klein, & Keel, Reference Brown, Klein, Keel, Smolak and Levine2015). For example, in an eight-year longitudinal study (Stice, Marti, & Rohde, Reference Stice, Marti and Rohde2013) of incident eating disorder cases in girls ages 12–15 years at baseline, 14 of 15 (93%) participants who met DSM-5 criteria for BED at some point during the study had gone at least a month without meeting full criteria for BED by 2-year follow-up. (Results for DSM-IV BED in the same sample were similar [Stice, Marti, Shaw, & Jaconis, Reference Stice, Marti, Shaw and Jaconis2009]). In a longitudinal study (Fairburn, Cooper, Doll, Norman, & O'Connor, Reference Fairburn, Cooper, Doll, Norman and O'Connor2000) of 48 women with DSM-IV BED who were age 16–35 years at baseline, seven (18%) of the 40 participants with follow-up data met criteria for an eating disorder at 5-year follow-up, including four (10%) participants who still had BED. In a study (Cachelin et al., Reference Cachelin, Striegel-Moore, Elder, Pike, Wilfley and Fairburn1999) of 31 women with DSM-IV BED and a mean age of 31.8 years, 11 (52%) of the 21 participants with follow-up data met full criteria for DSM-IV BED at six months follow-up, and the remaining 10 (48%) experienced less than one binge-eating episode/week. However, no participants experienced more than two months without binge-eating episodes at either 3-month or 6-month follow-up.
All of these prior studies have limitations: the retrospective studies are subject to recall bias, whereas the prospective studies have been extremely small, excluded males, and either lacked longer-term follow-up (Cachelin et al., Reference Cachelin, Striegel-Moore, Elder, Pike, Wilfley and Fairburn1999) or included participants who were not representative of individuals with BED in terms of age of onset or BMI (Fairburn et al., Reference Fairburn, Cooper, Doll, Norman and O'Connor2000; Stice et al., Reference Stice, Marti and Rohde2013). More specifically, mean age of onset was approximately 17 years in Fairburn et al. (Reference Fairburn, Cooper, Doll, Norman and O'Connor2000) and Stice et al. (Reference Stice, Marti and Rohde2013), which is earlier than the 23–25 years reported in large nationally representative, community-based studies (Kessler et al., Reference Kessler, Berglund, Chiu, Deitz, Hudson, Shahly and Xavier2013; Taylor et al., Reference Taylor, Caldwell, Baser, Faison and Jackson2007; Udo & Grilo, Reference Udo and Grilo2018). Similarly, of 36 participants with BED and recorded BMI in Fairburn et al. (Reference Fairburn, Cooper, Doll, Norman and O'Connor2000), only 22% had a BMI of 30 or greater, in contrast to 68% of participants with 12-month BED in the most recent nationally representative study (Udo & Grilo, Reference Udo and Grilo2018). Thus, larger multi-year prospective studies that include men, middle-aged adults, and more individuals with higher BMI are needed to better understand the natural course of BED.
Further, little is known about the demographic and clinical predictors of the natural course of BED. One study (Cachelin et al., Reference Cachelin, Striegel-Moore, Elder, Pike, Wilfley and Fairburn1999), which used multiple linear regression to identify predictors of binge-eating frequency at 3- and 6-months follow-up in 21 participants, found that older age predicted greater frequency of binge eating at 3-month, but not 6-month, follow-up.
Thus, in this study, our primary aim was to characterize the course of BED, including time to remission and time to relapse after remission, based on prospective data from a longitudinal study (Hudson et al., Reference Hudson, Lalonde, Coit, Tsuang, McElroy, Crow and Pope2010) of adults with BED. Our secondary aim was to apply machine learning methods to examine whether common, easily collected demographic and clinical measures from baseline could be used to predict time to remission. Note that the goal of prediction is fundamentally different from the goal of ‘explanation,’ or identification of causal structure (Diemer, Hudson, & Javaras, Reference Diemer, Hudson and Javaras2021; Yarkoni & Westfall, Reference Yarkoni and Westfall2017).
Methods
Participants
At baseline, 156 individuals (99 probands and 57 relatives) with current DSM-IV BED were selected for a longitudinal follow-up study (Hudson et al., Reference Hudson, Lalonde, Coit, Tsuang, McElroy, Crow and Pope2010) from among participants in a family study of BED (Hudson et al., Reference Hudson, Lalonde, Berry, Pindyck, Bulik, Crow and Pope2006). Previous investigations using the family study data have examined the (retrospectively assessed) age of onset and diagnostic stability of BED (Pope et al., Reference Pope, Lalonde, Pindyck, Walsh, Bulik, Crow and Hudson2006), the familial aggregation and heritability of BED (Hudson et al., Reference Hudson, Lalonde, Berry, Pindyck, Bulik, Crow and Pope2006; Javaras et al., Reference Javaras, Laird, Reichborn-Kjennerud, Bulik, Pope and Hudson2008a), and the co-occurrence of psychiatric and other medical disorders with BED (Javaras et al., Reference Javaras, Pope, Lalonde, Roberts, Nillni, Laird and Hudson2008b); in addition, a previous investigation using the longitudinal study data has examined the onset of metabolic syndrome components in BED (Hudson et al., Reference Hudson, Lalonde, Coit, Tsuang, McElroy, Crow and Pope2010). Notably, probands in the family study were required to have body mass index (BMI) greater than 25 kg/m2 (women) or 27 kg/m2 (men), because a major aim of that study was to examine the familial aggregation of BED controlling for familial aggregation of higher BMI.
Of the 156 individuals selected for longitudinal follow-up, 19 (12 probands and 7 relatives) could not be assessed at 2.5-year follow-up, and an additional eight participants (5 probands and 3 relatives) could not be assessed at 5-year follow-up, corresponding to follow-up rates of 87.8% (2.5 years) and 82.7% (5 years), respectively.
Procedures and materials
Participants completed interviews and questionnaires at baseline and at 2.5-year and 5-year follow-up timepoints. Baseline assessments were conducted from October 2002 to July 2004. The longitudinal study protocol specified that the 2.5- and 5-year follow-up assessments would occur 24–36 months and 54–66 months, respectively, after the baseline assessment. There were three deviations from this timeline (two participants interviewed at 36.1 and 36.3 months for the 2.5-year follow-up, and one participant interviewed at 71.5 months for the 5-year follow-up).
Participants were interviewed by the same of three psychiatrists (including JIH and HGP) at all three timepoints. At each timepoint, the interviewer administered the Structured Clinical Interview for DSM-IV (SCID-IV) (First, Spitzer, Gibbons, & Williams, Reference First, Spitzer, Gibbons and Williams2001), which included assessment of BED diagnostic status at that timepoint. At follow-up timepoints, the interviewer also gathered detailed information on any changes in BED symptoms since the last assessment, using questions based on the SCID-IV, together with the participants' retrospective chronology of these changes, which was recorded to the nearest month using a timeline that was reviewed with the participant. These assessments yielded information on: (a) BED diagnostic status (i.e. full DSM-IV BED, four sub-threshold forms of BED, and no BED; see Table 1) at each timepoint; and (b) changes in BED diagnostic status throughout the follow-up interval, along with the time at which the changes occurred, provided that the new status lasted for at least 3 months. The baseline SCID-IV information on other lifetime psychiatric disorders was used to create baseline predictors of time to remission from BED.
Table 1. Definitions of BED diagnostic statusa
Abbreviation: BED, binge-eating disorder.
a A (change in) status was recorded only if the new status was experienced for 3 or more months.
b Meets criteria for DSM-5 BED.
Additionally, at baseline, participants completed several questionnaires tapping constructs relevant to BED, including the Eating Disorder Inventory–2 (EDI-2; Garner, Reference Garner1991), the Three Factor Eating Questionnaire (Stunkard & Messick, Reference Stunkard and Messick1985), and the Barratt Impulsiveness Scale–11 (Patton, Stanford, & Barratt, Reference Patton, Stanford and Barratt1995). Subscale scores for these questionnaires were used as predictors of time to remission from BED.
This study was approved by the Mass General Brigham Institutional Review Board, and all participants provided informed consent after the study procedures had been fully explained.
Analyses
We used R version 4.3.2 to perform all data analyses.
Below, remission refers to the first occurrence of no BED within the follow-up period, and relapse refers to the first occurrence of sub-threshold or full BED within the follow-up period following remission.
To graphically display how BED diagnostic status differed at the three assessment timepoints (baseline, 2.5 years, and 5 years), we used the ggplot function from the ggalluvial package. To visualize time to remission and time to relapse after remission, we used the survfit function from the survival package to calculate Kaplan–Meier curves with confidence intervals that account for clustering by family. To investigate the relationship between individual variables (e.g. duration of BED) and time to remission, we used Cox proportional hazards models (i.e. Cox regression) with robust variance estimation that accounted for clustering of observations by family.
For the main machine learning analyses, we used 38 demographic and clinical variables (see online Supplemental Materials Table S1) from the baseline timepoint to predict the time to remission for the 137 participants with follow-up data. The predictors were selected based on statistical and, more importantly, practical considerations: our goal was to pose a well-defined prediction problem using common and easily collected demographic and clinical measures from baseline as predictors of time to remission. We considered four machine learning methods (standard Cox regression; penalized Cox regression with lasso or elastic-net penalties; and random survival forests), which represent a combination of regression-based and tree-based methods similar to those we have used previously (Hudson et al., Reference Hudson, Hudson, Kanayama, Schnabel, Javaras, Kaufman and Pope2023), and that have been used in a prior study predicting treatment outcomes in a randomized controlled trial in BED (Forrest, Ivezaj, & Grilo, Reference Forrest, Ivezaj and Grilo2023). Because 1.0% of the predictor values were missing, we first created five imputed datasets. Then, for each imputed dataset, we performed 100 replicates of 5-fold cross-validation with pre-selected seeds to examine the performance of the four methods. We tuned hyper-parameters for penalized Cox regression and random survival forests via nested 5-fold cross-validation. To obtain a summary of our two chosen performance statistics (the concordance index and R 2 based on Schmid's robust estimate of prediction error [Rahman, Ambler, Choodari-Oskooei, & Omar, Reference Rahman, Ambler, Choodari-Oskooei and Omar2017; Schmid, Hielscher, Augustin, & Gefeller, Reference Schmid, Hielscher, Augustin and Gefeller2011]), we then calculated the mean of the 500 performance statistics (5 folds times 100 replicates) for each method for a given imputed dataset, and then took the median of the mean performance statistics for each method across imputations (Marshall, Altman, Holder, & Royston, Reference Marshall, Altman, Holder and Royston2009; Wood, Royston, & White, Reference Wood, Royston and White2015). We also performed six sensitivity analyses. More details of the machine learning analyses are provided in the online Supplemental Materials.
Results
Sample characteristics
Table 2 presents summary statistics for participants with follow-up data (n = 137). For these participants, the mean reported age was 47.2 years (standard deviation [s.d.] = 13.2); 78.1% reported their sex as female, and 21.9% reported their sex as male; and 87.6%, 11.7%, and 0.7% reported identifying as White non-Hispanic, Black non-Hispanic, and White Hispanic, respectively. Gender identity was not assessed. Regarding education, 23.4% reported completing high school (or equivalent) or less education, 31.4% reported completing some college, 17.5% reported completing a 2- or 4-year college, 10.9% reported completing some graduate or professional school, and 16.8% reported completing graduate or professional school. Information on income or net assets was not available. The mean duration of BED was 22.3 years (s.d. = 15.4), and the mean BMI was 36.1 kg/m2 (s.d. = 7.5). Also, only a small proportion of participants (n = 8, 5.8%) had a history of DSM-IV anorexia nervosa (excluding the amenorrhea criterion) or bulimia nervosa at baseline.
Table 2. Summary statistics for baseline and 2.5- and 5-year follow-up variables
Abbreviations: BED, binge-eating disorder; BMI, body mass index.
a Summary statistics presented for n = 137 participants who contributed to follow-up analyses. For continuous variables, summary statistics are mean, (standard deviation), [range]. For categorical variables, summary statistics are number and (percent) belonging to each category.
b BMI suppression defined as highest lifetime BMI – current BMI.
c BED family history defined as the proportion of the participant's interviewed relatives who met criteria for (full) DSM-IV BED based on the SCID-IV at baseline.
d Past anorexia nervosa defined as meeting criteria for DSM-IV anorexia nervosa (excluding the amenorrhea criterion).
e Past bulimia nervosa defined as meeting criteria for DSM-IV bulimia nervosa.
f Any lifetime substance use disorder defined as meeting criteria for any of the following lifetime DSM-IV disorders based on the SCID-IV: alcohol abuse; alcohol dependence; drug abuse; drug dependence.
g Any lifetime major anxiety disorder defined as meeting criteria for any of the following lifetime DSM-IV disorders based on the SCID-IV: panic disorder without agoraphobia; panic disorder with agoraphobia; agoraphobia without panic disorder; social phobia; generalized anxiety disorder.
h Any lifetime major mood disorder defined as meeting criteria for any of the following lifetime DSM-IV disorders based on the SCID-IV: bipolar I disorder; bipolar II disorder; major depressive disorder.
Natural course of BED
We first present statistics calculated from participants' diagnostic status at the baseline and 2.5- and 5-year follow-up timepoints (see Fig. 1). The percentage of participants with full BED decreased over time, and the percentages with sub-threshold BED and no BED increased over time. Specifically, at the 2.5-year follow-up, 61.3%, 23.4%, and 15.3% of assessed participants had full BED, sub-threshold BED, and no BED, respectively, and, at the 5-year follow-up, 45.7%, 32.6%, and 21.7% of assessed participants had full BED, sub-threshold BED, and no BED, respectively. Notably, there was no diagnostic cross-over (e.g. from BED to anorexia nervosa or bulimia nervosa) at either of the follow-up timepoints. Additionally, for participants assessed at both follow-up timepoints, 35.0% of those experiencing remission (i.e. no BED) at 2.5-year follow-up had relapsed to either full or sub-threshold BED at 5-year follow-up.
Figure 1. Alluvial plot for diagnostic status at baseline and 2.5- and 5-year follow-up timepoints. The figure illustrates diagnostic status at each timepoint, as well as transitions among these statuses. The non-italicized numbers refer to timepoint percentages (i.e. out of all 156 participants, what percentage belonged to each diagnostic status at a given timepoint); for the sub-threshold BED status, the numbers in the parentheses further represent the percentage with each type of sub-threshold BED (i.e. [A | B | C | D]). The italicized numbers refer to transition percentages (i.e. out of all participants with a given diagnostic status at a given timepoint, what percentage transitioned to each diagnostic status at the next timepoint). Abbreviations: BED, binge-eating disorder; sub-BED, sub-threshold binge-eating disorder.
We next present statistics calculated from participants' diagnostic status transition data across the 5-year follow-up period (see Fig. 2). The median time to remission (i.e. no BED) exceeded 60 months. The median time to relapse (i.e. to full BED or sub-threshold BED) after remission was approximately 30 months.
Figure 2. Survival curves for time to remission and time to relapse after remission during the 5-year follow-up period. Remission is defined as not having either (full) BED or sub-threshold BED for at least 3 months. Relapse is defined as having either (full) BED or sub-threshold BED for at least 3 months after experiencing remission. In the plots, the solid gray line is the Kaplan–Meier survival curve, and the dotted gray lines are the corresponding 95% confidence interval. The horizontal black line represents the median.
Abbreviation: BED, binge-eating disorder.
Predictors of time to remission
In the main analyses and the sensitivity analyses, none of the four machine learning methods had a summary measure of concordance better than random guessing for both performance statistics (see online Supplemental Materials for method performance in the main and sensitivity analyses).
Similarly, in standard Cox regression models for time to remission as a function of each baseline predictor (see online Supplemental Materials Table S1), no predictors would survive correction for multiple comparisons.
Discussion
We examined the natural course of BED, including time to remission and relapse, over a 5-year follow-up period for a community-based sample of men and women with BED (by DSM-IV criteria) whose ages ranged from 19 to 74 years and whose BMI ranged from 22.8 to 63.4 kg/m2. We also used two classes of machine learning methods to predict time to remission as a function of baseline demographic and clinical variables.
In our sample, BED status improved from baseline to 2.5-year follow-up, and from 2.5-year follow-up to 5-year follow-up. However, almost two-thirds (61.3%) of participants assessed at 2.5-year follow-up and almost half (45.7%) of participants assessed at 5-year follow-up still met criteria for full (DSM-IV) BED at those follow-up timepoints. Moreover, the percentages meeting the broader DSM-5 BED criteria at these follow-up timepoints would be even higher (i.e. 67.9% and 58.9%, respectively). Further, only a minority of individuals assessed at 2.5-year follow-up (15.3%) and 5-year follow-up (21.7%) experienced remission at those timepoints, respectively. Similarly, the diagnostic status transition data suggested that the median time to remission exceeds 5 years. Taken together, these results suggest that, under naturalistic circumstances, the majority of individuals with BED will experience some improvement in binge eating within a 5-year period, but achieving full remission within this time frame is not common – at least for prevalent cases of BED in adults with generally higher BMI. These results are consistent with findings from community-based studies using retrospective lifetime interviews (Hudson et al., Reference Hudson, Hiripi, Pope and Kessler2007; Pope et al., Reference Pope, Lalonde, Pindyck, Walsh, Bulik, Crow and Hudson2006; Taylor, Reference Taylor2015; Udo & Grilo, Reference Udo and Grilo2018), including the family study from which our participants derived (Pope et al., Reference Pope, Lalonde, Pindyck, Walsh, Bulik, Crow and Hudson2006), all of which suggest that BED tends to have a multi-year course. These results are less consistent with findings from two prior prospective, community-based studies with multi-year follow-up periods (Fairburn et al., Reference Fairburn, Cooper, Doll, Norman and O'Connor2000; Stice et al., Reference Stice, Marti and Rohde2013, Reference Stice, Marti, Shaw and Jaconis2009), as well as those from two multi-year prospective, community-based studies that followed adolescent girls (Glazer et al., Reference Glazer, Sonneville, Micali, Swanson, Crosby, Horton and Field2019) or pregnant young adult women (Knoph et al., Reference Knoph, Von Holle, Zerwas, Torgersen, Tambs, Stoltenberg and Reichborn-Kjennerud2013) with BED-like presentations based on questionnaires that assessed some, but not all, BED criteria. Although different definitions of remission preclude fully comparing these prospective studies' results to our own, these prior studies suggest that BED is more transient, with few individuals meeting full criteria for BED at follow-up. In contrast to our study, which sampled prevalent cases from among adults with higher BMI or at familial risk for higher BMI (that is, having a proband relative with higher BMI), the prior prospective studies of BED focused on incident cases (Stice et al., Reference Stice, Marti and Rohde2013, Reference Stice, Marti, Shaw and Jaconis2009) and/or on adolescent (Stice et al., Reference Stice, Marti and Rohde2013, Reference Stice, Marti, Shaw and Jaconis2009) or young adult (Fairburn et al., Reference Fairburn, Cooper, Doll, Norman and O'Connor2000) females only, which may account for the more favorable prognosis in these other samples – especially given some evidence that younger age predicts better natural outcomes (Cachelin et al., Reference Cachelin, Striegel-Moore, Elder, Pike, Wilfley and Fairburn1999). Also of note, outcomes in our naturalistic study were generally worse than 4-year outcomes in a longitudinal follow-up study that recruited participants from both specialty clinics and the community (Agras, Crow, Mitchell, Halmi, & Bryson, Reference Agras, Crow, Mitchell, Halmi and Bryson2009). Of the 104 participants with DSM-IV BED at baseline in this latter study, 82% no longer met criteria for an eating disorder at some point during the 4-year follow-up period, although a high proportion of these experienced a subsequent relapse (i.e. meeting criteria for any eating disorder). Notably, 70% of participants with BED reported outpatient or inpatient treatment for BED at baseline (Agras et al., Reference Agras, Crow, Mitchell, Halmi and Bryson2009), which likely exceeds the prevalence of specialty treatment for BED in our sample (Hudson et al., Reference Hudson, Hiripi, Pope and Kessler2007). Likewise, outcomes of our naturalistic study are considerably worse than 6-year follow-up outcomes for 68 women who had DSM-IV BED at the time they received inpatient treatment for eating disorders (Fichter, Quadflieg, & Gnutzmann, Reference Fichter, Quadflieg and Gnutzmann1998)
Outcomes in our naturalistic study were also generally worse than long-term outcomes for specific psychological interventions for BED in clinical trials, which is not surprising given evidence for the effectiveness of psychotherapy for BED in the long term (Hilbert et al., Reference Hilbert, Bishop, Stein, Tanofsky-Kraff, Swenson, Welch and Wilfley2012), coupled with the finding that the majority of individuals with BED in the community do not receive treatment for their eating disorder (Coffino, Udo, & Grilo, Reference Coffino, Udo and Grilo2019; Hay et al., Reference Hay, Ghabrial, Mannan, Conti, Gonzalez-Chica, Stocks and Touyz2020; Kessler et al., Reference Kessler, Berglund, Chiu, Deitz, Hudson, Shahly and Xavier2013). More specifically, after receiving cognitive-behavioral therapy or interpersonal psychotherapy for BED, a larger proportion of individuals reported sustained remission at 2 years post-treatment (Wilson, Wilfley, Agras, & Bryson, Reference Wilson, Wilfley, Agras and Bryson2010) than at 2.5-year follow-up in this study, and a smaller proportion of individuals continued to binge eat at least once weekly at approximately 4 years post-treatment (Hilbert et al., Reference Hilbert, Bishop, Stein, Tanofsky-Kraff, Swenson, Welch and Wilfley2012) than at the 5-year follow-up in this study. Less is known about the long-term effectiveness of pharmacologic treatment for BED, including lisdexamfetamine (Feltner et al., Reference Feltner, Peat, Reddy, Riley, Berkman, Middleton and Jonas2022; McElroy et al., Reference McElroy, Hudson, Ferreira-Cornwell, Radewonuk, Whitaker and Gasior2016), the only medication that has gained regulatory approval for marketing for BED in the United States and other countries. However, in a placebo-controlled trial to assess relapse prevention after achieving response to a medication, treatment with lisdexamfetamine was associated with a markedly lower rate of relapse of BED (3.7%) than treatment with placebo (32.1%) over a six-month period (Hudson, McElroy, Ferreira-Cornwell, Radewonuk, & Gasior, Reference Hudson, McElroy, Ferreira-Cornwell, Radewonuk and Gasior2017). Our results thus provide additional evidence that more widespread dissemination of psychotherapy and potentially efficacious pharmacologic treatment for BED would reduce the burden of disease due to BED (Moessner & Bauer, Reference Moessner and Bauer2017), with the caveat that it is not known whether results from treatment-seeking samples generalize to the broader population.
Notably, there was no diagnostic crossover (e.g. from BED to anorexia nervosa or bulimia nervosa) in our study. This result contrasts with that of some prior prospective, community-based studies, which found that crossover from BED to bulimia nervosa was relatively common in adolescent females (Allen, Byrne, Oddy, & Crosby, Reference Allen, Byrne, Oddy and Crosby2013; Stice et al., Reference Stice, Marti and Rohde2013, Reference Stice, Marti, Shaw and Jaconis2009). However, another prospective, community-based study in young adult women found that crossover from BED to bulimia nervosa was relatively uncommon (Fairburn et al., Reference Fairburn, Cooper, Doll, Norman and O'Connor2000). Collectively, these results suggest that the likelihood of crossover from BED to bulimia nervosa may decrease with age and with longer BED duration.
In our sample, approximately one-third (35.0%) of individuals experiencing remission at 2.5-year follow-up subsequently relapsed to either full or sub-threshold BED at 5-year follow-up. Further, based on the diagnostic status transition data, time to relapse after remission was approximately 2.5 years. Taken together, these results suggest that relapse after remission is relatively common – at least among prevalent cases of BED in adults with generally higher BMI. Although methodological differences make it difficult to compare our relapse results to those of prior studies with similar follow-up periods (Fairburn et al., Reference Fairburn, Cooper, Doll, Norman and O'Connor2000; Stice et al., Reference Stice, Marti and Rohde2013, Reference Stice, Marti, Shaw and Jaconis2009), our relapse results appear similar to some (Stice et al., Reference Stice, Marti and Rohde2013, Reference Stice, Marti, Shaw and Jaconis2009), but higher than others (Fairburn et al., Reference Fairburn, Cooper, Doll, Norman and O'Connor2000).
Regarding the machine learning analyses, which aimed to build prediction models of time to remission from baseline demographic and clinical variables, no machine learning methods strongly or consistently outperformed random guessing. These results suggest that the demographic and clinical measures we included are not strongly predictive of time to remission for BED in community-based samples. These findings are consistent with a prior prospective, community-based study that found no consistent predictors of objective binge-eating episodes at both 3- and 6-month follow-up (Cachelin et al., Reference Cachelin, Striegel-Moore, Elder, Pike, Wilfley and Fairburn1999). Our findings are also consistent with machine learning results from a randomized controlled trial of behavioral weight loss v. stepped care in BED, where area under the receiver operator characteristic curve values from cross-validation ranged from 0.49 to 0.51 for traditional logistic regression, elastic-net, and random forest models that predicted binge-eating abstinence during the last month of treatment as a function of treatment condition, demographic information, and baseline clinical characteristics (Forrest et al., Reference Forrest, Ivezaj and Grilo2023). Thus, the factors that contribute to a better or worse naturalistic course of BED – or to binge-eating cessation in the treatment context – remain to be fully elucidated. One possibility is that the natural course of BED is influenced by changes in the individual's familial, social, professional, or broader environment that are not related to the individual's baseline characteristics.
Overall, our study has several strengths. To the best of our knowledge, it is the largest prospective, community-based study of individuals with BED and is more than three times larger than prior multi-year prospective, community-based studies of BED (Fairburn et al., Reference Fairburn, Cooper, Doll, Norman and O'Connor2000; Stice et al., Reference Stice, Marti and Rohde2013). Further, it includes men, who were excluded from all prior prospective, community-based studies of BED (Cachelin et al., Reference Cachelin, Striegel-Moore, Elder, Pike, Wilfley and Fairburn1999; Fairburn et al., Reference Fairburn, Cooper, Doll, Norman and O'Connor2000; Stice et al., Reference Stice, Marti, Shaw and Jaconis2009). Additionally, the mean age of BED onset for our participants was approximately 25 years, which is similar to the mean age of onset in large, community-based retrospective studies; in contrast, the mean age of onset was approximately 17 years in prior multi-year prospective, community-based studies (Fairburn et al., Reference Fairburn, Cooper, Doll, Norman and O'Connor2000; Stice et al., Reference Stice, Marti and Rohde2013). Likewise, 77% of participants in our study had a BMI of 30 or above at baseline, which is not far from the analogous percentage (68%) in the most recent nationally representative study (Udo & Grilo, Reference Udo and Grilo2018); in contrast, the analogous percentage was 22% (Fairburn et al., Reference Fairburn, Cooper, Doll, Norman and O'Connor2000) or possibly even lower (Stice et al., Reference Stice, Marti and Rohde2013) in prior community-based studies. Additionally, the collection of diagnostic status transition data allowed us to use survival methods to examine the time to remission and the time to relapse after remission. Further, a host of clinically relevant variables were measured at baseline, thereby providing numerous predictors for machine learning analysis. Also, our use of a machine learning framework – through dividing the dataset into a training and test set, and fitting multiple classes of models – allowed for more rigorous assessment of the baseline variables' predictive value, compared to simply fitting a regression model with all predictors to the entire dataset.
Our study findings should be interpreted in light of certain limitations. First, we sampled prevalent, rather than incident, cases of BED, and more transient cases of BED may have been underrepresented in our sample due to length bias in sampling. One would expect this bias to be worse for probands than relatives, given that the former were selected for inclusion in the original family study based on their BED status. However, proband status was not associated with time to remission (estimated Cox proportional hazard ratio for relative v. proband [95% confidence interval]: 0.97 [0.54–1.74]), nor was BED duration (for 5-year increase: 0.97 [0.88–1.07]), which suggests that this source of bias may be small. Second, in our study, diagnosis was based on DSM-IV criteria, which raises the question of whether the results would generalize to BED by DSM-5 criteria, which requires only one binge-eating episode per week over three months, as opposed to two days per week with binge-eating episodes over six months in DSM-IV. Notably, within our family study sample, only three additional individuals met the broader DSM-5 criteria for BED at baseline (Hudson, Coit, Lalonde, & Pope, Reference Hudson, Coit, Lalonde and Pope2012), and including those individuals in the present follow-up study would thus have little impact on the findings as a whole. However, as noted above, using DSM-5 criteria to assess diagnostic status at follow-up timepoints would increase the percentages of individuals with (full) BED at those timepoints. Third, as implemented in many clinical drug trials, the same rater who interviewed the participant at baseline also interviewed them at the follow-up timepoints, to minimize measurement error due to intra-participant assessment variability over time, because our focus was detecting change in BED status. Thus, interviewers were not blinded to participants' prior BED status, which could have introduced bias (e.g. if interviewers were predisposed to see participants as unchanged, even if they had improved). However, the baseline assessment included extensive consensus ratings of BED and sub-threshold BED among all three interviewers, which likely helped to standardize their assessments of BED status. Fourth, the questions used to assess diagnostic changes over follow-up intervals have not been published or validated, and the accuracy of participants' reports of these changes may have been reduced by the limitations of memory over an approximately 2.5-year period. However, the same approach (i.e. focusing on changepoints along a timeline) is utilized in a highly-regarded interview assessing longitudinal changes in eating disorders (Eddy, Murray, & Thomas, Reference Eddy, Murray and Thomas2015). Further, using only information on diagnosis status at follow-up timepoints leads to the same conclusions regarding remission and relapse in our study, with the exception of the median time to remission, which is greater than 2.5 years when using only this information. Fifth, we examined remission from BED based on the absence of binge-eating episodes, but we did not assess either researcher- or individually-defined recovery, which can reflect improvements in eating disorder cognitions and other domains instead of or in addition to cessation of regular binge eating (Bardone-Cone et al., Reference Bardone-Cone, Harney, Maldonado, Lawson, Robinson, Smith and Tosh2010; Silén et al., Reference Silén, Sipilä, Raevuori, Mustelin, Marttunen, Kaprio and Keski-Rahkonen2021; Slof-Op ’t Landt, Dingemans, de la Torre, Rivas, & van Furth, Reference Slof-Op ’t Landt, Dingemans, de la Torre, Rivas and van Furth2019). Sixth, importantly, although our sample was more representative of individuals with BED in the general population in terms of sex, age of onset, and BMI compared to other prospective studies (Fairburn et al., Reference Fairburn, Cooper, Doll, Norman and O'Connor2000; Stice et al., Reference Stice, Marti and Rohde2013, Reference Stice, Marti, Shaw and Jaconis2009), all but one participant in our study identified as White or Black non-Hispanic, limiting generalizability to other racial and ethnic groups (Goel et al., Reference Goel, Mathis, Egbert, Petterway, Breithaupt, Eddy and Graham2022). Likewise, only three participants in our sample had BMI less than 25 at baseline, limiting our findings' generalizability to individuals with lower BMI. Seventh, in our analyses, we considered four machine learning methods to predict time to remission, three of which belonged to the same class (i.e. Cox regression). It is possible that other methods not included in our a priori data analysis plan (such as support vector machines or neural networks) may have yielded better predictions, although standard Cox regression and random survival forests have demonstrated better predictive accuracy for survival outcomes than these other methods in simulations (Wang et al., Reference Wang, Greenberg, Forkert, Chekouo, Afriyie, Ismail and Sajobi2022). Further, it is possible that predictive power may have been better with measures not included among our baseline predictors (e.g. frequency of objective binge-eating episodes, shape/weight overvaluation (Forrest, Jacobucci, & Grilo, Reference Forrest, Jacobucci and Grilo2022), scores on the Binge Eating Scale (Gormally, Black, Daston, & Rardin, Reference Gormally, Black, Daston and Rardin1982) or the Yale-Brown Obsessive Compulsive Scale modified for Binge Eating (Deal, Wirth, Gasior, Herman, & McElroy, Reference Deal, Wirth, Gasior, Herman and McElroy2015), or measures of depressive symptoms), or for outcomes other than BED remission (e.g. overall eating disorder symptoms [Forrest et al., Reference Forrest, Ivezaj and Grilo2023]). Additionally, our sample size, although considerably larger than existing studies, was not large in absolute terms, and less than half the sample experienced remission, limiting our ability to detect weak effects of the baseline variables and to achieve high predictive power, especially for random survival forests.
In conclusion, these limitations notwithstanding, our results suggest that, among adults with generally higher BMI in naturalistic circumstances, BED improves with time, but full remission often takes multiple years, and relapse is relatively common.
Supplementary material
The supplementary material for this article can be found at https://doi.org/10.1017/S0033291724000977
Acknowledgments
The authors gratefully acknowledge M. Shafiqur Rahman for providing an excerpt of the Stata code used in Rahman et al. (Reference Rahman, Ambler, Choodari-Oskooei and Omar2017).
Disclaimers
Ortho-McNeil Janssen Scientific Affairs had no control over the content of this manuscript or the authors' freedom to publish it. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health
Funding statement
The family study and longitudinal study were supported in part by an investigator-initiated grant from Ortho-McNeil Janssen Scientific Affairs (JH). Further, KNJ’s work on this project was supported by the National Institute of Diabetes and Digestive and Kidney Diseases of the National Institutes of Health (KJ, grant number K23-DK120517) and by a NARSAD Young Investigator Grant from the Brain & Behavior Research Foundation. CMB’s work on this project was supported in part by the National Institute of Mental Health (CB, grant numbers R56-MH129437, R01-MH120170, R01-MH119084, R01-MH118278, and R01MH124871); and the Swedish Research Council (CB, Vetenskapsrådet, grant number 538-2013-8864l).
Competing interests
Author 1 has owned equity shares in Sanofi and Centene Corporation, served on the Clinical Advisory Board for Beanbag Health, and received research funding from the National Institute of Diabetes and Digestive and Kidney Diseases. Author 8 has received consulting fees from Otsuka and has received research grant support from Boehringer-Ingelheim and Idorsia. Author 4 reports the following relationships: Lundbeckfonden (grant recipient); Pearson (author, royalty recipient). Author 6 is or has been a consultant to or member of the scientific advisory boards of F. Hoffmann-La Roche Ltd., Idorsia, Myriad, Novo Nordisk, Otsuka, Sipnose, Sunovion, and Takeda. She is or has been a principal or co-investigator on studies sponsored by Brainsway, Idorsia, Janssen, Marriott Foundation, Myriad, National Institute of Mental Health, Novo Nordisk, Otsuka, and Sunovion. She is also an inventor on United States Patent No. 6,323,236 B2, Use of Sulfamate Derivatives for Treating Impulse Control Disorders, and along with the patent's assignee, University of Cincinnati, Cincinnati, Ohio, has received payments from Johnson & Johnson, which has exclusive rights under the patent. Author 5 has received royalties from Up to Date. Author 2, Author 3, and Author 7 report no financial relationships with commercial interests.
Ethical standards
The authors assert that all procedures contributing to this work comply with the ethical standards of the relevant national and institutional committees on human experimentation and with the Helsinki Declaration of 1975, as revised in 2008.
