Lessons learned from a randomised controlled trial of a digital intervention targeting eating disorder symptoms in cisgender sexual minoritized men Authors: Emma Austen, Matthew Fuller-Tyszkiewicz, Siân A McLean, Beth O’Gorman, Katie Bishop, Hannah K Jarman, Susan J Paxton, Rachel F Rodgers, An Dang, Emily A Harris, Felicity Angelopoulos, Wesley Grey, Suku Sukunesan, Alexandra Parker, Adrian Shatte, Long Khanh-Dao Le, Scott Griffiths Published: 2026 ABSTRACT AbstractBackgroundThere is an unmet need for evidence-based and accessible interventions designed to treat eating disorders in sexual minoritized men, given their heightened vulnerability to developing disordered eating. Digital interventions are a promising avenue for eating disorder treatment because of their accessibility and flexibility in delivery. Here, we present a randomised controlled trial of Seed, a 10-week treatment-focussed digital intervention for eating disorders, among sexual minoritized men.MethodsParticipants (N = 126 sexual minoritized men) with a binge-spectrum eating disorder diagnosis (e.g., bulimia nervosa, binge-eating disorder) were randomly allocated to one of three conditions: (1) a waitlist control (2) a personalised intervention, where participants received personalised content recommendations based on their individual symptom profile, or (3) a default intervention, in which participants received no personalised recommendations. Changes from baseline in eating disorder symptoms were evaluated via clinical interview and self-report at 10 week and 6 month follow up.ResultsRelative to the waitlist control and default conditions, participants in the personalised intervention condition reported significantly greater decreases in their global Eating Disorder Examination interview scores at post-intervention/10-week follow up. We did not observe any significant changes in self-reported eating disorder symptoms. Our results are tentative as we were underpowered to detect small-to-medium treatment effects.ConclusionsOur study provides preliminary evidence for the efficacy of Seed for treating eating disorders in sexual minoritized men, particularly when personalised content recommendations are provided. However, future research must provide greater quality evidence through well-powered, more rigorous clinical trials.Trial registration: This trial was retrospectively registered on 04/03/2026 (Trial ID ACTRN12626000271303).Supplementary InformationThe online version contains supplementary material available at 10.1186/s40337-026-01648-0.Keywords: Eating disorders treatment, Sexual minoritized men, Randomised controlled trial, mHealth, Digital intervention Background Sexual minoritized (i.e., non-heterosexual) men experience higher rates of eating disorders than heterosexual men [10], in part due to minority stressors (e.g., sexual orientation discrimination), alongside intracommunity pressures to have a muscular physique [5]. The vulnerability of sexual minoritized men to eating disorders, facilitated by unique risk factors (e.g., minority stress, intracommunity appearance pressures), warrants the development and evaluation of tailored interventions to assist this group. In this paper, we present results and reflections from a randomised controlled trial of Seed, a digital intervention for treating eating disorders, among sexual minoritized men. Notably, our study faced several methodological challenges, including recruitment challenges that led to underpowered analyses. We hope this paper can inform future efforts to develop and evaluate treatment-focussed interventions for sexual minoritized men, an at-risk and underserviced population. Psychological interventions for eating disorders among sexual minoritized men Few studies have evaluated psychological interventions for eating disorders among men (e.g [19, 21]), and fewer have been conducted with sexual minoritized men. Most studies of psychological interventions for eating disorders are prevention-focused and evaluated using non-clinical samples [25], whereas treatment-focused interventions typically involve individuals diagnosed with an eating disorder or who exceed clinical thresholds on eating disorder symptom measures. Hendricks and colleagues [17] systematically reviewed eight trials of prevention-focussed interventions for eating disorders and body image among men. The authors identified eight interventions, all of which were prevention-focused. Five interventions were dissonance-based, fostering cognitive dissonance between participants’ values and behaviours around appearance. The remaining three interventions included two psychoeducational lifestyle modification programs and one media literacy intervention. Across these studies, interventions were predominantly delivered by a mix of clinicians/psychologists, peer-leaders, and members of research team. Evidence for psychoeducation and media literacy interventions was limited; however, the dissonance-based interventions showed promise. The four dissonance-based interventions with moderate quality evidence reported significant improvements in body image concerns at 4-week follow up, including to drive for muscularity [2, 6, 7, 21], and global body dissatisfaction [7]. Some studies observed improvements up to 6 months follow up, including in muscle dysmorphia symptoms [2] and body fat dissatisfaction [21]. One study found that reductions in body ideal internalisation partially mediated reductions in eating disorder symptoms [6], suggesting the importance of addressing body ideal internalisation in intervention design and delivery. Of these dissonance-based interventions, Brown and Keel [7] is arguably the most relevant to the current study because it is the only intervention, to our knowledge, to have been evaluated with sexual minoritized men specifically. The authors conducted an RCT of a dissonance-based program, The PRIDE Body Project, among 87 young college-aged gay men in the US. The intervention focused on challenging prohibitive appearance ideals among sexual minoritized men across verbal, written, and behavioural activities, and was delivered to participants in groups of four to seven people across two 120-minute sessions by a clinician and a peer leader. Relative to a waitlist control, intervention participants reported moderate-to-large (Cohen’s d = 0.61–1.01) improvements in dietary restraint, bulimic symptoms, body image dissatisfaction, drive for muscularity, and body-ideal internalisation at 4-week follow-up. This limited evidence suggests that dissonance-based strategies to may be effective in reducing eating disorder symptoms in sexual minoritized men. In sum, existing interventions for eating disorders among men are prevention-focussed and employ dissonance-based approaches. Treatment-focussed interventions, evaluated among clinical samples, are required to support to those whose eating disorder symptoms exceed clinical thresholds. To date, there are no treatment-focussed interventions designed specifically for sexual minoritized men. Digital interventions for eating disorders Unlike the in-person intervention described above, digital interventions for eating disorders are psychological interventions delivered via the Internet, smartphones, or other technological means and may be particularly useful when trying to reach under-served minority groups, including sexual minority men. One meta-analysis identified eight RCTs of treatment-focused digital interventions for eating disorders [25], which primarily implemented cognitive-behavioural therapy (CBT) and dissonance-based strategies. Intervention participants reported moderate-to-large improvements to eating disorder symptoms (e.g., binge episodes, global eating disorder symptoms) relative to a control condition, with follow-up periods ranging from eight to 24 weeks across studies. No digital interventions have been evaluated in sexual minoritized men, despite this being a high-risk group. Digital mental health interventions can be personalised, tailoring content to users’ specific needs through feedback or recommendations; doing so may improve participant engagement and enhance treatment outcomes [3]. A recent review highlighted that few studies have examined the impact of personalisation of digital mental health interventions on treatment efficacy, relative to a non-personalised version of the same intervention [27]. The authors highlight users should be able to complete intervention content autonomously (i.e., in an order of their choosing), providing users a sense of agency. Several studies have explored the feasibility of personalisation in eating disorder treatment (e.g [24, 23]). In this body of work, individual symptom profiles are developed through network analysis of ecological momentary assessment data. These symptom profiles are then used to provide individualized treatment plans, delivered by a clinician. While these personalised treatments have been shown to be effective at reducing eating disorder symptoms, RCTS are required to determine whether they are more effective than existing evidence-based treatments (e.g., CBT). We are unaware of any existing research implementing personalization to digital eating disorder interventions. However, Jarman and colleagues [20] qualitatively explored the preferences of potential end-users of the digital intervention trialled in the current study, finding that participants would prefer smartphone treatment apps that are personalised and can be completed flexibly (e.g., in a time or order of their choosing). Overall, personalisation and flexibility in digital interventions might enhance participants’ engagement and improve treatment outcomes; however, evidence is limited. The current study We are unaware of any published randomised control trials evaluating treatment-focused interventions for eating disorders among sexual minoritized men despite the vulnerability and unmet treatment needs of this population. This study evaluates Seed, a digital treatment-focused intervention, among sexual minoritized men with binge-spectrum eating disorders (e.g., binge-eating disorder, bulimia nervosa). Seed is a self-guided app (i.e., with no clinician involvement) that implements CBT and dissonance-based approaches to treat eating disorder symptoms. This combined approach was based on evidence indicating that CBT and dissonance-based therapy are effective in the treatment and prevention of eating disorders (e.g., Grilo et al. [15]; [17]; Stice et al. [29]). We had three conditions: a waitlist control, a personalised intervention (where participants receive tailored recommendations for app content based on their eating disorder symptom profile), and a default, non-personalised intervention. We hypothesised that intervention participants would report improved outcomes relative to a waitlist control at post-intervention: (i) greater reduction in disordered eating, and (ii) greater improvement in quality of life, depressive symptoms, body image flexibility, and body image internalisation. We also hypothesised that (iii) personalised intervention participants would report greater improvements in outcomes than default intervention participants.1 Method Participants We recruited 132 sexual minoritized men living in Australia. Eligible participants: were cis-gender sexual minoritized men; ≥ 18 years old; not receiving treatment for eating or body image concerns; not at risk of suicide; and met criteria for an eating disorder diagnosis other than Atypical Anorexia Nervosa (AAN), Anorexia Nervosa (AN) and Avoidant Restrictive Food Intake Disorder (ARFID). These diagnoses were excluded either on ethical grounds due to medical risk (AN, AAN) or because the app did not incorporate evidence-based treatments for the specific diagnosis (ARFID). We excluded data for six participants who were provided a different version of the intervention than they were allocated to, resulting in a final sample of 126 participants (Table 1). Table 1Participant characteristics at baseline (N = 126)VariableConditionCategoryWaitlist Control(N = 44)Default(N = 39)Personalised(N = 43)n (%)n (%)n (%)Sexual identityGay39 (88.64)37 (94.87)32 (74.42)Bisexual2 (4.55)2 (5.13)9 (20.93)Pansexual2 (4.55)0 (0.00)0 (0.00)Questioning0 (0.00)0 (0.00)1 (2.33)Prefer not to say1 (2.27)0 (0.00)1 (2.33)Eating disorder diagnosisBED12 (27.27)11 (28.21)13 (30.23)BN8 (18.18)4 (10.26)6 (13.95)OSFED a22 (50.00)19 (48.72)21 (48.83)Sub-threshold BED7 (15.91)7 (17.95)7 (16.28)Sub-threshold BN8 (18.18)10 (25.64)11 (25.58)No specifier provided7 (15.91)2 (5.13)3 (6.98)Unspecified Feeding or Eating Disorder1 (2.27)5 (12.82)3 (6.98)Missing *1 (2.27)0 (0.00)0 (0.00)RaceWhite31 (70.45)29 (74.36)21 (48.84)Asian2 (4.55)2 (5.13)6 (13.95)South Asian0 (0.00)1 (2.56)2 (4.65)Middle Eastern or North African0 (0.00)1 (2.56)2 (4.65)Aboriginal or Torres Strait Islander2 (4.55)0 (0.00)0 (0.00)Hispanic, Latino, or Spanish3 (6.82)0 (0.00)3 (6.98)Pacific Islander0 (0.00)0 (0.00)1 (2.33)Multiple racial identities2 (4.55)0 (0.00)3 (6.98)Prefer not to say0 (0.00)1 (2.56)1 (2.33)Missing4 (9.09)5 (12.82)4 (9.30)M (SD)M (SD)M (SD)Age41.02 (11.61)38.90 (11.80)38.87 (12.58)SES6.10 (1.53)6.85 (1.46)5.82 (1.71)BMI31.26 (6.56)30.66 (5.27)31.77 (8.99)EDE Global3.06 (0.89)2.84 (0.98)3.12 (0.80)EPSI Body dissatisfaction18.89 (4.06)18.62 (4.71)17.84 (5.19)EPSI Binge20.32 (5.43)20.44 (5.99)21.70 (5.69)EPSI Cognitive Restraint6.50 (2.77)6.15 (2.75)6.53 (2.50)EPSI Purging5.27 (5.01)2.82 (4.36)4.74 (4.35)EPSI Restricting6.70 (5.35)5.46 (4.63)5.42 (4.20)EPSI Excessive Exercise7.75 (4.73)6.74 (4.82)9.12 (5.01)EPSI Obesity10.66 (3.97)10.18 (4.67)11.12 (4.31)EPSI Muscle build7.59 (5.23)6.54 (5.16)8.26 (5.16)EDE Eating Disorder Examination, BN Bulimia nervosa, BED Binge eating disorder, OSFED Other Specified Feeding or Eating Disorder, BMI Body Mass Index (kg/m2), SES socioeconomic status (a score of 1 indicates the lowest subjective SES, and 10 the highest possible subjective SES). * Uncategorised diagnosis due to server-side error in data collection. a Other OSFED specifiers were either not observed in our sample (e.g., Purging Disorder), or not assessed due to the EDE protocol not generating sufficient information to reach a diagnosis (e.g., Night Eating Syndrome) Participant characteristics at baseline (N = 126) EDE Eating Disorder Examination, BN Bulimia nervosa, BED Binge eating disorder, OSFED Other Specified Feeding or Eating Disorder, BMI Body Mass Index (kg/m2), SES socioeconomic status (a score of 1 indicates the lowest subjective SES, and 10 the highest possible subjective SES). * Uncategorised diagnosis due to server-side error in data collection. a Other OSFED specifiers were either not observed in our sample (e.g., Purging Disorder), or not assessed due to the EDE protocol not generating sufficient information to reach a diagnosis (e.g., Night Eating Syndrome) Materials Demographics Participants self-reported their age, gender identity, sex at birth, race/ethnicity, height and weight. We measured subjective socioeconomic status with the MacArthur Scale of Subjective Social Status [1], where scores range from 1 (lowest) to 10 (highest). Primary outcomes Eating disorder examination The Eating Disorder Examination (EDE; [12]) is a clinician-led interview that examines the frequency of disordered eating over the previous 28 days. The EDE can be used to derive DSM-5 eating disorder diagnoses, and a global eating disorder symptom score is calculated by averaging scores across four subscales. Scores range from 0 to 7, with higher scores reflecting greater symptom frequency/severity. The global score demonstrated excellent reliability (McDonald’s omega [ω]s: 0.87–91). Eating Pathology Symptoms Inventory The Eating and Pathology Symptoms Inventory (EPSI; [13]) captures self-reported eating disorder symptoms across eight domains: body dissatisfaction; binge eating; excessive exercise; cognitive restraint; purging; negative attitudes towards obesity; restricting; and muscle building. The EPSI subscales demonstrated good reliability (ωs: 0.70–0.89). Secondary outcomes Our secondary outcomes were: body image internalization (Sociocultural Attitudes Towards Appearance Questionnaire; [28]); (2) body image flexibility (Body Image Acceptance and Action Questionnaire-5; [4]); quality of life (Assessment of Quality of Life – 4 Dimension [AQoL-4D]; [16]); and depressive symptoms (Patient Health Questionnaire-2; [22]). All scales demonstrated acceptable reliability (ωs > 0.70), except for the AQoL-4D, which we excluded from our reporting. Our intervention: Seed Seed is a smartphone app-based intervention with eight modules that provide psychoeducation and guided exercises addressing key components of eating disorder pathology and maintaining factors (Table 2). The modules are intended to be completed at a rate of approximately one per week. Seed primarily uses CBT approaches, with some elements adapted from dissonance-based therapy. For example, the “Regulated Eating” and “Mood Management” modules reflect a CBT approach and incorporate cognitive restructuring and behavioural experiments to challenge unhelpful thoughts and behaviours related to food and eating, appearance ideals, and their overvaluation. The “Overvaluation” module uses a dissonance-based approach, encouraging users to critically evaluate and reject sociocultural appearance ideals. Table 2Overview of Seed’s content modules, therapeutic approaches employed, and recommended order of completion for personalised intervention participantsApp moduleDescription of contentTherapeutic approachPersonalised condition: Recommended module for high EPSI subscale score11. Values and MotivationsIntroduction to values, identifying values, and building motivation to change in accordance with valuesCBT; Dissonance-based activitiesNA – first module completed by all participants2. Regulated EatingPsychoeducation on regular eating, compensatory behaviours, strategies to help with regular eatingCBTRestriction2; Purging3; Binge eating43. Mood ManagementIdentifying, challenging, and balancing negative thoughtsCBTBinge eating44. Diet Rules and ExerciseIdentifying beliefs about dieting, challenging unhelpful beliefs, and generating positive attitudes about exerciseCBT; Dissonance-basedExcessive Exercise5; Muscularity Concerns6; Cognitive Control75. OvervaluationPsychoeducation on overvaluation of appearance, societal influences on appearance values, and challenging these societal messages about appearance and dietCBT; Dissonance-basedBody dissatisfaction86. Body Checking & Appearance ComparisonsBody checking, appearance comparisons, the role of social media in facilitating body checking & appearance comparisonsCBT; Dissonance-basedBody dissatisfaction87. Emotion RegulationLinking emotions to eating and appearance, learning emotion regulation techniques, distress toleranceCBTBinge eating48. Maintaining ProgressIdentifying and managing relapses, generating a relapse planCBTNA1 Participants in our personalised condition were provided module recommendations based on their level of specific eating disorder symptoms (as measured by the EPSI). EPSI ‘high’ scores for each subscale [13]: 2 Restriction scores > 15; 3 Purging scores > 9; 4 Binge Eating scores > 24; 5 Excessive Exercise scores > 18; 6 Muscle Building scores > 12; 7 Cognitive Restraint scores > 11; 8 Body Dissatisfaction scores > 25 Overview of Seed’s content modules, therapeutic approaches employed, and recommended order of completion for personalised intervention participants 1 Participants in our personalised condition were provided module recommendations based on their level of specific eating disorder symptoms (as measured by the EPSI). EPSI ‘high’ scores for each subscale [13]: 2 Restriction scores > 15; 3 Purging scores > 9; 4 Binge Eating scores > 24; 5 Excessive Exercise scores > 18; 6 Muscle Building scores > 12; 7 Cognitive Restraint scores > 11; 8 Body Dissatisfaction scores > 25 We evaluated default and personalised versions of Seed. Participants in the personalised condition were recommended to first complete modules whose content addressed the symptoms for which they had ‘high’ EPSI subscales scores (Table 3) but could follow any order of their choosing. For example, participants with high Binge Eating subscale scores (> 24; [13]) were recommended to first complete the Regulated Eating, Mood Management, and Emotion Regulation modules. Participants assigned to our default format were not provided personalised recommendations and had to complete the content modules in sequential order. All participants completed Module 1 (Values and Motivations) before moving to the next modules. This module was designed to help users clarify their personal values, explore their motivations for change, and begin to build commitment to the therapeutic process, thereby enhancing engagement and treatment adherence. Table 3Pooled standardised estimates from mixed effects regressions modelling changes in clinician-evaluated (EDE) and self-reported (EPSI) eating disorder symptoms across time and between conditions (top half of table), and mean difference estimates (bottom half of table)Regression parametersEDEEPSIGlobal scoreMuscle buildingPurgingBinge eatingBody dissatisfactionRestrictingExcessive exerciseNATOIntercept0.16[– 0.11, 0.43]0.16[– 0.13, 0.45] 0.32 [0.03, 0.61] 0.18 [0.11, 0.47] 0.29 [0.01, 0.58] 0.29 [0.01, 0.58]0.09[– 0.20, 0.38]0.11[– 0.18, 0.40]Time 2 a– 0.19[– 0.56, 0.18]– 0.36[– 0.69, 0.03] – 0.30 [– 0.48, – 0.12] – 0.48 [– 0.83, – 0.13] – 0.41 [– 0.76, – 0.06]– 0.26[– 0.59, 0.07]– 0.33[– 0.66, 0.00]– 0.18[– 0.51, 0.15]Default b– 0.21[– 0.62, 0.20]– 0.20[– 0.63, 0.23] – 0.54 [– 0.95, – 0.13]0.02[– 0.39, 0.43]– 0.05[– 0.46, 0.36]– 0.27[– 0.68, 0.14]– 0.21[– 0.62, 0.20]– 0.10[– 0.53, 0.33]Personalised b0.05[– 0.36, 0.46]0.13[– 0.28, 0.54]– 0.12[– 0.53, 0.29]0.20[– 0.21, 0.61]– 0.19[– 0.60, 0.22]– 0.28[– 0.69, 0.13]0.28[– 0.13, 0.69]0.10[– 0.31, 0.51]Time 2 × Default0.07[– 0.66, 0.80]0.01[– 0.52, 0.54]0.11[– 0.16, 0.38]– 0.22[– 0.83, 0.39]– 0.46[– 1.09, 0.17]– 0.29[– 0.82, 0.24]– 0.09[– 0.62, 0.44]– 0.03[– 0.64, 0.58]Time 2 × Personalised – 0.83 [– 1.46, – 0.20]0.14[– 0.39, 0.67]0.10[– 0.19, 0.39]– 0.34[– 0.91, 0.23]– 0.01[– 0.62, 0.60]0.16[– 0.39, 0.71]0.17[– 0.38, 0.72]– 0.26[– 0.79, 0.27] Standardised mean difference ( d ) Waitlist T1 – T20.19[– 0.18, 0.57]0.36[0.02, 0.70] 0.30 [0.11, 0.48] 0.49 [0.14, 0.84]0.41[0.04, 0.77]0.26[– 0.08, 0.61]0.33[– 0.02, 0.67]0.18[– 0.16, 0.53]Default T1 – T20.12[– 0.52, 0.75]0.36[– 0.08, 0.80]0.19[– 0.03, 0.40] 0.71 [0.19, 1.23] 0.88 [0.34, 1.40]0.56[0.11, 1.00]0.42[– 0.01, 0.86]0.21[– 0.30, 0.73]Personalised T1 – T2 1.02 [0.50, 1.54]0.22[– 0.20, 0.64]0.19[– 0.05, 0.44] 0.83 [0.36, 1.29]0.42[– 0.06, 0.90]0.10[– 0.34, 0.54]0.15[– 0.30, 0.61]0.44[0.02, 0.86]Δd Personalised – Default 0.90 [0.14, 1.66]– 0.14[– 0.73, 0.45]– 0.01[– 0.30, 0.32]0.12[– 0.55, 0.79]– 0.45[– 1.14, 0.24]– 0.46[– 1.07, 0.31]– 0.27[– 0.88, 0.34]0.23[– 0.42, 0.88]Estimates are pooled across 60 imputed datasets. Ns at T1/T2: Waitlist = 44/36; Default = 43/20; Personalised = 45/20. Regression estimates are formatted as β [95% CI], mean differences as Cohen’s d (i.e., standardised mean difference) [95% CI], and the difference in mean difference values for personalised and default conditions as Δ [95% CI]. Bold = significant (p < .05). The p values for our mean difference values are adjusted for the false-discovery rate (Verhoeven et al. 2005). a reference group = Time 1, b reference group = Waitlist Control. EDE Eating Disorders Examination, EPSI Eating Pathology Symptoms Inventory, NATO Negative Attitudes Towards Obesity Pooled standardised estimates from mixed effects regressions modelling changes in clinician-evaluated (EDE) and self-reported (EPSI) eating disorder symptoms across time and between conditions (top half of table), and mean difference estimates (bottom half of table) 0.16 [– 0.11, 0.43] 0.16 [– 0.13, 0.45] 0.32 [0.03, 0.61] 0.18 [0.11, 0.47] 0.29 [0.01, 0.58] 0.29 [0.01, 0.58] 0.09 [– 0.20, 0.38] 0.11 [– 0.18, 0.40] – 0.19 [– 0.56, 0.18] – 0.36 [– 0.69, 0.03] – 0.30 [– 0.48, – 0.12] – 0.48 [– 0.83, – 0.13] – 0.41 [– 0.76, – 0.06] – 0.26 [– 0.59, 0.07] – 0.33 [– 0.66, 0.00] – 0.18 [– 0.51, 0.15] – 0.21 [– 0.62, 0.20] – 0.20 [– 0.63, 0.23] – 0.54 [– 0.95, – 0.13] 0.02 [– 0.39, 0.43] – 0.05 [– 0.46, 0.36] – 0.27 [– 0.68, 0.14] – 0.21 [– 0.62, 0.20] – 0.10 [– 0.53, 0.33] 0.05 [– 0.36, 0.46] 0.13 [– 0.28, 0.54] – 0.12 [– 0.53, 0.29] 0.20 [– 0.21, 0.61] – 0.19 [– 0.60, 0.22] – 0.28 [– 0.69, 0.13] 0.28 [– 0.13, 0.69] 0.10 [– 0.31, 0.51] 0.07 [– 0.66, 0.80] 0.01 [– 0.52, 0.54] 0.11 [– 0.16, 0.38] – 0.22 [– 0.83, 0.39] – 0.46 [– 1.09, 0.17] – 0.29 [– 0.82, 0.24] – 0.09 [– 0.62, 0.44] – 0.03 [– 0.64, 0.58] – 0.83 [– 1.46, – 0.20] 0.14 [– 0.39, 0.67] 0.10 [– 0.19, 0.39] – 0.34 [– 0.91, 0.23] – 0.01 [– 0.62, 0.60] 0.16 [– 0.39, 0.71] 0.17 [– 0.38, 0.72] – 0.26 [– 0.79, 0.27] 0.19 [– 0.18, 0.57] 0.36 [0.02, 0.70] 0.30 [0.11, 0.48] 0.49 [0.14, 0.84] 0.41 [0.04, 0.77] 0.26 [– 0.08, 0.61] 0.33 [– 0.02, 0.67] 0.18 [– 0.16, 0.53] 0.12 [– 0.52, 0.75] 0.36 [– 0.08, 0.80] 0.19 [– 0.03, 0.40] 0.71 [0.19, 1.23] 0.88 [0.34, 1.40] 0.56 [0.11, 1.00] 0.42 [– 0.01, 0.86] 0.21 [– 0.30, 0.73] 1.02 [0.50, 1.54] 0.22 [– 0.20, 0.64] 0.19 [– 0.05, 0.44] 0.83 [0.36, 1.29] 0.42 [– 0.06, 0.90] 0.10 [– 0.34, 0.54] 0.15 [– 0.30, 0.61] 0.44 [0.02, 0.86] 0.90 [0.14, 1.66] – 0.14 [– 0.73, 0.45] – 0.01 [– 0.30, 0.32] 0.12 [– 0.55, 0.79] – 0.45 [– 1.14, 0.24] – 0.46 [– 1.07, 0.31] – 0.27 [– 0.88, 0.34] 0.23 [– 0.42, 0.88] Estimates are pooled across 60 imputed datasets. Ns at T1/T2: Waitlist = 44/36; Default = 43/20; Personalised = 45/20. Regression estimates are formatted as β [95% CI], mean differences as Cohen’s d (i.e., standardised mean difference) [95% CI], and the difference in mean difference values for personalised and default conditions as Δ [95% CI]. Bold = significant (p < .05). The p values for our mean difference values are adjusted for the false-discovery rate (Verhoeven et al. 2005). a reference group = Time 1, b reference group = Waitlist Control. EDE Eating Disorders Examination, EPSI Eating Pathology Symptoms Inventory, NATO Negative Attitudes Towards Obesity Tailoring Seed for sexual minoritized men The current study trials a version of Seed app tailored specifically to sexual minoritized men. Changes to the app included changes in language referring to dieting, exercise, and body ideals; for example, instead of referring exclusively to dieting and exercise in pursuit of thinness, this version of the app also referenced behaviours engaged in pursuit of muscularity (e.g., example, muscle building, bulking, and shredding). App content discussing dieting was edited to make more frequent references to protein intake and specific supplements more often taken by men (e.g., testosterone boosters). App content discussing appearance ideals was adjusted to refer to physiques idealized more (on average) by queer men; that is, physiques that are both lean and muscular. Following initial development of this version of Seed, sexual minoritized men with a history of eating disorders trialled the app. Participants discussed their experience with Seed and provided feedback via interview; their dislikes and preferences were compiled and informed subsequent revisions to the app. Procedure This study (trial registration: ACTRN12626000271303) is related to another RCT evaluating Seed among men and women (trial registration: ACTRN12621001244897). Unlike the larger RCT, this study examines the efficacy of a version of Seed tailored specifically for sexual minoritized men. We retrospectively registered the trial reported in this paper for clarity, both because of the changes to the app, along with the separate population from which we sampled from. Our Open Science Framework page contains our preregistration, data, code, and example intervention content. We received ethics approval from the University of Melbourne (IDs: 30177, 13046). Figure 1 presents our CONSORT diagram. We recruited participants via dating apps (e.g., Grindr), social media, and eating disorder organisations. Interested participants completed a screening survey, and eligible participants were contacted to book an EDE interview. Provisionally or generally registered psychologists with ≥ 6 months’ clinical experience administered the EDE – these interviewers were trained by a clinical psychologist with expertise in the EDE. CONSORT Diagram visualising the flow of our participant numbers from enrolment through to analysis. Notes: * Some participants in the default and personalised intervention conditions were erroneously provided the other version of the intervention (i.e., default participants provided personalised version, and vice-versa). We excluded data for these participants prior to analysis Eligible participants were randomised to an intervention condition using Qualtrics’ randomisation function. Participants who were randomised to the personalised or default conditions were then emailed app download instructions by the psychologist who administered their EDE interview. This study was conducted single-blind: participants were not told what condition they were allocated to. Participants re-completed the baseline measures and EDE at 10-weeks post-intervention, and baseline measures at 6-months. Participants were remunerated at follow-ups. Data collection periods were: recruitment (February–October 2024), post-intervention (April–December 2025), and 6-month follow up (August 2024–April 2025). We discovered an error in treatment allocation while cross-checking our randomisation record against app access instructions distributed to participants. Some participants randomised to our default (n = 4) and personalised (n = 2) conditions were provided access to the alternate version of the app (e.g., default participants given access to personalised format, and vice-versa). This was an administrative error, whereby a staff member on the project erroneously provided participants with instructions to download the opposite version of the app to which they were randomised. We excluded data for these participants prior to analysis. Data analysis Statistical power Our pre-registered power analysis (Supplementary Fig. 1) assumed 80% power, a standardised mean difference of 0.3 between intervention groups, α = 0.05, and 50% attrition, generating a recruitment target of 300 participants. This calculation did not account for the multilevel structure of our analyses and likely underestimated the required sample size. Compounding this, our recruitment period was limited due to funding constraints, and we did not meet our recruitment target. Given our achieved sample size and attrition, our analyses are underpowered: readers should prioritise the 95% confidence intervals around observed effects to interpret the range of plausible effect sizes. Missing data A Little’s Test provided no evidence against the assumption that our data were missing completely at random, χ2(375) = 358.10, p = .726. We handled missing data with multiple imputation via chained equations (N = 60 imputations; [9]). Intervention efficacy We deviated from our pre-registered analysis plan due to power constraints. Instead of modelling all timepoints together, we simplified our analyses into two stages: (i) intervention efficacy (Times 1 and 2; reported in our main text), and (ii) symptom maintenance (Times 2 and 3; Supplementary Table A). We used mixed effects regressions to examine changes in outcomes from Time 1 to Time 2. We regressed each outcome on time (Time 1, Time 2), condition (waitlist control, personalised, default), and their interaction. We included a random intercept for participant, and fixed effects for condition and timepoint. Our models assumed an unstructured covariance structure. The focal effects were the time x condition interactions: a significant negative interaction indicates that, relative to the waitlist control, intervention participants reported significantly greater reductions in outcomes from Time 1 to 2, suggesting that changes in outcomes were attributable to the intervention. To compare intervention groups, we computed mean differences in outcomes within each condition from Time 1 to 2, adjusting our p-values for the false discovery rate [30]. We then computed the between-group difference of these mean differences (i.e., Δ mean difference = Personalised – Default). Results Table 2 presents standardised estimates for eating disorder models (for secondary outcomes, see Supplementary Tables B and C). We observed a significant time x condition interaction on the EDE global score. Relative to the waitlist control and default conditions, personalised condition participants reported greater reductions in their global score from Time 1 to Time 2 (small to large effects). We did not observe significant time x condition interactions on EPSI scores or secondary outcomes (Supplementary Table B). We remind readers that these results are preliminary and must be interpreted with caution given that our analyses were underpowered. Discussion We evaluated Seed, a treatment-focussed digital intervention, among sexual minoritized men with binge-spectrum eating disorders. We hypothesised that, relative to a waitlist control, intervention participants would report significantly greater improvements in eating disorder symptoms and secondary outcomes. We expected participants in our personalised intervention would report greater symptom improvement than default intervention participants. As expected, participants in the personalised condition reported greater improvements in their clinician-evaluated symptoms than the default intervention and waitlist control; however, no significant changes in self-reported eating disorder symptoms or secondary outcomes were observed. Several methodological limitations impact the strength of evidence provided by this study, as discussed below. Evidence suggests users may prefer digital eating disorder interventions that are personalised and can be completed flexibly because they might be more engaging (Jarman et al. 2022). Supporting this, participants assigned to our personalised app, which offered tailored content recommendations and flexible completion, reported greater improvements to their clinician-evaluated symptoms than those assigned to the default, non-personalised app. Critically, Seed was not set up to record engagement data (e.g., time spent on app, amount of content completed; [26]); therefore, we could not evaluate whether heightened engagement accounted for greater improvements in the personalised condition. However, we observed a large attrition rate in our intervention groups (> 50% for intervention versus 19% for waitlist), which is considered a close proxy for engagement [26]; here, the large attrition rate in our intervention groups may reflect that these participants were not engaged with the app and therefore not motivated to complete the follow-up assessments. Supporting this, a meta-analysis of RCTs of digital eating disorders interventions highlighted attrition is 44% higher for interventions delivered via smartphone app than those delivered on a computer or website [26]. The authors contend that smartphones tend to be used in a less structured and more momentary manner than computers, “inadvertently reduc[ing] users’ … investment in the intervention and commitment to completing trial assessments” (p. 8). We note we observed significant treatment effects when measured via the EDE, but not the EPSI. The lack of treatment effects on self-reported eating disorder symptoms may result from differences in clinician-led and self-report measurement: clinician-led assessments may be more accurate than self-reports for symptoms whose identification benefits from clinical expertise (e.g., distinguishing objective from subjective binge episodes; [11]). Compounding this, the null effects observed across a range of outcomes may be due to low statistical power, resulting from our small sample and substantial attrition. Lower rates of recognition and treatment-seeking among men with eating disorders [14], combined with our short recruitment period (nine months) may explain our small sample size. Future directions: lessons learned from our study We recommend several avenues through which future research can provide stronger evidence for the efficacy of Seed among sexual minoritized men. First, our recruitment period was limited due to funding constraints, and we recruited primarily from online dating apps such as Grindr and Scruff. Future studies might accrue larger samples at enrolment with a longer recruitment period and more diversified recruitment strategy. Second, future studies can incorporate strategies demonstrated to improve engagement and participant retention – for example, through delivering Seed via a web- or computer-based format, and increasing human contact throughout the intervention (e.g., regular therapeutic support). Relatedly, any future trials must record participants’ level of engagement with Seed, for example, through recording the number of modules completed, number of program logins or views, and/or the number of days that participants actively use the program [26]. Future trials may also gather qualitative feedback from participants to supplement engagement data; for example, by providing insight into the specific barriers to their engagement with the app, and/or the elements of the app they found the most engaging. Third, we may have prevented or reduced the number of participants provided the incorrect version of the app through more frequent cross-checking of our randomisation record against the instructions provided to participants. We recommend that future studies evaluating Seed, or indeed any other intervention, engage in regular administrative cross-checking to reduce the incidence of similar errors. Fourth, we did not examine inter-rater reliability for the EDE; inter-rater variation may have reduced the precision of EDE model estimates [18]. Future research should account for inter-rater variation to generate more precise estimates. Lastly, the generalizability of our findings is limited, as our sample predominantly identified as gay, White, with average age of 38 to 41 years. Future research can evaluate the efficacy of Seed in more diverse populations (e.g., across intersections of age, sexual identity, and ethnicity), for whom the risk factors for, and presentation of, eating disorders vary [8]. Conclusions In conclusion, while our study provides preliminary evidence of the efficacy of Seed among sexual minoritized men, well-powered and more methodologically rigorous RCTs are required to reliably evaluate Seed’s efficacy. Electronic Supplementary Material Below is the link to the electronic supplementary material. Supplementary Material 1. Acknowledgements Not applicable. Author contributions Emma Austen: Formal Analysis, Investigation, Data Curation, Writing – Original Draft, Visualization, Project administration; Matthew Fuller-Tyszkiewicz: Conceptualization, Methodology, Software, Resources, Writing – Reviewing & Editing, Supervision, Project administration, Funding acquisition; Siân A. McLean: Conceptualization, Methodology, Resources, Writing – Review & Editing, Funding acquisition; Beth O’Gorman: Conceptualization; Methodology; Writing – Review & Editing, Supervision; Katie Bishop: Methodology, Software, Writing – Review & Editing; Hannah K. Jarman: Conceptualization, Writing – Review & Editing; Susan J. Paxton: Conceptualization, Methodology, Writing – Review & Editing, Funding acquisition; Rachel F. Rodgers: Conceptualization, Methodology, Writing – Review & Editing, Funding acquisition; An Dang: Methodology, Investigation, Resources, Writing – Review & Editing; Emily A. Harris: Methodology, Writing – Review & Editing, Project administration; Felicity Angelopoulos: Software, Writing – Review & Editing; Wesley Grey: Software, Writing – Review & Editing; Suku Sukunesan: Conceptualizaton, Software, Funding acquisition; Alexandra Parker: Conceptualizaton, Methodology, Funding acquisition; Adrian Shatte: Conceptualization, Software, Methodology, Funding acquisition; Long Khanh-Dao Le: Conceptualization, Writing – Review & Editing; Scott Griffiths: Conceptualization, Methodology, Formal analysis, Resources, Writing – Review & Editing, Supervision, Funding acquisition. Funding This study was supported by the Medical Research Future Fund (APP1179321). Data availability The data that support the findings of this study are openly available on the Open Science Framework at [https://osf.io/hzyqv/](https:/osf.io/hzyqv) . Declarations Ethics approval and consent to participate This study received ethics approval from the University of Melbourne (IDs: 30177, 13046). Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Note: Our Supplementary Materials reports outcomes up to 6-month follow-up. Our main text reports our pre- to post-intervention outcomes only, as our 6-month follow-up analyses were substantially underpowered. Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1. Data Availability Statement The data that support the findings of this study are openly available on the Open Science Framework at [https://osf.io/hzyqv/](https:/osf.io/hzyqv) .