Trial of an Online Group Psychotherapy Intervention for Common Mental Health Disorders

The study protocol - including a full statistical analysis plan was pre-registered and made publicly accessible in the Open Science Framework on the 4th of October 2024 (available here - https://osf.io/vup9c/).

A pragmatic, multi-site, cluster factorial randomised controlled trial, involving National Health Service (NHS) Talking Therapies (NHS-TT) services for anxiety and depression (formerly known as IAPT services) in England. The "UPLift" digitally enabled group intervention, based on the Unified Protocol, will be available to patients with common mental disorders (and comorbidities) that are eligible to receive high intensity Cognitive Behavioural Therapy (CBT) in NHS-TT services, following national guidelines (National Collaborating Centre for Mental Health, 2024). The Unified Protocol is a transdiagnostic treatment that was designed to be relevant to a wide range of emotional disorders, given the common co-occurrence of depression, anxiety, somatoform and dissociative symptoms. It is based on CBT and emotion regulation theories, and aims to support patients to apply a series of emotion regulation skills that target the core common vulnerabilities that underlie a range of mental health problems: neuroticism, low perceived control, and overestimation of threat. The protocol follows the principle of parsimony, by including the minimum number of empirically validated techniques that would make it accessible, uncomplicated, and effective for the maximum number of patients with various combinations of symptoms and problems. Contents are organised across 8 modules delivered online at 12 weekly 1.5 hour sessions and will include a combination of internet-based resources (self-monitoring questionnaires, psychoeducational videos, skills practices linked to each module) and group therapy sessions delivered by a qualified cognitive behavioural therapist.

The unit of randomisation (cluster) will be each participating NHS-TT service. NHS Trusts are wider organisations that manage one or more NHS-TT services, so the unit of randomisation is the local team. All patients treated in the same local area service will be members of the same cluster. Participating services will be randomly allocated to one of 5 arms. Each arm will include the same 12-session online group Unified Protocol treatment but organised in a different sequence. The first two modules will remain fixed at the start of the intervention, as these introduction sessions are necessary to orient all patients to the treatment and to establish a cohesive group process. Similarly, the final module is fixed at the end of the intervention, as it includes an overall review of all coping skills, and it focuses on relapse prevention. The middle modules cover specific emotion regulation skills and are most amenable to personalisation (e.g., some may be more or less effective for specific patients). These prescriptive modules will be delivered in different sequences across each of the trial arms.

Participants will complete baseline measures on entry to the study, weekly symptom outcome measures prior to each session and a post-treatment (session 12) and 6-month follow-up survey. The primary outcome will be Negative affectivity (NA), based on a composite score pooling items from Patient Health Questionnaire-9 (PHQ-9; depression), Generalized Anxiety Disorder-7 scale (GAD-7; anxiety) and Work and Social Adjustment Scale (WSAS; functioning). Secondary outcomes will be Quality of life (EuroQol - 5 dimensions; EQ5D), Wellbeing (Warwick and Edinburgh Mental Wellbeing Scale; WEMWBS) and drop out from treatment.

Researchers will also collect anonymised clinical health records for all patients accessing individually delivered CBT in the participating services during the period of the clinical trial. These data will include demographics, clinical care and clinical outcomes. This will enable researchers to undertake comparisons of clinical and cost-effectiveness using a case-control matching strategy described in the statistical analysis plan.

The statistical analysis plan will be carried out in three stages linked to each of the objectives of the study.

Objective 1: compare outcomes of different UP module sequences This will be the primary analysis, based on which the sample size calculation is based. Cases with missing post treatment (12th session) data-points in the dependent variable will be imputed using the MissForest R package using all baseline variables as predictors. Imputations will be carried out separately for each trial arm. The analysis will apply a linear multilevel (mixed effects) model, where patients will be nested within groups, entering a random effect for the group-level. The dependent variable will be the post-treatment Negative Affectivity (NA) score, entering trial arm as an independent variable, controlling for baseline NA. Models will be developed in sequential steps and model fit indices (AIC, BIC, -2LL) will be examined after each modelling step. Researchers will retain and interpret the best-fitting and most parsimonious model achieved through this stepwise process, using the -2 loglikelihood ratio test to compare models. The intra-cluster correlation coefficient (ICC) will be reported as an index of variability in treatment outcomes attributable to the group-level. Between-group effect sizes (Cohen's d) will be calculated and reported for each contrast (e.g., standard UP sequence vs. other sequences).

Secondary analysis: The above analysis plan will be repeated using a logistic multilevel model with dropout as the dependent variable. The dropout variable will be derived from group therapy attendance records, where cases that attended more than half (≥6) sessions will be grouped into the reference category (code = 0) and cases that attended less than half (≤5) will be classed as cases that dropped out (code = 1).

Objective 2: develop a personalised module selection method Cases with missing data-points in the dependent variable will be imputed using the MissForest R package using all baseline variables as predictors, with separate imputations in the training/validation partitions, and within each trial arm. Reliable improvement (RI) refers to a reduction in symptom severity in between two time-points, which is greater in magnitude than the reliable change index for the psychometric measure. Following this definition, the investigators will class cases using a binary variable denoting whether they did (code = 1) or did not (code = 0) have reliable improvement after exposure to a prescriptive Unified Protocol module. This coding will enable us to identify cases that experience reliable symptomatic improvements after exposure to each of the target modules, regardless of where the module is located (e.g., towards the start, middle or end of the intervention) in the group intervention they have been allocated to. This methodology will enable researchers to control for sequencing and early response effects, to identify the profile of patients who have a favourable response to each specific module.

Sample partitioning: The full trial sample will be randomly partitioned into training and validation subsets using a 60:40 split, with a balanced number from each of the 5 trial arms in each training-partition split. This will ensure that 396 cases will be randomly selected into the training partition, exceeding the minimum sample size of 300 necessary to train a reliable clinical prediction model using supervised machine learning.

Model development and optimisation (in the training partition): Researchers will produce a total of 5 machine learning prediction models, one for each prescriptive UP module. Each trained model will process inputs (patients' baseline features) and will output [a] a predicted classification and [b] a corresponding predicted probability pertaining to the classification. The logic of these algorithms is to answer the following question: Given these patient characteristics, how likely is it that this specific patient will experience reliable improvement after accessing module X? Cross-validation strategy (in the validation sample): The statistical accuracy and clinical utility of the machine learning models will be evaluated in the validation strategy by comparing the algorithm predictions against the observed outcomes and comparing outcomes for patients who received their optimal module first compared to patients who did not.

Secondary analysis: The above machine learning analysis and cross-validation strategy will be repeated to train a binary classifier with dropout as the dependent variable. The dropout variable will be derived from group therapy attendance records, where cases that attended more than half (≥6) sessions will be grouped into the reference category (code = 0) and cases that attended less than half (≤5) will be classed as cases that dropped out (code = 1).

Objective 3: evaluate the effectiveness of digitally enabled UP group therapy vs. individual CBT This analysis will include [a] a subgroup of clinical trial participants who were randomly assigned to the arm where the UP protocol was delivered in the standard sequence, and [b] a matched sample of patients who accessed individual high intensity CBT in the participating NHS-TT services. The data for the individual CBT cases will be sampled from electronic health records. CBT cases that meet the trial inclusion criteria listed will be selected for this analysis, only including cases where the primary presenting problem is identifiable in clinical records and where baseline severity scores (PHQ-9, GAD-7, WSAS) are available at their first attended CBT session. Cases with missing post-treatment (last attended session) data-points in the dependent variable will be imputed using the MissForest R package using all covariates as predictors. Imputations will be carried out separately for each group (individual CBT; standard UP group therapy). Post-treatment clinical outcomes (NA scores) will be compared between cases accessing the standard group UP protocol versus individual CBT cases using doubly robust estimation, which combines an adjustment for non-random allocation to treatments, and a covariate-adjusted regression model to test the causal effect of exposure to an intervention.

Secondary analyses: The above analysis will be repeated to compare outcomes using disorder-specific measures (PHQ-9, GAD-7, ADSM) for cases with the relevant diagnoses. Between-group effect sizes (Cohen's d) will be calculated and reported for each comparison.

Health economic analysis: An economic analysis will be conducted from the NHS and Personal Social Services perspective. A cost-utility analysis will use quality-adjusted life years (QALYs derived from the EQ-5D questionnaire, and tariff based on the UK public value set) as the measure of quality of life, using post-treatment (last-attended session) data on the NA scale in the primary analysis, and 3-month follow-up data in a secondary analysis. Health and social services resource use will be valued using NHS reference costs and the personal and social services resource use database (PSSRU). Investigators will estimate the incremental cost effectiveness ratio for group therapy versus individual CBT, using bootstrapping to estimate confidence intervals. Decision uncertainty will be presented on a cost-effectiveness acceptability curve. Additional sensitivity analyses will be conducted for resource use and unit costs. A sensitivity analysis will control for baseline costs. Scenario analyses will explore alternative costing perspectives; that is, NHS and NHS/PSS perspectives. Results of the cost-effectiveness analysis will be reported in line with the Consolidated Health Economic Evaluation Reporting Standards 2022 Statement.