Prognostic predictors of outcome following psychological treatment for anxiety or depression: a longitudinal statistical learning approach using electronic health record data

Lead Supervisor
Prof Thalia Eley
Professor of Developmental Behavioural Genetics
Social, Genetic and Developmental Psychiatry Centre, Institute of Psychiatry, Psychology and Neuroscience (IoPPN)
King’s College London

Dr Ewan Carr

Project Details

Anxiety and depression are chronic and debilitating psychiatric illnesses. Effective psychological treatments are available, but these do not work for all patients. Among individuals undergoing IAPT treatment (Improving Access to Psychological Therapies; England and Wales), around 5/10 do not achieve recovery and 3/10 do not see ‘reliable and substantial reductions’ in their symptoms[1]. Accurate prediction of prognostic outcomes for patients receiving psychological therapy is therefore vital – to inform adaptations to existing therapy or switching to alternative treatments.

There is considerable evidence of raised levels of mental health problems such as anxiety and depression in individuals self-identifying as being of non-standard gender and/or sexuality. However thus far, we know little about whether treatment outcomes also differ in these groups.


In this project the student will build a screening tool to identify patients undergoing IAPT treatment who are at risk of poor prognostic outcomes (e.g., not recovering, little or no reduction in symptoms). Building on the large evidence base on the predictors of treatment response, the project will address two important gaps:

== Aim 1: Using longitudinal data to improve prediction of treatment outcomes ==

Past studies have considered baseline predictors of treatment outcome among IAPT patients, but few have incorporated repeated measures. This project will consider how longitudinal assessments, collected weekly during a 12-week course of treatment, can improve predictive accuracy[2]. For example, using mixed effect random forests, time series support vector machines (SVM), or penalised generalised linear mixed effect models[3]. The student will consider how discrimination increases as additional assessments (at weeks 1, 2, 3, …, 12) are added to the model. For example, how many weeks into therapy are needed before a prognostic outcome can be predicted with sufficient accuracy? How can this be used to inform treatment plans?

These analyses will use data from the use data from the South London and Maudsley NHS Trust (SLaM) IAPT service (2019-2021; >90k observations). External validation of discrimination and calibration will be carried out using linked IAPT records from the Genetic Links to Anxiety and Depression (GLAD) study (2018-2021; >10k observations).

== Aim 2: Stratification by non-standard gender or sexuality ==

Past models have considered stratification by gender, but few to date have considered how (i) prognostic markers and (ii) predictive accuracy for psychological treatment outcomes vary by non-standard gender or sexuality. These groups have typically been excluded or overlooked in existing analyses. These analyses will be undertaken in the GLAD-IAPT dataset, where there is unusually detailed data about gender/sexuality and strong representation of individuals of non-standard sexuality in particular. As such, we are uniquely placed to study how the ability to screen for poor treatment outcomes varies between smaller but clinically important subgroups.


Findings will be disseminated to the public via social media, blogs, GLAD study participant newsletters, and public science events. We will also invite lived-experience experts and clinical service leads to discuss the potential use of our findings in decision making. The results will be presented at academic conferences and written-up for publication. Students in the first supervisor’s lab usually complete their PhD with 3-4 first author publications.


[1] NHS England » IAPT at 10: Achievements and challenges.
[2] Bull, L. M., Lunt, M., Martin, G. P., Hyrich, K. & Sergeant, J. C. Harnessing repeated measurements of predictor variables for clinical risk prediction: a review of existing methods. Diagn Progn Res 4, 9 (2020).
[3] Groll, A. & Tutz, G. Variable selection for generalized linear mixed models by L1-penalized estimation. Stat Comput 24, 137–154 (2014).


The student will work with two pre-existing datasets:

1) SLaM.

The South London and Maudsley (SLaM) NHS Trust IAPT service dataset includes 90,000 individuals treated between 2008-2021. Mean age 37 years, 67% female, reported ethnicities 68% White, 18% Black, 6% Mixed, 5% Asian, 3% Other.

2) GLAD.

The Genetic Links to Anxiety and Depression (GLAD) Study has collected baseline data from ~41,000 individuals across the UK with lifetime experience anxiety or depression. Mean age 38 years, 78% female, 95% white. Medical record linkage is underway and will be achieved by the end of 2022. Approximately 75% of GLAD participants report receiving psychological treatment, and we expect >10,000 to have undergone psychological therapy within the IAPT service. Data sharing agreements and ethical approval are in place for both datasets, although specific data requests will need to be made for this project. The student will need to apply for approval to access SLaM data but other students supervised by the primary supervisor have done so and are already working with this data.

In all samples, participants undergoing therapy complete weekly measures of anxiety (GAD-7), depression (PHQ-9) and social impairment (Work and Social Adjustment Scale). Sociodemographic factors including age, gender, ethnicity, nationality, and employment status are also measured. Notably, GLAD is well represented with regard to individuals of non-standard gender (3%, ~1,000 individuals) and sexuality (28% in total including 6% homosexual, 16% bisexual, 1% asexual and 3% prefer to self-define).


Anxiety; depression; psychological treatment; cognitive behavioural therapy; statistical learning; prognostic outcome; non-standard gender; prediction model