Suicidal behaviour following different treatments for depression
Mental Health Research UK Award

Mark Robinson PhD Scholarship 2014, King's College London, Academic Department of Psychological Medicine


Supervisors: Professor Robert Stewart (First Supervisor), Dr Rina Dutta (Second Supervisor)

Research Student: 
Andrea Fernandes

Hello all, my name is Andrea Fernandes. In 2013, the Department of Psychological Medicine at the Institute of Psychiatry, Psychology and Neurosciences, King’s College London (KCL) was awarded the MHRUK, Mark Robinson MRCVS Scholarship. Subsequently I secured this post.

I have now officially started my studentship (as of the 1st of October 2014)  I will be supervised by Professor Robert Stewart, Professor of Psychiatric Epidemiology and Clinical Informatics, and Dr Rina Dutta, Clinical Scientist Fellow and Academic Lecturer.

The project we’ll be starting on will use a bespoke mental health research database to investigate the effects on suicidal behaviour of treatments used in major depression. There is a lack of good quality research looking at how treatments, especially antidepressants, truly affect suicidal behaviour. In general, we aim to answer the following questions:

-    What factors determine the type of treatment a person is started on after being diagnosed with depression?

-    What treatments, or combinations of treatments, are most/least associated with suicidal behaviour?

-    How much can suicidal behaviour be said to be associated with the treatments received for depression rather than the earlier factors determining those treatments?

Although I have to admit I’m a little bit nervous this first month, I feel excited to start this PhD – working on the project, working with my supervisors, the training – and look forward to what the next three years bring.

Summary: 

Depression is the most important risk factor for suicide. Depression treatment includes medication (e.g. antidepressants), ‘talking treatments’ such as cognitive behavioural therapy (CBT), or a combination of both. Aside from reducing symptoms of depression, a short-term goal of treatment is to lower the risk of suicidal behaviours – which range from a person harming themselves in a relatively minor way to suicide itself. However, several research studies have found that some people are at higher risk of suicidal behaviours when they first start certain antidepressant medicines. This is clearly important, but it can be very difficult to disentangle whether particular treatments are actually directly increasing the risk of suicidal behaviour, or whether people already at risk of suicidal behaviour are given those particular treatments (i.e. whether the apparent ‘effect’ might be explained by earlier symptoms or other personal characteristics). Randomised trials would be helpful but are rarely large enough and are not going to be forthcoming for already-established treatments. Our proposed study will meet this challenge using a wealth of in-depth ‘real world’ information from a large number of people receiving routine care for depression (both medication and talking treatments). The student will begin by defining the most common treatments (single or in combination) and will compare the characteristics of people receiving them. Taking these characteristics into account, they will then investigate whether there are true differences in suicidal behaviour risk. The studentship will therefore provide important new information of high relevance for the mental health services providing treatment for depression.

Start date:
September 2014

Aim: To investigate relationships of different treatments with suicidal behaviour in people with depression receiving secondary mental health care.

Objectives:

To describe the most common profiles of depression treatment (e.g. antidepressants, cognitive behaviour therapy (CBT) and other co-prescribed psychotropic agents) in secondary care.

To investigate demographic and clinical factors, including prior suicidal ideation/behaviour, predicting receipt of individual depression treatment profiles.

Taking into account these relationships, to investigate variation in risk of future suicidal behaviour between groups receiving different depression treatments.

2017 Report

WORK IN THE LAST ACADEMIC YEAR

This year saw me conducting two studies, both of which contributed to the overall study aim. It has been a year of steep learning curves which led to major progress towards the research question, described below.

Measuring suicidality mentions in the text

Being able to measure suicidality within free-text clinical records - which is something that has not been done in our dataset and without which we would not be able to answer our main research question. In the past year, we have generated two algorithms that identify mentions of suicidal ideation and attempts. This study has been submitted for publication.

Investigating factors associated with antidepressant prescriptions in secondary care

We have also investigated factors that are involved in the prediction of antidepressant prescription to ensure that we can control for these to avoid biases in the analyses. This study is being submitted for publication.

Please see here for the presentation for results from both studies.

FUTURE WORK

I am now ready to plan the analysis to investigate the effect of antidepressants on suicidal behaviour in a cohort of patients with clinical depression receiving care from psychiatric services. This next year will involve planning and conducting the analysis involved in answering the main research question.

2016 Report


This project aims to study suicidal behaviours using free-text clinical records in a longitudinal manner. A large part of assessing suicidal behaviour following different antidepressant treatment in a clinical cohort is to ensure that we control for indication bias (i.e. how much the treatment choice has been influenced by perception of pre-treatment risk rather than being itself a determinant of subsequent suicidal behaviours).

Here’s a reminder of the main phases of this research:

Phase 1) Get familiar with the literature: identify gaps and the strengths and limitations of research in this area. (Brief summary of the results can be found at: https://github.com/andreafernandes/Appendices-for-MHRUK-report)

Phase 2) Understand the existing database and identify variables to utilize in analysis. In addition to this work, we have now completed extracting data on suicide attempts (the definition can be found at: https://github.com/andreafernandes/Defining_Suicide_Attempt_for_TextMining/blob/master/Readme.md) and will be using this as a main measure of suicidal behaviour (together with pre-existing data on deaths by suicide). The variables required for analyses after the current study is complete, are ready to be utilized.

Phase 3) Identify what the most common antidepressant prescriptions are (Results can be found at: https://github.com/andreafernandes/Appendices-for-MHRUK-report

Phase 4) Identify what factors are associated with each of the most common antidepressant prescriptions

Phase 5) Using propensity score analysis, ascertain what the association is of each of the common antidepressant prescriptions with suicide attempt and completed suicide.

Results from Phase 4 (work in progress)

After a four-month training course, I have started analyzing factors associated with the most common antidepressant user patient groups. This analysis is on-going. Here are the results so far:

Patients prescribed SSRIs or both SNRIs and SSRIs or TCAs and SSRIs or Mirtazapine are more likely to be on antipsychotics and would have been an inpatient in the past as well. Symptoms such as flat affect, hallucinations, anergia, anhedonia and hopelessness also are associated with these antidepressant prescriptions. These are only a brief summary and, given the similarities in factors associated with different antidepressant treatment, our study suggests that studies comparing antidepressant agents may not benefit greatly by further adjusting for risks associated with outcomes based on variables obtained from medical records.

Phase 4 is being completed in the two weeks and a paper produced in two additional weeks. After a brief period of training, analyses Phase 5 will begin which will answer the main question of this research project. 

2015 Report

Background


It is important to understand how treatments for depression affect one of its most lethal symptoms – suicide and suicidal behaviours (suicidality). There are several ways to treat depression – clinically and non-clinically. In this PhD I am focusing on the two main treatment forms - psychotherapy and antidepressant medication. 
With regards to antidepressants, literature to date suggests that in fact antidepressants may not have an overtly beneficial effect on symptoms of suicidality (Rihmer and Akiskal 2006, Simon, Savarino et al. 2006). However, this evidence is weak because suicidality is rare (but this does not dampen the importance of this research because of the lethality of extreme suicidal behaviour) and research is limited by the quality of information or data available. In addition, there may be biases to publish solely positive results. 
With psychotherapy, there is very little research to know what the effects of the therapy are on suicidality (Cuijpers, de Beurs et al. 2013). 

What we are attempting to do 

We are attempting to overcome the limitations faced by studies to date, to better understand the association of depression treatments with suicidality. We are proposing to do this by using a real clinical database as our data source. This database will be
mined - using text mining algorithms - for empirical data to answer the following questions: 
1. What characteristics dictate referral to psychotherapy or antidepressant therapy or both?
2. How does antidepressant or psychotherapy
treatment affect suicidality over time?
3. Do antidepressants give rise to increased suicidality?

Challenges we face and Future Work

The main challenge we face in this project is extracting data variables from our clinical database. During these last few months, I have been investigating two text-mining algorithms to extract data variables from the clinical database. Investing time in data extraction now, will mean efficient analysis in the next couple of years. To date, we have tested a bespoke machine learning support vector based algorithm, to extract defined suicidal ideation variables from our dataset. 
The results of this approach were not successful (i.e. we failed to extract data on suicidal ideation from the clinical database) which goes to demonstrate the difficulties in capturing suicidality. We are now investigating ways to extract suicidality, using an algorithm which works using rules assigned to it to extract data.

References
Cuijpers, P., D. P. de Beurs, B. A. van Spijker, M. Berking, G. Andersson and A. J. Kerkhof (2013). "The effects of psychotherapy for adult depression on suicidality and hopelessness: a systematic review and meta-analysis." J Affect Disord 144(3): 183-190. Rihmer, Z. and H. Akiskal (2006). "Do antidepressants t(h)reat(en) depressives? Toward a clinically judicious formulation of the antidepressant-suicidality FDA advisory in light of declining national suicide statistics from many countries." J Affect Disord 94(1-3): 3-13. Simon, G. E., J. Savarino, B. Operskalski and P. S. Wang
(2006). "Suicide risk during antidepressant treatment." Am J Psychiatry 163(1): 41-47.”