Clinical observations suggest that individuals with a diagnosis of bipolar face difficulties regulating emotions and impairments to their cognitive processing, which can contribute to high-risk behaviours. However, there are few studies which explore the types of risk-taking behaviour that manifest in reality and evidence suggests that there is currently not enough support for the management of these behaviours. This study examined the types of risk-taking behaviours described by people who live with bipolar and their access to support for these behaviours. Semi-structured interviews were conducted with n=18 participants with a lived experience of bipolar and n=5 healthcare professionals. The interviews comprised open ended questions and a Likert-item questionnaire. The responses to the interview questions were analysed using content analysis and corpus linguistic methods to develop a classification system of risk-taking behaviours. The Likert-item questionnaire was analysed statistically and insights from the questionnaire were incorporated into the classification system. Our classification system includes 39 reported risk-taking behaviours which we manually inferred into six domains of risk-taking. Corpus linguistic and qualitative analysis of the interview data demonstrate that people need more support for risk-taking behaviours and that aside from suicide, self-harm, and excessive spending, many behaviours are not routinely monitored. This study shows that that people living with bipolar report the need for improved access to psychologically informed care, and that a standardised classification system or risk-taking questionnaire could act as a useful elicitation tool for guiding conversations around risk-taking to ensure that opportunities for intervention are not missed. We have also presented a novel methodological framework which demonstrates the utility of computational linguistic methods for the analysis of health research data.Individuals living with bipolar disorder are likely to engage in behaviours which can be risky for themselves or others. This includes increased prevalence of suicide and self-harm, excessive spending, alcohol or drug use and risky sexual behaviour. Understanding more about this behaviour is crucial as with the right help people living with bipolar "have the potential to return to normal function with optimal treatment". Current psychological models of bipolar explain risky behaviour as an attempt to avoid low mood, a response to mood elevation or to impulsivity/sensitivity to reward. These approaches have informed the development of psychological interventions to improve coping strategies for mood change. However, the effectiveness of such approaches is mixed and evidence is lacking for improvements in the functional and recovery outcomes which qualitative research has shown are valued. Current research has relied on questionnaire measures of hypothesised processes, which limits what can be learnt about the subjective experiences of people living with bipolar. For instance, they tell us little about how such individuals define risk, why they chose to engage in some such behaviours and how socially normative such behaviour might be. It is clear therefore, that a mixed method approach is needed to understand the processes which underpin risk in bipolar. This should combine in-depth qualitative approaches with methods that explore how people describe their experiences in natural language, not constrained by typical research or clinical settings. This is particularly important for risky behaviour that is likely to have been stigmatised.

The interviews took place remotely via Microsoft Teams or by telephone and were either recorded directly through Microsoft Teams or through an encrypted device (in accordance with the ethics approval granted to this project by Lancaster University), and all recordings were converted to M4A format. The audio recordings were then imported to NVivo transcription software ​which processed the audio and attempted to automatically convert the speech to text. This process increased the efficiency and speed of the overall transcription process but the resulting text after automatic transcription still included errors. The transcripts were edited manually to correct errors and to remove hesitations, non-verbal utterances (e.g. ‘uh’ and ‘er’) and repetition, resulting in a naturalised form of transcription (deemed acceptable for this task because the analysis is semantic not phonetic). This study integrated two primary sources of data 1) textual data from interview transcripts and 2) statistical data from a risk-taking measure, with both sets of data generated during the interviews. The analysis framework for this data included three stages. First, content analysis was conducted on the interview transcripts to encode risk-taking behaviours described by participants and create a classification system. Second, we performed statistical analysis of a risk-taking questionnaire ​risk-taking questionnaire which was completed by the same interview participants as part of the interview process, and finally we used corpus analysis to identify how risk-taking behaviours were talked about in relation to the four key themes which formed the structure of the interview schedule.