Learning to Detect Bipolar Disorder and Borderline Personality Disorder with Language and Speech in Non-Clinical Interviews

Bipolar disorder (BD) is a recurrent chronic mental health condition which occurs in approximately 1% of the global population [1]. It is characterised by episodes of low and high mood which cause significant interference with everyday life. Borderline personality disorder (BPD) is characterised by a long-term pattern of constantly variable mood, self-image and behaviour. Although BD and BPD are two very different conditions they share some similar symptoms such as mood instability and impulsive behaviour [2]. A recent study [3] reported the high prevalence of comorbidity between the two conditions, with up to 21.6% of individuals with BD found to have comorbid BPD. As a result they can be difficult to distinguish, but accurate diagnosis is crucial as they require different treatment [4, 5]. Standard diagnostic assessment involves a psychiatrist asking a series of questions about symptoms and the person has to retrospectively describe their account of these symptoms. The success of the assessment also relies on how the psychiatrist interprets both the verbal and non-verbal cues drawn from the person’s responses. In this work, we aim to develop a method that extracts cues automatically from interviews conducted in a non-clinical setting, to assist the existing assessment framework, which is expensive and subjective.

Recent studies have explored data driven approaches to automatically screen patients, incorporating features extracted from multiple modalities in clinical interviews, showing diagnostic value for mental health conditions such as depression and bipolar disorder [6, 7, 8, 9]. [8] finds the performance of automatic mood detection to be much better in clinical interactions than in personal conversations, and there are significant differences in the features important to each type of interaction. While existing studies of BD have focused on recognising mood episodes, the distinction between BD and BPD remains understudied. In this paper, we aim to bridge this gap by presenting a multi-modal (i.e. speech and text) dataset containing interviews in a non-clinical setting involving individuals with a diagnosis of BD or BPD, and study the automatic assessment of the two mental health conditions.

Motivated to study the interaction between the interviewer and participant during the course of an interview from different aspects (including linguistic complexity, semantic content and dialogue structure), we investigate features extracted from different modalities. Path signatures, initially introduced in rough path theory as a branch of stochastic analysis, has been shown to be successful in a range of machine learning tasks involving modelling temporal dynamics [10, 11, 12]. We propose to apply path signatures for summarising features extracted from each utterance, sentence and speaker-turn into interview-level feature representations, given its ability to naturally capture the order of events. By doing so, we automatically include more non-linear prior knowledge in our final feature set, which leads to effective classification, even with a simple linear classifier.

The contributions of this work are as follows: (1) We present a new non-clinical interview dataset involving BD and BPD patients; (2), We investigate different feature types and propose using path signatures as a novel approach of summarising turn-level features; (3) We demonstrate a good linear model can be learnt for three classification tasks, and provide insights into the distinction between BD and BPD by analysing the importance of the selected features.

The original Automated Monitoring of Symptoms Severity (AMoSS) study [13, 10] was a longitudinal study during which a range of wearables in combination with a bespoke smartphone app were used for the daily self-monitoring of mood instability. Among the 139 participants enrolled in the study, 53 had a BD diagnosis, 33 had been diagnosed with BPD and 53 were healthy volunteers. All the diagnoses had been confirmed prior to the study using the structured clinical interview for DSM-IV (the 4th edition of Diagnostic and Statistical Manual of Mental Disorders) and the International Personality Disorder Examination (IPDE) [14]. The majority of the BD participants were euthymic while BPD participants were not in crisis but were still symptomatic as is the case with chronic experience of the condition. Exclusion criteria for BD and BPD were comorbidity of each diagnosis [15].

62 participants were interviewed halfway through the study to gather qualitative feedback and discuss potential improvement¹¹1The study protocol was approved by the NRES Committee East of England—Norfolk (13/EE/0288), and all 62 participants consented to be interviewed and for those interviews to be recorded.. Each participant was interviewed only once. These semi-structured, one-on-one qualitative interviews took place either in person or by telephone, conducted by 2 clinicians and 2 psychology graduates who were involved in the roll out of the AMoSS study. The natural conversations recorded between the interviewer and participant are usually within the scope of: 1) experience using the mood reporting app and the questionnaires in the app; 2) experience using different wearable devices; 3) benefits of taking part in the study and discussion of potential improvement, making the interviews semi-structured.

The AMoSS interview (AMoSS-I) dataset we study here consists of 50 randomly sampled interviews that were initially transcribed by the same interviewers. The audio recordings and manually transcribed text were then aligned by a Sakoe-Chiba Band Dynamic Time Warping based forced speech alignment model in aeneas²²2https://github.com/readbeyond/aeneas. This was followed by a convolutional neural network (CNN) based noise-robust speech segmentation [16], generating speaker-turn-level alignments in the time domain. The manual transcripts and automatically generated time alignments were then reviewed and improved by three research assistants, where each interview was reviewed twice to ensure quality and consistency. We used finetuneas³³3https://github.com/ozdefir/finetuneas as the annotation interface for reviewing. The demographic details of the participants are summarised in Table 1. We can see both BIS-11⁴⁴4The Barratt Impulsiveness Scale (BIS-11) [17] is a self-report questionnaire designed to assess the personality/behavioral construct of impulsiveness. Higher BIS-11 scores are indicative of higher impulsivity. and IPDE scores are higher among BD and BPD patients compared to controls. We also observe from the density plots in Figure 1 that distributions of the three user groups are very similar in interview length and number of participant responses.

We identify a set of features that are motivated by existing work in automatic mental health assessment and clinical studies of bipolar disorder symptoms. These features are selected to allow studying the interviews from different aspects : lexical diversity and density, syntax, semantic content and dialogue structure.

Linguistic complexity features (LING): Previous studies have shown language disturbances such as pressure of speech in mania and poverty of speech in depression are among the main symptoms of acute episodes in BD [20, 21]. We adopt a set of linguistic complexity measures used in [9], including measurements for lexical diversity such as moving average type-to-token ratio (MATTR), brunét’s index (BI) and honoré’s statistic (HS); lexical density such as number of function words per word (FUNC/W) and number of interjections per word (UH/W); and mean length sentence (MLS). These features are shown to be effective in distinguishing patients with schizophrenia and bipolar I disorder. We also use the dependency-based propositional idea density (DEPID), originally proposed in [22] for measuring the rate of propositions or ideas expressed per word in spontaneous speech transcripts.

In addition to the aforementioned measures we extract part-of-speech related variables including: first person pronouns (it has been shown that people suffering from depression use more first person pronouns in written [23] and spoken language [24]); swear words; speech disfluencies; filler words/phrases such as “okay” and “you know”, using the LIWC-2015 dictionary [25]. Finally, we add the number of absolutist words per word (ABS/W) as a feature using a 19-word absolutist dictionary curated by clinical psychologists [26], who found that the online anxiety, depression, and suicidal ideation forums contained more absolutist words than the control ones. Overall we extract 28 linguistic features, denoted as LING.

Semantic content features (CNT): We extract content related to psychological states from the transcripts by applying relevant categories of LIWC [27] such as emotions (e.g. anxiety), social processes (e.g. family) and drives (e.g. reward and risk). We also apply the empathetic concern and personal distress lexica [28], and an optimism lexicon (i.e. positive future-oriented thinking), by taking the sum over all weighted words multiplying their relative frequencies in the speaker turn. Overall we use 19 content features.

Dialogue features (DIAL): We use the set of high-level turn-taking behaviour related features proposed in [7], namely relative floor control, turn hold offset (i.e. short pauses that are less than half a second), number of consecutive turns (separated by longer than half a second pauses), turn switch offset and turn length, per speaker turn. Turn switch offset measures the latency between speaker turns, and it is shown that depressed people take more time to respond to questions by clinicians [29]. While Relative floor control measures the percentage of time an individual controlled the conversation floor up to the time of speaking, we also add relative turn length that measures the percentage of the length of the current turn relative to the average turn length up to the time of speaking. We compute both features in seconds as well as in number of words, following [30].

Talking over or interrupting others, is one of the characteristics for pressure speech, which has shown to be a significant feature in bipolar mania [31, 32]. We use a LSTM-based overlapping speech detection model proposed in [33] to extract the number and duration of speech overlaps in each speaker turn as features. Additionally, we add the number of words per second per turn, as a feature representing speaking speed.

We average turn hold offset and speech overlaps per turn, in addition to number of consecutive turns, relative floor control (time), relative floor control (words), relative turn length (time), relative turn length (words), turn switch offset, turn length and number of words per second, which results in 11 features representing each speaker turn.

Following previous work we chose leave-one-participant-out (LOOCV) as the evaluation scheme, and logistic regression as the classification model given our preference over interpretability and the size of our data. For each fold, we first apply signature transform to each type of turn-level features, and keep only the first three levels of the path signature⁶⁶6We use iisignature Python library, https://pypi.org/project/iisignature/, and set the maximal order to which iterated integration is performed in signature to be 3.. To avoid overfitting, we conduct feature selection on signature-transformed interview-level features through computing Pearson Correlation Coefficients (PCC) with the IPDE scores on the training data and retain the features with $p$-values less than 0.001. This results in a small number of features. The selected features are then fed to the classifier for 3 separate binary tasks: (1), BD vs. healthy controls, (2), BPD vs. healthy controls, and (3), BD vs. BPD patients. We conduct three separate experiments, extracting features from the speech of each participant and interviewer respectively, as well as the whole interview (as a sequence of turns) without speaker identification (denoted as ‘Both’)⁷⁷7For the ‘Interviewer’ and ‘Both’ experiments, we increase the p value threshold to 0.002..

In this paper, we demonstrate the potential of using features extracted from language and speech in non-clinical interviews to assist the assessment of bipolar disorder BD and borderline personality disorder BPD, which is challenging for clinicians to distinguish. We first presented a non-clinical interview dataset, named AMoSS-I, conducted partially by psychology graduates, for the task of detecting BD and BPD. We demonstrated good performance in three classification tasks using down-selected features and a new way of summarising these features based on path signatures. Lastly, we showed the importance of linguistic features in all three tasks and the benefits of feature fusion from different modalities. For future work, we plan to learn acoustic features, and investigate the effect of acoustic properties of the interviews and the impact of recording environments.

This work was supported by the MRC Mental Health Data Pathfinder award to the University of Oxford [MC_PC_17215], by the NIHR Oxford Health Biomedical Research Centre and by the The Alan Turing Institute under the EPSRC grant EP/N510129/1. We would also like to thank Zakaria Aldeneh and Kairit Sirts for sharing code and thoughts, and to Priyanka Panchal and Rota Silva for conducting the interviews. The views expressed are those of the authors and not necessarily those of the NHS, NIHR or the Department of Health.