Human language is a fundamental ability which, in addition to social interaction, primarily serves the exchange of feelings; however, in the sense of an inner dialogue, it also serves as a means of self-perception and self-reflection [1]. This establishes a relationship between pure content as an expression of cognitive performance and emotional content in the sense of affective meaning. This close connection between emotion and cognition, especially with language and speech, has been scientifically researched within the framework of relational frame theory for many years [2]. Human language as a form of expression, with human listening on the other side, has been the basis of professional therapeutic support for mental disorders for more than 100 years [3,4]. The recent treatment developments regarding linguistic and psychological research in relational frame theory have been acceptance-based treatments, such as acceptance and commitment therapy [5]. 

Due to the increasing global frequency of mental disorders, their safe detection and treatment is very important. According to World Health Organization calculations [6], depression alone will be the second greatest burden of disease in the world in 2020. Other mental disorders, such as neurodevelopmental disorders (attention deficit hyperactivity disorder [ADHD], autism spectrum disorders, etc [7,8]) or personality disorders [9], which have been increasing in recent years, also cause significant individual stress and have a societal impact due to the loss of potential productivity and treatment costs [10]. Early detection and treatment of mental disorders is, therefore, a major factor in preventing these problems, as is the avoidance of deterioration. However, there is currently no objective, ideal physiological parameter available for this task.

Since the 1990s, attempts have been made to determine the emotional state of a speaker by analyzing the content of their speech and the linguistic structure [11]. There have been positive results thus far, with high specificity and sensitivity in some subjects when using suitable parameters, like in cases where depressive disorders have been detected [12-14]. As a result, there has been intensive development in the automation of such analyses, with continuing positive results [14]. However, so far dissemination of such analyses, such as in the early detection, diagnosis, or assessment of mental disorders, has not occurred in broad clinical practice but has been described in a few single cases [15,16]. The reasons for this may be complex, but it may be due to the high technical effort required to perform these analyses that have led to them not being well reflected in outpatient care practice [17]. Also, resentment against machine diagnosis or very technical measurements may have played a role that should not be underestimated. Finally, the diagnostic and prognostic validity of such measures still needs to be proven.

In the last decade, electronic assistants have taken up considerable space in all areas of life, first through computerization and later via smartphones [18]. In medicine, and especially in psychiatry and psychotherapy, numerous programs, online tools, and coaching applications are now available and have seen increasing acceptance [19]. At the same time, the technical effort that needs to be expended to create a high-quality, medically or psychologically useful application for use in the clinic has decreased [20,21]. The applications that can be implemented in this way are even more complex and extensive than previously thought possible. It may, therefore, make sense to use these new technical possibilities and examine whether the analysis of language, especially the noncontent aspects (eg, speech flow, speech melody, expression), can be used by a differentiated algorithm to assess the mental state of a subject. It is possible that they could also measure the course of mental illness, the quality of life of patients with chronic mental illnesses, and patients' well-being [22].

The study presented here uses a noncontent linguistic analysis algorithm developed by VoiceSense, a company specializing in prosodic speech analysis, which links speech patterns to behavioral tendencies. The analysis is commercially used for personal risk assessment by banks and insurance companies, for candidate assessment by human resources companies, and customer analytics by enterprise call centers. The current study would provide a structured setting for the differentiated evaluation of psychopathological factors (effects, personality aspects, psychomotor factors, etc) by this speech pattern analysis.

The primary objective of this study is to evaluate the measurements of the mental state of the participants by comparing the analysis of noncontent speech parameters to the results of several psychological questionnaires (Symptom Checklist-90 [SCL-90], the Patient Health Questionnaire [PHQ], and the Big 5 Test). We hypothesized that the emotional state of mind, the behavioral patterns, the well-being, and some of the personality traits of the participants could be identified using the prosodic analysis algorithm. A second hypothesis is that the diagnosis could be confirmed by the application of the analysis program, and a distinction can be made between different psychopathologically-defined syndromes.

A total of 166 outpatients will be recruited in one center (Neuropsychiatric Center of Hamburg), following the invitation of their treating physician or psychotherapist, to use a new method to measure their mental state and distress. Inclusion and exclusion criteria for the study are listed in Textbox 1.

For the exclusion criteria, both schizophrenia and dementia have shown indications that a patient’s speech patterns are influenced by these disorders. Therefore, we decided to consider these two groups of patients separately in a future study to see whether there is a comparable difference to the general population on a content-free linguistic level.

This study is a case-controlled study with a case group and a control group to identify the best target population for measuring speech patterns with the VoiceSense app. We adhered to the Consolidated Standards of Reporting Trials (CONSORT) guidelines in the design of the trial.

In the first step, the participants will be grouped by the physician or psychotherapist into a control group (participants without mental disorders) or a case group (participants with mental disorders). All participants will provide written informed consent after being given detailed information by the treating therapist or physician. The original forms will be stored in a closed cabinet, but the participants will receive a copy. Second, they will receive a pseudonymized label for the blinded inhouse-rater to work with the questionnaires, and for the data center for the speech-analysis. Neither the rater nor the data center will know the name and diagnosis of the participant. Data security and availability will be ensured every time, according to the European Union (EU) rules [29]. The participants’ data will be stored in a MySQL-Database. Within the project database, identifying data will be stored separately from the collected data, with only project staff with specifically conferred access rights able to access the identifying data. It is envisioned that the distribution of participants will correspond to the clinical distribution in the Neuropsychiatric Center with different degrees of severity of mental disorders (about 25% mild, about 50% moderate, about 25% severe). This should be checked by the study nurse every 2 weeks.

For documentation purposes, the presence or absence (in the case of neurological outpatients) of any psychiatric diagnosis will be screened using the PHQ-D [30]. Personality traits will then be measured with the Big 5 Test [23]. In the case of ADHD, we added a Visual Analogue Scale (ADHD-VAS) to identify the three main affective states (impulsivity, inattention, and hyperactivity). This will be done at baseline only.

At baseline, at follow-up one (2-4 weeks after baseline), and follow-up two (2-4 weeks after follow-up 1), participants will be measured for mental distress using the SCL-90 [31]. They will then be asked to freely answer nine standardized questions, with a few sentences per question. These answers will be recorded for evaluation. If less than 120 seconds of free speech is possible, three more standardized questions will be asked and the answers to these will also be recorded. The recorded sentences will then be sent to the server to carry out the prosodic analysis. The results of the questionnaires and the results of the prosodic analysis will be stored in the patients’ records on file. Figure 1 shows the flow chart of the study.

It should be possible for the study team to evaluate the efficacy and effectiveness of the raw data from the linguistic analysis in correlation with the clinical data collected with the questionnaires. All statistical tests will be two-tailed and will be considered statistically significant at P<.05. A sample size of 166 subjects was calculated (effect size medium=0.30; alpha=0.05; 1–beta=0.95), and considering a drop-out rate of 15%, a total sample size of 190 patients was determined. 

Statistical analysis will be performed with R version 3.5.3 (The R Project, Vienna, Austria), and epidemiological data will be evaluated. Continuous variables will be described with the mean, standard deviation, median, minimum, maximum, and the 25th and 75th percentiles. Categorical variables will be described with percentages and absolute frequencies. The differences in continuous variables between the two groups will be evaluated with the Kruskal-Wallis test, followed by the Dunn multiple comparison test. The differences between the two diagnostic groups (psychiatric versus neurological diagnosis) for normally distributed data will be evaluated with a one-way analysis of variance (ANOVA), followed by the Newman-Keuls multiple comparison test. The normality of the distribution will be evaluated with the Kolmogorov-Smirnov test. Any correlation between the variables under evaluation will be assessed by the Spearman r correlation. To compare qualitative data, we will use the Chi-square test with the Yates correction or the Fisher exact test. Lastly, the variables will be grouped to their diagnostic groups according to the 10th edition of the International Statistical Classification of Diseases and Related Health Problems or the fifth edition of the Diagnostic and Statistical Manual of Mental Disorders, and a confirmative factor analysis will be done comparing these groups with the prosodic parameters.

This study protocol has been approved by the ethical committee of the Neuropsychiatric Center under the Declaration of Helsinki. Written informed consent will be collected from the participants before they will be included in the study.