Assessing the accuracy of automatic speech recognition for psychotherapy

Returning Individual Research Results from Digital Phenotyping in Psychiatry

Psychiatry is rapidly adopting digital phenotyping and artificial intelligence/machine learning tools to study mental illness based on tracking participants' locations, online activity, phone and text message usage, heart rate, sleep, physical activity, and more. Existing ethical frameworks for return of individual research results (IRRs) are inadequate to guide researchers for when, if, and how to return this unprecedented number of potentially sensitive results about each participant's real-world behavior. To address this gap, we convened an interdisciplinary expert working group, supported by a National Institute of Mental Health grant. Building on established guidelines and the emerging norm of returning results in participant-centered research, we present a novel framework specific to the ethical, legal, and social implications of returning IRRs in digital phenotyping research. Our framework offers researchers, clinicians, and Institutional Review Boards (IRBs) urgently needed guidance, and the principles developed here in the context of psychiatry will be readily adaptable to other therapeutic areas.

Using HIPAA (Health Insurance Portability and Accountability Act)-Compliant Transcription Services for Virtual Psychiatric Interviews: Pilot Comparison Study

BACKGROUND

Automatic speech recognition (ASR) technology is increasingly being used for transcription in clinical contexts. Although there are numerous transcription services using ASR, few studies have compared the word error rate (WER) between different transcription services among different diagnostic groups in a mental health setting. There has also been little research into the types of words ASR transcriptions mistakenly generate or omit.

OBJECTIVE

This study compared the WER of 3 ASR transcription services (Amazon Transcribe [Amazon.com, Inc], Zoom-Otter AI [Zoom Video Communications, Inc], and Whisper [OpenAI Inc]) in interviews across 2 different clinical categories (controls and participants experiencing a variety of mental health conditions). These ASR transcription services were also compared with a commercial human transcription service, Rev (Rev.Com, Inc). Words that were either included or excluded by the error in the transcripts were systematically analyzed by their Linguistic Inquiry and Word Count categories.

METHODS

Participants completed a 1-time research psychiatric interview, which was recorded on a secure server. Transcriptions created by the research team were used as the gold standard from which WER was calculated. The interviewees were categorized into either the control group (n=18) or the mental health condition group (n=47) using the Mini-International Neuropsychiatric Interview. The total sample included 65 participants. Brunner-Munzel tests were used for comparing independent sets, such as the diagnostic groupings, and Wilcoxon signed rank tests were used for correlated samples when comparing the total sample between different transcription services.

RESULTS

There were significant differences between each ASR transcription service's WER (P<.001). Amazon Transcribe's output exhibited significantly lower WERs compared with the Zoom-Otter AI's and Whisper's ASR. ASR performances did not significantly differ across the 2 different clinical categories within each service (P>.05). A comparison between the human transcription service output from Rev and the best-performing ASR (Amazon Transcribe) demonstrated a significant difference (P<.001), with Rev having a slightly lower median WER (7.6%, IQR 5.4%-11.35 vs 8.9%, IQR 6.9%-11.6%). Heat maps and spider plots were used to visualize the most common errors in Linguistic Inquiry and Word Count categories, which were found to be within 3 overarching categories: Conversation, Cognition, and Function.

CONCLUSIONS

Overall, consistent with previous literature, our results suggest that the WER between manual and automated transcription services may be narrowing as ASR services advance. These advances, coupled with decreased cost and time in receiving transcriptions, may make ASR transcriptions a more viable option within health care settings. However, more research is required to determine if errors in specific types of words impact the analysis and usability of these transcriptions, particularly for specific applications and in a variety of populations in terms of clinical diagnosis, literacy level, accent, and cultural origin.

Natural language processing for mental health interventions: a systematic review and research framework

Neuropsychiatric disorders pose a high societal cost, but their treatment is hindered by lack of objective outcomes and fidelity metrics. AI technologies and specifically Natural Language Processing (NLP) have emerged as tools to study mental health interventions (MHI) at the level of their constituent conversations. However, NLP's potential to address clinical and research challenges remains unclear. We therefore conducted a pre-registered systematic review of NLP-MHI studies using PRISMA guidelines (osf.io/s52jh) to evaluate their models, clinical applications, and to identify biases and gaps. Candidate studies (n = 19,756), including peer-reviewed AI conference manuscripts, were collected up to January 2023 through PubMed, PsycINFO, Scopus, Google Scholar, and ArXiv. A total of 102 articles were included to investigate their computational characteristics (NLP algorithms, audio features, machine learning pipelines, outcome metrics), clinical characteristics (clinical ground truths, study samples, clinical focus), and limitations. Results indicate a rapid growth of NLP MHI studies since 2019, characterized by increased sample sizes and use of large language models. Digital health platforms were the largest providers of MHI data. Ground truth for supervised learning models was based on clinician ratings (n = 31), patient self-report (n = 29) and annotations by raters (n = 26). Text-based features contributed more to model accuracy than audio markers. Patients' clinical presentation (n = 34), response to intervention (n = 11), intervention monitoring (n = 20), providers' characteristics (n = 12), relational dynamics (n = 14), and data preparation (n = 4) were commonly investigated clinical categories. Limitations of reviewed studies included lack of linguistic diversity, limited reproducibility, and population bias. A research framework is developed and validated (NLPxMHI) to assist computational and clinical researchers in addressing the remaining gaps in applying NLP to MHI, with the goal of improving clinical utility, data access, and fairness.

Machine learning-based speech recognition system for nursing documentation - A pilot study

PURPOSE

Considering the significant workload of nursing tasks, enhancing the efficiency of nursing documentation is imperative. This study aimed to evaluate the effectiveness of a machine learning-based speech recognition (SR) system in reducing the clinical workload associated with typing nursing records, implemented in a psychiatry ward.

METHODS

The study was conducted between July 15, 2020, and June 30, 2021, at Cheng Hsin General Hospital in Taiwan. The language corpus was based on the existing records from the hospital nursing information system. The participating ward's nursing activities, clinical conversation, and accent data were also collected for deep learning-based SR-engine training. A total of 21 nurses participated in the evaluation of the SR system. Documentation time and recognition error rate were evaluated in parallel between SR-generated records and keyboard entry over 4 sessions. Any differences between SR and keyboard transcriptions were regarded as SR errors.

FINDINGS

A total of 200 data were obtained from four evaluation sessions, 10 participants were asked to use SR and keyboard entry in parallel at each session and 5 entries were collected from each participant. Overall, the SR system processed 30,112 words in 32,456 s (0.928 words per second). The mean accuracy of the SR system improved after each session, from 87.06% in 1st session to 95.07% in 4th session.

CONCLUSION

This pilot study demonstrated our machine learning-based SR system has an acceptable recognition accuracy and may reduce the burden of documentation for nurses. However, the potential error with the SR transcription should continually be recognized and improved. Further studies are needed to improve the integration of SR in digital documentation of nursing records, in terms of both productivity and accuracy across different clinical specialties.

HEAR4Health: a blueprint for making computer audition a staple of modern healthcare

Recent years have seen a rapid increase in digital medicine research in an attempt to transform traditional healthcare systems to their modern, intelligent, and versatile equivalents that are adequately equipped to tackle contemporary challenges. This has led to a wave of applications that utilise AI technologies; first and foremost in the fields of medical imaging, but also in the use of wearables and other intelligent sensors. In comparison, computer audition can be seen to be lagging behind, at least in terms of commercial interest. Yet, audition has long been a staple assistant for medical practitioners, with the stethoscope being the quintessential sign of doctors around the world. Transforming this traditional technology with the use of AI entails a set of unique challenges. We categorise the advances needed in four key pillars: Hear, corresponding to the cornerstone technologies needed to analyse auditory signals in real-life conditions; Earlier, for the advances needed in computational and data efficiency; Attentively, for accounting to individual differences and handling the longitudinal nature of medical data; and, finally, Responsibly, for ensuring compliance to the ethical standards accorded to the field of medicine. Thus, we provide an overview and perspective of HEAR4Health: the sketch of a modern, ubiquitous sensing system that can bring computer audition on par with other AI technologies in the strive for improved healthcare systems.

A computational approach to measure the linguistic characteristics of psychotherapy timing, responsiveness, and consistency

Although individual psychotherapy is generally effective for a range of mental health conditions, little is known about the moment-to-moment language use of effective therapists. Increased access to computational power, coupled with a rise in computer-mediated communication (telehealth), makes feasible the large-scale analyses of language use during psychotherapy. Transparent methodological approaches are lacking, however. Here we present novel methods to increase the efficiency of efforts to examine language use in psychotherapy. We evaluate three important aspects of therapist language use - timing, responsiveness, and consistency - across five clinically relevant language domains: pronouns, time orientation, emotional polarity, therapist tactics, and paralinguistic style. We find therapist language is dynamic within sessions, responds to patient language, and relates to patient symptom diagnosis but not symptom severity. Our results demonstrate that analyzing therapist language at scale is feasible and may help answer longstanding questions about specific behaviors of effective therapists.

Evaluating Web-Based Automatic Transcription for Alzheimer Speech Data: Transcript Comparison and Machine Learning Analysis

Background

Speech data for medical research can be collected noninvasively and in large volumes. Speech analysis has shown promise in diagnosing neurodegenerative disease. To effectively leverage speech data, transcription is important, as there is valuable information contained in lexical content. Manual transcription, while highly accurate, limits the potential scalability and cost savings associated with language-based screening.

Objective

To better understand the use of automatic transcription for classification of neurodegenerative disease, namely, Alzheimer disease (AD), mild cognitive impairment (MCI), or subjective memory complaints (SMC) versus healthy controls, we compared automatically generated transcripts against transcripts that went through manual correction.

Methods

We recruited individuals from a memory clinic (“patients”) with a diagnosis of mild-to-moderate AD, (n=44, 30%), MCI (n=20, 13%), SMC (n=8, 5%), as well as healthy controls (n=77, 52%) living in the community. Participants were asked to describe a standardized picture, read a paragraph, and recall a pleasant life experience. We compared transcripts generated using Google speech-to-text software to manually verified transcripts by examining transcription confidence scores, transcription error rates, and machine learning classification accuracy. For the classification tasks, logistic regression, Gaussian naive Bayes, and random forests were used.

Results

The transcription software showed higher confidence scores (P<.001) and lower error rates (P>.05) for speech from healthy controls compared with patients. Classification models using human-verified transcripts significantly (P<.001) outperformed automatically generated transcript models for both spontaneous speech tasks. This comparison showed no difference in the reading task. Manually adding pauses to transcripts had no impact on classification performance. However, manually correcting both spontaneous speech tasks led to significantly higher performances in the machine learning models.

Conclusions

We found that automatically transcribed speech data could be used to distinguish patients with a diagnosis of AD, MCI, or SMC from controls. We recommend a human verification step to improve the performance of automatic transcripts, especially for spontaneous tasks. Moreover, human verification can focus on correcting errors and adding punctuation to transcripts. However, manual addition of pauses is not needed, which can simplify the human verification step to more efficiently process large volumes of speech data.

Experiences of a Speech-Enabled Conversational Agent for the Self-Report of Wellbeing Among People Living with Affective Disorders: An In-The-Wild Study

The growing commercial success of smart speaker devices following recent advancements in speech recognition technology has surfaced new opportunities for collecting self-reported health and wellbeing data. Speech-enabled conversational agents (CAs) in particular, deployed in home environments using just such systems, may offer increasingly intuitive and engaging means of self-report. To date, however, few real-world studies have examined users’ experiences of engaging in the self-report of mental health using such devices, nor the challenges of deploying these systems in the home context. With these aims in mind, this paper recounts findings from a four-week ‘in-the-wild’ study during which 20 individuals with depression or bipolar disorder used a speech-enabled CA named ‘Sofia’ to maintain a daily diary log, responding also to the WHO-5 wellbeing scale every two weeks. Thematic analysis of post-study interviews highlights actions taken by participants to overcome CAs’ limitations, diverse personifications of a speech-enabled agent, and unique forms of valuing of this system among users’ personal and social circles. These findings serve as initial evidence for the potential of CAs to support the self-report of mental health and wellbeing, while highlighting the need to address outstanding technical limitations in addition to design challenges of conversational pattern matching, filling unmet interpersonal gaps, and the use of self-report CAs in the at-home social context. Based on these insights, we discuss implications for the future design of CAs to support the self-report of mental health and wellbeing.

Automated evaluation of psychotherapy skills using speech and language technologies

With the growing prevalence of psychological interventions, it is vital to have measures which rate the effectiveness of psychological care to assist in training, supervision, and quality assurance of services. Traditionally, quality assessment is addressed by human raters who evaluate recorded sessions along specific dimensions, often codified through constructs relevant to the approach and domain. This is, however, a cost-prohibitive and time-consuming method that leads to poor feasibility and limited use in real-world settings. To facilitate this process, we have developed an automated competency rating tool able to process the raw recorded audio of a session, analyzing who spoke when, what they said, and how the health professional used language to provide therapy. Focusing on a use case of a specific type of psychotherapy called "motivational interviewing", our system gives comprehensive feedback to the therapist, including information about the dynamics of the session (e.g., therapist's vs. client's talking time), low-level psychological language descriptors (e.g., type of questions asked), as well as other high-level behavioral constructs (e.g., the extent to which the therapist understands the clients' perspective). We describe our platform and its performance using a dataset of more than 5000 recordings drawn from its deployment in a real-world clinical setting used to assist training of new therapists. Widespread use of automated psychotherapy rating tools may augment experts' capabilities by providing an avenue for more effective training and skill improvement, eventually leading to more positive clinical outcomes.

Preliminary Evaluation of Automated Speech Recognition Apps for the Hearing Impaired and Deaf

Objective

Automated speech recognition (ASR) systems have become increasingly sophisticated, accurate, and deployable on many digital devices, including on a smartphone. This pilot study aims to examine the speech recognition performance of ASR apps using audiological speech tests. In addition, we compare ASR speech recognition performance to normal hearing and hearing impaired listeners and evaluate if standard clinical audiological tests are a meaningful and quick measure of the performance of ASR apps.

Methods

Four apps have been tested on a smartphone, respectively AVA, Earfy, Live Transcribe, and Speechy. The Dutch audiological speech tests performed were speech audiometry in quiet (Dutch CNC-test), Digits-in-Noise (DIN)-test with steady-state speech-shaped noise, sentences in quiet and in averaged long-term speech-shaped spectrum noise (Plomp-test). For comparison, the app's ability to transcribe a spoken dialogue (Dutch and English) was tested.

Results

All apps scored at least 50% phonemes correct on the Dutch CNC-test for a conversational speech intensity level (65 dB SPL) and achieved 90–100% phoneme recognition at higher intensity levels. On the DIN-test, AVA and Live Transcribe had the lowest (best) signal-to-noise ratio +8 dB. The lowest signal-to-noise measured with the Plomp-test was +8 to 9 dB for Earfy (Android) and Live Transcribe (Android). Overall, the word error rate for the dialogue in English (19–34%) was lower (better) than for the Dutch dialogue (25–66%).

Conclusion

The performance of the apps was limited on audiological tests that provide little linguistic context or use low signal to noise levels. For Dutch audiological speech tests in quiet, ASR apps performed similarly to a person with a moderate hearing loss. In noise, the ASR apps performed more poorly than most profoundly deaf people using a hearing aid or cochlear implant. Adding new performance metrics including the semantic difference as a function of SNR and reverberation time could help to monitor and further improve ASR performance.

Expectations for Artificial Intelligence (AI) in Psychiatry

PURPOSE OF REVIEW

Artificial intelligence (AI) is often presented as a transformative technology for clinical medicine even though the current technology maturity of AI is low. The purpose of this narrative review is to describe the complex reasons for the low technology maturity and set realistic expectations for the safe, routine use of AI in clinical medicine.

RECENT FINDINGS

For AI to be productive in clinical medicine, many diverse factors that contribute to the low maturity level need to be addressed. These include technical problems such as data quality, dataset shift, black-box opacity, validation and regulatory challenges, and human factors such as a lack of education in AI, workflow changes, automation bias, and deskilling. There will also be new and unanticipated safety risks with the introduction of AI. The solutions to these issues are complex and will take time to discover, develop, validate, and implement. However, addressing the many problems in a methodical manner will expedite the safe and beneficial use of AI to augment medical decision making in psychiatry.

Mitigating Patient and Consumer Safety Risks When Using Conversational Assistants for Medical Information: Exploratory Mixed Methods Experiment

BACKGROUND

Prior studies have demonstrated the safety risks when patients and consumers use conversational assistants such as Apple's Siri and Amazon's Alexa for obtaining medical information.

OBJECTIVE

The aim of this study is to evaluate two approaches to reducing the likelihood that patients or consumers will act on the potentially harmful medical information they receive from conversational assistants.

METHODS

Participants were given medical problems to pose to conversational assistants that had been previously demonstrated to result in potentially harmful recommendations. Each conversational assistant's response was randomly varied to include either a correct or incorrect paraphrase of the query or a disclaimer message-or not-telling the participants that they should not act on the advice without first talking to a physician. The participants were then asked what actions they would take based on their interaction, along with the likelihood of taking the action. The reported actions were recorded and analyzed, and the participants were interviewed at the end of each interaction.

RESULTS

A total of 32 participants completed the study, each interacting with 4 conversational assistants. The participants were on average aged 42.44 (SD 14.08) years, 53% (17/32) were women, and 66% (21/32) were college educated. Those participants who heard a correct paraphrase of their query were significantly more likely to state that they would follow the medical advice provided by the conversational assistant (χ²₁=3.1; P=.04). Those participants who heard a disclaimer message were significantly more likely to say that they would contact a physician or health professional before acting on the medical advice received (χ²₁=43.5; P=.001).

CONCLUSIONS

Designers of conversational systems should consider incorporating both disclaimers and feedback on query understanding in response to user queries for medical advice. Unconstrained natural language input should not be used in systems designed specifically to provide medical advice.

Boys-Specific Text-Comprehension Enhancement With Dual Visual-Auditory Text Presentation Among 12-14 Years-Old Students

Quality of language comprehension determines performance in all kinds of activities including academics. Processing of words initially develops as auditory, and gradually extends to visual as children learn to read. School failure is highly related to listening and reading comprehension problems. In this study we analyzed sex-differences in comprehension of texts in Spanish (standardized reading test PROLEC-R) in three modalities (visual, auditory, and both simultaneously: dual-modality) presented to 12-14-years old students, native in Spanish. We controlled relevant cognitive variables such as attention (d2), phonological and semantic fluency (FAS) and speed of processing (WISC subtest Coding). Girls' comprehension was similar in the three modalities of presentation, however boys were importantly benefited by dual-modality as compared to boys exposed only to visual or auditory text presentation. With respect to the relation of text comprehension and school performance, students with low grades in Spanish showed low auditory comprehension. Interestingly, visual and dual modalities preserved comprehension levels in these low skilled students. Our results suggest that the use of visual-text support during auditory language presentation could be beneficial for low school performance students, especially boys, and encourage future research to evaluate the implementation in classes of the rapidly developing technology of simultaneous speech transcription, that could be, in addition, beneficial to non-native students, especially those recently incorporated into school or newly arrived in a country from abroad.

Smartphone-Detected Ambient Speech and Self-Reported Measures of Anxiety and Depression: Exploratory Observational Study

Background

The ability to objectively measure the severity of depression and anxiety disorders in a passive manner could have a profound impact on the way in which these disorders are diagnosed, assessed, and treated. Existing studies have demonstrated links between both depression and anxiety and the linguistic properties of words that people use to communicate. Smartphones offer the ability to passively and continuously detect spoken words to monitor and analyze the linguistic properties of speech produced by the speaker and other sources of ambient speech in their environment. The linguistic properties of automatically detected and recognized speech may be used to build objective severity measures of depression and anxiety.

Objective

The aim of this study was to determine if the linguistic properties of words passively detected from environmental audio recorded using a participant’s smartphone can be used to find correlates of symptom severity of social anxiety disorder, generalized anxiety disorder, depression, and general impairment.

Methods

An Android app was designed to collect periodic audiorecordings of participants’ environments and to detect English words using automatic speech recognition. Participants were recruited into a 2-week observational study. The app was installed on the participants’ personal smartphones to record and analyze audio. The participants also completed self-report severity measures of social anxiety disorder, generalized anxiety disorder, depression, and functional impairment. Words detected from audiorecordings were categorized, and correlations were measured between words counts in each category and the 4 self-report measures to determine if any categories could serve as correlates of social anxiety disorder, generalized anxiety disorder, depression, or general impairment.

Results

The participants were 112 adults who resided in Canada from a nonclinical population; 86 participants yielded sufficient data for analysis. Correlations between word counts in 67 word categories and each of the 4 self-report measures revealed a strong relationship between the usage rates of death-related words and depressive symptoms (r=0.41, P<.001). There were also interesting correlations between rates of word usage in the categories of reward-related words with depression (r=–0.22, P=.04) and generalized anxiety (r=–0.29, P=.007), and vision-related words with social anxiety (r=0.31, P=.003).

Conclusions

In this study, words automatically recognized from environmental audio were shown to contain a number of potential associations with severity of depression and anxiety. This work suggests that sparsely sampled audio could provide relevant insight into individuals’ mental health.

Readiness for voice assistants to support healthcare delivery during a health crisis and pandemic

To prevent the spread of COVID-19 and to continue responding to healthcare needs, hospitals are rapidly adopting telehealth and other digital health tools to deliver care remotely. Intelligent conversational agents and virtual assistants, such as chatbots and voice assistants, have been utilized to augment health service capacity to screen symptoms, deliver healthcare information, and reduce exposure. In this commentary, we examined the state of voice assistants (e.g., Google Assistant, Apple Siri, Amazon Alexa) as an emerging tool for remote healthcare delivery service and discussed the readiness of the health system and technology providers to adapt voice assistants as an alternative healthcare delivery modality during a health crisis and pandemic.

Illuminating the dark spaces of healthcare with ambient intelligence

Advances in machine learning and contactless sensors have given rise to ambient intelligence-physical spaces that are sensitive and responsive to the presence of humans. Here we review how this technology could improve our understanding of the metaphorically dark, unobserved spaces of healthcare. In hospital spaces, early applications could soon enable more efficient clinical workflows and improved patient safety in intensive care units and operating rooms. In daily living spaces, ambient intelligence could prolong the independence of older individuals and improve the management of individuals with a chronic disease by understanding everyday behaviour. Similar to other technologies, transformation into clinical applications at scale must overcome challenges such as rigorous clinical validation, appropriate data privacy and model transparency. Thoughtful use of this technology would enable us to understand the complex interplay between the physical environment and health-critical human behaviours.