Multimodal fusion of brain signals for robust prediction of psychosis transition.

The prospective study of youths at clinical high risk (CHR) for psychosis, including neuroimaging, can identify neural signatures predictive of psychosis outcomes using algorithms that integrate complex information. Here, to identify risk and psychosis conversion, we implemented multiple kernel learning (MKL), a multimodal machine learning approach allowing patterns from each modality to inform each other. Baseline multimodal scans (n = 74, 11 converters) included structural, resting-state functional imaging, and diffusion-weighted data. Multimodal MKL outperformed unimodal models (AUC = 0.73 vs. 0.66 in predicting conversion). Moreover, patterns learned by MKL were robust to training set variations, suggesting it can identify cross-modality redundancies and synergies to stabilize the predictive pattern. We identified many predictors consistent with the literature, including frontal cortices, cingulate, thalamus, and striatum. This highlights the advantage of methods that leverage the complex pathophysiology of psychosis.

Fusion of biomedical imaging studies for increased sample size and diversity: a case study of brain MRI.

Data collection, curation, and cleaning constitute a crucial phase in Machine Learning (ML) projects. In biomedical ML, it is often desirable to leverage multiple datasets to increase sample size and diversity, but this poses unique challenges, which arise from heterogeneity in study design, data descriptors, file system organization, and metadata. In this study, we present an approach to the integration of multiple brain MRI datasets with a focus on homogenization of their organization and preprocessing for ML. We use our own fusion example (approximately 84,000 images from 54,000 subjects, 12 studies, and 88 individual scanners) to illustrate and discuss the issues faced by study fusion efforts, and we examine key decisions necessary during dataset homogenization, presenting in detail a database structure flexible enough to accommodate multiple observational MRI datasets. We believe our approach can provide a basis for future similarly-minded biomedical ML projects.

Detecting at-risk mental states for psychosis (ARMS) using machine learning ensembles and facial features.

AIMS: Our study aimed to develop a machine learning ensemble to distinguish "at-risk mental states for psychosis" (ARMS) subjects from control individuals from the general population based on facial data extracted from video-recordings. METHODS: 58 non-help-seeking medication-naïve ARMS and 70 healthy subjects were screened from a general population sample. At-risk status was assessed with the Structured Interview for Prodromal Syndromes (SIPS), and "Subject's Overview" section was filmed (5-10 min). Several features were extracted, e.g., eye and mouth aspect ratio, Euler angles, coordinates from 51 facial landmarks. This elicited 649 facial features, which were further selected using Gradient Boosting Machines (AdaBoost combined with Random Forests). Data was split in 70/30 for training, and Monte Carlo cross validation was used. RESULTS: Final model reached 83 % of mean F1-score, and balanced accuracy of 85 %. Mean area under the curve for the receiver operator curve classifier was 93 %. Convergent validity testing showed that two features included in the model were significantly correlated with Avolition (SIPS N2 item) and expression of emotion (SIPS N3 item). CONCLUSION: Our model capitalized on short video-recordings from individuals recruited from the general population, effectively distinguishing between ARMS and controls. Results are encouraging for large-screening purposes in low-resource settings.

ChatGPT and Bard exhibit spontaneous citation fabrication during psychiatry literature search.

ChatGPT (Generative Pre-Trained Transformer) is a large language model (LLM), which comprises a neural network that has learned information and patterns of language use from large amounts of text on the internet. ChatGPT, introduced by OpenAI, responds to human queries in a conversational manner. Here, we aimed to assess whether ChatGPT could reliably produce accurate references to supplement the literature search process. We describe our March 2023 exchange with ChatGPT, which generated thirty-five citations, two of which were real. 12 citations were similar to actual manuscripts (e.g., titles with incorrect author lists, journals, or publication years) and the remaining 21, while plausible, were in fact a pastiche of multiple existent manuscripts. In June 2023, we re-tested ChatGPT's performance and compared it to that of Google's GPT counterpart, Bard 2.0. We investigated performance in English, as well as in Spanish and Italian. Fabrications made by LLMs, including erroneous citations, have been called "hallucinations"; we discuss reasons for which this is a misnomer. Furthermore, we describe potential explanations for citation fabrication by GPTs, as well as measures being taken to remedy this issue, including reinforcement learning. Our results underscore that output from conversational LLMs should be verified.

Differential Expression of Anomalous Self-Experiences in Spontaneous Speech in Clinical High-Risk and Early-Course Psychosis Quantified by Natural Language Processing.

BACKGROUND: Basic self-disturbance, or anomalous self-experiences (ASEs), is a core feature of the schizophrenia spectrum. We propose a novel method of natural language processing to quantify ASEs in spoken language by direct comparison to an inventory of self-disturbance, the Inventory of Psychotic-Like Anomalous Self-Experiences (IPASE). We hypothesized that there would be increased similarity in open-ended speech to the IPASE items in individuals with early-course psychosis (PSY) compared with healthy individuals, with clinical high-risk (CHR) individuals intermediate in similarity. METHODS: Open-ended interviews were obtained from 170 healthy control participants, 167 CHR participants, and 89 PSY participants. We calculated the semantic similarity between IPASE items and "I" sentences from transcribed speech samples using S-BERT (Sentence Bidirectional Encoder Representation from Text). Kolmogorov-Smirnov tests were used to compare distributions across groups. A nonnegative matrix factorization of cosine similarity was performed to rank IPASE items. RESULTS: Spoken language of CHR individuals had the greatest semantic similarity to IPASE items when compared to both healthy control (s = 0.44, p < 10-14) and PSY (s = 0.36, p < 10-6) individuals, while IPASE scores were higher among PSY than CHR group participants. In addition, the nonnegative matrix factorization approach produced a data-driven domain that differentiated the CHR group from the others. CONCLUSIONS: We found that open-ended interviews elicited language with increased semantic similarity to the IPASE by participants in the CHR group compared with patients with psychosis. This demonstrates the utility of these methods for differentiating patients from healthy control participants. This complementary approach has the capacity to scale to large studies investigating phenomenological features of schizophrenia and potentially other clinical populations.

Expansive linguistic representations to predict interpretable odor mixture discriminability.

Language is often thought as being poorly adapted to precisely describe or quantify smell and olfactory attributes. In this work, we show that semantic descriptors of odors can be implemented in a model to successfully predict odor mixture discriminability, an olfactory attribute. We achieved this by taking advantage of the structure-to-percept model we previously developed for monomolecular odorants, using chemical descriptors to predict pleasantness, intensity and 19 semantic descriptors such as "fish," "cold," "burnt," "garlic," "grass," and "sweet" for odor mixtures, followed by a metric learning to obtain odor mixture discriminability. Through this expansion of the representation of olfactory mixtures, our Semantic model outperforms state of the art methods by taking advantage of the intermediary semantic representations learned from human perception data to enhance and generalize the odor discriminability/similarity predictions. As 10 of the semantic descriptors were selected to predict discriminability/similarity, our approach meets the need of rapidly obtaining interpretable attributes of odor mixtures as illustrated by the difficulty of finding olfactory metamers. More fundamentally, it also shows that language can be used to establish a metric of discriminability in the everyday olfactory space.

Natural Language Processing Markers for Psychosis and Other Psychiatric Disorders: Emerging Themes and Research Agenda From a Cross-Linguistic Workshop.

This workshop summary on natural language processing (NLP) markers for psychosis and other psychiatric disorders presents some of the clinical and research issues that NLP markers might address and some of the activities needed to move in that direction. We propose that the optimal development of NLP markers would occur in the context of research efforts to map out the underlying mechanisms of psychosis and other disorders. In this workshop, we identified some of the challenges to be addressed in developing and implementing NLP markers-based Clinical Decision Support Systems (CDSSs) in psychiatric practice, especially with respect to psychosis. Of note, a CDSS is meant to enhance decision-making by clinicians by providing additional relevant information primarily through software (although CDSSs are not without risks). In psychiatry, a field that relies on subjective clinical ratings that condense rich temporal behavioral information, the inclusion of computational quantitative NLP markers can plausibly lead to operationalized decision models in place of idiosyncratic ones, although ethical issues must always be paramount.

Linguistic correlates of suicidal ideation in youth at clinical high-risk for psychosis.

Suicidal ideation (SI) is prevalent among individuals at clinical high-risk for psychosis (CHR). Natural language processing (NLP) provides an efficient method to identify linguistic markers of suicidality. Prior work has demonstrated that an increased use of "I", as well as words with semantic similarity to "anger", "sadness", "stress" and "lonely", are correlated with SI in other cohorts. The current project analyzes data collected in an SI supplement to an NIH R01 study of thought disorder and social cognition in CHR. This study is the first to use NLP analyses of spoken language to identify linguistic correlates of recent suicidal ideation among CHR individuals. The sample included 43 CHR individuals, 10 with recent suicidal ideation and 33 without, as measured by the Columbia-Suicide Severity Rating Scale, as well as 14 healthy volunteers without SI. NLP methods include part-of-speech (POS) tagging, a GoEmotions-trained BERT Model, and Zero-Shot Learning. As hypothesized, individuals at CHR for psychosis who endorsed recent SI utilized more words with semantic similarity to "anger" compared to those who did not. Words with semantic similarity to "stress", "loneliness", and "sadness" were not significantly different between the two CHR groups. Contrary to our hypotheses, CHR individuals with recent SI did not use the word "I" more than those without recent SI. As anger is not characteristic of CHR, findings have implications for the consideration of subthreshold anger-related sentiment in suicidal risk assessment. As NLP is scalable, findings suggest that language markers may improve suicide screening and prediction in this population.

Emergence of Language Related to Self-experience and Agency in Autobiographical Narratives of Individuals With Schizophrenia.

BACKGROUND AND HYPOTHESIS: Disturbances in self-experience are a central feature of schizophrenia and its study can enhance phenomenological understanding and inform mechanisms underlying clinical symptoms. Self-experience involves the sense of self-presence, of being the subject of one's own experiences and agent of one's own actions, and of being distinct from others. Self-experience is traditionally assessed by manual rating of interviews; however, natural language processing (NLP) offers automated approach that can augment manual ratings by rapid and reliable analysis of text. STUDY DESIGN: We elicited autobiographical narratives from 167 patients with schizophrenia or schizoaffective disorder (SZ) and 90 healthy controls (HC), amounting to 490 000 words and 26 000 sentences. We used NLP techniques to examine transcripts for language related to self-experience, machine learning to validate group differences in language, and canonical correlation analysis to examine the relationship between language and symptoms. STUDY RESULTS: Topics related to self-experience and agency emerged as significantly more expressed in SZ than HC (P < 10-13) and were decoupled from similarly emerging features such as emotional tone, semantic coherence, and concepts related to burden. Further validation on hold-out data showed that a classifier trained on these features achieved patient-control discrimination with AUC = 0.80 (P < 10-5). Canonical correlation analysis revealed significant relationships between self-experience and agency language features and clinical symptoms. CONCLUSIONS: Notably, the self-experience and agency topics emerged without any explicit probing by the interviewer and can be algorithmically detected even though they involve higher-order metacognitive processes. These findings illustrate the utility of NLP methods to examine phenomenological aspects of schizophrenia.

Ethical Implications of the Use of Language Analysis Technologies for the Diagnosis and Prediction of Psychiatric Disorders.

Recent developments in artificial intelligence technologies have come to a point where machine learning algorithms can infer mental status based on someone's photos and texts posted on social media. More than that, these algorithms are able to predict, with a reasonable degree of accuracy, future mental illness. They potentially represent an important advance in mental health care for preventive and early diagnosis initiatives, and for aiding professionals in the follow-up and prognosis of their patients. However, important issues call for major caution in the use of such technologies, namely, privacy and the stigma related to mental disorders. In this paper, we discuss the bioethical implications of using such technologies to diagnose and predict future mental illness, given the current scenario of swiftly growing technologies that analyze human language and the online availability of personal information given by social media. We also suggest future directions to be taken to minimize the misuse of such important technologies.

Predictive Modeling of Huntington's Disease Unfolds Thalamic and Caudate Atrophy Dissociation.

BACKGROUND: Atrophy in the striatum is a hallmark of Huntington's disease (HD), including the period before clinical motor diagnosis (before-CMD), but it extends to other subcortical structures. The study of the covariation of these structures could improve the detection of disease-related longitudinal progression before-CMD, provide mechanistic insights of the disease, and potentially be used to obtain accurate prospective estimates of atrophy before-CMD and early after-CMD. METHODS: We analyzed data from 337 before-CMD individuals, 236 healthy control subjects, and 95 early after-CMD individuals from three studies, and we used nine subcortical regions volumes in two analyses. First, we discriminated before-CMD from healthy control trajectories by integrating volume changes from these regions. Second, we estimated prospective atrophy before-CMD and early after-CMD by considering the influence of a region's present volume over the future volume of another one. RESULTS: Before-CMD progression was robustly detected across studies. Indeed, detection of before-CMD progression improved when multiple structures were integrated, as opposed to analyzing the striatum alone, likely because of the reduced partial correlation between caudate and thalamic volume change before-CMD. Our multivariate atrophy prediction model found a thalamus-caudate association that is consistent with this pattern, which yields an improved caudate atrophy prediction in early after-CMD. CONCLUSIONS: This study is the first attempt to validate before-CMD multivariate subcortical change detection across studies and to do multivariate prospective atrophy prediction in HD. These models achieve improved performance by detecting a dissociation between caudate and thalamic atrophy trajectories, and they provide a possible mechanistic understanding of the dynamics of HD. © 2022 International Parkinson and Movement Disorder Society.

Negative symptoms and speech pauses in youths at clinical high risk for psychosis.

Aberrant pauses are characteristic of schizophrenia and are robustly associated with its negative symptoms. Here, we found that pause behavior was associated with negative symptoms in individuals at clinical high risk (CHR) for psychosis, and with measures of syntactic complexity-phrase length and usage of determiners that introduce clauses-that we previously showed in this same CHR cohort to help comprise a classifier that predicted psychosis. These findings suggest a common impairment in discourse planning and verbal self-monitoring that affects both speech and language, and which is detected in clinical ratings of negative symptoms.

Identifying signals associated with psychiatric illness utilizing language and images posted to Facebook.

Prior research has identified associations between social media activity and psychiatric diagnoses; however, diagnoses are rarely clinically confirmed. Toward the goal of applying novel approaches to improve outcomes, research using real patient data is necessary. We collected 3,404,959 Facebook messages and 142,390 images across 223 participants (mean age = 23.7; 41.7% male) with schizophrenia spectrum disorders (SSD), mood disorders (MD), and healthy volunteers (HV). We analyzed features uploaded up to 18 months before the first hospitalization using machine learning and built classifiers that distinguished SSD and MD from HV, and SSD from MD. Classification achieved AUC of 0.77 (HV vs. MD), 0.76 (HV vs. SSD), and 0.72 (SSD vs. MD). SSD used more (P < 0.01) perception words (hear, see, feel) than MD or HV. SSD and MD used more (P < 0.01) swear words compared to HV. SSD were more likely to express negative emotions compared to HV (P < 0.01). MD used more words related to biological processes (blood/pain) compared to HV (P < 0.01). The height and width of photos posted by SSD and MD were smaller (P < 0.01) than HV. MD photos contained more blues and less yellows (P < 0.01). Closer to hospitalization, use of punctuation increased (SSD vs HV), use of negative emotion words increased (MD vs. HV), and use of swear words increased (P < 0.01) for SSD and MD compared to HV. Machine-learning algorithms are capable of differentiating SSD and MD using Facebook activity alone over a year in advance of hospitalization. Integrating Facebook data with clinical information could one day serve to inform clinical decision-making.

Using Language Processing and Speech Analysis for the Identification of Psychosis and Other Disorders.

Increasingly, data-driven methods have been implemented to understand psychopathology. Language is the main source of information in psychiatry and represents "big data" at the level of the individual. Language and behavior are amenable to computational natural language processing (NLP) analytics, which may help operationalize the mental status examination. In this review, we highlight the application of NLP to schizophrenia and its risk states as an exemplar of its use, operationalizing tangential and concrete speech as reductions in semantic coherence and syntactic complexity, respectively. Other clinical applications are reviewed, including forecasting suicide risk and detecting intoxication. Challenges and future directions are discussed, including biomarker development, harmonization, and application of NLP more broadly to behavior, including intonation/prosody, facial expression and gesture, and the integration of these in dyads and during discourse. Similar NLP analytics can also be applied beyond humans to behavioral motifs across species, important for modeling psychopathology in animal models. Finally, clinical neuroscience can inform the development of artificial intelligence.

Language as a biomarker for psychosis: A natural language processing approach.

Human ratings of conceptual disorganization, poverty of content, referential cohesion and illogical thinking have been shown to predict psychosis onset in prospective clinical high risk (CHR) cohort studies. The potential value of linguistic biomarkers has been significantly magnified, however, by recent advances in natural language processing (NLP) and machine learning (ML). Such methodologies allow for the rapid and objective measurement of language features, many of which are not easily recognized by human raters. Here we review the key findings on language production disturbance in psychosis. We also describe recent advances in the computational methods used to analyze language data, including methods for the automatic measurement of discourse coherence, syntactic complexity, poverty of content, referential coherence, and metaphorical language. Linguistic biomarkers of psychosis risk are now undergoing cross-validation, with attention to harmonization of methods. Future directions in extended CHR networks include studies of sources of variance, and combination with other promising biomarkers of psychosis risk, such as cognitive and sensory processing impairments likely to be related to language. Implications for the broader study of social communication, including reciprocal prosody, face expression and gesture, are discussed.

Loss of nucleus accumbens low-frequency fluctuations is a signature of chronic pain.

Chronic pain is a highly prevalent disease with poorly understood pathophysiology. In particular, the brain mechanisms mediating the transition from acute to chronic pain remain largely unknown. Here, we identify a subcortical signature of back pain. Specifically, subacute back pain patients who are at risk for developing chronic pain exhibit a smaller nucleus accumbens volume, which persists in the chronic phase, compared to healthy controls. The smaller accumbens volume was also observed in a separate cohort of chronic low-back pain patients and was associated with dynamic changes in functional connectivity. At baseline, subacute back pain patients showed altered local nucleus accumbens connectivity between putative shell and core, irrespective of the risk of transition to chronic pain. At follow-up, connectivity changes were observed between nucleus accumbens and rostral anterior cingulate cortex in the patients with persistent pain. Analysis of the power spectral density of nucleus accumbens resting-state activity in the subacute and chronic back pain patients revealed loss of power in the slow-5 frequency band (0.01 to 0.027 Hz) which developed only in the chronic phase of pain. This loss of power was reproducible across two cohorts of chronic low-back pain patients obtained from different sites and accurately classified chronic low-back pain patients in two additional independent datasets. Our results provide evidence that lower nucleus accumbens volume confers risk for developing chronic pain and altered nucleus accumbens activity is a signature of the state of chronic pain.

Resting-state connectivity stratifies premanifest Huntington's disease by longitudinal cognitive decline rate.

Patient stratification is critical for the sensitivity of clinical trials at early stages of neurodegenerative disorders. In Huntington's disease (HD), genetic tests make cognitive, motor and brain imaging measurements possible before symptom manifestation (pre-HD). We evaluated pre-HD stratification models based on single visit resting-state functional MRI (rs-fMRI) data that assess observed longitudinal motor and cognitive change rates from the multisite Track-On HD cohort (74 pre-HD, 79 control participants). We computed longitudinal performance change on 10 tasks (including visits from the preceding TRACK-HD study when available), as well as functional connectivity density (FCD) maps in single rs-fMRI visits, which showed high test-retest reliability. We assigned pre-HD subjects to subgroups of fast, intermediate, and slow change along single tasks or combinations of them, correcting for expectations based on aging; and trained FCD-based classifiers to distinguish fast- from slow-progressing individuals. For robustness, models were validated across imaging sites. Stratification models distinguished fast- from slow-changing participants and provided continuous assessments of decline applicable to the whole pre-HD population, relying on previously-neglected white matter functional signals. These results suggest novel correlates of early deterioration and a robust stratification strategy where a single MRI measurement provides an estimate of multiple ongoing longitudinal changes.

Detection of acute 3,4-methylenedioxymethamphetamine (MDMA) effects across protocols using automated natural language processing.

The detection of changes in mental states such as those caused by psychoactive drugs relies on clinical assessments that are inherently subjective. Automated speech analysis may represent a novel method to detect objective markers, which could help improve the characterization of these mental states. In this study, we employed computer-extracted speech features from multiple domains (acoustic, semantic, and psycholinguistic) to assess mental states after controlled administration of 3,4-methylenedioxymethamphetamine (MDMA) and intranasal oxytocin. The training/validation set comprised within-participants data from 31 healthy adults who, over four sessions, were administered MDMA (0.75, 1.5 mg/kg), oxytocin (20 IU), and placebo in randomized, double-blind fashion. Participants completed two 5-min speech tasks during peak drug effects. Analyses included group-level comparisons of drug conditions and estimation of classification at the individual level within this dataset and on two independent datasets. Promising classification results were obtained to detect drug conditions, achieving cross-validated accuracies of up to 87% in training/validation and 92% in the independent datasets, suggesting that the detected patterns of speech variability are associated with drug consumption. Specifically, we found that oxytocin seems to be mostly driven by changes in emotion and prosody, which are mainly captured by acoustic features. In contrast, mental states driven by MDMA consumption appear to manifest in multiple domains of speech. Furthermore, we find that the experimental task has an effect on the speech response within these mental states, which can be attributed to presence or absence of an interaction with another individual. These results represent a proof-of-concept application of the potential of speech to provide an objective measurement of mental states elicited during intoxication.