Key Points
Question
Are naturalistic speech features associated with concurrent variation in psychotic symptom severity?
Findings
In this cohort study of 356 patients with schizophrenia spectrum disorders in the Netherlands, with replication in 72 US patients, artificial intelligence–derived voice and text features from 938 speech recordings were associated with positive and negative symptom severity at both the subscale and item level. Symptoms such as hallucinations, suspiciousness, and blunted affect were associated with a less than 1-point error, within the range considered acceptable for human raters.
Meaning
These findings suggest that psychotic symptoms leave specific, quantifiable, and interpretable signatures in speech, supporting scalable, low-burden tools for real-time monitoring in psychosis.
This cohort study examines which features of naturalistic speech are associated with concurrent variation in symptom severity in adult patients with schizophrenia spectrum disorders in Dutch and US samples.
Abstract
Importance
Speech carries subtle indicators of current mental state, a phenomenon routinely used in psychiatric assessment. Despite its potential, quantitative speech analysis is not yet integrated into clinical care.
Objective
To identify which features of naturalistic speech are associated with concurrent variation in symptom severity in patients with psychotic disorders.
Design, Setting, and Participants
This longitudinal, multicenter cohort study was conducted in the Netherlands (June 7, 2017, to July 31, 2025) in adult patients (aged ≥18 years) with schizophrenia spectrum disorders who completed repeated clinical and speech assessments for up to 8 years. Findings were replicated in a longitudinal cohort from the US (March 1, 2021, to December 1, 2022).
Main Outcomes and Measures
Individual symptom severity was measured using the Positive and Negative Syndrome Scale (PANSS) (Dutch cohort) or the Brief Psychiatric Rating Scale (US cohort). Speech samples were converted into interpretable artificial intelligence–derived voice and text features, reduced using principal component analysis. Associations were estimated using linear mixed models adjusted for demographic characteristics and time point. Estimative accuracy was quantified using mean absolute error (MAE) and R2.
Results
In the Dutch cohort, 773 speech recordings from 356 participants (mean [SD] age, 30.4 [10.3] years; 257 male [72.2%]) were analyzed, and in the US cohort, 165 speech recordings from 72 participants (mean [SD] age, 26.4 [5.2] years; 46 male [64.8%]) were analyzed. In the Dutch cohort, speech-based models detected individual psychotic symptoms with clinically meaningful accuracy, with an item-level MAE of less than 1 point (scale, 1-7), comparable to the 1-point agreement margin used in PANSS rater training. Models were associated with PANSS positive and negative subscale scores, with MAE of 2.85 and 3.22, explaining 13.4% and 17.8% of the variance, respectively, and well below the 20% threshold used to flag unreliable PANSS ratings. In the US replication cohort, Brief Psychiatric Rating Scale thought disturbance and withdrawal scores were associated with similar MAEs (3.0 and 2.4, respectively), explaining 31.0% and 21.5% of the variance. In the Dutch cohort, negative symptoms were associated with more reduced speech output (estimate, 2.46 [95% CI, 1.05-3.88]) and flatter acoustic profiles (estimate, 0.29 [95% CI, 0.10-0.48]), whereas positive symptoms were marked by longer utterances (estimate, 1.18 [95% CI, 0.08-2.29]) and altered discourse organization (estimate, −0.25 [95% CI, −0.48 to −0.02]).
Conclusions and Relevance
This cohort study of individuals with schizophrenia found that psychotic symptoms left specific, interpretable signatures in naturalistic speech that could be quantified, tracked longitudinally, and replicated across cohorts. Speech-based modeling achieved clinically meaningful, symptom-level estimations, providing a solid basis for scalable, low-burden tools for real-time monitoring in psychosis.
Introduction
You call a patient you have been treating for years. She answers with a simple, “Hello.” From that single word alone, you instantly know something is wrong. Her tone is flat, her rhythm slightly off, her volume low. Before she speaks another word, your brain has processed dozens of subtle acoustic cues signaling that something in her mental state has changed. This kind of intuitive detection is something clinicians do every day. We read moods, intentions, and psychological states not only from what people say but also from how they say it. In psychiatry, this intuition is essential for clinical practice—speech is the primary medium through which patients express their mental states and through which clinicians assess and monitor symptoms.
Despite its centrality, changes in speech have been difficult to quantify. In psychosis, speech alterations, such as disorganization, flat affect, and reduced verbal output, often precede relapse,1 but tracking these changes requires contact far more frequent than routine care allows. Recent advances in automated speech recognition and natural language processing now enable scalable extraction of linguistic features,2,3,4 supporting high-frequency, low-burden monitoring to detect early signs of relapse or recovery.5
Prior work has suggested that specific linguistic markers, including reduced syntactic complexity, lower speech quantity, and altered semantic similarity, track symptom fluctuations in psychosis,6,7,8 yet few studies have assessed these fluctuations at the level of individual symptoms. In this longitudinal study, we tested whether short, naturalistic speech samples are associated with concurrent variation in psychotic symptoms using the Positive and Negative Syndrome Scale (PANSS) in Dutch and the Brief Psychiatric Rating Scale (BPRS) in English. We evaluated this approach in a large, multicenter cohort in the Netherlands and examined reproducibility in an independent longitudinal US cohort, which allowed us to examine the following fundamental questions for clinical use: How accurately can speech-based models detect symptom severity? Can they move beyond overall scale scores to identify specific symptoms, such as hallucinations or blunted affect? Do they show face validity by aligning with what clinicians observe when rating symptoms? Do they capture psychopathology above and beyond background characteristics, such as age, sex, and education? Answering these questions is essential for the safe and meaningful use of speech technology in psychiatry. By treating speech not just as expression but also as a dynamic signal of mental state, we may advance efforts toward scalable, precise, and real-time monitoring of psychotic symptoms.
Methods
This cohort study included a Dutch cohort recruited from the University Medical Centers in Utrecht and Groningen, the Netherlands (HAMLETT [June 7, 2017, to July 31, 2025], RAPSODI [August 1, 2016, to July 1, 2021]), and a US cohort (LPOP study) recruited from acute inpatient units at Zucker Hillside Hospital in New York (March 1, 2021, to December 1, 2022). All studies were approved by the Ethics Committees of University Medical Center Groningen or Utrecht for the Dutch cohort and the Northwell Health Institutional Review Board for the US cohort. All participants provided written informed consent. The study followed the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) reporting guideline for cohort studies.
Eligible participants were aged 18 years or older, fluent in Dutch or English, and had diagnoses of schizophrenia spectrum or bipolar I disorder confirmed by structured diagnostic interviews according to Diagnostic and Statistical Manual of Mental Disorders (Fourth Edition) and Diagnostic and Statistical Manual of Mental Disorders (Fifth Edition) criteria. Study-specific inclusion criteria are shown in eAppendix 1 in Supplement 1. In the US cohort, race (Asian, Black or African American, White, other, or not reported) was self-identified by participants. In the Dutch cohort, migration background was determined based on participant-reported country of birth and parental country of birth. Race and migration background were included as covariates because they may be associated with variation in speech-derived features.
Spontaneous speech samples were collected using open-ended questions at baseline and follow-up. Dutch interviews were conducted by trained researchers; US samples were collected via a touch-screen tablet using the Winterlight application (Winterlight Labs) (eAppendix 2 in Supplement 1).7,9 Longitudinal follow-up frequency varied by parent study protocol.
Primary Outcome
Symptom severity was assessed using PANSS (Dutch cohort) or BPRS (US cohort). Assessments focused on positive, negative, thought disturbance, and withdrawal subscales.10
Feature Extraction
Speech recordings were processed using automated pipelines to extract acoustic, syntactic, semantic, and sentiment features (eAppendixes 3 [Dutch cohort] and 4 [US cohort] in Supplement 1).8,11,12,13,14 A total of 160 features were extracted for the Dutch cohort, and 138 features for the US cohort.
Statistical Analysis
The skewness of voice and text features was assessed, with absolute coefficients greater than 2 considered indicative of substantial nonnormality. Features exceeding this threshold were winsorized at the 10th and 90th percentiles. Since acoustic features are often affected by background noise, recording artifacts, and speaker variability resulting in large outliers, lower levels of winsorization (1.0-99.0, 2.5-97.5, 5.0-95.0) were insufficient to normalize the data.15 Sensitivity analyses using winsorization at the 5th and 95th percentiles are reported in eAppendix 5 in Supplement 1.
To reduce the dimensionality of the feature space while preserving interpretability, separate principal component analyses were conducted for each feature type: acoustic, semantic, sentiment, and syntactic. Principal component analyses reduced features to interpretable speech-language components (eAppendix 6 in Supplement 1), which were tested against symptom scores using linear mixed-effects models adjusted for age, sex, education, racial or migration background, and time point (eAppendix 6 in Supplement 1). Associations with symptom severity were evaluated at both the subscale (PANSS positive and negative; BPRS thought disturbance and withdrawal) and item level. The BPRS thought disturbance and withdrawal factors approximate the PANSS positive and negative subscales, respectively. The BPRS thought disturbance factors include conceptual disorganization, hallucinatory behavior, and unusual thought content; conceptual disorganization and hallucinatory behavior correspond directly to PANSS positive items P2 and P3. The BPRS withdrawal factors include emotional withdrawal, blunted affect, and motor retardation; blunted affect and emotional withdrawal correspond directly to PANSS negative items N1 and N2. Thus, BPRS thought disturbance may align more closely with thought disorder within the broader positive symptom domain, whereas BPRS withdrawal may additionally reflect psychomotor slowing in addition to negative symptoms.
For analyses in the Dutch cohort, multiple comparisons were controlled using the Benjamini-Hochberg false discovery rate applied across 22 variables for the subscales and 11 variables for the item-level analyses. Statistical significance was defined as a false discovery rate–adjusted 2-sided P < .05. In the US replication cohort, analyses were restricted to prespecified associations from the primary Dutch cohort and were considered confirmatory; therefore, no correction for multiple comparisons was applied. Statistical significance was defined as a 2-sided P < .05.
Out-of-sample performance was assessed using 5-fold participant-level cross-validation reporting of the mean absolute error (MAE) and marginal R2. This framework enabled us to (1) evaluate whether the speech longitudinal model estimates PANSS total subscales severity; (2) identify which speech-language components contribute most strongly to that estimation; and (3) evaluate which speech-language components were associated with which specific symptoms. To examine potential differences in MAE across demographic groups, linear models were fitted with participant-level out-of-fold MAE as the dependent variable and demographic characteristics as estimators. Analyses were conducted using R, version 4.4.3 (R Foundation for Statistical Computing).
Results
In the Dutch cohort, a total of 773 audio recordings from 356 participants (mean [SD] age, 30.4 [10.3] years; 99 female [27.8%] and 257 male [72.2%]; 234 nonmigrants [65.7%] and 10 first-generation [2.8%] and 108 second-generation [30.3%] migrants) were included (Table). Figure 1 displays mean PANSS total positive and total negative scores across the 6 time points for the Dutch cohort. The US cohort included 165 audio recordings from 72 participants (mean [SD] age, 26.4 [5.2] years; 20 female [28.3%], 46 male [64.8%], 1 intersex [1.5%], 4 not reporting sex [5.6%]; 12 identifying as Asian [16.9%], 29 as Black or African American [40.8%], 16 as White [22.5%], 9 as other [12.7%], and 4 not reporting [7.0%] race).
Table. Baseline Demographic and Clinical Characteristics.
| Characteristic | Cohort participants, No. (%) | |
|---|---|---|
| Dutch (n = 356) | US replication (n = 72)a | |
| No. of recordings across time points | 773 | 165 |
| Age, mean (SD), y | 30.4 (10.3) | 26.4 (5.2) |
| No. of visits, median (IQR) | 2 (1-3) | NR |
| Interval between consecutive visits, median (IQR), d | 108 (91-360) | NR |
| Sex | ||
| Female | 99 (27.8) | 20 (28.3) |
| Male | 257 (72.2) | 46 (64.8) |
| Intersex | 0 | 1 (1.5) |
| Not reported | 0 | 4 (5.6) |
| Education, mean (SD), y | 13.5 (2.97) | 14.1 (1.9) |
| Migration status | ||
| Nonmigrant | 234 (65.7) | NA |
| Migrant | ||
| First generation | 10 (2.8) | NA |
| Second generation | 108 (30.3) | NA |
| Race | ||
| Asian | NA | 12 (16.9) |
| Black or African American | NA | 29 (40.8) |
| White | NA | 16 (22.5) |
| Otherb | NA | 9 (12.7) |
| Not reported | NA | 4 (7) |
| DSM-IV or DSM-5 diagnosis | ||
| Bipolar with psychotic symptoms | 0 | 4 (6.0) |
| Psychosis not otherwise specified | 224 (62.9) | 10 (14.9) |
| Schizoaffective disorder | 27 (7.6) | 15 (22.4) |
| Schizophrenia | 79 (22.2) | 34 (50.7) |
| Schizophreniform disorder | 26 (7.3) | 4 (6.0) |
| Psychotic symptom severity, mean (SD)c | ||
| PANSS positive total (range, 7-49) | 12.2 (4.61) | NA |
| PANSS negative total (range, 7-49) | 10.3 (4.04) | NA |
| BPRS thought disturbance factor (range, 7-21) | NA | 12.1 (3.5) |
| BPRS withdrawal factor (range, 7-21) | NA | 5.6 (3.0) |
Abbreviations: BPRS, Brief Psychiatric Rating Scale; DSM-5, Diagnostic and Statistical Manual of Mental Disorders (Fifth Edition); DSM-IV, Diagnostic and Statistical Manual of Mental Disorders (Fourth Edition); NA, not applicable; NR, not reported; PANSS, Positive and Negative Syndrome Scale.
Missing data in the US cohort: n = 1 for sex, n = 2 for race, and n = 5 for DSM-IV or DSM-5 diagnosis.
Other racial and ethnic categories were not reported.
Dutch cohort: PANSS; US cohort: BPRS. Higher PANSS and BPRS scores indicate greater symptom severity.
Figure 1. Line Graph of Positive and Negative Syndrome Scale (PANSS) Total Scores Across Time Points (Dutch Cohort).
Error bars indicate SEs.
Dutch Cohort
Eleven speech-language components were extracted (eAppendix 6 in Supplement 1). The speech-based models detected individual psychotic symptoms with clinically meaningful accuracy, with an item-level MAE of less than 1 point (scale, 1-7), comparable to the 1-point agreement margin used in PANSS rater training. Models estimated PANSS positive and negative symptoms, with cross-validated MAEs of 2.85 and 3.22, explaining 13.4% and 17.8% of the variance, well below the 20% threshold used to flag unreliable PANSS ratings (Figure 2A and B).
Figure 2. Bar Graph of Cross-Validation Performance for Estimation of Positive and Negative Symptom Severity (Dutch Cohort).
A, The horizontal line indicates mean MAE (mean absolute error) at 2.85. B, The horizontal line indicates the mean MAE at 3.23.
Positive symptoms were associated with more negative expressed sentiment (estimate, −1.18 [95% CI, −2.12 to −0.23]; P = .002) and lower maximum semantic similarity (estimate, −0.25 [95% CI, −0.48 to −0.02]; P = .03) and at uncorrected thresholds, with longer sentences (estimate, 1.18 [95% CI, 0.08-2.29]; P = .03) and less reduced verbal output (estimate, −1.33 [95% CI, −2.48 to −0.19]; P = .02) (Figure 3A). Negative symptoms were robustly associated with shorter sentences (estimate, −3.11 [95% CI, −4.46 to −1.75]; P < .001) and more reduced verbal output (estimate, 2.46 [95% CI, 1.05-3.88]; P < .001) (Figure 3B). Formants reached uncorrected significance (estimate, 1.58 [95% CI, 0.21-2.94]; P = .02). Positive and negative symptoms (estimates, −1.02 [95% CI, −1.93 to −0.05; P = .04] and −1.33 [95% CI, −2.38 to −0.27; P = .01], respectively) were higher in men and associated with lower education (estimates, −0.81 [95% CI, −1.17 to −0.46] and −0.72 [95% CI, −1.09 to −0.35], respectively) (P < .001); positive symptoms also increased with age (estimate, 0.97 [95% CI, 0.61-1.33]; P < .001) (eAppendix 7 in Supplement 1). A linear model examining whether demographic variables were associated with out-of-fold MAE for the positive symptom model showed that age (estimate, 0.52 [95% CI, 0.29-0.74]; P < .001), education (estimate, −0.32 [95% CI, −0.54 to −0.11]; P = .003), and sex (estimate, −0.63 [95% CI, −1.15 to −0.13]; P = .02) were significantly associated with estimation error, whereas migration background was not. Demographic variables jointly explained a modest proportion of variance in MAE (R2 = 0.08), suggesting limited, but detectable variation in model performance across subgroups. For the negative symptom model, higher education was associated with lower out-of-fold MAE (estimate, −0.32 [95% CI, −0.55 to −0.10]; P = .005). Age, sex, and migration background were not significantly associated with MAE. Demographic variables explained 3.2% of the variance in estimation error.
Figure 3. Forest Plot of Language Components Estimating Total Symptom Severity (Dutch Cohort).
Standardized fixed-effects estimates from linear mixed-effects models estimating Positive and Negative Syndrome Scale (PANSS) symptom severity from individual language components, adjusting for demographic variables.
aUncorrected P < .05.
bFalse discovery rate–corrected P < .05.
Item-level cross-validated MAEs ranged from 0.21 to 0.94 (marginal R2, 4%-18%). Significant item-level estimations are shown in Figure S9 of eAppendix 8 in Supplement 1; full model performance metrics are provided in eAppendix 8 in Supplement 1.
Hallucinatory behavior was associated with reduced loudness (estimate, −0.24 [95% CI, −0.37 to −0.11]; P = .004), lower pitch variability (estimate, −0.54 [95% CI, −0.91 to −0.18]; P = .01), and increased voice quality (estimate, 0.29 [95% CI, 0.10-0.48]; P = .01), describing a profile of softer, more monotone, whisper-like speech (Figure 4). A similar pattern was observed for suspiciousness or persecution. Blunted affect was associated with lower loudness (estimate, −0.21 [95% CI, −0.38 to −0.05]; P = .04) and shorter sentences (estimate, −0.57 [95% CI, −0.96 to −0.18]; P < .04). Emotional withdrawal (estimate, −0.51 [95% CI, −0.84 to −0.19]; P = .02), lack of spontaneity (estimate, −0.84 [95% CI −1.19 to −0.49]; P = .03), and difficulty in abstract thinking (estimate, −0.45 [95% CI, −0.76 to −0.14]; P = .04) were all associated with shorter sentence length. Lack of spontaneity and flow of conversation were associated with reduced semantic variability (estimate, −0.15 [95% CI, −0.25 to −0.04]; P = .03), shorter sentence length (estimate, −0.84 [95% CI, −1.19 to −0.49]; P < .001), and more reduced verbal output (estimate, 0.69 [95% CI, 0.32-1.05]; P = .001).
Figure 4. Forest Plot of Language Components Estimating Individual Symptom Severity (Dutch Cohort).
Standardized fixed-effects estimates from linear mixed-effects models estimating individual Positive and Negative Syndrome Scale symptoms from speech-language components, adjusting for demographic variables.
aFalse discovery rate–corrected P < .05.
bUncorrected P < .05.
US Replication Cohort
A distinct set of 8 speech-language components was derived for the US cohort (eAppendix 6 in Supplement 1). Models estimated BPRS thought disturbance and withdrawal factors with cross-validated MAEs of 3.00 and 2.44, explaining 31% and 21.5% of variance.
Thought disturbance was associated with higher sentence similarity (estimate, 1.10 [95% CI, 0.37-1.84]; P = .003) and longer sentences (estimate, 0.86 [95% CI, 0.16-1.56]; P = .02). Withdrawal was associated with reduced loudness (estimate, −1.30 [95% CI, −1.95 to −0.66]; P = .001), lower formants or timing scores (estimate, −0.88 [95% CI, −1.61 to −0.15]; P = .02), and lower sentence similarity (estimate, −0.65 [95% CI, −1.28 to −0.02]; P = .04). Demographic associations with thought disturbance and withdrawal severity are reported eAppendix 7 in Supplement 1. In the US sample, none of the demographic variables estimated MAE, suggesting no clear evidence of substantial performance differences across demographic groups in this sample.
Item-level MAEs ranged from 0.83 to 1.76 (marginal R2, 17.0%-31.5%). Full performance metrics and associations between speech-language components and individual symptoms are reported in eAppendix 8 in Supplement 1.
Discussion
As clinicians routinely infer mental state from speech, this cohort study examined that intuition by using interpretable artificial intelligence–derived features. From brief naturalistic speech samples, we estimated concurrent severity of individual psychotic symptoms with clinically meaningful accuracy. Models estimated PANSS positive and negative subscales with an approximate 3-point error in the Dutch cohort (range, 7-49 points), with similar performance for BPRS thought disturbance and withdrawal subscales in the US cohort. Given that deviations of more than 20% from expert ratings are considered unreliable in rater training studies, model errors of automated speech and language analyses fall well within ranges acceptable for human raters.
Several individual items, including hallucinations, suspiciousness, blunted affect, and poor rapport, were estimated with a less than 1-point error, within the 1-point margin used to qualify PANSS raters (range, 1-7). US item-level errors (range, 0.8-1.8) were slightly higher for positive symptoms but remained modest, which may reflect greater positive symptom severity in the US cohort. US cohort participants were recruited from acute inpatient units and were required to have at least moderate positive or disorganized symptoms at inclusion, while one of the Dutch parent studies specifically required participants to be in remission of psychotic symptoms for 3 to 6 months. Across cohorts, negative symptoms were marked by reduced verbal output, limited articulatory variability, and lower loudness, consistent with alogia and flattened prosody.16 These associations were replicated in the US sample and extend prior findings.8,14 Positive symptoms were associated with longer utterances and altered semantic structure, aligning with thought disorder.17 Associations remained after demographic adjustment, indicating that speech features added estimative value beyond stable demographic characteristics and may include a state-related component. However, the models did not separate within-person from between-person effects, making it important for future research to clarify to what extent these associations reflect within-person symptom fluctuations vs stable between-person differences. Moreover, while our analyses examined concurrent associations, prospective estimation of relapse or symptom trajectories warrants further investigation. Overall, rather than a unitary language deficit, these findings support distinct, symptom-specific linguistic signatures in patients with psychotic disorders that mirror clinical evaluation and validate automated speech analyses as a dynamic marker of psychopathology.
Strengths and Limitations
Strengths of this study included the large longitudinal and clinically diverse Dutch cohort, which allowed us to investigate associations at the single-symptom level over time. The findings were tested in an independent replication sample, which also had repeated assessments that captured change over time. The replication was conducted in an English-speaking cohort at a different clinical stage, supporting generalizability across language and illness stage.
Several limitations should be noted. Direct 1-to-1 replication was not possible because the Dutch and US cohorts differed in language, clinical setting, diagnostic composition, outcome measures, and component structure. The US sample included bipolar I disorder with psychotic features in addition to schizophrenia spectrum disorders and required at least moderate positive or disorganized symptoms at admission, whereas the Dutch studies focused only on schizophrenia spectrum disorders and applied additional study-specific criteria, including antipsychotic treatment requirements in 1 parent study. In addition, although the speech elicitation tasks and analytic pipelines were broadly comparable across cohorts, they were not identical. Replication should therefore be understood as consistency in the direction of speech-symptom associations and in estimative performance across clinically distinct samples rather than replication of identical constructs. These speech-symptom associations reflect adjusted effects within multivariable models, and their clinical interpretation should therefore be tempered by interview context, recording conditions, medication, extrapyramidal symptoms, or sedation. In addition, for symptoms evaluated largely based on speech behavior, such as lack of spontaneity and flow of conversation, associations with speech-derived features may partly reflect construct overlap rather than fully independent estimator relationships.
Information on race was unavailable in the Dutch cohort, so migration background was used as a proxy measure, limiting direct comparability. In addition, both cohorts were predominantly male, which may limit generalizability and may have influenced speech-related findings given known sex differences in acoustic characteristics. Finally, demographic data were missing for some participants, including sex, migration background, and race, which may have limited subgroup analyses.
Conclusions
This cohort study of individuals with schizophrenia found that psychotic symptoms left measurable traces in language. Quantifying these traces during brief open-ended tasks enabled objective, interpretable, and real-time assessment of symptom dynamics at the single-feature level. Speech-based tools, once validated in practice, may support precision monitoring and provide a window of opportunity to prevent relapse and other undesirable outcomes.
eAppendix 1. Inclusion Criteria
eAppendix 2. Semistructured Interview
eAppendix 3. Audio Processing and Feature Extraction (Dutch Cohort)
eAppendix 4. Audio Processing and Feature Extraction (US Cohort)
eAppendix 5. Robustness Analysis (Winsorization)
eAppendix 6. Principal Components Analysis
eAppendix 7. Effects of Demographics
eAppendix 8. Individual Symptom Prediction
eReferences.
Data Sharing Statement
References
- 1.Silva AM, Limongi R, MacKinley M, Ford SD, Alonso-Sánchez MF, Palaniyappan L. Syntactic complexity of spoken language in the diagnosis of schizophrenia: a probabilistic Bayes network model. Schizophr Res. 2023;259:88-96. doi: 10.1016/j.schres.2022.06.011 [DOI] [PubMed] [Google Scholar]
- 2.Palaniyappan L. More than a biomarker: could language be a biosocial marker of psychosis? NPJ Schizophr. 2021;7(1):42. doi: 10.1038/s41537-021-00172-1 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3.Corcoran CM, Cecchi GA. Anomalies of language in schizophrenia: a trajectory of mathematical and computational analyses. World Psychiatry. 2024;23(2):212-213. doi: 10.1002/wps.21192 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Mota NB, Furtado R, Maia PPC, Copelli M, Ribeiro S. Graph analysis of dream reports is especially informative about psychosis. Sci Rep. 2014;4:3691. doi: 10.1038/srep03691 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Hüppi RM, Bautista L, Cecere G, et al. TRUSTING: an international multicenter observational study of speech-based relapse prediction in psychosis using explainable AI. medRxiv. Preprint posted online November 17, 2025. doi: 10.1101/2025.11.14.25339774 [DOI] [PMC free article] [PubMed]
- 6.Corcoran CM, Carrillo F, Fernández-Slezak D, et al. Prediction of psychosis across protocols and risk cohorts using automated language analysis. World Psychiatry. 2018;17(1):67-75. doi: 10.1002/wps.20491 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.Tang SX, Spilka MJ, John M, et al. Automated speech and language markers of longitudinal changes in psychosis symptoms. NPP Digit Psychiatry Neurosci. 2025;3(1):13. doi: 10.1038/s44277-025-00034-z [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Worthington M, Efstathiadis G, Yadav V, et al. Measurement of schizophrenia symptoms through speech analysis from PANSS interview recordings. Front Psychiatry. 2025;16:1571647. doi: 10.3389/fpsyt.2025.1571647 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9.de Boer JN, Voppel AE, Brederoo SG, Wijnen FNK, Sommer IEC. Language disturbances in schizophrenia: the relation with antipsychotic medication. NPJ Schizophr. 2020;6(1):24. doi: 10.1038/s41537-020-00114-3 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Shafer A. Meta-analysis of the Brief Psychiatric Rating Scale factor structure. Psychol Assess. 2005;17(3):324-335. doi: 10.1037/1040-3590.17.3.324 [DOI] [PubMed] [Google Scholar]
- 11.de Boer JN, Voppel AE, Brederoo SG, et al. Acoustic speech markers for schizophrenia-spectrum disorders: a diagnostic and symptom-recognition tool. Psychol Med. 2023;53(4):1302-1312. doi: 10.1017/S0033291721002804 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Ciampelli S, de Boer JN, Voppel AE, et al. Syntactic network analysis in schizophrenia-spectrum disorders. Schizophr Bull. 2023;49(suppl 2):S172-S182. doi: 10.1093/schbul/sbac194 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13.Ciampelli S, de Boer JN, Voppel AE, et al. Syntactic network analysis in first-episode psychosis: toward generalizability. Schizophrenia (Heidelb). 2025;11(1):147. doi: 10.1038/s41537-025-00693-z [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14.He R, Palominos C, Zhang H, Alonso-Sánchez MF, Palaniyappan L, Hinzen W. Navigating the semantic space: unraveling the structure of meaning in psychosis using different computational language models. Psychiatry Res. 2024;333:115752. doi: 10.1016/j.psychres.2024.115752 [DOI] [PubMed] [Google Scholar]
- 15.Eyben F, Weninger F, Gross F, Schuller B. Recent developments in openSMILE, the Munich open-source multimedia feature extractor. In: MM ’13: Proceedings of the 21st ACM International Conference on Multimedia. Association for Computing Machinery; 2013:835-838. doi: 10.1145/2502081.2502224 [DOI] [Google Scholar]
- 16.Cohen AS, Schwartz E, Le TP, et al. Digital phenotyping of negative symptoms: the relationship to clinician ratings. Schizophr Bull. 2021;47(1):44-53. doi: 10.1093/schbul/sbaa065 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17.Pintos AS, Hui CLM, De Deyne S, et al. A longitudinal study of semantic networks in schizophrenia and other psychotic disorders using the word association task. Schizophr Bull Open. 2022;3(1):sgac054. doi: 10.1093/schizbullopen/sgac054 [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
eAppendix 1. Inclusion Criteria
eAppendix 2. Semistructured Interview
eAppendix 3. Audio Processing and Feature Extraction (Dutch Cohort)
eAppendix 4. Audio Processing and Feature Extraction (US Cohort)
eAppendix 5. Robustness Analysis (Winsorization)
eAppendix 6. Principal Components Analysis
eAppendix 7. Effects of Demographics
eAppendix 8. Individual Symptom Prediction
eReferences.
Data Sharing Statement
