Language in schizophrenia: relation with diagnosis, symptomatology and white matter tracts

The study investigates the relationship between spoken language characteristics and schizophrenia, focusing on diagnostic classification, symptom severity, and neurobiological underpinnings in white matter language pathways. Language disturbances (formal thought disorder) are core features of schizophrenia, encompassing positive symptoms (e.g., idiosyncratic associations, neologisms) and negative symptoms (e.g., poverty of speech, reduced grammatical complexity). Traditional observer-based ratings lack sensitivity for subtle language form deviations, while computational linguistic tools enable objective, reproducible analyses. Given evidence for widespread white matter abnormalities and disordered connectivity in schizophrenia, and dual-stream models of language processing (dorsal: SLF/AF; ventral: ILF/IFOF/UF), the authors test whether quantitative speech measures can classify patients vs controls, relate to PANSS symptom domains, and predict white matter integrity (FA/MD) within language tracts.

Prior research evidences language disorder as a key symptom of schizophrenia, typically assessed via subjective FTD scales. Recent computational approaches allow quantitative analyses across phonetics, syntax, and semantics, but have been sparsely applied in schizophrenia. Neurobiologically, schizophrenia is associated with widespread white matter differences (ENIGMA DTI meta-analysis) and altered functional/structural connectivity. FTD has been linked to aberrations in language networks and speech graph connectedness relates to structural and functional markers in psychotic disorders. Dual-stream models assign sound-to-meaning mapping to ventral tracts (ILF, IFOF, UF) and auditory-motor integration to dorsal tracts (SLF, AF). However, specific associations between white matter microstructure of language pathways and quantitative language disturbances remain underexplored.

Design: Cross-sectional study including 26 patients with schizophrenia spectrum disorders and 22 healthy controls (2015–2018, University Medical Center Utrecht). Inclusion criteria: age ≥18, native Dutch speakers. Patients met DSM-IV diagnoses (schizophrenia, schizophreniform, schizoaffective, psychotic disorder NOS) confirmed by treating psychiatrist and CASH interview; controls screened with CASH to exclude psychiatric history and family history of psychosis. Exclusions: uncorrected hearing/speech deficits, MRI contraindications, left-handedness. PANSS assessed in patients; antipsychotic dosages converted to chlorpromazine equivalents. Ethics approved; informed consent obtained. Language data: Semi-structured, 15-minute neutral-topic interviews elicited spontaneous speech; participants blinded to language focus until after. Audio recorded via head-worn microphone to 44.1 kHz/16-bit. Interviewer and participant speech streams separated manually (inter-rater reliability 97.7%); files normalized to 60 dB. Praat Script Syllable Nuclei v2 extracted speech/articulation metrics; pauses defined as silences >200 ms. Transcripts followed CHILDES-CHAT; CLAN tools (EVAL, FLUCALC) computed measures of fluency and complexity. Language variables: articulation rate; average pause duration; average speaking turn duration; percentage of time speaking; mean length of utterance (MLU, morphemes); type-token ratio (TTR); clauses per utterance; noun-verb ratio; open-closed ratio; disfluencies (per word); pause-to-word ratio. DTI acquisition: Philips Achieva 3T, two diffusion scans (b=1000 s/mm², 30 directions, 5 b0; 2 mm slices; TE=68 ms; TR=7011 ms; SENSE=3; reversed readout on second scan for susceptibility correction). Participants watched TV and were instructed to minimize movement. DTI preprocessing: FSL (5.0.9) pipeline—brain extraction; TOPUP for susceptibility; eddy for motion/eddy current correction; tensor fitting to compute FA and MD. TBSS: non-linear registration to FMRIB58_FA, skeletonized at FA>0.2. ROI masks from JHU ICBM-DTI-81 and tractography atlases for bilateral SLF, UF, AF (approximated using temporal branch of SLF), ILF, IFOF; mean FA/MD extracted for ROIs; voxel-wise group comparisons with FSL randomize. Statistics: SPSS v25. Group differences in demographics via t-tests/χ²/Mann–Whitney. One-way MANCOVA (covariate: age) tested group differences in language measures. Backward binary logistic regression predicted group membership from language variables with age, gender, education as covariates. Correlations between PANSS and language variables examined (item-level exploratory); FDR correction applied. Backward multivariate linear regressions assessed whether language variables predicted mean FA (per ROI, mean of language tracts, and whole-brain FA); analogous analyses for MD. Multiple comparison control via FDR.

• Sample: 26 patients (mean age 26.7±5.43 years; 76.9% male), 22 controls (24.3±4.40 years; 86.4% male); no significant group differences in age, gender, or education.
• Medication: Chlorpromazine-equivalent dose showed no significant correlations with PANSS subscores, language measures, or language tract metrics.
• Language group differences: MANCOVA (age as covariate) showed significant main effect of group on language characteristics (F(11,35)=2.565, Pillai’s trace=0.446, p=0.017). Post hoc tests indicated patients spoke more slowly, spent a smaller proportion of time speaking, produced shorter utterances (lower MLU), had higher TTR, and used fewer clauses per utterance than controls (text reports p<0.05). Table 3 univariate results show significant differences for: total words (p=0.008), speaking turn duration (p=0.011), percentage of time speaking (p=0.034), MLU (p=0.009), TTR (p=0.001); articulation rate and clauses per utterance did not reach univariate significance in Table 3.
• Diagnostic classification: Binary logistic regression (predictors: MLU, clauses per utterance; covariates: age, education) achieved Nagelkerke R²=0.733; Hosmer–Lemeshow p=0.874 (good fit). Sensitivity 88.5% and specificity 81.8% (abstract rounded to 89% and 82%).
• Symptom correlations: PANSS negative subscale correlated negatively with articulation rate (r=-0.414, p=0.036), speaking turn duration (r=-0.420, p=0.033), percentage of time speaking (r=-0.715, p<0.001), and MLU (r=-0.393, p=0.047), and positively with open-closed ratio (r=0.397, p=0.044). After FDR correction, only percentage of time speaking remained significant (p<0.001). Exploratory item-level: conceptual disorganization correlated with turn duration, percentage of time speaking, MLU, open-closed ratio; excitement with articulation rate; grandiosity with percentage of time speaking.
• White matter integrity (group level): No significant group differences in mean FA (F(11,35)=0.783, p=0.655) or mean MD (F(11,33)=1.351, p=0.242) across language tracts or whole brain. TBSS voxel-wise analyses revealed significantly decreased FA clusters in patients across all language ROIs and additional regions (corpus callosum, cingulum, corona radiata) after multiple comparison correction.
• Language–FA associations: Multivariate regressions showed language measures predicted mean FA substantially. • Patients: Adjusted R²=0.467 for mean FA of language tracts (significant predictors included pause duration [+], MLU [−], noun-verb ratio [−], clauses per utterance [−]); Adjusted R²=0.516 for whole-brain FA (pause duration [+], speaking turn duration [+], MLU [−], noun-verb ratio [−]). • Controls: Adjusted R²=0.483 for mean FA of language tracts (MLU [+], clauses per utterance [−]); Adjusted R²=0.331 for whole-brain FA (speaking turn duration [+], pause-to-word ratio [−]).
• ROI-specific models (Table 5): In controls, left-hemisphere tracts often showed stronger associations (e.g., left AF adjusted R²=0.753; predictors included MLU [+], pause duration [+], disfluencies [+], clauses per utterance [−], pause-to-word ratio [−]). In patients, significant predictors frequently included pause duration (+) and noun-verb ratio (−) across multiple tracts; right UF in patients was strongly associated with disfluencies and pauses.

Quantitative analysis of spontaneous speech distinguishes schizophrenia patients from healthy controls with high sensitivity and specificity using simple linguistic predictors (MLU and clauses per utterance), supporting clinical utility. Language disturbances relate predominantly to negative symptom severity; the robust association after correction was with reduced percentage of time speaking, indicating diminished output/engagement reflects negative symptoms. Despite no group differences in mean FA/MD over atlas-based ROIs, TBSS revealed microstructural FA reductions in patients across language pathways and other major tracts, indicating subtle, spatially localized abnormalities consistent with diffuse developmental connectivity alterations. Language variables captured both complexity (MLU, clauses per utterance, noun-verb ratio) and speaking efficiency/processing speed (pause duration, pause-to-word ratio). Their associations with FA align with known developmental trajectories (increasing white matter integrity correlates with more complex, efficient language) and with prior links between FA and processing efficiency. Notably, in controls, language measures predicted FA more specifically within language tracts than whole brain, suggesting specialization; in patients, prediction was similar or stronger for whole-brain FA, implying reduced specialization or broader cognitive dysfunction impacting language. Lateralization patterns also differed: controls showed stronger left-hemispheric tract associations, consistent with typical language lateralization; patients lacked clear left dominance, aligning with evidence of reduced lateralization in schizophrenia. The findings underscore that functional language output is more discriminative of case status than gross structural white matter metrics, consonant with literature showing modest structural differences but prominent functional alterations in schizophrenia.

Quantitative, computational analyses of spontaneous speech provide sensitive and specific markers distinguishing schizophrenia patients from controls and reflect negative symptom severity. Linguistic measures robustly predict white matter integrity within language pathways in both groups, revealing meaningful brain–behavior relationships despite limited group differences in average FA/MD. These results support incorporating objective language analytics into schizophrenia research and clinical workflows, aligning with precision psychiatry and RDoC frameworks. Future work should replicate in larger, independent and high-risk cohorts, apply cross-validation, longitudinally track illness course and treatment effects, refine tractography (especially AF), include medicated control groups, and account for cognitive/IQ influences to clarify specificity and generalizability.

• Approximation of arcuate fasciculus: no dedicated atlas mask; temporal branch of SLF used as proxy, warranting cautious interpretation for AF-related results.
• Small sample size, recent-onset patients; limits power to detect group ROI mean FA/MD differences and generalizability; need replication in larger samples.
• Most patients were medicated and relatively stable; likely reduced correlations with positive symptoms; absence of a medicated non-psychosis control group.
• No direct IQ assessment/control (education used as proxy); potential influence of intelligence on language measures cannot be excluded.
• Lack of cross-validation of the diagnostic model; performance estimates may be optimistic.
• Exclusion of left-handed participants limits generalizability regarding lateralization.
• Topic-neutral interview design may limit generalization to more naturalistic narrative speech with richer prosody/emotion.