2026 ECNP Congress

PS02-2035

Automated linguistic examination across stages of psychosis

Berat Arslan — E. Kizilay, Y.E. Turan, E. Cesim, M. Demir, M.S. Eyuboglu, B. Verim, C. Demirlek, B. Yalincetin, N. Zorlu, E. Bora · Dokuz Eylul University, Department of Neurosciences, İzmir, Türkiye

2026 ECNP Congress

Introduction

Stages of psychosis encompass a clinical progression from clinical high-risk for psychosis (CHR-P) to transition to first-episode psychosis (FEP) and chronic schizophrenia (SCH). Individuals in the prodromal phase often exhibit subtle abnormalities in language that potentially worsen along this trajectory. While natural language processing (NLP) methods offer objective tools to detect these abnormalities [1] [2], studies simultaneously examining automated linguistic features across these stages are limited.

We aimed to apply NLP methods such as semantic similarity as well as normative thematic and lexical choice across these stages. This approach may help to map the abnormalities in language from prodromal phase to chronic illness and identify potential speech-based biomarkers.

Methods

Speech samples were collected from 77 individuals at CHR-P, 89 patients with FEP, 112 patients with SCH and 113 healthy controls (HCs) during descriptions of Thematic Apperception Test (TAT) pictures. Transcripts were preprocessed using the NLTK for tokenization into words and sentences, and punctuation removal. Word embeddings were obtained using a FastText Turkish model [3], from which local similarity (neighboring words) and all-pair similarity were computed, including their mean, variance, 5th and 95th percentiles.

Lexical and thematic overlap metrics were used to investigate the degree of normative choice of participants. Lexical overlap with HCs was quantified using the Jaccard similarity coefficient, while thematic overlap was calculated via cosine similarity between SBERT-derived centroid embeddings [4]. Group differences were analyzed using ANOVA with Games-Howell post-hoc tests for pairwise comparisons.

Results

Significant group differences were observed for nine out of ten linguistic metrics, except for mean local similarity. Post-hoc comparisons showed local similarity variance was higher across all clinical stages compared to HCs (p: .027 to < .001). The 5th percentile of local similarity was lower in SCH than in CHR-P (p = .004), while the 95th percentile was higher than HCs (p = .003). Mean all-pair similarity was significantly elevated in all clinical groups relative to HCs (p: .046 to < .001). All-pair similarity variance was higher in FEP and SCH than HCs (p’s < .001), with SCH also exceeding the CHR-P group (p = .002). SCH patients exhibited a lower 5th percentile and a higher 95th percentile of all-pair similarity compared to both HCs and CHR-P (p’s < .05). Lexical overlap followed a nearly stepwise decline (HC > FEP = CHR-P > SCH), whereas thematic overlap was reduced only in the SCH group compared to HCs (p <.001).

Conclusion

This study extends current literature by examining automated linguistic metrics across stages of psychosis in a Turkish-speaking population. The combination of higher mean all-pair similarity and elevated 95th percentiles may reflect repetitive and restricted speech in patients, potentially serving as objective proxies for clinical symptoms such as perseveration, circumstantiality, and overall cognitive decline, which should be investigated in future work. The increased variance and reduced 5th percentiles highlight a speech irregularity and incoherence particularly in SCH. The observed stepwise decline in lexical overlap further supports this trajectory. These findings demonstrate that NLP-derived metrics may help to identify speech-based biomarkers from the prodromal to chronic stages of psychosis.

References

  1. Arslan, B., Kizilay, E., Verim, B., Demirlek, C., Demir, M., Cesim, E., Eyuboglu, M.S., Ozbek, S.U., Sut, E., Yalincetin, B., Bora, E., 2024. Computational analysis of linguistic features in speech samples of first-episode bipolar disorder and psychosis. J. Affect. Disord. 363, 340–347. https://doi.org/10.1016/j.jad.2024.07.102
  2. Kizilay, E., Arslan, B., Verim, B., Demirlek, C., Demir, M., Cesim, E., Eyuboglu, M.S., Uzman Ozbek, S., Sut, E., Yalincetin, B., Bora, E., 2024. Automated linguistic analysis in youth at clinical high risk for psychosis. Schizophr. Res. 274, 121–128. https://doi.org/10.1016/j.schres.2024.09.009
  3. Grave, E., Bojanowski, P., Gupta, P., Joulin, A., Mikolov, T., 2018. Learning Word Vectors for 157 Languages. https://doi.org/10.48550/ARXIV.1802.06893
  4. Reimers, N., Gurevych, I., 2020. Making Monolingual Sentence Embeddings Multilingual using Knowledge Distillation, in: Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP). Association for Computational Linguistics, Online, pp. 4512–4525. https://doi.org/10.18653/v1/2020.emnlp-main.365