2026 ECNP Congress

PS01-1259

Early detection of psychosis with speech-language analysis using speech-to-text technology and AI in Jakarta Indonesia

Khamelia Malik — P.A. Yusuf, A.A. Krisnadhi, O. Riandi · Faculty of Medicine Universitas Indonesia, Department of Psychiatry Cipto Mangunkusumo National Hospital Jakarta - Neurocience & Brain Development IMERI Faculty of Medicine Universitas Indonesia, Jakarta, Indonesia

2026 ECNP Congress

Introduction

Schizophrenia prevalence in Indonesia is 6.7 per 1,000 households. Prevalence of people over 15 who suffer from mental disorders is increased, around 9.8% of the total population Duration of untreated psychosis lasts between 6 months and 2 years. Adolescent with prodromal psychosis can experience worsening of the condition into a severe mental disorder, so early recognition of this phase is very important to prevent the incidence of schizophrenia. The importance of language as a biomarker in psychiatry is recognized, where language impairment reflects disorganized thinking. Through an approach incorporating automated analysis, this study overcomes the subjectivity often inherent in traditional clinical assessments and provides more consistent monitoring in a high-risk population.

Methods

The methods include syntactic and semantic analysis to understand complex language patterns. This technique measures various factors, including sentence length, syntactic complexity, and emotional elements in speech. StethoSoul plays a key role in automating this analysis, facilitating its use in clinical settings. Phase 1: 150 adolescents aged 14–19 years using the Prodromal Questionnaire-Brief Indonesian version, the subjects were split into two groups: prodromal and normal. All participants were voice-recorded during interviews using an open-ended qualitative questionnaire. Syntactic and semantic analysis was carried out on all data which amounted to 1017 phrase segments and classified by machine learning. To evaluate semantic features, the probability of forming a group of words (language model incoherence, or LMI) and the representation of words in vectors (latent semantic analysis, or LSA, and windowed semantic similarity vector, or WSSV) were used. Phase 2: The classification model enhanced by adding 2 syntactic features, 9 semantic coherence features, and 3 determiner features was tested on 1,020 subjects at high school, hospital and social community facility.

Discussion

There was a significant difference in the minimum value of coherence between sentence phrases using latent semantic analysis between prodromal psychosis and normal adolescents. Subordinating conjunctions, personal pronouns, nouns, adjectives, prepositions, and proper nouns are language expressions with a lower frequency of use in the prodromal group. Negation was found to be more prevalent in the prodromal group The word-to-topic vector in the prodromal group had a greater value than the normal group. The decrease in coherence in prodromal psychosis often does not take the form of loose associations or apparent incoherence due to the mild degree of thought process impairment that occurs. Increasing the number of features from 4 to 14 raised the accuracy of the previous model from 57% to 82%. Early indications of discordance in speech not only help detect but also predict the potential development of psychotic symptoms. With the application of this technology, symptom detection can be performed more quickly and on a larger scale.

Conclusion

The analysis results indicate significant differences in syntactic and semantic parameters between adolescents with prodromal psychosis. Our AI-based language analysis technology, apps StethoSoul, offers an effective solution for early detection of psychosis. With its ability to reduce bias in assessment and quickly provide objective data, this method has the potential to improve treatment outcomes for individuals at risk of developing mental disorders.

References

Ministry of Health, Republic of Indonesia, 2019. National Report of Riskesdas 2018. 1st edn. Edited by Tim Riskesdas 2018. Jakarta: Lembaga Penerbit Badan Penelitian dan Pengembangan Kesehatan. Malik K et al.2023. Differences in syntactic and semantic analysis based on machine learning algorithms in prodromal psychosis and normal adolescents. Asian Journal of Psychiatry 85 (2023) 103633 Bechdolf, A., et al., 2012. Preventing progression to first-episode psychosis in early initial prodromal states. Br. J. Psychiatry 200, 22–29. https://doi.org/10.1192/bjp. bp.109.066357. Loewy, R.L., et al., 2012. Prodromal psychosis screening in adolescent psychiatry clinics. Early Interv. Psychiatry 6 (June 2010), 69–75. https://doi.org/10.1111/j.1751- 7893.2011.00286.x.