2026 ECNP Congress

PS02-2040

Computational analysis of speech acoustics in first-episode bipolar disorder and psychosis

Elif Kizilay — B. Arslan, E. Cesim, M. Demir, B. Verim, S. Uzman Ozbek, Y.E. Turan, C. Demirlek, B. Yalincetin, E. Bora · Dokuz Eylul University, Department of Neurosciences, İzmir, Türkiye

2026 ECNP Congress

Introduction

Schizophrenia and bipolar disorder present with both overlapping and distinct characteristics, often leading to misclassification during their first episodes. These shared and distinct features also extend to their speech and linguistic patterns [1]. Recently, automated computational methods have emerged as promising tools, enabling a more objective assessment of patients’ acoustic features compared to traditional clinical evaluations [2].

We hypothesise that patients with first-episode psychosis (FEP) and first-episode bipolar disorder (FEBD) would demonstrate quantifiable acoustic patterns that highlight their shared and distinct pathologies. Furthermore, we aim to classify patients and controls, as well as the clinical groups from each other, using machine learning models.

Methods

Audio recordings were collected from 47 Turkish-speaking patients with FEP, 36 patients with FEBD, and 67 healthy controls (HCs) during a free conversational speech task. Speech segments of participants were isolated using Praat software [3]. To quantify acoustic patterns, the openSMILE toolkit was used to extract the eGeMAPS feature set [4], comprising 88 parameters related to frequency, energy/amplitude, spectral, and temporal features.

Group differences were analyzed using Welch’s ANOVA, followed by Games-Howell post-hoc tests for pairwise comparisons. Additionally, machine learning models were developed using 10-fold stratified cross-validation to classify across groups (All Patients vs. HC, FEBD vs. HC, FEP vs. HC, and FEBD vs. FEP). Logistic regression with L2 regularization was applied to perform the classifications and prevent overfitting.

Results

Our analysis revealed significant differences across the groups in 26 out of 88 acoustic features (ANOVA, F range = [3.16 - 9.16], p < 0.05), with 23 features remaining significant in post-hoc analyses (p < 0.05). Compared to HCs, both patient groups exhibited a reduced pitch range and lower jitter with larger variability, indicating overlapping abnormalities in prosody and voice quality.

Additionally, FEP demonstrated widespread acoustic differences compared to HC, characterized by decreased shimmer with greater variability, reduced relative energy of the first, second and third formant compared to fundamental frequency with greater variability, and a lower loudness peak rate. FEP exhibited an increased spectral slope, reflecting a tenser voice, as well as longer and more variable pausing patterns compared to HC. FEBD was differentiated from FEP by demonstrating significantly higher values across 8 out of 10 loudness-related features.

The logistic regression model achieved the highest performance in distinguishing FEP from HC (accuracy = 0.720, ROC/AUC = 0.804), followed by all patients vs. controls (accuracy = 0.653, ROC/AUC = 0.718) and FEBD vs. HC (accuracy = 0.672, ROC/AUC = 0.715). The classification between FEBD and FEP yielded a lower performance (accuracy = 0.556, ROC/AUC = 0.589).

Conclusion

Patients demonstrated both shared and distinct acoustic patterns compared to HCs. Only loudness-related features differed between FEP and FEBD. Machine learning results indicate that acoustic speech features hold promise for differentiating patients from controls; however, classification performance between the two clinical groups remained at chance level due to their profound acoustic overlap. While computational acoustic analysis serves as a promising objective tool for characterizing FEP and FEBD, our findings also highlight the current challenges in identifying distinct digital biomarkers for these overlapping disorders.

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. De Boer, J.N., Voppel, A.E., Brederoo, S.G., Schnack, H.G., Truong, K.P., Wijnen, F.N.K., Sommer, I.E.C., 2023. Acoustic speech markers for schizophrenia-spectrum disorders: a diagnostic and symptom-recognition tool. Psychol. Med. 53, 1302–1312. https://doi.org/10.1017/S0033291721002804
  3. Boersma, P., Weenink, D., 2013. Praat software. University of Amsterdam: Amsterdam.
  4. Eyben, F., Scherer, K.R., Schuller, B.W., Sundberg, J., Andre, E., Busso, C., Devillers, L.Y., Epps, J., Laukka, P., Narayanan, S.S., Truong, K.P., 2016. The Geneva Minimalistic Acoustic Parameter Set (GeMAPS) for Voice Research and Affective Computing. IEEE Trans. Affective Comput. 7, 190–202. https://doi.org/10.1109/TAFFC.2015.2457417