Speech analysis for differentiating bipolar disorder and major depressive disorder during euthymic states.

Abstract

OBJECTIVE: Differentiating major depressive disorder (MDD) from bipolar disorder (BP) is crucial for early diagnosis and targeted treatment. This study investigates speech, linguistic biomarkers, and machine learning models to differentiate between the MDD and BP patients using speech analysis during symptom-free periods. METHODS: Voice recordings from 191 patients (41 BP, 150 MDD) during euthymic phases were analyzed. Linguistic metrics, sentiment scores, and acoustic features were extracted and compared between the groups. Various machine learning models were employed to classify diagnoses based on these features. RESULTS: Significant differences were observed between BP and MDD patients. BP patients exhibited a shorter pause durations per total speaking duration and total word count (0.19 vs. 0.29; p < 0.05, 0.07 vs. 0.12; p < 0.05). MDD patients showed greater variability in speech intensity and peak amplitude during euthymic states. Sentiment analysis revealed no significant differences in broader emotional dimensions. Among the models tested, the Random Forest classifier achieved the highest predictive performance (AUC = 0.807), underscoring its effectiveness in differentiating BP from MDD. CONCLUSION: Speech analysis during euthymic periods reveals distinct linguistic and acoustic patterns between MDD and BP patients. These findings highlight the potential of speech features as non-invasive biomarkers for mood disorder classification, offering promising avenues for enhancing clinical decision-making. However, given the sample size, the results are preliminary and warrant replication in larger, independent cohorts.