JMIR Publications

ABSTRACT

Background:

Automated speech analysis offers a low-burden approach for identifying objective markers of psychiatric assessment. While alterations in speech have been documented in both schizophrenia (SZ) and major depressive disorder (MDD), direct comparisons between these conditions within the same cohort remain limited.

Objective:

This study aimed to examine whether automatically extracted acoustic and linguistic speech features differentiate individuals with SZ, individuals with MDD, and healthy controls (HC), and whether these features are associated with clinical symptom severity and diagnostic classification performance.

Methods:

A total of 66 participants (22 with SZ, 22 with MDD, and 22 HC) completed a tablet-based speech assessment, at two time points. Speech tasks included a semi-structured picture description task and emotional storytelling prompts. Automatic extraction of acoustic, temporal, spectral, and linguistic features was performed. Group differences were tested using Kruskal-Wallis and Mann-Whitney U tests, symptom–speech associations were examined using correlation analyses with correction for multiple testing, and diagnostic classification was evaluated using machine learning models with leave-one-out cross-validation. Speech-based models were compared with demographic and symptom-based baseline models.

Results:

The clearest group differences emerged in the picture description task. Participants with SZ exhibited reduced speech output, lower informational content, and altered acoustic features compared to those with MDD or HC. Omnibus group effects survived correction for multiple testing for certain speech features, such as correct concepts, number of pauses and word count (adjusted P<.05). In contrast, univariate differences in emotional storytelling tasks were limited and mainly observed for the positive storytelling task at the second time point. Symptom-speech associations were selective after correction for multiple testing, including associations between depressive symptoms and lower loudness as well as between PANSS scores and jitter-related measures. Within the present sample, speech-based machine learning models showed the most consistent classification performance for distinguishing SZ from MDD, with area under the curve (AUC) values of approximately 0.91 to 0.96 across tasks, whereas performance for distinguishing HC from SZ was moderate.

Conclusions:

Automated speech features captured diagnostically relevant information across SZ, MDD, and HC, with the clearest differentiation observed between SZ and MDD. Semi-structured picture description appeared particularly sensitive to disorder-related differences in speech output and informational content. These results suggest that automated speech analysis could be used as a complementary tool for differential diagnosis and symptom monitoring, but larger multisite studies with external validation are required before clinical implementation.