JMIR Publications

ABSTRACT

Background:

Bipolar disorder (BP) affects 1–3% of adults worldwide, with diagnostic delays averaging 6 years and high initial misdiagnosis rates. The absence of objective diagnostic measures contributes to clinical uncertainty and suboptimal outcomes.

Objective:

This study aimed to (1) develop a machine learning model for BP diagnosis and mood state classification using linguistic and acoustic features, (2) identify features associated with different mood states by comparing each with healthy controls (HC), and (3) explore whether features distinguish between mood states within BP.

Methods:

We conducted a six-year prospective study involving 61 BP patients and 99 HC across Japan. Speech samples were collected via semi-structured interviews over 2–5 visits. We extracted 6,373 acoustic features (OpenSMILE 3.0, ComParE 2016) and 211 linguistic features. LASSO-regularized logistic regression with nested cross-validation and age- and sex-matched undersampling was used. Mood states were classified as depressive (DS: HAMD-17 > 7, YMRS < 8), manic/mixed symptoms (MS/MX: YMRS > 7), or euthymic (ES: HAMD-17 < 8, YMRS < 8). Seven classification tasks included BP vs. HC, each mood state vs. HC, and inter-state comparisons. Feature importance was validated via permutation testing (p < 0.05).

Results:

The model achieved high accuracy for BP vs. HC (AUC = 0.916, 95% CI 0.854–0.958, sensitivity = 0.940, specificity = 0.898) and strong performance for mood states vs. HC (AUC = 0.805–0.874). Only acoustic features achieved statistical significance (p < 0.05), including Mel-Frequency Cepstral Coefficients, spectral characteristics, and voice quality measures, showing strong correlations with diagnostic probabilities (|rs| = 0.6–0.85) and symptom severity (HAMD-17: rs = -0.51; YMRS: rs = -0.37). With sample independence ensured for inter-state comparisons, the depressive vs. euthymic comparison significantly exceeded chance (AUC = 0.760, permutation p = 0.030), while manic/mixed symptoms vs. euthymic (AUC = 0.601, p = 0.198) and depressive vs. manic/mixed symptoms (AUC = 0.553) did not reach significance, suggesting markers primarily reflect persistent disease characteristics with partial state-associated components in depressive states.

Conclusions:

Multimodal speech analysis, primarily driven by acoustic features, provides objective tools for BP diagnosis (AUC > 0.9), with features detectable even during remission. Although depressive states showed partial distinguishability from euthymic states, overall limitations remain for monitoring mood state changes. Future research should validate findings in diverse populations and examine specificity against other psychiatric conditions.