Acoustic features from speech as markers of depressive and manic symptoms in bipolar disorder: A prospective study.
Level 4 - case-series / case-control
Prospective longitudinal observational study (single-group cohort/case series) without an external control group.
PubMed 39118422 · doi:10.1111/acps.13735
What was done
In a prospective observational study, 51 patients with bipolar disorder used a dedicated smartphone app (BDmon) that passively extracted acoustic parameters (prosodic, spectral, and voice quality features) from daily phone calls over an average of 208 days. Clinicians evaluated psychiatric symptom severity using the Hamilton Depression Rating Scale-17 (HDRS-17) and the Young Mania Rating Scale (YMRS). Generalized mixed-effects models were used to analyze the relationship between acoustic metrics and clinical state and to predict the bipolar disorder phase.
What was found
The predictive generalized mixed-effects model achieved an accuracy of 70.9% to 71.4% for predicting bipolar disorder phase. Symptom-voice relationships differed substantially by sex. In males, higher mania severity correlated with louder speech (β = 1.6), higher vocal pitch (β = 0.71), clearer speech (β = 1.35), sharper speech (β = 0.95), and longer conversations (β = 1.64), whereas higher depressive severity in males correlated with quieter (β = -1.07) and less clear speech (β = -1.00). In females, mania severity was associated with opposite patterns (quieter voice: β = -0.27; lower pitch: β = -0.21; less clear speech: β = -0.25; no call length association), and no distinct acoustic correlations were identified for depressive symptom severity.
Why it matters
Passive voice analysis from everyday smartphone use may serve as an objective digital biomarker for tracking affective states in bipolar disorder, though diagnostic algorithms must account for marked sex differences in acoustic patterns.
Limits
The sample size was relatively small (n = 51) and lacked a healthy control group. Acoustic features in female patients showed weak or null associations with depressive symptoms. The abstract does not report external out-of-sample validation, medication effects, ambient noise controls, or specifics on missing call data over the 208-day observation window.