SPEECH FEATURES IN A DEVELOPMENTAL PERSPECTIVE: A SCREENING TOOL FOR PSYCHOSIS, DEPRESSION AND ANXIETY

Language production provides a window into cognitive development, social cognition, and mental health. To translate this potential into clinical practice, we developed a scalable speech-based platform integrating graph-theoretical, psycholinguistic, and emotional features from brief spontaneous narratives to support developmental monitoring, mental wellness screening, and clinical decision-making. Normative developmental trajectories were derived from 2149 participants spanning 0-23 years of formal education. Narratives elicited using positively valenced pictures were transformed into word graphs to derive two composite indices: Narrative Complexity (NC), reflecting structural language organization, and Emotional Narrative Complexity (ENC), integrating narrative structure and emotional content. Both indices were modeled using asymptotic functions calibrated by schooling and converted into percentile-based normative curves. NC (R 2 = 0.29, τ = 7.77) and ENC (R 2 = 0.31, τ = 6.97) increased with schooling, approaching saturation around middle school. These trajectories provide educationally calibrated references for evaluating linguistic and socioemotional development. Clinical utility was investigated in a transdiagnostic sample of 120 participants, comprising healthy controls (n = 38), mood disorders (n = 28), affective psychosis (n = 36), and non-affective psychosis (n = 18). Although ENC alone did not discriminate between symptom-risk or diagnostic groups, classifiers combining its constituent variables with education and sex achieved AUCs of 0.726 for symptom risk and 0.800 for diagnosis, increasing to 0.774 and 0.858, respectively, with additional linguistic features. Machine-learning models integrating graph-derived, psycholinguistic, and emotional features further estimated symptom dimensions assessed by PANSS, GAD-7, and PHQ-9, achieving AUCs of 0.74-0.92 across severity thresholds while maintaining sensitivity ≥ 0.80. Individual linguistic variables also differed significantly across symptom-severity strata, suggesting distinct associations with psychopathological dimensions. Together, these findings establish an integrated speech-based framework spanning normative developmental assessment, mental wellness screening, and symptom-informed clinical support. By combining educationally calibrated trajectories with interpretable linguistic features and machine learning, this approach offers scalable digital markers to identify atypical developmental patterns, inform preventive strategies, and assist clinical assessment and decision-making.