2026 ECNP Congress

S20.04

The geometry of thought: mapping psychiatric cognition from speech and multimodal artificial intelligence inference

Baihan Lin — Icahn School of Medicine at Mount Sinai, Departments of AI and Human Health- Psychiatry- and Neuroscience, New York, United States

2026 ECNP Congress

Language-derived psychiatric scores may reflect the person, the rater or the instrument. Computational Mapping of Psychometric Assessment in Speech and Semantics (COMPASS) maps dialogue into questionnaire-informed representations, initially applied to patient–therapist interactions [1].

Across 275 interviews, a systematic evaluation of depression ratings finds that model choice contributes more score variation than stable person differences (unpublished preprint) [2]. Similar discrimination conceals different screening decisions; calibration reduces disagreement without eliminating it.

COMPASS 2.0 uses psychometric representational similarity analysis (RSA) to compare item relationships in speech, self-report and wording (unpublished preprint) [3]. Interview questions and wording generate reproducible structure. Registered tests do not establish residual symptom structure after centering and whitening. Yet correctly paired scores retain information about overall distress. Shuffling person pairings distinguishes population structure from individual validity.

Our related acoustic work in cognitive impairment [4] and temporal speech analysis in mania [5] motivate future comparisons of language alone, acoustics alone and their combination, evaluated against independent clinical criteria and change within individuals.

[1] Lin et al. COMPASS: computational mapping of patient-therapist alliance strategies with language modeling. Translational Psychiatry (2025). [2] Lin. Language-model ratings of depression reflect the rater more than the patient. preprint. arXiv:2610.08501 (2026). [3] Lin. COMPASS 2.0: psychometric representational similarity analysis distinguishes symptom structure from personal signal. preprint. arXiv:2610.06615 (2026). [4] Colonel et al. Acoustic Analysis of Primary Care Patient–Clinician Conversations to Screen for Cognitive Impairment. JAMA Neurology (2026). [5] Joyce et al. Temporal elements of speech in mania. Translational Psychiatry (2026).