From Symptom Networks to Conversation Networks: A Cross-Sectional Study Mapping the Topology of Suicide-Related Clinical Dialogue

Key Findings

• — Session-level LLM C-SSRS severity agreed with independent expert ratings on 30 sessions (Spearman rho 0.885, ICC(2,1) 0.708, quadratic-weighted kappa 0.70; 90.0% within one level); no case rated suicidal by the expert was rated non-suicidal by the model, which tended to overestimate (bias +0.57)
• — Of 55,796 utterance segments, 364 suicidality-relevant utterances came from therapists (71.9%) and 142 from patients (28.1%); 25 sessions had none
• — Active ideation (n=67) and passive wish (n=33) were the most frequent patient categories; non-suicidal self-injury, interrupted attempt, preparatory behaviour and third-person mention never occurred
• — Active suicidal ideation acted as a bridge node from which escalation pathways to method, plan and attempt categories originated; closeness centrality fell monotonically with C-SSRS severity
• — The network was centred on therapist inquiry with high reciprocity and small-world characteristics

Methodology Notes

Adults aged 18-65 with schizophrenia-spectrum or affective disorders, University Hospital Frankfurt and Privatklinik Dr. Amelung; trial registration DRKS00027878; interviews from 2023-11-01 (the v1 text leaves the end date as an unfilled placeholder); WhisperX transcription with manual pseudonymisation. Utterance-level LLM labels were not validated; validation was at session level only. Clinician-patient dialogue, no chatbot involved. Posted 2026-09-29 (v1), not peer reviewed. Funded by LOEWE (DYNAMIC), ATHENE (PADAM) and the Reiss Foundation; two authors report pharmaceutical honoraria.

Sources