LLM Questionnaire Completion for Automatic Psychiatric Assessment

Evaluation and psychometric validity2024ACL AnthologyApproved editorial review

Authors: Gony Rosenman, Talma Hendler, Lior Wolf

Keywords: Large Language Models, Psychiatric Assessment, Mental Health, Depression (PHQ-8), PTSD (PCL-C), Psychological Interviews

Source: Open primary source (opens in a new tab)

3
Authors
17
Findings
57
Limitations
16
Evidence

Editorial summary

English

The paper presents LMIQ, a pipeline for transforming psychological-interview transcripts into numerical variables. An LLM receives a transcript, adopts the interviewee's perspective, and answers questions with values from 1 to 5. The resulting 135 responses feed a Random Forest that predicts two self-report scores: PHQ-8 for depressive symptoms and PCL-C for post-traumatic stress symptoms. The main experiment uses the 275 E-DAIC interviews and its predefined splits. Features comprise 70 mental-health questions, 25 Big Five personality questions, 15 therapeutic questions, and 25 questions called direct. The non-direct items were initially generated with GPT-4 and manually refined; the paper provides no clinical or psychometric validation of these new questionnaires.

With GPT-3.5 Turbo Instruct, LMIQ obtains test MSEs of 20.42 for PHQ and 192.93 for PTSD, compared with 24.16 and 195.55 when the same model is asked to complete only the direct questionnaires and its answers are summed. Mixtral improves the main results to 18.05 and 163.75. These are regression errors on questionnaire scores, not measures of clinical diagnosis. The paper also compares direct GPT-3.5 prediction, Ada-002 embeddings, MentaLLaMA, TF-IDF, and the dataset mean. LMIQ-Mixtral is the strongest variant in the main table, but no repeated runs, confidence intervals, or statistical tests are reported to quantify the stability or significance of the differences.

The ablation qualifies the claim that pooling every domain is most effective. With GPT-3.5, the 135-feature set is not the best test configuration for either task: personality plus direct questions reaches 20.26 PHQ MSE and 166.62 PTSD MSE, versus 20.42 and 192.93 with all features. Therapeutic plus direct also improves PTSD to 168.62 and mental health plus direct to 177.15. Variants without direct questions are substantially worse. This supports the presence of predictive signal in simulated answers, but it also reveals heavy dependence on questions that closely proxy the target instruments and shows that adding every domain can be harmful, especially for PTSD.

The appendix audit finds an instrument-fidelity problem. The paper declares eight PHQ-8 questions plus seventeen PCL-C questions, yet the published PHQ list contains nine items because it adds “How likely are you to volunteer your time to help others?”, which is not part of the PHQ-8. The official PHQ-8 asks about frequency during the past two weeks and uses 0–3 scoring; LMIQ's uniform prompt requests 1–5 agreement and removes that time anchor. These features are therefore not faithful PHQ-8 completions but model-generated ordinal proxies. The official PCL-C does use 17 items rated 1–5, although the paper likewise does not precisely document preservation of its time and stressful-experience anchors. The paper itself acknowledges that impersonated responses are not validated against human answers.

The defensible contribution is an exploratory study showing that structured responses generated from an interview can be more predictive than several text representations on one benchmark split. It does not establish automatic diagnosis or clinical validity. Targets are self-report severity scores rather than clinician diagnoses; thresholds, sensitivity, specificity, calibration, subgroup analysis, uncertainty, and prospective evaluation are absent. No code, derived data, or executable configuration is released, and the exact Mixtral provider/version, snapshots, seeds, cost, failures, and retries are not documented. Sending sensitive transcripts to a commercial API is justified with a general statement about de-identification and compliance, without documenting consent for this processing, ethics review, retention, residency, a processing agreement, or re-identification testing. In its present form, LMIQ should be understood as experimental feature extraction for benchmark regression, not a system ready for clinical decisions.

Español

El artículo presenta LMIQ, un proceso para transformar transcripciones de entrevistas psicológicas en variables numéricas. Un LLM recibe una transcripción, adopta la perspectiva de la persona entrevistada y contesta preguntas con valores de 1 a 5. Las 135 respuestas resultantes alimentan un Random Forest que predice dos puntuaciones de autoinforme: PHQ-8 para síntomas depresivos y PCL-C para síntomas de estrés postraumático. El experimento principal usa las 275 entrevistas de E-DAIC y sus particiones predefinidas. Las variables incluyen 70 preguntas de salud mental, 25 de personalidad Big Five, 15 terapéuticas y 25 denominadas directas. Las preguntas no directas se generaron inicialmente con GPT-4 y se refinaron manualmente; el artículo no aporta una validación clínica o psicométrica de esos nuevos cuestionarios.

Con GPT-3.5 Turbo Instruct, LMIQ obtiene en prueba MSE 20,42 para PHQ y 192,93 para PTSD, frente a 24,16 y 195,55 al pedir al mismo modelo que complete únicamente los cuestionarios directos y sumar sus respuestas. Mixtral mejora los resultados principales hasta 18,05 y 163,75. Son errores de regresión sobre puntuaciones de cuestionarios, no medidas de diagnóstico clínico. El trabajo también compara con predicción directa de GPT-3.5, embeddings Ada-002, MentaLLaMA, TF-IDF y la media del conjunto. LMIQ-Mixtral es la mejor variante de la tabla principal, pero no se publican repeticiones, intervalos de confianza ni pruebas estadísticas que permitan cuantificar la estabilidad o significación de las diferencias.

La ablación matiza la tesis de que reunir todos los dominios sea lo más eficaz. Con GPT-3.5, las 135 variables no logran el mejor resultado en ninguna de las dos tareas de prueba: personalidad más preguntas directas obtiene MSE 20,26 en PHQ y 166,62 en PTSD, frente a 20,42 y 192,93 con todas las variables. Terapéutico más directo también mejora PTSD hasta 168,62 y salud mental más directo hasta 177,15. Las variantes sin preguntas directas son claramente peores. Esto apoya que las respuestas simuladas contienen señal predictiva, pero también muestra una fuerte dependencia de preguntas que aproximan directamente los instrumentos objetivo y que añadir todos los dominios puede perjudicar, especialmente en PTSD.

La auditoría del apéndice detecta un problema de fidelidad instrumental. El artículo declara ocho preguntas PHQ-8 más diecisiete PCL-C, pero la lista publicada bajo PHQ contiene nueve ítems porque añade «How likely are you to volunteer your time to help others?», que no pertenece al PHQ-8. Además, el PHQ-8 oficial pregunta por frecuencia durante las dos últimas semanas y puntúa 0–3; el prompt uniforme de LMIQ solicita acuerdo de 1 a 5 y elimina ese anclaje temporal. Por tanto, estas variables no son cumplimentaciones fieles del PHQ-8, sino aproximaciones ordinales generadas por un modelo. El PCL-C sí usa oficialmente 17 ítems de 1 a 5, aunque el artículo tampoco documenta con precisión cómo conserva sus anclajes de tiempo y experiencia. El propio trabajo reconoce que no valida las respuestas personificadas frente a respuestas humanas.

La contribución defendible es un estudio exploratorio que muestra que respuestas estructuradas generadas desde una entrevista pueden resultar más predictivas que varias representaciones textuales en una partición de benchmark. No demuestra diagnóstico automático ni validez clínica. Los objetivos son puntuaciones de autoinforme, no diagnósticos de profesionales; no se informan umbrales, sensibilidad, especificidad, calibración, análisis por subgrupos, incertidumbre ni evaluación prospectiva. Tampoco se publican código, datos derivados o configuración ejecutable, y no se detallan proveedor y versión exacta de Mixtral, snapshots, semillas, coste, fallos o reintentos. El envío de transcripciones sensibles a una API comercial se justifica mediante una mención general a desidentificación y cumplimiento, sin describir consentimiento para este tratamiento, revisión ética, retención, residencia, acuerdo de procesamiento o comprobaciones de reidentificación. En su estado actual, LMIQ debe entenderse como extracción experimental de variables para regresión sobre un benchmark, no como sistema listo para decisiones clínicas.

Research question

Can an LLM, by personifying the person described in a psychological interview and completing structured questionnaires, produce variables that improve the prediction of their PHQ-8 and PCL-C scores compared to textual representations and direct predictions?

Method

LMIQ combines the transcription of each interview with five-item questionnaires per topic and asks the LLM for integer responses from 1 to 5 from the interviewee's perspective. It concatenates 135 responses from four domains and trains a RandomForestRegressor to predict two continuous scores. It evaluates GPT-3.5 Turbo Instruct and Mixtral on the predefined splits of E-DAIC; selects n_estimators among 100/200/300 and max_depth among 10/20/30 on development, and compares MSE on test with direct prediction, sum of direct responses, analysis plus embeddings, direct embeddings, MentaLLaMA, TF-IDF, and mean. It adds a secondary comparison of PHQ on DAIC-WOZ and a domain ablation with GPT-3.5. The audit read and rendered the thirteen pages, verified tables and appendices, contrasted PHQ-8 and PCL-C with official documentation, and searched for public artifacts linked to the work.

Sample: E-DAIC provides 275 semi-clinical interviews conducted by a virtual interviewer and associated PHQ-8 and PCL-C scores. The article uses the predefined splits, but does not present in the main body a demographic, representativeness, power, or subgroup analysis. DAIC-WOZ is used solely as a secondary comparison for PHQ and does not constitute an independent external clinical validation of the DAIC ecosystem.

Findings

  • LMIQ converts each interview into a vector of 135 ordinal responses generated by an LLM.
  • With GPT-3.5, LMIQ obtains test MSE of 20.42 on PHQ and 192.93 on PTSD.
  • With Mixtral, LMIQ improves to MSE 18.05 on PHQ and 163.75 on PTSD.
  • Mixtral is the best variant in the main table on both scores.
  • Direct prediction by GPT-3.5 is worse: 33.42 on PHQ and 336.49 on PTSD.
  • Asking GPT-3.5 to personify and only sum the direct questionnaires reaches 24.16 and 195.55.
  • The gain of GPT-3.5 LMIQ over that direct baseline is small for PTSD.
  • Direct Ada-002 embeddings obtain 38.23 on PHQ and 389.65 on PTSD.
  • On DAIC-WOZ, LMIQ reports MAE 3.46 and RMSE 4.30 on test for PHQ.
  • The DAIC-WOZ comparisons mix different systems and modalities.
  • The full configuration of 135 variables is not the best in any test task of the GPT-3.5 ablation.
  • Personality plus direct questions improves PHQ to 20.26 and PTSD to 166.62.
  • Therapeutic plus direct obtains 168.62 on PTSD and mental health plus direct 177.15.
  • All configurations without direct questions perform worse than the best combinations with direct questions.
  • Sleep variables dominate the published importances for PHQ and panic variables for PTSD.
  • The appendix lists nine questions under PHQ despite declaring PHQ-8, including a question about volunteering that is foreign to the instrument.
  • The targeted search found no public code or implementation associated with the article.

Limitations

  • A single main benchmark of 275 interviews is evaluated.
  • DAIC-WOZ is a related set and not an independent external clinical cohort.
  • The splits are fixed and no replications with other seeds are published.
  • No confidence intervals, deviations across runs, or statistical tests are reported.
  • There is no power analysis or sample size justification.
  • The targets are self-report scores, not clinical diagnoses.
  • The language of diagnostic precision exceeds the variable actually evaluated.
  • Clinical thresholds, sensitivity, specificity, calibration, or decision utility are not evaluated.
  • No predictions with uncertainty or abstention mechanism are provided.
  • There is no comparison with professional assessments or prospective evaluation.
  • Demographic subgroups, languages, cultures, or performance disparities are not examined.
  • Robustness to incomplete transcripts, ASR errors, negation, or contradictions is not reported.
  • Generalization outside DAIC interviews is not tested.
  • The mental health and therapy questionnaires were generated by GPT-4 and manually refined.
  • It is not reported who refined the items, their credentials, agreement, or protocol.
  • The new questionnaires lack clinical and psychometric validation.
  • Reliability, factor structure, convergent validity, or discriminant validity are not evaluated.
  • The uniform prompt transforms all responses into a 1–5 agreement scale.
  • The official PHQ-8 uses 0–3 frequency over the last two weeks.
  • The two-week temporal anchoring of the PHQ-8 is lost.
  • The appendix lists nine items under PHQ instead of eight.
  • The added question about volunteering does not belong to the PHQ-8.
  • The contradiction leaves uncertain which 25 direct variables were actually run.
  • The PCL-C is compatible with 1–5, but its contextual and temporal anchoring is not faithfully documented.
  • The simulated responses are not compared item by item with the actual responses of the interviewee.
  • The article itself acknowledges that it does not validate the personification of the LLM.
  • The direct questions closely approximate the target instruments.
  • The improvement may depend on proxy scoring of symptoms already present in the interview.
  • The ablation shows that the direct domains contribute much of the useful signal.
  • Adding all domains worsens PTSD compared to three smaller combinations.
  • It is not described whether a multi-objective regressor or separate models were trained.
  • Hyperparameter selection uses a single small development set.
  • The configurations finally chosen per task and model are not published.
  • Importance by impurity reduction can be biased with correlated variables.
  • No intervals or stability are provided for the variable importances.
  • Predictive importances do not demonstrate clinical reasoning or causality.
  • GPT-3.5 Turbo Instruct is not identified by an immutable snapshot.
  • The API defaults are described as consistent and unbiased without evidence.
  • Temperature, top-p, execution date, or resolved provider changes are not documented.
  • Mixtral is named ambiguously as 7Bx8 and no provider, checkpoint, or quantization is specified.
  • The MentaLLaMA checkpoint is not reproducibly identified.
  • No seeds for the Random Forest or for any generation are published.
  • Invalid responses, missing responses, retries, or repair rules are not reported.
  • Cost, latency, or number of calls are not reported.
  • The text is ambiguous between one question per session and five grouped questions.
  • The prompt used to initially generate the questionnaires with GPT-4 is not published.
  • The final items are only available in the PDF and not as a versioned artifact.
  • No code, lockfile, environment, executable configuration, or derived data are published.
  • It cannot be verified that tables and ablations come from a reproducible pipeline.
  • Sending sensitive transcripts to the OpenAI API creates an additional trust boundary.
  • The mention of de-identification and compliance does not describe the technical procedure.
  • No re-identification, minimization, retention, or data residency tests are reported.
  • No specific consent or ethical approval for processing with a commercial API is documented.
  • No processing agreement, provider access, or verifiable deletion is described.
  • No human oversight, usage limits, monitoring, or response to clinical harm are proposed.
  • Hallucinations, harmful responses, or attribution failures of the LLM are not evaluated.
  • The limitations section does not cover the PHQ-8 discrepancy or the dependence on direct questions.

What the study does not establish

  • It does not demonstrate that the system diagnoses depression or PTSD.
  • It does not replace an interview or assessment by a mental health professional.
  • It does not clinically validate the questionnaires generated with GPT-4.
  • It does not demonstrate that the LLM faithfully reproduces the interviewee's responses.
  • It does not demonstrate that the variables called PHQ-8 constitute a valid administration of the PHQ-8.
  • It does not establish psychometric validity of the personality or therapy responses.
  • It does not test generalization to real patients outside DAIC, other languages, or other contexts.
  • It does not demonstrate equity across demographic groups.
  • It does not establish statistically significant superiority over each baseline.
  • It does not demonstrate that using the 135 items is better than smaller subsets.
  • It does not separate the value of direct questions from the incremental value of broader domains.
  • It does not turn Random Forest importances into causal clinical explanations.
  • It does not offer calibration, uncertainty, or safety sufficient for clinical decisions.
  • It does not demonstrate effective privacy or health regulatory compliance.
  • It does not allow exact reproduction with public artifacts.

Traceability

Scope: Full text

Version: ACL Anthology 2024.findings-emnlp.23; Findings of ACL: EMNLP 2024; DOI 10.18653/v1/2024.findings-emnlp.23; pp. 403–415; CC BY 4.0

Consulted source: https://aclanthology.org/2024.findings-emnlp.23.pdf

Review: Codex full-text, visual, clinical-instrument, statistical, privacy and reproducibility audit, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • GPT-3.5 Turbo Instruct through the OpenAI API, exact snapshot and resolved defaults not reported
  • Mixtral 7Bx8 as named by the paper, provider, exact checkpoint and serving configuration not reported
  • GPT-4 used to generate initial custom questionnaire topics and items
  • MentaLLaMA 7B baseline, exact checkpoint not reported
  • OpenAI Ada-002 embedding baseline
  • RandomForestRegressor with development-set hyperparameter selection

Instruments and metrics

  • PHQ-8 depression self-report target
  • PCL-C 17-item DSM-IV PTSD self-report target
  • Seventy GPT-4-generated and manually refined mental-health items across fourteen topics
  • Twenty-five Big Five personality items across five traits
  • Fifteen therapeutic-domain items covering family history, trauma history and resilience
  • Twenty-five declared direct items combining PHQ-8 and PCL-C
  • Uniform 1–5 LLM response prompt
  • Mean squared error for E-DAIC
  • Mean absolute error and root mean squared error for DAIC-WOZ comparison
  • Random-forest impurity feature importance

Data used

  • E-DAIC: 275 semi-clinical interview transcripts with PHQ-8 and PCL-C scores and predefined train/development/test partitions
  • DAIC-WOZ predecessor benchmark for secondary PHQ-only comparison
  • LLM-generated 135-dimensional questionnaire-response vectors; not released
  • No public code or implementation linked by ACL, arXiv or the authors in the paper as of the 15 July 2026 audit

Evidence and location

  • Objective, LMIQ, and main claim: Findings of ACL: EMNLP 2024 paper, abstract and section 1, pp. 403–405
  • E-DAIC, 275 interviews, and PHQ/PCL-C targets: Paper, section 3.1 and Figure 1, pp. 405–406
  • Domains, 135 variables, and question generation/refinement: Paper, section 3.2 and Table 1, pp. 405–406
  • Personification prompt and grouping ambiguity: Paper, section 3.3 and Figure 1, p. 405; Appendix A, p. 411
  • Random Forest and hyperparameter search: Paper, section 3.2, pp. 405–406
  • Models and baselines: Paper, sections 4.1–4.2, pp. 406–407
  • Main E-DAIC results: Paper, Table 2 and section 5, p. 407
  • Secondary DAIC-WOZ comparison: Paper, Table 3 and section 5, p. 407
  • Domain ablation: Paper, Table 4 and section 5.1, p. 408
  • Importances for PHQ and PTSD: Paper, Tables 5–6 and section 5.2, p. 408
  • Recognized limitations and declared privacy: Paper, Limitations and Ethical Considerations, p. 409
  • Question list and nine-item discrepancy under PHQ-8: Paper, Appendix B.4, pp. 413–414; visual audit of rendered pages 11–12
  • Official PHQ-8: eight items, two weeks, and 0–3 frequency: CDC National Health Statistics Reports No. 172; checked 15 Jul 2026
  • Official PCL-C: 17 items and 1–5 scale: U.S. Department of Veterans Affairs National Center for PTSD, PCL handout DSM-IV; checked 15 Jul 2026
  • Metadata, DOI, pages, and license: ACL Anthology 2024.findings-emnlp.23; DOI 10.18653/v1/2024.findings-emnlp.23; CC BY 4.0
  • Absence of linked public implementation: ACL, arXiv, paper links, Papers with Code and targeted GitHub/web search; checked 15 Jul 2026