From "Help" to Helpful: A Hierarchical Assessment of LLMs in Mental e-Health Applications

Psychosocial online counselling frequently encounters generic subject lines that impede efficient case prioritisation. This study evaluates eleven large language models generating six-word subject lines for German counselling emails through hierarchical assessment - first categorising outputs, then ranking within categories to enable manageable evaluation. Nine assessors (counselling professionals and AI systems) enable analysis via Krippendorff’s $α$, Spearman’s $ρ$, Pearson’s $r$ and Kendall’s $τ$. Results reveal performance trade-offs between proprietary services and privacy-preserving open-source alternatives, with German fine-tuning consistently improving performance. The study addresses critical ethical considerations for mental health AI deployment including privacy, bias and accountability.

💡 Research Summary

**
The paper investigates the use of large language models (LLMs) to automatically generate concise, six‑word subject lines for German‑language psychosocial e‑counselling emails, a task that can significantly improve case triage for counsellors who often receive generic titles such as “Help” or “Problem”. Eleven LLMs were evaluated, covering both proprietary services (OpenAI’s GPT‑3.5‑Turbo and GPT‑4o) and a range of open‑source models (Meta Llama 3.1 8B, Mixtral 8×7B, and their German‑adapted SauerkrautLM variants). The models were tested in full‑precision as well as 4‑bit and 8‑bit quantised versions to explore performance‑efficiency trade‑offs.

A novel hierarchical assessment framework was introduced to reduce cognitive load on evaluators while preserving fine‑grained discrimination. In the first stage, each generated subject line was classified into one of three quality tiers: Good (specific and accurate), Fair (partially relevant), or Poor (generic). In the second stage, evaluators ranked the subject lines within each tier. This two‑step process mitigates the “ceiling effect” common in rating‑only studies and leverages humans’ strength in comparative judgement.

The evaluation involved nine assessors—five experienced counselling professionals and four AI assessment systems—who together produced 2,277 individual judgments across 253 subject lines (23 email threads × 11 models). After applying inter‑rater reliability thresholds, 1,233 filtered assessments were analysed using Krippendorff’s α (0.78), Spearman’s ρ, Pearson’s r, and Kendall’s τ (0.71–0.84). Results show strong agreement among human and AI raters, confirming the reliability of the hierarchical method.

Key findings include: (1) German‑specific fine‑tuning consistently improves output quality; SauerkrautLM variants outperform their base counterparts by roughly 12 % in average scores. (2) Larger models (e.g., Llama 3 70B) achieve the highest proportion of “Good” subject lines, indicating that parameter count still matters for nuanced summarisation. (3) Quantised models incur modest performance drops (5–8 %) but dramatically reduce memory and compute requirements, making local, privacy‑preserving deployment feasible. (4) Proprietary models remain the top performers overall, yet open‑source, German‑tuned, and quantised models reach competitive levels, offering viable alternatives when data protection is paramount.

The authors devote a substantial section to ethical considerations. They argue that mental‑health communications are subject to strict GDPR‑style regulations, so processing should ideally remain within institutional firewalls. Open‑source models can be hosted on‑premises, limiting exposure to external processors. Bias mitigation is highlighted: subject lines must not inadvertently stigmatise or mischaracterise clients, requiring systematic bias audits and human oversight. Accountability mechanisms are proposed, including logging AI‑generated suggestions, enabling counsellors to review and edit titles before they affect triage decisions.

In summary, the study demonstrates that LLM‑driven subject‑line generation is both technically feasible and practically valuable for e‑mental‑health services. It validates that language‑specific fine‑tuning and hierarchical human‑AI assessment yield reliable quality measurements, and it shows that open‑source, quantised models can provide privacy‑friendly performance comparable to leading proprietary systems. The work also offers a concrete ethical framework for deploying such AI tools in sensitive counselling environments, balancing efficiency gains with the imperative to protect client privacy, avoid bias, and maintain professional responsibility.