From Evaluation to Enhancement: A Decision Support Framework for Quality Assurance in Therapeutic AI Systems (Preprint)
Bibliographic record
Abstract
Abstract Background Therapeutic chatbots are increasingly deployed across digital mental health services, yet most evaluation efforts remain diagnostic rather than actionable. Organizations lack structured pathways to translate evaluation findings into validated quality improvements aligned with health care quality assurance requirements. Objective This study aims to introduce EvaluationPlus, a decision support framework that operationalizes a reproducible evaluation-to-enhancement loop for therapeutic artificial intelligence systems. We aimed to demonstrate its feasibility through expert-guided diagnosis, multi-large language model (LLM) enhancement mapping, and within-subject validation. Methods Using the bilingual mental health chatbot Dr. CareSam (GPT 4.0–based), we conducted 3 iterative enhancement cycles. Two licensed clinical psychologists performed structured diagnostic reviews using think-aloud protocols to identify competency-specific deficits across a 7-dimension therapeutic competency rubric. Three LLMs (GPT 4.0, Claude 4.0 Sonnet, Gemini 2.5 Flash) generated prescriptive enhancement strategies aligned with identified gaps. A participant-blinded, within-subject A/B validation study with Korean graduate students from the Department of Applied Artificial Intelligence, Sungkyunkwan University (N=15; 16 recruited, 1 excluded; IRB-approved) compared baseline and enhanced versions across standardized clinical scenarios spanning mild anxiety, to crisis-level presentations. Results The enhanced system demonstrated substantial improvement in overall therapeutic quality, with mean scores increasing from 5.40 to 7.63 (Δ =+2.23 points, 41%; dz=0.881; 95% bootstrap CI [0.32-2.20]). Prespecified target dimensions — active listening and appropriate questions, personalization, and complex thinking — showed large-effect improvements (mean gain +3.04; dz range 0.96‐1.08), significantly exceeding gains in nontargeted dimensions (+1.62; targeting differential +1.42 points). Directional improvement was observed in 13 of 15 participants (86.7%). User preference strongly favored the enhanced system (13/15, 86.7%), and expert clinical evaluation confirmed maintained safety and therapeutic appropriateness across four scenario severity levels (preference rate 75%; 3 of 4 scenarios). Cross-participant rating consistency improved substantially (coefficient of variation: 20% → 8.1%). Conclusions EvaluationPlus demonstrates feasibility as a structured framework for iterative quality assurance of therapeutic artificial intelligence systems. By linking expert diagnostic procedures with prescriptive multi-LLM enhancement mapping and multistakeholder validation, the framework supports reproducible improvement cycles relevant to organizational oversight of digital mental health tools. Limitations include a small pilot sample, single-culture focus, and simulated crisis scenarios; future work should extend validation to diverse clinical populations and longitudinal outcome assessment.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.162 | 0.164 |
| Meta-epidemiology (narrow) | 0.002 | 0.001 |
| Meta-epidemiology (broad) | 0.001 | 0.003 |
| Bibliometrics | 0.005 | 0.003 |
| Science and technology studies | 0.003 | 0.012 |
| Scholarly communication | 0.015 | 0.013 |
| Open science | 0.004 | 0.010 |
| Research integrity | 0.004 | 0.004 |
| Insufficient payload (model declined to judge) | 0.007 | 0.001 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".