Layperson-Friendly AI Translation of Medical Documents to Improve Doctor–Patient Communication: Protocols for the AI-INFOCARE and AI-MEDTALK Randomized Controlled Trials
Bibliographic record
Abstract
BACKGROUND: Many patients struggle to understand referral letters and discharge summaries; low health literacy is prevalent, and short consultations limit explanations. Large language models (LLMs) can translate clinical jargon into layperson language, but their impact on doctor-patient communication in real care remains untested. OBJECTIVE: This study aims to determine whether providing an artificial intelligence (AI)-generated, medically validated layperson-language translation of key medical documents before a consultation improves the quality of doctor-patient communication. METHODS: We plan to conduct two single-center, parallel-group randomized controlled trials in neurosurgery: AI-INFOCARE (outpatients prior to treatment discussions) and AI-MEDTALK (inpatients at surgical discharge). Adults (aged 18 years or older, German-speaking, and able to consent) are randomized 1:1 to receive either (1) an AI-generated layperson-friendly summary of their referral or discharge document in addition to usual care or (2) usual care only. Summaries are produced with a Claude-based system (Simply Onno GmbH) and undergo human verification against the source document using a predefined checklist. The primary outcome is patient-rated interaction quality (Fragebogen zur Arzt-Patient-Interaktion/Questionnaire on the Quality of Physician-Patient Interaction) immediately postconsultation. Secondary outcomes are perceived autonomy support (brief Health Care Climate Questionnaire), self-rated understanding (5-point item), physician-rated encounter difficulty (Difficult Doctor-Patient Relationship Questionnaire, 10-item), and consultation length (minutes). Randomization uses a computer-generated allocation list with concealed, sequential assignment at the point of inclusion; recruiting clinicians have no access to the sequence, and analysts are blinded to group codes. Fidelity is assured by standard operating procedures, rater training, and periodic dual review with interrater agreement estimates. Safety monitoring defines information-related adverse events (eg, distress requiring unplanned clinical support or clinically relevant inaccuracies) and includes an independent safety overseer. Ethics approvals were obtained in February 2025 from the ethics committee of Technische Universität Dresden, and both trials are registered in the German Clinical Trials Register. RESULTS: Each trial targets 300 participants (150 per arm), providing greater than 80% power for an effect size d≈0.4 on the primary outcome (α=0.05). Outcomes are assessed immediately after the index consultation, with no additional follow-up planned in this protocol. We hypothesize that providing layperson-friendly summaries will significantly improve patients' understanding and satisfaction with information, foster a more autonomy-supportive communication climate, and reduce physicians' perceived difficulty in the encounter without unduly prolonging consultation time. Neither study received external funding. Both trials are currently in the recruitment phase, with patient enrollment scheduled to begin in May 2025 and expected to conclude by February 2026. Results are anticipated to be published in summer 2026. CONCLUSIONS: These pragmatic randomized controlled trials test a scalable AI intervention to strengthen understanding and interaction quality without adding clinician burden. If effective, AI-assisted layperson summaries could be integrated into routine workflows to advance health literacy and patient-centered care. TRIAL REGISTRATION: Deutsches Register Klinischer Studien DRKS00036810; https://www.drks.de/search/de/trial/DRKS00036810 and Deutsches Register Klinischer Studien DRKS00036814; https://drks.de/search/de/trial/DRKS00036814. INTERNATIONAL REGISTERED REPORT IDENTIFIER (IRRID): PRR1-10.2196/77204.
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 | 0.055 | 0.094 |
| Meta-epidemiology (narrow) | 0.005 | 0.002 |
| Meta-epidemiology (broad) | 0.008 | 0.008 |
| Bibliometrics | 0.003 | 0.003 |
| Science and technology studies | 0.002 | 0.003 |
| Scholarly communication | 0.004 | 0.004 |
| Open science | 0.003 | 0.002 |
| Research integrity | 0.007 | 0.006 |
| Insufficient payload (model declined to judge) | 0.034 | 0.005 |
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".