Neural Conversational Agents for Multilingual Medical Communication Translation

Thesis Abstract

The study addresses a critical gap in multilingual medical communication by examining the efficacy and reliability of neural conversational agents (NCAs) in translating patient–clinician interactions across languages, with a focus on preserving clinical meaning, patient safety, and user trust. The aim is to develop and evaluate an NCA-driven translation framework that supports real-time multilingual medical consultations, ensuring semantic fidelity and cultural competence. Specific objectives include (1) assessing the translation quality of NCAs for medical discourse across English, Spanish, Mandarin, and Arabic; (2) evaluating user trust, perceived empathy, and satisfaction in NCA-mediated consultations; (3) analyzing the impact of domain-specific fine-tuning on translation accuracy for symptom description, medication instructions, and consent conversations; (4) identifying failure modes and risk factors for miscommunication and proposing mitigation strategies; and (5) formulating best-practice guidelines for clinical deployment and ethical governance. The methodology adopts a mixed-methods design comprising three phases. In Phase 1, a controlled laboratory evaluation will recruit 120 volunteer participants representing four language groups (English, Spanish, Mandarin, Arabic) alongside 25 professional clinicians. Data collection instruments include standardized medical vignettes, scripted and unscripted patient–clinician dialogues, and a bilingual evaluation rubric capturing semantic equivalence, pragmatics, and safety constraints. Phase 2 entails field testing in a tertiary care hospital with 400 consenting patients across the four languages, where NCAs are integrated into triage and consent workflows. Instrumentation includes the Multilingual Medical Communication Quality Instrument (MMCQI), the Trust in Automation scale, and the Patient Comprehension Assessment (PCA). Phase 3 uses post hoc diagnostic analytics and a qualitative component for error analysis. The analysis plan employs both quantitative and qualitative techniques. For translation quality, BLEU, METEOR, and biomedical-specific TER metrics will be computed, complemented by clinical semantic adequacy assessment conducted by bilingual medical professionals using a calibrated 5-point Likert rubric. Hypothesis testing will apply repeated-measures ANOVA to compare translation quality across language pairs and clinical domains, and multivariate regression to examine relationships among translation fidelity, clinician satisfaction, and patient comprehension. In-depth interviews and thematic analysis will explore perceived empathy, trust, and cultural appropriateness, guided by the Cultural-Loci Theory of Medical Communication and Communicative Action Theory to interpret normative expectations around care and consent. A confounding-control model will adjust for participant health literacy and prior exposure to technology. Model specification includes an error-aware neural translator with domain-adaptation layers, augmented by ESCALATE checks to flag potential safety risks in high-stakes terms (e.g., dosage, allergy, consent). Expected findings anticipate that domain-tuned NCAs will achieve meaning-preserving translations with clinically relevant terminology across all languages, but performance gaps may emerge in nuanced symptom descriptions or culturally bound expressions of pain and risk. It is anticipated that standardized post-editing by clinicians will improve safety-critical utterances, while user trust will correlate positively with perceived transparency about machine limitations and the availability of human-in-the-loop escalation. The study will identify specific language pairs with higher error rates and delineate categories of miscommunication ripe for remediation, such as negation handling, polysemy in medical terms, and discourse structure differences in procedural instructions. The study contributes to knowledge by empirical evaluation of multilingual NCAs in real-world medical settings, advancing understanding of translation quality, safety, and user–machine interaction in high-stakes communication. It proposes an integrated framework for clinical deployment that combines domain-adaptive NCA architectures, post-editing protocols, and governance mechanisms informed by bioethics and data-privacy principles. Findings are expected to inform standard-setting for multilingual patient care, influence policy on AI-assisted medical communication, and guide future research on multilingual NLP applications in healthcare. The main conclusion is that while domain-specific fine-tuning and human-in-the-loop strategies substantially enhance reliability and safety, robust multilingual medical communication requires transparent disclosure of limitations, ongoing clinician oversight, and culturally sensitive design to ensure equitable access and patient safety across language communities. Recommendations include implementing standardized bilingual evaluation benchmarks, developing clinician-facing tooling for rapid correction of NCA outputs, and establishing regulatory guidelines for accountability and patient consent when AI-mediated translation is involved.