Feasibility of Real-time Artificial Intelligence–based Language Translation for Bilingual Informed Consent in Interventional Radiology:An Australian Proof-of-concept Authors: Goran Mitreski, Philip Chan, Cephas Tamburayi Kamba, Nathan Ng, David Duong, Rose Thomas, Jyothirmayi Velaga, Miranda Siemienowicz, Hong Kuan Kok Published: 2026 ABSTRACT Purpose: Language barriers in healthcare lead to miscommunication, reduced comprehension, and adverse outcomes. Professional medical interpreters are the gold standard, but are often unavailable for interventional radiology consent discussions. Recent advances in large language models such as GPT-4o have shown high accuracy in multilingual tasks. We conducted a proof-of-concept internal validation of GPT-4o as a live bilingual translation tool in interventional radiology consent simulations across nine languages.Material and Methods: Thirty-six simulated consent sessions were conducted across nine languages - Shona, Macedonian, Polish, Tamil, Cantonese, Mandarin, Malay (Bahasa Malaysia), Afrikaans, and Vietnamese, with four procedural scenarios evaluated per language (Transjugular Intrahepatic Portosystemic Shunt, Yttrium-90 Radioembolization, liver biopsy, large-bore mechanical thrombectomy) using GPT-4o to perform real-time translation between English and native language speaking Interventional Radiologists/Diagnostic Radiologists. A 13-domain rubric of translation adequacy (each scored 1-5) produced a maximum possible total of 65 points. Sessions scoring ≥50/65 were considered acceptable.Results: Across 36 sessions, mean scores were 56.2 ± 6.3 (86.5% of maximum). Acceptability (≥50) was achieved in 31/36 sessions (86.1%). Seven languages (Mandarin, Macedonian, Polish, Tamil, Malay, Afrikaans, and Vietnamese) consistently exceeded the threshold, while Cantonese (mean ~40/65) and Shona (mean ~42/65) lagged. Translation speed (time to output) averaged 4.8/5 across all languages.Conclusions: GPT-4o achieved high translation quality across multiple interventional radiology consent simulations, demonstrating the feasibility of real-time artificial intelligence interpretation for many languages. However, performance was language-dependent; certain languages (Cantonese and Shona) remained sub-optimal and merit human-interpreter supervision. These findings represent a platform-specific proof-of-concept evaluation rather than validation of artificial intelligence-based medical translation more broadly. Introduction Effective communication is critical to obtaining informed consent for interventional radiology (IR) procedures. Patients with limited proficiency in the dominant language face an increased risk of misunderstanding, adverse events, and reduced satisfaction with healthcare. Interpreter services are critical to the functioning and flow of a major hospital; this resource can be scarce, and options for unplanned interpreter services are limited. Systemic barriers to interpreter services exist, including inflexible booking processes, long waiting times for phone interpreters, and low availability for in-person interpreters [1]. Additionally, human translators are costly for institutions, and remote services rely on adequate telecommunication access for both parties across multiple locations. In Victoria, Australia, over 30% of the population is born overseas, with over 27% of people speaking a language other than English as their first language [2]. Machine translation has evolved rapidly, with earlier commercial translators (such as Google and Microsoft translators) achieving 80%-90% accuracy for general phrases. In healthcare settings, accuracy tends to be higher for English-to-other-language translations (83%-98%) than the reverse (36%-76%) [3]. The GPT-4 (OpenAI) large language model has demonstrated improved performance in medical and multilingual tasks, including medical licensing examination questions and generating readable procedure explanations. More recent models have increased output accuracy, leading to higher performance on the United States medical licensing examination (USMLE) and radiology board examinations, increasing scores from 61% with ChatGPT-3 to 81% with ChatGPT-4 [4, 5]. GPT-4o has further been evaluated for multilingual translation of radiology reports across imaging modalities, demonstrating robust performance in translating reports across multiple languages [6, 7]. The study aimed to evaluate the feasibility of GPT-4o for real-time bilingual translation during the informed consent process in IR clinical practice. We hypothesized that GPT-4o would deliver translations achieving a defined quality threshold in select languages that are common in our local institution, and we further aimed to assess low-resource or dialect-heavy languages. Low-resource languages such as Shona have a scarcity of labeled and unlabeled language data and poor-quality data that is not sufficiently representative of the language and its sociocultural context [8]. Heavy-dialect, tone-rich languages such as Cantonese, known for its complex phonology, idiomatic slang, and limited shared intelligibility with Mandarin, although represented by a large population volume, remain low-resourced languages in the large language model (LLM) space [9, 10]. Material and Methods An internal simulation-based validation study was performed. Thirty-six consenting simulations were conducted in person in an IR practice environment, each involving an English-speaking Interventional Radiologist or Diagnostic Radiologist who was medically bilingual in both English and another language (one of the nine tested). Translations were delivered via ChatGPT (GPT-4o, March update, 2025 version) in real time using a commercially available smartphone (iPhone 16, Apple, Cupertino, CA, USA) with internet access. Four IR procedures were selected for assessment: (1) Transjugular Intrahepatic Portosystemic Shunt (TIPS); (2) Yttrium-90 Radioembolization (Y-90); (3) Ultrasound-guided liver biopsy; and (4) Large-bore mechanical thrombectomy (LBMT). Translation for each procedure was simulated once in each language (9 × 4 = 36 sessions). The order of translation for each procedure was kept consistent. This study was designed as an exploratory feasibility evaluation rather than a definitive validation study. The number of sessions was therefore limited to representative simulations across multiple languages and procedures to establish proof-of-concept feasibility prior to larger-scale investigations. The radiologists were fluent in English and familiar with the procedural content. They acted as both physician and patient, switching between English and their native language, in which they had at least professional working fluency, including medical-level grammar, speech, and writing. All participants provided consent for their time and involvement. As the study was stimulatory and involved no patients, formal ethics approval was not required. The radiologist spoke in English and their other language whilst having the voice module on GPT-4o active. The translation function translated English into the target language and the target language into English during each dialogue point. A dialogue was created for each procedure. The physician would read the English (doctor script) dialogue exactly as written. The patient script, which contained simulated questions and comments about the procedure, was spoken in the target language based on a written English prompt (See Appendix 1). The consent scripts were intentionally standardized and simplified to enable consistent cross-language comparison of translation accuracy. These scripts represent baseline procedural explanations rather than the full depth of real-world consent discussions, which typically involve iterative dialogue and individualized risk explanation. Nine languages were evaluated (Mandarin, Cantonese, Macedonian, Polish, Tamil, Malay, Afrikaans, Vietnamese, and Shona). Each language underwent four simulated consent sessions corresponding to four IR procedures, resulting in 36 total evaluated sessions. These languages were chosen due to their complexity in comprehension, grammar, and tonality. Additionally, many non-English–speaking patients within our health network speak one of these as their primary language. Each simulation was also carried out after ensuring adequate network coverage and internet connectivity to ensure seamless communication with ChatGPT. A 13-domain rubric was created, with each domain scored 1 (poor) to 5 (excellent), yielding a maximum score of 65 points (See Appendix 2). 1. Fluency 2. Grammar 3. Accuracy 4. Lay language 5. Risk Clarity 6. Comprehension 7. Response to Questions 8. Bidirectional accuracy 9. Cultural appropriateness 10. Phrasing 11. Back-translation alignment 12. Comprehensibility 13. Time to output (speed) The 13-domain rubric was developed based on established frameworks used in medical communication and translation assessment, including linguistic accuracy, completeness of risk explanation, preservation of clinical intent, and conversational flow. Each domain was scored on a 0-5 scale, producing a maximum score of 65. While an ROC-AUC of approximately 0.8 indicates good discriminatory ability [11], ROC curves do not specify an optimal operational cut-off for multidomain translation accuracy. Contemporary language models, including GPT (OpenAI) and Gemini (Google/DeepMind), define translation and safety acceptability using human-expert consensus and performance-band classification, as multidomain linguistic accuracy does not map to binary labels suitable for ROC/AUC analysis [12, 13]. A threshold of ≥50/65 (approximately 75% adequacy) was predefined as representing clinically acceptable translation quality based on pilot testing and consensus among the evaluating radiologists. This threshold was selected to approximate the minimum level at which translated explanations preserved core clinical meaning, safety messaging, and procedural risks. After each session, the physician would mark the procedure against the scoring rubric. A follow-up binary question was asked of the assessors: “Would you consider this translation safe and valid for real-world use?” Free-text comment space was also provided for any peculiarities or concerns identified during assessment. Statistical analysis was undertaken for qualitative and descriptive variables. The mean and standard deviation of the total scores were calculated overall, by language, and by procedure. Further descriptive analyses were performed of the domains to identify outliers. Qualitative comments were reviewed for error patterns. Heatmaps were generated in Microsoft Excel using the raw rubric scores for each language and evaluation domain. No normalization or statistical transformation was applied. Conditional formatting with a color gradient was used, where darker colors (green) represent higher rubric scores and lighter colors (red) represent lower scores. This approach allowed visual comparison of translation performance across languages and evaluation domains. Given the exploratory proof-of-concept design and small sample size, visual descriptive analysis using heatmaps was chosen rather than formal comparative statistical testing. Results Overall performance Across 36 sessions, the mean total score was 56.2 ± 6.3 out of 65 (86.5%). The threshold for satisfactory performance (≥50) was reached in 31 of 36 sessions (86.1%). Translation speed (time to output domain) scored 4.8 ± 0.3 out of 5 (96%) in all languages. A heatmap of the mean scores (Figure 1) demonstrates the averages for each language across all domains and highlights those that underperformed. Heatmap of the mean scores for each language in each simulation of procedural consent. TIPS: Transjugular intrahepatic portosystemic shunt, Y90: Radioembolization using Ytrrium90 Liver Bx: liver biopsy; LBMT: large-bore mechanical thrombectomy Language-specific performance Seven out of nine (77%) languages exceeded the threshold in all four procedure sessions: Mandarin, Macedonian, Polish, Tamil, Malay, Afrikaans, and Vietnamese. In contrast, Cantonese averaged ~41/65 and Shona ~43/65, with multiple sessions falling below the satisfactory performance threshold. Cantonese sessions evidenced frequent switching to Mandarin and non-colloquial phrasing; Shona sessions had mildly awkward grammar and lay-language phrasing. For a minority of languages assessed, comments were noted that the tone and accent of the language model (British voice; Arbor; GPT-4o) could be understood, but were recognizably non-native when providing non-English verbal output. Procedure-specific performance Mean scores by procedure were: TIPS 57.0 ± 5.8, Y-90 55.5 ± 6.8, liver biopsy 56.4 ± 5.2, LBMT 55.0 ± 6.5. Acceptability rates (≥50) by procedure ranged from 83% (Y-90) to 89% (TIPS). Assessors identified that the language model improved across the four procedures, with LBMT having the most ‘natural’ and ‘evolved’ language translation compared to TIPS, which was assessed first by each participant. Domain-specific findings The highest mean domain ratings were time to output (~4.8/5), Accuracy (~4.7/5), and Fluency (~4.5/5). Lower domain means included Cultural Appropriateness (~4.1/5) and Phrasing (~4.2/5), suggesting minor stylistic rather than clinically relevant deficiencies. A radar plot summarizing the mean scores across the 13 evaluation domains is shown in Figure 2. A radar plot summarizing the mean scores across the 13 evaluation domains. Qualitative observation Evaluator comments highlighted excellent medical-term reliability in supported languages, seamless back-and-forth dialogue, and minimal delay. For Cantonese, the evaluator flagged the use of Mandarin vocabulary, despite explicit instruction to maintain Cantonese translation and non-native cadence. For Shona, whilst content was accurate, phrasing sometimes felt mechanical and less conversational. Polish translations preserved overall meaning but showed consistent limitations, including imperfect accent and enunciation, occasional mistranslations for culturally specific phrases, inconsistent use of formal address, and repeated misgendering of the clinician. Discussion This study evaluated a single artificial intelligence (AI) platform (GPT-4o, March 2025 version), and the findings should therefore not be generalized to AI-based medical translation systems broadly. Performance may vary across different models, architectures, and training datasets. GPT-4o translation can achieve high-quality bidirectional bilingual consent across multiple languages and procedures. The mean score of ~56/65 and acceptability rate of 86% suggest that language learning models and voice chat AI programs may approach the human-interpreter standard in many languages. Importantly, performance was consistent across four distinct procedural scripts, indicating the model’s robustness to varying content and complexity. Our findings challenge previous reports on language learning models describing clinician hesitancy around quality, reliability, and comprehensibility [3]. This internal validation study exceeded previous accuracy rates for earlier communication tools (80%-90% in general translation) and suggests an avenue to move forward where interpreter services are either unavailable or logistically challenging. For procedural consent, where technical terms and risk-benefit discussions are complex, GPT-4o handled both content and conversational context proficiently in supported languages. However, our data underscores a key caveat: translation quality is strongly language-dependent. While high-resource and well-supported languages (Mandarin, Tamil, Afrikaans, etc.) performed very well, low-resource languages and some dialects (e.g., Cantonese and Shona) performed below threshold. This likely relates to a degree of limited model training data, particularly for Shona, although this was a surprising finding for Cantonese, which is widely spoken. Code-switching or the use of non-native phrasing in Cantonese may undermine patient comprehension and trust. While the Cantonese dialogue was contaminated with the tones and pronunciations of Mandarin, this may still be comprehensible for a Cantonese-speaking patient. Therefore, AI translation should not be assumed to be safe across all languages without prior validation. From a clinical workflow perspective, the speed and convenience of AI translation are major advantages. Every simulation session demonstrated near instant translation (~4.8/5). In urgent clinical settings where an interpreter is not promptly available, GPT-4o could serve as a valuable time-saving adjunct. Additionally, it may help promote equity in access to informed consent for patients who speak languages for which interpreter services are scarce. Logistically, it is important to understand the environment in which consent takes place. The AI translation app requires relative silence during translation; otherwise, it staggers its cadence and translation, picking up adjacent conversation. Therefore, any clinical use would have to factor in an appropriate low-noise environment for accurate translation. Nevertheless, acceptable clinical, ethical, and medicolegal oversight remains essential. This was a simulation with prepared scripts to ensure key content was discussed. The simplified consent scripts used in this study do not fully capture the complexity of real-world consent conversations, which frequently include unscripted patient questions, emotional responses, and individualized risk discussion. In actual practice, conversations may be more free-form, and the AI may face unexpected questions or need to accommodate patients speaking in slang and colloquial terms that may not be fully captured in our simulation. We did not include unscripted patient questions, although the ‘patient’ assessment of the working dialogue required some degree of liberty from the assessor to convey the message in the native tongue as accurately as possible. AI-based translation tools should therefore be considered adjunctive communication aids rather than replacements for clinician-patient interaction. Human verification of AI-generated content remains essential to ensure the accuracy and completeness of information. Limitations As a simulation study with 36 sessions, our findings may not fully generalize to patients experiencing stress or emotional distress, or to all clinical scenarios. A key limitation is the use of bilingual radiologists acting as both physician and patient participants. While this design allowed consistent linguistic evaluation across languages, it may introduce bias because medically trained participants possess higher baseline knowledge of IR procedures and medical terminology than typical patients. This could artificially increase perceived translation adequacy and comprehension scores. Given that the assessor was both patient and clinician, we could not assess true compression outcomes (e.g., teach-back) or compare directly to professional human interpreters. Inter-rater reliability could not be assessed, and our bilingual participants may have been more forgiving of AI translation than non-English–speaking patients. The inclusion of diagnostic radiologists may reduce procedural realism compared with interventional radiologists who routinely conduct consent discussions. The 13-domain rubric, whilst comprehensive, is novel and has not been validated externally. Future studies should incorporate trained simulated patients or prospective patient cohorts without a medical background to better evaluate real-world consent comprehension and acceptability. Finally, the specific GPT-4o version used may evolve over time as updates are rolled out, which may alter its translation performance. Future direction Further methodological work is required before prospective clinical trials can be considered. Future research should first involve larger simulation studies incorporating non-medical simulated patients and comparative evaluation across multiple AI translation platforms. Such work would help determine whether the performance observed in this study is platform-specific or generalizable across AI systems. When appropriate, prospective clinical trials should evaluate actual patient-doctor dialogues using AI translation, measuring comprehension, satisfaction, time to consent, and error rates compared to gold standard human interpreters. Expanding language coverage, especially to truly low-resource and rarer languages, is also critical. Furthermore, training or fine-tuning language models via forced feedback to specific dialects (e.g., Cantonese) or domains in IR may further improve performance. Nonetheless, this proof of concept has yielded promising results and suggests that real-time AI translation in a clinical setting could be a valuable adjunct to clinical environments where timely access to human interpreters is challenged. Conclusion In summary, GPT-4o demonstrated strong feasibility for real-time bidirectional translation in simulated IR consent discussions, across several high-resource languages. However, performance was language-dependent, with reduced reliability in dialect-heavy and low-resource languages such as Cantonese and Shona. These findings support AI translation as a workflow adjunct rather than a replacement for human interpreters and represent a platform-specific proof of concept requiring further validation prior to broader clinical implementation. Author Contributions: Contributed to the manuscript in methodology and data collection, write-up, and revision: all authors. Conceptualization, resources, data curation, writing, and editing: lead author. Conflicts of Interest: There are no conflicts of interest. Supplementary Material