Lost in translation? Evaluating associations between English proficiency and next generation sequencing completion, and comparing understandability across different translation modalities among Spanish speakers with metastatic breast cancer.

Abstract

1524 Background: Disparities in next-generation sequencing (NGS) utilization persist among patients with metastatic breast cancer (MBC), disproportionately affecting Hispanic/Latino populations. We hypothesized that the challenges in explaining NGS, particularly when mediated through translation, could be contributing to this observed discrepancy in NGS testing rates within the Hispanic/Latino populations. We evaluated the associations between limited English proficiency and NGS testing by comparing the understandability between English, certified Spanish translations, and artificial intelligence (AI)-generated Spanish translations of provider explanations of NGS testing. Methods: We analyzed 191 patients with recurrent MBC from the Dallas Metastatic Breast Cancer Study. Ethnicity was categorized as Hispanic vs non-Hispanic. NGS completion between 2014–2022 was assessed. Three versions of NGS explanations (original English, certified Spanish translation, and AI-generated Spanish translation) were evaluated using the Flesch-Kincaid Readability Score (English) and the Fernández-Huerta Index (Spanish) using 10 provider-patient interactions. Scores range from 0–100, with higher scores indicating greater understandability; CDC/NIH-recommended patient materials target scores of 60–70. Readability discordance < 10 points was considered acceptable. Results: 81.7% (156) of patients were proficient in English and were 3.2 times more likely to complete NGS testing than Spanish speakers (95% CI 1.35-7.73, p < 0.01). Mean readability score between the English, certified Spanish, and AI Spanish were 47.9, 51.4, and 54.1, respectively. The mean discordance was 6.1 points between English and certified Spanish, 5.9 points between certified Spanish and AI Spanish, and 8.6 points between English and AI Spanish. All ten original English explanations fell below CDC/NIH readability targets, and only 4 of 10 certified Spanish translations met the recommended standards. When refining the prompts to maximize patient understanding, AI-generated text generated perfect understandability scores. Conclusions: Although translation fidelity was preserved, most provider-generated explanations failed to meet recommended readability standards, suggesting that communication quality, not translation alone, may limit patient understanding. AI-generated explanations may serve as a tool to mitigate language-related barriers when certified interpreters or optimized explanations are unavailable and enhance informed decision-making across varied clinical discussions, not limited to NGS testing.