Quality evaluation of digital voice assistants for diabetes management

Background Digital voice assistants (DVAs) are increasingly used to search for health information. However, the quality of information provided by DVAs is not consistent across health conditions. From our knowledge, there have been no studies that evaluated the quality of DVAs in response to diabetes-related queries. The objective of this study was to evaluate the quality of DVAs in relation to queries on diabetes management. Materials and methods Seventy-four questions were posed to smartphone (Apple Siri, Google Assistant, Samsung Bixby) and non-smartphone DVAs (Amazon Alexa, Sulli the Diabetes Guru, Google Nest Mini, Microsoft Cortana), and their responses were compared to that of Internet Google Search. Questions were categorized under diagnosis, screening, management, treatment and complications of diabetes, and the impacts of COVID-19 on diabetes. The DVAs were evaluated on their technical ability, user-friendliness, reliability, comprehensiveness and accuracy of their responses. Data was analyzed using the Kruskal-Wallis and Wilcoxon rank-sum tests. Intraclass correlation coefficient was used to report inter-rater reliability. Results Google Assistant (n = 69/74, 93.2%), Siri and Nest Mini (n = 64/74, 86.5% each) had the highest proportions of successful and relevant responses, in contrast to Cortana (n = 23/74, 31.1%) and Sulli (n = 10/74, 13.5%), which had the lowest successful and relevant responses. Median total scores of the smartphone DVAs (Bixby 75.3%, Google Assistant 73.3%, Siri 72.0%) were comparable to that of Google Search (70.0%, p = 0.034), while median total scores of non-smartphone DVAs (Nest Mini 56.9%, Alexa 52.9%, Cortana 52.5% and Sulli the Diabetes Guru 48.6%) were significantly lower (p < 0.001). Non-smartphone DVAs had much lower median comprehensiveness (16.7% versus 100.0%, p < 0.001) and reliability scores (30.8% versus 61.5%, p < 0.001) compared to Google Search. Conclusions Google Assistant, Siri and Bixby were the best-performing DVAs for answering diabetes-related queries. However, the lack of successful and relevant responses by Bixby may frustrate users, especially if they have COVID-19 related queries. All DVAs scored highly for user-friendliness, but can be improved in terms of accuracy, comprehensiveness and reliability. DVA designers are encouraged to consider features related to accuracy, comprehensiveness, reliability and user-friendliness when developing their products, so as to enhance the quality of DVAs for medical purposes, such as diabetes management.