BACKGROUND
The demand for medical advice by mobile artificial intelligence voice assistants (AI-VAs) during the ongoing COVID-19 pandemic is increasing, especially for the diagnosis and screening of seemingly insignificant degenerative diseases such as early knee osteoarthritis (KOA).
OBJECTIVE
The objective of this study is to: measure the ability of mobile AI-VA to provide responses to questions associated with KOA; identify the impact of different software, regions, and languages on such responses; and explore the accuracy of medical advice given by AI-VA when compared with those obtained through conventional methods.
METHODS
AI-VA software popular in China (such as, Huawei’s Xiaoyi and Apple’s Siri) were applied to test the responses to 15 most frequently-asked patient-centered questions associated with KOA in both the English and Chinese languages. Voice recognition and response abilities, as well as the accuracy of the medical advice given by different AI-VAs, were measured. Siri was further tested in four international medical centers with regards to the KOA questions using different languages and search engines. A questionnaire comprising the KOA questions was also sent to seven medical practitioners not specializing in KOA and five junior orthopedic surgeons as positive controls.
RESULTS
The speech recognition and voice response capabilities of the two aforementioned AI-VAs were similar in the Chinese language test, and Siri's accuracy of medical advice was significantly better than that of Xiaoyi's (Z=2.762, p=0.006). Siri performed much better than Xiaoyi in the English language test, especially in terms of voice response capability (53.3% vs. 0%). Siri also showed satisfactory agreement across the cities and languages tested, but there were differences in the accuracy of medical advice given by the AI-Vas when using different search engines. The accuracy of Siri's medical advice was higher than that obtained from direct web searches (95.6% vs. 53.3%, Z=4.290, p<0.001), suggesting higher ability of Siri to filter web searches. The accuracy of medical advice given by Siri was also higher than that given by clinicians not specializing in KOA and it was not found to be inferior to the medical advice given by senior orthopedic surgeons.
CONCLUSIONS
Current performance of AI-VAs for providing voice responses is generally known to be unsatisfactory. It is limited to presenting web search results to users, and the accuracy of medical advice given by them varies between countries and search engines used. The medical advice given by AI-VAs is deemed informative only due to their ability to filter direct web search results. The diagnoses of early degenerative diseases using AI-VA is advantageous, especially during the COVID-19 pandemic which has limited users’ access to hospitals and the available medical resources are scarce. Still, mobile AI-VAs generally fail to meet expectations in the field of medicine.