Preferences for Artificial Intelligence Clinicians Before and During the COVID-19 Pandemic: Discrete Choice Experiment and Propensity Score Matching Study (Preprint)

Abstract

BACKGROUND The COVID-19 pandemic poses a great threat to the public health system globally and has squeezed medical and doctor resources. Artificial intelligence (AI) has potential uses in virus detection and relieving the public health pressure caused by the pandemic. In the case of a shortage of medical resources caused by the pandemic, whether people’s preference for AI doctors and traditional clinicians has changed is worth exploring. OBJECTIVE We aim to quantify and compare people’s preference for AI medicine and traditional clinicians before and after the COVID-19 pandemic to check whether people’s preference is affected by the pressure of pandemic METHODS The propensity score matching (PSM) method was applied to match two different groups of respondents recruited in 2017 and 2020 with similar demographic characteristics. A total of 2048 respondents (1520 from 2017 and 528 from 2020) completed the questionnaire and were included in the analysis. The Multinomial Logit Model (MNL) and Latent Class Model (LCM) were used to explore people’s preferences for different diagnosis methods. RESULTS Among these respondents, 84.7% in 2017 and 91.3% in 2020 were confident that AI diagnosis would outperform human clinician diagnoses in the future. Both groups of respondents matched from 2017 and 2020 attached most importance to the attribute ‘accuracy’, followed by ‘diagnosis expense’, and they prefer the combined diagnosis of AI and human clinicians (2017: odds ratio [OR] 1.645; 95% CI 1.535,1.763, p < 0.001; 2020: OR 1.513, 95% CI 1.413, 1.621, p < 0.001, Reference level: Clinician). LCM identified three classes with different attribute priorities. In Class 1, the preference for combination diagnosis and accuracy remains constant in 2017 and 2020, and higher accuracy (e.g., 2017 OR for 100% 1.357; 95% CI 1.164, 1.581) is preferred. People in 2017 and 2020 prefer 0 min outpatient waiting time and 0 RMB diagnosis expense. In Class 2, the 2017 matched data is also very similar to class 2 in 2020, AI combined with human clinicians (2017: OR 1.204, 95% CI 1.039, 1.394, p = 0.011; 2020: OR 2.009, 95% CI 1.826, 2.211, p < 0.001, Reference level: Clinician) and 20 minutes (2017: OR 1.349, 95% CI 1.065, 1.708, p < 0.001; 2020: OR 1.488, 95% CI 1.287, 1.721, p < 0.001, Reference level, 0 min) of outpatient waiting time were consistently preferred. In Class 3, the respondents in 2017 and 2020 had different preferences for diagnosis method; respondents in Class 3 of 2017 prefer clinicians, whereas respondents in Class 3 of 2020 prefer AI diagnosis. The odds ratios of accuracy continued increasing with the increasing of accuracy, like other classes of 2017 and 2020. As for the latent class segmented according to different sexes, all of the male and female respondent classes from 2017 and 2020 rank accuracy as the most important attribute. CONCLUSIONS Individual preference for clinical diagnosis between AI and human clinicians were very similar and mostly unaffected by the burden of the public health system caused by the pandemic. Diagnosis accuracy and expense for diagnosis were of the most important attributes of choice of the type of diagnosis. These findings can provide guidance for policymaking relevant to the development of AI-based healthcare.