Efficacy of a US-developed machine-learned coronary artery disease algorithm in China

Background Cardiac computed tomography angiography (CCTA), the premier coronary artery disease (CAD) rule-out test, is less available in rural areas (vs urban) in the US. To address this gap, we previously developed a test using non-invasive signals and a machine-learned algorithm, usable anywhere with internet. We trained the algorithm using US populations: (1) subjects enrolled before CCTA for specificity and (2) subjects undergoing invasive coronary angiography (ICA), for both sensitivity and specificity. US validation used separate, identically constructed, blinded populations (N = 1816). In US validation, sensitivity was 88% (95% CI 85–91%) and specificity was 51% (48–55%). Although China is urbanizing rapidly, 510 million (36% of the population) still reside in rural areas. Thus, China might benefit from the test, but given the demographic differences, the efficacy of the test in the Chinese population is unknown. The aim of the present study was to assess test performance within a cohort of Chinese ICA subjects. Methods Subjects were enrolled at Zhongshan Hospital, Fudan University (China), using the same inclusion and exclusion criteria as in the US study. The same device as in the US captured signals pre-ICA. We applied the unchanged algorithm from US validation to the signals, then matched with the ICA results. Results 558 subjects had both ICA results and a signal, 458 passed signal quality and outlier checks, and 348 had significant CAD. Compared to the US sensitivity cohort, the Chinese cohort was significantly younger, had a lower BMI, a higher proportion of males, and lower rates of tobacco use, hypertension, and diabetes. Race was Asian only in China, while the US was primarily White/Caucasian. Chinese cohort sensitivity was 89% (86–93%), p = 0.46 vs the US. Discussion Our study was limited by the lack of a Chinese CCTA specificity cohort. However, sensitivity was high, and thus robust to the significant demographic shifts. Therefore, we anticipate similar behavior in specificity; to explore that, we combined the Chinese ICA negatives (N = 110) with the US CCTA negatives from the previous work, in the same proportion as used in US validation, which models the population of patients under assessment for cardiac symptoms. The resultant specificity was 51% (48–55%). With the Chinese ICA positives, the area under the receiver operator characteristic curve was 0.79 (0.77–0.81). These results suggest the test’s potential effectiveness in the Chinese population. Trial registration : NCT04034537.