Trends in cardiovascular risk in the United States 1999 to 2018

Abstract

Introduction: The ongoing COVID-19 outbreak has severely affected the healthcare service delivery, including routine endoscopy services. We have previously shown a 37% to 46% drop in gastrointestinal cancers diagnosed during the initial phase of COVID-19 in Hong Kong. As the situation would unlikely to resolve soon, we have to develop a new strategy to monitor and predict our endoscopic service requirement, in order to provide robust guidance on optimal endoscopy volume to minimise the delay in gastrointestinal cancers diagnosis. Methods: We retrieved from the CDARS of the Hospital Authority the number of patients who had upper and lower endoscopy and the number of new cases of gastrointestinal cancers diagnosed, as well as the realtime reproductive number of local COVID-19 case and the number of new COVID-19 cases from the School of Public Health dashboard. The number of patients with potential delay in cancer diagnosis was estimated with the autoregressive integrated moving average (ARIMA) model. Six different machine learning models: traditional linear regression (LR), random forest (RF), support vector machine (SVM), stochastic gradient boosting (SGB), neural network (NN) and extreme gradient boosting (XGBoost) were used to estimate the requirement of endoscopy service for gastrointestinal cancer diagnosis during different phases of COVID-19 in Hong Kong. Results: There were a total of 337903 upper endoscopies and 254588 lower endoscopies performed between October 2016 and June 2020. The model with the best performance in terms of prediction of minimal upper and lower endoscopy volume without delay in diagnosis of gastric and colorectal cancer is the XGBoost (MAPE±95% confidence interval [CI]=3.0±2.6 and 2.4±0.01, respectively). The minimal average weekly number of upper endoscopy to be performed in the subsequent month is 1781±149 (mean±95% CI), which is still significantly lower than usual upper endoscopy volume in the pre-COVID-19 period (1974±818, P<0.01). Accordingly, the minimal average weekly number of lower endoscopy required is 1115±386 as compared with usual lower endoscopy volume of 1330±456 (P<0.01). Conclusion: Machine learning model, particularly the XGBoost, can be applied in the prediction of minimal number of endoscopy service required to minimise delay in gastrointestinal cancer diagnosis during the ongoing COVID-19 outbreak in Hong Kong.