Clinical Characteristics and Prediction Model of Re-positive Nucleic Acid Tests Among Omicron Infections by Machine Learning: a Real-world Study of 35,488 Cases

Background: During the Omicron BA.2 variant outbreak in Shanghai, China, from April to May 2022, PCR nucleic acid test re-positivity (TR) occurred frequently, yet the risk factors and predictive models for TR remain unclear. This study aims to identify the factors influencing Omicron TR and to develop machine learning models to predict TR risk. Accurately predicting re-positive patients is crucial for identifying high-risk individuals, optimizing resource allocation, and developing personalized treatment and management plans, thereby effectively controlling the spread of the epidemic, reducing community burden, and ensuring public health.

Methods: A retrospective study was conducted among individuals infected with Omicron BA.2 variant from April 12 to May 25, 2022, in the largest Shanghai Fangcang shelter hospital. Five machine learning models were compared, including k-nearest-neighbors (KNN), logistic regression (logistic), bootstrap aggregation (bagging), error back-propagation (BP) neural network, and support vector machines (SVM), to select the best prediction model for the TR risk factors.

Results: A total of 35,488 cases were included in this real-world study. The TR and control groups comprised of 6,171 and 29,317 cases respectively, with a re-positive rate of 17.39%. Higher occurrence of TR was observed in young age, males, those with obvious symptoms, underlying diseases, and a low Ct value. The KNN model proved to be the best in predicting the prognosis in the overall evaluation (accuracy = 0.8198, recall = 0.8026, and AUC = 0.8110 in the test set).

Interpretation: Higher TR risk was found in infected cases who were underage or with underlying diseases; vaccine brand and inoculation status were not significantly associated with TR. KNN was the most effective machine learning model to predict TR occurrence in isolation.