Machine learning model for prediction of coronavirus disease 2019 within 6 months after three doses of BNT162b2 in Hong Kong.

IF 3.1 4区医学 Q1 MEDICINE, GENERAL & INTERNAL

Hong Kong Medical Journal Pub Date : 2025-06-23 DOI:10.12809/hkmj2411879

J T Tan, R Zhang, K H Chan, J Qin, I F N Hung, K S Cheung

{"title":"Machine learning model for prediction of coronavirus disease 2019 within 6 months after three doses of BNT162b2 in Hong Kong.","authors":"J T Tan, R Zhang, K H Chan, J Qin, I F N Hung, K S Cheung","doi":"10.12809/hkmj2411879","DOIUrl":null,"url":null,"abstract":"Introduction: We aimed to develop a machine learning (ML) model to predict the risk of coronavirus disease 2019 (COVID-19) among three-dose BNT162b2 vaccine recipients in Hong Kong.Methods: A total of 304 individuals who had received three doses of BNT162b2 were recruited from three vaccination centres in Hong Kong between May and August 2021. The dataset was randomly divided into training (n=184) and testing (n=120) sets in a 6:4 ratio. Demographics, co-morbidities and medications, blood tests (complete blood count, liver and renal function tests, glycated haemoglobin level, lipid profile, and presence of hepatitis B surface antigen), and controlled attenuation parameter (CAP) were used to develop six ML models (logistic regression, linear discriminant analysis, random forest, naïve Bayes, neural network [NN], and extreme gradient boosting models) to predict COVID-19 risk. Model performance was assessed using area under the receiver operating characteristic curve (AUC), sensitivity, specificity, and positive predictive value (PPV) and negative predictive value (NPV).Results: Among the study population (median age: 50.9 years [interquartile range=43.6-57.8]; men: 30.9% [n=94]), 27 participants (8.9%) developed COVID-19 within 6 months. Fifteen clinical variables were used to train the models. The NN model achieved the best performance, with an AUC of 0.74 (95% confidence interval [95% CI]=0.60-0.88). Using the optimal cut-off value based on the maximised Youden index, sensitivity, specificity, PPV, and NPV were 90% (95% CI=55%-100%), 58% (95% CI=48%-68%), 16% (95% CI=8%-29%), and 98% (95% CI=92%-100%), respectively. The top predictors in the NN model include age, prediabetes/diabetes, CAP, alanine aminotransferase level, and aspartate aminotransferase level.Conclusion: An NN model integrating 15 clinical variables effectively identified individuals at low risk of COVID-19 following three doses of BNT162b2.","PeriodicalId":48828,"journal":{"name":"Hong Kong Medical Journal","volume":" ","pages":""},"PeriodicalIF":3.1000,"publicationDate":"2025-06-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Hong Kong Medical Journal","FirstCategoryId":"3","ListUrlMain":"https://doi.org/10.12809/hkmj2411879","RegionNum":4,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q1","JCRName":"MEDICINE, GENERAL & INTERNAL","Score":null,"Total":0}

引用次数: 0

Abstract

Introduction: We aimed to develop a machine learning (ML) model to predict the risk of coronavirus disease 2019 (COVID-19) among three-dose BNT162b2 vaccine recipients in Hong Kong.

Methods: A total of 304 individuals who had received three doses of BNT162b2 were recruited from three vaccination centres in Hong Kong between May and August 2021. The dataset was randomly divided into training (n=184) and testing (n=120) sets in a 6:4 ratio. Demographics, co-morbidities and medications, blood tests (complete blood count, liver and renal function tests, glycated haemoglobin level, lipid profile, and presence of hepatitis B surface antigen), and controlled attenuation parameter (CAP) were used to develop six ML models (logistic regression, linear discriminant analysis, random forest, naïve Bayes, neural network [NN], and extreme gradient boosting models) to predict COVID-19 risk. Model performance was assessed using area under the receiver operating characteristic curve (AUC), sensitivity, specificity, and positive predictive value (PPV) and negative predictive value (NPV).

Results: Among the study population (median age: 50.9 years [interquartile range=43.6-57.8]; men: 30.9% [n=94]), 27 participants (8.9%) developed COVID-19 within 6 months. Fifteen clinical variables were used to train the models. The NN model achieved the best performance, with an AUC of 0.74 (95% confidence interval [95% CI]=0.60-0.88). Using the optimal cut-off value based on the maximised Youden index, sensitivity, specificity, PPV, and NPV were 90% (95% CI=55%-100%), 58% (95% CI=48%-68%), 16% (95% CI=8%-29%), and 98% (95% CI=92%-100%), respectively. The top predictors in the NN model include age, prediabetes/diabetes, CAP, alanine aminotransferase level, and aspartate aminotransferase level.

Conclusion: An NN model integrating 15 clinical variables effectively identified individuals at low risk of COVID-19 following three doses of BNT162b2.

查看原文本刊更多论文

在香港注射三剂BNT162b2后6个月内预测2019冠状病毒病的机器学习模型。

前言：我们旨在开发一种机器学习（ML）模型，以预测香港三剂BNT162b2疫苗接种者2019冠状病毒病（COVID-19）的风险。方法：在2021年5月至8月期间，从香港的三个疫苗接种中心招募了304名接种了三剂BNT162b2的个体。数据集以6:4的比例随机分为训练集（n=184）和测试集（n=120）。使用人口统计学、合并症和药物、血液检查（全血细胞计数、肝肾功能检查、糖化血红蛋白水平、血脂和乙型肝炎表面抗原的存在）和控制衰减参数（CAP）建立6个ML模型（逻辑回归、线性判别分析、随机森林、naïve贝叶斯、神经网络[NN]和极端梯度增强模型）来预测COVID-19风险。采用受试者工作特征曲线下面积（AUC）、敏感性、特异性、阳性预测值（PPV）和阴性预测值（NPV）评估模型性能。结果：在研究人群中(中位年龄：50.9岁[四分位数间距=43.6-57.8]；男性：30.9% [n=94])， 27名参与者（8.9%）在6个月内出现COVID-19。使用15个临床变量来训练模型。NN模型取得了最好的性能，AUC为0.74（95%置信区间[95% CI]=0.60-0.88）。使用基于最大约登指数的最佳临界值，敏感性、特异性、PPV和NPV分别为90% （95% CI=55%-100%）、58% （95% CI=48%-68%）、16% （95% CI=8%-29%）和98% （95% CI=92%-100%）。神经网络模型中最重要的预测因子包括年龄、前驱糖尿病/糖尿病、CAP、丙氨酸转氨酶水平和天冬氨酸转氨酶水平。结论：整合15个临床变量的神经网络模型有效识别了三剂BNT162b2后的低风险个体。

本文章由计算机程序翻译，如有差异，请以英文原文为准。

求助全文

约1分钟内获得全文求助全文

来源期刊

Hong Kong Medical Journal MEDICINE, GENERAL & INTERNAL-

CiteScore

1.50

自引率

14.80%

发文量

117

审稿时长

10 weeks

期刊介绍： The HKMJ is a Hong Kong-based, peer-reviewed, general medical journal which is circulated to 6000 readers, including all members of the HKMA and Fellows of the HKAM. The HKMJ publishes original research papers, review articles, medical practice papers, case reports, editorials, commentaries, book reviews, and letters to the Editor. Topics of interest include all subjects that relate to clinical practice and research in all branches of medicine. The HKMJ welcomes manuscripts from authors, but usually solicits reviews. Proposals for review papers can be sent to the Managing Editor directly. Please refer to the contact information of the Editorial Office.