Prediction of Ki-67 Expression in HIV-Associated Lung Adenocarcinoma Patients Using Multiple Machine Learning Models Based on CT Imaging Radiomics

Chang Song; Jingsong Chen; Chunyan Zhao; Shulin Song; Tong Yang; Aichun Huang; Renhao Liu; Yanxi Pan; Chaoyan Xu; Canling Chen; Qingdong Zhu

doi:10.2147/CMAR.S505390

Prediction of Ki-67 Expression in HIV-Associated Lung Adenocarcinoma Patients Using Multiple Machine Learning Models Based on CT Imaging Radiomics

Cancer Manag Res. 2025 Apr 25:17:881-892. doi: 10.2147/CMAR.S505390. eCollection 2025.

Authors

Chang Song^#¹, Jingsong Chen^#², Chunyan Zhao^#¹, Shulin Song^#³, Tong Yang⁴, Aichun Huang¹, Renhao Liu¹, Yanxi Pan³, Chaoyan Xu¹, Canling Chen¹, Qingdong Zhu¹

Affiliations

¹ Tuberculosis Department, Nanning Fourth People's Hospital, Nanning, Guangxi, 530023, People's Republic of China.
² Gastroenterology Department, Hepu County People's Hospital, Beihai, Guangxi, 536100, People's Republic of China.
³ Radiology Department, Nanning Fourth People's Hospital, Nanning, Guangxi, 530023, People's Republic of China.
⁴ Rehabilitation Department, Hepu County People's Hospital, Beihai, Guangxi, 536100, People's Republic of China.

^# Contributed equally.

Abstract

Purpose: The incidence of lung adenocarcinoma (LUAD) in HIV-infected individuals is significantly increased. However, invasive procedures for Ki-67 assessment may increase the risk of complications. Therefore, developing a non-invasive and accurate method for Ki-67 prediction holds significant clinical importance. This study aims to explore the feasibility and value of a radiomics model based on preoperative CT images in predicting Ki-67 expression levels in HIV-associated LUAD.

Patients and methods: A total of 237 patients with HIV-associated LUAD were included. Of these, 102 were classified into the high Ki-67 expression group, and 135 into the low Ki-67 expression group. The patients were randomly divided into a training group (n=189) and a validation group (n=48) in a 4:1 ratio. Feature selection was based on intra-class correlation coefficient (ICC), Spearman correlation coefficient, and Least Absolute Shrinkage and Selection Operator (LASSO) regression, yielding 16 optimal radiomic features for building a logistic regression model. Model performance was evaluated by sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), F1 score, and the area under the receiver operating characteristic curve (AUC).

Results: 1834 CT image features were extracted, with 16 retained for further analysis. The Support Vector Machine (SVM) model demonstrated the most balanced and optimal performance among the seven developed models. It achieved robust sensitivity (training set: 0.89; testing set: 0.86), specificity (training set: 0.92; testing set: 0.89), PPV (training set: 0.89; testing set: 0.86), NPV (training set: 0.92; testing set: 0.89), F1 score (training set: 0.89; testing set: 0.86), and AUC (training set: 0.975; testing set: 0.905), indicating excellent predictive accuracy.

Conclusion: This study first demonstrates that a preoperative CT-based radiomics model can non-invasively predict Ki-67 expression levels in HIV-associated LUAD patients. This finding not only provides a precise assessment tool for the HIV-infected population to avoid the risks of invasive examinations but also paves new interdisciplinary research avenues for exploring tumor heterogeneity under immunodeficiency conditions.

Keywords: HIV; Ki-67; SVM; lung adenocarcinoma; machine learning; radiomics.