Development of an exploratory prediction model for preoperative CK19 expression in esophageal cancer driven by
Jiafu Yang1, Sicheng Wang2, Xin Chen2
1Department of Radiology, Guangdong Engineering Research Center of AI-Powered Precision Cancer Diagnostics and Therapeutics(Proposed)/GuangDong Engineering Technology Research Center of AI-Powered Precision Cancer Diagnostics and Therapeutics, Cancer Hospital of Shantou University Medical College, Shantou, China.
Background:
Esophageal cancer ranks among the most lethal malignancies worldwide, particularly prevalent in the Guangdong-Chaoshan region of China due to regional dietary habits. Cytokeratin 19 (CK19) is an important immunohistochemical marker reflecting tumor invasiveness and metastatic potential; however, noninvasive preoperative prediction of CK19 expression remains unavailable. This study aimed to develop a CT-based radiomics model combined with machine learning to predict CK19 expression preoperatively.
Methods:
This study included 134 patients with primary esophageal cancer. All patients underwent enhanced CT scans before surgery, and CK19 expression was evaluated by pathological analysis after surgery. Radiomics technology was used to extract multidimensional image features including shape, texture, and first-order features from CT images. A prediction model was established by combining machine learning models such as gradient boosted decision tree (GBDT), random forest (RF), extreme gradient boosting (XGB), and lightweight gradient boosting machine (LGBM), and the interpretability of the model was analyzed by SHAP value.
Results:
The random forest model showed relatively higher accuracy and precision among the compared models, with an AUC value of 0.6765 and an accuracy of 0.8293. GBDT demonstrated a more balanced performance (AUC: 0.6597), while XGB (AUC: 0.6744) and LGBM (AUC: 0.6807) showed comparable but overall slightly lower discriminative ability. Feature importance analysis showed that the features after wavelet transformation made a significant contribution to the prediction results. The results verified the potential of radiomics combined with machine learning technology in the preoperative prediction of CK19 expression.
Conclusion:
This study developed a preoperative noninvasive prediction model based on radiomics and machine learning, which showed modest predictive performance in an exploratory setting in the evaluation of CK19 markers in patients with esophageal cancer, may provide preliminary support for further exploration in precision medicine. In the future, the clinical applicability of this model needs to be further verified and its promotion and application in a larger population needs to be optimized.
