Integrating CT radiomics and transcriptomics: a biologically-informed machine learning model for predicting
Xin Xiong1, Xiaodong Ji2, Xilong Yang3
1Department of Otorhinolaryngology Head and Neck Surgery, Tianjin First Central Hospital, Institute of Otolaryngology of Tianjin, Key Laboratory of Auditory Speech and Balance Medicine, Key Medical Cultivation Discipline of Tianjin (Otolaryngology), Quality Control Centre of Otolaryngology, Tianjin, China.
Background:
Predicting response to induction chemotherapy (IC) in advanced laryngeal cancer (LC) remains a clinical challenge. This study aimed to develop a non-invasive, interpretable model integrating CT radiomics and clinical features to predict chemotherapy outcomes.
Methods:
We retrospectively analyzed 161 advanced LC patients treated with IC. From pre-treatment CT images, 1,321 radiomics features were extracted, and a radiomics score (Rad-score) was constructed using LASSO regression. Transcriptomic analysis explored the biological basis of Rad-score. Independent predictors were identified via multivariate logistic regression and used to build five machine learning models. Model performance was evaluated using AUC, accuracy, and specificity. SHAP analysis was applied to interpret the optimal model.
Results:
Four robust radiomics features were selected to construct the Rad-score. The Rad-score demonstrated satisfactory discrimination with an Area Under the Curve (AUC) of 0.715 in the training set and 0.707 in the validation set. In multivariate analysis, the Rad-score (Odds Ratio [OR]=2.89, 95% CI: 1.29-6.48, P = 0.010), gap invasion and validation were identified as independent predictors of chemotherapy response. Among the machine learning models, the Random Forest model achieved the best performance, yielding an AUC of 0.914 in the training set, 0.856 in the validation set, and 0.810 in the external test set. Decision curve analysis confirmed the clinical utility of the model. SHAP analysis confirmed Rad-score and fat space invasion as core predictors, with synergistic effects.
Conclusions:
We developed a highly accurate and interpretable Random Forest model that integrates radiomics and clinical features to predict IC response in advanced LC. This tool enables precise risk stratification and personalized treatment decisions, sparing non-responders from ineffective therapy. Prospective studies are needed to validate its clinical utility.
