Objective To develop a predictive model of acute hematologic toxicity (HT) based on preradiotherapy CT radiomic features combined with clinical and dosimetric characteristics for patients with cervical cancer undergoing definitive radiotherapy.

Methods Data from 82 patients with cervical cancer who received definitive radiotherapy were retrospectively analyzed. The patients were divided into training and test sets at an 8:2 ratio, with grade ≥2 HT as the study endpoint. Clinical characteristics and dosimetric parameters of pelvic regions were collected, and 1 046 radiomic features were extracted from the overall pelvic structure using PyRadiomics. Radiomic features were processed using Z-score normalization, Spearman correlation analysis (>0.9), and LASSO regression for feature selection. A radiomics model was constructed using K-nearest neighbor machine learning with fivefold cross-validation. For 56 clinical and dosimetric features, independent predictors were identified through univariate and multivariate analyses to establish a clinical-dose model. A hybrid model was further constructed by combining the two sets of features. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), decision curve analysis, and calibration curves.

Results Forty patients (48.8%) developed grade ≥2 HT. Univariate and multivariate analyses showed that V20 of the lumbosacral spine was significantly associated with grade ≥2 HT (P<0.05). The AUC values for the radiomics, clinical-dose, and hybrid models were 0.802, 0.750, and 0.861 in the training set, respectively, and 0.703, 0.633, and 0.781 in the test set, respectively.

Conclusion Preradiotherapy CT radiomic features combined with dosimetric characteristics can predict the occurrence of acute hematologic toxicity in patients with cervical cancer, potentially facilitating early clinical intervention.