2891 - A Multi-Omics Model Predicting Acute Hematologic Toxicity (HT) in Locally Advanced Cervical Cancer Undergoing External-Beam Radiotherapy

Purpose/Objective(s): To develop a multi-omics model integrating clinical, dosimetric, and radiomic features to predict acute hematologic toxicity (HT) in patients with locally advanced cervical cancer (LACC) undergoing external-beam radiotherapy.

Materials/Methods: 187 LACC patients with definitive chemoradiotherapy were retrospectively studied. Acute HT was graded weekly by CTCAE 5.0. Clinical factors were BMI, FIGO stage, and chemotherapy status. Dosimetric features were dose of pelvic bone marrow (BM), including 2D dose (V10-40Gy, mean dose). Radiomic features were extracted from the BM on plan CT. 7 predictive models were: (1) Clinical, (2) Dosimetric, (3) Radiomic, (4) Clinical-Dosimetric, (5) Dosimetric-Radiomic, (6) Clinical-Radiomic, and (7) Multi-Omics (integrating all features). 4 machine learning —Random Forest (RF), Logistic Regression(LR), Naive Bayes(NB), and Support Vector Machine (SVM)—were evaluated. Model performance was assessed using Area Under Curve (AUC), accuracy, sensitivity, and specificity.

Results: Of 187 patients, 100 met inclusion criteria after excluding incomplete datasets. Patients with grade =3 was 49%. 6 Dosimetric features selected from 5174 features and 5 radiomic features selected from 1326 features by Pearson Correlation Analysis and Random Forest Importance Ranking. Compared to the combination of two-omics and one-omic models, the multi-omics predictive model demonstrates the best AUC (RF: 0.982, LR: 0.894, NB: 0.847, SVM: 0.819), the same as the best accuracy, sensitivity,specificity for all the 4 ML algorithms. Details can be found in Table 1.