Prediction of Radiotherapy-Induced Hematologic Toxicity in Cervical Cancer with Cohort-Aware Framework

Abstract: Hematologic toxicity (HT) is a major dose-limiting complication of pelvic radiotherapy for cervical cancer. Although radiomic and dosiomic features improve HT prediction beyond dosimetric metrics, their performance is highly sensitive to contour variability, limiting generalizability. We developed a cohort-aware representation-learning framework to address this challenge. We retrospectively analyzed 152 cervical cancer patients treated with pelvic radiotherapy without concurrent chemotherapy. Patients were divided into two cohorts based on the operators performing pelvic bone segmentation. HT prediction models were developed using cohort-specific training, pooled training, statistical harmonization, and a cohort-aware neural network (CANN) that learns shared and cohort-specific representations with contrastive regularization. Performance was evaluated using cross-validation and an independent test set. Cohort-specific models achieved test AUCs of 0.77 and 0.71, outperforming a dosimetry-only model (AUC=0.58). Directly pooling cohorts reduced performance (test AUC=0.64). Statistical harmonization provided limited benefit, while adversarial and correlation-based alignment further degraded performance. CANN achieved the best balance between robustness and generalizability (test AUC=0.72), with ablation studies confirming the importance of cohort-specific representations and contrastive alignment. These results demonstrate that cohort-aware representation learning effectively mitigates contour variability and improves the generalizability of radiomic and dosiomic models for HT prediction.

• The paper presents a cohort-aware neural network that integrates radiomic, dosiomic, and clinical features to improve prediction of radiotherapy-induced hematologic toxicity in cervical cancer.
• The study demonstrates that combining advanced imaging biomarkers with clinical data yields higher AUCs (~0.77-0.81) compared to traditional dosimetric models.
• The proposed framework effectively mitigates cohort-specific segmentation biases and enhances model generalizability across heterogeneous patient data.

Prediction of Radiotherapy-Induced Hematologic Toxicity in Cervical Cancer with a Cohort-Aware Framework

Introduction and Motivation

This study addresses the persistent challenge of hematologic toxicity (HT) in cervical cancer patients undergoing pelvic radiotherapy, a primary dose-limiting factor adversely affecting treatment outcomes. Standard predictors, largely reliant on dosimetric parameters and basic clinical features, have consistently demonstrated limited discriminatory power and a lack of individualized risk stratification, due in part to significant interindividual differences in hematopoietic reserve and bone marrow radiosensitivity.

Prior research has validated the additive predictive value of radiomics and dosiomics—high-dimensional, quantitative descriptors extracted from routine imaging and dose distributions—for capturing anatomical and biological heterogeneity in the pelvic bone marrow microenvironment. However, the integration of such biomarkers into robust predictive models has been thwarted by their pronounced sensitivity to cohort-specific segmentation protocols, operator-dependent delineation practices, and systematic inter-observer variability, leading to poor model generalizability and unreliable cross-cohort performance.

Methodological Framework

A retrospective analysis was performed on 152 cervical cancer patients treated with external beam radiotherapy (EBRT) alone, partitioned into two cohorts according to segmentation operator and practice temporalities. This controlled for confounding due to chemotherapy and ensured that cohort differences were predominantly the result of systematic contouring protocols.

Feature extraction encompassed three domains: (1) radiomic and dosiomic features from CT images and dose maps using PyRadiomics, (2) conventional dosimetric parameters, and (3) clinical metadata. Segmentation of pelvic bone subregions (iliac bone, lower pelvis, lumbosacral vertebrae) was performed using a semi-automatic workflow, introducing a reproducible yet inherently cohort-specific source of anatomical annotation.

Feature selection utilized SelectKBest (ANOVA F-test), with dimensionality constrained to 100 features based on cross-validated AUC optimization across several model architectures. Baseline predictive models included logistic regression, random forest, neural networks, and gradient boosting, all benchmarked with and without feature harmonization procedures (ComBat, CovBat, domain-adversarial neural network [DANN], ICC filtering).

Central to the study is the development of the Cohort-Aware Neural Network (CANN), which eschews naive domain invariance in favor of explicitly conditioning prediction on cohort identifier. CANN employs a parallel-MLP architecture: independent "cohort heads" learn cohort-specific representations, weighted dynamically by a site probability estimator (MLP-based), and predictions are synthesized in a soft, learnable fashion. Contrastive regularization further enforces separation of shared and private subspaces, mitigating negative transfer and preserving biologically relevant cohort-specific signals.

Results

Impact of Feature Domains

Radiomic and dosiomic features, when combined with clinical and dosimetric variables, consistently outperformed conventional models. AUCs for cohort-specific models reached 0.81 (Cohort A) and 0.77 (Cohort B) in cross-validation, with independent test AUCs of 0.77 and 0.71, respectively. These results represent a substantial (~0.20) improvement over dosimetry-only models (AUC=0.58), delineating the insufficiency of traditional metrics for HT risk prediction in this clinical context.

Cohort Effects and Generalization

Pooling data across cohorts without alignment induced notable degradation in model performance (test AUC=0.64 for pooled training). Neither statistical harmonization (ComBat/CovBat) nor DANN remedied this deficiency; harmonization provided only marginal improvements, and adversarial invariance often exacerbated predictive decline due to negative transfer, confirming that domain shifts reflect systematic, non-stochastic variation rather than technical noise.

CANN, in contrast, robustly rectified these generalizability issues, achieving cross-validation AUCs of 0.77 and test AUCs of 0.72 while maintaining calibration and balanced error profiles. Importantly, permutation testing of cohort labels ruled out direct exploitation of cohort identity, establishing that improvements arose from conditional modeling of outcome predictors.

Analysis of Domain Shift and Biological Relevance

Kolmogorov-Smirnov testing demonstrated that regional radiomic feature distribution shifts were most severe in the lumbosacral vertebrae and lower pelvis, corresponding to areas of greatest delineation ambiguity. SHAP interpretation identified baseline platelet and neutrophil counts (clinical), and a range of wavelet-based GLCM/GLSZM (radiomic/dosiomic) features as dominant contributors to model predictions, with lower pelvis-derived features comprising 34% of the most influential predictors. Traditional dosimetric variables, including V40Gy and mean dose, had negligible predictive utility.

Volume analysis attributed cross-cohort discrepancies primarily to systematic segmentation differences, particularly at the superior lumbosacral boundary, resulting in up to 5.9% volume variation. These findings validate that observed feature drift emanates from reproducible, practice-related annotation differences, not patient anatomy or imaging protocol.

Decision Curve Analysis

CANN achieved strictly higher net clinical benefit (decision curve analysis) than both unharmonized and ComBat-harmonized models throughout the clinically relevant threshold range, indicating both theoretical and practical superiority for patient-specific stratification and treatment planning.

Implications and Future Directions

This work provides compelling evidence for the necessity of cohort-aware, conditionally adaptive frameworks in radiomics-driven prediction of radiotherapy-induced HT. Failure to account for segmentation-derived domain shifts or simplistic reliance on global harmonization critically undermines model validity and cross-cohort utility. The CANN approach—modeling both cohort-invariant and cohort-specific feature structure—represents a significant methodological advance for integrating heterogeneously annotated datasets and supports the translation of high-dimensional imaging biomarkers into robust, clinically actionable models.

Practically, the results suggest that precise segmentation protocol documentation and consideration of cohort context should be standard in multi-institutional radiomics modeling. The dominance of lower pelvic features in predictive utility further encourages focused anatomical subregion analysis for future imaging biomarker discovery.

Future work should generalize this framework to continuous and hierarchical representations of cohort attributes, incorporate multi-institutional and multi-modality (e.g., PET/MRI) datasets, and explore translation to other toxicity endpoints and tumor sites. Model interpretability strategies (e.g., SHAP, counterfactuals) should be coupled with prospective, adaptive radiotherapy protocols to enable truly personalized treatment risk estimation.

Conclusion

Integrating CT radiomics and 3D dosiomics markedly enhances prediction of radiotherapy-induced hematologic toxicity in cervical cancer. Model robustness and generalizability are fundamentally constrained by segmentation-dependent domain shifts, which are not effectively addressed by traditional harmonization or adversarial techniques. The proposed Cohort-Aware Neural Network framework conditionally incorporates cohort information, preserving essential clinical dependencies and achieving superior performance across heterogeneous patient datasets. This approach advances the reliability and translation of radiomics-driven models into individualized radiation oncology.