Longitudinal Risk Prediction in Mammography with Privileged History Distillation
arXiv:2603.15814v1 Announce Type: new Abstract: Breast cancer remains a leading cause of cancer-related mortality worldwide. Longitudinal mammography risk prediction models improve multi-year breast cancer risk prediction based on prior screening exams. However, in real-world clinical practice, longitudinal histories are often incomplete, irregular, or unavailable due to missed screenings, first-time examinations, heterogeneous acquisition schedules, or archival constraints. The absence of prior exams degrades the performance of longitudinal risk models and limits their practical applicability. While substantial longitudinal history is available during training, prior exams are commonly absent at test time. In this paper, we address missing history at inference time and propose a longitudinal risk prediction method that uses mammography history as privileged information during training and distills its prognostic value into a student model that only requires the current exam at infere
arXiv:2603.15814v1 Announce Type: new Abstract: Breast cancer remains a leading cause of cancer-related mortality worldwide. Longitudinal mammography risk prediction models improve multi-year breast cancer risk prediction based on prior screening exams. However, in real-world clinical practice, longitudinal histories are often incomplete, irregular, or unavailable due to missed screenings, first-time examinations, heterogeneous acquisition schedules, or archival constraints. The absence of prior exams degrades the performance of longitudinal risk models and limits their practical applicability. While substantial longitudinal history is available during training, prior exams are commonly absent at test time. In this paper, we address missing history at inference time and propose a longitudinal risk prediction method that uses mammography history as privileged information during training and distills its prognostic value into a student model that only requires the current exam at inference time. The key idea is a privileged multi-teacher distillation scheme with horizon-specific teachers: each teacher is trained on the full longitudinal history to specialize in one prediction horizon, while the student receives only a reconstructed history derived from the current exam. This allows the student to inherit horizon-dependent longitudinal risk cues without requiring prior screening exams at deployment. Our new Privileged History Distillation (PHD) method is validated on a large longitudinal mammography dataset with multi-year cancer outcomes, CSAW-CC, comparing full-history and no-history baselines to their distilled counterparts. Using time-dependent AUC across horizons, our privileged history distillation method markedly improves the performance of long-horizon prediction over no-history models and is comparable to that of full-history models, while using only the current exam at inference time.
Executive Summary
The article presents a novel approach to longitudinal breast cancer risk prediction in mammography, addressing the critical challenge of missing prior screening history at inference time. The authors propose the Privileged History Distillation (PHD) method, which leverages full longitudinal history as privileged information during training to distill prognostic cues into a student model that relies solely on the current exam at deployment. Validated on the CSAW-CC dataset, PHD achieves performance comparable to full-history models for long-horizon predictions while maintaining the practical advantage of requiring only the current exam during inference, thus bridging the gap between training and real-world clinical constraints.
Key Points
- ▸ The study tackles a pervasive issue in clinical practice: incomplete or unavailable longitudinal mammography histories at inference time, which degrades risk prediction model performance.
- ▸ The proposed PHD method uses a privileged multi-teacher distillation framework, where each teacher model specializes in a specific prediction horizon using full longitudinal history, while the student model distills this knowledge using only reconstructed history from the current exam.
- ▸ Empirical validation on the CSAW-CC dataset demonstrates that PHD significantly outperforms no-history baselines and approaches the performance of full-history models across multi-year prediction horizons, as measured by time-dependent AUC.
Merits
Innovative Methodology
The introduction of privileged history distillation as a means to transfer prognostic value from full longitudinal histories to a student model requiring only current exams is both conceptually elegant and practically impactful.
Clinical Relevance
The method directly addresses a real-world clinical limitation—missing prior exams—by enabling robust risk prediction without sacrificing performance, which could improve early detection and patient outcomes.
Empirical Rigor
The validation on a large longitudinal dataset (CSAW-CC) with multi-year outcomes and the use of time-dependent AUC metrics provide strong evidence of the method's effectiveness across prediction horizons.
Demerits
Dependence on Training Data Completeness
The method assumes that comprehensive longitudinal histories are available during training, which may not always be feasible in resource-constrained settings or where archival data is incomplete.
Reconstruction Assumptions
The reliance on reconstructed history from the current exam for distillation may introduce biases or inaccuracies, particularly in cases with highly irregular screening intervals or missing data patterns.
Generalizability Concerns
While validated on CSAW-CC, the method's performance in diverse populations or healthcare systems with differing screening protocols remains to be thoroughly tested.
Expert Commentary
The authors present a compelling solution to a long-standing challenge in longitudinal risk prediction: the mismatch between training data availability and real-world clinical constraints. By leveraging privileged information during training to distill prognostic cues into a more practical student model, the PHD method not only advances the technical frontier of mammography-based risk prediction but also aligns with the practical realities of healthcare delivery. The use of horizon-specific teachers and time-dependent AUC metrics further demonstrates a nuanced understanding of the temporal dynamics in cancer risk stratification. However, the reliance on reconstructed history and the assumption of complete training data warrant cautious optimism. Future work should explore the method's robustness in heterogeneous populations and its integration with federated learning frameworks to address data fragmentation across institutions. Additionally, the ethical and regulatory implications of training models with privileged information—while deploying them with limited inputs—merit deeper scrutiny to ensure equitable and safe deployment in clinical settings.
Recommendations
- ✓ Conduct further validation of PHD across diverse populations and healthcare systems to assess generalizability and robustness, particularly in low-resource settings where archival data may be incomplete.
- ✓ Explore the integration of PHD with federated learning to enable collaborative training across institutions without centralizing sensitive longitudinal data, thereby enhancing privacy and scalability.
- ✓ Develop explainable AI frameworks to interpret the distilled student model's predictions, ensuring clinical transparency and trust in AI-driven risk stratification tools.
- ✓ Engage regulatory bodies early in the development and validation process to align PHD with evolving standards for AI in medical diagnostics, particularly regarding the use of privileged information in training.
