Work

Physics-informed neural surrogate learning predicts cervical cancer progression with mechanistic interpretability

Year: 2026
Type: article
Abstract: Deep learning models achieve high predictive accuracy in oncology but operate as black boxes, while mechanistic models offer biological interpretability but fail to capture patient heterogeneity. Prior work has explored hybrid approaches combining these paradigms, yet fundamental challenges remain in ensuring biological consistency and clinical trustworthiness. We developed a Physics-Informed Neural Surrogate Learning (PI-NSL) framework that incorporates a system of biophysical partial differential equations (PDEs) governing tumor-immune-viral dynamics as soft constraints within a deep neural network, guiding learned representations toward biological plausibility. The framework comprises a variational physics encoder that infers patient-specific PDE parameters from routine clinical data, a differentiable PDE solver for efficient spatiotemporal simulation, and a dynamic risk operator for survival prediction. We applied PI-NSL to a longitudinal cohort of \(\text{20,287}\) cervical cancer patients treated at Apollo Proton Cancer Centre (Chennai, India, \(2018-2023)\) . The development cohort was partitioned \(70/15/15\) . All \(\text{2,663}\) patients enrolled in \(2023\) were assigned to the held-out test set to enable temporal internal validation; the \(2023\) subset is therefore nested within the \(\text{3,043}\) -patient test set, not additional to the \(\text{20,287}\) -patient analytical cohort. External validation used two independent institutions: Tata Memorial Centre (TMC, \(n=\text{8,432})\) and AIIMS \(\left(n=\text{7,535}\right)\) ; total external \(n=\text{15,967}\) . PI-NSL achieved a time-dependent AUC of \(0.91\pm 0.03\) for \(2-year\) progression-free survival prediction, with a C-index of \(0.89\pm 0.04\) . Unsupervised analysis of the model’s latent representations identified three clinically distinct clusters with significantly different survival outcomes \(\left(log-rank p<0.001\right)\) : Cluster 1 (high immune activity, \(34\%\) , \( 84\%, 2-year PFS\) ), Cluster 2 (peripheral immune accumulation, \(41\%, 45\%\) ), and Cluster 3 (high proliferation, \(25\%, 28\%\) ). Model calibration demonstrated good agreement between predicted and observed survival probabilities (integrated Brier score \(=0.082\pm 0.015\) ). On the \(2023\) temporal subset nested within the test set \(\left(n=\text{2,663}\right)\) , performance was stable (time-dependent AUC \(0.89\pm 0.05; \Delta AUC=-0.01\) versus the non- \(2023\) test subset, \(p=0.72)\) . External evaluation provided evidence of generalizability within the Indian healthcare context, with time-dependent AUC values of \(0.87\pm 0.05\) at TMC and \(0.86\pm 0.05\) at AIIMS. PI-NSL demonstrates that incorporating mechanistic biological knowledge as soft constraints within neural architectures can achieve strong predictive performance while providing clinically interpretable outputs. The identified clusters generate hypotheses about mechanisms of treatment resistance that warrant prospective validation. This framework represents a step toward AI systems that are both accurate and mechanistically grounded for clinical oncology applications.
Language: English
Open Access status: gold
Estimated APC paid: $2,690
OpenAlex ID: W7237051161
FWCI:
Cites: 26
Cited by:
Related to: 10