Selection-Bias-Aware Multimodal Survival Learning for Molecular Pathological Epidemiology: A Stabilized IPW–Cox Method with a Differentiable Etiologic-Heterogeneity Head
DOI:
https://doi.org/10.56147/jbhs.3.3.152Keywords:
- Selection bias,
- Inverse probability weighting,
- Survival analysis,
- Molecular pathological epidemiology,
- Multimodal machine learning,
- Etiologic heterogeneity,
- NHANES,
- Colorectal cancer,
- Computational pathology,
- Digital health
Abstract
Background: Deep survival models for tumor-tissue and molecular studies are trained only on patients whose specimens are collected, archived, retrieved, assayed and pass quality control. In Molecular Pathological Epidemiology (MPE) this analyzable subset is a covariate-dependent draw from the incident-case population, because tissue retrieval, assay success and molecular-subtype ascertainment all depend on prognostic variables. A model that ignores this targets the analyzable-sample relationship rather than the population one.
Methods: We developed SBASURV, an architecture-agnostic survival head that (1) replaces the ordinary Cox partial likelihood with a stabilized Inverse-Probability-Weighted (IPW) Cox objective in which availability weights enter both the event term and the risk-set denominator and (2) implements the Lunn-McNeil duplication method as a differentiable subtype-specific module emitting exposure-by-subtype log-hazard ratios and a Wald statistic for etiologic heterogeneity. We validated the mechanism on five public survival datasets with induced covariate-dependent selection, then deployed the identical head on TCGA-COAD/READ colorectal cancer across three modalities clinical, tumor omics and whole-slide histopathology (features from 441 diagnostic slides via a public foundation model) using real tissue analyzability weights.
Findings: The differentiable stabilized-IPW head improved held-out concordance over a naive neural head on all five public datasets, while weight diagnostics made the bias-variance regime explicit. On real colorectal cancer, clinical+omics fusion reached held-out Harrell C 0.68 (Uno 0.66) in 530 cases; adding pathology improved discrimination over clinical alone (0.59 to 0.62) in the 367-case three-modality sub cohort. Availability weighting preserved discrimination while shifting the estimand with well-conditioned weights (mean ≈ 1, effective sample size 96–99%). A tumor-mutation-burden/MSI proxy subtype yielded significant stage-by-subtype heterogeneity (Q=10·2, p=0·001), but a genuine methylation-defined CIMP subtype did not (Q=0·01, p=0·92).
Interpretation: SBASURV provides a principled, differentiable mechanism for population-targeted multimodal survival learning. The contrast between proxy and assay-grounded subtypes shows that built-in heterogeneity tests are valuable as much for the nulls they return as for the signals and must be read with weight and sample-size diagnostics in view. The method is a research instrument, not a clinical tool.
Downloads
References
Cox DR (1972) Regression models and life-tables. Journal of the Royal Statistical Society: Series B 34: 187-202.
Jared LK, Uri S, Alexander C, Jonathan B, Tingting J, et al. (2018) DeepSurv: Personalized treatment recommender system using a Cox proportional hazards deep neural network. BMC Medical Research Methodology 18: 24.
Ching T, Zhu X, Lana XG (2018) Coxnnet: An artificial neural network method for prognosis prediction of high-throughput omics data. PLoS Computational Biology 14: e1006076.
Lee C, William RZ, Jinsung Y, Mihaela S (2018) DeepHit: A deep learning approach to survival analysis with competing risks. AAAI Conference on Artificial Intelligence 32: 2314-2321.
Kvamme H, Ørnulf B, Scheel I (2019) Time-to-event prediction with neural networks and Cox regression. Journal of Machine Learning Research 20: 1-30.
Wang Z, Sun J (2022) SurvTRACE: Transformers for survival analysis with competing events. 13th ACM International Conference on Bioinformatics, Computational Biology and Health Informatics: 1-9.
Ishwaran H, Kogalur UB, Blackstone EH, Lauer MS (2008) Random survival forests. Annals of Applied Statistics 2: 841-860.
Gensheimer MF, Narasimhan B (2019) A scalable discrete-time survival model for neural networks. PeerJ 7: e6257.
Ogino S, Stampfer MJ (2010) Lifestyle factors and microsatellite instability in colorectal cancer: The evolving field of molecular pathological epidemiology. Journal of the National Cancer Institute 102: 365-367.
Ogino S, Chan AT, Fuchs CS, Giovannucci E (2011) Molecular pathological epidemiology of colorectal neoplasia: An emerging transdisciplinary and interdisciplinary field. Gut 60: 397-411.
Ogino S, Lochhead P, Chan AT, Nishihara R, Cho E, et al. (2013) Molecular pathological epidemiology of epigenetics: Emerging integrative science to analyze environment, host and disease. Modern Pathology 26: 465-484.
Ogino S, Ugai S, Hamada T, Ugai T (2025) The early-onset cancer epidemics: Evidence synthesis using the prospective cohort incident-tumor biobank method. European Journal of Epidemiology 40: 1405-1417.
Horvitz DG, Thompson DJ (1952) A generalization of sampling without replacement from a finite universe. Journal of the American Statistical Association 47: 663-685.
Rosenbaum PR, Rubin DB (1983) The central role of the propensity score in observational studies for causal effects. Biometrika 70: 41-55.
Robins JM, Rotnitzky A, Zhao LP (1994) Estimation of regression coefficients when some regressors are not always observed. Journal of the American Statistical Association 89: 846-866.
Hernán MA, Hernández-Díaz S, Robins JM (2004) A structural approach to selection bias. Epidemiology 15: 615-625.
Cole SR, Hernán MA (2008) Constructing inverse probability weights for marginal structural models. American Journal of Epidemiology 168: 656-664.
Robins JM, Hernán MA, Brumback B (2000) Marginal structural models and causal inference in epidemiology. Epidemiology 11: 550-560.
Lin DY, Lee-Jen W (1989) The robust inference for the Cox proportional hazards model. Journal of the American Statistical Association 84: 1074-1078.
Binder DA (1992) Fitting Cox’s proportional hazards models from survey data. Biometrika 79: 139-147.
Shu D, Young JG, Toh S, Wang R (2021) Variance estimation in inverse probability weighted Cox models. Biometrics 77: 1101-1117.
Liu L, Nevo D, Nishihara R, Cao Y, Song M, et al. (2018) Utility of inverse probability weighting in molecular pathological epidemiology. European Journal of Epidemiology 33: 381-392.
Lunn M, McNeil D (1995) Applying Cox regression to competing risks. Biometrics 51: 524-532.
Wang M, Kuchiba A, Ogino S (2015) A meta-regression method for studying etiological heterogeneity across disease subtypes classified by multiple biomarkers. American Journal of Epidemiology 182: 263-270.
Chen RJ, Ding T, Lu MY, Williamson DFK, Jaume G, et al. (2024) Towards a general-purpose foundation model for computational pathology. Nature Medicine 30: 850-862.
Filiot A, Ghermi R, Olivier A, Jacob P, Fidon L, et al. (2023) Scaling self-supervised learning for histopathology with masked image modeling.
Xu H, Usuyama N, Bagga J, Zhang S, Rao R, et al. (2024) A whole-slide foundation model for digital pathology from real-world data. Nature 630: 181-188.
Lu MY, Chen B, Williamson DFK, Chen RJ, Liang I, et al. (2024) A visual-language foundation model for computational pathology. Nature Medicine 30: 863-874.
Wang X, Yang S, Zhang J, Wang M, Zhang J, et al. (2022) Transformer-based unsupervised contrastive learning for histopathological image classification. Medical Image Analysis 81: 102559.
Zhou J, Wei C, Wang H, Shen W, Xie C, et al. (2022) iBOT: Image BERT pre-training with online tokenizer. International Conference on Learning Representations.
Lu MY, Williamson DFK, Chen TY, Chen RJ, Barbieri M, Mahmood F (2021) Data-efficient and weakly supervised computational pathology on whole-slide images. Nature Biomedical Engineering 5: 555-570.
Ilse M, Tomczak JM, Welling M (2018) Attention-based deep multiple instance learning. 35th International Conference on Machine Learning 80: 2127-2136.
Campanella G, Hanna MG, Geneslaw L, Miraflor A, Silva VWK, et al. (2019) Clinical-grade computational pathology using weakly supervised deep learning on whole slide images. Nature Medicine 25: 1301-1309.
Kather JN, Pearson AT, Halama N, Jäger D, Krause J, et al. (2019) Deep learning can predict microsatellite instability directly from histology in gastrointestinal cancer. Nature Medicine 25: 1054-1056.
Sirinukunwattana K, Domingo E, Richman SD, Redmond KL, Blake A, et al. (2021) Image-based Consensus Molecular Subtype (imCMS) classification of colorectal cancer using deep learning. Gut 70: 544-554.
Pei-Chen T, Tsung-Hua L, Kun-Chi K, Fang-Yi S, Tsung-Lu ML, et al. (2023) Histopathology images predict multi-omics aberrations and prognoses in colorectal cancer patients. Nature Communications 14: 2102.
Chen RJ, Lu MY, Wang J, Williamson DFK, Rodig SJ, et al. (2022) Pathomic fusion: An integrated framework for fusing histopathology and genomic features for cancer diagnosis and prognosis. IEEE Transactions on Medical Imaging 41: 757-770.
Chen RJ, Lu My, Williamson DFK, Chen TY, Lipkova J, et al. (2022) Pan-cancer integrative histology-genomic analysis via multimodal deep learning. Cancer Cell 40: 865-878.
Moertel CG, Fleming TR, Macdonald JS, Haller DG, Laurie JA, et al. (1990) Levamisole and fluorouracil for adjuvant therapy of resected colon carcinoma. New England Journal of Medicine 322: 352-358.
Drysdale E (2022) SurvSet: An open-source time-to-event dataset repository.
Harrell FE, Califf RM, Pryor DB, Lee KL, Rosati RA (1982) Evaluating the yield of medical tests. JAMA 247: 2543-2546.
The Cancer Genome Atlas Network (2012) Comprehensive molecular characterization of human colon and rectal cancer. Nature 487: 330-337.
Uno H, Cai T, Pencina MJ, D’Agostino RB, Wei LJ (2011) On the C-statistics for evaluating overall adequacy of risk prediction procedures with censored survival data. Statistics in Medicine 30: 1105-1117.
Weisenberger DJ, Siegmund KD, Campan M, Young J, Long TI, et al. (2006) CpG island methylator phenotype underlies sporadic microsatellite instability and is tightly associated with BRAF mutation in colorectal cancer. Nature Genetics 38: 787-793.
Guinney J, Dienstmann R, Wang X, Reyniès A, Schlicker A, et al. (2015) The consensus molecular subtypes of colorectal cancer. Nature Medicine 21: 1350-1356.
Graf E, Schmoor C, Sauerbrei W, Schumacher M (1999) Assessment and comparison of prognostic classification schemes for survival data. Statistics in Medicine 18: 2529-2545.
Vickers AJ, Elkin EB (2006) Decision curve analysis: A novel method for evaluating prediction models. Medical Decision Making 26: 565-574.
Collins GS, Moons KGM, Dhiman P, Riley RD, Beam AL, et al. (2024) TRIPOD+AI statement: Updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ 385: e078378.
