Background and objective. Gleason grading of prostate cancer is a critical clinical task inherently defined by an ordinal progression of tissue deterioration. However, most automated computational approaches simplify this problem into a standard multi-class classification problem, which ignores its ordinal structure. Furthermore, the suitability of foundation models—which currently provide state-of-the-art latent representations of histological images— for ordinal decision making remains underexplored. Specifically, it is unclear how model design should adapt when these embeddings fail to naturally capture disease progression. Methods: In this context, we investigate the concept of rankability, defined as the degree to which the embedding space of a foundation model inherently preserves the natural ordering of Gleason grades. We systematically analyze how this property impacts downstream ordinal predictions. Consequently, we propose a novel framework that integrates pre-trained foundation model embeddings with Ordinal Gaussian Processes (OGPs). By explicitly incorporating an ordinal likelihood, our approach mathematically enforces the strictly ordered nature of disease severity within the probabilistic model. Results: Evaluated across two publicly available prostate cancer datasets exhibiting different levels of rankability, our experiments show the benefits of OGPs for ordinal problems. While both ordinal and standard multi-class approaches achieve similar performance in naturally highly rankable embedding spaces, the proposed OGP framework significantly outperforms its non-ordinal counterparts in low-rankability scenarios. In these challenging settings, explicitly integrating an ordinal likelihood yields substantial improvements in both predictive agreement metrics and calibration. Conclusion: Our findings highlight that explicitly modeling ordinality acts as a crucial safeguard in automated grading systems. When state-of-the-art foundation models fail to capture the sequential nature of disease progression, integrating an ordinal likelihood improves the predictive robustness and clinical reliability necessary for deploying computational pathology tools in practice.
Background Digital twin (DT) methodologies for type 1 diabetes require personalization of physiological models using continuous glucose monitoring and exogenous insulin data. The imbalance between data information content and model complexity often necessitates Bayesian estimation for model identification, requiring informative a priori parameter distributions that are difficult to obtain from the literature, especially for parameters describing gastrointestinal glucose absorption. This study aims to provide such priors by modeling inter-subject variability of meal glucose rate of appearance (Ra). Methods Four glucose absorption models used in published DT were formulated within a nonlinear mixed-effects (NLME) framework. Models were identified using Ra derived from gold-standard tracer data collected in 43 adults undergoing either a mixed meal (MTT) or an oral glucose tolerance test (OGTT). A stepwise, parsimony guided, model-building strategy was adopted to define the random-effects structure and incorporate covariate effects. Parameter estimation was performed using likelihood-based methods. Model adequacy was assessed through parameter precision and visual predictive checks. Results All models achieved adequate fits of Ra with acceptable parameter precision. Test type and body mass index were the covariates explaining most of the inter-subject variability. While all models well reproduced mixed-meal Ra, differences were observed in their representations of OGTT responses. For each model, fixed effects, random-effects covariance matrices, and covariate relationships were estimated, enabling the construction of ready-to-use prior parameter distributions for Bayesian DT implementations. Conclusions The proposed NLME-based framework provides physiologically grounded prior distributions that will support researchers in designing robust and physiologically consistent Bayesian DT algorithms.
Objective: Accurate and robust survival prediction is essential for personalised breast cancer prognosis and treatment decision-making. However, existing machine learning-based survival models often lack stability under cohort heterogeneity and distribution shift, while tabular foundation models typically do not support censored time-to-event data. This study aims to develop a foundation-model-based survival framework that explicitly handles right censoring and enables stable deployment across heterogeneous cohorts without task-specific retraining. Methods: We propose TabSurv, a foundation-model-based framework for survival analysis that adapts pretrained tabular foundation models to censored time-to-event prediction. Built on TabPFN, TabSurv operates via in-context learning without gradient-based training or hyperparameter tuning. To address censoring, TabSurv introduces a two-stage, censoring-aware strategy. In the first stage, the model learns from patients with observed events and generates multiple plausible survival times for each censored patient, with all imputed times occurring after the observed censoring time. In the second stage, the model uses each imputed dataset as context and makes predictions through in-context learning. These predictions are then combined to produce the final result while accounting for uncertainty in the imputed survival times. For treatment recommendation, TabSurv fits a TabPFN regressor in context to risk scores from a random survival forest, which handles censoring, and estimates each patient’s predicted risk under hypothetical treatment reassignment. Results: TabSurv was evaluated on twelve breast cancer gene expression cohorts under in-distribution and out-of-distribution settings. It achieved competitive or superior prognostic performance compared with eight established survival models and demonstrated the highest predictive stability under cohort heterogeneity. In exploratory treatment-recommendation experiments on METABRIC, patients whose treatment matched TabSurv’s recommendation showed the largest average restricted mean survival time advantage among the compared methods. Conclusion: TabSurv provides an efficient and stable foundation-model-based approach to censored survival analysis, supporting robust prognosis and personalised treatment recommendation in heterogeneous breast cancer cohorts.
