Tabular Foundation Model for Breast Cancer Prognosis using Gene Expression Data

Tabular Foundation Model for Breast Cancer Prognosis using Gene Expression Data | medRxiv /* */ /* */ <!-- <!-- /*! * yepnope1.5.4 * (c) WTFPL, GPLv2 */ (function(a,b,c){function d(a){return"[object Function]"==o.call(a)}function e(a){return"string"==typeof a}function f(){}function g(a){return!a||"loaded"==a||"complete"==a||"uninitialized"==a}function h(){var a=p.shift();q=1,a?a.t?m(function(){("c"==a.t?B.injectCss:B.injectJs)(a.s,0,a.a,a.x,a.e,1)},0):(a(),h()):q=0}function i(a,c,d,e,f,i,j){function k(b){if(!o&&g(l.readyState)&&(u.r=o=1,!q&&h(),l.onload=l.onreadystatechange=null,b)){"img"!=a&&m(function(){t.removeChild(l)},50);for(var d in y[c])y[c].hasOwnProperty(d)&&y[c][d].onload()}}var j=j||B.errorTimeout,l=b.createElement(a),o=0,r=0,u={t:d,s:c,e:f,a:i,x:j};1===y[c]&&(r=1,y[c]=[]),"object"==a?l.data=c:(l.src=c,l.type=a),l.width=l.height="0",l.onerror=l.onload=l.onreadystatechange=function(){k.call(this,r)},p.splice(e,0,u),"img"!=a&&(r||2===y[c]?(t.insertBefore(l,s?null:n),m(k,j)):y[c].push(l))}function j(a,b,c,d,f){return q=0,b=b||"j",e(a)?i("c"==b?v:u,a,b,this.i++,c,d,f):(p.splice(this.i++,0,a),1==p.length&&h()),this}function k(){var a=B;return a.loader={load:j,i:0},a}var l=b.documentElement,m=a.setTimeout,n=b.getElementsByTagName("script")[0],o={}.toString,p=[],q=0,r="MozAppearance"in l.style,s=r&&!!b.createRange().compareNode,t=s?l:n.parentNode,l=a.opera&&"[object Opera]"==o.call(a.opera),l=!!b.attachEvent&&!l,u=r?"object":l?"script":"img",v=l?"script":u,w=Array.isArray||function(a){return"[object Array]"==o.call(a)},x=[],y={},z={timeout:function(a,b){return b.length&&(a.timeout=b[0]),a}},A,B;B=function(a){function b(a){var a=a.split("!"),b=x.length,c=a.pop(),d=a.length,c={url:c,origUrl:c,prefixes:a},e,f,g;for(f=0;f<d;f++)g=a[f].split("="),(e=z[g.shift()])&&(c=e(c,g));for(f=0;f<b;f++)c=x[f](c);return c}function g(a,e,f,g,h){var i=b(a),j=i.autoCallback;i.url.split(".").pop().split("?").shift(),i.bypass||(e&&(e=d(e)?e:e[a]||e[g]||e[a.split("/").pop().split("?")[0]]),i.instead?i.instead(a,e,f,g,h):(y[i.url]?i.noexec=!0:y[i.url]=1,f.load(i.url,i.forceCSS||!i.forceJS&&"css"==i.url.split(".").pop().split("?").shift()?"c":c,i.noexec,i.attrs,i.timeout),(d(e)||d(j))&&f.load(function(){k(),e&&e(i.origUrl,h,g),j&&j(i.origUrl,h,g),y[i.url]=2})))}function h(a,b){function c(a,c){if(a){if(e(a))c||(j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}),g(a,j,b,0,h);else if(Object(a)===a)for(n in m=function(){var b=0,c;for(c in a)a.hasOwnProperty(c)&&b++;return b}(),a)a.hasOwnProperty(n)&&(!c&&!--m&&(d(j)?j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}:j[n]=function(a){return function(){var b=[].slice.call(arguments);a&&a.apply(this,b),l()}}(k[n])),g(a[n],j,b,n,h))}else!c&&l()}var h=!!a.test,i=a.load||a.both,j=a.callback||f,k=j,l=a.complete||f,m,n;c(h?a.yep:a.nope,!!i),i&&c(i)}var i,j,l=this.yepnope.loader;if(e(a))g(a,0,l,0);else if(w(a))for(i=0;i (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0];var j=d.createElement(s);var dl=l!='dataLayer'?'&l='+l:'';j.src='//www.googletagmanager.com/gtm.js?id='+i+dl;j.type='text/javascript';j.async=true;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-P4HH5NV'); Skip to main content Home About Submit ALERTS / RSS Search for this keyword Advanced Search Tabular Foundation Model for Breast Cancer Prognosis using Gene Expression Data View ORCID Profile Tuyen Vu , Ha X. Tran , View ORCID Profile Xiaomei Li , View ORCID Profile Lin Liu , View ORCID Profile Jiuyong Li , View ORCID Profile Jia Tina Du , View ORCID Profile Thuc D. Le doi: https://doi.org/10.1101/2025.10.03.25337265 Tuyen Vu 1 UniSA STEM, University of South Australia , Adelaide, SA, Australia Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Tuyen Vu Ha X. Tran 2 Accenture , Adelaide, SA, Australia Find this author on Google Scholar Find this author on PubMed Search for this author on this site Xiaomei Li 3 Commonwealth Scientific and Industrial Research Organisation (CSIRO) , Marsfield, NSW, Australia Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Xiaomei Li Lin Liu 1 UniSA STEM, University of South Australia , Adelaide, SA, Australia Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Lin Liu Jiuyong Li 1 UniSA STEM, University of South Australia , Adelaide, SA, Australia Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Jiuyong Li Jia Tina Du 1 UniSA STEM, University of South Australia , Adelaide, SA, Australia 4 School of Information and Communication Studies, Charles Sturt University , NSW, Australia Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Jia Tina Du Thuc D. Le 1 UniSA STEM, University of South Australia , Adelaide, SA, Australia Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Thuc D. Le For correspondence: thuc.le{at}unisa.edu.au Abstract Full Text Info/History Metrics Data/Code Preview PDF Abstract Survival analysis is essential in oncology for modeling time-to-event outcomes such as overall survival and disease recurrence. Traditional approaches, such as the Cox Proportional Hazards (CPH) model, have been widely used due to their interpretability but rely on restrictive assumptions of linearity and proportional hazards, which limit their ability to capture the nonlinear relationships present in high-dimensional genomic data. Recent machine learning methods, including Random Survival Forests (RSF) and deep learning models such as DeepSurv, have improved flexibility and predictive performance but require extensive training data, hyperparameter tuning, and computationally expensive optimization, which hinder their practical use. We propose TabSurv, a novel survival prediction framework that leverages a foundation model for tabular data tasks using in-context learning. TabSurv predicts survival times in a regression setting, allowing rapid adaptation to new datasets with minimal computational cost. The model is trained using only uncensored samples and evaluated with the concordance index (C-index) and stability metrics to assess both accuracy and robustness. We benchmark TabSurv against seven state-of-the-art survival models across 12 breast cancer datasets. The results demonstrate that TabSurv achieves competitive or superior performance, obtaining the best C-index on six datasets and the highest overall stability score. These findings highlight TabSurv as a powerful and efficient tool for breast cancer prognosis using high-dimensional molecular data. 1 Introduction Survival analysis plays a crucial role in oncology research and clinical practice by modeling time-to-event outcomes such as overall survival, disease-free survival, or recurrence risk. 1 , 2 Accurate breast cancer prognosis can significantly tailor treatments to individual patients, thereby improving survival outcomes. Traditionally, breast cancer prognosis has predominantly relied on clinical features such as tumor size, patient age, lymph node status, and histological grading. However, tumors exhibiting identical clinical characteristics may vary substantially in both prognosis and treatment responses due to the underlying molecular heterogeneity of breast cancer, driven by diverse mechanisms of carcinogenesis and tumor development. 3 , 4 Therefore, incorporating gene expression data to perform breast cancer prognosis at the molecular level has become highly desirable to address these complexities effectively. Various machine learning-based survival models, such as random survival forests (RSF) 5 and deep neural networks like DeepSurv, 6 have been developed to capture complex nonlinear interactions between features and hazard risks for breast cancer patients. These models have improved flexibility and predictive accuracy compared to traditional Cox Proportional Hazards (CPH) 1 models but require extensive training data, hyperparameter tuning, and computationally expensive optimization processes, limiting their real-world application. To overcome these limitations, we introduce TabSurv , a novel survival prediction framework that leverages a foundation model specifically adapted for tabular survival data. TabSurv builds upon TabPFN, 7 a transformer-based model originally designed for rapid and efficient zero-shot prediction through in-context learning on tabular datasets. TabPFN approximates Bayesian inference by learning to implicitly perform posterior predictive inference through exposure to millions of synthetic classification tasks, capturing realistic causal structures and data-generating processes. We have modified TabPFN to create TabSurv, adapting it specifically for survival data analysis and prognostic modeling in breast cancer. TabSurv predicts survival times directly in a regression setting, enabling swift adaptation to new datasets without extensive computational resources. The main contributions of this paper are: Development of TabSurv, a novel adaptation of the TabPFN foundation model designed specifically for survival analysis and breast cancer prognosis. Demonstration of TabSurv’s ability to predict breast cancer patient outcomes accurately using high-dimensional gene expression data, validated on 12 diverse breast cancer datasets. Benchmarking of TabSurv against state-of-the-art survival models, including Random Survival Forest, DeepSurv, Logistic Hazard, PMF, DeepHit, PCHazard, and MTLR, showing competitive or superior performance while requiring minimal task-specific training, thereby highlighting its robustness and efficiency. Through these contributions, TabSurv presents a significant advancement in breast cancer prognostic modeling, facilitating personalized treatment decisions informed by molecular-level insights and demonstrating the promising integration of survival analysis and foundation models. The rest of this paper is structured as follows: Section 2 presents a review of related studies on survival analysis techniques, breast cancer prognosis approaches, and causal methods for survival data. Section 3 describes the proposed TabSurv methodology and its underlying algorithm. Section 4 reports the experimental findings on real-world breast cancer datasets, together with performance comparisons against state-of-the-art benchmarks. Section 5 concludes the study by highlighting the effectiveness of the proposed TabSurv framework. 2 Background and Related Work 2.1 Survival Analysis Methods The goal of survival analysis is to predict the time-to-event, defined as the time at which an event of interest occurs. 8 , 9 It can also be viewed as estimating the probability of an event occurring over a given time period. 10 Applications of survival analysis span diverse domains. In clinical research, the event often represents patient death, disease recurrence, or treatment failure. 6 , 11 , 12 In business analytics, it may be used to forecast customer churn or the timing of the next purchase. 13 , 14 For each subject, survival data typically include the observed time t i and an event indicator δ i ∈ {0, 1}, where δ i = 1 indicates that the event occurred at t i and δ i = 0 denotes censoring, meaning the true time-to-event is unknown beyond the observed period. 15 Censoring poses a significant challenge for predictive algorithms originally designed for fully observed outcomes. 16 Survival analysis methods are therefore designed to model time-to-event data while explicitly accounting for censoring during model estimation. The survival function S ( t ) represents the probability that a subject survives beyond a specified time t : 9 , 15 Another key quantity is the hazard function h ( t ), which quantifies the instantaneous risk of the event occurring at time t given that the subject has survived up to t : 17 The hazard function provides insight into the timing and dynamics of risk, whereas the survival function summarises the overall survival probability. The Cox Proportional Hazards (CoxPH) model 1 is one of the most widely used methods for estimating the hazard function while incorporating subject-specific covariates: where h 0 ( t ) is the baseline hazard function and β are the regression coefficients. CoxPH is semi-parametric, making no assumption about h 0 ( t ), but it requires the proportional hazards assumption and models covariate effects linearly. Machine learning (ML) and Deep Learning (DL) methods have been developed to address the limitations of classical approaches. The RSF, 5 an extension of Breiman’s Random Forests to right-censored data, builds an ensemble of survival trees on bootstrap samples and aggregates their predictions to estimate survival functions. DeepSurv 6 embeds a neural network within the CoxPH framework to learn complex non-linear covariate–risk relationships. Other architectures, such as DeepHit 10 and Neural Multi-Task Logistic Regression (N-MTLR), 18 , 19 frame survival prediction as a time-dependent classification or multi-task learning problem, directly estimating S ( t ) without proportional hazards assumptions. Despite advances in modelling flexibility and predictive accuracy, significant challenges remain. First, many models either rely on restrictive assumptions (e.g., proportional hazards) or require large, high-quality datasets to train effectively, which are often unavailable in clinical contexts. Second, deep learning and complex ensemble methods demand extensive hyperparameter tuning and substantial computational resources, which can hinder deployment in time-sensitive healthcare environments. Finally, robustness to dataset shifts, varying censoring rates, and heterogeneous patient populations remains limited. 2.2 Breast Cancer Prognosis Methods Breast cancer remains one of the most prevalent cancers among women worldwide and is a leading cause of cancer-related mortality. 20 Accurate prognosis estimation is essential for guiding treatment decisions and informing patients about likely outcomes. In clinical practice, prognosis is often used to tailor treatment strategies such as chemotherapy, hormone therapy, or targeted therapy to improve the survival outcomes. 21 Traditionally, prognostic assessment has relied on clinical and pathological features, including tumour size, lymph node status, histological grade, and receptor status (ER, PR, HER2). 22 Tools such as the Nottingham Prognostic Index (NPI) 23 combine these factors into risk scores. However, these tools assume linear effects and do not fully capture the molecular heterogeneity of breast cancer. Tumours with identical TNM stage 24 can differ substantially in prognosis and treatment response due to variations in their underlying biological mechanisms. With the advent of high-throughput technologies, large-scale transcriptomic profiling from microarray and RNA-sequencing platforms has revolutionised breast cancer prognosis. Gene expression signatures such as Oncotype DX, 25 MammaPrint, 26 and PAM50/Prosigna 27 provide refined stratification by identifying intrinsic molecular subtypes (e.g., Luminal A/B, HER2-enriched, Basal-like). Many computational models, particularly those based on Cox regression, 28 have been developed to exploit these profiles. Examples include rorS, 27 which applies ridge-penalised Cox regression to subtype-specific gene clusters, and EndoPredict, 29 which combines selected gene expression levels into weighted risk scores. ML and DL approaches have emerged as powerful alternatives for modelling the high-dimensional, non-linear relationships between molecular features and survival outcomes. Examples include Cox regression with elastic net regularisation, 30 RSF, 5 support vector regression for censored data (SVRc), 31 and deep neural models such as DeepSurv, 6 DeepHit, 10 and N-MTLR. 18 Many existing methods assume that training and test datasets share the same distribution, an assumption that is frequently violated in practice. In real-world scenarios, gene expression data may be generated using different experimental platforms, from patients in different geographic regions, or from cohorts with distinct stage distributions. 32 , 33 This distribution shift can lead to substantial degradation in predictive performance, even for models that perform well on internal validation. Ensuring stability, that is consistent predictive performance across datasets from different sources, is crucial for clinical utility. Approaches to improving stability include multi-cohort training with robust feature selection 21 , 34 and transfer learning techniques 35 , 36 that leverage external cancer datasets to learn domain-invariant representations. However, these approaches require access to multiple large, well-annotated datasets with survival outcomes, which are often difficult and time-consuming to obtain due to the need for long-term patient follow-up. 2.3 Causal Methods for Survival Data Traditional survival models primarily estimate time-to-event outcomes without explicitly accounting for the causal effects of treatments. In clinical decision-making, particularly in oncology, optimising treatment recommendations requires estimating how an individual patient’s survival outcome would change under different treatment options. This motivates the development of causal survival analysis methods, which aim to estimate individualised treatment effects and recommend the treatment that maximises the expected survival benefit. Causal inference for survival data is often framed under the potential outcomes framework, 37 , 38 where each patient is assumed to have counterfactual survival times corresponding to different treatment assignments. A common approach is to estimate counterfactual survival curves 39 , 40 for each treatment arm and select the treatment yielding the most favourable survival probability. This approach differs from standard predictive modelling because it explicitly targets the causal effect of treatment on survival, rather than only modelling observed associations. For example, DeepSurv 6 can be trained separately for each treatment arm to estimate treatmentspecific risk scores, enabling treatment comparisons at the individual level. Building on this idea, multi-task learning (MTL) frameworks 41 and deep counterfactual networks 42 jointly learn survival probabilities for multiple treatment arms within a shared architecture, improving statistical efficiency by sharing information across arms while retaining treatment-specific parameters. Another line of work applies uplift modelling or conditional average treatment effect estimation to censored data. Survival Causal Forests 43 extend the Causal Forest framework 44 to survival settings by adapting the splitting criterion to maximise heterogeneity in treatment effects while accounting for censoring. Similarly, Survival Causal Trees (SCT) 45 recursively partition patients into subgroups with distinct treatment effects by leveraging decision tree structures tailored for survival outcomes. More recent methods, such as 2SWSA, 46 incorporate reweighting strategies to mitigate the effects of censoring and remove conditional spurious correlations, thereby improving stability in survival analysis. These causal-based methods have shown promise for treatment recommendation, particularly when applied to structured clinical datasets with moderate feature dimensions. However, these methods struggle to scale to high-dimensional feature spaces, such as transcriptomic or wholegenome expression data, due to the curse of dimensionality, limited sample sizes, and the difficulty of modelling complex. 3 Methodology 3.1 Overview of TabSurv The gaps in the existing methods identified in Section 2 , including restrictive modelling assumptions, the need for large high-quality datasets, computationally expensive training, and reduced robustness under distribution shift, motivate the development of new survival prediction frameworks that are both data-efficient and generalisable. Existing machine learning and causal survival models often require extensive retraining and hyperparameter tuning on each new dataset, and their performance can deteriorate substantially when applied to unseen cohorts or high-dimensional molecular data, such as gene expression data. To this end, we propose TabSurv, a novel survival prediction framework that leverages a foundation model specifically adapted for tabular survival data. TabSurv builds on TabPFN, 7 a transformer-based Prior-Data Fitted Network (PFN) originally designed for small-scale tabular classification and regression tasks. TabPFN is uniquely suited for survival analysis in heterogeneous, low-data regimes because it replaces conventional training with an in-context learning mechanism, enabling zero-shot prediction without any gradient-based updates or hyperparameter tuning on new datasets. At its core, TabPFN approximates Bayesian inference over a wide distribution of tabular tasks. During pretraining, the model is exposed to millions of synthetically generated tasks that capture realistic statistical structures, such as non-linear feature–outcome relationships, heterogeneous treatment effects, and low-dimensional manifolds. This pretraining allows TabPFN to implicitly perform posterior predictive inference by conditioning on the provided context—comprising both training examples and features—and directly producing predictions for test samples. 7 , 47 , 48 In other words, given a new survival prediction task, TabPFN behaves like a Bayesian model updating its beliefs in real-time, without explicit retraining. In the TabSurv framework, survival analysis is formulated as a regression problem over uncensored observations, as discussed in Section 2 . The model is first provided with a set of training examples ( x i , t i ), where x i are covariates and t i is the observed survival time for uncensored patients. For counterfactual treatment recommendation, each test instance is duplicated across all candidate treatment arms, with treatment indicators modified accordingly while keeping all other covariates fixed. TabPFN processes these augmented instances in the same input context as the training data, producing counterfactual survival time estimates for each treatment option. By combining the interpretability and flexibility of survival analysis with the adaptability of a pretrained foundation model, TabSurv addresses key gaps in existing methods, offering a stable, efficient, and generalisable solution for high-dimensional and heterogeneous breast cancer prognosis tasks. 3.2 TabSurv Algorithm Algorithm 1 outlines the TabSurv workflow, which leverages a pretrained TabPFN foundation model to perform survival prediction by reframing the task as a regression problem over uncensored samples. Given a dataset 𝒟 = ( x i , t i , δ i ) i = 1 N , where x i ∈ ℝ d denotes the covariates, t i ∈ ℝ+ is the observed time, and δ i ∈ 0, 1 is the event indicator, TabSurv begins with uncensored sample selection. In this step, the dataset is filtered to retain only instances where the event of interest has occurred ( δ i = 1), producing a subset 𝒟 1 = ( x i , t i ) | δ i = 1. This ensures that the model is trained exclusively on fully observed time-to-event outcomes, thereby avoiding the complexities associated with directly modelling censored data. The uncensored subset 𝒟1 is then divided into training and testing sets, (𝒟train, 𝒟test). The covariates and observed times from 𝒟train serve as the context input to the pretrained TabPFN model. Unlike conventional machine learning approaches that require gradient-based optimisation, TabPFN operates via in-context learning, where the entire supervised learning task— consisting of training examples and test features—is provided as a single sequence to the transformer-based Prior-Data Fitted Network. This enables the model to produce predictions without parameter updates, retraining, or hyperparameter tuning. TabSurv aims to minimise the discrepancy between predicted and observed survival times in the training set. At inference time, the model predicts survival times for each test instance directly from the provided training context. Model performance is first evaluated using the concordance index, which measures how well the predicted survival times preserve the correct ordering of observed events. To assess robustness across heterogeneous datasets, TabSurv also computes a stability score that combines the average concordance index with its variability across multiple test sets. This metric rewards models that perform both accurately and consistently in diverse clinical scenarios. A higher stability score reflects strong average performance and low variability, which is critical for clinical applicability. Finally, the predicted survival times, evaluation metrics, and corresponding patient covariates from the test set are saved for downstream analyses, such as survival curve estimation, subgroup comparisons, and benchmarking against baseline methods. Algorithm 1 TabSurv: Foundation Model-Based Survival Analysis Download figure Open in new tab 3.3 Censoring Data Handling Censoring is a fundamental challenge in survival analysis, arising when the true event time for an individual is only partially observed. In right-censoring, the most common form, a patient’s follow-up period ends before the event of interest (e.g., death, recurrence) occurs. In such cases, we know only that the event time exceeds a certain threshold, but not its exact value. Formally, for each individual i , the data are represented as a triplet ( x i , t i , δ i ), where x i ∈ ℝ d denotes the covariates, t i ∈ ℝ + is the observed time (either the event time or censoring time), and δ i ∈ {0, 1} is the event indicator, with δ i = 1 if the event occurred and δ i = 0 if censored. The proposed TabSurv framework addresses censoring by reframing survival prediction as a regression task over uncensored observations. Only instances with δ i = 1, where the true event time is fully observed, are retained for training, forming a reduced but fully informative dataset: This approach avoids the complexities of modelling censored observations directly, such as estimating hazard or survival functions, and instead focuses on learning from complete time-to-event data. TabSurv leverages TabPFN, a pretrained transformer-based foundation model for tabular data, which performs few-shot regression through in-context learning. Unlike conventional survival models that require gradient-based optimisation, TabPFN conditions on the provided training examples and directly predicts survival times for unseen instances without parameter updates. This eliminates the need for retraining or hyperparameter tuning and allows effective learning from small datasets. While discarding censored samples can drastically reduce the number of training instances,often to a fraction of the original dataset, TabSurv maintains strong predictive performance. Most survival models, especially those relying on parametric or semi-parametric estimation, suffer sub-stantial performance degradation in this scenario due to the increased variance in parameter estimates caused by reduced sample size. In contrast, TabPFN benefits from a strong Bayesian inductive prior learned during large-scale pretraining. Formally, if we denote the model’s prediction as: Then the variance of is reduced not only by the sample size n but also by the prior p ( θ ), which encodes structural assumptions about plausible data-generating processes. This allows Tab-Surv to achieve a more favourable bias–variance trade-off than conventional estimators, even when n is small. This property is particularly advantageous in high-dimensional clinical datasets, such as gene expression profiles, where the number of features far exceeds the number of uncensored samples. By exploiting structural priors and in-context learning, TabSurv remains robust to extreme data sparsity, making it well-suited for real-world clinical settings where complete follow-up data are rare but personalised predictions are crucial. 3.4 Evaluation Metrics Evaluating cancer prognosis models requires metrics that capture not only their predictive accuracy but also their robustness across different datasets and clinical scenarios. Prognostic models must provide reliable risk stratification for patients, regardless of the diversity in patient populations and their varying clinical and molecular characteristics. The first evaluation method discussed here describes conventional performance metrics used to quantify the accuracy of survival predictions. We then introduce additional stability measures to assess how consistently a model performs across multiple test datasets, providing crucial insight into its generalisation and reliability in real-world applications. 3.4.1 Concordance Index The Concordance Index (C-index) 49 is one of the most widely used metrics in survival analysis to evaluate the discriminative power of prognostic models. It measures the ability of a model to correctly rank patients according to their risk of experiencing an event (e.g., death or disease recurrence). Formally, for each patient i , let Y i denote the true (but typically unobservable) survival time and r i denote the risk score predicted by the model. For a randomly selected pair of patients ( i, j ), if Y j r i ). The C-index is thus the proportion of all comparable pairs for which the ordering of predicted risk matches the ordering of actual survival times. However, survival data are often right-censored, meaning that for some patients the event of interest has not occurred by the end of follow-up. Instead of the true survival time Y i , we observe the follow-up time T i and an event indicator E i , where E i = 1 indicates the event occurred and E i = 0 denotes censoring. To account for censoring, the C-index is computed as: where I ( · ) is the indicator function. This formulation considers only those pairs where patient i experienced an event before patient j ’s follow-up ended, ensuring that comparisons are valid. The C-index ranges from 0 to 1, where 0.5 corresponds to random predictions, 1.0 represents perfect concordance between predicted risks and survival outcomes, and 0 indicates that the model systematically predicts risk in the opposite direction. Achieving a high C-index is critical in cancer prognosis because it reflects a model’s ability to accurately rank patients according to their likelihood of experiencing the event. 3.4.2 Stability Metrics While predictive accuracy is essential, a clinically useful prognostic model must also be robust and generalizable to different datasets that may differ in sample size, feature distribution, and outcome prevalence. To capture this notion of robustness, we introduce stability metrics that quantify the consistency of model performance across multiple independent test datasets. Consider a collection of M test datasets, denoted as D 1 , D 2 , …, D M . Let CI m be the C-index obtained by the model on dataset D m . We define the stability of the model based on both the average performance and its variability across datasets as: where mean( · ) denotes the average C-index and sd( · ) denotes the standard deviation across the M datasets. A higher stability score indicates that a model achieves not only strong average predictive performance but also low variability when evaluated on diverse datasets. Models with high stability are more reliable for clinical applications because they are less sensitive to dataset-specific factors such as cohort composition, data quality, or measurement variability. This stability formulation is inspired by, 50 which assessed model robustness by considering both prediction accuracy and error variability. By penalizing high variability, this metric rewards models that perform consistently well across heterogeneous clinical environments. In the context of cancer prognosis, a stable model is highly desirable as it implies that the model can be confidently applied to new patient populations without substantial loss in predictive accuracy. Consequently, both the C-index and the stability metric together provide a comprehensive evaluation of a prognostic method’s accuracy and reliability. 4. Experiment Results 4.1 Datasets We used 12 breast cancer gene expression datasets from 21 for this study. As shown in Table 1 , these datasets were compiled from various public sources and include a total of 4,428 breast cancer patients. The METABRIC dataset is accessible through the European Genomephenome Archive (EGA) under accession number EGAS00000000083 1 , with access restrictions. TCGA500 are provided by The Cancer Genome Atlas (TCGA) program 2 . Datasets in-cluding MAINZ, TRANSBIG, UPP, UNT , and NKI are obtained via Bioconductor 3 using the data packages breastCancerMAINZ, breastCancerTRANSBIG, breastCancerUPP, breastCancerUNT, and breastCancerNKI , respectively. The remaining datasets were retrieved from the Gene Ex-pression Omnibus (GEO) database 4 . View this table: View inline View popup Download powerpoint Table 1: Summary statistics of the datasets used in this study. The clinical endpoints across these datasets vary: relapse-free survival (RFS) is used in UPP, GSE6532, GEO, TCGA500, METABRIC , and UK ; distant metastasis-free survival (DMFS) in TRANS-BIG, UNT, MAINZ , and NKI ; and death or DMFS in HEL . All survival times are recorded in years, with maximum durations ranging from 5 to 29.6 years. Events of interest include disease recur-rence, metastasis, or death, with event rates ranging between 9.00% and 47.40%. Due to differences in experimental batches, microarray/RNA-sequencing platforms, and preprocessing pipelines, the datasets exhibit heterogeneous distributions in both covariates and outcomes. This dataset shift complicates the development of models that generalize well to new, unseen cohorts. 33 , 51 To address this challenge, we propose a robust cancer prognosis model designed to perform reliably under distributional variability. 4.2 Baselines To evaluate the effectiveness of our proposed method, we benchmark it against several wellestablished survival analysis models: RSF (Random Survival Forrest), 5 DeepSurv , 6 Logistic Hazard , 52 PMF (Probability Mass Function), 52 DeepHit , 10 PCHazard , 52 MTLR (Multi-Task Logistic Regression). 19 These baseline methods are applied to predict time-to-event outcomes for breast cancer prognosis, using both censored and uncensored clinical data. The benchmarking process involves standard preprocessing steps, including encoding categorical features, imputing missing values, and normalizing continuous variables. All models are trained to predict the survival time and are evaluated using concordance index (C-index) as the standard performance metric and the Stability Score metric to compare with the TabSurv. The following baseline survival models are used: RSF : 5 A non-parametric ensemble method that constructs multiple survival trees using bootstrap samples and aggregates them to estimate cumulative hazard functions. DeepSurv : 6 A deep neural network that extends the Cox proportional hazards model to capture nonlinear relationships between features and survival risk. Logistic Hazard (LH) : 52 A discrete-time survival model based on logistic regression that models the conditional probability of failure at each time interval. PMF : 52 A flexible discrete-time model that learns survival distributions directly by parameterizing the probability mass function over time intervals. DeepHitSingle (DeepHS) : 10 A fully parametric model that directly predicts the probability of an event occurring at each time point, optimized using a ranking loss and likelihood loss. PCHazard : 52 A piecewise constant hazard model implemented via neural networks, allowing for time-varying hazards across intervals. MTLR : 19 A discrete-time survival model that treats survival prediction as a series of binary classification problems across time bins, jointly learned via logistic regression. All deep learning-based models are implemented using the pycox library, 52 and RSF is implemented using sksurv library. 53 This diverse set of models ensures a fair and comprehensive comparison of predictive performance across different modeling paradigms. 4.3 Implementation We train the models on a high performance compute (HPC) facility with an NVIDIA GPU (Tesla V100S-PCIE-32 GB). The maximum wall time of the HPC is 8 hours, which means the model training need to be finished within this period of time. Our method is implemented using Python with the TabPFN model from TabPFN package 5 . The baselines are implemented using Python with LogisticHazard (LH), PMF, DeepHitSingle (DeepHS), PCHazard, MTLR, CoxPH models from pycox package 6 , and RandomSurvivalForest from sksurv package 7 . Codes and public datasets are available for public access at GitHub 8 . 4.4 Model Performance Comparisons 4.4.1 Concordance Index Table 2 summarizes the benchmarking results of various survival models across multiple datasets using the concordance index (C-index) as the evaluation metric. It presents the C-index values of the proposed method TabSurv and seven baseline survival models ( LogisticHazard, PMF, DeepHitSingle, PCHazard, MTLR, DeepSurv , and RSF ) across 12 breast cancer datasets. The best result for each dataset is highlighted in bold. View this table: View inline View popup Download powerpoint Table 2: Benchmarking Survival Models Across 12 Datasets using Concordance Index (C-index). As observed, TabSurv achieves the best C-index on six datasets ( METABRIC, GSE19783, NKI, TRANSBIG, MAINZ , and UPP ). The other baseline methods achieve the highest score on at most three datasets: RSF on four datasets ( METABRIC, GEO, GSE6532, UK ), DeepSurv on one dataset ( HEL ), DeepHitSingle on one dataset ( UNT ) and MTLR on one dataset ( TCGA500 ). A good survival model should achieve a C-index greater than 0.5 on each dataset, indicating that predicted risk scores are concordant with real survival outcomes. From Table 2 , most models meet this threshold on the majority of datasets. However, TabSurv demonstrates the most consistent performance, with C-indices exceeding 0.7 on several datasets. Some baseline methods, such as DeepHitSingle, have C-index values below 0.5 on certain datasets (e.g., GSE19783, MAINZ ), indicating poor concordance between predicted risks and patient outcomes. These inconsistent results may arise from distributional differences across datasets, which traditional survival models struggle to address. 4.4.2 Stability Metric Table 3 presents the mean C-index, standard deviation, and stability score of each model across all datasets. The stability metric is defined as the mean C-index minus its standard deviation, capturing both predictive accuracy and robustness. View this table: View inline View popup Download powerpoint Table 3: Comparison of the Stability Score of all Models. TabSurv achieves the highest stability score ( Stability = 0.6139), outperforming all baseline methods. This demonstrates that TabSurv provides not only strong predictive accuracy but also reliable generalisation across diverse datasets. Although some models, such as PMF and RSF , show relatively low variability, their mean C-index values are lower, limiting their overall utility. These results confirm that TabSurv consistently delivers accurate and stable predictions across heterogeneous breast cancer cohorts. 5 Conclusion In this study, we introduced TabSurv, a novel survival prediction framework that leverages a tabular foundation model to perform time-to-event prediction in a regression setting. Unlike conventional survival models that require task-specific training and extensive hyperparameter tuning, TabSurv utilises in-context learning to make zero-shot predictions, enabling rapid adaptation to new datasets with minimal computational overhead. By training only on uncensored samples and directly predicting survival times, TabSurv effectively handles high-dimensional genomic data without the need for complex survival-specific architectures. Extensive benchmarking on 12 breast cancer datasets demonstrated that TabSurv achieves competitive or superior predictive performance compared to state-of-the-art survival models, including Random Survival Forests, DeepSurv, Logistic Hazard, PMF, DeepHit, PCHazard, and MTLR. TabSurv not only attained the best C-index on multiple datasets but also achieved the highest overall stability score, indicating both strong predictive accuracy and consistent performance across diverse cohorts. These results highlight the potential of foundation models as a promising direction for survival analysis in oncology. By combining in-context learning with high-dimensional gene expression data, TabSurv provides a practical and efficient solution for prognostic modeling in precision medicine. Future work will explore extending TabSurv to integrate multi-omics data, incorporate treatment effect estimation, and evaluate its applicability in prospective clinical decision-support settings. Data Availability All data produced in the present study are available upon reasonable request to the authors https://www.cancer.gov/tcga https://www.ebi.ac.uk/ega/ Footnotes 1 https://www.ebi.ac.uk/ega/ 2 https://www.cancer.gov/tcga 3 https://bioconductor.org/ 4 https://www.ncbi.nlm.nih.gov/geo/ 5 https://github.com/PriorLabs/TabPFN 6 https://github.com/havakv/pycox 7 https://scikit-survival.readthedocs.io/ 8 https://github.com/vntuyen/tabsurv References [1]. ↵ Cox , D. R. Journal of the Royal Statistical Society: Series B (Methodological) 1972 , 34 , 202 –. OpenUrl [2]. ↵ Therneau , T. M. ; Grambsch , P. M. Modeling Survival Data: Extending the Cox Model; Springer Science & Business Media , 2000 . [3]. ↵ Bair , E. ; Tibshirani , R. PLoS Biology 2004 , 2 , e108 . OpenUrl CrossRef PubMed [4]. ↵ Royston , P. ; Altman , D. G. BMC Medical Research Methodology 2013 , 13 , 15 –. OpenUrl PubMed [5]. ↵ Ishwaran , H. ; Kogalur , U. B. ; Blackstone , E. H. ; Lauer , M. S. The Annals of Applied Statistics 2008 , 2 , 860 –. OpenUrl [6]. ↵ Katzman , J. L. ; Shaham , U. ; Cloninger , A. ; Bates , J. ; Jiang , T. ; Kluger , Y. BMC Medical Research Methodology 2018 , 18 , 24 . OpenUrl CrossRef PubMed [7]. ↵ Hoyer , S. ; Arjona-Medina , J. A. ; Zimmer , L. ; Pfisterer , F. ; Nieke , T. ; Mohr , F. ; Hutter , F. ; Lindauer , M. arXiv preprint arxiv: 2207.01848 2023 , [8]. ↵ Wu , M. ; Zhang , Y. ; Davidian , M. ; Tsiatis , A. A. Journal of the American Statistical Association 2015 , 110 , 251 –. OpenUrl [9]. ↵ Lee , E. T. ; Wang , J. W. Statistical Methods for Survival Data Analysis ; John Wiley & Sons , 2003 . [10]. ↵ Lee , C. ; Zame , W. R. ; Yoon , J. ; van der Schaar , M. Deephit: A deep learning approach to survival analysis with competing risks . Proceedings of the AAAI Conference on Artificial Intelligence . 2018 . [11]. ↵ Zhu , H. ; Ghosh , D. ; Cho , H. IEEE journal of biomedical and health informatics 2017 , 22 , 1600 –. OpenUrl [12]. ↵ Luck , M. ; Sylvain , T. ; Cardinal , H. ; Lodi , A. ; Bengio , Y. arXiv preprint arxiv: 1708.04649 2017 , [13]. ↵ Jing , L. ; Smola , A. J. T-PARE: temporal pattern augmented ranking estimation . Proceedings of the 26th International Conference on World Wide Web Companion . 2017 ; pp 1025 – 1034 . [14]. ↵ Grob , E. ; Wuthrich , K. ; Engel , P. Customer lifetime value prediction using embeddings . Proceedings of the Eleventh ACM International Conference on Web Search and Data Mining . 2018 ; pp 586 – 594 . [15]. ↵ Klein , J. P. ; Moeschberger , M. L. Survival Analysis: Techniques for Censored and Truncated Data ; Springer Science & Business Media , 2003 . [16]. ↵ Wang , W. ; Li , S. ; Dagne , G. Statistical Methods in Medical Research 2019 , 28 , 2309 –. OpenUrl [17]. ↵ Dunn , G. ; Clark , V. A. Applied Regression Analysis ; John Wiley & Sons , 2009 . [18]. ↵ Fotso , S. arXiv preprint arxiv: 1801.05512 2018 , [19]. ↵ Yu , C. ; Greiner , R. ; Lin , H. ; Baracos , V. Learning patient-specific cancer survival distributions as a sequence of dependent regressors . Advances in Neural Information Processing Systems . 2011 . [20]. ↵ Siegel , R. L. ; Miller , K. D. ; Fuchs , H. E. ; Jemal , A. CA: A Cancer Journal for Clinicians 2022 , 72 , 33 –. OpenUrl [21]. ↵ Li , X. ; Liu , L. ; Li , J. ; Le , T. D. IEEE Transactions on Computational Biology and Bioinformatics 2025 , [22]. ↵ Cuzick , J. ; Dowsett , M. ; Pineda , S. ; others Breast Cancer Research and Treatment 2011 , 125 , 289 –. OpenUrl [23]. ↵ Galea , M. H. ; Blamey , R. W. ; Elston , C. E. ; Ellis , I. O. Breast cancer research and treatment 1992 , 22 , 219 –. OpenUrl [24]. ↵ Park , Y. ; Lee , S. ; Cho , E. ; Choi , Y. L. ; Lee , J. ; Nam , S. ; Yang , J.-H. ; Shin , J. ; Ko , E. ; Han , B.-K. ; others Annals of Oncology 2011 , 22 , 1560 –. OpenUrl [25]. ↵ Paik , S. ; others New England Journal of Medicine 2004 , 351 , 2826 –. OpenUrl [26]. ↵ van ‘t Veer , L. J. ; others Nature 2002 , 415 , 536 –. OpenUrl [27]. ↵ Parker , J. S. ; Mullins , M. ; Cheang , M. C. ; others Journal of Clinical Oncology 2009 , 27 , 1167 –. OpenUrl [28]. ↵ Cox , D. R. Journal of the Royal Statistical Society. Series B (Methodological) 1972 , 34 , 220 –. OpenUrl [29]. ↵ Filipits , M. ; Rudas , M. ; Jakesz , R. ; Dubsky , P. ; Fitzal , F. ; Singer , C. F. ; Dietze , O. ; Greil , R. ; Jelen , A. ; Sevelda , P. ; others Clinical Cancer Research 2011 , 17 , 6020 –. OpenUrl [30]. ↵ Simon , N. ; Friedman , J. ; Hastie , T. ; Tibshirani , R. Journal of statistical software 2011 , 39 , 13 –. OpenUrl [31]. ↵ Khan , F. M. ; Zubek , V. B. Support vector regression for censored data (SVRc): a novel tool for survival analysis. 2008 Eighth IEEE International Conference on Data Mining . 2008 ; pp 863 – 868 . [32]. ↵ Shen , Z. ; Liu , J. ; He , Y. ; Zhang , X. ; Xu , R. ; Yu , H. ; Cui , P. arXiv preprint arxiv: 2108.13624 2021 , [33]. ↵ Wang , J. ; Lan , C. ; Liu , C. ; Ouyang , Y. ; Qin , T. ; Lu , W. ; Chen , Y. ; Zeng , W. ; Yu , P. S. IEEE transactions on knowledge and data engineering 2022 , 35 , 8072 –. OpenUrl [34]. ↵ Abraham , G. ; Kowalczyk , A. ; Loi , S. ; Haviv , I. ; Zobel , J. BMC bioinformatics 2010 , 11 , 15 –. OpenUrl CrossRef PubMed [35]. ↵ Li , Y. ; Wang , L. ; Wang , J. ; Ye , J. ; Reddy , C. K. Transfer learning for survival analysis via efficient l2, 1-norm regularized cox regression . 2016 IEEE 16th International Conference on Data Mining (ICDM) . 2016 ; pp 231 – 240 . [36]. ↵ Kim , S. ; Kim , K. ; Choe , J. ; Lee , I. ; Kang , J. Bioinformatics 2020 , 36 , i389 – i398 . OpenUrl CrossRef PubMed [37]. ↵ Rubin , D. B. Journal of Educational Psychology 1974 , 66 , 701 –. OpenUrl [38]. ↵ Hernán , M. A. ; Robins , J. M. Causal Inference: What If ; Chapman & Hall/CRC , 2020 . [39]. ↵ Lee , C. ; Zame , W. R. ; Yoon , J. ; van der Schaar , M. Proceedings of the AAAI Conference on Artificial Intelligence 2018 , 32 . [40]. ↵ Zhao , Y. ; Li , Y. ; Li , X. ; others Deep Survival Forests for Counterfactual Predictions of Cancer Treatment Outcomes . Proceedings of the ACM Conference on Health, Inference, and Learning . 2021 . [41]. ↵ Zhang , Y. ; Chen , W. ; others IEEE Transactions on Pattern Analysis and Machine Intelligence 2020 , [42]. ↵ Schwab , P. ; Linhardt , L. ; Karlen , W. Learning counterfactual representations for estimating individual dose-response curves . AAAI Conference on Artificial Intelligence . 2020 . [43]. ↵ Lu , C. ; Tian , T. ; Li , L. ; others Journal of Machine Learning Research 2021 , 22 , 46 –. OpenUrl [44]. ↵ Wager , S. ; Athey , S. Journal of the American Statistical Association 2018 , 113 , 1242 –. OpenUrl [45]. ↵ Zhang , W. ; Le , T. D. ; Liu , L. ; Zhou , Z.-H. ; Li , J. Bioinformatics 2017 , 33 , 2378 –. OpenUrl [46]. ↵ Tran , H. X. ; Le , T. D. ; Li , J. ; Liu , L. ; Li , X. ; Liu , J. ; Waters , T. Stabilising job survival analysis for disability employment services in unseen environments . Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining . 2023 ; pp 4970 – 4980 . [47]. ↵ Hoyer , S. ; Lindauer , M. ; Hutter , F. arXiv preprint arxiv: 2207.01848 2023 , [48]. ↵ Künzel , S. R. ; Sekhon , J. S. ; Bickel , P. J. ; Yu , B. Proceedings of the National Academy of Sciences 2019 , 116 , 4165 –. OpenUrl [49]. ↵ Harrell , F. E. ; Lee , K. L. ; Mark , D. B. Statistics in medicine 1996 , 15 , 387 –. OpenUrl [50]. ↵ Kuang , K. ; Cui , P. ; Athey , S. ; Xiong , R. ; Li , B. Stable prediction across unknown environments . Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining . 2018 ; pp 1617 – 1626 . [51]. ↵ Liu , J. ; Shen , Z. ; He , Y. ; Zhang , X. ; Xu , R. ; Yu , H. ; Cui , P. arXiv preprint arxiv: 2108.13624 2021 , [52]. ↵ Kvamme , H. ; Borgan , . ; Scheel , I. Journal of Machine Learning Research 2019 , 20 , 30 –. OpenUrl [53]. ↵ Pölsterl , S. Journal of Machine Learning Research 2020 , 21 , 6 –. OpenUrl View the discussion thread. Back to top Previous Next Posted October 05, 2025. Download PDF Data/Code Email Thank you for your interest in spreading the word about medRxiv. NOTE: Your email address is requested solely to identify you as the sender of this article. Your Email * Your Name * Send To * Enter multiple addresses on separate lines or separate them with commas. You are going to email the following Tabular Foundation Model for Breast Cancer Prognosis using Gene Expression Data Message Subject (Your Name) has forwarded a page to you from medRxiv Message Body (Your Name) thought you would like to see this page from the medRxiv website. Your Personal Message CAPTCHA This question is for testing whether or not you are a human visitor and to prevent automated spam submissions. Share Tabular Foundation Model for Breast Cancer Prognosis using Gene Expression Data Tuyen Vu , Ha X. Tran , Xiaomei Li , Lin Liu , Jiuyong Li , Jia Tina Du , Thuc D. Le medRxiv 2025.10.03.25337265; doi: https://doi.org/10.1101/2025.10.03.25337265 Share This Article: Copy Citation Tools Tabular Foundation Model for Breast Cancer Prognosis using Gene Expression Data Tuyen Vu , Ha X. Tran , Xiaomei Li , Lin Liu , Jiuyong Li , Jia Tina Du , Thuc D. Le medRxiv 2025.10.03.25337265; doi: https://doi.org/10.1101/2025.10.03.25337265 Citation Manager Formats BibTeX Bookends EasyBib EndNote (tagged) EndNote 8 (xml) Medlars Mendeley Papers RefWorks Tagged Ref Manager RIS Zotero Tweet Widget Facebook Like Google Plus One Subject Area Health Informatics Subject Areas All Articles Addiction Medicine (568) Allergy and Immunology (863) Anesthesia (299) Cardiovascular Medicine (4425) Dentistry and Oral Medicine (443) Dermatology (382) Emergency Medicine (607) Endocrinology (including Diabetes Mellitus and Metabolic Disease) (1507) Epidemiology (15221) Forensic Medicine (30) Gastroenterology (1123) Genetic and Genomic Medicine (6588) Geriatric Medicine (667) Health Economics (997) Health Informatics (4524) Health Policy (1368) Health Systems and Quality Improvement (1612) Hematology (540) HIV/AIDS (1264) Infectious Diseases (except HIV/AIDS) (15910) Intensive Care and Critical Care Medicine (1103) Medical Education (623) Medical Ethics (145) Nephrology (667) Neurology (6588) Nursing (346) Nutrition (998) Obstetrics and Gynecology (1143) Occupational and Environmental Health (956) Oncology (3331) Ophthalmology (970) Orthopedics (369) Otolaryngology (420) Pain Medicine (435) Palliative Medicine (129) Pathology (663) Pediatrics (1690) Pharmacology and Therapeutics (691) Primary Care Research (710) Psychiatry and Clinical Psychology (5440) Public and Global Health (9220) Radiology and Imaging (2195) Rehabilitation Medicine and Physical Therapy (1369) Respiratory Medicine (1196) Rheumatology (593) Sexual and Reproductive Health (710) Sports Medicine (529) Surgery (710) Toxicology (99) Transplantation (289) Urology (265) (function(){function c(){var b=a.contentDocument||a.contentWindow.document;if(b){var d=b.createElement('script');d.innerHTML="window.__CF$cv$params={r:'9ffd28719af752ad',t:'MTc3OTQ2NjkyOA=='};var a=document.createElement('script');a.src='/cdn-cgi/challenge-platform/scripts/jsd/main.js';document.getElementsByTagName('head')[0].appendChild(a);";b.getElementsByTagName('head')[0].appendChild(d)}}if(document.body){var a=document.createElement('iframe');a.height=1;a.width=1;a.style.position='absolute';a.style.top=0;a.style.left=0;a.style.border='none';a.style.visibility='hidden';document.body.appendChild(a);if('loading'!==document.readyState)c();else if(window.addEventListener)document.addEventListener('DOMContentLoaded',c);else{var e=document.onreadystatechange||function(){};document.onreadystatechange=function(b){e(b);'loading'!==document.readyState&&(document.onreadystatechange=e,c())}}}})();