Preprint
Brief Report

This version is not peer-reviewed.

Limited Prognostic Value of DeepSurv Over Cox Regression for Survival Prediction in Glioblastoma: A TCGA Analysis

Submitted:

04 November 2026

Posted:

09 November 2026

You are already at the latest version

Abstract
As newer neural-network methods are applied to survival analysis, it is important to determine whether they improve prediction in relatively small clinical datasets. Our study compared model discrimination between conventional Cox regression and a DeepSurv neural network implemented in PyTorch to predict survival in historical TCGA glioblastoma patients with reported IDH-wildtype status. Both models used expression of the same four genes: OSMR, HOXC10, SCARA3, and SLC39A10. The primary measure of model performance was the C-index, with time-dependent AUC and Brier prediction error at one year as secondary measures. Age and sex were added in exploratory analyses to assess changes in performance. DeepSurv did not outperform conventional Cox regression in the main 139-patient four-gene analysis, returning a C-index of 0.567 compared with 0.591 for the refitted Cox model. The fixed published score developed by Cao et al. using the same genes returned a C-index of 0.614. Because the Cao score was developed using overlapping TCGA data, it served as a contextual benchmark rather than an independent validation. Adding age and sex yielded small numerical increases in the main-cohort C-index (DeepSurv, 0.577; Cox, 0.597). These findings show no observed advantage from the tested neural network over Cox regression in this internal comparison. Potential explanations that are not explored in this study but can be theorized are data missingness, overfitting, and limited data set size. The small historical cohort and lack of independent validation limit broader conclusions or clinical use at this time.
Keywords: 
;  ;  ;  

Introduction

Along with the increased use of clinicopathologic data in the diagnosis and prognosis of primary CNS tumors, there has been interest in generating risk scores for glioblastoma (GBM). Public resources such as The Cancer Genome Atlas (TCGA) have enabled retrospective studies combining genetic and clinical data to estimate survival risk. Many scores use traditional statistical methods, including Cox regression. For example, Cao et al. generated a four-gene risk score using OSMR, HOXC10, SCARA3, and SLC39A10 in TCGA GBM patients [1]. Other studies have explored neural-network survival models as an alternative way to relate these measurements to risk [2,3]. Such models can represent more complex relationships, but whether that additional flexibility improves prediction in a relatively small cohort remains uncertain. Our study aimed to explore the difference in prognostic performance between conventional Cox regression and a DeepSurv neural network implemented in PyTorch [4,5]. Sourcing from TCGA, we compared models built from the same four-gene expression values on the same patients and held-out validation groups. We then added age and recorded sex in exploratory analyses to observe changes in model performance. An important contextual benchmark was Cao et al.’s published four-gene Cox score, which was also developed using TCGA data. Because those data overlap our cohort, this score cannot serve as independent validation. Our study provides context for whether the greater complexity of a neural model yielded an observable benefit in this dataset.

Methods

Study Design and Objective

This portion of the study was designed as a retrospective survival-analysis model comparison using public TCGA data. The primary objective was to compare a DeepSurv neural Cox model implemented in PyTorch with conventional Cox regression using the same four gene-expression predictors and the same patients and held-out folds. The published Cao et al. four-gene score served as a fixed benchmark using overlapping TCGA data, not as an independent validation set. Further exploratory model comparisons included age and recorded sex. The outcome of interest was overall survival. Model performance was assessed primarily with Harrell’s C-index, with one-year time-dependent AUC and Brier prediction error as secondary measures.

Data Source and Patient Cohorts

Overall survival outcome and time were reconstructed from the TCGA Clinical Data Resource (TCGA-CDR). Gene expression came from UCSC Xena TCGA GBM HiSeqV2 and was used on its archived log2(RSEM normalized count + 1) scale. Inclusion required a primary-tumor specimen, a valid curated overall-survival endpoint, and values for all four genes. The expression dataset contained 172 specimens: 154 primary tumors, 13 recurrent tumors, and five normal-tissue specimens. One primary-tumor patient lacked a valid curated overall-survival endpoint, leaving 153 patients for the four-gene analysis. Of these, 139 had reported IDH-wildtype status and formed the main cohort. The broader sensitivity cohort included all 153 patients: the main 139, ten with reported IDH-mutant status, and four without a published IDH classification. These historical annotations do not constitute a new integrated pathology diagnosis.
Recurrent and normal-tissue specimens were excluded because the models were intended to predict survival from diagnosis using a primary tumor. Patient identifiers were checked before clinical and molecular datasets were linked, and one primary-tumor specimen per patient was retained. Among patients with a primary tumor and all four expression values, one was excluded because curated overall-survival information was unavailable. No additional patients were removed from the resulting 153-patient cohort for the four-gene analysis. The main cohort contains 139 of these historical TCGA-GBM patients with reported IDH-wildtype status.

Outcome and Time Origin

The outcome of interest is overall survival (OS) which uses TCGA-CDR OS and OS.time. Time is in days; death is coded as an event (1), and no death recorded during follow-up is coded as censored (0). Time origin is from time of diagnosis and endpoint is overall survival time, defined as death or the end of recorded observations.

Predictors and Data Preparation

The main comparison utilized four continuous gene-expression predictors based on the model created by Cao et al.: OSMR, HOXC10, SCARA3, and SLC39A10. Age and recorded sex were added in an exploratory extension. MGMT was not included in the reported model comparisons because source annotations were missing or conflicting for 32 of the 139 main-cohort patients. KPS was omitted because availability and timing were inconsistent. Post-diagnosis treatment was not treated as a baseline predictor.
Table 1. Gathered data and the published Cao score used or considered, with sources and dataset explanations.
Table 1. Gathered data and the published Cao score used or considered, with sources and dataset explanations.
Gathered Data/Model Explanation
Four gene data (OSMR, HOXC10, SCARA3, and SLC39A10) UCSC Xena TCGA GBM HiSeqV2; log2 RSEM normalized count + 1.
Published Cao score 0.279×OSMR + 0.089×HOXC10 + 0.238×SCARA3 - 0.424×SLC39A10.
Age and sex Reported complete and checked against original public clinical records for all 153 patients.
MGMT Source-annotated methylation status was usable and nonconflicting for 107 of 139 main-cohort patients; not included in the reported models.
KPS Not used in current models; only 49 main-cohort patients have explicitly preoperative KPS.

Model Comparisons

Table 2. Model specifications and comparison groups, with descriptions of their training and use.
Table 2. Model specifications and comparison groups, with descriptions of their training and use.
Model Description
Refitted four-gene Cox Conventional Cox proportional-hazards regression using the four continuous gene values; refitted inside outer training folds.
Four-gene DeepSurv Tested Cox-based neural survival model; candidate hidden layers [8] or [16,8], ReLU activation, two L2-penalty choices, Adam optimizer, inner-validation early stopping, one initialization per fit.
Fixed Cao score Published four-gene coefficient score applied as a fixed benchmark, not refitted. For one-year Brier calculations, baseline survival was estimated from each outer training group.
Age/sex extensions Age/sex Cox; age/sex/four-gene Cox; matched age/sex/four-gene DeepSurv.

Internal Validation and Tuning

Internal validation was accomplished with three repetitions of five-fold outer validation, with folds stratified by death versus censoring status. Inner three-fold tuning was employed for DeepSurv within each outer training group. Both Cox and DeepSurv were evaluated on the same held-out patients, while scaling, tuning, model fitting, and baseline-survival estimation used training data only. Primary concordance was compared within shared test folds and combined across folds according to the number of comparable patient pairs, rather than pooling raw risk scores from separately fitted models. The four genes themselves were selected in prior work using TCGA, so this analysis does not independently validate the gene-selection process.

Model Performance and Uncertainty

Harrell concordance (C-index) was the principal measure of survival risk-ranking. Censoring-aware time-dependent AUC and Brier prediction error were summarized at one year (365 days), with censoring estimates derived from the relevant training data. Higher C-index and AUC indicate better discrimination; lower Brier error indicates less overall prediction error. Two-year summaries were withheld because follow-up in the main cohort did not support reliable estimates in every validation group. The current 95% intervals were calculated by resampling saved held-out predictions and do not include uncertainty from refitting or retuning the models.

Sample Size, Missingness, Ethics, and Reproducibility

The current sample size was dictated by the available reconstructed TCGA data. The main cohort required valid curated survival data, complete four-gene expression values, and reported IDH-wildtype status. KPS was omitted because of missingness and inconsistent timing. Public source tables and saved model outputs were checked; an independent R analysis reproduced the refitted Cox calculations and held-out concordance results for the new reconstruction. This does not validate the original GIMPAS results.
This analysis used de-identified public data therefore IRB approval was not needed in this study. This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors. The authors have no conflicts of interest to report.

Results

Study Cohorts

Table 3. Cohort counts and analysis populations with their roles in the study.
Table 3. Cohort counts and analysis populations with their roles in the study.
Analysis group Patients Deaths/events Censored Role
Main: historical TCGA-GBM with reported IDH-wildtype status 139 113 26 Primary cohort for matched model comparison.
Broader historical TCGA-GBM cohort 153 121 32 Sensitivity analysis; includes all 139 main-cohort patients; not independent replication.

Primary Comparison: Four-Gene Cox Versus DeepSurv

Table 4. C-index for the fixed Cao score, refitted Cox, and DeepSurv in the main and broader cohorts; differences are DeepSurv minus refitted Cox.
Table 4. C-index for the fixed Cao score, refitted Cox, and DeepSurv in the main and broader cohorts; differences are DeepSurv minus refitted Cox.
Four-gene model/score Main cohort n=139 C-index Broader cohort n=153 C-index
Fixed published Cao score 0.614 0.636
Refitted four-gene Cox 0.591 0.627
Four-gene DeepSurv 0.567 0.620
DeepSurv minus refitted Cox -0.0238 -0.0071
Note: C-index measures the ranking of survival risk, not the percentage of patients correctly classified or an individual’s probability of survival. A value of 0.5 corresponds to chance-level ranking.

Paired Model Differences

Table 5. DeepSurv-minus-Cox C-index differences with descriptive conditional 95% bootstrap intervals.
Table 5. DeepSurv-minus-Cox C-index differences with descriptive conditional 95% bootstrap intervals.
Cohort / contrast DeepSurv minus Cox C-index 95% empirical-Bayesian coverage intervals
Main n=139; four genes only -0.0238 -0.0608 to +0.0111
Broader n=153; four genes only -0.0071 -0.0302 to +0.0159
Main n=139; age/sex/four genes -0.0204 -0.0558 to +0.0165
Note: These intervals describe variation in the saved predictions only; they are not full-pipeline confidence intervals.

One-Year Performance

Table 6. Secondary measures of model performance: time-dependent AUC and Brier prediction error reported at one year.
Table 6. Secondary measures of model performance: time-dependent AUC and Brier prediction error reported at one year.
Cohort Model Time-dependent AUC at 365 days (higher indicates better predictive discernment) Brier error at 365 days (lower indicates less prediction error)
Main n=139 Fixed Cao score 0.723 0.216
Main n=139 Refitted Cox 0.677 0.230
Main n=139 DeepSurv 0.657 0.242
Broader n=153 Fixed Cao score 0.758 0.204
Broader n=153 Refitted Cox 0.739 0.212
Broader n=153 DeepSurv 0.712 0.228
Note: Two-year performance is not reported because follow-up support was insufficient in the main cohort. Brier error reflects overall prediction error and is not, by itself, a calibration assessment.

Clinical Predictor Extension: Age and Sex

Table 7. C-index results with additional clinical factors in the main and broader cohorts.
Table 7. C-index results with additional clinical factors in the main and broader cohorts.
Model / inputs Main cohort n=139 C-index Broader cohort n=153 C-index
Four-gene Cox (prior four-gene run) 0.591 0.627
Age + sex + four-gene Cox 0.597 0.646
Age + sex + four-gene DeepSurv 0.577 0.635
Table 8. Main-cohort C-index differences with clinical predictors and matched DeepSurv-versus-Cox models.
Table 8. Main-cohort C-index differences with clinical predictors and matched DeepSurv-versus-Cox models.
Comparison C-index difference 95% empirical-Bayesian coverage intervals
Adding age/sex to four genes (Cox) +0.0063 -0.0243 to +0.0367
DeepSurv vs Cox, age/sex/four genes -0.0204 -0.0558 to +0.0165

Discussion

The principal finding of this study was the lack of observed DeepSurv superiority over conventional Cox regression in from-diagnosis, retrospective GBM survival analysis. Our main measure, the C-index, represents survival risk-ranking, and all models showed modest discrimination. We evaluated each model in the 139-patient main cohort and the broader 153-patient cohort. The broader cohort includes the main cohort and is a sensitivity analysis, not an independent replication. In the basic four-gene comparison, DeepSurv returned C-indexes of 0.567 and 0.620 for the main and broader cohorts, compared with 0.591 and 0.627 for the refitted Cox model. The paired DeepSurv-minus-Cox differences were -0.0238 and -0.0071, respectively. The descriptive conditional bootstrap intervals for these differences included zero, but they were calculated from saved predictions and did not account for model refitting or retuning. The results therefore show no observed improvement from the tested neural model.
Additionally, both refitted models had lower C-indexes than the published Cao et al. score, which employed Cox regression methods and returned a C-index of 0.614 for the main cohort and 0.636 for the broader cohort. Because Cao’s genes and coefficients were developed using overlapping TCGA data, this score is useful as a published benchmark but is not an independent validation comparison. Its higher C-index cannot by itself tell us whether the difference arose from model construction, cohort overlap, or another cause, but further analysis can investigate these likely culprits.
Other measures of model performance we employed were time-dependent AUC and Brier prediction error at one year. These followed the same general pattern as the C-index in the main cohort: Cox had a time-dependent AUC of 0.677 and Brier error of 0.230, compared with 0.657 and 0.242 for DeepSurv. The published Cao score had an AUC of 0.723 and Brier error of 0.216. Higher AUC and lower Brier error favor Cox and Cao in these data, although Brier error measures overall prediction error rather than discrimination alone.
The exploratory expanded analyses including age and sex showed small numerical increases in C-index. The four-gene Cox model increased from 0.591 to 0.597, while DeepSurv increased from 0.567 to 0.577. The overall pattern observed in the four-gene analysis was preserved, with the age/sex/four-gene Cox model still slightly higher than the matched DeepSurv model. These changes are small, however, and the available conditional intervals do not establish a clear benefit from adding age and sex in the main cohort.
The observed findings, particularly the slightly lower C-index for DeepSurv than for refitted Cox on the same patients, are not strong enough to establish general superiority, inferiority, or equivalence. The relatively small historical dataset, missing or conflicting information, inconsistently timed KPS measurements, and absence of an independent external test group limit the analysis. The four genes were selected in earlier TCGA work, so our internal comparison does not independently validate the complete signature-development process. The network architecture and tuning choices were also limited to the tested configuration; other designs could perform differently. Nonetheless, the direction of results across cohorts and performance measures suggests that additional model complexity did not yield an observed benefit in this particular analysis. Future work should assess whether this pattern holds in independent patient data. This retrospective comparison is insufficient to guide individual clinical decisions without further validation.

Conclusions

In conclusion, the tested DeepSurv model did not outperform conventional Cox regression in this reconstructed TCGA GBM survival analysis. The small historical cohort and lack of independent validation prevent definitive conclusions about performance in other patient groups. These limitations highlight the need for a careful assessment of uncertainty and for external validation before clinical claims are made.

References

  1. Cao, M.; Cai, J.; Yuan, Y.; et al. A four-gene signature-derived risk score for glioblastoma: prospects for prognostic and response predictive analyses. Cancer Biol. Med. 2019, 16(3), 595–605. [Google Scholar] [CrossRef] [PubMed]
  2. Pohl, L.C.; Leitheiser, M.; Obrecht, D.; et al. Molecular characteristics and improved survival prediction in a cohort of 2023 ependymomas. Acta Neuropathol. (Berl) 2024, 147(1), 24. [Google Scholar] [CrossRef] [PubMed]
  3. Familiari, P.; Lapolla, P.; Picotti, V.; et al. Role of 1p/19q Codeletion in Diffuse Low-grade Glioma Tumour Prognosis. Anticancer Res. 2023, 43(6), 2659–2670. [Google Scholar] [CrossRef] [PubMed]
  4. Kvamme, H. havakv/pycox. 28 October 2026. Available online: https://github.com/havakv/pycox.
  5. jaredleekatzman. jaredleekatzman/DeepSurv. 16 October 2026. Available online: https://github.com/jaredleekatzman/DeepSurv (accessed on 28 October 2026).
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.
Copyright: This open access article is published under a Creative Commons CC BY 4.0 license, which permit the free download, distribution, and reuse, provided that the author and preprint are cited in any reuse.