Machine Learning Reveals What Really Predicts Survival in Dogs With Mammary Tumors

Canine mammary tumors are the most common cancer in female dogs, and they remain one of the most lethal. Even after surgical removal of the affected gland, roughly 45 percent of dogs with malignant tumors experience recurrence or distant metastasis within two years of diagnosis, and mortality reaches about 45 percent within a single year. For decades, veterinarians have relied on the WHO Tumor-Node-Metastasis staging system combined with histopathological examination to decide how aggressively to treat each patient. Yet this approach captures only part of the biological complexity of a disease that is notoriously heterogeneous. A new study published in Veterinary Oncology has now put that conventional wisdom to a rigorous computational test, asking whether modern machine learning and gene expression data can outperform the clinical variables veterinarians already measure at the bench.

The research team, led by Hedda Fjell Scheel of the Norwegian University of Life Sciences together with colleagues at Oslo University Hospital and the University of Oslo, took advantage of a publicly available dataset of whole transcriptome sequencing data from canine mammary tumors. After quality control and the removal of animals with incomplete survival records, the final cohort comprised 146 dogs aged between two and nineteen years, representing twenty different breeds. Thirty-eight of these dogs had a recorded time of death after surgery, ranging from one day to 709 days with a median of 205.5 days, while the remaining 108 were right censored, meaning they were still alive at their last follow-up. This substantial censoring proportion, roughly 74 percent of the cohort, posed a genuine statistical challenge that shaped every subsequent analytical decision.

The investigators built three competing Cox proportional hazards survival models. The first, a clinical model, used only variables measurable through standard veterinary workup. The second, a gene expression model, relied exclusively on standardized, pre-filtered transcriptomic features. The third combined both data types. Because the gene expression matrix contained 18,579 genes, vastly more predictors than the 38 survival events available, the team applied careful feature pre-filtering: they retained the 1,000 most variable genes by mean absolute deviation, removed highly correlated pairs above a threshold of 0.95, and selected genes significant in univariate Cox models at a false discovery rate below 0.1. This left forty candidate genes, none of which belonged to the PAM50 subtyping panel used in human breast cancer. Regularized elastic net regression with cross-validated tuning of the alpha and lambda parameters then selected the most informative features for the gene-based models.

Model performance was evaluated with Uno’s C-index, a censoring-robust concordance measure in which 0.5 represents random prediction and 1.0 perfect discrimination, and with time-dependent ROC-AUC at six months, one year, and two years after surgery. Across 100 iterations of Monte Carlo cross-validation with stratified train-test splits of 4:1, the results were strikingly clear. The clinical model achieved a median C-index of 0.646, marginally ahead of the gene expression model at 0.635 and the combined model at 0.621, though none of these differences reached statistical significance. More tellingly, at the clinically critical timepoints of six months and one year after surgery, the clinical model significantly outperformed both transcriptomic models, with a median ROC-AUC of 0.811 at six months compared with 0.727 for the gene expression model and 0.768 for the combined model.

What emerged as decisive were three humble, readily obtainable clinical variables. Stepwise regression favored a multivariate model containing age, lymphatic invasion assessed by tumor histology, and estrogen receptor status evaluated by immunohistochemistry. Lymphatic invasion exerted the largest effect on the hazard of death, confirming its previously reported prognostic power, while estrogen receptor positivity was associated with improved survival, mirroring the pattern well established in human breast cancer. Notably, histopathological grade and molecular subtype, both significant in univariate analyses, dropped out of the multivariate model, likely because they were strongly correlated with lymphatic invasion and estrogen receptor negativity and therefore added no independent information. The final clinical model explained 17.8 percent of the overall variation in survival.

The molecular subtyping analysis itself yielded biologically meaningful results. Using the single-sample MPAM50 classifier, every tumor in the cohort was assigned to either the luminal A or the basal-like subtype, with luminal A accounting for 58.2 percent of samples. Subtype was significantly associated with age, breed, lymphatic invasion, grade, and estrogen receptor status. Basal-like tumors were more frequently grade 2 or 3, more often showed lymphatic vessel invasion, and were overrepresented among estrogen receptor negative tumors, with 69.6 percent of the 23 receptor-negative tumors classified as basal-like. Dogs with luminal A tumors were on average one year younger, and Maltese dogs were more likely to harbor luminal A tumors. These findings indicate that the canine PAM50 subtypes capture tumor biology beyond what estrogen receptor status alone reveals, echoing their prognostic significance in human patients.

Perhaps the most intriguing result came when the researchers used the best-performing clinical model to assign each dog a risk score and split the cohort at the median into high-risk and low-risk groups of 73 dogs each. Survival differed significantly between the groups, with a log-rank p-value of 0.0036. Differential gene expression analysis then uncovered 433 genes significantly differentially expressed between the groups, 279 upregulated in high-risk tumors and 154 in low-risk tumors. Gene set enrichment analysis using the Hallmark collection from the Molecular Signatures Database revealed twenty-four significantly enriched pathways. High-risk tumors were enriched for signatures of proliferation, including E2F targets and the G2M checkpoint, and for metabolic programs such as oxidative phosphorylation and glycolysis, while low-risk tumors showed enrichment of immune-related pathways, including TNF alpha signaling, IL6-JAK-STAT signaling, inflammatory response, and interferon alpha response.

That immune activity marks the low-risk group is a finding with real translational resonance. In human breast cancer, both innate and adaptive immune components are associated with recurrence-free survival, and immune gene expression carries prognostic weight particularly in hormone receptor negative tumors. The canine data suggest that the immune system plays an equally central role in anti-tumor response across species, reinforcing the value of dogs as a comparative model for human breast cancer research. The two diseases share spontaneous onset, histological subtypes, underlying genetic alterations, and gene expression changes, yet molecular markers that transformed human breast cancer management, from PAM50 subtyping to gene expression signatures like MammaPrint, have never been incorporated into routine canine staging. This study represents one of the most systematic attempts to close that gap.

Why, then, did gene expression fail to improve prediction? The authors offer several candid explanations. Adding thousands of transcriptomic features to a small, heavily censored dataset may introduce more noise than signal, a phenomenon documented in large-scale benchmark studies showing that multi-omics survival models are widely sensitive to noise. Linear Cox models may also miss non-linear interactions between genomic and clinical features. Moreover, with only 38 events, statistical power is inherently limited, and a larger number of events might reveal differences that this cohort cannot detect. None of the models exceeded a time-independent C-index of 0.65, which, while comfortably above random, falls short of what would be needed for clinical implementation, though comparable multi-omics studies in humans rarely exceed 0.7 either.

The practical message for veterinary oncology is nonetheless concrete. Estrogen receptor status measured by immunohistochemistry and lymphatic invasion assessed by routine histology could be valuable additions to current TNM staging, potentially refining survival predictions and informing decisions such as whether ovariohysterectomy at the time of tumor removal might benefit a particular patient, since dogs with high serum estradiol and receptor-positive tumors appear more likely to benefit from that intervention. The authors also emphasize the limitations of using overall survival rather than cancer-specific survival in an elderly cohort, and they call for future studies to record cause of death, complete TNM parameters, and richer genomic data. Alternative modeling approaches such as boosting methods and random survival forests may yet extract more from transcriptomic data. For now, the study stands as a sobering and instructive reminder that in the era of big data, the humble microscope slide and the immunohistochemistry stain still hold their ground.

Subject of Research: Machine learning-based identification of prognostic factors for survival in canine mammary gland tumors

Article Title: Unraveling prognostic factors in canine mammary gland tumors using machine learning

Article References: Scheel, H. F., Zobolas, J., Lien, T. G., Lingaas, F., & Bergholtz, H. (2025). Unraveling prognostic factors in canine mammary gland tumors using machine learning. Veterinary Oncology, 2(1), Article 16. https://doi.org/10.1186/s44356-025-00030-7

Image Credits: AI Generated

DOI: 10.1186/s44356-025-00030-7

Keywords: canine mammary tumors, machine learning, Cox regression, gene expression, estrogen receptor, lymphatic invasion, survival prediction, TNM staging, PAM50 subtyping, gene set enrichment analysis, veterinary oncology, precision medicine