Routine Blood Tests Fail as Reliable Early Cancer Screen in Dogs, Landmark Machine Learning Study Finds

A comprehensive new study has delivered a sobering verdict on one of veterinary medicine’s most tantalizing ideas: that the humble, routinely collected blood test could be mined by artificial intelligence to catch cancer early in dogs. Drawing on more than 22,000 laboratory visits from over 3,000 Golden Retrievers followed for years, the research systematically tested 126 different machine learning pipelines and found that while routine bloodwork does carry a statistically detectable signal associated with malignancy, that signal is far too weak and non-specific to support a clinically useful screening tool. The findings, published in the journal Veterinary Oncology, establish what the author describes as a performance ceiling for this data modality when used in isolation, and they carry important lessons for both veterinary and human cancer research.

The motivation behind the study is compelling. Cancer is a leading cause of death in companion dogs, and its incidence climbs steeply with age. A 2025 survey by the veterinary imaging company HT Vista found that 62 percent of masses seen in US veterinary clinics go undiagnosed, underscoring a glaring diagnostic gap. Because dogs also serve as valuable spontaneous models for human oncology, a cheap, scalable screening method built on data that veterinarians already collect every year, the complete blood count and serum biochemistry panel, would be transformative. The central hypothesis was that although individual blood parameters are uninformative on their own, subtle multivariate patterns hidden across dozens of values might harbor a presymptomatic signature of malignancy that machine learning could unlock.

To test that hypothesis rigorously, the study harnessed the Golden Retriever Lifetime Study, a large prospective observational cohort of 3,044 purebred Golden Retrievers enrolled between 2012 and 2015 and followed for their entire lives. Annual visits generate full physical examinations, owner and veterinarian questionnaires, and biospecimens processed for complete blood counts and chemistry profiles on standardized laboratory instruments. Cancer diagnoses were assembled from two complementary sources covering all three tiers of diagnostic confidence, from histologically confirmed cases down to presumptive clinical suspicions, yielding a final cancer cohort of 659 unique subjects. The most common tumor types were hemangiosarcoma at just over a quarter of cases, followed by mast cell tumors and lymphoma.

The resulting analytical dataset of 22,460 veterinary visits presented a formidable statistical challenge: only 6.3 percent of visits were cancer-positive, a degree of class imbalance notorious for derailing machine learning algorithms. The study design deliberately embraced real-world messiness. Visits were labeled positive only if they occurred on or after the recorded diagnosis date, and for the 41.4 percent of cancer subjects with no post-diagnosis laboratory work, only the final pre-diagnosis visit was labeled positive. Crucially, the model was trained without discriminating for treatment status, meaning it was exposed to laboratory results from both treated and untreated periods after diagnosis, exactly as real-world screening data would be.

The methodological framework was exhaustive. Six base algorithms were compared, spanning logistic regression, random forests, two gradient boosting methods, a neural network, and Naïve Bayes. These were crossed with three feature selection strategies, including recursive feature elimination and a manually curated panel of fifteen biomarkers tied to paraneoplastic syndromes such as anemia, thrombocytopenia, and hypercalcemia, and with six resampling techniques for handling imbalance, from SMOTE to class weighting. Every one of the 126 pipelines underwent five-fold cross-validated grid search optimized for the Matthews Correlation Coefficient, a metric well suited to imbalanced data. Data were split at the patient level to prevent information leakage, ensuring that no dog’s visits appeared in more than one partition and that the held-out test set provided a genuinely unbiased estimate of generalization.

The winning pipeline, a logistic regression model with class weighting and recursive feature elimination, achieved a respectable area under the ROC curve of 0.815 on the test set, with a 95 percent confidence interval of 0.793 to 0.836. That figure means a randomly chosen cancer-positive visit had an 81.5 percent chance of receiving a higher risk score than a randomly chosen cancer-negative visit, proof that a genuine signal exists in the data. But ranking ability is not the same as classification, and here the model failed decisively. Its F1-score was just 0.25, and its positive predictive value a mere 0.15, meaning that of all visits flagged as high risk, only 15 percent were actual cancer cases while 85 percent were false alarms.

The precision-recall curve told the clinical story plainly. Sitting only slightly above the no-skill baseline set by the 6.3 percent cancer prevalence, it showed that achieving any meaningful sensitivity would drive precision to levels that would flood clinics with false positives. The model’s high negative predictive value of 0.98 initially suggested promise as a rule-out test, but that hope collapsed under scrutiny of its recall: at 0.79, the model would miss roughly one in five dogs with cancer, an unacceptable false-negative rate for any screening application. The study’s conclusion was unambiguous: the model, and likely any model built solely on this data modality, is unsuitable for clinical deployment.

Explainable AI analysis using SHAP, a game-theory-based method that quantifies each feature’s contribution to individual predictions, revealed why the model fell short. The single most powerful predictor was patient age, followed by markers of anemia, such as lower hemoglobin, and systemic inflammation, such as elevated band neutrophils and a higher neutrophil-to-lymphocyte ratio. These are biologically plausible features, but they are profoundly non-specific. In effect, the model had learned to identify older dogs showing signs of chronic illness rather than a distinctive signature of cancer, functioning more as an old-dog detector than a malignancy biomarker. Cancer-induced blood changes are simply too entangled with those caused by common geriatric conditions, from chronic kidney disease to inflammatory bowel disease, for a bloodwork-only model to separate them.

A further and arguably more insidious problem was time-varying confounding by treatment. Because the dataset included post-diagnosis visits, the model could associate the iatrogenic effects of therapy with the cancer label. A dog with lymphoma treated with a multi-agent chemotherapy protocol will develop a stress leukogram and possible hyperglycemia; a dog with a mast cell tumor on a tyrosine kinase inhibitor may develop hypoalbuminemia from gastrointestinal toxicity. The model thus learned a confounded causal pathway, from cancer to diagnosis to treatment to blood changes, rather than the desired direct pathway from cancer to blood changes. Addressing this will require advanced causal inference methods such as marginal structural models, the study notes.

The study’s limitations are candidly acknowledged. Grouping dozens of histologically distinct malignancies into a single multi-cancer label, a concession to limited case numbers for any one tumor type, inevitably biased the model toward generic markers of systemic illness. The population was restricted to a single breed with known genetic predispositions to certain cancers, limiting generalizability, and the model lacked access to physical exam findings, comorbidities, and owner-reported signs. Yet these constraints are precisely what make the benchmark valuable: it quantifies, under realistic conditions, the maximum performance achievable from routine laboratory data alone. The path forward, the author argues, lies not in more complex algorithms but in multi-modal data integration, combining bloodwork with imaging, medical records, molecular diagnostics, and clinical context to emulate the reasoning of an expert clinician, validated in large, diverse, external cohorts before any clinical use.

Subject of Research: Machine learning assessment of routine laboratory data for early canine cancer detection on an imbalanced dataset

Article Title: Assessing the feasibility of early cancer detection using routine laboratory data: an evaluation of machine learning approaches on an imbalanced dataset

Article References: Assessing the feasibility of early cancer detection using routine laboratory data: an evaluation of machine learning approaches on an imbalanced dataset. (n.d.). https://doi.org/10.1186/s44356-025-00054-z

Image Credits: AI Generated

DOI: 10.1186/s44356-025-00054-z

Keywords: veterinary oncology, machine learning, canine cancer, early detection, imbalanced data, routine laboratory data, Golden Retriever Lifetime Study, predictive modeling, SHAP, class imbalance, bloodwork screening, logistic regression