Scholarlink

Background <p>Early cancer risk assessment in the aging canine population is a critical unmet need. Routine laboratory data are a ubiquitous, low-cost data source for developing screening tools, but their utility is limited by the non-specificity of biomarkers and the low prevalence of cancer in screening populations, which creates severe dataset imbalance for machine learning.</p> Methods <p>Using the Golden Retriever Lifetime Study cohort, the feasibility of building a cancer risk classification model from routine laboratory data was assessed. This study design intentionally reflected real-world data constraints, including the grouping of diverse cancer types and the inclusion of post-diagnosis visits. Data were partitioned at the patient level to prevent information leakage. This study systematically compared 126 analytical pipelines, evaluating various machine learning models, feature selection methods, and data balancing techniques. The best model’s predictions were interpreted using SHapley Additive exPlanations (SHAP).</p> Results <p>The optimal approach was a Logistic Regression model with class weighting and recursive feature elimination. On the test set, the model showed moderate patient ranking ability (AUROC = 0.815; 95% CI: 0.793–0.836) but failed as a clinical classification tool, with a low F1-score (0.25) and Positive Predictive Value (0.15). A high Negative Predictive Value (0.98) was undermined by insufficient recall (0.79), precluding its use as a rule-out test. SHAP analysis confirmed that predictions were driven by non-specific features, primarily age and markers of inflammation and anemia.</p> Conclusion <p>This benchmark study demonstrates that routine laboratory data contain a statistically detectable but clinically unreliable signal for canine cancer. The signal is too weak and non-specific to distinguish malignancy from normal aging or other inflammatory conditions, and is likely confounded by treatment effects. The resulting model is unsuitable for clinical use. These findings establish a performance ceiling for this data modality used in isolation and highlight that meaningful progress in computational veterinary oncology will require the integration of multi-modal data.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

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

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年2月1日

查看全文
  • Shumin Li

摘要

Background

Early cancer risk assessment in the aging canine population is a critical unmet need. Routine laboratory data are a ubiquitous, low-cost data source for developing screening tools, but their utility is limited by the non-specificity of biomarkers and the low prevalence of cancer in screening populations, which creates severe dataset imbalance for machine learning.

Methods

Using the Golden Retriever Lifetime Study cohort, the feasibility of building a cancer risk classification model from routine laboratory data was assessed. This study design intentionally reflected real-world data constraints, including the grouping of diverse cancer types and the inclusion of post-diagnosis visits. Data were partitioned at the patient level to prevent information leakage. This study systematically compared 126 analytical pipelines, evaluating various machine learning models, feature selection methods, and data balancing techniques. The best model’s predictions were interpreted using SHapley Additive exPlanations (SHAP).

Results

The optimal approach was a Logistic Regression model with class weighting and recursive feature elimination. On the test set, the model showed moderate patient ranking ability (AUROC = 0.815; 95% CI: 0.793–0.836) but failed as a clinical classification tool, with a low F1-score (0.25) and Positive Predictive Value (0.15). A high Negative Predictive Value (0.98) was undermined by insufficient recall (0.79), precluding its use as a rule-out test. SHAP analysis confirmed that predictions were driven by non-specific features, primarily age and markers of inflammation and anemia.

Conclusion

This benchmark study demonstrates that routine laboratory data contain a statistically detectable but clinically unreliable signal for canine cancer. The signal is too weak and non-specific to distinguish malignancy from normal aging or other inflammatory conditions, and is likely confounded by treatment effects. The resulting model is unsuitable for clinical use. These findings establish a performance ceiling for this data modality used in isolation and highlight that meaningful progress in computational veterinary oncology will require the integration of multi-modal data.