🎉 New Publication — Now Online!
Excited to share that my paper "Machine Learning Insights for Cardiovascular Risk Prediction in Diabetic Patients: Emphasis on Renal and Cardiac Markers Using Random Forests" has been published in Artificial Intelligence in Health (AccScience Publishing).
🔬 What the study does:Evaluates logistic regression and random forest models for heart failure mortality prediction using a leakage-resistant, fully reproducible analytic pipeline with stratified five-fold cross-validation.
📊 Key findings:→ Random forest achieved AUC 0.91 vs. 0.86 for logistic regression under pooled out-of-fold evaluation → Serum creatinine, ejection fraction, age, and follow-up time consistently emerged as dominant predictors across models and SHAP analyses → The two models exhibited distinct clinical error profiles: logistic regression favored sensitivity, random forest favored specificity → All metrics derived exclusively from out-of-fold predictions — no in-sample inflation
🧠 Why it matters:Too many clinical ML studies report inflated performance from weak validation. This work prioritizes methodological discipline over model complexity, establishing a transparent baseline for cardiometabolic risk modeling that aligns with FDA Good Machine Learning Practice principles.
🛡️ Why conservative validation matters in clinical AI:
Conservative validation is a foundational requirement for clinical AI because it prevents inflated performance estimates and ensures models are safe and reliable for real-world medical use. Here is why it is essential:
🔒 Mitigating performance inflation and information leakageMany AI studies report inflated accuracy because they allow information leakage, where evaluation data contaminates the training process. Conservative validation prevents this by ensuring all preprocessing, scaling, and class imbalance handling are restricted strictly to training folds and never involve held-out data.
⚠️ Addressing shortcut learningModels can achieve high accuracy by exploiting spurious correlations or dataset-specific artifacts rather than identifying genuine clinical signals. Rigorous, leakage-resistant evaluation helps identify if a model is relying on these brittle artifacts, which could lead to unsafe behavior when deployed in new populations.
📐 Establishing credible benchmarksBefore moving to highly complex architectures, it is vital to establish transparent and defensible performance baselines using standard models. This ensures that performance gains reflect actual model improvements rather than methodological weaknesses.
🎯 Evaluating decision-level behaviorConservative validation moves beyond aggregate metrics like AUC to examine threshold-dependent classification behavior. Using techniques like McNemar's test on pooled out-of-fold predictions, researchers can determine if model differences are statistically significant at the decision level — which is more operationally meaningful for clinicians.
🏥 Building clinical trust and regulatory alignmentMethodological rigor and transparency are foundational prerequisites for clinical impact. Following conservative practices aligned with FDA Good Machine Learning Practice supports the development of ethically deployable and trustworthy AI systems that clinicians can rely on for high-stakes decisions like cardiovascular risk stratification.
💡 Ultimately, rigorous evaluation and appropriate validation practices may be more consequential for trustworthy clinical machine learning than the complexity of the model itself.
📄 DOI: 10.36922/AIH025490111 Open Access | Artificial Intelligence in Health
#MachineLearning #CardiovascularDisease #Diabetes #HeartFailure #RandomForest #ClinicalAI #Reproducibility #HealthInformatics #MedicalResearch #OpenAccess #ValidationMatters #FDAGuidance #ClinicalDecisionSupport 😀😀😀
ALT https://julianborgesmd.blogspot.com/2026/05/blog-post.html