Machine Learning Prediction of BNP, hs-Troponin, and CRP from Routine Blood Tests in Heart Failure Patients: Evidence from Abakaliki, Ebonyi State, Nigeria

Abstract

Heart failure (HF) is a significant cause of morbidity and mortality worldwide, with a hefty burden in low-resource settings where diagnostic assays remain costly and limited in availability. This study examined whether routine haematology, coagulation, and biochemistry parameters can predict B-type Natriuretic Peptide, High-sensitivity Troponin and C-reactive Protein in HF patients using machine learning (ML). Methods: We prospectively enrolled 579 adults with HF at Alex Ekwueme Federal University Teaching Hospital, Abakaliki. We collected 10 mL of venous blood for laboratory investigations and obtained demographic and clinical data from their medical records. We processed the data in Python 3.12 and applied feature selection techniques including correlation thresholds, recursive feature elimination, its standard libraries. We trained and evaluated nine models, choosed the best model for each biomarker, and conducted sex-stratified analyses to compare performance between male and female participants. Results: Important predictors included Urea, creatinine, eGFR, D-dimer, fibrinogen, and neutrophil-to-lymphocyte ratio (NLR). Models that combined all the parameters outperformed single-domain models. CatBoost produced the best results for BNP (R² 0.30–0.33), ElasticNetCV for hs-Troponin (R² 0.09–0.12), and Ridge/ElasticNetCV for CRP (R² 0.53–0.54). SHAP analysis indicated that Urea and D-dimer strongly influenced BNP, while NLR, eGFR, and fibrinogen contributed most to the predictions of hs-Troponin and CRP. Sex-stratified models showed consistent behaviour across algorithms, with only minor differences in predictive strength. Conclusion: Routine laboratory data can estimate BNP, hs-Troponin, and CRP using ML in patients with HF from low-resource settings.