From prediction to phenotype: machine learning classification of heart failure and its relation to imaging derived left ventricular structure in the UKBiobank

Abstract

We aimed to utilise data-driven machine learning (ML) algorithms to identify the most discriminative features for classifying clinical heart failure (HF) and to investigate how these features associate with quantitative cardiac magnetic resonance (CMR) imaging phenotypes in individuals free from HF.

We employed XGBoost, LightGBM, and Elastic Net to classify prevalent HF cases at UK Biobank baseline (2,731 events / 502,129 participants). A feature set of 77 variables was included across all models, encompassing socio-demographic characteristics, physical measurements, lifestyle factors, comorbid conditions, laboratory tests, and HF risk score (calculated using PRS-CS). Datasets were partitioned into training and test sets, pre-processed separately to prevent data leakage between sets and an oversampling technique (SMOTE) was applied exclusively to the training dataset. The most overlapped features appearing in the top 20 ranked features across all models were identified. These features were then evaluated in a separate UK Biobank study population free from HF at baseline, using linear regression models to assess associations with 12 left ventricular (LV) imaging measurements derived from CMR imaging.

All ML models demonstrated excellent performance, with area under the receiver operating characteristic curve often exceeding 0.90 across diverse evaluation settings. Notably, XGBoost achieved specificity of 99% and Elastic Net achieved sensitivity of 97%. The 21 overlapping top-ranked features across models comprised conventional risk factors (age, height, HTN, AF, sedentary behaviour, smoking status, …), and some non-traditional biomarkers. Calcium, total bilirubin and urea demonstrated remarkably neutral associations across all LV indices. Higher creatinine was associated with subtly increased LV mass (body surface area [BSA]-indexed LVM: +1.5% per SD) and marked reductions in global function (LVGFI: -1.35% per SD, LVMCF: -3.08% per SD). Higher resting heart rate exhibited statistically significant and consistent pattern, relating to lower EF (-1.18% per SD), smaller LV volumes (BSA indexed LVEDV: -4.5% per SD, BSA indexed LVSDV: -4.07% per SD, BSA indexed LVSV: -4.81% per SD) and mass (BSA indexed LVM: -3.26% per SD), and poorer functional indices (LVGFI: -1.26% per SD, LVMCF: -2.69% per SD).

This study demonstrates that ML models trained to classify HF distil a biologically coherent signature of early LV remodelling. This work provides empirical support for the mechanistic relevance of data-driven feature rankings and offers a framework for understanding how multifactorial risk operates at the subclinical level.Overlapping top-ranked featuresFor image description, please refer to the figure legend and surrounding text.Multivariable linear regression modelsFor image description, please refer to the figure legend and surrounding text.