Multicenter machine learning based risk assessment model, integrating omics, clinical and patient centred data to prevent recurrent heart failure events

Background

Despite advances in heart failure (HF) management, mortality, hospital readmissions and healthcare-related costs remain markedly high. Although several integrated transitional care models have been developed, most of these focus on mortality prediction and none of these has specifically assessed to predict HF-related hospitalizations. There is a growing need for multimodal predictive approaches that better capture the biological and clinical complexity of HF.

Purpose

The ORACLE multicentric study (ClinicalTrials.gov ID NCT05679713) aimed to develop a machine learning (ML) model integrating omics, clinical and psychosocial data to improve HF risk stratification and prediction of HF-related hospitalizations.

Methods

RNA-sequencing was performed in 60 patients (30 cases, 30 controls) to identify candidate blood biomarkers associated with HF. Differential expression analysis and GLM models (Lasso, Ridge) selected 22 genes. Their expression was then quantified by qPCR in an independent cohort of 119 patients. These gene-expression values were integrated with clinical and psychosocial variables and used as predictors in supervised learning models.

Multiple algorithms (including logistic regression, random forest and CatBoost) were trained using 5-fold nested cross-validation, outer loop for unbiased estimation and inner for loop for hyperparameter tuning. Internal performance was assessed using AUC and balanced accuracy (BA).

Feature importance was examined using usual (SHAP) and model-intrinsic metrics.

Results

RNA-seq identified 821 putatively differentially expressed genes, from which 22 showed significant associations with HF and were validated by qPCR. In the integrated prediction model including genomic, clinical and patient-centred data, the CatBoost model achieved the best performance, with a mean AUC of ~0.75 and BA of ~0.70 on the internal test set. SHAP analysis showed that Gene 7 was the most influential variable across all three models, while the 6-minute walk test ranked as the second most important feature in both CatBoost and Random Forest.

Conclusions

Our findings suggest that these 22 genes represent potential novel blood-based biomarkers of HF, and that integrating them with patient centred data in a ML model may enhance risk stratification and prediction of HF-related hospitalizations, more accurately than current clinical methods.ROC curves from 5-fold CV and SHAP analysisHeatMap of the 22 selected genes.