Neuro-symbolic AI model with heart failure ontology for precise phenotyping and 1-year prognosis prediction in chinese heart failure patients: a multicenter study
T Zou, J Wang, J Yang, D Liu, Z Niu, H QiAbstract
Background
Precise phenotyping and risk stratification of heart failure (HF) remain major challenges in Chinese populations with prominent etiological heterogeneity. Traditional data-driven artificial intelligence (AI) models have critical limitations of poor interpretability and insufficient semantic consistency across multicenter heterogeneous data.
Purpose
To develop and validate an interpretable neuro-symbolic AI model integrated with a standardized heart failure ontology, to achieve precise phenotyping and accurate prognosis prediction in Chinese HF patients.
Methods
We first constructed a standardized Heart Failure Ontology (HFO) covering 8 core domains: etiology, phenotype, comorbidities, biomarkers, imaging, treatment, endpoint events, and semantic associations, for standardized governance of real-world data from the China Heart Failure Alliance (CHFA) database. We then developed a Neuro-Symbolic Reasoning Model (NSRM), which integrates high-dimensional feature representation of deep neural networks with symbolic logic rules embedded in the 2024 ESC Heart Failure Guidelines, to simultaneously perform unsupervised phenotyping and supervised prediction of the 1-year composite endpoint (all-cause mortality and HF rehospitalization). Performance was compared with conventional machine learning models.
Results
A total of 48,267 hospitalized HF patients from 128 multicenter sites nationwide in the CHFA database (2018–2024) were included. HFO achieved 98.2% semantic alignment of heterogeneous clinical variables across centers. NSRM identified 5 HF phenotypes with significant prognostic differences, with 1-year composite endpoint event rates ranging from 8.3% to 37.1% (P<0.001). In the validation cohort, NSRM reached an AUC-ROC of 0.89 (95%CI 0.88–0.90) for composite endpoint prediction, significantly outperforming the Lasso-Cox model (0.79), random forest model (0.81), and pure CNN-LSTM model (0.85, all P<0.001), with optimal calibration and clinical net benefit.
Conclusion
The neuro-symbolic AI model integrated with a heart failure ontology enables precise, interpretable phenotyping and prognosis prediction in Chinese HF populations, overcomes the core limitations of traditional black-box AI models, and provides a powerful tool for individualized clinical HF management.Figure 1:Architecture of the NSRM