Clinical Data and Tabular DL Models to Identify HFpEF Patients

Georgios Nikolaos Tsakonas1, Athanasios Samaras2, Dimitris Fotopoulos1, Georgios Giannakoulas3, Ioanna Chouvarda4
1Lab of Computing, Medical Informatics and Biomedical - Imaging Technologies, School of Medicine, Aristotle University of Thessaloniki, 2Second Department of Cardiology, Aristotle University of Thessaloniki, General Hospital ‘Hippokration', 3First Department of Cardiology, Aristotle University of Thessaloniki, AHEPA University Hospital, 4Aristotle University of Thessaloniki

Abstract

Heart Failure with Preserved Ejection Fraction (HFpEF) is a complex and increasingly prevalent form of heart failure. Diagnosis of HFpEF is faced with a current issue that is frequently misdiagnosed, underdiagnosed or diagnosed late.

This study aims to develop a binary classification model capable of distinguishing healthy individuals from patients with heart failure with preserved ejection fraction (HFpEF), and to investigate whether such a model can generalize to the detection of underdiagnosed HFpEF cases.

The dataset includes 20,000 patient records containing 3 major groups, healthy subjects, HFpEF patients and underdiagnosed HFpEF patients from the CardioMining database. The data derives from discharge letters in cardiology hospitalization cases transformed into tabular format and it contains over 400 features including categories like medical history, symptoms, discharge diagnosis, etc.

A binary classification model was trained using healthy individuals and diagnosed HFpEF patients, with 80% of the data used for training and 20% reserved for testing.

Both tabular transformer-based approaches and table-to-image conversion techniques were explored to train the binary classification model, succeeding to classify test-set patients with over 88% accuracy and recall. The evaluation of the developed healthy-patient model was further extended to underdiagnosed HFpEF cases. The proposed model identifies underdiagnosed HFpEF cases, detecting more than 75% of such cases, indicating strong discriminative performance.

Further evaluation is being conducted to determine which features influence each class and therefore, the discrimination ability of the model. SHAP (SHapley Additive exPlanations) was applied to the trained model to quantify feature contributions both globally, identifying the clinical characteristics most consistently associated with HFpEF across patients, and locally, explaining individual predictions in terms of each patient's specific features. Initial global analysis indicates that features related to cardiac function and structural findings, including LV systolic function and dyspnea, emerge as the strongest drivers of the model's HFpEF classification.