Work

Diagnosing Congenital Heart Disease Using Clinical and Auscultation Data: a Bayesian approach

Year: 2026
Type: dissertation
Abstract: Introduction: Congenital heart disease (CHD) is the most common group of congenital anomalies worldwide and remains a major cause of morbidity and mortality in paediatric populations.Early detection is essential to improve outcomes, yet access to specialised diagnostic tools such as echocardiography is often limited, particularly in resource-constrained settings.There is, therefore, a need for reliable and interpretable decision-support tools to support clinical screening.Aim: To develop and evaluate interpretable probabilistic models for paediatric CHD screening using routinely collected clinical and auscultatory data.Methods: A retrospective modelling study was conducted using data from 865 paediatric subjects enrolled in a cardiovascular screening programme.Following outcome definition, 701 subjects with sufficient diagnostic information to define the outcome were included in supervised analyses.Clinically relevant predictors available at the time of screening were selected and discretised.Bayesian Network (BN), Naïve Bayes (NB), and Tree-Augmented Naïve Bayes (TAN) models were developed and evaluated using cross-validation.Model performance was assessed using area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, precision, and F1-score.Threshold optimisation was performed to explore clinical applicability in a screening context.Results: All models demonstrated comparable discriminative performance for binary CHD classification.The Naïve Bayes model achieved the highest observed AUC (0.781), closely followed by the TAN model (0.778) and the core Bayesian Network model (0.772).Expanding the predictor set did not improve performance and reduced the effective sample size due to missing data.Murmur-related variables consistently emerged as the most informative predictors.Threshold optimisation increased sensitivity to clinically relevant levels (≥85%), at the cost of reduced specificity and higher referral rates.Four-class diagnostic classification showed substantially lower performance.Discussion: Probabilistic models provide an interpretable and flexible framework for supporting congenital heart disease screening using routinely collected clinical, physiological, and auscultation data.The findings suggest that carefully selected clinical predictors may have a greater influence on screening performance than increasing model complexity.These models show potential as decision-support tools for early risk stratification and referral, particularly in settings with limited access to specialised care.
Language: English
Open Access status: green
OpenAlex ID: W7233517749
Cites: 51
Cited by:
Related to: 10