A Bayesian network to support differential diagnosis in patients with acute chest pain
D Di Vece, V Schweiger, V Capone, M Meyer Roqueni, M Wuerdinger, S Kraler, F A Wenzl, E A Berrios Barcenas, C Mauro, L Liberale, T F Luescher, J R Ghadri, E Bossone, M Wischnewsky, C TemplinAbstract
Background
Undifferentiated acute chest pain is one of the most challenging scenarios in emergency medicine. Several life-threatening conditions, including acute coronary syndromes (ACS), aortic dissection, pulmonary embolism (PE), myocarditis, spontaneous coronary artery dissection (SCAD), and takotsubo syndrome (TTS), may present with overlapping clinical and biochemical features but necessitate fundamentally different management strategies. Existing disease-specific scores do not account for simultaneous multi-disease probabilities and do not reflect the iterative nature of clinical reasoning.
Purpose
To develop and internally validate an interpretable Bayesian network capable of simultaneous differential diagnosis across major life-threatening cardiovascular diseases in patients presenting with acute chest pain.
Methods
Seven diagnostic cohorts of patients presenting with acute chest pain were analysed, including ACS, aortic dissection, myocarditis, PE, TTS, SCAD, and non-life-threatening conditions. A probabilistic graphical model (Bayesian Network) was developed using a hybrid modelling strategy that combined an expert-defined clinical structure with data-driven structure learning. Multiple structure learning algorithms were evaluated, with final model selection being based on performance optimisation. The network incorporated 43 clinical variables spanning demographic characteristics, medical history, ECG features, and laboratory findings, enabling simultaneous estimation of posterior disease probabilities and inference irrespective of data completeness. Internal validation was performed using 15-fold cross-validation and leave-one-out validation. Discriminative performance for each disease entity was assessed using area under the receiver-operating-characteristic curves (AUC).
Results
A total of 1,954 patients were included: 701 with ACS (35.9%), 273 with aortic dissection (14.0%), 153 with myocarditis (7.8%), 128 with PE (6.6%), 173 with SCAD (8.9%), 386 with TTS (19.8%), and 140 with non-life-threatening conditions (7.2%). In 15-fold cross-validation, the Bayesian network demonstrated high discriminative performance across all diagnostic entities: ACS (AUC 0.918), aortic dissection (0.915), TTS (0.931), myocarditis (0.895), SCAD (0.889), and PE (0.826). Non-life-threatening conditions were also accurately identified (AUC 0.914). Performance remained stable in leave-one-out validation (ACS 0.919; aortic dissection 0.915; TTS 0.932; myocarditis 0.896; SCAD 0.889; PE 0.831; non-life-threatening conditions 0.918). Derivation AUCs were consistently high across entities.
Conclusions
In patients presenting with acute chest pain, simultaneous probabilistic modelling of multiple life-threatening conditions is feasible and accurate. Pending external validation, this approach may support emergency decision-making, reduce diagnostic uncertainty, optimise resource allocation, and ultimately improve patient outcomes.Figure 1