Introduction Atrial fibrillation (AF) is a cardiological disease, its risk factors and mecha-nisms are, however, often rooted in non-cardiological comorbidities, intro-ducing complexity in patient treatment. We present a clinical decision sup-port system (CDSS), which aims to mitigate potential negative impact of the cross-disciplinarity of AF, by predicting comorbidities related to AF mecha-nisms based on provided symptoms.
Methods MIMIC-III, a critical care database containing electronic health records (EHRs) from 38,597 patients is used. Diagnostic codes are captured, and symptoms are identified from free-text clinical notes using named entity recognition, and entity linking. Diagnoses and symptoms are mapped to a domain ontology, acting as a knowledge base. The obtained dataset of symptoms and corresponding diagnoses for each patient is employed to train an ensemble of Naive Bayes classifiers, with each classifier predicting the probability of a given diagnosis using the present symptoms. Performance is assessed using top-k accuracy and 5-fold cross validation. A CDSS integrat-ing the knowledge base, classifier ensemble, and a communication interface is implemented.
Results 10,277 patients with AF were identified in the database. Patients had 7.58+-4.19 symptoms, and 2.92+-1.96 comorbidities. The ensemble of Naive Bayes classifiers predicts the probability of 18 comorbidities from up to 43 symptoms with a top-k accuracy of 0.51, 0.66, and 0.77 for k = 1, 3, and 5 respectively.
Conclusions The present work demonstrates the ability of probabilistic classifiers to iden-tify AF related comorbidities from patient reported symptoms. The use of such systems could support clinicians in the identification of comorbidities and mechanisms driving the arrhythmia, thus providing support in the selec-tion of the optimal treatment.