Machine Learning Models for Differentiating Causes of Dyspnea in Emergency Department Patients: A Retrospective Study.

Background: Dyspnea is a frequent symptom in emergency departments (ED) with multifactorial causes, including cardiac and pulmonary conditions. Accurate and timely diagnosis is crucial to guide appropriate management and improve patient outcomes. Machine learning (ML) may aid this process. Methods: This retrospective study analyzed 787 adult patients presenting with dyspnea to the ED at Kantonsspital Baselland (Switzerland) in 2022. Clinical, laboratory, diagnostic data and final diagnoses were collected. ML models including decision trees, random forest and boosted decision trees were trained to classify dyspnea etiologies (cardiac vs respiratory) and predict final diagnoses. Performance metrics included accuracy, sensitivity, and specificity. Results: The most common diagnoses were decompensated heart failure (28.4%), pneumonia (26.4%), and COVID-19 (17%). Binary classification into cardiac vs respiratory causes achieved the highest performance (accuracy: 89.6% with boosted trees). Multiclass prediction of specific diagnoses yielded lower performance (accuracy: 36.8%). CRP, BNP, and cough emerged as key predictive features consistent with established clinical knowledge, supporting model interpretability. Comorbidities, though clinically relevant, showed limited predictive value. Conclusion: ML algorithms show promise in supporting triage by distinguishing broad etiological categories of dyspnea, such as respiratory versus cardiac origins. While the models demonstrate useful classification performance, their limited sensitivity for specific diagnoses underscores the need for larger, more diverse datasets and advanced modeling approaches.