Diagnosing heart failure (HF) presents a significant challenge in healthcare, often requiring the expertise of specialized clinicians. The complexity of HF diagnosis and the increasing prevalence of the condition necessitate innovative approaches to improve diagnostic accuracy and efficiency. Artificial intelligence (AI), particularly Clinical Decision Support Systems (CDSS), emerges as a promising tool to aid physicians in this critical task. This article delves into the diagnostic capabilities of an AI-CDSS specifically designed for heart failure, evaluating its performance through rigorous retrospective and prospective studies.
The development of this AI-CDSS for cardiology employed a hybrid knowledge acquisition strategy, blending expert-driven insights with machine-learning-driven analysis. This approach allows the system’s knowledge base to evolve continuously, enhancing its diagnostic precision in the complex landscape of heart failure. To rigorously assess its effectiveness, the AI-CDSS underwent testing using a retrospective cohort of 1198 patients, encompassing both individuals with and without heart failure. This cohort was divided into a training dataset (n=600) for system development and a test dataset (n=598) to evaluate its performance in a controlled setting.
The results from the retrospective analysis were compelling. The AI-CDSS demonstrated a high degree of concordance with established diagnoses, achieving an impressive 98.3% concordance rate within the test dataset. Breaking down the performance across different heart failure classifications revealed even more granular insights. The system achieved perfect concordance (100%) in diagnosing heart failure with reduced ejection fraction (HFrEF) and heart failure with mid-range ejection fraction (HFmrEF). For heart failure with preserved ejection fraction (HFpEF), a particularly challenging subtype to diagnose, the concordance rate remained exceptionally high at 99.6%. Even in cases of no heart failure, the AI-CDSS maintained a strong concordance rate of 91.7%, indicating its ability to accurately rule out the condition.
To further validate the AI-CDSS in a real-world clinical setting, a prospective pilot study was conducted involving 97 patients presenting with dyspnea (shortness of breath) in an outpatient clinic. Dyspnea is a common symptom with diverse underlying causes, including heart failure, making it a relevant presentation for evaluating diagnostic tools. In this prospective cohort, 44% of patients were ultimately diagnosed with heart failure. The AI-CDSS exhibited a remarkable 98% concordance rate with the diagnoses made by heart failure specialists. Notably, when comparing the diagnostic accuracy of non-heart failure specialists to heart failure specialists, the concordance rate was significantly lower at 76%. This highlights the potential of AI-CDSS to bridge the diagnostic gap, particularly in settings where specialist expertise may not be readily available.
In conclusion, the AI-CDSS demonstrated exceptional diagnostic accuracy for heart failure across both retrospective and prospective evaluations. These findings suggest that AI-CDSS holds significant promise as a valuable tool for heart failure diagnosis, potentially improving diagnostic precision and efficiency, especially in scenarios where access to heart failure specialists is limited. The integration of artificial intelligence into diagnostic workflows can empower clinicians, enhance patient care, and contribute to more timely and accurate identification of heart failure.
