Introduction
Research reveals that artificial intelligence can identify acute heart failure more precisely than existing blood testing alone. Artificial intelligence (AI) uses methods such as Machine Learning and its variant Deep Learning to do tasks that need human intelligence. AI is rapidly being used to improve diagnosis, therapy, risk prediction, clinical care, and drug development in the field of cardiovascular medicine. Heart failure (HF) is prevalent and has been on the rise. HF is associated with a significant morbidity and death rate. The development of medicinal and surgical therapies tailored to HF patients depends critically on the early diagnosis of heart failure. Since HF is a complicated illness caused by structural and functional heart disorders rather than a single disease entity, even HF specialists may struggle to make an accurate diagnosis. Nevertheless, an accurate diagnosis is required before adequate therapy can begin.
Also, modern clinicians are challenged by quickly changing scientific information, new medicines, and the complexity of HF care recommendations, particularly in outpatient clinics. Artificial intelligence (AI) has become increasingly significant in cardiology as information and communication technology have advanced, allowing for the simple storage, acquisition, and recovery of large amounts of data and information. Through the implementation of an algorithm, AI may be of significant assistance in evaluating raw image data from cardiac imaging modalities (such as echocardiography, computed tomography, and cardiac MRI, among others) and ECG recordings.
There are two types of AI decision systems: white-box-based and black-box based. A white-box AI-based decision system uses supervised algorithms such as the decision tree algorithm to analyze gathered data and incorporates correlations and transparency among rules. A black-box AI features opaque algorithms, and the method and logic used to get the corresponding results are difficult to understand.
The use of electronic medical records (EMRs) with computer-based physician order entry (CPOE) capacity, in conjunction with clinical decision support systems (CDSS), has been proposed as a possible solution to the main issues in HA diagnosis. An effective CDSS must match individual patient characteristics with a clinical knowledge base, provide patient-centered assessments and recommendations, and ultimately present white-box proposals to physicians for their final selection.
In this paper, we report how the study analyzed the level of agreement between HF experts and AI-CDSS at a tertiary facility in Korea about HF diagnosis to identify the three forms of HF, namely HFrEF, HFmrEF, and HFpEF. During the study, an AI-CDSS was initially constructed utilizing a hybrid strategy combining expert-driven knowledge acquisition and ML-driven rule creation. Second, as a pilot clinical investigation, the AI-CDSS diagnostic concordance (degree of agreement) was tested in a test sample of patients with and without HF.
In the third step, the AI-CDSS diagnostic performance was prospectively assessed in consecutive patients coming to the outpatient clinic with dyspnea.
Study Methods
In this section, we describe the automated rules and the expected workflow of the system in order to describe how the CDSS was supposed to operate. Also, included is a summary of the assessment that was carried out to understand how the system operated and the issues that were discovered during the pilot implementation.
Data collection was conducted based on the following methods:
Retrospective cohort and Prospective pilot cohort. In the study, AI-CDSS was constructed using data from 1198 patients with and without HF, showing that AI-CDSS had extremely good diagnostic accuracy in these patients. AI-CDSS showed a remarkably high diagnosis accuracy in a prospective cohort of patients presenting with dyspnea to the outpatient clinic. In contrast, non-HF experts displayed a rather low level of HF diagnosis accuracy. As a result, AI-CDSS may be effective for HF diagnosis, particularly when HF specialists are unavailable. CDSS has been used in a variety of settings, including clinical diagnosis, preventative care, and chronic illness management.
Echocardiography. All photos were captured using normal ultrasound equipment and a 2.5-MHz probe. M-mode, two-dimensional, and Doppler measurements were obtained using conventional methods in compliance with the standards of the American Society of Echocardiography.
Generation of cardiovascular AI-CDSS. Following ML algorithm processing, the AI-CDSS employs patient data as the second crucial source of information (ML-driven approach). The AI-CDSS eliminates doctors’ reliance on knowledge engineers by focusing on a hybrid method of expert-driven knowledge acquisition and ML-driven rule creation. The clinical knowledge model (CKM), a traditional top-down decision tree, is built by domain experts (physicians) utilizing guidelines and their expertise in AI-CDSS; this is referred to as Expert-Driven Knowledge.
Study parameters. Patients with HF were characterized as having signs or symptoms of HF as well as either lung congestion, objective indications of left ventricular (LV) systolic failure, or structural heart disease. Two independent HF specialists with more than ten years of clinical experience verified the diagnosis of HF. The experts’ diagnosis was regarded as the gold standard. Patients were categorized as having HFrEF (LVEF 40%), HFmrEF (40% LVEF 50%), or HFpEF (LVEF 50%) based on their left ventricular ejection fraction (LVEF) on echocardiography.
Results
The study assessed indicators of the development of the cardiovascular system AI-CDSS. The overall total of 100 patients who attended the outpatient clinic with dyspnea was included. Three patients’ data were incomplete, hence the data from 97 individuals were included in the final study. 43 (44%) of the 97 patients had HF. The concordance rate of non-HF experts in this prospective cohort was 76%, whereas AI-CDSS was 98%. Particularly, among non-HF specialists, the diagnosis of HFmrEF and HFpEF was rare, but the diagnosis of no-HF was fairly common.
Conclusion
Despite a few limitations, researchers found that the AI-CDSS demonstrated good HF diagnosis accuracy across all HF subtypes.
Also, became clear the contribution of medical professionals is essential to the generation and confirmation of knowledge. Although, the level of competence differs amongst physicians, hence the clinical knowledge model (CKM) established by physicians in one hospital may differ from that developed by physicians in another hospital. Moreover, because the qualities in prediction mode (PM) are dependent on the patient data, they may differ from the variables specified in the recommendations. As a result, more research is required to verify the AI-CDSS in various study groups. Therefore, AI-CDSS may be effective for the diagnosis of HF, particularly in the absence of HF specialists.
Reference