AI-Enhanced Cardiology Takes Another Step Forward

By integrating a convolutional neural network with regular electrocardiograms (ECGs), the presence of low ejection fraction, a recognised indicator of Asymptomatic Left Ventricular Systolic Dysfunction, was identified.

This post was authored by John Halamka, M.D., who serves as the president of the Mayo Clinic Platform, and Paul Cerrato, a senior research analyst and communications specialist at the Mayo Clinic Platform.

Although not well recognised in the field of medicine, asymptomatic left ventricular systolic dysfunction (ALVSD) is associated with an elevated risk of heart failure and mortality for affected individuals. Regrettably, the detection of ALVSD is not a straightforward task. The condition is distinguished by a low ejection fraction (EF), which quantifies the amount of blood expelled by the heart each contraction. It can be easily detected through the use of an echocardiography. However, due to the high cost associated with the technique, it is not advisable to utilise it as a standard screening method for the general population. The integration of a newly created algorithm, boosted by artificial intelligence (AI), with an electrocardiogram (ECG) has enabled the identification of low ejection fraction (EF). This represents one of several advancements that will inevitably render machine learning an indispensable component of every clinician's repertoire.

The recently developed algorithm, which was collaboratively developed by multiple clinical departments at Mayo Clinic and Mayo Clinic Platform, has been made available online through publication in the esteemed journal Nature Medicine. The EAGLE experiment encompassed a total of 22,000 participants, who were allocated into distinct intervention and control groups. The trial was overseen by a cohort of 358 clinicians, who were affiliated with 45 clinics and institutions. The algorithm/electrocardiogram (ECG) was employed for the assessment of patients in both cohorts. However, only doctors assigned to the intervention group were granted access to the artificial intelligence (AI) outcomes while making decisions regarding the necessity of an echocardiogram. Upon further examination, it was determined that 49.6% of patients, whose physicians were granted access to the artificial intelligence (AI) data, underwent echocardiogram. In contrast, only 38.1% of patients did the same procedure without access to the AI data. 

The odds ratio was calculated to be 1.63, with a p-value of less than 0.001. Xiaoxi Yao, along with colleagues from the Kern Centre for the Science of Health Care Delivery at Mayo Clinic, conducted a study in which they observed an increase in the diagnosis of low ejection fraction (EF) in the entire cohort. The control arm had a diagnosis rate of 1.6%, while the intervention arm had a diagnosis rate of 2.1%. This increase was also observed among individuals who were identified as having a high probability of having low EF. The utilisation of the AI tool resulted in a significant 32% increase in the overall diagnosis of poor EF among primary care physicians, as compared to the diagnosis rate observed among patients who got standard care. In terms of absolute figures, the AI system produced a total of five additional low EF diagnoses per 1,000 patients tested, as compared to the standard care.

Previous studies conducted on the neural network utilised in the development of the artificial intelligence tool have demonstrated robust empirical backing. There has been an increasing amount of criticism from prominent figures in the medical field over the expeditious development of AI-driven algorithms. This critique stems from the observation that a significant proportion of these algorithms lack the necessary scientific basis to adequately support their use in the context of direct patient care. The AI developers are currently facing criticism regarding several concerns. These concerns include the utilisation of algorithms that are derived from a dataset lacking validation from an external dataset, excessive dependence on retrospective analysis, limited generalizability, and the presence of various types of bias. 

These issues are thoroughly examined and discussed in our publication titled "The Digital Reconstruction of Healthcare." The concerns raised were effectively addressed by the investigators of the EAGLE trial by the use of their methodology on many patient cohorts. A previous investigation employed the aforementioned method to train a convolutional neural network using a dataset including more than 44,000 patients from the Mayo Clinic. Subsequently, the network's performance was evaluated by conducting tests on a separate cohort of almost 53,000 patients. Although the present study adopted a retrospective methodology, subsequent investigations have corroborated the efficacy of the algorithm in clinical settings by the implementation of a prospective design. 

The study referenced at the outset of our blog post is not only characterised by its prospective design, but also by its pragmatic approach, thereby mirroring the practical context in which healthcare professionals operate. Conventional randomised controlled trials are characterised by substantial budget allocation, prolonged duration of implementation, and typically entail an extensive set of patient eligibility requirements. In contrast, the EAGLE trial was conducted among patients in real-world clinical settings.

Bagikan Artikel ini