Why the study?
Does an artificial intelligence algorithm using single- and 12-lead ECG improve the detection of left ventricular systolic dysfunction in hospital cohorts?
Does an artificial intelligence algorithm using single- and 12-lead ECG improve the detection of left ventricular systolic dysfunction in hospital cohorts?
Answers are grounded in this paper and cite their sources.
An AI algorithm applied to single- or 12-lead ECGs demonstrates high accuracy (AUC >0.9) for detecting left ventricular systolic dysfunction, highlighting its potential as a scalable screening tool.
“By far the greatest danger of Artificial Intelligence is that people conclude too early that they understand it.” Eliezer Yudkowsky The notion of early disease detection is located at the center of preventive medicine. This effort can ultimately culminate in mass-scale screening testing for common diseases such as cancer, heart disease, etc. Yudkowsky, a world-renowned artificial intelligence (AI) pioneer, has upheld this technology as the scientific method for refining the art of human rationality by overcoming bias. However, AI has become more of a contemporary buzz word in medicine with extremes of reception ranging from excessive enthusiasm to casual dismissal because of a myriad of human biases. Nevertheless, numerous researchers are investing resources to devise and evaluate modern AI tools. Artificial intelligence has been applied in prognostic algorithms aiming to potentiate early detection, homogenize preventive practices and eliminate human error. Artificial intelligence has already penetrated into clinical practice with the FDA recently approving an algorithm to prioritize critical chest x-ray review.1 Such has also been the case with AI applied in the automated echocardiographic recognition and interpretation of common 2D and 3D structures.2,3 Consequently, a bedside medical assistant (rather than a formally trained sonographer) can now scan and acquire high fidelity images involving the majority of commonly assessed clinically relevant echocardiographic variables. Heart failure—despite our best efforts to prevent or diagnose it before significant disease burden progression—continues to impact a significant population of patients leading to a major societal and economic burden because of disability, deranged quality of life, frequent hospitalizations, and reduced life expectancy, with a prognosis worse than most forms of cancer.4 One of the first applications of AI in a population of heart failure patients has been the implementation of machine learning algorithms for the phenomapping of heart failure with preserved ejection fraction.5 This has allowed for breaking down preconceived notions about a homogenized subset of patients with the heart failure syndrome, allowing for improved phenotyping of patients with heart failure with preserved ejection fraction into distinct mutually exclusive clinical syndromes. Furthermore, the concept of electromechanical coupling in heart failure is an established scientific paradigm allowing for electrocardiographic biomarkers to be leveraged for understanding myocardial mechanics. We previously demonstrated that electrocardiographic markers of repolarization heterogeneity are independently associated with echocardiographic metrics of diastolic dysfunction.6 Recently, Sugrue et al.7 showed that electromechanical window negativity, a novel echocardiographic derived biomarker, outperformed QT interval as predictor of life-threatening events in patients with long QT syndrome. This electromechanical coupling association was further established in a population without underlying QT syndrome.8 Others have identified electrocardiographic markers including QRS duration and abnormal QRS–T-wave axis to be associated with heart failure.9 In the current issue of ASAIO, Cho et al.10 bring AI and electromechanical coupling to practical fruition, as they report their results on a novel algorithm using single- and 12-lead ECG for the early detection of HFrEF by analyzing the data from 24,211 patients included in two large volume Korean hospital cohorts. Once initially derived in one cohort, their predictive model was internally and externally validated in the other cohort to confirm the model robustness. The authors sequentially employed the short-time Fourier transform followed by a convolutional neural network, a method applied in image processing, to detect left ventricular (LV) systolic dysfunction (ejection fraction <40%) based on subtle ECG patterns.10 Saliency maps were adopted to indicate which regions of the ECG were important to predict the result. Not surprisingly, high correlations were observed between markers of abnormal electromechanical coupling (QT interval, etc.) and the model-derived features associated with reduced ejection fraction. The performance of the AI algorithm was evaluated using the area under the curve (AUC) for the receiver operating characteristics to identify sensitivity and specificity. The authors found that the 12-lead ECG-guided AI algorithm could identify LV dysfunction with an AUC of 0.91 (95% confidence interval [CI], 0.902–0.925) in one hospital (derivation and internal validation), whereas AUC was 0.961 (95% CI, 0.951–0.971) at the second hospital (external validation). In addition, the single-lead ECG ability to detect LV dysfunction was comparable with an AUC of 0.874 (95% CI, 0.856–0.890) and 0.929 (95% CI, 0.911–0.946), respectively. The authors are to be commended for their work. By using a massive amount of electrocardiographic and echocardiographic data, they have been able to derive a quick, timely, and likely cost-effective tool for the initial screening for LV systolic dysfunction. It is important to acknowledge the opportunity for generalizability of the generated AI algorithm that was evaluated in different hospitals, including a diverse cohort of patient presentations, with truly outstanding performance for a screening test (AUC >0.9). Importantly, the authors demonstrated a comparable performance with a single-lead AI algorithm. The applicability of such a potential population screening algorithm cannot be missed. The development of a single-lead AI algorithm along with the advancement of the technology of wearable devices, such as smartwatches and smartphones, could prove to be a game changer in the detection of progressive or even transient myocardial dysfunction (Figure 1). This transformation will be twofold. First, the clinical setting of screening can shift from the clinic or ER/hospital to the patient’s home. This would provide us with lead time in the detection of LV dysfunction, potentially leading to earlier intervention. Whether this is feasible remains to be seen. Second, early screening would be applied to a larger population as these wearable devices become increasingly popular among the general public. Consequently, this could mean more impactful population-wide interventions, perhaps even allowing strategies to prevent development of overt (stage C) heart failure to be more readily employed for patients with asymptomatic LV systolic dysfunction (stage B HFrEF).Fig. 1Caution must be exercised when it comes to interpreting these results. Artificial intelligence is a scientific discipline that uses machine learning to remove human bias and help identify data patterns and associations.11 Models generated with the assistance of AI are not exempt from the general rules of prediction model derivation. The Transparent Reporting of a Multivariable Prediction Model for individual Prognosis or Diagnosis published in 2015 provides an outline of the key reportable elements when developing, evaluating, and validating prediction models.12 If not applied with clear and specified methodology and under appropriate guidance, AI outputs can lead to poor transparency and reproducibility.13 These latter methodological shortcomings have caused AI to be considered a “black box” by some; guidelines that shed light on the differences of traditional statistical validation methods versus methods adopted by AI, the basics of machine learning, and the implementation in clinical medicine attempt to address this “black box” reputation.14 Ultimately, AI has to be thought of as another tool for assistance of clinical reasoning, particularly advantaged by its ability to reduce human bias. Keeping this frame of mind, without excessively inflating expectations, is perhaps the balanced view.
No takes yet. Share an insight, caveat, or question.
Mastoris et al. (2020) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: