Predicting heart failure with preserved ejection fraction: revisiting an old friend with new knowledge

This research-article refers to ‘Artificial intelligence assessment for early detection of heart failure with preserved ejection fraction based on electrocardiographic features’ by J. Kwon et al., on page 106. Heart failure (HF) is a growing health problem affecting millions of people worldwide. Heart failure with preserved ejection fraction (HFpEF) accounts for at least 50% of HF diagnoses and its prevalence is expected to increase due to an aging population and a growing burden of obesity, hypertension, and diabetes mellitus.1 Although HFpEF and HF with reduced ejection fraction (HFrEF) share a common phenotypic spectrum of symptoms, clinical signs, and poor health status, they are distinct entities.2 The heterogeneity and poorly understood pathophysiology of HFpEF makes it a complex syndrome with to date no therapies demonstrating clear benefits in large trials.3 Moreover, adverse events related to HFpEF have been steadily increasing which accounts for substantial morbidity, mortality, and increasing health care cost.4,5 Hence, the need for new techniques to define and diagnose this population with the possibility of precision treatment and improving prognosis. Machine learning (ML) is everywhere around us and has made its way into clinical practice with emerging applications to interpret echocardiography, radiologic images, and pathology slides. Yet, the earliest form of ML within cardiology is the machine read an electrocardiogram (ECG) over 40 years ago. Where at the time accuracy was low, current neural networks achieve a high diagnostic performance outperforming cardiology residents.6 Therefore, it is not surprising that we are revisiting this old friend with new knowledge. In this issue of European Heart Journal – Digital Health, Kwon et al. propose an ensemble neural network-based deep learning model (DLM) combining electrocardiographic data and four clinical variables to predict HFpEF.7 A total of 22 148 patients from a Korean hospital were retrospectively included of whom data of 20 169 patients were used to derive the DLM. Internal validation was performed in the remaining 1979 patients and external validation in 11 955 patients from a second institution. All patients underwent an ECG and echocardiogram, and left ventricular diastolic dysfunction (LVDD) was defined according to current guidelines.3 Cases with LVDD in the presence of a (near) normal ejection fraction with symptoms or signs of heart failure were deemed as having HFpEF. The algorithm resulted in an area under the curve (AUC) of 0.87 [95% confidence interval (CI) 0.86–0.88], sensitivity of 0.77 (95% CI 0.75–0.79), and specificity of 0.82 (95% CI 0.81–0.83) in the external validation cohort, results that are similar to a recently published clinical risk score.8 Additionally, in patients initially classified by the DLM as having HFpEF without evidence of LVDD by echocardiographic assessment (false positives), a significant proportion of ‘high-risk’ patients eventually developed HFpEF during follow-up (33.6% vs. 8.4%, P < 0.001). Overall, the study by Kwon et al. confirms the feasibility of a DLM to discriminate between patients with and without HFpEF and highlights the possibility of screening and early diagnosis by means of the ECG. In contrast to available risk scores, the strength of this approach lies within the utilization of an inexpensive and almost universal available tool in combination with four patient characteristics to predict a prevalent disease. Their analysis regarding the ‘false positives’ suggests that subtle ECG abnormalities may identify early stages of HFpEF when overt disease is not yet present, a phenomenon also described by Attia et al.9 in their work on HFrEF. Furthermore, by applying guided gradient backpropagation and a sensitivity map analysis the authors visualize the electrocardiographic regions(R-wave and T-wave axis) that significantly affected the DLM. Not only does this help to unravel the ‘black box’ myth surrounding ML, it also expands our knowledge on subtle disease-specific electrocardiographic features improving ECG interpretation in daily practice.10,11 Although showing great promise, we should be careful before applying the presented algorithm in clinical practice. This DLM was developed upon a retrospective analysis which increases the possibility of incidental findings and correlations. The authors already noted this limitation that will be addressed in a prospective study designed to reproduce the current findings. More importantly, defining HFpEF is challenging and only combining the presence of LVDD with signs or symptoms of HF might be an oversimplification of a complex disease. This raises the question if the DLM truly predicts HFpEF or LVDD, a problem that should be addressed by using the HFA-PEFF diagnostic algorithm in future studies.12 By using the Youden J-statistic sensitivity and specificity were optimized leading to a very good diagnostic accuracy (AUC) of the DLM. However, in the case of a screening test one might be more interested in maximizing either the sensitivity or the specificity in order to effectively rule-out or rule-in disease, respectively. Moreover, the trade-off between these variables introduces false positives or false negatives leading to unnecessary testing or false reassurance of patients. These problems are not apparent when focusing on the diagnostic accuracy of a model but become important when regarded from a clinical and economical perspective. Last, DLMs may demonstrate a lower accuracy when used beyond their training and validation data. Consequently, extensive external validation of the model is necessary prior to clinical implementation. This process should not only be repeated in a different institution, as was the case here, but also in a different geographic location (beyond Korea) to provide evidence of its applicability. Multicentre collaboration or open data sharing are necessary to facilitate this on a large scale. Despite these challenges, the findings by Kwon et al. offer a glimpse into a future in which we are able to identify those at high risk of developing HFpEF in a simple but efficient way. Perhaps, this may even be applicable in regions with limited cardiovascular care with the use of wearables or lifestyle devices. Early recognition of these patients can initiate lifestyle intervention and treatment of known risk factors. The latter may lead to improved quality of life, significant reductions in HF associated adverse events and reduced costs benefiting patients, physicians, and healthcare systems. Also, sub-classification of patients with HFpEF using deep phenotyping may help to approach this heterogeneous disease and selectively direct individual patients towards clinical trials that match their underlying pathophysiologic features. A strategy eventually leading to effective patient management.13,14 Additionally, follow-up of untreated patients might help delineate the complex natural course of this multisystem disease. Machine learning is becoming increasingly important in clinical medicine and medical research, yet we must remember to have a healthy amount of scepticism when interpreting its results. Hopefully, coupled with new imaging technology and molecular analysis, machine learning will assist in a more accurate diagnosis, targeted and efficacious treatment, and a true personalized approach for HFpEF. The opinions expressed in this article are not necessarily those of the Editors of the European Heart Journal – Digital Health or of the European Society of Cardiology. I thank dr.Rutger-Jan Nuis and dr. Isabella Kardys for their advice regarding the preparation of this research-article. Conflict of interest: none declared.