Moving Forwards from Clinical Trials to a More Individualized Management of Treatments in Heart Failure? Great Value of Doppler Echocardiography Data

This article refers to ‘A machine learning derived echocardiographic algorithm identifies people at risk of heart failure with distinct cardiac structure, function, and response to spironolactone: Findings from the HOMAGE trial’ by M. Kobayashi et al., published in this issue on pages 1284–1289. The management of heart failure (HF) has increasingly improved these recent years.1 New therapeutic options have emerged. Nevertheless, the prescription of these treatments is not systematic and perhaps should not be systematic but adapted to individual specificities.2 In these past years, the relevance of the ‘heart team decision-making process’ has been emphasized and even recommended by guidelines.1, 3 Heart teams are gathering colleagues, experts, that are supposed to decide based on the clinical files of a patient what should be the optimal management. It is much more efficient than a decision-making process based on a single view. However, it could be biased by many factors or parameters considered by the heart team or not. The decisions made are most often appropriate but the same patient discussed in different heart teams or at different heart team meetings may end up with different medical managements. Therefore, risk stratification and management of complex diseases like HF could take advantage of new approaches. Recently, a ‘decision support tool’ has been demonstrated to improve the appropriateness of mineralocorticoid receptor antagonist (MRA) prescription.4 An automated, patient-specific, electronic health record-embedded alert was found to increase MRA prescription. It has been validated in a study including 2211 patients using a cluster randomized methodology.4 In this issue of the Journal, the work done by Kobayashi et al.5 is going in that direction. We can expect from our computers (using large databases and solid statistical algorithms) some help for guiding us in being more relevant in managing our patients. We can get from the clinical file of a dedicated patient some characteristics, some quantitative characteristics like to ones we get from Doppler echocardiography reports. These Doppler echocardiography data have a great value in better phenotyping patients.6 This phenotyping based on simple measurements, like it is the case in Kobayashi et al. work, or more advanced ones (that could be calculated automatically by our Doppler echocardiography software) could help us making a large step towards an individualized patient management.5 The content of these data is huge. The remarkable work reaching to “e'VM algorithm” is nice because it is a really simple and robust echocardiographic data-driven algorithm (based on e', and left ventricular volume and mass). It was first demonstrated to allow the identification of echocardiographic phenotypes with a high risk of developing HF.6 It has been demonstrated for cardiac resynchronization therapy or other disease management.7, 8 According to Kobayashi et al.,5 we can expect from our simple assessment of the diastolic function characteristics, a guide for insuring the impact of a treatment on the course of HF (Figure 1). It means that we could better individualize our management. Some patients required MRA and it is nicely demonstrated in the present manuscript. Some are less relevant for this treatment and then might take advantage from other strategies. The e'VM algorithm identified patients at risk of HF with different characteristics who had different responses to spironolactone treatment.5 This simple approach is reported to perform better than the guidelines and recommendation algorithms. e'VM classification-guided preventive HF trial would be worthy of investigation to ascertain whether this new echocardiographic classification can be applied in a decision support system to identify individuals at risk of HF, and to identify those who require early initiation of specific disease-modifying agent(s). We are at the beginning and many challenges remain but thanks to works like the one done by Kobayashi et al., we confirm that individualized management of medical treatment could be proposed and that could be implemented using a computer-based guide. This guide seems to take largely advantage of cardiac imaging and, particularly, echocardiographic data. Authors should be applauded for their achievement.5 Although artificial intelligence has a huge potential to help the clinicians to better stratify cardiac disease risk and to cluster phenotypes that could benefit from precision medicine in terms of management and treatment, several issues remain unsolved, including the ethical and regulatory aspects, the selection bias in terms of patients and parameters, and the external validation. Transparency and reproducibility as well as impact on the outcome are often missing too. The next step is to confirm these results on ongoing databases with more gender balance, coming from different sites or healthcare systems. The clustering should adapt to the moving characteristics of the populations we are dealing with, also to the change in guidelines and the emergence of new therapies. The algorithm should be regularly retrained and revalidated. Coming back to cardiac resynchronization therapy, it has been observed how the guidelines have been relatively stable over the years but the characteristics of treated patients have changed because the phenotypes of our patients are rapidly changing.9 Clinical decision support systems embody one of the eagerly awaited applications of artificial intelligence and machine learning for patient care.10 By leveraging the power of ‘big data’, these technologies promise the most effective practice medicine ever.11 Nevertheless, artificial intelligence algorithms, or the decision support tools alerting clinicians to their complex outputs, will still largely depend on small to moderate size of the validation cohorts. Conflict of interest: none declared.