Mike Marber , MBBS, PhD, FRCP
Professor of Cardiology, St Thomas’ Hospital Campus, King’s College London, London, UK
Big data, AI, and mHealth- the digital evolution of cardiology
This issue of our journal is dedicated to Digital Cardiology, a term that is difficult to define, with a surprisingly long history. 1,2 Over the last 60 years the complexity of the cardiac signals we digitize has increased from the voltage of the single-lead electrocardiogram to electromagnetic radiation encoding 4-dimensional MRI datasets of high anatomical and temporal density. 1-3 Similarly, the duration of digital recordings has increased from seconds to many years, with implantable devices. In addition, we now acquire data in free-living people through direct-to-consumer wearables (eg, Apple Watch) and diagnostics (eg, AliveCor). These changes have been accompanied by a transition to digital electronic health records, whole-genome sequences, and footprints of our everyday social and commercial behaviors. Collectively, these have created huge lakes of interconnected data that can be mined by computer algorithms to reveal patterns that indicate current or future disease. The term “big data” is often used as a collective noun to describe these data lakes. In parallel, the analysis of big data has been supercharged by sophisticated computer algorithms with “artificial intelligence” (AI) learning from the information they examine. Mobile health (mHealth) refers to the mobile phones and other portable devices that can be used both to collect, process, and/or feed back information to improve an individual’s health and happiness. All these topics are discussed in the current issue under the umbrella of Digital Cardiology.
In navigating this issue, the article on artificial intelligence in cardiology by Prof Lionel Tarassenko (p 8) is a good place to start, since it is written from a historical perspective by an engineer specializing in health care applications. It introduces the concepts that underpin AI and how the algorithms have advanced from simple, understandable “if-this-then-that” to the complex “black-box” multi-layered, deep-learning systems of today. This added complexity means we can no longer use our framework of how the cardiovascular system works to understand how an expert system has processed the data to come to a conclusion. In common with many cardiologists, I was attracted to the specialty by the strong logic of the underpinning physiology and biochemistry. This science is validated by its ability to shape successful drugs and devices. I therefore find the black-box nature of AI uncomfortable and suspicious. This cautious undertone ripples through all the articles in this issue.
The common flaws of AI and how to reality check its claims are picked up in the Refresher Corner by Dr Pablo Lamata, also a bioengineer (p 33). This article further examines the gap between AI and the inductive-deductive reasoning of humans. The article also provides practical advice on how to assess AI and highlights the common pitfall of results that only work in the dataset used to train the AI, (aka overfitting, lack of generalizability, spurious correlations, bespoke output).
One cause of overfitting is the big data training set containing incorrect entries, leading to noise. As Prof Gennaro Galasso explains (p 4), Big Data has the attraction of the Four Vs (volume, variety, velocity, and value). However, the way the data are collected is often not standardized and it therefore lacks the crucial V of veracity! This gives rise to the derogatory acronym GIGO (garbage in, garbage out). Dr Lamata encourages us to examine the revelations of AI through the lens of causality; do they make sense? In common with Prof Tarassenko, Dr Lamata predicts a future of augmented intelligence where AI is used to reduce the monotony of repetitive human tasks, such as drawing endocardial contours. Humans, and their abilities, lie above the AI and are in charge!
The practical applications of AI and augmented intelligence in echocardiography are covered by Dr Jonathan Sen and Prof Thomas Marwick (p 21) in their in-depth article. This article also covers some of the statistical methods used to train the AI. After reading this I was convinced that echocardiography is an area where AI has been proven to improve both data acquisition (transducer position) and interpretation – for example by removing the subjectivity and variability that creeps in with the routine measurements to calculate ejection fraction. However, it is also clear that the areas of success so far have been in augmenting, rather than replacing, human function.
The remaining articles of this issue concern non-AI aspects of digital cardiology. Dr Arvind Singhal and Prof Martin Cowie (p 12) cover wearable devices and discuss where data are stored, who owns the data, and general data protection regulation. These are complex areas best left to experts who can evaluate the compliance of apps and/or devices and list those approved on official sites. This process, and where to access compliant apps, is discussed in this article and that by Dr Itzhak Gabizon (p 17). Both highlight the bewildering number of apps and devices available and the rarity of any formal external validation or assessment of efficacy; making it difficult to separate useless plaything from useful intervention. Even with substantial investment and a very defined purpose, data protection and technical difficulties can frustrate delivery, as Dr Gabizon illustrates with COVID tracking apps. Dr Gabizon also provides useful practical advice about the patient characteristics that should be used to match them to the most appropriate app and provides examples of apps delivering augmented cardiac rehabilitation.
Finally, for light relief, there is a highly original perspective on the dissemination of medical information through social, digital media, and in particular Twitter. Prof Darrel Francis (p 36) is an academic cardiologist with 13K followers on Twitter.4 Together with Dr Pranev Sharma he has categorized tweets and discusses the advantages and disadvantages of the “twittersphere” and of “twitterati” with cogent and amusing case studies. I like their ideas on collective peer review and the fact that many brains can make light work, but also agree that the conventional hierarchy of cardiology politics and peer review has merits, even though it lacks the dynamism, excitement, and instant gratification of social media.
I enjoyed reading this issue and I think the authors have balanced enthusiasm with scrutiny. It’s all too easy to get caught up by the bandwagon of digital cardiology. In the words of Prof Tarassenko, “the state of AI hype has so far exceeded the state of AI science.” Digital cardiology needs to be taken with a pinch of analog salt!■
Disclosure/Acknowledgments: The author has no conflict of interest to declare.
- 1. Pipberger, HV, Freis ED, Taback L, Mason HL. Preparation of electrocardiographic data for analysis by digital electronic computer.
- 2. Stallmann FW, Pipberger HV. Automatic recognition of electrocardiographic waves by digital computer.
- 3. Knott KD, Seraphim A, Augusto JB. et al. The prognostic significance of quantitative myocardial perfusion: an artificial intelligence-based approach using perfusion mapping.
- 4. https://twitter.com/ProfDFrancis?s=20. Accessed June 2020.