info@biomedres.us
+1 (502) 904-2126
One Westbrook Corporate Center, Suite 300, Westchester, IL 60154, USA
Site Map
ISSN: 2574 -1241
ISSN: 2574 -1241
Received: September 03, 2024; Published: September 13, 2024
*Corresponding author: Malav Bhimpuria, GP Principal, Clinical Lead, Older People’s Mental Health, Cambridgeshire and Peterborough Integrated Care System, UK
DOI: 10.26717/BJSTR.2024.58.009180
Abbreviations: AI: Artificial Intelligence; ML: Machine Learning; DL: Deep Learning; FFNN: Feed Forward Neural Networks; MLP: Multilayer Perception; RNN: Recurring Neural Network; RBFN: Radial Basis Functional Network; MNN: Modular Neural Network; CNNs: Convolutional Neural Networks; ECG/EKG: Electrocardiography/ Electrocardiogram; STEMI: ST Elevation Myocardial Infarction
The future of health and healthcare is of concern to many, with the worldwide increase in both chronic diseases, like heart disease but also infectious diseases such as Covid-19. Cardiovascular disease for instance, is the principal cause of morbidity and mortality globally [1]. One of the lessons of the recent pandemic is that the world is increasingly connected and with a growing demand for healthcare, it is imperative that medical advances in healthcare are rapidly scalable and actionable, to facilitate clinicians in their decision-making processes. This is where artificial intelligence (AI) could help [2]. Cardiology is at the forefront of this AI revolution in medicine [1], enabling better preventative cardiology, improved diagnostic imaging, more readily available arrythmia monitoring, new protocols in heart failure and more sophisticated interventional cardiology. It is one of the few medical specialities where AI has been examined systematically [3]. At the heart of this new dawn is “big data” and digitising healthcare records is imperative. Indeed, in the UK, the NHS has set the ambitious target to digitise the entire healthcare field and have digital health and social care records for most of the population by 2025 [4]. There are many benefits to this, such as seamless transfer of medical information between various professionals to improve patient care, reducing over-investigation and treatment (as previous tests are now readily available to view) and therefore reducing costs to the system.
Using AI in healthcare could also eliminate human error and lead to diagnoses that are more accurate. Indeed, in the future, it has even been mooted by researchers at Oxford and Yale [5] that AI could replace surgeons in thirty years! There are however some negatives to using AI. Some argue that as AI has been trained on data developed by humans, it could actually enhance existing human biases [6]. It is incredibly good at processing vast amounts of data in short amounts of time and thus identifying patterns, but does it ignore the outliers? Therefore, its generalisability to new populations may cause poor outcomes and unwise decision making [7]. It may even make the entire healthcare system more complex for clinicians and patients to navigate and far from reducing healthcare costs, it could increase them. Will this therefore widen health inequalities across the globe? Finally, the public (and some clinicians) may not fully understand AI and there are valid concerns over its ethics and privacy, with a fear that large commercial organisations could “steal” sensitive patient data and use it inappropriately [7].
In this essay, I will endeavour to answer these questions and discover ways in which AI is used in cardiology today and could be used in the next ten years. I will also discuss some of the ethical drawbacks and propose a common AI regulatory framework to address these concerns. Big data gives big solutions but to avoid too many unintended consequences, we also need overarching regulation and quality control. This does not necessarily have to be “big brother Governments” but frameworks that are both robust enough to reassure the public and light-touch enough for the sector to innovate. But first, let me first define my terms.
AI is a computer science, where” …a machine has an ability to generalise learning in order to efficiently achieve complex tasks autonomously” [8]. The Oxford Dictionary has a more philosophical definition, stating AI “…as the capacity of computers or other machines to exhibit or simulate intelligent behaviour” [9]. One could argue that AI could be used by computers to copy the behaviour of humans, with potentially far-reaching consequences. Alan Turing, the English mathematician and World War Two codebreaker, is regarded by many as the “father of AI” following his landmark article in 1950 entitled “Computing Machinery and Intelligence” [10], (Figure 1), where he posed his famous question “Can machines think?” He recommended definitions for “machine” and “think” and proposed the “Turing Test”: if a human is interacting with another human and a machine cannot distinguish between the two, then the machine is “intelligent.” This is the gold-standard method to identify intelligence in an artificial system. Turing himself believing that by the year 2000, computers would pass this test. This has still not happened, and many experts believe we will have to wait many more years before AI can truly replicate human behaviour.
Figure 1
Machine Learning
Another AI pioneer (McCarthy, 1956) used the expressions “machine learning” (ML) and “deep learning” (DL) [11]. ML is a subdivision of AI using sophisticated computing and statistical algorithms to enable computers to analyse datasets in a timely and efficient manner. It has three types:
1. Supervised learning, based on analysing previously human
labelled datasets to develop models to predict future events,
2. Unsupervised learning, where the datasets have not been
categorised before and the models seek to discover hidden
relationships and
3. Reinforcement learning which is a reward-based system of
positive or negative reinforcement within the system environment.
With repeated positive reinforcement, the AI model
can improve its performance [12].
Deep Learning
DL however uses neural networks, like in the human brain to extract more meaningful patterns from complex datasets. It is a subfield of ML. This could be particularly useful in other medical specialities where pattern recognition is paramount, like dermatology. DL can also be used for classification, clustering of data, reduction of data, language processing, computer vision and predictions.
There are many types of neural network, such as
a. Feed Forward Neural Networks (FFNN): the simplest type,
where data travels in just one direction. Can be single layered or
multi-layered and used in simple classification and facial recognition.
b. Multilayer Perception (MLP): data travels in various layers
of artificial neurones. Used in more complex classification and
speech recognition.
c. Recurrent Neural Network (RNN): the output of one layer
becomes the input of the next layer. This forms a feedback loop
and has internal memory. It is used in text to speech processing.
d. Radial Basis Function Network (RBFN): the input is in the
centre and the output combines outputs from the radial basis
function. It is used in predictive analysis.
e. Modular Neural Network (MNN): there are multiple different
networks working independently and undertaking different
tasks to achieve the output. This can manage substantial amounts
of data.
f. Convolutional Neural Networks (CNNs): the most popular
in DL as it can build networks which respond to visual inputs like
the visual cortex in humans [13].
Drawbacks of AI Models
As they need “big data,” small datasets can pose problems, meaning that techniques such as data augmentation need to be applied to “train” the model more efficiently [14]. This can cause further problems as more assumptions are used in the models and occasionally irrelevant data patterns memorised. Outliers (data at the extremes) may carry too much or too little weight in the final model and as all human beings are different, how can we have a “one size fits all” algorithm? Additionally, with big data sets more things will become statistically significant though in a real-world analysis not everything that one counts needs to be counted. So certain outputs need less emphasis than others. Finally, as more data is used, variation and bias also increase, which could compromise the ultimate outputs in AI.
Cardiology is a branch of medicine which specialises in diagnosis and treatment of diseases and conditions in the heart. It was initially established as a medical field in 1628 [15] by William Harvey who was able to demonstrate the circulation of blood. Although not initially using the term “cardiology,” Harvey researched the anatomy and physiology of the heart itself and overall interest in this field blossomed, helped by the invention of the stethoscope by Laennec [15], electrocardiography (ECGs) in 1903 by Einthoven (Figure 2) and cardiac catheterisation by Forssman (1929) which enabled direct visualisation of the four chambers of the heart and direct measurement of its pressures. The invention of echocardiography in 1950 [16] which enabled the heart to be visualised using ultrasound waves was a further step change in Mankind’s understanding of this organ, enabling cardiologists to detect congenital heart disease (birth defects affecting the functionality of the heart), heart failure (where the heart muscle does not pump the blood at the right pressures) and many other conditions [16].
Figure 2
With an estimated 620 million people globally living with some form of heart or circulatory disease [17] and with an ageing population, that number is due to rise. To try to mitigate against the disease burden of these conditions and improve prognosis and prevention, the application of AI to cardiology is becoming increasingly popular. From predictive risk scores to arrythmia detection, heart failure treatment to imaging and interventional cardiology, AI is increasingly making its presence felt both today and will do in the next decade, leading to a change in basic assumptions in cardiology: the AI paradigm shift (Figure 3).
Figure 3
AI and Preventative Cardiology
Using ML to provide a data-driven approach, patients can be risk stratified to see who is at an elevated risk of complications following myocardial infarction (Shetty, et al. [18]). These algorithms use several variables but avoid prior assumptions. They incorporate a plethora of risk factors, including some non-traditional and unknown ones. (D’Ascenzo, et al. [19]). Indeed, a recently proposed MERC model as improved 30-day mortality prediction after an ST elevation myocardial infarction (STEMI) and is thought to be of value in low- and middle- income countries. In such resource-limited countries like India, if patients at high risk of mortality can be identified, then they might benefit from intensive treatment after their heart attacks. This study (NORIN-STEMI) enrolled 3,635 patients, so had a large sample size. The ML model had a predictive ability with a sensitivity of 85% and an accuracy of 75% [18]. These are important statistics which could guide future management. In the next decade, as ML algorithms become increasingly sophisticated, accuracy of these models should increase, and more work can be done in other populations around the world. This could reduce health inequalities in some of our poorest countries and regions.ML methods and MLNN have also been applied to risk stratification in patients with newly diagnosed atrial fibrillation (AF). In this condition, the heart beats irregularly due to the random contraction of the atria.
This is because the signal to start the heartbeat does not originate in the sino-atrial node (SAN or pacemaker), as it should, but somewhere else in the atrium, causing the small area to contract, or fibrillate, erratically. This does not allow the heart muscle to relax optimally between contractions and thus reduces efficiency. Consequently, the heart cannot pump as much blood as it should and could go into heart failure (Figures 4). In the UK alone, it is estimated that there are 1.4 million people with AF [20] and is often associated with stroke, heart failure and premature mortality. Therefore, stratifying risk in AF is crucial as one of the mainstays of treatment is antithrombotic therapy which help prevent blood clots which can cause strokes. However, these drugs carry risks and can cause bleeding, hence accurate risk prediction tools are imperative to enable informed decision making between the clinician and the patient. The traditional risk scoring methods are CHA2DS2-VASc (which measures future possibility of an ischaemic stroke or “brain attack”) and HAS-BLED (which predicts a patient’s bleeding risk.) A recent retrospective cohort study of patients with AF compared a new ML method to traditional ones for predicting outcomes in AF [21].
Figure 4
They looked at 9,670 patients (so a sizeable number) and followed them up for up-to one year (a reasonable period of time). They showed that GBM, the best performing ML algorithm showed a modest performance for stroke compared to CHA2DS2-VASc (Area under Curve {AUC} = 0.685 vs 0.652) but a significant improvement in predicting major bleeding to HAS-BLED (AUC = 0.709 vs 0.522) and death in comparison to CHA2DS2-VASc (AUC = 0.765 vs. 0.606).With more sophisticated ML in development, AI could potentially transform these risk scoring methods in the next decade (Figure 5).
Figure 5
AI and Arrhythmia Detection The ECG is the core of cardiology, it is quite literally its “beating heart.” ML models fare better than humans in identifying various abnormalities on ECGs. This is particularly true for AF and long QT syndrome [22,23]. Viskin, et al. [22] showed that even world-renowned experts and arrhythmia specialists could not accurately calculate a long QT interval and identify it on an ECG. Their study used 902 physicians, so a decent sample [22]. Meanwhile, Attia, et al. [23] have demonstrated that neural networks applied to a traditional ECG can improve ECG interpretation, although their sample size was small, only 42 [23]. AI has also been shown to be useful in other arrhythmias [24]. This could help in the future and empower non cardiologists like primary care physicians to diagnose and manage more patients in the community. Other models have detected AF with good sensitivity (79%) and specificity (79.5%) [25]. A further study by the Mayo clinic [26] used an AI enhanced ECG network to read the ECGs of around one million people with no AF (the control group) followed by ECGs of people with episodic AF. The AI network was never shown the ECGs of people with persistent AF, only episodic AF, and the control group. The AI network detected AF with a high degree of accuracy 79.4%, AUC 0.87. At evaluation 31 days after the first ECG, the network’s accuracy increased further to 83.3% with AUC 0.90.
This study with a huge sample size showed the benefits of AI in AF detection and is a potential game changer. There were some caveats however as the experimental and control groups were not identical and there were differences in their physical characteristics such as age, height, and weight. These might have influenced the results which may not be replicable. Additionally, the network was diagnosing only episodic and not persistent AF so translating these results to all types of AF may not be valid, AF however can be difficult to detect with traditional methods of pulse palpation, ECG, or 24-hour Holter monitoring (where the ECG is attached to the patient for a day.) However mobile devices, known as “wearables,” such as smart bands, smartwatches and smartphones are playing an increasingly significant role in its detection. The KardiaBand from AliveCor uses a smartphone application based on ML to recognise AF [26]. A randomised control trial (RCT) of 1001 ambulatory patients aged 65 years and above (and were high risk of stroke) were considerably more likely to have AF detected on this device rather than traditional routine monitoring, over a twelve-month period [26]. Additionally, the Apple Heart Study [27] showed that using smartphones was effective in identifying patients who had asymptomatic AF. This large study had 420,000 participants and 0.5% of patients had an irregular pulse, which was confirmed on ECG testing in 34%.
These patients were then treated appropriately, and strokes and other complications potentially avoided. The University of California [28] conducted a similar study looking at passive detection of AF from a smartwatch. They had 9750 participants and validated their findings against a 12 lead ECG. Over 139 million heart rate measurements were recorded, and the neural network had a sensitivity of 98% and specificity of 90.2%. This study was also revolutionary and empha
Figure 6
They showed that GBM, the best performing ML algorithm showed a modest performance for stroke compared to CHA2DS2-VASc (Area under Curve {AUC} = 0.685 vs 0.652) but a significant improvement in predicting major bleeding to HAS-BLED (AUC = 0.709 vs 0.522) and death in comparison to CHA2DS2-VASc (AUC = 0.765 vs. 0.606). With more sophisticated ML in development, AI could potentially transform these risk scoring methods in the next decade. However, one could argue that the sensitivity and specificity of these experiments prove that these devices are still not accurate enough as we are dealing with people’s lives. Thus, inaccuracies should be minimised. Nevertheless, with further refinements over the next decade, this AI network could be transformative. As the popularity of these devices increase, the costs could also come down. Currently, the most popular smartwatch that offers ECG functionality is the Apple watch series 8 [29] which sells on the Apples website for £419. This might limit the market of such watches to a very small user group as many people may not be able to justify spending this much on a watch. There are currently over ten commercial wearable devices which have received FDA approval in the USA, which will revolutionise treatment if this condition in the future (Figure 7). Others have used an AI model to detect reduced ejection fraction (a parameter of heart failure) with excellent sensitivity (86.3%) and specificity (85.7%) [30]. This comprehensive study looked at 52,870 patients and the model revealed an excellent AUC of 0.93, prompting the authors to conclude that adding AI to the ECG is a “...ubiquitous, low-cost test” and can be a powerful screening tool. I would anticipate that these algorithms will be more readily used in the next decade.
Figure 7
AI and Heart Failure (HF)
HF is a common disease, affecting 1-2% of the adult population in developed countries [31]. A significant proportion (more than 10%) are older than 70. ML has been used for the early detection, classification, and anticipation of adverse events like 30-day in HF [32,33]. showed that ML algorithms have a superior predictive ability of future HF events than current ones and incorporating them in the electronic medical record is advantageous. These AI based Clinical Decision Support Systems have a high accuracy for HF and this may be important in poorer countries [33]. Moreover, AI-based stethoscopes which can also record ECGs have reported good diagnostic accuracy for detecting HF with a reduced ejection fraction [34]. This prospective observational study was one of the first of its kind and recruited 1050 patients. The authors felt that there was potential for this inexpensive and non-invasive point of care screening and certainly could be the future of HF diagnosis in the next decade. Other studies have highlighted the value of AI algorithms to better identify those patients at high risk of readmission [35]. Although the numbers were small (1653 patients), there was a 17.8% reduction in readmission after one month in those patients who identified on the ML model. Finally, AI models have also been used to identify which patients would be suitable for certain types of treatment in HF, such as cardiac resynchronisation therapy. This is a major treatment in HF, but these patients need specific ECG characteristics to be eligible. ML models have been developed with scoring systems which reliably predict which patients will respond to this treatment and which will not [36]. With more data and more sophisticated modelling this application of AI in HF could be extremely significant in the next decade.
AI and Diagnostic Imaging
AI is already playing a vital role in diagnostic imaging in cardiology. It helps in patient selection for the right treatments and helps in diagnosis. DL has assisted in analysis of echocardiograms, cardiac
Figure 8
AI and Interventional Cardiology
AI has a growing place in interventional cardiology also. A recent trial: CEREBRIA-1 compared ML to expert human opinion to determine what were the best revascularisation strategies to restore blood flow to diseased coronary arteries [40]. The ML algorithms were no worse than expert consensus opinion in deciding what was the best treatment strategy for the patient. Other studies have echoed these findings. Once again, over the next ten years, these ML algorithms should become more common place, and will be a clinical aid to decision making in tricky situations. This could free up clinician time and provide a robust reason for any treatment decisions undertaken.
AI and Heart Murmurs
Murmurs are due to turbulent blood flow across heart valves. They can be innocent or pathological. An AI based algorithm has recently been developed [41] using 3180 heart sound recordings to detect and classify these murmurs. The algorithm showed a sensitivity of 93% and a specificity of 81% with an accuracy of 88%. This trial was the first to objectively evaluate an AI based algorithm to detect murmurs. It had a good sample size and surprisingly good accuracy. In the next decade it could be a useful tool in screening for heart disease.
Virtual Wards
Currently in their infancy in the UK, AI could be used to identify cardiac emergencies remotely such as worsening HF or heart attacks. Wolgast et al., (2016) used DL to measure ECG signals and transmit via Bluetooth to a smartphone. This raised the alarm if a heart attack was identified. Others have used wearable sensors and an ML model to detect HF exacerbations in the community [42]: the vary basis of remote monitoring in these “virtual wards.” They will become an increasing feature of the healthcare landscape in many countries in the future as they are less resource intensive, and patients often prefer them to hospital admission [42] (Figure 9).
Figure 9
From these myriad examples, encompassing preventative cardiology, to smartwatches, from atrial fibrillation to heart failure diagnosis, management, and risk stratification, from imaging to interventional cardiology and finally to murmur detection, it can be seen that AI in cardiology is a huge and burgeoning field. There are however some ethical considerations which should be addressed. Since the dawn of civilisation, humankind has been fascinated and fearful of the new and unfamiliar in equal measure. Any proponent of AI needs to acknowledge this. The studies in AI, though numerous and increasing, can have methodological flaws, some of which have been highlighted like small sample sizes which could lead to bias. There can also be selection bias in the population studied which can include sampling and observer bias. The gold standard for research is meta-analysis of randomised controlled trials, but there are few of them in this field. Many studies are retrospective in design which can be problematic in terms of validity and reproducibility. The population itself that is being studied is dynamic, not static, and there maybe problems with under-representation of minority ethnic groups and discrimination against this population. This could exacerbate existing health inequalities. The data chosen to train the algorithms in AI could be investigator specific and reflect their own prejudices and biases and this needs to be accounted for when interpreting the research findings. There are broader ethical dilemmas too, concerning the huge amount of data needed to “train” these machines and validate these algorithms.
This data is often stored in large repositories and only the companies with the biggest resources can afford to access it. This can lead to reduced opportunities for others and a potential democratic deficit, where the future is ruled by “big tech” and “big pharma.” This may worsen health inequalities between richer and poorer nations. Additionally, many people would feel uncomfortable with their personal data being used by large, for-profit organisations where it could be mis-used or even stolen. These are all genuine concerns which need to be addressed. One should also understand how outliers are treated by the ML algorithm, we are all, by our natures, inherently different and being “abnormal” is just a part of being “normal.” Therefore, the clinician who interprets the information needs to be cognisant of that fact. There will be in many cases a skills and knowledge gap between the humans and AI, which need to be addressed in medical schools where computational sciences and algorithmics could be embedded into the curriculum. The public also should be educated into the advantages and drawbacks of AI in cardiology in particular and medicine in general.
AI in cardiology can clearly be a game-changer in the next ten years and its applications are dazzling. As AI is constantly learning and improving, this positive feedback loop will make it more accurate and useful in the future. It will help in digital biomarkers for risk prediction and early, targeted interventions in certain diseases. This is precision medicine at its best. However unregulated can have disastrous consequences and the fear remains amongst many of humankind being overtaken by machines. I would therefore like to propose a common overarching framework for the regulation and quality control of AI in general. This would be implemented by an international body whose sole remit would be to ensure that AI works for the benefit of humanity in a controlled way. AI has a lot of potential for the next decade and the innovations are only just beginning. An unfettered, unregulated AI environment serves nobody and therefore this body would be the “global AI policeman.” It will ensure that AI is allowed to innovate, but on our terms. There will be regulation but not enough to stifle innovation. This body would then hopefully reassure the public and professionals alike and address their fears and concerns but also fully grasp the opportunities.
We live in an ever-changing world where time is precious, and resources limited. Health and healthcare have come increasingly under strain, especially since the Covid-19 pandemic. Knowledgeable patients expect the best treatment and clinicians need to access the latest in evidence-based medicine while tailoring the treatment to the individual patient.AI can and will play an integral role in this in the field of cardiology in the next ten years. From detecting abnormal heart rhythms through traditional ECGs augmented by machine learning, to using innovative technologies such as wearable smartwatches, or accurate predictive risk tools to prevent readmission from heart failure AI is slowly becoming integrated into the cardiologist’s daily routine. By rapidly processing echocardiograms or assessing coronary artery disease risk scores, determining the best strategies for coronary artery revascularisation to detecting heart murmurs, AI has in many instances proven itself to be no worse than expert cardiologists. These applications of AI will only improve in the next decade due to the positive feedback algorithms. There may even be avenues that we have barely explored where AI could play a key role, such as virtual wards. Ethical dilemmas and fears remain however that AI will replace humans and there are some significant concerns over the data published to date. There are also valid anxieties over the widening of health inequalities between richer and poorer nations. Therefore, a common overarching framework is proposed for regulation and quality control of AI in general.
This framework will enable AI to thrive but in a regulated environment. It will address the “democratic deficit” in the current global AI space where “big tech” and “big data” is king. On a final note, AI can certainly improve work efficiency in many areas, detect patterns and predict future events. However, one should bear in mind that AI is just one of a suite of tools used to improve clinical judgement and nothing can replace the role of the physician, who integrates a multitude of variables, some scientific, others less so, into the decision-making process to deliver evidence-based and person-centred care. The old adage “the doctor will see you now” will still ring true today, tomorrow and in ten years’ time.
I would like to take this moment to extend my heartfelt gratitude to the individuals who have played a pivotal role in the assisting of me in the completion of this essay. First and foremost, I would like to thank Dr Sarah Clark for first introducing me to the field of cardiology and catalysing the start of my interest in this field. I would also like to thank my supervisor, Dr Mitchell for her amazing support throughout this project, for imparting her knowledge and expertise, offering valuable insights into how to research well, and pushing me to explore new horizons in my research. Lastly, I would like to mention my friends and family for their unwavering encouragement and understanding, constantly pushing me to give this project my all. Without the collective efforts of these remarkable individuals, this essay would not have been possible.
Abstract
PDF
