Heterogeneity, risk factors and personalised profiles of patients with HF
HF is a complex disease entity, being typified by a common ‘downstream’ clinical syndrome but with a broad range of ‘upstream’ structural, functional and biochemical phenotypes as well as diverse disease courses and prognoses . Assessment and treatment of patients presenting with stable or decompensated HF must be performed in a manner tailored to the patient’s individual phenotype [37,38,39], which is often not (yet) the case.
Patients with HF typically have multiple comorbidities  and crossover symptoms (e.g. breathlessness in angina, chronic obstructive pulmonary disease, pulmonary vascular disease, anaemia, uncontrolled arrhythmia) . In addition, certain clinical signs (such as tachycardia) may be symptom-generating in themselves (breathlessness), may reflect an underlying pathophysiological process (such as infection or fluid overload), or may even be the underlying cause of HF, which itself may also result in similar symptomatology. Teasing out these subtleties by history and physical examination can be extremely challenging, even by specialists . Therefore, better stratification of patients with chronic diseases is essential in order to define personalised profiles that are required to achieve the paradigm shift to PPPM.
Challenges in HF care
Choosing the wrong therapy may have potentially deleterious effects. Severe side effects and even death can occur, e.g. in patients wrongly deemed to be systemically congested due to inappropriate diuretic administration , or intensifying rate controlling medications in tachycardia related to congestion without addressing the underlying volume overload . Equally, delaying appropriate therapy may reduce the efficacy of the intended therapy, e.g. early intervention with diuretic therapy to reduced volume overload is more likely to be effective whereas once systemic volume has been retained the oral route is less effective necessitating intravenous therapy [45, 46]. In complex fields with a broad library of research, case histories and potential therapies, it can be difficult for individuals to remain up to date with best evidence—especially in complex diseases where software such as the IBM Watson has shown some promise in guiding appropriate therapy .
Added to this is the concept of a patient acting as the gatekeeper of their own baseline health status and an active invigilator in detecting when their baseline has changed. This may be the ultimate trigger in deciding when to change therapy. A patient’s subjective awareness of what different disease processes feel, e.g. does chest pain feel like the angina previously experienced?, may be important in accurate diagnosis and treatment.
Computer-guided algorithms have the advantage of processing many data points quickly. However, they rely on accurate contemporaneous patient input and engagement. A computer may also be less accurate in certain elements of patient care, which may be difficult to quantify in an algorithmic fashion, e.g. patient frailty, difficulty to tolerate medications, wishes regarding ceiling of care and end of life care.
Healthcare professionals and computers may be complementary to determine response to therapy. For example, clinical response to therapy may be earlier than objective measures of response such as biochemical  or weight  parameters in some but later in other patients. On the other hand, clinically obvious parameters may be obscured and alternative diagnostic tools may be more sensitive in detecting congestions . Thus, remote diagnostics with appropriate analyses techniques may be superior to clinical findings; however, this might not be uniform in all HF patients.
Similarly, distinguishing side effects of medications from the underlying disease process and comorbidities may also be challenging for both machines and humans. Examples may include, does tiredness reflect low cardiac output or effect of treatment such as beta blockers? Does deterioration in renal function represent progression of the underlying HF, an expected response to angiotensin-blocking agents or overdosing diuretic therapy? It is possible that AI may be able to improve distinction, prediction, and prevention significantly, but this has not yet been properly tested.
These questions highlight the complexity of applying a structured diagnostic algorithm to the clinical course of a disease, as well as the inherent complexities in its personalised management. Given the broad phenotypic variation on both a personal and population level, it is likely that a combined approach of well-informed ‘baseline’ clinical algorithms along with personalisation to individual patients are needed for maximum efficacy. In particular, an AI-delivered therapeutic strategy needs to learn how the patient responds to treatment, how complications of both treatment and disease present, and be able to alter future responses based on accumulated knowledge—similar to how a physician becomes familiar with an individual patient over time.
Participatory medicine: new roles for both patients and healthcare professionals
In recent years, the doctor-patient relationship has evolved beyond the paternalistic ‘doctor knows best’ model, toward one where the patients wish to be more actively engaged and expect a degree of input in their care [51,52,53]. This, in turn, represents an opportunity to not only involve the patient in decisions around their care but also to actively involve them in self-administering care.
Increased patient engagement in their own care has many potential benefits:
Improvement of patient adherence 
Improvement of patient satisfaction 
Reduction of direct patient costs (e.g. travel, parking) 
Reduction of indirect costs (e.g. productivity days missed for patients and carers) 
Improvement of patient outcomes and reduction of healthcare costs .
Patient-centred care represents a paradigm shift in how patients, doctors, nurses and other health professions think about the processes of treatment and healing. Defined by the US Institute of Medicine as the act of ‘providing care that is respectful of, and responsive to, individual patient preferences, needs and values, and ensuring that patient values guide all clinical decisions’,  patient-centred care prizes transparency, compassion and empowerment. The rise of patient-centred care makes way for a healthcare system designed to optimise the agency and comfort of the most important and vulnerable people in the equation: patients, their families and their communities .
Patient-centred care is a far-ranging new view of healthcare that resists simple summation. But there are a few consistent core ideas that guide this new style of care . Researchers from Harvard Medical School, on behalf of the Picker Institute and The Commonwealth Fund, identify Picker's eight principles of patient-centred care (Fig. 2) . However, as it turned out, the development from theory to practice is a lengthy and unruly process.
Successful patient engagement
Current concepts imply that key to the success of patient engagement in self-care is educating patients about the clinical manifestations of their disease process, effects of medications, lifestyle modifications required and clinical particulars regarding the rationale for a prescribed treatment regimen . However, for patients or their carers to truly take charge of day-to-day disease management, they must also be sufficiently supported and equipped with the essential tools, e.g. to deal with a change in clinical status—such as a structured approach to identification and management of deterioration, which may be ASYMPTOMATIC (detected by clinical measurement tools) or SYMPTOMATIC (which may be recognised by the patient or carer).
Detection of asymptomatic deterioration would effectively take the place following a physical examination performed in the clinic and may involve the use of simple measures such as weight and blood pressure, but may also expand in suitable patients to involve more advanced metrics such as implantable sensor data (e.g. CardioMEMS™ , OptiVol™ ). Effective monitoring may be improved by novel eHealth technology and sensors, which is a rapidly growing field.
Detection of symptomatic deterioration would take the place of ‘doctor-initiated’ history taking. So far, key to this process is educating and coaching the patient and carer in awareness of symptoms indicating deterioration of their condition, e.g. worsening breathlessness in HF . It would also require educating the patient about symptoms, which may indicate an alternative pathology warranting alternative treatments, e.g. cough with sputum production in the setting of worsening breathlessness may indicate an infective process rather than deteriorating HF. However, telemonitoring can be more efficient than education alone in supporting patients and carers in detecting and interpreting symptoms . In particular, the majority of patients with chronic diseases who are usually elderly may not be able to detect and interpret symptoms sufficiently without additional support.
Additional measures are required for a successful personalised self-guided care strategy. There must be (i) a relatively simple treatment algorithm in place to respond to changes in clinical status which the patient can follow, and (ii) backup from a clinical specialist nurse or physician if needed. Basic therapies, such as adjustment of diuretic doses  or adjustment of anti-hypertensive medications , may assist in preventing a hospital admission with acute deterioration. Modern eHealth technology using AI may be more efficient and may allow a much larger proportion of the HF population to apply personalised self-guided care strategies.
New technologies enabling data generation by the patient will increase the data volume, which will improve timely information and education for patients and doctors in personalising therapy and predicting response in a personalised manner. This can also be enhanced by different approaches including the use of gameful design of support systems  to increase patients’ engagement in therapeutic processes and adherence and help identify and address their requirements and preferences in the context of self-management of chronic conditions, like HF. In turn, this will offer new ways of identifying patients early with very high likelihood for acute deterioration and death. One example might be that early indicators like changes in daily behaviours may prevent fatalities of individual patients by an early alarm system . Remote patient management may detect early signs and symptoms of cardiac decompensation, thus enabling prompt initiation of specific treatment before full manifestation of HF decompensation. Ideally, this is initiated by the patients themselves instead of being professional driven, which is currently the case in telemonitoring systems. However, this means that patient participation, and their perspectives, need to be included in the development of new technical devices and healthcare approaches. This new approach in patients with a life-threatening chronic disease will be a milestone for approaches in integrated care, PPPM and managing patients with multimorbid diseases.
Therefore, the uniqueness of such novel strategy aiming to be a comprehensive patient-centred clinical support system for taking care of patients with HF, can also become a blueprint for managing patients with different other chronic diseases using eHealth solutions via intelligent algorithms. This will pave the way for overdue advancements in daily medical practice, namely, improved efficiency in patient care coupled with high potential for cost reduction.
Support for changes in the care process using eHealth
There are numerous definitions for eHealth in the scientific literature. In short, eHealth is often defined as the use of information and communication technology in healthcare. In the context of this strategic paper, a more comprehensive definition is used; ‘eHealth is an emerging field in the intersection of medical informatics, public health and business, referring to health services and information delivered or enhanced through the Internet and related technologies. In a broader sense, the term characterises not only a technical development, but also a state-of-mind, a way of thinking, an attitude, and a commitment for networked, global thinking, to improve healthcare locally, regionally, and worldwide by using information and communication technology’. 
The development of new eHealth products is evolving fast, with the immense availability of healthcare related applications. These developments are not only supported by the industry who participate in their development, but also by patients, national and international societies of physicians, and governments [70, 71]. A vast majority of patients are willing to engage in these new practices with high patient satisfaction and improved health outcomes; these are the beginnings of a fundamental paradigm shift in healthcare . A consultancy agency that explored the potential of medical technologies with regard to patient centred care needs, by order of the Ministry of Health, Welfare and Sports (VWS) of the Netherlands, concluded that the attitudes of patients and healthcare professionals for these new technologies are positive, but there is a reluctance in full adoption . In addition to a lack of well-established evidence and reluctance of healthcare professionals to redesign care processes, insufficient funding by healthcare insurances or by Government , and the difficulty of linking existing information and communications technology (ICT) systems with the new technologies  are important barriers. The minister of VWS of the Netherlands has set up actions to encourage insurance companies, healthcare organisations, stakeholders and developer to cooperate and create a successful implementation and upscaling of the use of eHealth  indicating the support from the government for changes in the usage of eHealth in the care process. The expectation worldwide is that eHealth will be the way forward in achieving solutions for sustainable healthcare .
Patient-interface: how to facilitate eHealth for patient centred care
eHealth technologies are playing an increasingly important role in providing internet-based disease management, including self-management support, facilitating information exchange among professionals and with patients, and monitoring the performance of disease management programs. In the approach of Eysenbach , eHealth is seen as more than just a sophisticated tool: it is about co-creating and evaluating an infrastructure for knowledge dissemination, communication, user engagement and the organisation of care with all involved stakeholders—patients, doctors, nurses and technology developers. This requires careful embedding of technology into the care process, with attention on added value in a dedicated context. Co-creating sustainable eHealth technologies thus requires a holistic development and evaluation approach that takes into account the triad between technology, users and context of implementation .
Patients newly diagnosed with a chronic disease are often overwhelmed with their situation. They feel insecure, sometimes anxious and often confused by the quantity and diverse provision of information. Affected patients have to learn to deal with their new living conditions, including the reaction of their relatives. Thanks to new technology it is possible to offer them remote support and online coaching with which they can complete efficient self-care. Practice has shown the importance of relevant knowledge for personal processing of a diagnosis, behavioural improvement and timely recognition of symptoms to prevent exacerbations. The mission of eCoaching is to reach optimal HRQoL. ‘Patients make better medical choices with coaching’, emphasises Jeff Belkora, founder of Patient Support Corps and Professor of Surgery and Health Policy at the University of California.
Usual care for chronic diseases has been limited to episodic visits to various health services in which interim results and events are reviewed, and plans are made for adjusting care as needed, based on the information acquired during the consultation. In order to unburden both patients and healthcare professionals, modern care should be provided at the convenience of the engaged patient, taking advantage of modern technology. Against this background, the concept of eCoaching in care of chronic diseases may help to optimise patient independence and outcomes using eHealth. The combination of the various aspects of eHealth including remote patient monitoring, eCoaching, and decision support enables patients to be treated outside of conventional clinical settings, potentially improving HRQoL, engagement and compliance, and outcomes in chronic diseases.
Gamification to improve patient engagement
Incentives and rewards are one of the major means to attract users’ attention and ensure long-term engagement. Through proper and carefully thought through design, reward programs can be of high importance, as they provide (necessary) motivation for behaviour change and better long-term engagement. There are two elements of a reward that need careful planning; timing and nature of the reward. Ultimately, the goal is to maximise motivation and effort prior to receiving the reward, as well as happiness after receiving the reward . Gamification employs a seamless combination of a number of mechanisms, tailored to specific situations on a case-by-case basis in which users find themselves, taking into account their personal values and expectations.
Success of gamification is driven by various factors, which include emergence of mobile devices, rise of big data and wearable computing . Gamification in combination with AI found its use in various contexts including intelligent tutoring systems, health applications with decision support systems and a number of industrial applications that benefit from gamification mechanics and dynamics elements . Thus, the inclusion of gameplay elements and systems into non-game contexts, such as healthcare, education, training, and work can enhance the user experience, and encourage behaviour modification and long-term participation. Various stakeholders including healthcare providers are investigating how to ‘gamify’ their processes and systems to increase engagement . Some examples include the use of gamification to run a global awareness campaign to prevent cancer , and the DIDGET Blood Glucose Meter, that plugs into a Nintendo DS or DS™ Lite gaming system to reward kids for consistent testing .
Gamification can also be used to provide engagement to use AI based platforms. The growth of digital media has seen an explosive growth in the exchange of data, with its ability to collect, store and sort data, that is often an impossible task for humans. Crucial to the success of AI, are the algorithms that can learn and advance their intelligence. One way to inform these algorithms is to use gamification. Games are mainly complex systems, but tracking and storing data from them is easy and straightforward. Still, such an approach not only enables AI to learn and to improve, but also brings benefits to the user, who is motivated to engage with the application through the various gamification features.
Gamification can be used to track and measure progress. Using the various elements mentioned above, stakeholders can store and track the performance, progress and aptitude, and can make further decisions based on this data. In healthcare, this may result in using gamification and AI for diagnostic purposes. There are various benefits of using gamification to provide engagement in AI-based platforms, such as :
Flexibility in addressing problems and challenges
Knowledge being more transparent
More data for AI to learn and improve from
Low-cost and low maintenance required for AI.
Eventually, this will result in better patient adherence to using of the eHealth application and simultaneously to gather additional diagnostic information to better monitor and guide patients. Uniquely, via data analytics, AI can learn how best to enhance patients’ self-care capability. Thus by combining AI and gamification, an eHealth application may be smart, intelligent and entertaining at the same time.
Understanding the role of artificial intelligence in the paradigm change
When a machine mimics ‘cognitive’ functions that humans associate with the human mind, the term ‘artificial intelligence’ is applied. Examples are data analysis and prediction, image analysis, human speech recognition, self-driving cars and many others. Tools in AI are based on statistics and specific methods like logistic regression, neuronal networks (deep learning), random forest, ensemble tree, fuzzy partition, support vector machines, self-organising maps and more. Many AI algorithms are capable of improving from data, to enhance themselves, by learning new strategies that have worked well in the past. AI algorithms consider possible hypotheses and match it against the data [84,85,86,87,88,89].
Supervised learning means learning a function that maps an input to an output based on example input-output pairs. Therefore, a supervised learning algorithm analyses the training data and produces an inferred function, capable for mapping new examples . Unsupervised learning stands for inferring a function that describes the structure of data that has not been classified or categorised. In this case, there is no straightforward way to evaluate the accuracy of the result which has been produced by the algorithm.
McKinsey predicts healthcare as one of the top 5 industries with more than 50 use cases using AI . AI will change the healthcare industry, support doctors in improved provision of healthcare and empower patients in controlling and maintaining their health. AI is perfectly suited to support PPPM. There exist already many examples which provide a clearer view as to how AI will affect healthcare. The direction of impact is inclusive of making hidden information into data usable, gain of efficacy and efficiency, risk identification, prevention, triage of patients, telehealth, virtual nursing, robot surgery, therapy guidance, self-treatment, higher precision, prediction, personalisation and relief from routine tasks.
Complex chronic diseases like HF require personalised diagnostics and therapy, which may be supported by AI to improve prediction, therapy guidance and prevention of deterioration. Nowadays, diagnostic companies still focus on approaches using a limited number of markers, leaving interpretation to the healthcare professional. However, the dynamics of complex diseases are often only determined by the use of a broad range of markers using different diagnostic modalities. In contrast to the human brain, AI is able to deal with the exponential increase of possible combinations related to multi-marker approaches. Therefore, the paradigm change using AI requires the availability of a broad range of data, including patient history and characteristics, signs and symptoms of not only heart failure but also accompanying comorbidities, results of diagnostic tests, input from multiple markers (e.g. biomarkers) and sensors, medication and other aspects of treatment. In essence, the algorithms require the same information as healthcare providers use for decision-making.
AI introduction in healthcare is starting with hybrid models. Physicians are supported in diagnosis, treatment planning, identifying risk factors, but retain ultimate responsibility for the patient’s care. This will result in faster adoption by healthcare providers by mitigating perceived risk and delivering measurable improvements in patient outcomes and efficiency at scale. Long term, there will be a shift from decisions exclusively made by physicians and other healthcare providers to a hybrid decision model, where unambiguous decisions are made by AI and directly advised to the patients but leaving the responsibility for more complex or ambiguous decisions to the healthcare professionals.
Expected “added value” of AI services in healthcare and potential harm
The face of modern healthcare is changing dramatically and AI has a significant impact on this in two main ways; assisting clinicians in the delivery of care and the extraction of meaning from ‘Big Data’ . Task delegation and sharing is becoming significantly more important with the global shortage of healthcare professionals and the increasing number of patients. AI may reduce healthcare expenditure and ultimately improve quality of care. A prospective analysis concluded that the use of AI applications could save American healthcare by approximately $150 billion annually by 2026 .
Despite several examples of AI already successfully introduced in clinical practice, adoption is not yet broad, with some exceptions, and implications to clinical practice is limited. This will, however, rapidly change as the following examples illustrate. AI has been ‘visible’ through clinical decision support systems (CDSS), in care delivery across specialties. Castaneda et al. diagrammatically illustrated a typical clinical consultation and the effect a CDSS may have . The amalgamation of information to support clinical decisions was publicly evident in IBM’s ‘Watson’ , which had the capability of replicating up to 90% of the decisions made by healthcare professionals. In terms of delivering personalised medicine, AI has a distinct advantage in its ability to pool massive data sets and extrapolate their relevance to an individual—as evidenced in a novel program designed to calculate a woman’s ‘Hereditary Breast and Ovarian Cancer’ risk result and offer recommendations and questions for her to raise with her healthcare provider .
Furthermore, the algorithm for the detection of cardiac rhythm changes (ST elevation) is commonly available with the majority of 12 lead ECG machines offering analysis and enabling the provisional diagnosis of myocardial infarction and cardiac arrhythmia . Other recent work has demonstrated the ability to predict risk of cardiovascular disease based purely on a still image of a patient’s retina . An AI-based test of multiple blood derived biomarkers detected coronary artery disease with an accuracy which so far was only achievable by advanced imaging technologies such as modern contrast CT angiography . A mobile Atrial Fibrillation App incorporated clinical decision-support tools alongside educational material, patient involvement strategies and follow-up . Patients reported the app was user friendly and significantly improved patients’ knowledge, drug adherence and anticoagulant satisfaction, increased quality of life, and reduced anxiety and depression. Indeed, AI has the capability for real-time continued monitoring (whether in the home setting or within healthcare settings) with built-in care pathways for escalation and interventions when needed, ultimately empowering patients in self-management to improve clinical outcomes [101, 102].
Further advancements in AI have led to the development of embodied conversational agents or avatars, which can further improve user engagement and effectiveness of an intervention, for example directing patients to the most appropriate care service. Since availability is 24/7, avatars can monitor patients and provide quick and timely answers. Thus, integrated AI technology provides health state analysis, decision support and treatment recommendations. For example, an animated conversational agent playing the role of a health counsellor, through a series of daily conversations over two months, had a positive impact on fruit and vegetable consumption .
AI may also improve the diagnostic process of many diseases. A recent Stanford University study tested an AI algorithm to detect skin cancer against dermatologists, and found it to perform at the level of the humans. Further, Baidu Research announced that the results of its deep learning algorithm can outperform humans with identifying breast cancer metastasis . AI is providing several solutions for triage of patients that provide quick, scalable access for basic questions and medical issues. Additionally, unnecessary trips to the GP can be avoided, reducing the rising demand on primary healthcare providers and providing basic guidance that otherwise would not be available for populations in rural areas. In radiology, AI solutions can automate image analysis and diagnosis. These solutions drive efficiency and reduce human error. Improved tumour detection on magnetic resonance imaging (MRI) and computed tomography illustrates the progress. The US Food and Drug Administration (FDA) has given clearance for an AI platform which analyses and interprets Cardiac MRI images. For humans, image analysis is very time consuming. A Massachusetts Institute of Technology led research team developed a machine-learning algorithm that can analyse 3D scans up to 1000 times faster than before, making it possible to study changes almost in real time .
‘Big data’ involves the integration and interpretation of large volumes of healthcare information, such as biomedical and clinical data to generate robust scenarios applicable for everyday life. In the UK, information from electronic health records has been used to inform prediction of drug effects and interaction, identification of type II classifications and discovery of comorbidity clusters in autism spectrum . Furthermore, there is increasing evidence on the benefit of computer-based decision support within pharmacology with the ability to reduce physician errors and quicken decision-making, thereby potentially saving lives. AI may also help to identify new potential therapies from vast databases which could be redesigned to target critical threats. This could improve the efficiency and success rates of drug development, accelerating the process to bring new drugs into market.
AI-assisted robotic surgery gets an increasing amount of attention. At present, Da Vinci is the most advanced surgical robot that allows doctors to perform complex procedures with greater control than conventional approaches. Heart surgeons are assisted by heart lander, a miniature robot that enters a small incision on the chest to perform mapping and therapy over the surface of the heart. Machine learning has been instrumental in orthopaedics, in terms of understanding biomechanics, orthopaedic implant design , prediction of progression of osteoarthritis  and robotic surgery .
A number of future avenues for the use of AI in the ‘big data’ domain includes the use of unsupervised learning techniques to more precisely phenotype complex disease  and facilitate earlier prediction of an epidemic. It is anticipated that the World Health Organisation (WHO) will be able to monitor big data in real time from a wide range of international sources, thus mitigating the progression of an epidemic . AI can already address healthcare challenges within resource-poor settings, for example Onu described signal processing and machine learning in Nigeria, whereby mobile phone data has been used to predict birth asphyxia .
The large potential of AI for improving healthcare is indisputable. The question is how to integrate it safely and successfully into our everyday life and healthcare systems. According to the authoritative online publisher The Medical Futurist in the e-book ‘A guide to Artificial Intelligence in Healthcare’ a range of hurdles must be tackled before AI redesigns healthcare.
Potential harm and hurdles of AI introduction in healthcare
It is the promise of AI has to be balanced by possible risks and harms. One of the greatest concerns of introducing AI into healthcare is the potential for error and fraud. Despite an increasing emphasis on healthcare quality and safety, medical errors are not uncommon and pose a serious public health problem. Indeed, hospital medical errors are the third leading cause of death in the USA . AI is being used or trialled for a range of healthcare and research purposes and though it has the potential to address important health challenges, its use also raises ethical issues such as the potential to make erroneous decisions, compromise or fail to safeguard patient health data, and be used for malicious purposes . For example, in using AI, medicine may be susceptible to adversarial attacks both in terms of monetary incentives and technical vulnerability, subsequently caution in deploying AI in clinical settings has been urged .
Embracing advances enabled by AI is likely to incur cost savings and efficiency. Whilst it may be argued that AI might be expected to make healthcare safer and more efficient, we need to ensure that our datasets are robust, have continuation of reliable data supply, and have sufficient breadth and depth for accurate diagnosis . However, healthcare is still under-digitised in many countries. Many primary care physicians still keep their medical records on paper (e.g. Germany 34%, Switzerland 41%) . In many countries, the digitisation of the health system corresponds to the level that financial services and industry had more than 20 years ago. While in other industries digital tools are widely used and decision support tools are seen as essential aid, some doctors do not yet see the added value, with potentially tremendous negative effects on productivity and accuracy.
Along with a wealth of information comes responsibility inclusive of issues of privacy, ethics, data security and accountability. Although AI can virtually remove man-made error from processes, it can still exist in the programming: being largely algorithm based, the technology is not exempt from bias or prejudice . Hence, AI is only as intelligent and discerning as those responsible for its initial programming, noting that later improvement by self-learning algorithms may be possible. Currently, there is a dearth of regulations and lack of standards to assess the safety and efficacy of AI systems. However, the FDA has made some inroads in an attempt to provide guidance for assessing AI systems . Further to this, by succeeding in creating ethical standards, developing measures of success and effectiveness and by making it available to the mainstream, we can reduce many of the privacy concerns and misapprehensions surrounding AI .
Perhaps one of the biggest trepidations about AI is that it will become so sophisticated it will take over our lives. Is there the potential for AI to take control away from humans, de-humanise actions, reducing compassion and empathy? On the other hand, the complex and unique circumstances surrounding HF often require clinicians taking judgement calls into account when making decisions; as such the question then becomes, Can AI be programmed to successfully manage complex long-term conditions? The most probable answer to this question would be: yes, in part.
Clinicians need to be prepared for a future where their intellect and clinical discernment can be replaced—at least in part—with systems that are much more robust. This also means that the tasks of the healthcare professionals will change. We need to acknowledge the benefits and identify how best to cope with any perceived or real drawbacks of AI in the management of chronic conditions such as HF. For successful integration in the care process to enable PPPM, close ‘collaboration’ of technology using AI, complex algorithms, avatars and clinicians will, therefore, be key.
Towards a cost-effective and sustainable economic model for integrated HF care
As argued above, in an ideal situation of healthcare maximum effort should be invested into (1) facilitation of therapeutic decision making for and by the patients along with their local carers in a home setting, while taking social factors into account; (2) early detection of critical cases and/or new disease(s) which require interventions by patients or specialist time and hospital capacity, to ensure optimal use of these scarce resources; and (3) prevention at home, promoting healthy lifestyle choices especially for those who are predisposed to chronic non-communicable disease conditions.
To achieve this, a paradigm shift in healthcare is needed, which is now possible from the advances and new developments in AI evidence-based medicine. Innovation in medical diagnosis and treatment is bringing new opportunities to change the landscape of evidence-based medical treatment. In practice, AI is not one new technology, but a variety of technologies, algorithms and software tools. While the clinical advantages seem to be obvious, there are a number of issues that need to be addressed and overcome for future sustainable improvement in patient health outcomes.
Expectations vs commercialization and sustainability
Having a new technology or device developed and tested in a controlled environment (laboratory or hospital with carefully selected patient groups) does not mean that it is ready for use in a wider ‘real world’ context. For broad exploitation in the healthcare market, several issues need to be addressed:
Personalised technology or device is not fully validated or not applicable to or not tested in the relevant context—e.g. decision making not applicable to the specific contexts; poor quality (low signal to noise ratio), poor prediction, lack of robustness (drop outs, maintenance issues).
Accessibility cannot be guaranteed—e.g. distribution problems, price too high, device not certified, no reimbursement, lack of knowledge by care professionals of its use.
Despite advances in new AI-supported systems, proper clinical validation is still essential—e.g. one aim of validation is to prove the benefit but also to identify and address flaws and biases of the system.
Implementation in clinical care is mostly lacking and not properly tested—e.g. several additional challenges need to be addressed to enable commercialisation.
Some of these aspects are specifically related to AI based systems, others are of a more general market-based nature. Several additional challenges need to be addressed to enable commercialisation:
Regulatory authorisation—procedures may evolve much more slowly than the technology developments in this field and acceptability of a new personalised approach may be limited.
Ethics—what do we allow the AI system to handle, decisions will have to be agreed and incorporated into (ethical) guidelines and contracts.
Responsibilities—decision-making procedures need to be adapted, professional liabilities should be well described, perhaps even insurance coverage expanded, hence legal input is essential.
Social challenges—resistance by medical staff to change their work practices, overcoming perceptions that AI may replace/change jobs, resistance by patients and their carers (accepting as substitute for care professionals, which type of advice is acceptable, e.g. lifestyle changes, therapy decisions), institutional resistance to change (need to change processes) and practical difficulties (specific training required, investments to be made, risks taken).
Safeguarding—the position of patients as vulnerable adults and safeguarding the involvement of carers in the AI-supported systems.
Skills—specifically, trained staff who can manage, maintain and/or work together with AI-supported systems. Also, the patients and their carers will need education on how to optimally benefit
Financial—the financial structures, accounting conventions, financing availability and health reimbursement models involved vary by region or country.
Legal—governance and data management to maintain patient confidentiality and manage data exchange in appropriate ways. The use of Common and Civil law in different countries across Europe may require the use of a range of legal solutions that may differ country by country.
Valuing social-economic impact
The discussion about who bears which costs is one of the most basic elements in developing a sustainable Business Model in healthcare commissioning and reimbursement. Investing in preventive measures and improving HRQoL in their home setting makes a lot of sense and can have a significant economic impact. However, quantification of socio-economic impact may be a challenge (for lack of generally accepted indicators and standardised methodologies); Which participant in the health commissioning framework will realise the very real financial benefits that will accrue over time and therefore be able to accept the cost of provision?
A more fundamental underlying issue is to focus on health outcome and long-term impact. This is especially true when seeking to improve the access to long-term healthcare treatment and quality of life for patients who require chronic care. This goes beyond the more common challenges of commercialisation in the healthcare sector and requires acceptability by all involved stakeholders. Sustainability in this context requires insight into the financial/economic context. It is important to develop a specific Business Case for these integrated and personalised systems considering the different routes to commercialisation, which depends on who is the consumer, who actually is paying for it:
Patient—can pay for a sensor or smart phone App, but paying the full price to access an integrated therapy support system may be beyond a patient’s reach unless the AI-system lowers the access fee to a reasonable level (volume pricing).
Care provider—a group of practitioners provide specific tools or aid devices for home care or a clinical centre buys diagnostic or imaging equipment (often the quality is the first indicator).
Insurance company—a device, therapy or care service is reimbursed (often cost reduction is the first indicator).
Community—a city, region or other commissioning area decides to provide certain forms of healthcare (often negotiated as package) to all patients in their region (volume pricing).
Government or intermediary bodies—after formal assessment (health technology assessment (HTA)) access and prices for (a package of) certain products, therapies or provision of services with are recommended, facilitated or established as sole choice (formal tendering process). Obviously, combinations of these are common.
Two other elements are integral to sustainability, i.e. scalability, achieving critical mass so that prices can go down, and transferability from one environment to another environment. Unfortunately, across Europe there is a huge variety of structural conditions, and the legal, regulatory, governance and reimbursement requirements, opportunities and constraints on health budgets are quite different. The regulatory frame is hindering healthcare development towards higher efficacy. Guidelines are one example that does not even consider the use of AI. Guidelines are history-oriented and do often not meet the requirements of modern medicine. A report by the US Institute of Medicine suggested that it takes on average 17 years before new knowledge generated in randomised trials is incorporated into practice and even then, acceptance varies considerably among centres . Novel AI-based technologies may help to facilitate implementation of novel therapy but may face the same lack of acceptance even if validated in clinical studies.
In developing a sustainable, scalable and transferable AI-based solution all these factors must be comprehensively considered in early development, recognising that many of the building blocks are at different stages of technical, operational or social development. Failing to do so may result in restrictions on the Business Case, significantly impact the required balance of (co-) financing and reduce the potential level of technological innovations that are implemented in the market. This will in turn restrict the application of AI across the healthcare market perspectives and its consequential long-term growth potential.
The bottleneck is, therefore, not necessarily the technology (although new developments may open up new approaches, as discussed above). It is essential to involve all stakeholders at an early stage for proper validation and implementation to enable financially sustainable models with lasting impact on our healthcare system and on the patient’s life. In fact, it is probably true that a mediocre technology pursued within a great business model may be more valuable than a great technology exploited via a mediocre business model. Unless a suitable model can be found, these technologies will yield less value to the firm than they otherwise might .
Legal, ethical and societal issues
As the acceptance of AI grows, so do the ethical and societal questions concerning its implementation into the health system. The future impact of the new technology provides various challenges. Possible ethical, legal and societal issues could arise, e.g. in regard to a possible discrimination on the part of an AI system and the question of accountability, in the case of mistakes and the patient-physician relationship.
An important ethical issue is the potential for AI to unfairly discriminate between patients, coming from the training data that contains human biases . Thus, a recent study found that some facial recognition programs incorrectly classify the gender of less than one percent of light-skinned men, but more than of one third in dark-skinned women . The most widely used cardiovascular risk score developed using data from mostly white patients may be less precise in minorities . Further, most evidence-based treatment recommendations in HF result from studies in white men . An uncontrolled AI algorithm could make therapeutic decisions based on preexisting biases, especially when used for complex conditions with a high degree of uncertainty. In HF, this may be particularly true in patients in the palliative phase of the disease course or in patients with HF and preserved LVEF, where solid evidence is lacking. The clinical consequences of such potential misinterpretation are not known. Still, clinical decision-making faces the same biased evidence, where we accept extrapolation without support by evidence . Thus, it will be important to include testing routines for detection of potential bias in programming of AI and in critically reviewing the results of it to achieve true PPPM that is superior to current care. In addition, circumstances need to be considered where AI may not provide the required results and human intervention is required, such as in the palliative phase of care.
Associated with this arises the question of accountability. When an AI system fails at a certain assigned task, who should be responsible? The programmer, the data owner, or the end user ? The question of who is responsible if AI makes a mistake is still unanswered . Self-driving cars provide an example. Those vehicles could be involved in accidents, just like human drivers today. The difference is that we have a clear understanding of fault and blame for human drivers, but this does not yet exist for AI. A car could be programmed to act in the safest way for the passenger, or it could be programmed to protect the people in the other vehicles. Whether or not the manufacturer or the owner makes that decision, the responsibility for the fate of people involved in a car crash is not yet resolved. The same principle can be applied to AI systems in healthcare. The scope and content of these restrictions, e.g. whether and how AI can be intelligible and will apply, remain uncertain and contested . Questions of accountability and liability are easier to answer when the reasons that lead to a certain decision or action are comprehensible. Therefore, transparency is another aspect that needs to be considered when discussing AI . AI algorithms may be seen as a kind of ‘black box’ . Thus, AI is difficult to understand or interpret and it may be impossible to determine how AI has reached its decision. This could lead to bad adherence, e.g. if patients do not know why the AI system suggests a particular diagnosis, treatment, recommendation or outcome prediction . Also, medical research has to be transparent, requiring ways to uncover and show the ‘inside’ of an AI algorithm. Considering the expected complexity of such algorithms, this will be an almost impossible endeavour.
Besides that, there are also important issues regarding data sharing and protection of AI system, which need to be addressed, taking the strict ‘General Data Protection Regulation’ of the European Union into account. What should be the code of conduct? What information is really necessary to ensure the best treatment for the patients, without exposing their sensitive data regarding their health? What could be the risks of automatic profiling of patients [129, 131]?
The impact of AI tools on the patient-physician relationship regarding the decision-making process may be an additional concern . Shared decision making is considered an important prerequisite of this relationship. One difficulty is seen in the increasing availability of information on the internet, some of them with questionable content and limited scientific reliance [133, 134], and the increased use of this option by the patients . The variation in quality of information or misunderstandings may generate conflicts between the patient and the physician and may require a significant amount of time to resolve [133, 134]. However, the implementation of AI systems developed by patients and physicians together may help to overcome this problem and may help patients to make well-informed decisions about complex medical issues. Another way could be quality regulation or an evaluation and certification program for AI systems and health technology in general . Thus, AI-driven decision support tools can significantly change the way treatment decisions are made. Importantly, to assist in achieving sustainable healthcare systems the relationship of patients and physicians needs to be redefined and the future role of physicians clarified [132, 137].
Several studies have shown that good doctor-patient relationship positively effects health outcome [138,139,140]. Therefore, even if AI is successfully introduced in clinical practice human interaction may still be required in healthcare, particularly in areas with social interaction and a demand for a holistic perspective. However, the exact new roles of healthcare providers and their effect on patient outcome need to be tested. Diagnostics and therapy recommendation should also consider soft factors like patient fears and worries, the social environment, lifestyle and other conditions. It remains to be determined to what extent AI can address these ‘soft’ factors.
In view of the imminent collapse of the current healthcare system, relieving physicians of administrative and routine tasks could be the key . These tasks could be performed by AI systems, leaving more room and time for physicians to spend time with their patients and being empathetic. In addition, some specific medical professions, such as radiologists, could in future be replaced (in part) by algorithms able to interpret images even better than human doctors . However, it is very unlikely that AI will completely replace allied professionals in the health sector, yet doctors who use AI will likely replace those who do not . Therefore, knowledge and a basic understanding of the key principles of AI systems will need to be a crucial part for the future generations of all healthcare professionals.
Who are the beneficiaries of the novel approach?
The proposed PPPM concepts, based on multi-professional expertise, foresee a facilitated knowledge transfer between innovative sciences and advanced medical services, and incorporate strategies considering interests of:
Individuals in suboptimal health conditions predisposed to chronic pathologies such as HF—by personalised innovative screening programmes and preventive measures tailored to the individualised profiles
Patient cohorts—by precise patient stratification/prediction according to the disease subtype, risk factors, collateral pathologies etc.
Individual patients and their relatives—by measures adapted to personalised needs, including treatment algorithm tailored to the person, self-monitoring and active involvement of the closest environment in the treatment process
Healthcarers—by innovative educational programmes, digitalisation of routine procedures and decision-making process
Diagnostic and pharmaceutical industry—by creation of extended market opportunities
Biomedical sciences—by motivating innovative research in the context of PPPM
Healthcare systems—by improved operation processes and positive economy
Society as a whole—by advanced ethical and socio-economical solutions.
The complex measure of this PPPM approach may be considered as the medicine of the future . Contextually, the management of chronic diseases benefits particularly from application of AI technologies. This is due to huge potential in machine learning, data processing, computation analysis, monitoring and treatment of complex and collateral pathologies such as HF, which represents a highly heterogeneous patient cohort and may strongly benefit from improved subtyping in order to better characterise its pathophysiology and to develop novel targeted therapies . In this way, AI can help clinicians deliver more accurate care and protects patients against potential harm by treatment mistakes linked to the disease complexity . AI is becoming a mandatory technology in clinical practice. Machine learning and big data analytics have been proposed specifically for cardiology for predicting individual risks and applying genomic information for precision medical approach. Currently, run projects employ machine-learning techniques to address the problem of classification of HF subtypes and unbiased clustering analysis using dense pheno-mapping to identify phenotypically distinct HF categories .
The HF case report of the future
Although not experienced with computer technology, Mr. Johnson is using a physician avatar on a tablet computer. Access is available and data protection guaranteed by iris detection, which is so easy for him that he does not have to worry how to use it. Once weekly, he performs an outpatient visit by himself with the help of the physician avatar. The interval of the visits could also be longer depending on how stable his condition is, but he feels safer with this interval. The avatar advises him how to take medication, sends new prescriptions to the local pharmacy if required and tells him when to have his blood tested. To do this, he has to go to the local healthcare point in the village. However, he was told that in the near future, the avatar will be able to this at home. He already uses tools to measure his health—they call them sensors—and obviously, blood testing is going to be an addition to these tools. If he feels bad and has more symptoms, he can let the avatar do a check. Recently, he had more shortness of breath. The avatar advised him to take more diuretics and the symptoms disappeared rather quickly. After this episode, the avatar adjusted his treatment and now he feels very well. Before, there was an episode with very fast heart rate. The avatar told him that he needs to seek advice from his cardiologist because he had tachycardic atrial fibrillation. The cardiologist sent the ambulance to his home for a cardioversion. This went well and he felt much better.
The avatar gives him a lot of confidence and safety. He now has access to advice whenever needed at home. He has less contact with the healthcare providers but if he needs them, waiting time is much less. He gets advice regarding many other aspects of his daily life and has a much healthier lifestyle. In fact, exercise has become fun with the use of the avatar. He can hardly understand anymore, why he was so reluctant to use the avatar in the beginning.