Introduction

     The application of Artificial Intelligence in healthcare was first described in 1976 when a computer algorithm was used to identify causes of acute abdominal pain. Since then, there have been diverse and manifold applications of AI in medicine proposed. 

These range from aiding in the detection of diseases, such as detecting skin cancers in dermatology or diabetic retinopathy to improved pathology, for example in classifying scans in radiology or delineating electrocardiogram features in cardiology to predicting disease patterns and epidemiology. 

However, despite the healthcare industry’s investment in AI technology, the adoption of AI solutions and their implementation in healthcare remains in its infancy. Some of the most pressing challenges facing healthcare are reduced expenditure, physician shortage and burnout, and the shift toward chronic disease management. 

As the workforce appears to be critically stretched, it has been proposed that AI, in particular deep learning, could be the key to filling this gap. If AI systems are more widely adopted, not only could they reduce the workload but also increase the quality of patient care. The question, therefore, remains if such opportunities for AI in healthcare do exist, why they remain untapped, and what hinders their implementation.

What is Artificial Intelligence?

    Artificial intelligence is the catch-all name for machine learning, deep learning, and robotics. As such, it is a bit of a misnomer – AI isn’t a system, it is a tool implemented in a system. At its core, AI is a tool that strives to mimic human brain power and decision-making processes by the initial creation of algorithms that learn from themselves and that continue and continue to learn from their own experience much like humans. This allows the system to harness that experience in order in order to perform certain tasks better, faster or more efficiently than humans. 

Machine Learning

This is an AI application whereby the system uses an iterative process until it learns based on past experience. An example could be an application that processes similar data repeatedly until it can predict an answer.

Deep Learning

This is a more complex form of AI whereby the system is fed large amounts of data until it learns by example deep learning is inspired by the human neural network, allowing the system to discover patterns, face recognition software, and self-driving cars are examples of this. 

Robotics 

This is a machine performing a human task, usually with accuracy and precision beyond what a human is capable of. A new generation Roomba vacuum is an example of this.    

Much like humans, AI thrives on and improves with experience. In the case of AI-driven systems, experience equates to more and more data, more possess through data, and more neural connectivity. An easy human analogy is a parent keeping a child from touching a hot stove. 

The parent will do that repeatedly, until the child knows not to touch the hot stove eventually the child will know through experience and neural connectively that hot things of any kind will hurt and, without being told by a parent, will know not to touch an open flame.

To some extent, AI has ceased to be important to most users and most Americans 85% since 2017 report using at least one AI device, such as a navigation device. But while we embrace and accept AI in our daily lives, the adoption rate has been slower. A 2017 survey pegged its use by healthcare systems at less than 50%. 

Artificial Intelligence and Healthcare 

    Artificial intelligence in health care garners greater interest and the investment sector has taken note. AI in healthcare is an investment favorite; about $1 billion was invested in healthcare AI in recent years. The AI health market hit $6.6 billion by 2021 eleven times the size of the 2014 market. According to consulting giant Accenture, the top 5 healthcare applications are: Orthopedic Robot-Assisted Surgery, Virtual Nursing Assistants, Administrative Workflow Assistance, Fraud Detection, and Dosage Error Reduction. Connected machines, clinical trial participant identifiers, preliminary diagnosis, automated image diagnosis, and cybersecurity round out the top 10.  The top AI application according to Accenture, is robot-assisted surgery. Cognitive robotics can integrate information from pre-op medical records with real-time operating metrics to physically guide and enhance the physician’s instrument precision. 

    While investment interest is broad across all AI healthcare applications, the application currently in use, or for which it is great academic and research interest, involves particularly radiology and pathology, and those around workflows and population health. In other words, applications that can augment human activity by doing tedious, or by relieving clinicians of non-clinical tasks, or determining risk factors, are the ones currently in use the most.     Some examples:

• Screening for diabetic retinopathy where an image of the patients retina is uploaded to the software system in the cloud, the system reviews the scan and returns a result of either ‘More than mild diabetic retinopathy detected: refer to an eye care professional’ or ‘negative for more than mild diabetic retinopathy and re-screen again in 12 months.  

• Post-partum depression in developing regions to see if live and automated text messaging augmented by AI could deliver care to a country in which mental health providers are very scarce. 

• Predicting adverse health events in individuals by algorithms that can predict an in-surgical and post-discharge complications in a patients already are developed are algorithms that detect heart failure and sepsis at their earliest stages. 

• Virtual nursing assistants: nurse avatars answer questions 24x7, automate data collection, and generally free up human nurses to perform clinical tasks rather than administrative tasks. 

• Workflow applications that automate administrative workflow tasks that allow clinicians to save time. These applications can include voice-to-text transcriptions for charting, prescriptions, and test orders as well as computer-assisted physician documentation programs. 

Mission control command centers for example John Hopkins Hospital created a command center powered by AI technology predictive analytics which helps the command center staff to know, among other things, how to reduce wait times by predicting room turns and patient discharges. 

Humber River Hospital in Toronto, Canada has gone even further by going all-digital, employing tracking devices on patients and staff to monitor and manage the flow of patient activity, using robots to prepare to and deliver supplies and medications, and automating 75% of its non-clinical operations like the pharmacy, laundry, and food delivery. According to Humber River Hospital, it bills itself as the world’s first all-digital, the impact of its command center on patients and healthcare delivery systems will be equivalent to adding 45 new beds, enabling them to serve 4,000 more patients per year. 

The Drawbacks of Artificial Intelligence

    There are diverse challenges to the successful implementation of any information technology in health care let alone Artificial Intelligence. These challenges occur at all stages of AI implementation: data acquisition, technology development, clinical implementation, and ethical and social issues. 

Data Challenges

The first barrier is data availability where deep learning models require large databases to accurately classify or predict different tasks. Sectors, where AI has seen immense progression, are those with large datasets available to enable more complex, precise algorithms. 

In health care, however, the availability of data is a complex issue. On the organizational level, health data is not only expensive but there is ingrained reluctance towards sharing with other hospitals as they are considered the property of each hospital to manage their individual patients. A further issue faced is the continued availability of data following the introduction of the algorithm analyzing it.

 Ideally, AI-based systems; require continuous improvements from training with progressively bigger datasets. However, due to organizational resistance, this is often difficult to achieve. It has been suggested that what is required for information technology and AI to progress in health care is a transformative shift from focusing on individual patient treatment to overall patient outcomes, 

Additionally, technical developments may alleviate the challenge of limited datasets, for example through improved algorithms that can work on or less expansive basis as opposed to multimodal learning, as well as the overuse challenge of storing these ever-increasing datasets. 

    AI-based applications bring concerns about data privacy and security. Health data is sensitive and a frequent target for data breaches. The protection of patient data is thus paramount. With the development of AI come additional concerns regarding data privacy, as individuals may mistake artificial systems for humans and allow further unconscious data collection. Patient consent is therefore a crucial component in data privacy concerns, as healthcare organizations may permit the large-scale use of patient data for AI training without sufficient individual patient consent. 

Potential solutions to this issue include tightening regulations and laws with respect to personal data, such as the General Data Processing Regulations and Health Research Regulations enforced across Europe in 2018 as an example.

Developer Challenges

    Biases may occur in the collection of data used to train models, racial biases may be introduced in the creation of databases, with minorities being under-represented thereby leading to lower-than-expected prediction performance. Various methods exist to counteract this bias, such as a recent bias-resilient neural network that reduces the effect of such confounding variables. Only time will tell whether such approaches will be successful in eliminating biases practices. This is just one of many issues that are of major challenge with respect to drawbacks in Artificial Intelligence. 

Clinical Implementation Practices

    The first barrier to successful implementation is the lack of empirical evidence proving the efficacy of AI-based interventions in prospective clinical trials. Empirical research remains scarce and largely pertains to AI in the general workforce, not it’s an effect in healthcare is mostly preclinical, and occurs in artificial environments. Results are therefore difficult to extrapolate to reality. Randomized controlled trials are considered the gold standard in medicine but are lacking in proving the efficacy of AI in healthcare. 

Ethical Challenges

    Ethical concerns and protests have beleaguered AI since its inception. Aside from those data privacy and safety, the main concern is accountability. Particularly in healthcare, poor decisions carry heavy consequences and the current paradigm is that some people must be held accountable. AI is often viewed as a “black box” where one can’t discern why the algorithm arrived at a particular prediction or recommendation. 

One could argue that the “black-box” phenomenon need not be concerning for algorithms in applications with lower stakes at hand, such as those that are not patient-centered but instead focused on efficacy or improved operations. However, the question of accountability is far more crucial when considering AI applications that aim to improve outcomes for patients, especially when things go wrong. 

Therefore, it is unclear who should take responsibility should the system be wrong. To hold the physician accountable may be unfair as the algorithm is neither developed nor controlled in any manner by them, yet to hold the developer accountable seems too far removed from the clinical context. In China, it is illegal for AI to make any decision in healthcare, requiring some form of human input such that they are held accountable. 

Social Challenges

    A longstanding concern regarding AI in healthcare is the fear that it will replace jobs, thus rendering healthcare workers obsolete. The threat of replacement translates to distrust and dislike of AI-based interventions. 

However, this belief is based largely on a misunderstanding of AI in its various forms. Even when disregarding the years that it would hypothetically take for AI to be advanced enough too successfully to replace healthcare workers, the introductions of AI do not mean that jobs become obsolete, but rather re-engineered. 

Many aspects of medicine are innately human and unpredictable and cannot ever be completely linear or structured like an algorithm.  However, the damaging effect of distrust in AI is clear and represents a further challenge to its adoption. Conversely, an insufficient understanding of AI may lead to unrealistically high expectations of its results and efficacy. 

The general public may misunderstand the current capabilities of AI, and their resulting disappointment may give way to reluctance to trust such technology. Therefore again, a more public debate about AI in health care is needed to address these beliefs amongst patients and health care professionals. 

Future of Artificial Intelligence 

    AI is advancing in many fields. AI has the capacity to have an enormous impact on doctors and patients in healthcare. Because of its capability to collect and analyze a huge amount and various forms of data AI has a bright future in healthcare. 

AI could yield considerably quicker and much more accurate diagnoses for a broader section of the population. Individuals without access to extremely specialized healthcare will achieve greater results and care with AI. Artificial intelligence has already changed the shape of healthcare. 

However, there are many details and challenges that need to be addressed before its implementation to the clinical practice. The current regulations and lack of standards to evaluate the safety and efficacy of AI algorithms must be changed. Before incorporating AI into clinical practice, legislative issues should be solved. In the future cognitive computers will be assisting clinicians in their decision-making and determining and predicting patient outcomes. 

The massive amount of data generated by routine daily work-up necessitates the application of AI into practice. It is important not to fear AI but to embrace it as healthcare becomes more digitalized every day. AI will provide clinicians the skills to interpret patient-level data in greater depth than ever before. Physicians should prepare themselves for the era of AI and acquire the needed skills on when to apply AI models and how to interpret results properly.  

Conclusion

    Artificial intelligence technology is well on its way to enhancing the scope to enhance technology at many levels, leading to much better, faster patient outcomes. 

Health organizations must quickly adapt to evolving technologies, regulations, and consumer demands. AI machine learning, and deep learning can assist with proactive patient care, reduce future risk, and streamline work processes. It is also helpful in robotic-assisted surgery, diagnosing diseases at their earlier stages. 

It is possible to outsource data storage, leverage an advanced theoretical understanding of data, and take advantage of computers that can execute complex tasks at high speeds and lower costs. It is used as a virtual nursing assistant, clinical judgment or diagnosis, image analysis, workflow, and administrative tasks.