There are many ways to process data to inform clinical decision-making in many areas within medicine from medicine management, to surgical precision to aid with treatment decisions [10] . The use of AI can help by 1) improving data collection and extraction within existing electronic health record systems to improve decision making; 2) enhancing measurement fidelity; 3) harnessing smartphone and biosensor data to inform clinical decision-making; 4) and finally helping co-ordinate care through analysis of the available data [10] . However, data sharing, privacy and standardization prevent AI integration in many systems within healthcare with each hospital within the UK operating a different system rich in complexity and limited in functionality often due to cost pressures [13] [14] . The data collected must be accurate and meaningful in order to produce outputs that are similar otherwise the adage of “rubbish in rubbish out” in computing still stands [15] . It becomes impossible to develop useful models that facilitate testing before changes are implemented for instance patient flow in Accident and Emergency (A & E) Departments if the data source is not clear or standardised and if stakeholders such as patients or clinicians are not involved in modelling or facilitating the understanding of the information required to achieve reliable effective decision making [16] .

In order to ensure reliability each area of consideration almost requires starting from scratch rather than trying to adapt to the information available. Inferences about the data required for decisions and the outputs generated in themselves can limit the usefulness of the system being adopted. There is however a pressure to develop as we try to embrace the technology available and look to improve evidence-based medical practice with technology.

[17] There is big data available within the National Health Service (NHS) to support clinical decision making and clinicians need systems of support by having this information available and presented to them in a digestible easily translatable form [18] .

AI requires a foundation of specialized hardware and software for writing and training machine learning algorithms. Those algorithms themselves need human input to achieve any result and can be flawed as “to err is to be human” and similar results happen with computer systems [19] [20] .

In general, AI systems work by ingesting large amounts of labelled training data, analysing the data for correlations and patterns, and using these patterns to make predictions about future states [21] . This input is one of the fundamental needs to begin developing AI models that support clinical decision-making, the use of the word support rather than make clinical decisions is deliberate [22] [23] .

As the hype around AI has accelerated, vendors have been scrambling to promote how their products and services use AI [24] . Often what they refer to as AI is simply one component of AI, such as machine learning [25] . AI requires a foundation of specialized hardware and software for writing and training machine learning algorithms [26] . No one programming language is synonymous with AI, but a few, including Python, R and Java, are popular [27] . So, there is heterogenicity in the software systems, this sometimes gives developers more flexibility in their approach as seldom does “one size fit all”.

In general, AI systems work by ingesting large amounts of labelled training data, analysing the data for correlations and patterns, and using these patterns to make predictions about future states where it has been trained in that function [28] . In this way, a chatbot that is fed examples of text chats can learn to produce lifelike exchanges with people [29] [30] , or an image recognition tool can learn to identify and describe objects in images by reviewing millions of examples and this can be even available on a smartphone [31] .

AI programming focuses on three cognitive skills: learning, reasoning and self-correction.

Learning processes. This aspect of AI programming focuses on acquiring data and creating rules for how to turn the data into actionable information. The rules, which are essentially algorithms, provide computing devices with step-by-step instructions for how to complete a specific task. These algorithms assist clinicians in decision making for instance when to operate and how best to deliver a successful operation by using a specific approach [32] .

[30] Algorithms have existed to provide this guidance before the availability of AI to manage what is an ever-increasing set of rules based on evidence gained from research and practice [33] . However, there are some difficulties in using this approach to achieve optimal patient care, including the magnitude of evidence available for often marginal statistical gains with poor generalisability to those with complex multimorbidity. The technology-prompted care was felt to be less patient focussed by some [34] .

Reasoning processes. This aspect of AI programming focuses on choosing the right algorithm to reach a desired outcome. Clinical reasoning is difficult to model, and “a doctor’s reasoning is built around a temporal unfolding of information” [35] . Reaching a medical diagnosis on which to base treatment decisions often involves ill-structured processes as the number of explanations for presenting conditions can be immense [36] . Computers traditionally use “structured decision-making approaches” whereas more “open-ended reasoning methods” are used by medical artificial intelligence (AI) programs” [36] . Further complexity is introduced when the AI suggests a route of treatment in addition to aiding diagnosis [36] . Sometimes these decisions are difficult to explain by the clinician [37] , who cannot easily check the validity of the decision [36] and the outcome may still be adverse making it difficult for a clinician to unpick how the decision was made [38] . There remain some ethical and legal dilemmas about where the responsibility for the decision lies [39] .

Self-correction processes. This aspect of AI programming is designed to continually fine-tune algorithms and ensure they provide the most accurate results possible. Part of how an AI improves relates to the processes that the AI employs to deal with uncertainty and this impacts on how much “trust” a clinician can attribute to that decision [40] .

In order to understand how AI had been implemented locally we surveyed local trust IT departments about the use of AI within their trusts and the practical applications being implemented to support clinical practice.

Is artificial intelligence important in aiding medical decision-making in current practice?

AI is important because it can give enterprises and healthcare insights into operational performance and potential improvements not previously evident [41] . In some cases, AI can perform tasks virtual or robotic-assisted better than humans [42] [43] . Particularly when it comes to repetitive, detail-oriented tasks like analyzing large numbers of legal documents to ensure relevant fields are filled in properly, AI tools often complete jobs quickly and with relatively few errors [44] . Despite the attractiveness of improved accuracy and performance, there is a general distrust of AI performance amongst patients in contrast to the media’s perception or government [45] .

In a way, it’s every clinician’s dream to automate data collection analysis when it is standard, duplicated, and monotonous even though detecting abnormality can be key and potentially lifesaving such as observation monitoring [46] .

Initially, the machines were used with simple algorithms to process healthcare information and were reactive with limited memory [24] , however, the AI machines used have improved and now are used to provide precision medicine, error reduction, drug development to name a few [47] .

Type 1: Reactive machines. These AI systems have no memory and are task-specific. An example is Deep Blue, the IBM chess program that beat Garry Kasparov in the 1990s. Deep Blue can identify pieces on the chessboard and make predictions, but because it has no memory, it cannot use past experiences to inform future ones [48] .

Type 2: Limited memory. These AI systems have memory, so they can use past experiences to inform future decisions. Some of the decision-making functions in self-driving cars are designed this way and the analytics performed by healthcare apps installed on mobile phones [49] .

Type 3: Theory of mind. Theory of mind is a psychology term [50] but when applied to AI, it means that the system would have the social intelligence to understand emotions. This type of AI will be able to infer human intentions and predict behaviour, a necessary skill for AI systems to become integral members of human teams [51] .

Type 4: Self-awareness. In this category, AI systems have a sense of self, which gives them consciousness [52] . Machines with self-awareness understand their own current state. At present, this type of AI does not yet exist [53] .

What examples of AI technology are used today in the UK in healthcare?

To understand what practical applications AI was being used within the United Kingdom a survey of 20 local hospitals within the Midlands was undertaken with 12 responding to a short questionnaire.

All the trusts that replied had used AI in the trust mainly in administrative areas rather than care areas. There were projects underway that used machine learning algorithms to look at results of anemia levels in pregnancy [54] , to help analyse results in patients with acute kidney injury (AKI) [55] and imaging review for scaphoid fractures [56] .

Nine of the trusts felt that clinical decision-making was being positively impacted by the introduction of AI improving decision-making. It was felt to have been cost effective by half although the consensus was that it was a little too early to be evaluating its effectiveness as the AI being used was still being calibrated and evaluated.

Automation (not currently utilized in UK practice locally). When paired with AI technologies, automation tools can expand the volume and types of tasks performed [3] . An example is robotic process automation (RPA), a type of software that automates repetitive, rules-based data processing tasks traditionally done by humans [57] . When combined with machine learning and emerging AI tools, RPA can automate bigger portions of enterprise jobs, enabling RPA’s tactical bots to pass along intelligence from AI and respond to process changes improving healthcare access for instance and record keeping [58] .

Machine learning. This is the science of getting a computer to act without programming [59] [60] . Deep learning is a subset of machine learning that, in very simple terms, can be thought of as the automation of predictive analytics for instance predicting psychosis from recorded dialogue [60] . This type of AI application is proving useful in diagnostics and treatment finding a place within practice locally. There are three types of machine learning algorithms:

Supervised learning. Data sets are labelled so that patterns can be detected and used to label new data sets [61] . For instance, the detection of acute kidney injury from laboratory results is being developed or used in 4 of the 15 trusts surveyed [62] .

Unsupervised learning. Data sets are not labelled and are sorted according to similarities or differences [63] , such as the detection of Sepsis in Intensive Care Units (ICU) [64] . At present no local units are utilising programmes with this design.

Reinforcement learning. Data sets are not labelled but, after performing an action or several actions, the AI system is given feedback [65] . This had no applications in clinical practice locally as safe practice is the prime goal and therefore AI is being used to build on previous learning and ensure rules are adhered to every time [66] [67] .

Machine vision. This technology gives a machine the ability to see [68] . Machine vision captures and analyses visual information using a camera, analog-to-digital conversion and digital signal processing [69] . It is often compared to human eyesight, but machine vision is not bound by biology and can be programmed to see through walls [70] . It is used in a range of applications from signature identification to medical image analysis and is being used to look for scaphoid fractures on a trial basis and to help scan medical notes in three trusts.

[56] Computer vision, which is focused on machine-based image processing [71] , is often conflated with machine vision but these are two separate things.

Natural language processing (NLP). This is the processing of human language by a computer program [72] . One of the older and best-known examples of NLP is spam detection, which looks at the subject line and text of an email and decides if it’s junk [73] . Current approaches to NLP are based on machine learning facilitating text translation, sentiment analysis and speech recognition [74] .

Robotics. This field of engineering focuses on the design and manufacturing of robots to perform tasks that are difficult for humans to perform consistently [75] . For example, robots are used in surgery to improve the precision in alignment of joint replacement [76] or tumour excision [77] . Researchers are also using machine learning to build robots that can interact in social settings and help with patient rehabilitation like Cyberdyne’s Hybrid Assistive Limb (HAL) exoskeleton designed for the rehabilitation of patients with spinal cord injuries and stroke [3] [78] . Autonomous surgery uses a combination of computer vision, image recognition, deep learning and robotics to build automated skills for piloting a robot to deliver precision surgery [75] . At present most trusts eight of those that had responded employed a form of robotic-assisted surgery to improve precision and outcomes.