Artificial intelligence in healthcare
Artificial intelligence in healthcare is the application of artificial intelligence (AI) to copy human cognition in the analysis, presentation, and understanding of complex medical and health care data, or to exceed human capabilities by providing new ways to diagnose, treat, or prevent disease.[1][2] Specifically, AI is the ability of computer algorithms to arrive at approximate conclusions based solely on input data.
The primary aim of health-related AI applications is to analyze relationships between clinical data and patient outcomes.[3] AI programs are applied to practices such as diagnostics, treatment protocol development, drug development, personalized medicine, and patient monitoring and care. What differentiates AI technology from traditional technologies in healthcare is the ability to gather larger and more diverse data, process it, and produce a well-defined output to the end-user. AI does this through machine learning algorithms and deep learning. Because radiographs are the most common imaging tests conducted in most radiology departments, the potential for AI to help with triage and interpretation of traditional radiographs (X-ray pictures) is particularly noteworthy.[4] These processes can recognize patterns in behavior and create their own logic. To gain useful insights and predictions, machine learning models must be trained using extensive amounts of input data. AI algorithms behave differently from humans in two ways: (1) algorithms are literal: once a goal is set, the algorithm learns exclusively from the input data and can only understand what it has been programmed to do, (2) and some deep learning algorithms are black boxes; algorithms can predict with extreme precision, but offer little to no comprehensible explanation to the logic behind its decisions aside from the data and type of algorithm used.[5]
As widespread use of AI in healthcare is relatively new, research is ongoing into its application in various fields of medicine and industry. Additionally, greater consideration is being given to the unprecedented ethical concerns related to its practice such as data privacy, automation of jobs, and representation biases.[6] Furthermore, new technologies brought about by AI in healthcare are often resisted by healthcare leaders, leading to slow and erratic adoption.[7]
In recent years, AI has played a leading role in the use and valuation of extensive collections of data, Google and the Mayo Clinic, for example, have announced a partnership to solve complex medical problems using data-driven medical innovation, or a team from the University of California, San Diego was able to create a diagnostic program by training AI on medical records from 1.3 million patients under the age of 18.80 .[8]
Research in the 1960s and 1970s produced the first problem-solving program, or expert system, known as Dendral.[9][10] While it was designed for applications in organic chemistry, it provided the basis for a subsequent system MYCIN,[11] considered one of the most significant early uses of artificial intelligence in medicine.[11][12] MYCIN and other systems such as INTERNIST-1 and CASNET did not achieve routine use by practitioners, however.[13]
The 1980s and 1990s brought the proliferation of the microcomputer and new levels of network connectivity. During this time, there was a recognition by researchers and developers that AI systems in healthcare must be designed to accommodate the absence of perfect data and build on the expertise of physicians.[14] Approaches involving fuzzy set theory,[15] Bayesian networks,[16] and artificial neural networks,[17][18] have been applied to intelligent computing systems in healthcare.
Medical and technological advancements occurring over this half-century period that have enabled the growth of healthcare-related applications of AI to include:
AI algorithms can also be used to analyze large amounts of data through electronic health records for disease prevention and diagnosis. Medical institutions such as The Mayo Clinic, Memorial Sloan Kettering Cancer Center,[26][27] and the British National Health Service,[28] have developed AI algorithms for their departments. Large technology companies such as IBM[29] and Google,[28] have also developed AI algorithms for healthcare. Additionally, hospitals are looking to AI software to support operational initiatives that increase cost saving, improve patient satisfaction, and satisfy their staffing and workforce needs.[30] Currently, the United States government is investing billions of dollars to progress the development of AI in healthcare.[5] Companies are developing technologies that help healthcare managers improve business operations through increasing utilization, decreasing patient boarding, reducing length of stay and optimizing staffing levels.[31]
Clinical applications[edit]
Cardiovascular[edit]
Artificial intelligence algorithms have shown promising results in accurately diagnosing and risk stratifying patients with concern for coronary artery disease, showing potential as an initial triage tool.[32][33] Other algorithms have been used in predicting patient mortality, medication effects, and adverse events following treatment for acute coronary syndrome.[32] Wearables, smartphones, and internet-based technologies have also shown the ability to monitor patients' cardiac data points, expanding the amount of data and the various settings AI models can use and potentially enabling earlier detection of cardiac events occurring outside of the hospital.[34] Another growing area of research is the utility of AI in classifying heart sounds and diagnosing valvular disease.[35] Challenges of AI in cardiovascular medicine have included the limited data available to train machine learning models, such as limited data on social determinants of health as they pertain to cardiovascular disease.[36]
A key limitation in early studies evaluating AI were omissions of data comparing algorithmic performance to humans. Examples of studies which assess AI performance relative to physicians includes how AI is noninferior to humans in interpretation of cardiac echocardiograms[37] and that AI can diagnose heart attack better than human physicians in the emergency setting, reducing both low-value testing and missed diagnoses.[38]
In cardiovascular tissue engineering and organoid studies, AI is increasingly used to analyze microscopy images, and integrate electrophysiological read outs.[39]
Dermatology[edit]
Dermatology is an imaging abundant speciality[40] and the development of deep learning has been strongly tied to image processing. Therefore, there is a natural fit between the dermatology and deep learning. There are three main imaging types in dermatology: contextual images, macro images, micro images.[41] For each modality, deep learning showed great progress.[42] Han et al. showed keratinocytic skin cancer detection from face photographs.[43] Esteva et al. demonstrated dermatologist-level classification of skin cancer from lesion images.[44] Noyan et al. demonstrated a convolutional neural network that achieved 94% accuracy at identifying skin cells from microscopic Tzanck smear images.[45] A concern raised with this work is that it has not engaged with disparities related to skin color or differential treatment of patients with non-white skin tones.[46]
According to some researchers, AI algorithms have been shown to be more effective than dermatologists at identifying cancer.[47] However, a 2021 review article found that a majority of papers analyzing the performance of AI algorithms designed for skin cancer classification failed to use external test sets.[48] Only four research studies were found in which the AI algorithms were tested on clinics, regions, or populations distinct from those it was trained on, and in each of those four studies, the performance of dermatologists was found to be on par with that of the algorithm. Moreover, only one study[49] was set in the context of a full clinical examination; others were based on interaction through web-apps or online questionnaires, with most based entirely on context-free images of lesions. In this study, it was found that dermatologists significantly outperformed the algorithms. Many articles claiming superior performance of AI algorithms also fail to distinguish between trainees and board-certified dermatologists in their analyses.[48]
It has also been suggested that AI could be used to automatically evaluate the outcome of maxillo-facial surgery or cleft palate therapy in regard to facial attractiveness or age appearance.[50][51]
Gastroenterology[edit]
AI can play a role in various facets of the field of gastroenterology. Endoscopic exams such as esophagogastroduodenoscopies (EGD) and colonoscopies rely on rapid detection of abnormal tissue. By enhancing these endoscopic procedures with AI, clinicians can more rapidly identify diseases, determine their severity, and visualize blind spots. Early trials in using AI detection systems of early gastric cancer have shown sensitivity close to expert endoscopists.[52]
Obstetrics and gynaecology[edit]
Artificial intelligence, or AI, utilises massive amounts of data to help with predicting illness, prevention, and diagnosis, as well as patient monitoring. In obstetrics, artificial intelligence is utilised in magnetic resonance imaging, ultrasound, and foetal cardiotocography. AI contributes in the resolution of a variety of obstetrical diagnostic issues.[53]
Infectious diseases[edit]
AI has shown potential in both the laboratory and clinical spheres of infectious disease medicine.[54] As the novel coronavirus ravages through the globe, the United States is estimated to invest more than $2 billion in AI-related healthcare research by 2025, more than 4 times the amount spent in 2019 ($463 million).[55] While neural networks have been developed to rapidly and accurately detect a host response to COVID-19 from mass spectrometry samples, a scoping review of the literature found few examples of AI being used directly in clinical practice during the COVID-19 pandemic itself.[56] Other applications include support-vector machines identifying antimicrobial resistance, machine learning analysis of blood smears to detect malaria, and improved point-of-care testing of Lyme disease based on antigen detection. Additionally, AI has been investigated for improving diagnosis of meningitis, sepsis, and tuberculosis, as well as predicting treatment complications in hepatitis B and hepatitis C patients.[54]
Musculoskeletal[edit]
AI has been used to identify causes of knee pain that doctors miss, that disproportionately affect Black patients.[57] Underserved populations experience higher levels of pain. These disparities persist even after controlling for the objective severity of diseases like osteoarthritis, as graded by human physicians using medical images, raising the possibility that underserved patients’ pain stems from factors external to the knee, such as stress. Researchers have conducted a study using a machine-learning algorithm to show that standard radiographic measures of severity overlook objective but undiagnosed features that disproportionately affect diagnosis and management of underserved populations with knee pain. They proposed that new algorithmic measure ALG-P could potentially enable expanded access to treatments for underserved patients.[58]
Neurology[edit]
The use of AI technologies has been explored for use in the diagnosis and prognosis of Alzheimer's disease (AD). For diagnostic purposes, machine learning models have been developed that rely on structural MRI inputs.[59] The input datasets for these models are drawn from databases such as the Alzheimer's Disease Neuroimaging Initiative.[60] Researchers have developed models that rely on convolutional neural networks with the aim of improving early diagnostic accuracy.[61] Generative adversarial networks are a form of deep learning that have also performed well in diagnosing AD.[62] There have also been efforts to develop machine learning models into forecasting tools that can predict the prognosis of patients with AD. Forecasting patient outcomes through generative models has been proposed by researchers as a means of synthesizing training and validation sets.[63] They suggest that generated patient forecasts could be used to provide future models larger training datasets than current open access databases.
Oncology[edit]
AI has been explored for use in cancer diagnosis, risk stratification, molecular characterization of tumors, and cancer drug discovery. A particular challenge in oncologic care that AI is being developed to address is the ability to accurately predict which treatment protocols will be best suited for each patient based on their individual genetic, molecular, and tumor-based characteristics.[64] AI has been trialed in cancer diagnostics with the reading of imaging studies and pathology slides.[65]
In January 2020, Google DeepMind announced an algorithm capable of surpassing human experts in breast cancer detection in screening scans.[66][67] A number of researchers, including Trevor Hastie, Joelle Pineau, and Robert Tibshirani among others, published a reply claiming that DeepMind's research publication in Nature lacked key details on methodology and code, "effectively undermin[ing] its scientific value" and making it impossible for the scientific community to confirm the work.[68] In the MIT Technology Review, author Benjamin Haibe-Kains characterized DeepMind's work as "an advertisement" having little to do with science.[69]
In July 2020, it was reported that an AI algorithm developed by the University of Pittsburgh achieves the highest accuracy to date in identifying prostate cancer, with 98% sensitivity and 97% specificity.[70][71] In 2023 a study reported the use of AI for CT-based radiomics classification at grading the aggressiveness of retroperitoneal sarcoma with 82% accuracy compared with 44% for lab analysis of biopsies.[72][73]
Ophthalmology[edit]
Artificial intelligence-enhanced technology is being used as an aid in the screening of eye disease and prevention of blindness.[74] In 2018, the U.S. Food and Drug Administration authorized the marketing of the first medical device to diagnose a specific type of eye disease, diabetic retinopathy using an artificial intelligence algorithm.[75] Moreover, AI technology may be used to further improve "diagnosis rates" because of the potential to decrease detection time.[76]
Systems applications[edit]
Disease diagnosis[edit]
An article by Jiang, et al. (2017) demonstrated that there are several types of AI techniques that have been used for a variety of different diseases, such as support vector machines, neural networks, and decision trees. Each of these techniques is described as having a "training goal" so "classifications agree with the outcomes as much as possible…".
To demonstrate some specifics for disease diagnosis/classification there are two different techniques used in the classification of these diseases including using artificial neural networks (ANN) and Bayesian networks (BN). It was found that ANN was better and could more accurately classify diabetes and cardiovascular disease.
Through the use of machine learning classifiers (MLCs), artificial intelligence has been able to substantially aid doctors in patient diagnosis through the manipulation of mass electronic health records (EHRs).[109] Medical conditions have grown more complex, and with a vast history of electronic medical records building, the likelihood of case duplication is high.[109] Although someone today with a rare illness is less likely to be the only person to have had any given disease, the inability to access cases from similarly symptomatic origins is a major roadblock for physicians.[109] The implementation of AI to not only help find similar cases and treatments, such as through early predictors of Alzheimer's disease and dementias,[110] but also factor in chief symptoms and help the physicians ask the most appropriate questions helps the patient receive the most accurate diagnosis and treatment possible.[109]
Recent developments in statistical physics, machine learning, and inference algorithms are being explored for their potential in improving medical diagnostic approaches.[111] Combining the skills of medical professionals and machines can help overcome decision-making weaknesses in medical practice. To do so, one needs precise disease definitions and a probabilistic analysis of symptoms and molecular profiles. Physicists have been studying similar problems for years, using microscopic elements and their interactions to extract macroscopic states of various physical systems. Physics inspired machine learning approaches can thus be applied to study disease processes and to perform biomarker analysis.
The trend of large health companies merging allows for greater health data accessibility. Greater health data lays the groundwork for the implementation of AI algorithms.
A large part of industry focus of implementation of AI in the healthcare sector is in the clinical decision support systems. As more data is collected, machine learning algorithms adapt and allow for more robust responses and solutions.[95] Numerous companies are exploring the possibilities of the incorporation of big data in the healthcare industry. Many companies investigate the market opportunities through the realms of "data assessment, storage, management, and analysis technologies" which are all crucial parts of the healthcare industry.[129]
The following are examples of large companies that have contributed to AI algorithms for use in healthcare:
Digital consultant apps use AI to give medical consultation based on personal medical history and common medical knowledge. Users report their symptoms into the app, which uses speech recognition to compare against a database of illnesses. Babylon then offers a recommended action, taking into account the user's medical history. Entrepreneurs in healthcare have been effectively using seven business model archetypes to take AI solution[buzzword] to the marketplace. These archetypes depend on the value generated for the target user (e.g. patient focus vs. healthcare provider and payer focus) and value capturing mechanisms (e.g. providing information or connecting stakeholders).
IFlytek launched a service robot "Xiao Man", which integrated artificial intelligence technology to identify the registered customer and provide personalized recommendations in medical areas. It also works in the field of medical imaging. Similar robots are also being made by companies such as UBTECH ("Cruzr") and Softbank Robotics ("Pepper").
The Indian startup Haptik recently developed a WhatsApp chatbot which answers questions associated with the deadly coronavirus in India.
With the market for AI expanding constantly, large tech companies such as Apple, Google, Amazon, and Baidu all have their own AI research divisions, as well as millions of dollars allocated for acquisition of smaller AI based companies.[129] Many automobile manufacturers are beginning to use machine learning healthcare in their cars as well.[129] Companies such as BMW, GE, Tesla, Toyota, and Volvo all have new research campaigns to find ways of learning a driver's vital statistics to ensure they are awake, paying attention to the road, and not under the influence of substances or in .[129]
Expanding care to developing nations[edit]
Artificial intelligence continues to expand in its abilities to diagnose more people accurately in nations where fewer doctors are accessible to the public. Many new technology companies such as SpaceX and the Raspberry Pi Foundation have enabled more developing countries to have access to computers and the internet than ever before.[132] With the increasing capabilities of AI over the internet, advanced machine learning algorithms can allow patients to get accurately diagnosed when they would previously have no way of knowing if they had a life-threatening disease or not.[132]
Using AI in developing nations that do not have the resources will diminish the need for outsourcing and can improve patient care. AI can allow for not only diagnosis of patient in areas where healthcare is scarce, but also allow for a good patient experience by resourcing files to find the best treatment for a patient.[133] The ability of AI to adjust course as it goes also allows the patient to have their treatment modified based on what works for them; a level of individualized care that is nearly non-existent in developing countries.[133]
Ethical concerns[edit]
Data collection[edit]
In order to effectively train Machine Learning and use AI in healthcare, massive amounts of data must be gathered. Acquiring this data, however, comes at the cost of patient privacy in most cases and is not well received publicly. For example, a survey conducted in the UK estimated that 63% of the population is uncomfortable with sharing their personal data in order to improve artificial intelligence technology.[138] The scarcity of real, accessible patient data is a hindrance that deters the progress of developing and deploying more artificial intelligence in healthcare.
Furthermore, the lack of current regulations surrounding AI in the United States has generated concerns about mismanagement of patient data, such as with corporations utilizing patient data for financial gain. For example, Roche, a Swiss healthcare company, was found to have purchased healthcare data for approximately 2 million cancer patients at an estimated total cost of $1.9 billion.[147] Naturally, this generates questions of ethical concern; Is there a monetary price that can be set for data, and should it depend on its perceived value or contributions to science? Is it fair to patients to sell their data? These concerns were addressed in a survey conducted by the Pew Research Center in 2022 that asked Americans for their opinions about the increased presence of AI in their daily lives, and the survey estimated that 37% of Americans were more concerned than excited about such increased presence, with 8% of participants specifically associating their concern with "people misusing AI".[148] Ultimately, the current potential of artificial intelligence in healthcare is additionally hindered by concerns about mismanagement of data collected, especially in the United States.
Automation[edit]
A systematic review and thematic analysis in 2023 showed that most stakeholders including health professionals, patients, and the general public doubted that care involving AI could be empathetic.[149]
According to a 2019 study, AI can replace up to 35% of jobs in the UK within the next 10 to 20 years.[150] However, of these jobs, it was concluded that AI has not eliminated any healthcare jobs so far. Though if AI were to automate healthcare-related jobs, the jobs most susceptible to automation would be those dealing with digital information, radiology, and pathology, as opposed to those dealing with doctor-to-patient interaction.[150]
Automation can provide benefits alongside doctors as well. It is expected that doctors who take advantage of AI in healthcare will provide greater quality healthcare than doctors and medical establishments who do not.[151] AI will likely not completely replace healthcare workers but rather give them more time to attend to their patients. AI may avert healthcare worker burnout and cognitive overload.
Recently, there have been many discussions between healthcare experts in terms of AI and elder care. In relation to elder care, AI bots have been helpful in guiding older residents living in assisted living with entertainment and company. These bots are allowing staff in the home to have more one-on-one time with each resident, but the bots are also programmed with more ability in what they are able to do; such as knowing different languages and different types of care depending on the patient's conditions. The bot is an AI machine, which means it goes through the same training as any other machine - using algorithms to parse the given data, learn from it and predict the outcome in relation to what situation is at hand[152]
Bias[edit]
Since AI makes decisions solely on the data it receives as input, it is important that this data represents accurate patient demographics. In a hospital setting, patients do not have full knowledge of how predictive algorithms are created or calibrated. Therefore, these medical establishments can unfairly code their algorithms to discriminate against minorities and prioritize profits rather than providing optimal care.[153] A recent scoping review identified 18 equity challenges along with 15 strategies that can be implemented to help address them when AI applications are developed using many-to-many mapping.[154]
There can also be unintended bias in these algorithms that can exacerbate social and healthcare inequities.[153] Since AI's decisions are a direct reflection of its input data, the data it receives must have accurate representation of patient demographics. White males are overly represented in medical data sets.[155] Therefore, having minimal patient data on minorities can lead to AI making more accurate predictions for majority populations, leading to unintended worse medical outcomes for minority populations.[156] Collecting data from minority communities can also lead to medical discrimination. For instance, HIV is a prevalent virus among minority communities and HIV status can be used to discriminate against patients.[155] In addition to biases that may arise from sample selection, different clinical systems used to collect data may also impact AI functionality. For example, radiographic systems and their outcomes (e.g., resolution) vary by provider. Moreover, clinician work practices, such as the positioning of the patient for radiography, can also greatly influence the data and make comparability difficult.[157] However, these biases are able to be eliminated through careful implementation and a methodical collection of representative data.
A final source of bias, which has been called "label choice bias", arises when proxy measures are used to train algorithms, that build in bias against certain groups. For example, a widely used algorithm predicted health care costs as a proxy for health care needs, and used predictions to allocate resources to help patients with complex health needs. This introduced bias because Black patients have lower costs, even when they are just as unhealthy as White patients.[158] Solutions to the "label choice bias" aim to match the actual target (what the algorithm is predicting) more closely to the ideal target (what researchers want the algorithm to predict), so for the prior example, instead of predicting cost, researchers would focus on the variable of healthcare needs which is rather more significant. Adjusting the target led to almost double the number of Black patients being selected for the program.