Artificial intelligence in healthcare

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,[25][26] though few studies have directly compared the accuracy of machine learning models to clinician diagnostic ability.[27] Other algorithms have been used in predicting patient mortality, medication effects, and adverse events following treatment for acute coronary syndrome.[25] 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.[28] Another growing area of research is the utility of AI in classifying heart sounds and diagnosing valvular disease.[29] 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.[30]

Dermatology is an imaging abundant speciality[31] 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 3 main imaging types in dermatology: contextual images, macro images, micro images.[32] For each modality, deep learning showed great progress.[33] Han et al. showed keratinocytic skin cancer detection from face photographs.[34] Esteva et al. demonstrated dermatologist-level classification of skin cancer from lesion images.[35] Noyan et al. demonstrated a convolutional neural network that achieved 94% accuracy at identifying skin cells from microscopic Tzanck smear images.[36]

Recent advances have suggested the use of AI to describe and evaluate the outcome of maxillo-facial surgery or the assessment of cleft palate therapy in regard to facial attractiveness or age appearance.[37][38]

In 2018, a paper published in the journal Annals of Oncology mentioned that skin cancer could be detected more accurately by an artificial intelligence system (which used a deep learning convolutional neural network) than by dermatologists. On average, the human dermatologists accurately detected 86.6% of skin cancers from the images, compared to 95% for the CNN machine.[39]

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.[40]

AI has shown potential in both the laboratory and clinical spheres of infectious disease medicine.[41] 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).[42] Neural networks have been developed to rapidly and accurately detect a host response to COVID-19 from mass spectrometry samples. 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.[41]

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.[43] Through its ability to translate images to mathematical sequences, AI has been trialed in cancer diagnostics with the reading of imaging studies and pathology slides.[44] In January 2020, researchers demonstrated an AI system, based on a Google DeepMind algorithm, capable of surpassing human experts in breast cancer detection.[45][46] 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.[47][48]

For many diseases, pathological analysis of cells and tissues is considered to be the gold standard of disease diagnosis. AI-assisted pathology tools have been developed to assist with the diagnosis of a number of diseases, including breast cancer, hepatitis B, gastric cancer, and colorectal cancer. AI has also been used to predict genetic mutations and prognosticate disease outcomes.[40] AI is well-suited for use in low-complexity pathological analysis of large-scale screening samples, such as colorectal or breast cancer screening, thus lessening the burden on pathologists and allowing for faster turnaround of sample analysis.[49] Several deep learning and artificial neural network models have shown accuracy similar to that of human pathologists,[49] and a study of deep learning assistance in diagnosing metastatic breast cancer in lymph nodes showed that the accuracy of humans with the assistance of a deep learning program was higher than either the humans alone or the AI program alone.[50] Additionally, implementation of digital pathology is predicted to save over $12 million for a university center over the course of five years,[51] though savings attributed to AI specifically have not yet been widely researched. The use of augmented and virtual reality could prove to be a stepping stone to wider implementation of AI-assisted pathology, as they can highlight areas of concern on a pathology sample and present them in real-time to a pathologist for more efficient review.[49] AI also has the potential to identify histological findings at levels beyond what the human eye can see,[49] and has shown the ability to utilize genotypic and phenotypic data to more accurately detect the tumor of origin for metastatic cancer.[52] One of the major current barriers to widespread implementation of AI-assisted pathology tools is the lack of prospective, randomized, multi-center controlled trials in determining the true clinical utility of AI for pathologists and patients, highlighting a current area of need in AI and healthcare research.[49]

Primary care has become one key development area for AI technologies.[53][54] AI in primary care has been used for supporting decision making, predictive modelling, and business analytics.[55] Despite the rapid advances in AI technologies, general practitioners' view on the role of AI in primary care is very limited–mainly focused on administrative and routine documentation tasks.[54][56] There are only few examples of AI decision support systems that were prospectively assessed on clinical efficacy when used in practice by physicians. But there are cases where the use of these systems yielded a positive effect on treatment choice by physicians.[57]

In psychiatry, AI applications are still in a phase of proof-of-concept.[58] Areas where the evidence is widening quickly include predictive modelling of diagnosis and treatment outcomes,[59] chatbots, conversational agents that imitate human behaviour and which have been studied for anxiety and depression.[60]

Challenges include the fact that many applications in the field are developed and proposed by private corporations, such as the screening for suicidal ideation implemented by Facebook in 2017.[61] Such applications outside the healthcare system raise various professional, ethical and regulatory questions.[62] Another issue is often with the validity and interpretabiltiy of the models. Small training datasets contain bias that is inherited by the models, and compromises the generalizability and stability of these models. Such models may also have the potential to be discriminatory against minority groups that are underrepresented in samples.[63]

AI is being studied within the field of radiology to detect and diagnose diseases through Computerized Tomography (CT) and Magnetic Resonance (MR) Imaging.[64] It may be particularly useful in settings where demand for human expertise exceeds supply, or where data is too complex to be efficiently interpreted by human readers.[65] Several deep learning models have shown the capability to be roughly as accurate as healthcare professionals in identifying diseases through medical imaging, though few of the studies reporting these findings have been externally validated.[66] AI can also provide non-interpretive benefit to radiologists, such as reducing noise in images, creating high-quality images from lower doses of radiation, enhancing MR image quality, and automatically assessing image quality.[67] Further research investigating the use of AI in nuclear medicine focuses on image reconstruction, anatomical landmarking, and the enablement of lower doses in imaging studies.[68]

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 include using "Artificial Neural Networks (ANN) and Bayesian Networks (BN)". It was found that ANN was better and could more accurately classify diabetes and CVD.

Through the use of Medical Learning Classifiers (MLC's), Artificial Intelligence has been able to substantially aid doctors in patient diagnosis through the manipulation of mass Electronic Health Records (EHR's).[69] Medical conditions have grown more complex, and with a vast history of electronic medical records building, the likelihood of case duplication is high.[69] 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.[69] The implementation of AI to not only help find similar cases and treatments, 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.[69]

An elderly man using a pulse oximeter to measure his blood oxygen levels

The increase of telemedicine, the treatment of patients remotely, has shown the rise of possible AI applications.[70] AI can assist in caring for patients remotely by monitoring their information through sensors.[71] A wearable device may allow for constant monitoring of a patient and the ability to notice changes that may be less distinguishable by humans. The information can be compared to other data that has already been collected using artificial intelligence algorithms that alert physicians if there are any issues to be aware of.[71]

Another application of artificial intelligence is chat-bot therapy. Some researchers charge that the reliance on chatbots for mental healthcare does not offer the reciprocity and accountability of care that should exist in the relationship between the consumer of mental healthcare and the care provider (be it a chat-bot or psychologist), though.[72]

Since the average age has risen due to a longer life expectancy, artificial intelligence could be useful in helping take care of older populations.[73] Tools such as environment and personal sensors can identify a person's regular activities and alert a caretaker if a behavior or a measured vital is abnormal.[73] Although the technology is useful, there are also discussions about limitations of monitoring in order to respect a person's privacy since there are technologies that are designed to map out home layouts and detect human interactions.[73]

Electronic health records (EHR) are crucial to the digitalization and information spread of the healthcare industry. Now that around 80% of medical practices use EHR, the next step is to use artificial intelligence to interpret the records and provide new information to physicians.[74]

One application uses natural language processing (NLP) to make more succinct reports that limit the variation between medical terms by matching similar medical terms.[74] For example, the term heart attack and myocardial infarction mean the same things, but physicians may use one over the over based on personal preferences.[74] NLP algorithms consolidate these differences so that larger datasets can be analyzed.[74] Another use of NLP identifies phrases that are redundant due to repetition in a physician's notes and keeps the relevant information to make it easier to read.[74] Other applications use concept processing to analyze the information entered by the current patient's doctor to present similar cases and help the physician remember to include all relevant details.[75]

Beyond making content edits to an EHR, there are AI algorithms that evaluate an individual patient's record and predict a risk for a disease based on their previous information and family history.[76] One general algorithm is a rule-based system that makes decisions similarly to how humans use flow charts.[77] This system takes in large amounts of data and creates a set of rules that connect specific observations to concluded diagnoses.[77] Thus, the algorithm can take in a new patient's data and try to predict the likeliness that they will have a certain condition or disease.[77] Since the algorithms can evaluate a patient's information based on collective data, they can find any outstanding issues to bring to a physician's attention and save time.[76] One study conducted by the Centerstone research institute found that predictive modeling of EHR data has achieved 70–72% accuracy in predicting individualized treatment response.[78] These methods are helpful due to the fact that the amount of online health records doubles every five years.[76] Physicians do not have the bandwidth to process all this data manually, and AI can leverage this data to assist physicians in treating their patients.[76]

Improvements in natural language processing led to the development of algorithms to identify drug-drug interactions in medical literature.[79][80][81][82] Drug-drug interactions pose a threat to those taking multiple medications simultaneously, and the danger increases with the number of medications being taken.[83] To address the difficulty of tracking all known or suspected drug-drug interactions, machine learning algorithms have been created to extract information on interacting drugs and their possible effects from medical literature. Efforts were consolidated in 2013 in the DDIExtraction Challenge, in which a team of researchers at Carlos III University assembled a corpus of literature on drug-drug interactions to form a standardized test for such algorithms.[84] Competitors were tested on their ability to accurately determine, from the text, which drugs were shown to interact and what the characteristics of their interactions were.[85] Researchers continue to use this corpus to standardize the measurement of the effectiveness of their algorithms.[79][80][82]

Other algorithms identify drug-drug interactions from patterns in user-generated content, especially electronic health records and/or adverse event reports.[80][81] Organizations such as the FDA Adverse Event Reporting System (FAERS) and the World Health Organization's VigiBase allow doctors to submit reports of possible negative reactions to medications. Deep learning algorithms have been developed to parse these reports and detect patterns that imply drug-drug interactions.[86]

In order to effectively train Machine Learning and use AI in healthcare, massive amounts of data must be gathered.[92] 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.[90] The scarcity of real, accessible patient data is a hindrance that deters the progress of developing and deploying more artificial intelligence in healthcare.

According to a recent study, AI can replace up to 35% of jobs in the UK within the next 10 to 20 years.[93] 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.[93]

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.[94] 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

AI will ultimately help contribute to progression of societal goals which include better communication, improved quality of healthcare, and autonomy.[95]

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.[96]

There can also be unintended bias in these algorithms that can exacerbate social and healthcare inequities.[96]  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.[97] 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.[98] 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.[97] 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.[99] However, these biases are able to be eliminated through careful implementation and a methodical collection of representative data.