these stakeholders. These examples are not exhaustive. The following sections explore the promise of AI in health care in more detail.

AI SOLUTIONS FOR PATIENTS AND FAMILIES

AI could soon play an important role in the self-management of chronic diseases such as cardiovascular diseases, diabetes, and depression. Self-management tasks can range from taking medications, modifying diet, and getting more physically active to care management, wound care, device management, and the delivery of injectables. Self-management can be assisted by AI solutions, including conversational agents, health monitoring and risk prediction tools, personalized adaptive interventions, and technologies to address the needs of individuals with disabilities. We describe these solutions in the following sections.

Conversational Agents

Conversational agents can engage in two-way dialogue with the user via speech recognition, natural language processing (NLP), natural language understanding, and natural language generation (Laranjo et al., 2018). AI is behind many of them. These interfaces may include text-based dialogue, spoken language, or both. They are called, variously, virtual agents, chatbots, or chatterbots. Some conversational agents present a human image (e.g., the image of a nurse or a coach) or nonhuman image (e.g., a robot or an animal) to provide a richer interactive experience. These are called embodied conversational agents (ECAs). These visible characters provide a richer and more convincing interactive experience than non-embodied voice-only agents such as Apple’s Siri, Amazon’s Alexa, or Microsoft’s Cortana. The imagistic entities can communicate nonverbally through hand gestures and facial expressions.

In the “self-management” domain, conversational agents already exist to address depression, smoking cessation, asthma, and diabetes. Although many chatbots and ECAs exist, evaluation of these agents has, unfortunately, been limited (Fitzpatrick et al., 2017).

The future potential for conversational agents in self-management seems high. While simulating a real-world interaction, the agent may assess symptoms, report back on outputs from health monitoring, and recommend a course of action based on these varied inputs. Most adults say they would use an intelligent virtual coach or an intelligent virtual nurse to monitor health and symptoms at home. There is somewhat lower enthusiasm for mental health support delivered via this method (Accenture, 2018).

Social support improves treatment outcomes (Hixson et al., 2015; Wicks et al., 2012). Conversational agents can make use of humans’ propensity to

treat computers as social agents. Such support could be useful as a means of combating loneliness and isolation (Stahl and Coeckelbergh, 2016; Wetzel, 2018). In other applications, conversational agents can be used to increase the engagement and effectiveness of interventions for health behavior change. Most studies of digital interventions for health behavior change have included support from either professionals or peers. It is worth noting that professionally supported interventions cost two to three times what technology interventions cost. A conversational agent could provide some social support and increased engagement while remaining scalable and cost-effective. Moreover, studies have shown that people tend to be more honest when interacting with technology than with humans (Borzykowski, 2016).

In the next decade, conversational AI will probably become more widely used as an extender of clinician support or as a stopgap where other options are not available (see Chapter 1). Now under development are new conversational AI strategies to infer emotion from voice analysis, computer vision, and other sources. We think it is likely that systems will thus become more conversant in the emotional domain and more effective in their communication.

Health Monitoring and Risk Prediction

AI can use raw data from accelerometers, gyroscopes, microphones, cameras, and other sensors, including smartphones. Machine learning algorithms can be trained to recognize patterns from the raw data inputs and then categorize these patterns as indicators of an individual’s behavior and health status. These systems can allow patients to understand and manage their own health and symptoms as well as share data with medical providers.

The current acceptance of wearables, smart devices, and mobile health applications has risen sharply. In just a 4-year period, between 2014 and 2018, the proportion of U.S. adults reporting that they use wearables increased from 9 percent to 33 percent. The use of mobile health apps increased from 16 percent to 48 percent (Accenture, 2018). Consumer interest is high (~50 percent) in using data generated by apps, wearables, and IoT devices to predict health risks (Accenture, 2018). Since 2013, AI startup companies with a focus on health care and wearables have raised $4.3 billion to develop, for example, bras designed for breast cancer risk prediction and smart clothing for cardiac, lung, and movement sensing (Wiggers, 2018).

Timely Personalized Interventions

AI-driven adaptive interventions are called JITAIs, or “just-in-time adaptive interventions.” These are learning systems that deliver dynamic, personalized

treatment to users over time (Nahum-Shani et al., 2015; Spruijt-Metz and Nilsen, 2014). The JITAI makes decisions about when and how to intervene based on response to prior intervention, as well as on awareness of current context, whether internal (e.g., mood, anxiety, blood pressure) or external (e.g., location, activity). JITAI assistance is provided when users are most in need of it or will be most receptive to it. These systems can also tell a clinician when a problematic pattern is detected. For example, a JITAI might detect when a user is in a risky situation for substance abuse relapse—and deliver an intervention against it.

These interventions rely on sensors, rather than a user’s self-report, to detect states of vulnerability or intervention opportunity. This addresses two key self-management challenges: the high user burden of self-monitoring and the limitations of self-awareness. As sensors become more ubiquitous in homes, in smartphones, and on bodies, the data sources for JITAIs are likely to continue expanding. AI can be used to allow connected devices to communicate with one another. (Perhaps a glucometer might receive feedback from refrigerators regarding the frequency and types of food consumed.) Leveraging data from multiple inputs can uniquely enhance AI’s ability to provide real-time behavioral management.

Assistance for Individuals with Cognitive Disabilities

According to the Centers for Disease Control and Prevention, 16 million individuals are living with cognitive disability in the United States alone. Age is the single best predictor of cognitive impairments, and an estimated 5 million Americans more than 65 years old have Alzheimer’s disease. These numbers are expected to increase due to the growth of an aging population: currently nearly 9 percent of all adults are more than 65 years old, a percentage expected to double by 2050 (CDC, 2018; Family Caregiver Alliance, 2019). Critically, 15.7 million family members provide unpaid care and support to individuals with Alzheimer’s disease or other dementias (Alzheimers.net, 2019). The current system of care is unprepared to handle the current or future load of patient needs, or to allow individuals to “age in place” at their current homes rather than relocating to assisted living or nursing home facilities (Family Caregiver Alliance, 2019).

Smart home monitoring and robotics may eventually use AI to address these challenges (Rabbitt et al., 2015). Home monitoring has the potential to increase independence and improve aging at home by monitoring physical space, falls, and amount of time in bed. (Excessive time in bed can be both the cause and outcome of depression, and places the elderly at high risk for bedsores, loss of mobility, and increased mortality.) Currently available social robots such as PARO, Kabochan, and PePeRe provide companionship and stimulation for dementia patients. Recently, the use of robotic pets has been reported to be helpful in

reducing agitation in nursing home patients with dementia (Schulman-Marcus et al., 2019; YellRobot, 2018).

For example, PARO is a robot designed to look like a cute, white baby seal that helps calm patients in hospitals and nursing homes. Initial pilot testing of PARO with 30 patient–caregiver dyads showed that PARO improved affect and communication among those dementia patients who interacted with the robot. This benefit was especially seen among those with more cognitive deficits. Larger clinical trials have also demonstrated improvements in patient engagement, although the effects on cognitive symptoms remain ambiguous (Moyle et al., 2017).

Although socially assistive robots are designed primarily for older adult consumers, caregivers also benefit from them because they relieve caregiver burden a bit and thus improve their well-being. As the technology improves, it may be that robots will do increasingly sophisticated tasks. Future applications of robotics are being developed to provide hands-on care. Platforms for smart home monitoring may eventually incorporate caregiver and patient needs in one seamless experience to ensure a family-wide experience rather than individual experiences. Designers of smart home monitoring should consider the ethics of equitable access by designing AI for the elderly, the dependent, and the short- or long-term disabled (Johnson, 2018).

AI SOLUTIONS FOR THE CLINICIAN CARE TEAM

There are two main areas of opportunity for AI in clinical care: (1) enhancing and optimizing care delivery, and (2) improving information management, user experience, and cognitive support in electronic health records (EHRs). Strides have been made in these areas for decades, largely through rule-based, expert-designed applications typically focused on specific clinical areas or problems. AI techniques offer the possibility of improving performance further.

Care Delivery

The amount of relevant data available for patient care is growing and will continue to grow in volume and variety. Data recorded digitally through EHRs only scratch the surface of the types of data that (when appropriately consented) could be leveraged for improving patient care. Clinicians are beginning to have access to data generated from wearable devices, social media, and public health records; to data about consumer spending, grocery purchase nutritional value, and an individual’s exposome; and to the many types of -omic data specific to an individual. AI will probably, we think, have a profound effect on the entire clinical care process, including prevention, early detection, risk/benefit identification, diagnosis, prognosis, and personalized treatment.

been shown to improve the safety of interventions where clinicians are exposed to high doses of ionizing radiation and makes surgery possible in anatomic locations not otherwise reachable by human hands (Shen et al., 2018; Zhao et al., 2018). As autonomous robotic surgery improves, it is likely that surgeons will in some cases oversee the movements of robots (Shademan et al., 2016).

Personalized Management and Treatment

Precision medicine allows clinicians to tailor medical treatment to the individual characteristics of each patient. Clinicians are testing whether AI will permit them to personalize chemotherapy dosing and map patient response to a treatment so as to plan future dosing (Poon et al., 2018). AI-driven NLP has been used to identify polyp descriptions in pathology reports that then trigger guideline-based clinical decision support to help clinicians determine the best surveillance intervals for colonoscopy exams (Imler et al., 2014). Other AI tools have helped clinicians select the best treatment options for complex diseases such as cancer (Zauderer et al., 2014).

The case of clinical equipoise—when clinical practice guidelines do not present a clear preference among care treatment options—also has significant potential for AI. Using retrospective data from other patients, AI techniques can predict treatment responses of different combinations of therapies for an individual patient (Brown, 2018). These types of tools may serve to help select a treatment immediately and may also provide new knowledge to future practice guidelines. Possibly useful will be dashboards demonstrating predicted outcomes along with cost of treatment and expected changes based on patient behavior, such as increased exercise. These may provide an excellent platform for shared decision making involving the patient, family, and clinical team. AI could also support a patient-centered medical homes model (Jackson et al., 2013).

As genome-phenome integration is realized, the use of genetic data in AI systems for diagnosis, clinical care, and treatment planning will probably increase. To truly impact routine care, though, genetic datasets will need to better represent the diversity of patient populations (Hindorff et al., 2018).

AI can also be used to find similar cases from patient records in an EHR to support treatment decisions based on previous outcomes (Schuler et al., 2018).

Improving Information Management, User Experience, and Cognitive Support in EHRs

The following sections describe a few areas that could benefit from AI-supported tools integrated with EHR systems, including information management (e.g., clinical documentation, information retrieval), user experience, and cognitive support.

applied convolutional neural network analytic approaches to quantify associations between the built environment and obesity prevalence. They have shown that physical characteristics of a neighborhood can be associated with variations in obesity prevalence across different neighborhoods (Maharana and Nsoesie, 2018). Shin et al. (2018) applied a machine learning approach that uses both biomarkers and sociomarkers to predict and identify pediatric asthma patients at risk of hospital revisits.

Without knowing specific symptom-related features, the sociomarker-based model correctly predicted two out of three patients at risk. Once identified, population or regions can be targeted with computational health campaigns that blur the distinction between interpersonal and mass influence (Cappella, 2017). However, the risks of machine learning in these contexts have also been described (Cabitza et al., 2017). They include (1) the risk that clinicians become unable to recognize when the algorithms are incorrect, (2) lack of an ability for the algorithms to address the context of care, or (3) the intrinsic lack of reliability of some medical data. However, many of these challenges are not intrinsic to machine learning or AI, but rather represent misuse of the technologies.

Population Health Improvement Through Chronic Disease Management

A spectrum of market-ready AI approaches to support population health programs already exists. They are used in areas of automated retinal screening, clinical decision support, predictive population risk stratification, and patient self-management tools (Contreras and Vehi, 2018; Dankwa-Mullan et al., 2019). Several solutions have received regulatory approval; for example, the U.S. Food and Drug Administration approved Medtronic’s Guardian Connect, marking the first AI-powered continuous glucose monitoring system. Crowd-sourced, real-world data on inhaler use, combined with environmental data, led to a policy recommendations model that can be replicated to address many public health challenges by simultaneously guiding individual, clinical, and policy decisions (Barrett et al., 2018).

There is an alternative approach to standard risk prediction models that applies AI tools. For example, predictive models using machine learning algorithms may facilitate recognition of clinically important unanticipated predictor variables that may not have previously been identified by “traditional” research approaches that rely on statistical methods testing a priori hypotheses (Waljee et al., 2014). Enabled by the availability of data from administrative claims and EHRs, machine learning can enable patient-level prediction, which moves beyond average population effects to consider personalized benefits and risks. Large-scale, patient-level

prediction models from observational health care data are facilitated by a common data model that enables prediction researchers to work across computer environments. An example can be found from the Observational Health Data Sciences and Informatics collaborative, which has adopted the Observational Medical Outcomes Partnership common data model for patient-level prediction models using observational health care data (Reps et al., 2018).

Another advantage of applying AI approaches to predictive models is the ability not only to predict risk but also the presence or absence of a disease in an individual. As an example, successful use of a memetic pattern-based algorithm approach was demonstrated in a broad risk spectrum of patients undergoing coronary artery disease evaluation and was shown to successfully identify and exclude coronary artery disease in a population instead of just predicting the probability of future events (Zellweger et al., 2018). In addition to helping health care organizations identify individuals with elevated risks of developing chronic conditions early in the disease’s progression, this approach may prevent unnecessary diagnostic procedures in patients where procedures may not be warranted and also support better outcomes.

Not all data elements needed to predict chronic disease can be found in administrative records and EHRs. Creating risk scores that include a blend of social, behavioral, and clinical data may help give providers the actionable, 360-degree insight necessary to identify patients in need of proactive, preventive care while meeting reimbursement requirements and improving outcomes (Kasthurirathne et al., 2018).

AI Solutions for Public Health Program Management

Public health professionals are focused on solutions for more efficient and effective administration of programs, policies, and services; disease outbreak detection and surveillance; and research. Relevant AI solutions are being experimented with in a number of areas.

Disease Surveillance

The range of AI solutions that can improve disease surveillance is considerable. For a number of years, researchers have tracked and refined the options for tracking disease outbreaks using search engine query data. Some of these approaches rely on the search terms that users type into Internet search engines (e.g., Google Flu Trends). At the same time, caution is warranted with these approaches. Relying on data not collected for scientific purposes (e.g., Internet search terms) to predict flu outbreaks has been fraught with error (Lazer et al., 2014). Nontransparent search algorithms that change constantly cannot be easily replicated and studied.

These changes may occur due to business needs (rather than the needs of a flu outbreak detection application) or due to changes in search behavior of consumers. Finally, relying on such methods exclusively misses the opportunity to combine them and co-develop them in conjunction with more traditional methods. As Lazer et al. (2014) detail, combining traditional and innovative methods (e.g., Google Flu Trends) performs better than either method alone.

Researchers and solution developers have experimented with the integration of case- and event-based surveillance (e.g., news and online media, sensors, digital traces, mobile devices, social media, microbiological labs, and clinical reporting) to arrive at dashboards and analysis approaches for threat verification. Such approaches have been referred to as digital epidemiological surveillance and can produce timelier data and reduce labor hours of investigation (Kostokova, 2013; Zhao et al., 2015). Such analyses rely on AI’s capacities in spatial and spatiotemporal profiling, environmental monitoring, and signal detection (i.e., from wearable sensors). They have been successfully implemented to build early warning systems for adverse drug events, falls detection, and air pollution (Mooney and Pejaver, 2018). The ability to rely on unstructured data such as photos, physicians’ notes, sensor data, and genomic information, when enabled by AI, may lead to additional, novel approaches in disease surveillance (Figge, 2018).

Moreover, participatory systems such as social media and listservs could be relied on to solicit information from individuals as well as groups in particular geographic locations. For example, such approaches may encourage a reduction in unsafe behaviors that put individuals at risk for human immunodeficiency virus (HIV) infection (Rubens et al., 2014; Young et al., 2017). For example, it has been demonstrated that psychiatric stressors can be detected from Twitter posts in select populations through keyword-based retrieval and filters and the use of neural networks (Du et al., 2018). However, how such AI solutions could improve the health of populations or communities is less clear, due to the lack of context for some tweets and because tweets may not reflect the true underlying mental health status of a person who tweeted. Studies that retroactively analyze the tweeting behavior of individuals with known suicide attempts or ideation, or other mental health conditions, may allow refinement in such approaches.

Finally, AI and machine learning have been used to develop a dashboard to provide live insight into opioid usage trends in Indiana (Bostic, 2018). This tool enabled prediction of drug positivity for small geographic areas (i.e., hot spots), allowing for interventions by public health officials, law enforcement, or program managers in targeted ways. A similar dashboarding approach supported by AI solutions has been used in Colorado to monitor HIV surveillance and outreach interventions and their impact after implementation (Snyder et al., 2016). This tool integrated data on regional resources with near-real-time visualization of

complex information to support program planning, patient management, and resource allocation.

Environmental and Occupational Health

AI has already made inroads into environmental and occupational health by leveraging data generated by sensors, nanotechnology, and robots. For example, water-testing sensors with AI tools have been paired with microscopes to detect bacterial contamination in treatment plants through hourly water sampling and analysis. This significantly reduces the time traditionally spent sending water samples for laboratory testing and lowers the cost of certain automated systems (Leider, 2018).

In a similar fashion, remote sensing from meteorological sensors, combined with geographic information systems, has been used to measure and analyze air pollution patterns in space and over time. This evolving field of inquiry has been termed geospatial AI because it combines innovations in spatial science with the rapid growth of methods in AI, including machine learning and deep learning. In another approach to understanding environmental factors, images of Google Street View have been analyzed using deep learning mechanisms to analyze urban greenness as a predictor and enabler of exercise (e.g., walking, cycling) (Lu, 2018).

Robots enabled by AI technology have been successfully deployed in a variety of hazardous occupational settings to improve worker safety and prevent injuries that can lead to costly medical treatments or short- and long-term disability. Robots can replace human labor in highly dangerous or tedious jobs that are fatiguing and could represent health risks to workers—reducing injuries and fatalities. Similarly, the construction industry has relied on AI-enabled robots for handling hazardous materials, handling heavy loads, working at elevation or in hard-to-reach places, and completing tasks that require difficult work postures, risking injury (Hsiao et al., 2017). Some of these robotic deployments are replacing human labor; in other instances, humans collaborate with robots to carry out such tasks. Hence, the deployment of robots requires workers to develop new skills in directing and managing robots and managing interactions between different types of robots or equipment, all operating in dynamic work environments.

AI SOLUTIONS FOR HEALTH CARE BUSINESS ADMINISTRATORS

Coordination and payment for care in the United States is highly complex. It involves the patient, providers, health care facilities, laboratories, hospitals, pharmacies, benefit administrators, payers, and others. Before, during, and after a patient encounter, administrative coordination occurs around scheduling, billing,

and payment. Collectively, we call this the administration of care, or administrative workflow. AI might be used in this setting in the form of machine learning models that can work alongside administrative personnel to perform mundane, repetitive tasks in a more efficient, accurate, and unbiased fashion.

Newer methods in AI, known generally as deep learning, have advantages over traditional machine learning, especially in the analysis of text data. The use of deep learning is particularly powerful in a workflow where a trained professional reviews narrative data and makes a decision about a clear action plan. A vast number of prior authorizations exist, with decisions and underlying data; these data provide the ideal substrate to train an AI model. Textual information used for prior authorization can be used for training a deep learning model that reaches or even exceeds the ability of human reviewers, perhaps with even more consistency than human reviewers. AI for health care administration will likely be utilized extensively, even beyond those solutions deployed for direct clinical care. “While all AI solutions give some false positives and false negatives, in administration, these will mostly produce annoyances, but should be monitored closely to ensure that patient safety is never impacted.”

As AI solutions become more sophisticated and automated, it may happen that a variety of AI methodologies will be deployed for a given solution. For example, a request by a patient to refill a prescription might involve speech recognition or AI chatbots, a rules-based system to determine if prior authorization is required, automated provider outreach, and a deep learning system for prior authorization when needed. It is worth noting that deep learning systems already drive many of today’s speech recognition, translation, and chatbot programs.

We provide illustrative (non-exhaustive) examples in Table 3-2 for different types of applications of AI solutions to support routine health care administration processes.

Prior Authorization

Most health plans and pharmacy benefit managers require prior authorization of devices, durable equipment, labs, and procedures. The process includes the submission of patient information along with the proposed request, along with justification. Determinations require professional skill, analysis, and judgment. Automating this process can reduce biased decisions and improve speed, consistency, and quality of decisions.

There are a number of different ways that AI is applied today. For example, AI could simply be used to sort cases to the appropriate level of reviewer (e.g., nurse practitioner, physician advisor, medical director). Or, AI could identify and highlight the specific, relevant information in long documents or narratives to

produce information regarding estimated costs or benefit/risk assessment to aid a consumer in a decision.

Automation of prior authorization could reduce administrative costs, frustration, and idle time for provider and payer alike. Ultimately, such a process could lead to fewer appeals as well, which is a costly outcome of any prior authorization decision. A prior authorization model would need to work in near real time, because the required decisions are typically time sensitive.

data comprehensiveness, public acceptance, algorithm accuracy, and appropriate regulatory frameworks.

AI Solutions for Research and Development Professionals

The uses of AI technologies in research are broad; they frequently drive the development of new machine learning algorithms. To narrow this massive landscape, we focus our discussion on research institutions with medical training facilities. These medical school–affiliated health care providers often house massive and multiple large-scale data repositories (such as biobanks, Digital Imaging and Communications in Medicine or DICOM systems, and EHR systems).

Mining EHR Data

Research with EHR data offers promising opportunities to advance biomedical research and improve health care by interpreting structured, unstructured (e.g., free text), genomic, and imaging data.

Machine Learning

Extracting practical information from EHR data is challenging because such data are highly dimensional, heterogeneous, sparse, and often of low quality (Jensen et al., 2012; Zhao et al., 2017). Nevertheless, AI technologies are being applied to EHR data. AI techniques used on these data include a vast array of data mining approaches, from clustering and association rules to deep learning. We focus our discussion in the sections below on areas of present key importance.

Deep Learning

Deep learning algorithms rely on the large quantities of data and massive computer resources, both of which are newly possible in this era. Deep learning can identify underlying patterns in data well beyond the pattern-perceiving capacities of humans. Deep learning and its associated techniques have become popular in many data-driven fields of research. The principal difference between deep and traditional (i.e., shallow) machine learning paradigms lies in the ability of deep learning algorithms to construct latent data representations from a large number of raw features, often through deep architectures (i.e., many layers) of artificial neural networks. This “unsupervised feature extraction” sometimes permits highly accurate predictions. Recent research on EHR data has shown that deep learning predictive models can outperform traditional clinically used predictive

models for predicting early detection of heart failure onset (Choi et al., 2017), various cancers (Miotto et al., 2016), and onset of and weaning from intensive care unit interventions (Suresh et al., 2017).

Nevertheless, the capacity of deep learning is a double-edged sword. The downside of deep learning comes from exactly where its superiority to other learning paradigms originates—that is, its ability to build and learn features. Model complexity means that human interpretability of deep learning models is almost nonexistent, because it is extremely hard to infer how the model makes its predictions so well. Deep learning models are “black box” models, where the internal workings of the algorithms remain unclear or mysterious to users of these models. As in other black box AI approaches, there is significant resistance to implementing deep learning models in the health care delivery process.

Applications to Imaging Data

Detecting abnormal brain structure is much more challenging for humans and machines than detecting a broken bone or a fracture. Exciting things are being done in this area with deep learning. One recent study predicted age from brain images (Cole et al., 2017). Multimodal image recognition analysis has discovered novel impairments not visible from a single view of the brain (e.g., structural MRI versus functional MRI) (Plis et al., 2018). Companies such as Avalon AI¹ are commercializing this type of work.

Effective AI use does not always require new modeling techniques. Some work at Massachusetts General Hospital in Boston uses a large selection of images and combines established machine learning techniques with mature brain-image analysis tools to explore what is normal for a child’s developing brain (NITRC, 2019; Ou et al., 2017). Other recent applications of AI to radiology data include using machine learning on electrocardiogram data to characterize types of heart failure (Sanchez-Martinez et al., 2018). In addition, AI can aid in reducing noise in real images (e.g., endoscopy) via “adversarial training.” It can smooth out erroneous signals in images to enhance prediction accuracy (Mahmood et al., 2018). AI is also being applied to moving images; gait analysis has long been done by human observation alone, but it now can be performed with greater accuracy by AI that uses video and sensor data. These techniques are being used to detect Parkinson’s disease, to improve geriatric care, for sports rehabilitation, and in other areas (Prakash et al., 2018). AI can also improve video-assisted surgery, for example, by detecting colon polyps in real time (Urban et al., 2018)

___________________

¹ See http://avalonai.mystrikingly.com.

datasets to train supervised algorithms requires involvement from domain experts, which hampers generalizability and scalability of phenotyping algorithms. As a result, the classification algorithms for phenotyping research has been limited so far to regularized algorithms that can address overfitting, which is what happens when an algorithm that uses many features is trained on small training sets.

Regularized classifiers penalize more features in favor of model parsimony. To overcome this limitation, new research is investigating the application of hybrid approaches (known as semi-supervised) to create semi-automatically annotated training sets and the use of unsupervised methods to scale up EHR phenotyping. The massive amounts of data in EHRs, if processed through deep neural networks, may soon permit the computation of phenotypes from a wider vector of features.

As mentioned previously, large national initiatives are now combining biobanks of genomic information with these phenotypes. For example, eMerge frequently uses genomic analyses such as genome-wide association studies in combination with phenotyping algorithms to define the gold-standard cohort or to study genetic risk factors in the phenotyped population (Gottesman et al., 2013; McCarty et al., 2011).

Drug Discovery

Machine learning has the capacity to make drug discovery faster, cheaper, and more effective. Drug designers frequently apply machine learning techniques to extract chemical information from large compound databases and to design drugs with important biological properties. Machine learning can also improve drug discovery by permitting a more comprehensive assessment of cellular systems and potential drug effects. With the emergence of large chemical datasets in recent years, machine and deep learning methods have been used in many areas (Baskin et al., 2016; Chen et al., 2018; Lima et al., 2016; Zitnik et al., 2018). These include

• predicting synthesis,
• biological activity of new ligands,
• drug selectivity,
• pharmacokinetic and toxicological profiles,
• modeling polypharmacy side effects (due to drug–drug interactions), and
• designing de novo molecular structures and structure-activity models.

Large chemical databases have made drug discovery faster and cheaper. EHR databases have brought millions of patients’ lives into the universe of statistical learning. Research initiatives to link structured patient data with biobanks,

radiology images, and notes are creating a rich and robust analytical playground for discovering new knowledge about human disease. Deep learning and other new techniques are creating solutions that can operate on the scale required to digest these multiterabyte datasets. The accelerating pace of discovery will probably challenge the research pipelines that translate new knowledge back into practice.

KEY CONSIDERATIONS

Although it is difficult to predict the future in a field that is changing so quickly, we offer the following ideas about how AI will be used and considerations for optimizing the success of AI for health.

Augmented Intelligence

AI will change health care delivery less by replacing clinicians than by supporting or augmenting clinicians in their work. AI will support clinicians with less training in performing tasks currently relegated to specialists. It filters out normal or noncomplex cases so that specialists can focus on a more challenging case load.

AI will support humans in tasks that suffer from inattention, cause fatigue, and are physically difficult to perform. AI will substitute for humans by facilitating screening and evaluation in areas with limited access to medical expertise. Some AI tools, like those for self-management or population health support, will be useful in spite of lower accuracy.

When an AI tool assists human cognition, it will initially need to explain the connections it has drawn, allowing for an understanding of a pathway to effects. With sufficient accuracy, humans will begin to trust the AI output and will require less transparency and explanation. In situations where AI substitutes for medical expertise, the workflow should include a human in the loop to identify misbehavior and provide accountability (Rahwan, 2018).

Partnerships

The central focus of health care will continue to expand from health care delivery systems to a dispersed model that aggregates information about behavior, traits, and environment in addition to medical symptoms and test results.

Market forces and privacy concerns or regulations may impede data sharing and analysis (Roski et al., 2014). Stakeholders will need to creatively balance

competing demands. More national-level investments similar to the National Institutes of Health’s All of Us program can facilitate and accelerate these partnerships. More ethical and legal guidelines are needed for successful data sharing and analysis (see Chapters 1 and 7).

Interoperability

AI tools will continue to be developed by industry, research, government, and individuals. With emerging standards such as SMART on FHIR (Substitutable Medical Apps, Reusable Technology on Fast Healthcare Interoperability Resource), these tools will increasingly be implemented across platforms regardless of the EHR vendor, brand of phone, etc. This will most likely speed the adoption of AI.

Clinical Practice Guidelines

AI will probably discover associations that have not yet been detected by humans and make predictions that differ from prevailing knowledge and expertise. As a result, some currently accepted practices may be abandoned, and best practice guidelines will be adjusted.

If the output of AI systems is going to influence international guidelines, developers of the applications will require fuller and more representative datasets for training and testing.

Clinical Evidence and Rate of Innovation

The dissemination of innovation will occur rapidly, which on the one hand may advance the adoption of new scientific knowledge but on the other may encourage the rushed adoption of innovation without sufficient evidence.

Bias and Trust

As AI increasingly infiltrates the field of health, biases inherent in clinical practice will appear in the datasets used in AI models. The discovery of existing bias will open the door to changing practices, but it may also produce public disillusionment and mistrust.

Important growth areas for AI include platforms designed for and accessible by the people most in need of additional support. This includes older adults, people living with multiple comorbid conditions, and people in low-resource settings.

the workplace. https://blogs.cdc.gov/niosh-science-blog/2017/12/05/robot_safety_conf (accessed December 13, 2019).

Huang, J., C. Osorio, and L. W. Sy. 2019. An empirical evaluation of deep learning for ICD-9 code assignment using MIMIC-III clinical notes. Computer Methods and Programs in Biomedicine 177:141–153.

Imler, T. D., J. Morea, T. F. Imperiale. 2014. Clinical decisions support with natural language processing facilitates determination of colonscopy surveillance intervals. Clinical Gastroenterology and Hepatology 12(7):1130–1136.

Ingold, D., and S. Soper. 2016. Amazon doesn’t consider the race of its customers. Should it? Bloomberg. https://www.bloomberg.com/graphics/2016-amazon-same-day (accessed November 12, 2019).

Jackson, G. L., B. J. Powers, R. Chatterjee, J. P. Bettger, A. R. Kemper, V. Hasselblad, R. J. Dolor, J. Irvine, B. L. Heidenfelder, A. S. Kendrick, R. Gray, and J. W. Williams. 2013. The patient-centered medical home: A systematic review. Annals of Internal Medicine 158:169–178.

Jensen, P. B., L. J. Jensen, and S. Brunak. 2012. Mining electronic health records: Towards better research applications and clinical care. Nature Reviews Genetics 13(6):395–405.

Jiang, S., K. S. Chin, and K. L. Tsui. 2018. A universal deep learning approach for modeling the flow of patients under different severities. Computer Methods and Programs in Biomedicine 154:191–203.

Joffe, E., O. Havakuk, J. R. Herskovic, V. L. Patel, and E. V. Bernstam. 2012. Collaborative knowledge acquisition for the design of context-aware alert systems. Journal of the American Medical Informatics Association 219(6):988–994.

Johnson, J. 2018. Designing technology for an aging population. Presentation at Stanford Center on Longevity meeting at Tresidder Oak Lounge. http://longevity.stanford.edu/2018/10/17/designing-technology-for-an-aging-population (accessed December 7, 2019).

Kasthurirathne, S. N., J. R. Vest, N. Menachemi, P. K. Halverson, and S. J. Grannis. 2018. Assessing the capacity of social determinants of health data to augment predictive models identifying patients in need of wraparound social services. Journal of the American Medical Informatics Association 25(1):47–53.

Kindig, D., and G. Stoddart. 2003. What is population health? American Journal of Public Health 93(3):380–383.

Klann, J. G., P. Szolovits, S. M. Downs, and G. Schadow. 2014. Decision support from local data: Creating adaptive order menus from past clinician behavior. Journal of Biomedical Informatics 48:84–93.

Kostokova, P. 2013. A roadmap to integrated digital public health surveillance: The vision and the challenges. In WWW ‘13 Companion Proceedings of the

22nd International Conference on World Wide Web. New York: CMS. Pp. 687–694. https://www.researchgate.net/publication/250963354_A_roadmap_to_integrated_digital_public_health_surveillance_The_vision_and_the_challenges (accessed November 12, 2019).

Laranjo, L., A. G. Dunn, H. L. Tong, A. B. Kocaballi, J. Chen, R. Bashir, D. Surian, B. Gallego, F. Magrabi, and A. Coiera. 2018. Conversational agents in healthcare: A systematic review. Journal of the American Medical Informatics Association 25(9):1248–1258.

Lazer, D., R. Kennedy, G. King, and A. Vespignani. 2014. The parable of Google flu: Traps in big data analysis. Science 343(6176):1203–1205.

Leider, N. 2018. AI could protect public health by monitoring water treatment systems. AI in Healthcare News. https://www.aiin.healthcare/topics/artificial-intelligence/ai-public-health-monitoring-water-treatment (accessed November 12, 2019).

Li, M., Z. Fei, M. Zeng, F. Wu, Y. Li, Y. Pan, and J. Wang. 2018. Automated ICD-9 coding via a deep learning approach. IEEE/ACM Transactions on Computational Biology and Bioinformatics 16(4):1193–1202. https://doi.org/10.1109/TCBB.2018.2817488.

Lima, A. N., E. A. Philot, G. H. G. Trossini, L. P. B. Scott, V. G. Maltarollo, and K. M. Honorio. 2016. Use of machine learning approaches for novel drug discovery. Expert Opinion on Drug Discovery 11(3):225–239.

Liu, F., Z. Zhou, A. Samsonov, D. Blankenbaker, W. Larison, A. Kanarek, K. Lian, S. Kambhampati, and R. Kijowski. 2018. Deep learning approach for evaluating knee MR images: Achieving high diagnostic performance for cartilage lesion detection. Radiology 289(1):160–169.

Lu, Y. 2018. The association of urban greenness and walking behavior: Using Google Street View and deep learning techniques to estimate residents’ exposure to urban greenness. International Journal of Environmental Health Research and Public Health 15:1576.

Maharana, A., and E. O. Nsoesie. 2018. Use of deep learning to examine the association of the built environment with prevalence of neighborhood adult obesity. JAMA Network Open 1(4):e181535. https://doi.org/10.1001/jamanetworkopen.2018.1535.

Mahmood, F., R. Chen, and N. J. Durr. 2018. Unsupervised reverse domain adaptation for synthetic medical images via adversarial training. IEEE Transactions on Medical Imaging 37(12):2572–2581. https://doi.org/10.1109/TMI.2018.2842767.

Matheson, R. 2018. Machine-learning system determines the fewest, smallest doses that could still shrink brain tumors. MIT News. http://news.mit.edu/2018/artificial-intelligence-model-learns-patient-data-cancer-treatment-less-toxic-0810.

Suggested citation for Chapter 3: Roski, J., W. Chapman, J. Heffner, R. Trivedi, G. Del Fiol, R. Kukafka, P. Bleicher, H. Estiri, J. Klann, and J. Pierce. 2020. How artificial intelligence is changing health and health care. In Artificial intelligence in health care: The hope, the hype, the promise, the peril. Washington, DC: National Academy of Medicine.