Challenges to Healthcare

Here are some of the current major challenges in healthcare that new computer technologies are poised to help address:

• Continued rising costs of healthcare must be slowed to make healthcare more accessible, in part by significantly reducing manual processes with technology automation.

• Healthcare data must be increasingly digitized to eliminate many paper-based processes and better address patient and clinician needs.

• Interoperability must improve to integrate new sources of data to enable the use of new technologies.

• Value-based care must be expanded to reward clinicians for keeping patients healthy rather than treating them after a diagnosis.

• The patient experience must be transformed to be closer aligned to that of a retail consumer experience, with price and quality transparency, modern and responsive customer service, and convenient at-home options for monitoring and treating ongoing healthcare concerns.

• Healthcare data analysis must shift from a historical or retrospective approach to one that is near real time, prospective, and predictive to help address issues before they become chronic conditions.

Technology for Transformation

Conversational artificial intelligence, in which humans and AI systems conduct conversations using everyday language, has improved tremendously in recent years, and is expected to become the primary and preferred way that people interact with many future technologies. This kind of interface will be at the center of many healthcare applications, device apps, and systems.

AI includes many subfields, such as natural language processing (NLP), machine learning (ML), deep learning (DL), and knowledge representation (KR). Figure 1 illustrates the relationships between various technologies (e.g., cloud, RPA, graph, IoT and blockchain) with AI as the engine, playing a role in the transformation of healthcare.

Figure 1. The many collaborative technologies at play for healthcare

AI will play a significant role in solving some of the challenges in healthcare and in driving a future vision of the patient, provider, and clinician experience. While this report organizes the topic of AI into different technological concepts, they are not distinct sets—there is significant overlap.

The cloud will be the dominant way in which healthcare organizations deliver value to internal and external customers. Virtual assistants and chats will become more useful as NLP feeds text-to-speech and speech-to-text interactions to begin the pathway toward conversational AI.

Technologies like blockchain are set to usher in a new era of cooperation among competing entities toward a common goal. For example, inaccurate doctor directories are currently a barrier to accessing healthcare because each insurer, payer, and health plan maintains its own directories, often with outdated content. If they all used a shared and trusted source of information through the use of a blockchain, however, a single accurate provider directory could significantly reduce the friction in the healthcare system across many stakeholders.

Robotic process automation (RPA) technology and chats will improve the customer experience. RPA software using software bots can mimic human behavior by navigating existing systems the same way as an employee.

The Internet of Things (IoT), using sensors and other devices to monitor health and enhance the patient’s experience, is on the rise with the healthcare system. IoT devices will be increasingly deployed and managed in combination with ML and DL models, giving rise to so-called edge computing, where computer technologies are deployed at the “edge,” such as homes, clinics, or hospitals. For example, a nurse engaging with a patient at home may use a mobile device that transparently uses NLP and ML/DL for quality note taking, enabling the nurse to maintain focus on the patient. Noninvasive sensors used to track movements or detect falls can be installed in homes and set to alert family and/or caregivers, helping to improve patient care.

Business capabilities such as benefit inquiry or ML-based disease prediction models can be built as microservices—Lego-like building blocks that get invoked or consumed as services by other systems. Graph databases are used for real-time queries and rapid flexible changes to the organization of the data. The ability to combine multiple data dimensions with graphs is ideal for many healthcare queries, such as a patient journey, which allows providers or payers (or anyone authorized) to see the steps a patient makes while interacting with the healthcare system.

Next we dive deeper into the application of these emerging technologies for healthcare.

What It Is

There’s no consensus, but most would agree that artificial intelligence is a field of computer science heavily focused on building intelligent things: systems, apps, applications, and more. AI is or will be deployed in a variety of settings, including cars, homes, offices, factories, and hospitals.

As mentioned earlier, AI contains subfields including deep learning, machine learning, natural language understanding, natural language processing, and computer vision. It includes a lot of distinct classes of ML systems such as generative adversarial networks (GANs) and reinforcement learning. This report only covers a few classes of machine learning systems.

How It Solves Healthcare Challenges

AI can and will solve healthcare challenges first and foremost as an aid or tool for clinicians. Adoption of AI improves decision making, minimizes diagnostic errors, improves recommended care pathways, provides a holistic view of a patient, reduces friction and delays in the healthcare system, and, most importantly, enables clinicians to spend more time with patients. AI eliminates some back-office processes while accelerating others, and creates more meaningful, frictionless interactions among stakeholders in the healthcare ecosystem. AI can be used to analyze orders of magnitude more data than ever before, such as data coming from a variety of new sources like genomics, wearables, and intelligent things.

Deep learning is making a huge difference by improving results and outcomes when applying computer vision, natural language processing, natural language understanding and machine learning. DL networks have driven breakthroughs in self-driving cars and image recognition. Deep neural networks provide optimal results for computer vision and image classification, such as helping clinicians quickly identify patients with pneumonia. Studies have found that providers can diagnose conditions significantly faster using tools in which AI-embedded image recognition is used, enabling treatment to start sooner, which is of course vital for severely ill patients.

The ImageNet competition prompted a flurry of activity in image recognition and spurred innovation in ML and computer vision, leading to an AI renaissance. Healthcare is now having its own ImageNet moment for NLP. A few healthcare examples include advances in deep learning NLP to redefine manual work, medical billing and coding, and reviewing bills for accuracy. AI can scan enormous quantities of bills and find—and in some cases, correct—errors, which is an improvement over the current approach of sampling a small percentage of bills for inaccuracies. Deep learning neural networks “learn” from large and complex sets of data, not people, to correctly solve a wide variety of business problems.

Deep learning is well suited for healthcare because complex decisions often require skilled professionals like physicians and nurses. AI is being trained on previous decisions made by these professionals about policy changes, prior authorizations, and chart abstraction, and deep learning neural networks are helping to make new decisions. In all of these examples, AI helps the clinician be more productive.

Use Cases

Another way to frame AI is to say that it enables machines to make judgments and infer meaning about data. Imagine all the areas in healthcare today in which humans need to make judgments, infer meaning, and make decisions. Here are just a few:

• Provider network management

• Prior authorization

• Claims processing

• Provider payments

• Clinical decision support

The goal is to create augmentation tools to improve clinician productivity, not to replace them. The rest of this section looks at specific cases in a little more detail.

Where It’s Headed

There is no doubt that computers can do some things better than people, as long as that thing is well defined, clearly explainable, and reproducible. Machines outperform people for any task where computation makes a difference. AI will be leveraged across the entire $3.5 trillion US healthcare system to keep people well, predict and diagnose conditions, recommend treatment paths, reduce friction in the ecosystem, improve and eliminate manual processes, and more. It may even perform, or help perform, surgery.

Usage of AI can help reduce the almost $500 billion spent every year on billing and insurance-related (BIR) costs. Automation using AI can help standardize and optimize care protocols and move the system toward preventative versus reactive treatment. Using AI with the help of clinicians and researchers may help identify potential cures for various forms of disease.

Imagine a future in which AI diagnostic tools are available for physicians and consumers. AI-powered diagnostic tools could use speech, biometrics, keyboard dynamics, touchscreen patterns, face recognition or visual cues to detect behavioral disorders. Real-time monitoring of health issues becomes more common as wearable become ubiquitous.

Deep learning and natural language processing are subfields of AI. Deep learning, the topic of the next section, is a big factor in the emerging success of AI in healthcare.

What It Is

How It Solves Healthcare Challenges

The most well-publicized applications of DL in healthcare have been in classifying medical images. There are lots of examples of medical images that have been previously evaluated by clinicians (e.g., radiologists and dermatologists) to use to train DL neural networks. Initial seminal research studies focused on training a DL solution to recognize cancerous skin lesions.

Other types of medical diagnostic data are also well suited for classification with a DL model to assist clinical diagnosis. EKGs, MRIs, and genetic sequences have all been successfully classified by DL neural networks. These breakthroughs are likely to transform clinical care going forward.

Administrative use cases in healthcare, such as confirmation of benefits and reimbursement, are also well suited for automation with DL models. Other examples of administrative functions include prior authorization, third-party physician reviews, revenue cycle management, claims processing and adjudication, and risk adjustment.

These are complex processes and evaluations that ensure the appropriate reimbursement of care. The healthcare industry has developed very large sets of rules or guidelines in an attempt to automate processing. But these rules have grown substantially as the complexity of evaluating care and reimbursement has increased. Skilled professionals are relied upon to manually complete complex transactions and care as the rules break down. Physicians, nurses, and medical coders are all counted on to review and determine the appropriate level of care.

Records of claims and charts reviewed by skilled professionals, and the decisions they have made, represent valuable training data for DL neural networks. As policies, medical innovation, and regulations change, DL neural networks will need to be retrained. By training DL neural networks with the decisions of skilled professionals, the industry can build substantial, up-to-date, and collective knowledge and understanding.

Use Cases

This section covers two use cases involving deep learning.

Where It’s Headed

As noted, DL models have driven amazing breakthroughs in the use and adoption of artificial intelligence. With each breakthrough, the question arises: when will DL automate clinical diagnosis and treatment? The barriers and risks to automation in these areas are numerous. One of the biggest challenges is explaining why the model recommended a particular diagnosis or treatment.

The question of why is probably essential to ensure that appropriate care is recommended. Physicians will likely need to understand how the model arrives at a recommendation in explicit terms. Was it blood pressure, genetic sequence, or familial history that led an AI solution to determine a diagnosis and recommend treatment? Often this is referred to as causality—something not inherent to DL models. For example, you can’t point to the specific pixel or set of pixels in the image of a cat to say why the model classified it as such.

Some techniques, as seen in the first case study, can illuminate key parts of the input that may explain how the inputs are impacting the classification, but that isn’t true causality. Work is under way to begin enabling the creation of causal models. Invariants, causal graphs, and deep probabilistic programming are all promising efforts intended to address these limitations. These and other breakthroughs will be essential to building true causal models that would be one key step toward automating clinical diagnosis and treatment.

What It Is

NLP is an interdisciplinary field of computer science and linguistics, and also a subfield of artificial intelligence, as Figure 2 shows. NLP enables a computer to read unstructured data, analyze keywords and phrases, extract meaning from them based on experience, generate language that people use, and create actionable insights. ML and DL models have become critical components of many NLP systems.

Figure 2. How NLP fits into artificial intelligence

Since their advent in 2012, DL models have made significant advances in the fields of image processing and speech recognition, as discussed in the section “Deep Learning”. However, progress in the area of language understanding was relatively slow. Then, in 2018, one of the most path-breaking developments for NLP occurred: The creation of Bidirectional Encoder Representations from Transformers (BERT).

BERT, and models like it, can represent words and sentences in a way that captures underlying meanings and relationships better than past models. In NLP, word embedding models are used to generate vector representations of words to use in a neural network model. Traditional word embedding models like word2vec generate a single context-free representation for each word in a document’s vocabulary (called a word embedding). Word embeddings are context-free because only one representation is allowed per word regardless of the number of senses the word can have. But models like BERT can represent a word based on both the context before and after the word, and is therefore bidirectional. The embeddings these models generate are described as context-aware. BERT and similar language models have shown exceptional results in NLP.

Models like BERT facilitate the use of transfer learning techniques in NLP. Transfer learning allows for models trained on larger datasets to be used on tasks that have smaller datasets in a related domain. DL techniques are data hungry and therefore don’t perform well on smaller datasets. BERT-like models (also called pretrained language models) can be trained on much larger, unlabeled datasets, which can then be fine-tuned on smaller datasets for specific tasks like document classification, question answering, and machine translation.

BERT is an instance of a neural network architecture called a transformer. Transformers are better than other DL models like recurrent neural networks (RNNs) at dealing with long-range dependencies (where words that are not close to each other in a text should be used together to extract the text’s meaning), which are very critical to a few NLP applications like question answering.

Generative Pretrained Transformer (GPT) was the first transformer-based pretrained language model, where the model predicts the next word based on the prior context. BERT is the first bidirectional language model, but a surge of other transformer models for NLP have appeared in the past year, including XLNet and RoBERa, each having more parameters than the one before. Nvidia’s Megatron includes 8 billion parameters and is the largest transformer model as of this writing. These larger models provide better accuracy but take a long time to train, require more computational power, and take longer to run at processing time. Therefore, there has also been an influx of smaller transformer models, like ALBERT and DistillBERT. Though these models are much smaller, they retain much of the same computational accuracy. For example, DistillBERT is 40% smaller than BERT, but retains almost 97% of accuracy. ALBERT has 1/18 of BERT’s parameters.

How It Solves Healthcare Challenges

A large portion of healthcare data is unstructured. For example, most clinical information (such as history and physical notes, operative notes, procedure notes, and discharge summaries) is narrative—free-form writing that computers have never been good at interpreting. Deep NLP models (i.e., ones that use deep learning) can help identify important clinical elements and establish meaningful relationships between them from free-form clinical data in electronic health records (EHRs). This information can then be used in clinical decision support systems such as prior authorization, computer-assisted coding, clinical documentation improvement, and extraction of risk adjustment and hierarchical condition categories.

NLP can also be useful in extracting key elements from patient-physician interaction and autopopulating parts of EHRs, thereby reducing the time physicians need to spend on clinical documentation.

Use Cases

Some cases in which NLP is being used in the healthcare field are described next. (Note that we use deep NLP to indicate the use of deep learning with NLP, which often provides outcomes superior to NLP without deep learning.)

Where It’s Headed

NLP investment for healthcare is projected to reach $4.3 billion by 2024. New and innovative applications of NLP are likely in many areas, including the following:

• Improving customer experience

• Supporting care delivery

• Understanding clinical language

NLP will transform the healthcare continuum of care by extracting information from unstructured text to enhance the end-user experience. Virtual personal health assistants will create a better provider and patient experience using technologies like conversational AI. NLP will be the key technology to develop successful CAC and CACDI products to help patient care centers improve clinical documentation and free up their staff to spend more time with patients.

What It Is

Bringing complex data together for display or analysis is a relationship-intensive query for a traditional relational database management system (RDBMS). We now know that these calculations can be run more efficiently using permanent in-memory pointers in a graph database. This is the key difference between the old and new ways to store clinical data.

They are called graph databases because they use the same structures as the well-established branch of mathematics called graph theory—a connected network of nodes and edges. Using in-memory pointer jumping is a sharp contrast to relational databases, which use JOIN operations to indicate relationships. Pointer jumping uses a memory location to store the memory address of a related object. Traversal across the data is fast because all the necessary information is directly in memory. No other data structures get in the way and slow down the relationship traversals.

Table 1 shows some of the primary differences between relational and graph databases, along with why graph databases are preferred for many healthcare applications.

Graph databases are only part of an emerging trend in healthcare as can be seen in Figure 3, which shows a trendline in graph adoption since 2014. Graph analytics, graph visualization, and graph discovery are also core components to integrated healthcare solutions:

Graph analytics

The process of running complex queries over large amounts of data to find new insights. Because graph databases are so much more efficient at relationship traversal, graph developers can use many analytics algorithms to more efficiently find groupings of data, similar structures, and paths between structures.

Graph visualization

A collection of tools and processes for visualizing highly connected data. This includes using automatic layout tools to view thousands of nodes and their connections to see patterns within the data.

Graph discovery

The process of looking for unusual patterns in graphs that may not be evident by simple inspection. Graph discovery algorithms understand “normal” structures and highlight structures that might deviate from these norms.

Figure 3. Popularity trendline for various database architectures (source: DB-Engines.com, used with permission)

How It Solves Healthcare Challenges

Gathering all the related datasets for a patient in a central location is key to suggesting a care plan. The accuracy of care path recommendations improves as a longitudinal health record (LHR)—a comprehensive clinical summary of a patient-based clinical experience—becomes more complete.

Building the LHR data requires gathering data from disparate sources in many forms and loading it into a centralized graph structure, as seen in Figure 4, which shows an example of a graph model for an LHR. This data needs to be normalized (made consistent) so that all records can be queried with consistent results to be most effective. Graphs are superior to other forms of databases in that the retrieval of these complex LHR data structures requires only the simple and fast graph traversal of in-memory pointers.

Figure 4. LHRs are characterized by many complex relationships, which makes them ideally suited for storage in graph databases

Use Cases

This section discusses other areas of healthcare where graph technologies are already being used.

Where It’s Headed

Although graph technologies are already helping to lower healthcare costs, this is really only the beginning of the healthcare graph story. There are some particular areas to watch.

What It Is

The human genome functions as a blueprint for the human body. DNA in the genome is located within chromosomes, comprises four building blocks, and represents the biological “plan” for each person. That plan tells each cell what to do and when to do it. A potential for disease occurs when a person’s biological blueprint is flawed. These flaws are called variants. Matching variants with clinical signs and symptoms can establish the linkage with the disease. This matching process requires the ability to search millions of medical records and genetic data to establish a linkage between a disease and one or more variant.

The process of mapping a person’s genome has become significantly cheaper. The first human genome, sequenced in 2003, cost $3 billion and took 13 years to complete. The cost is now in the range of $1,000, and the sequencing time is measured in hours. In addition, the data collected is more precise, and the insights obtained through DNA sequencing impact more conditions.

How It Solves Healthcare Challenges

As more genomic data is integrated into healthcare, clinicians and providers envision using genomic information in conjunction with other biomarkers to tailor a personalized approach to care known as precision medicine. Today, genomics is primarily used in oncology, rare diseases, infectious diseases, and pharmacology. Precision medicine is changing the way diseases are identified, prevented, and managed. It helps in personalizing treatments for cancer, diagnosing rare diseases, improving the success rate of organ transplants, and predicting disease outbreaks.

Use Cases

There are numerous genomic use cases that are helping advance healthcare and enabling patients to live healthier lives.

Where It’s Headed

As the pace of genetic sequencing accelerates and more genomic data is integrated into the health system, precision medicine will become a routine part of every person’s medical care.

Right now, most physicians and patients aren’t very familiar with genetic tests. In general, physicians don’t know what tests are available, how and when to order them, whether they are covered by insurance plans, how to interpret the test results, or how to talk to patients about the tests. This will change as the adoption of genetic tests grows thanks to education in key disease areas, medical research centers publishing evidence-based guidelines and scientific papers, and genetic testing being included in EHR systems.

What It Is

A blockchain is a shared, distributed digital ledger on which transactions are chronologically recorded in a cooperative and tamper-resistant manner. Transactions, such as money transfers between two banks, are recorded on a list, or ledger. Each bank maintains its own ledger of all account debits and credits, along with when the transactions occurred. A transfer between banks may take several days while the two banks reconcile and confirm that the transfer amount is properly debited from the sender account and credited to the receiver account.

This latency on transfers could be shortened if the banks could share a single ledger, because then there would be no need to reconcile multiple ledgers. However, having a single shared ledger for all banks is impractical due to technical and business considerations. But if each bank had its own copy of the shared information that was automatically synchronized, that ledger would be a shared, distributed ledger that addressed the business considerations. Additional technologies can also be used to ensure the transactions are valid and audit-proof.

There’s a lot of hype and misinformation about what blockchain can and can’t do today. Businesses are experimenting with the latest version of blockchain technology to track, trade, find, collect, synchronize, and validate data. Key to these efforts is finding ways to reduce the costs associated with data management. Blockchain makes it possible to trace information back to its original source and view every detail in between. It simplifies commercial relationships by letting an organization track and trade something of value (such as data) without an intermediary.

Enterprises have learned that blockchain technology is most useful when independent or loosely coupled organizations, such as those that do business together or are otherwise linked, want to confidently share and audit information and automate mutually beneficial processes. With built-in transparency and verifiability of transactions, control is decentralized, and all participating organizations are peers. These are important prerequisites if diverse stakeholders are to collaborate productively.

Blockchain technology is best suited to solve business problems involving the following:

• Multi-enterprise islands of data

• High cost to reconcile inconsistent data

• Lack of complete trust between companies

Further, blockchain enables organizations to collaborate on problems that pose little or no risk to competitive positions. But even where there is willingness, incentives may need to be created to encourage participation.

The rise of blockchain has generated a lot of excitement, with many companies attracted to it as the “shiny new thing” and experimenting with it as a solution to a variety of business challenges. However, companies would do better to focus on whether blockchain technology should be used for a particular use case and let the business problem drive the technology selection. This approach is called fit for purpose.

How It Solves Healthcare Challenges

As blockchain rapidly matures, many speculate that it will solve a lot of data problems in healthcare, such as improving data interoperability, helping prevent data breaches, and shifting ownership of personal health data from health systems to consumers. Healthcare organizations and big tech companies are experimenting with blockchain, seeking to understand what it is and how to effectively use it to derive real value. Figure 5 illustrates some of the stakeholders that could participate in and benefit from a future healthcare blockchain.

Figure 5. Many parties could participate in a permissioned healthcare blockchain

Use Cases

The use of blockchain is still in the early stages in healthcare. Due to the nature of the technology, blockchain requires density of adoption, also known as the network effect; that is, to be productive, a group of organizations must be willing to cooperate to test and leverage the efficacy of the technology. So, a good place to start with blockchain is on business problems shared by many organizations in an ecosystem. This section looks at a few examples.

Where It’s Headed

In most cases, an alliance between organizations is necessary to effectively leverage blockchain. As noted earlier, blockchain works when organizations come together to share and automate common processes. Additionally, blockchain requires a different business mindset. All alliance members are responsible for the success of the alliance, but no one member owns the blockchain. Instead, every participant is working together to solve an issue for the industry.

As you build or join an alliance, find potential allies who have an appetite for early adoption of new technology. Look for candidates that are open and willing to learn, as these are often more risk-tolerant and flexible. Proceed deliberately and quickly to explore how blockchain may help advance your business. Blockchain has the potential to give you and your business partners improved velocity with higher-quality transactions and new business models.

What It Is

In healthcare, the most widespread form of IoT is wearables. These include fitness-monitoring devices (such as the Fitbit and Apple Watch) and health-monitoring devices (such as a continuous blood glucose monitor or a Bluetooth-connected blood pressure monitor). More and more fitness- and health-related devices are appearing all the time.

Ambient computing works a little differently, in that a human speaks into the environment, and the computing system responds by answering a question or performing some action. Another category of ambient computing is video camera–based systems used to detect movement, identify falls, accumulate information about how an individual moves through the environment, and so forth.

How It Solves Healthcare Challenges

Ambient computing is a fundamentally new form of computing. Based on advances in artificial intelligence—specifically deep learning—voice and video data is now available for healthcare systems to process. For example, doctors can analyze video of a person’s gait to detect signs of early-onset osteoarthritis or other conditions. Face recognition can be used to infer a person’s emotions, and a telemedicine caregiver can determine whether an explanation of a medical procedure was confusing or frightening to the patient. Analysis of a person’s voice can even detect signs of depression.

New forms of interaction are available with ambient computing. Conversational AI uses ambient computing to provide a more natural and low-friction way for individuals to interact with healthcare systems. Gestures can be used to augment voice, and keyboard and touch technologies can enrich interaction with healthcare systems.

The healthcare industry uses the term patient-generated health data (PGHD) to refer to health-related data collected from IoT and ambient computing systems. PGHD becomes truly operational when it is integrated into the healthcare system. This involves merging the PGHD with each individual’s EHR, using that additional data to generate insights about the individual’s healthcare journey, and then assisting the individual’s care team to tailor their treatment.

Healthcare will begin to transition from episodic care to continuous care as PGHD is incorporated on a near-continuous basis as shown in Figure 6.

Figure 6. Differences in the verity, velocity, and volume of healthcare data with continuous care

In Figure 6, the space above the timeline running through the middle depicts the current situation. A patient sees a family doctor once a year. A couple of labs are run during this visit, and the patient’s EHR is updated. If all is well with the patient, there is no further update of the health system for a year or so. The space under the timeline indicates that the patient has one or more IoT devices, or maybe there is an in-home video camera. This continuous stream of data can be analyzed much more frequently, allowing for a richer longitudinal picture of the patient’s health.

Use Cases

The use cases for IoT are as varied as the IoT devices themselves. Many devices help people with their fitness goals and with chronic care management.

Where It’s Headed

As healthcare systems adapt to leverage the continuous stream of data, and as more IoT devices and ambient computing systems are deployed, healthcare companies could begin adopting what we’ll call an ambient healthcare flywheel, as illustrated in Figure 7.

Figure 7. Ambient healthcare flywheel

The system would collect data from ambient systems and IoT devices and integrate that data with all the other healthcare data available about the individual (EHR data, labs data, prescription history, genomics data, claims data, etc.). The combined data would yield a digital twin of the individual—a model of the individual’s past and current health situation. Analytics, many of which are based on AI technologies, could be used to generate insights about the individual, addressing questions such as the following:

• Do the signals in the data suggest the onset of Parkinson’s?

• Does the data indicate that the individual is moving enough after his hip surgery?

• Is the individual able to do all the activities of daily living in order to reassure her loved ones that she is safe to continue living at home?

These insights may be disseminated to those in the healthcare system who need to know about and act upon them. If the individual slips and falls, this flywheel system can notify family members and 911. If the insight strongly suggests early onset or worsening of a disease condition, the healthcare team can be alerted.

With the flywheel, the actions performed by the healthcare system can be used in response to the insights derived from analytics run on the digital twin to indicate what new kinds of healthcare data would be worthwhile to gather. For example, should new cameras be installed to be more precise about gait analysis? Should the patient have a wearable blood pressure monitor?

More data yields a richer digital twin, which allows for better analytics, generating more accurate insights and in turn more timely and effective healthcare actions. The iteratively improving system suggested by the ambient healthcare flywheel will deliver the continuous, anticipatory, and personalized form of healthcare called continuous care.

Differing Worldviews

The opportunity for healthcare to reinvent itself through technology is abundantly clear, but exactly how and when it all will happen remains unclear. Robert M. Wachter summarizes this dilemma aptly in his book, The Digital Doctor: Hope, Hype, and Harm at the Dawn of Medicine’s Computer Age (McGraw-Hill Education, 2015). While holding tremendous promise for improving healthcare, Wachter says, “technology sometimes overpromises, underdelivers, and can even cause harm.” With clinician oversight and keeping humans in the loop, we should be able to limit or avoid harm with these new technologies.

Vinod Khosla, a venture capitalist and cofounder of Sun Microsystems, writes that healthcare will see the “evolution from an entirely human-based healthcare system to an increasingly automated system that enhances human judgment.” Physician and author Abraham Verghese, in a 2011 TED Talk, laments that the patient has become a mere icon for the “real” patient who lives in the computer, which he calls the iPatient. Instead of interacting and engaging with the patient, physicians are spending more time looking at EHRs, images, and data on the computer—treating the iPatient.

Dr. Verghese points out that no two patients are alike, and no two doctors are alike. Often the first signals of healthcare needs are received from one of these two sources, patient or clinician. This makes the science of which care pathway to use problematic. A lot of human judgment is often needed. How do you explain pain? Level of discomfort? The observations, eye contact, and person-to-person contact with a doctor continue to be of paramount importance for improved patient care.

So, iPatient or human-centered healthcare? Wachter, in The Digital Doctor, predicts a future where technologies will “transform the work, the people who do it, and their relationships with each other and patients.” He sees a future of human-centered healthcare, dominant now and in the next five years.

The Race Is On

The last five years have seen AI as a general-purpose technology powering innovations and new use cases for genomics, ambient computing, IoT, blockchain, graph, and more, as outlined in this report. Technology companies, startups, and consumers will continue to force change. New entrants into healthcare may, for example, lease a provider network in order to collect data about how networks work today, and then disrupt them by creating a true, real-time marketplace between consumers and providers, as has been the case with ecommerce in other sectors of the economy.

Deep learning will become the de facto approach for building intelligent learning systems rather than the existing rules-based systems. Forward-thinking companies will always choose machine learning over creation of a complex heuristic, reducing their technical debt and increasing their agility while building learning systems. In his keynote speech at the 2019 Conference on Neural Information Processing Systems, Yoshua Bengio, Canadian computer scientist and winner of the 2018 ACM Turing Award for his work on deep learning, sees a world where DL is extended from its current state to a future state.

The race is on to provide better patient care, deal with the increasing flood of data, reduce the variations in practice that diminish the quality of care, and get patients to engage more actively in their health. Healthcare and technology companies, startups, and others will look to do all of the following:

• Reinvent healthcare

• Be the smartest at healthcare

• Excel at healthcare user experiences

• Empower players in the healthcare ecosystem to do more

Asymmetrical technology-driven assaults on the existing profit models will force change as compelling tools and systems are introduced. Healthcare will improve because of the technologies and methods discussed in this report, such as ML- and DL-powered diagnostic tools, improved aging at home, and healthcare on the go with wearables and computer vision.

Technology will help transform healthcare as more patients and consumers become more engaged. It will augment the work of clinicians, reducing their workload and allowing them to focus on providing actual healthcare and staying engaged with their patients. Diagnoses will become more accurate, and treatment will become more personalized. In summary, the technologies described in this report will play a significant role in making the healthcare system work better for everyone.