The Uni healthcare sector is ripe for disruption. In particular, the healthcare technology industry is highly fragmented in the United States, with interoperability challenges (within and between medical organizations) limiting the ability to fully aggregate and analyze data. Fortunately, in terms of data and technology, artificial intelligence (AI) has a particularly large potential to positively impact healthcare across the U.S. health ecosystem. Within the AI domain, a growing field of interest is generative artificial intelligence (GAI), which has the capability to impact multiple parts of healthcare for U.S. provider organizations, including administration, diagnosis, research, and treatment. It can both enhance quality and improve productivity (i.e., attain both effectiveness and efficiency). The integration of AI in certain areas of healthcare, such as medical imaging (e.g., detect cancer on medical scans), is fairly advanced, benefitting greatly from previous research in computer vision, but in many other areas, AI is in the infancy stages of development, testing, and implementation. For example, IBM’s AI-driven supercomputer Watson tried to address a broad set of healthcare challenges, including helping physicians diagnose diseases and recommending participants for clinical trials, but it ultimately has experienced only mixed success. Adopting AI in healthcare will be slow and difficult in the U.S., including with GAI. However, as we discuss, the potential for GAI to positively impact multiple domains in the healthcare ecosystem is tremendous.
Importance of GAI
Generative AI models are sometimes called “foundation models”. In some ways, these models are an entirely new approach to AI, in part because they are able to have a large impact in working with comparatively smaller data sets. Generative AI is currently at a nascent stage, but its impact is likely to grow as new applications continually develop. GAI does have the potential to generate trillions of dollars in economic value in the U.S. and internationally. By 2025, generative AI is projected to account for 10% of all data produced globally. In the long run, there is little doubt that AI will usher in fundamental, world changing innovations.
GAI is not a new idea, but it has reached a new threshold. Over the past year, machines rather suddenly --or so the pace seemed--became really good at generating completely new images and writing “original” text. Why is this emerging now? There are at least four reasons. First, the corpus of content (i.e., data) continues to increase rapidly, pulling from audio, video, and text sources. Second, there is greater storage capability. Third, processing power (i.e., speed of computation) continues to increase substantially. Fourth, algorithms have gotten considerably better.
Many experts and investors believe generative AI technology will usher in a productivity revolution across various industries, giving birth to massive companies in the process. GAI does hold the potential to eventually transform application software and workflows across many American hospitals and other healthcare organizations. While GAI models have shown rapid innovations over the past couple of years, that does not guarantee continued improvement at the same speed. AI does have a history of overpromising and underdelivering on the investor, practitioner, and societal expectations. For example, self-driving car technology has been promised and promoted for years in Europe and the U.S., but thus far, autonomous vehicles have only been deployed in a few limited pilot programs. However, the impact of GAI on fields such as healthcare is certainly more a matter of “when” not “if”.
What is GAI?
GAI are programs that allow computers/machines to use data elements such as audio files, images, and text to produce content. As such, it is artificial intelligence that aims to be creative rather than just to process and synthesize data. Generative AI can be used to produce original content on behalf of healthcare organizations. For instance, a GAI application looks at 100,000 samples and then aims to create the 100,001st sample itself. Generative AI is one of the most promising advances in the AI ecosystem in the past decade. GAI results in the creation of higher quality outputs as the machine self-learns from every data set. This permits the computer to comprehend, evaluate, and leverage new abstract, conceptual, and ideational principles. Unsupervised learning allows AI to more quickly acquire adaptable transferable skills that increase accuracy, effectiveness, and speed relative to typical human efforts where little employee training is required. As such, this technology has the capability to be tremendously beneficial for U.S. healthcare, life sciences, and medical organizations.
Generative AI creates new content by powerfully utilizing existing content. It references AI techniques processing data (text, image, audio, and video), utilizing them to create new original content that preserves similarities to original data. GAI models can pull from almost all available information on the internet, a data resource that obviously continues to grow rapidly, roughly doubling in size every two years. For example, text-oriented GAI models can examine thousands of digital books and trillions of words on the internet. GAI can enable machines to automatically control and regulate environments by taking action to accomplish definable goals. It can even create neural networks that imitate and mimic human brain activity. These capabilities are growing at a steady pace, including learning complex patterns in a repetitive nature similar to the human brain.
Leveraging state-of-the-art generative technology that learns new patterns, structures, and variations automatically using existing data, the industry is now developing realistic simulations and representative synthetic data at scale. Also, via video synthesis technology, GAI companies are implementing innovative visual content creation that reduces cost, improves quality, and lowers language barriers. GAI can leverage unsupervised learning algorithms to build new plausible content based on existing content, allowing machines to understand patterns in the input content and then generate similar content. Typically, during the training, GAI models are provided with limited parameters; this allows the model to yield its own conclusions about the most essential characteristics of the data. In theory, which makes GAI more capable of creating results free of the normal biases attendant to human comprehension, experience, and thought processes. In contrast, many traditional AI learning models have proven vulnerable to becoming discriminatory and skewed.
Generative AI can train neural networks even without access to all the training examples. Instead, we can provide the network with sufficient examples to learn the underlying structure of a problem. Then, after the model has learned this structure, we can generate more samples leveraging this knowledge. GAI can be taught to generate fake examples of underrepresented data that help develop and educate the model. Overall, one primary aim of generative intelligence is to identify new cases before they materialize, while also developing a recommended course of action.
Artificial intelligence, including GAI, can be used to disrupt the healthcare industry. In terms of healthcare specifically, how are and how will U.S. healthcare organizations leverage GAI? These can be placed into five large buckets: (i) clinical, (ii) privacy/security, (iii) administrative, (iv) fraud detection, and (v) drug discovery.
First, from a clinical perspective, generative AI can help enhance patient treatment. GAI-powered applications enable computers to produce new content based on existing information, so they can be used to create fake cases for underrepresented data, which substantially enhances development and training of the model. For example, in the case of X-ray images, it can supply various additional “fake” angles to visualize potential tumor growth outcomes, or it can be utilized to contrast the image of healthy organs with affected ones to help detect malignant developments. In the case of retinopathy diagnosis, this can be leveraged to create new medical images for diagnosis and testing (e.g., helping detect malignant developments by contrasting images of healthy organs from the databank to the affected one or to visualize possible tumor expansion by computing different angles of an x-ray image). GAI can also enable early detection of certain conditions for effective treatment options.
Second, GAI can strengthen privacyprotecting applications. Data de-identification is a significant problem for healthcare analysts. The reversal process can compromise sensitive and valuable patient records as it can lead to full identification. GAI provides a potentially promising solution. It can strengthen privacy and security for patients. It can help secure the reversal process to make it more penetration-proof and less susceptible to data de-identification. Also, generative AI avatars can be used to protect the identity of patients in the virtual care setting, particularly for patients who may not be fully comfortable with disclosing their identities in certain cases. GAI holds significant potential for American healthcare organizations to enhance the privacy and security of patient data.
Third, GAI can help with the administration and management dimensions of healthcare organizations. Primary care physicians spend twice as much on administrative and clerical tasks, including EHR entry, as direct patient care. Leveraging GAI-based technologies, (e.g., natural language processing and voiceto-text transcription), administrative and clerical activities such as completing EHR notes, filling prescriptions, and ordering tests can be substantially automated. This would also help improve physician satisfaction and reduce burnout rates. It would be particularly advantageous if one could leverage computer vision sensors and microphones hardware, coupled with a deep learning model software, to automatically and simultaneously convert patient-physician conversations into clinical documentation. Relatedly, healthcare organizations can use GAI to create new patient records that are then inserted into the system to enhance accuracy. In the business development domain, some companies are using AI to help automate highly repetitive tasks such as blog posts or sales emails. Thinking even more ambitiously, GAI can use audio synthesis to generate a computer voice very similar to a human voice, which can be used for communicating with patients and narration for videos.
Fourth, GAI can be used to help organizations detect the significant fraud, waste, and abuse that afflicts the U.S. healthcare ecosystem, as it does globally. GAI can automate fraud analysis to detect fraudulent transactions leveraging predefined algorithms. According to an analysis of one healthcare market from McKinsey & Company, nearly three-fourths of received health claims are flagged as unusual for manual review via audits, but only about one-tenth of these cases are successfully investigated. GAI can be used to identify and isolate malevolent activity.
Fifth, GAI can play a crucial role in drug discovery. By 2025, it is estimated that half of drug development initiatives will rely on generative AI. GAI is able to create molecular structures of drugs employed in curing certain indications. The treatment of new diseases can be accelerated when GAI is used to perform a rapid database search of compounds applicable for utilization. GAI can also be used to create organic molecules and prosthetic limbs from scratch with 3D printing. GAI can also help with drug development. For example, GAI was used to research antimicrobial peptides (AMP) to find drugs for Covid-19. Over the long-term, this means GAI has the power to improve the quality and speed of drug development, which has positive implications for both providers (who are the entities that tend to administer clinical trials and prescribe drugs) and pharmaceuticals (who are the entities that tend to design and pay for clinical trials).
While the overall future is exceptionally bright, GAI does present several challenges for full adoption by American health leaders and managers. First, because of their underlying complexity, GAI models can sometimes be unstable and difficult to control – often leading to unexpected, unexplainable outputs. This stochastic element is conducive to creativity and originality, but the healthcare environment generally prefers controllable, expected, explainable, and stable outputs. Thus, there will be an ongoing challenge for healthcare institutions in the U.S. who adopt GAI to resolve this inherent tension of creativity/originality versus unpredictable/unexpected.
Second, a significant amount of training data is needed to train GAI, so it is more expensive to implement given the requirement of significantly greater processing power. Third, while GAI can create new content by combining data provided in new ways, it cannot create completely new things. Machines struggle to comprehend highly abstract concepts encountered in real and virtual healthcare environments. Fourth, just as it can be a force for good, GAI can also be used by bad actors for untoward purposes such as generating fake news stories, developing sophisticated fraud schemes, and violating patient privacy. Fifth, artificial intelligence applications in healthcare, including GAI, can seem like a “black box” and thus untrustworthy to patients and providers. Thus, over time, a lack of trust by patients and providers will likely be a key roadblock to achieving greater use of AI across the healthcare ecosystem.
The future of generative artificial intelligence in the U.S. healthcare ecosystem is indeed promising. The most compelling use case for Generative AI-powered applications is original content creation based on existing information. GAI programs can use deep learning techniques to train themselves on massive amounts of data from actual patients and then generate new images based on those patterns. This technique allows GAI to create new data sets that humans could never have developed within any limited space and time. To further improve accuracy, the apps can then compare generated content against real-world data, allowing them to analyze huge quantities of data quickly and efficiently. In turn, this can lead to previously unreached levels of insights into diagnostics, diseases, and even treatments. GAI can be the basis for many applications in healthcare, and as the technology continues to develop and evolve, it will be finetuned to integrate more advanced applications.