Artificial intelligence (AI) has transformed many aspects of our lives. Common examples of AI include chatbots, digital assistants, facial recognition, self-driving vehicles, and social media platforms. Within the healthcare industry, AI innovations have improved monitoring and streamlined processes.
AI functions generally fall into two broad categories: deductive and generative. Deductive AI processes have received much of the attention. These algorithms seek patterns by analyzing complex data and performing deductive reasoning.
Generative AI Defined
Generative AI takes a different path. Beyond analyzing what already exists, this approach to AI creates new data by applying algorithms to training data. The new generative AI data may include images, text, sounds, and other datasets that have statistical similarities to the original data from which they were created.
With generative AI, the underlying technology teaches machines to synthesize new “fake” data after determining models, patterns, and properties in real data. This contrived but statistically valid data can expand datasets so that no single data point can be traced to an individual, while also compensating for data types that are underrepresented.
In healthcare, generative AI has a wide variety of applications throughout the spectrum of medical services. From medical education to disease diagnosis, and from patient care to documentation, innovations in generative AI in healthcare hold great promise.
Two generative AI processes that will continue to transform medical education and care are Generative Adversarial Networks (GANs) and Large Language Models (LLMs).
Generative Adversarial Networks
Often associated with the creation of “deepfakes,” GANs fabricate synthetic data as two AI neural networks compete against each other and learn from one another. The “generator” manufactures fake images, content, or data, with a pathway toward producing more realistic data. The “discriminator” evaluates data that is generated, and attempts to determine whether it is fake or real.
The goal of GANs is to learn to discern real data from fake “random noise” content. As these two networks compete, the generator improves in its ability to produce realistic data. As applied to healthcare, GANs can parse underlying data from extensive repositories of electronic health records (EHR), create data within realistic parameters, and apply these datasets to applications and processes.
Large Language Models
Known popularly for their ability to compose text that seems human-created, LLMs handle natural language processing (NLP) tasks. They absorb copious amounts of textual data, parse content, seek word relationships, and generate effective summaries.
LLMs have the potential to automate timely processes, especially related to transcription and documentation. An LLM can ingest various information in a patient’s EHR, synthesize everything from medical images and notes to lab results, and paint a summative picture. Content generated by LLMs may reduce the need for time-consuming documentation practices.
Frameworks to Consider
Similar to any increased reliance on technology, the use of generative AI in healthcare must take into account data integrity, data privacy, and the expertise of the medical practitioner.
Within the healthcare field, AI can become another tool, similar to a thermometer or stethoscope, that offers accurate readings. In this scenario, the doctor interprets readings while making decisions. The trust of human participants–the doctor, patient, and clinical team–remains a limiting factor.
Ongoing research in generative AI opens a different door, however, one where machine learning through GANs and LLMs transforms medical education and practice by expanding datasets so that physicians can make decisions on larger models. The goal is not technology for technology’s sake, but to improve the quality, affordability, and accessibility of healthcare.
In order for generative AI algorithms to make a broader, positive impact, these systems must meet the needs of healthcare consumers and practitioners within this medical framework:
- The system is believed to be useful and helpful.
- The system is flexible, and both easy to use and understand.
- The system offers opportunities for training, support, and improved efficiency.
- The system expands research, improves diagnoses, and creates personalized solutions.
Benefits of Generative AI in Healthcare
If incorporated responsibly, generative AI can parse and evaluate large datasets in a way that offers enhanced patient care and suggests better treatment alternatives. In addition to better research and improved diagnoses, AI can reduce paperwork, streamline billing, create important documentation, and extract helpful summaries for providers and patients.
Benefits span across the medical spectrum, in areas such as these:
Medical Education: Students in medical schools benefit from customized simulations and training materials. Virtual patients created by GANs allow for consultations, diagnoses, and treatment plans that create real-world scenarios without risk to living patients. Datasets and chatbots through LLMs can also account for uncommon illnesses not witnessed during their practicum.
Patient Education:Generative AI can design visual guides, provide helpful information, and offer a method for patients to interact without having to talk directly with a medical staff member about a sensitive issue.
Patient Privacy: By using anonymized and created data, GANs allow for simulations of medical diagnoses, testing, and other procedures with less potential for the sharing of an individual’s unique and sensitive information.
Disease Diagnosis: Generative AI models have shown an increased ability to make correct diagnoses based upon evaluating text responses or analyzing health records. Ongoing research should consider unanticipated “diagnostic blind spots” or biases in datasets.
Drug Discovery: Rapid data analysis allows for modeling to predict a drug’s safety, efficacy, and possible side effects by looking at interactions in datasets rather than with human subjects. Clinical trials that occur later can be modeled on results from GANs.
Mental Health/Behavioral Practices: Practitioners are aware of patient anxiety during medical procedures. Cognitive behavioral therapy (CBT) also tries to change and manage thoughts and behaviors. Generative AI can create triggers, and teach patients how to cope, respond, and take control of their reactions.
Creating Documentation: LLMs that review a variety of sources, including images, notes, reports, and other EHRs, have the ability to generate drafts of discharge summaries, progress notes, or patient instructions. Doctors must document for regulatory compliance and billing, and to reduce the risk of litigation. By creating clinical documentation, generative AI improves physician productivity.
Reviewing Documentation: Generative AI processes that comb through data can summarize, track changes in the patient’s condition across time, and put into place criteria to make sure something is not overlooked. By gathering, sorting, and generating data, AI allows practitioners to analyze patient progress faster and with less potential for errors.
Claims, Billing, and Recordkeeping: In addition to scheduling appointments, improved generative AI in healthcare facilitates the automation of insurance claims, reduces billing errors, and handles clerical tasks with improved efficiency.
Limitations of Generative AI in Healthcare
Although generative AI offers many benefits, the process remains imperfect. Computers can compile data rapidly, but imperfections in datasets may have grave consequences.
Areas of concern and limitations include the following:
Accuracy of Datasets: Algorithms that create new content from training data allow researchers to make hypotheses about diseases and drug effectiveness. However, human intervention and an understanding of the computational processes that create this fake data remain important to ensure that it reflects the populations under consideration.
Need for Labeling: GANs require extensive computing power, and technology has generally kept pace. However, think about the real-world liability of using unknown medicine in an unlabeled container. If data is unlabeled or improperly labeled, it may produce bad data or suggest inappropriate techniques or results.
Questions of Trust: If a doctor shares lab test results or common readings (such as body temperature) from samples received directly from the patient, they are usually trusted. If a doctor does an analysis based on computer-generated data, does the patient have the same level of trust?
Ethical Issues: Synthetic data predicts trends, but it cannot entirely replicate an individual. Privacy laws must be accounted for in the creation of fake data from agglomerated records. AI and machine learning are not identical. Machine learning involves massive and complex computation; AI “mimics humans without being alive.” At what point does the human doctor defer to the machine?
Regulatory Matters: As the presence of generative AI in healthcare increases, regulatory bodies will have to get involved. Do tests on synthetically derived datasets equate with laboratory practices on living beings? Will bodies such as the U.S. Food and Drug Administration approve drugs when their effectiveness was largely tested by GANs?
Remembering the Human Factor
Generative AI has great potential to improve healthcare practices. However, technology alone cannot transform medical care. Real change in all facets of healthcare, from medical education to critical care, requires that generative AI practices continue to “put people first.”
The bond of trust between doctors and patients must be maintained. Generative AI may assist, complement, and re-imagine aspects of the doctor-patient relationship, but the foundational trust must remain intact. Patients often seek medical care due to concerns and vulnerabilities, and they select practitioners based upon trust.
Enhancements of generative AI in healthcare will continue. Human empathy also remains a top priority.
