On November 30, 2022, OpenAI launched ChatGPT, and within a week, over a million users had registered. As a result, generative AI has garnered significant attention across various industries. What potential applications does generative AI have in the medical field? Has its application in healthcare reached maturity? In an exclusive interview, Kuo, Chang-Fu, Director of the Medical Artificial Intelligence Core Laboratory at Chang Gung Memorial Hospital, offers insights from a clinical perspective.
Generative AI is transforming various industries, but its impact in healthcare is just beginning. The two core elements of generative AI are “understanding” and “generation”; without effective understanding, it is impossible to produce useful outcomes.
This is particularly crucial in healthcare, where daily tasks for medical staff are mainly divided into two categories: communication and documentation. Communication involves interactions with patients, family members, and collaboration between doctors and nurses, while documentation is essential because medical activities are legal actions, and clinical decisions must be recorded. Both tasks are time-consuming, making the potential application of generative AI in these areas quite appealing.
Generative AI can assist healthcare workers by automatically generating records and documents, significantly reducing their workload. Additionally, clinical treatment increasingly emphasizes providing care according to established standards, consensus guidelines, or protocols. For instance, the Rheumatology and Allergy Immunology Department reviews nearly 20 guidelines each year, and if doctors can use generative AI to quickly extract key points and summaries, it could save a substantial amount of time.
Generative AI Research Still in Early Stages; Caution Necessary for Medical Applications
While generative AI appears to be a hot topic currently, it is still primarily in the academic research phase. The development in healthcare has only begun around late 2021 and remains in its infancy. According to statistics from the PubMed database, there were zero medical papers on generative AI in 2021, seven in 2022, and a rapid increase to 500 in 2023. However, most of these are prospective review articles, with only a few dozen being original research papers. This reflects a relatively conservative adoption of generative AI in the medical field, which is due to the stringent demands of healthcare.
The characteristics of the medical industry lead to a slower introduction of new technologies. Unlike the ICT industry, the medical field has very high requirements for the accuracy and safety of new technologies. Therefore, I believe that any significant impact of generative AI on the medical field will be difficult to observe in the short term.
Current Medical Applications Predominantly Use Discriminative AI; Generative AI Effects Still Awaiting Observation
Generative AI is fundamentally different from the more mature discriminative AI currently in use in the medical field. Discriminative AI is primarily utilized for diagnosis and image interpretation, such as in radiology and pathology. In contrast, generative AI can further support clinical decision-making by interpreting and generating unstructured data. However, due to the complexity of text generation and data integration involved, its clinical applications are still not mature.
For AI to be used in clinical settings, it must comply with legal regulations, obtain certifications like ISO and GMP, and pass reviews from Institutional Review Boards (IRB) and clinical trials before it can be effectively implemented.
In the past five years, about 90% of mature AI products in clinical practice have been related to imaging applications in radiology and pathology, while there are only about 20 to 30 applications related to non-imaging areas, primarily involving signal processing and some digital psychotherapeutic applications.
A recent article published in NEJM AI noted that the adoption of medical AI devices is still in its infancy, with most usage driven by a few specific devices, such as those assessing coronary artery disease and diagnosing diabetic retinopathy, with over 10,000 claims.
At Chang Gung Memorial Hospital, the AI software that has received approval from the Ministry of Health and Welfare primarily consists of discriminative AI systems, including software for detecting wrist scaphoid fractures, atrial fibrillation, and a heart failure screening detection tool (ECG). These applications are predominantly image-based and do not involve generative AI.
Traditional AI functions similarly to facial recognition, indicating whether something exists or not, and primarily assists doctors in making interpretations. If AI software helps doctors make a diagnosis and then generates a complete report, that qualifies as the use of generative AI. However, many current reports rely more on structured data solutions, fitting input into predefined templates based on the patient’s gender, age, medical history, and findings, which does not truly reflect the capabilities of generative AI. The real potential of generative AI lies in generating unstructured data and producing different outcomes based on various inquiries. As I mentioned earlier, the two core focuses of generative AI are understanding and generation, with understanding being even more critical; the current challenge lies in achieving effective understanding before proceeding to action.
Considerations of Patient Privacy and Processes in Medical Record Generation and Data Extraction Using Generative AI
Healthcare professionals often spend excessive time on communication, and there is immense potential for generative AI to alleviate the burdens of communication and documentation. I believe that utilizing generative AI to assist with voice recognition, documentation, and medical records could reasonably reduce the workload for healthcare workers. For example, using generative AI for voice recognition to document patients’ reports can significantly lower the workload. However, introducing new technology also brings challenges, such as needing to obtain patient consent before recording their voice. The process of generating medical records from voice recognition is fundamentally different from mere voice recognition; ensuring privacy and data security during the record generation process is a key concern.
The interaction between a patient and a doctor, followed by recording via computer, may seem like two steps, but they are not. Generative AI must listen to the voice, comprehend it, and then document it; when generating records, it needs to determine whether to produce a verbatim transcript or extract key questions and generate records based on medical requirements, followed by a review and consolidation by the physician. Since manual processes vary across different hospitals and specialties, replacing them with computer systems requires careful consideration.
However, for knowledge acquisition, particularly in rapid inquiry and real-time comprehension to extract vital information, generative AI can be especially beneficial for junior doctors or residents. For instance, when deciding on antibiotic treatment, a senior physician may quickly determine which medication to use, while a younger doctor may need to reference information; generative AI can provide immediate answers to assist healthcare workers in making swift decisions.
Nevertheless, generative AI may encounter issues such as "hallucination," where it generates false or misleading responses. This is absolutely unacceptable in healthcare. Several solutions exist, such as fine-tuning generative AI to limit its creativity; another approach involves feeding AI a range of possible answers similar to an open-book exam. Techniques like Retrieval-Augmented Generation (RAG) can help AI generate answers based on actual data instead of "inventing" new content. For example, if the physician knows the question pertains to an infection, the generative AI should respond based only on literature related to infections rather than fabricating an answer. The key difference between AI and humans is that humans may research slowly, while AI can provide rapid responses and synthesize appropriate answers.
Prioritizing Safety and Effectiveness in Medical AI Applications
Regardless of how promising AI technology appears in healthcare applications, we must not implement AI for the sake of AI itself. The most crucial prerequisite for any medical tool is ensuring safety before considering effectiveness. For patients, the application of generative AI must improve their prognosis, and for healthcare workers, it should reduce time costs and alleviate their workloads. If a technology increases the workload of healthcare professionals, its value will be significantly diminished. Therefore, validating its impact in real-world scenarios and conducting fair assessments remain vital; ultimately, it must comply with regulations and require healthcare or insurance coverage for stable operation in actual medical environments.
We cannot apply the rapid development perspective of the ICT industry to healthcare. Generative AI, as an emerging technology, is still in a trial and error phase. Overcoming the conservative and closed nature of the medical industry requires patience; finding a "killer application" that genuinely enhances healthcare efficiency will be key to achieving widespread adoption of generative AI in the medical field, allowing it to truly exert its influence.
Resource (mandarin): 應用仍在試錯 長庚AI實驗室建議符合安全再考慮有效
