Generative AI, a machine learning framework that creates text, images, music, or videos using artificial intelligence, has gained widespread attention since the launch of OpenAI's ChatGPT. But how is this technology evolving in the biomedical field?

The Biomedical Science and Technology Center recently interviewed Yao-Fang Hsieh, Senior Manager of Market Research at Quanta Computer, who is responsible for strategy planning and investments. From a tech industry perspective, Hsieh shared insights on the current and future applications of generative AI in the biomedical sector. The interview is presented in the first person:

Generative AI is a branch of deep learning that, compared to earlier deep learning techniques, focuses more on generating text and understanding the context of language. In recent years, the development of generative AI has been rapid, extending beyond text generation into the field of medical imaging. For instance, some healthcare institutions are experimenting with applying generative AI to medical imaging interpretation, leveraging its ability to generate similar images as samples to reduce the need for manual labeling, thereby improving accuracy and efficiency in image interpretation.

Currently, the primary application of generative AI is in text processing. Microsoft, for example, is using this technology to streamline medical record keeping and simplify healthcare administration. In telemedicine, generative AI is showing potential—whereas traditional chatbots provided rigid, standardized responses or required a nurse to be available 24/7, generative AI can now provide more personalized and user-friendly replies. However, these AI-generated responses still need continuous training and guidance to ensure accuracy and appropriateness, avoiding repetitive or irrelevant answers (a phenomenon known as "looping").

Potential and Challenges of Generative AI Applications

Professional Challenges: Generative AI Still Relies on Guidance and Expertise

One limitation of generative AI is its inability to fully understand highly complex issues, requiring professional guidance to generate responses that meet user needs. This highlights that the domain knowledge of doctors and nurses remains irreplaceable, as generative AI struggles to independently assess the accuracy of information and cannot handle highly specialized problems. Therefore, integrating domain expertise into AI training in medical applications can further improve its accuracy and practicality.

Additionally, the issue of "looping," where AI produces repetitive or irrelevant content when faced with complex questions, underscores the current limitations of the technology and the indispensable role of healthcare professionals in medical decision-making.

Text Generation Challenges: Content Coherence and Algorithm Limitations

The longer the text generated by generative AI, the more its coherence and logical flow can be challenged. This is similar to video generation—creating a 15-second clip is relatively easy, but generating a 5-minute video with consistent plot and logic presents challenges, which is a current bottleneck of the algorithms. While generative AI performs well in text generation, it still faces difficulties in maintaining semantic consistency and logical coherence in longer texts.

Chip Technology Development: Enhancing Computational Efficiency is Key

The efficiency of generative AI heavily depends on underlying hardware, such as chip technology. For example, GPT-3 has 96 neural network layers and 175 billion parameters. Using older chip technology for processing, even the most advanced cancer model algorithms could take days to run. By the time the AI completes its analysis, the physician would have already made their diagnosis. Therefore, the development of AI chips, especially those optimized for medical applications, will be a key factor in advancing generative AI.

Exploring AI in Medical Imaging: Comparing Deep Learning and Generative AI

It remains uncertain whether generative AI can surpass deep learning in image processing, but its advantages in text generation are already evident. Deep learning struggles to generate coherent reports because it relies on natural language processing (NLP) to establish relationships between words, resulting in poor performance when generating reports. However, with the advent of generative AI, this gap has been closed, enabling more natural and efficient report generation. This presents a powerful tool for automating the generation of medical texts.

Four Key Areas for Generative AI in Healthcare

While the potential for generative AI in healthcare is vast, its development is still in the exploratory phase. For now, we should observe its application in the following four areas to evaluate its effectiveness:

Electronic Medical Records (EMR) and Hospital Administration

Microsoft and other international companies are developing generative AI to automate the creation of EMRs and simplify administrative tasks, which could enhance operational efficiency and reduce the workload of medical staff.

Insurance Claims and Underwriting

Some foreign insurance companies are using generative AI for underwriting and claims management. These systems can automatically analyze customer data to detect fraud and determine payout amounts. Moreover, generative AI can design new insurance products, enhancing the flexibility and personalization of insurance analysis.

Telemedicine Enhancements

Generative AI has the potential to improve telemedicine, particularly in providing more flexible and accurate responses for initial diagnoses and patient inquiries, reducing the burden on healthcare professionals. However, continuous optimization is necessary to ensure that responses meet actual patient needs and avoid misleading information.

Optimizing Clinical Trials and Drug Development

In drug development, generative AI shows potential to accelerate clinical trials. Traditionally, recruiting large numbers of patients for trials is costly and time-consuming. By using generative AI to identify suitable patient groups, the recruitment process can be more efficient, reducing the time and cost of trials. Additionally, AI can help identify risks in clinical trials and optimize trial design to increase the likelihood of success.

In summary, while generative AI is still in the early stages of exploration in the healthcare sector, its potential to bring significant change over the next 5 to 10 years is promising. However, to ensure accuracy and reliability in applications, generative AI must be combined with the expertise of healthcare professionals. As AI technology continues to evolve, generative AI will have an increasingly profound impact on the biomedical field, leading the development of digital health technologies.

Resource (mandarin): 晶片發展影響AI準確度 專家籲專業人員仍不可取代