The wind of AI has finally reached the healthcare industry.
Healthcare has always been a large and highly typical industry scenario in urgent need of AI support. AI technologies represented by large models are now converging deeply with biopharmaceuticals and medical diagnosis and treatment, setting off a new wave of artificial intelligence. This article offers an in-depth reading of the 2023 Healthcare AI Large Model Industry Research Report for our readers.
Healthcare AI large models are a branch of artificial intelligence in medicine. By learning from vast amounts of data, they generate new data instances and show strong application potential in medical imaging, drug discovery, assisted prevention and medical text analysis. They can improve both the quality and the accessibility of medical services while driving innovation across the entire healthcare industry.
1
In 1956, the concept of "artificial intelligence" was first put forward. As mathematical tools such as Bayes' theorem and Boolean algebra were applied in medicine, an era of machine learning driven by statistical algorithms and prior knowledge began. The growth of clinical knowledge bases, electronic medical records and AI+imaging, together with the establishment of systematic standard datasets, pushed China's medical AI into the era of deep learning under a data-driven model, shifting from single-modality data to multimodal data fusion. After 2017 it evolved into the era of large models built on the Transformer framework and incorporating new algorithms. In recent years, major universities have launched research and development of healthcare AI large models, providing a solid technical foundation for the industry and helping China's AI+healthcare move into a stage of stable development.
Clinical application scenarios
Application scenarios for healthcare AI large models
Offering differential diagnosis for patients and diagnostic suggestions for physicians
Using image recognition and deep learning, intelligent image recognition helps physicians quickly extract the disease information presented in an image, improving diagnostic accuracy and efficiency, while providing clinical decision support and helping physicians formulate appropriate treatment and follow-up plans.
Raising the probability of success and lowering development costs
AI large models extract relevant information from scientific publications, linking genes, diseases, compounds and biological processes, and deeply analyzing disease molecular mechanisms and biological activity in the disease context to support drug target discovery. AI drug screening models can rapidly and accurately mine and screen suitable compounds, using computer simulation to predict drug activity and safety, improving the efficiency of drug development, reducing costs, and enhancing accuracy and personalized treatment.
Application to disease prevention holds enormous potential
Predicting an individual's risk of chronic diseases such as diabetes and heart disease based on factors including genetics, lifestyle habits and environment.
Delivering accurate, fast and personalized diagnosis and recommendations
Built on AI large models, medical text can be processed quickly and accurately: automating the handling of clinical records and building personalized health management plans; parsing medical literature to support medical research; helping convert clinical services into standardized medical billing codes, improving billing efficiency and facilitating operations management and decision support; and enabling intelligent forecasting of epidemics, reducing national disease prevention and control costs.
As healthcare AI large models attract widespread attention, a growing number of model testing studies have emerged. In an examination of 10 sets of images by the Royal College of Radiologists, the overall accuracy of AI was 79.5%, compared with 84.8% for radiologists. GPT-4V achieved 62% accuracy on the United States Medical Licensing Examination. Among the seven AI models tested against China's medical licensing examination, six platforms scored above 60%, even surpassing the general level of real candidates (whose overall written examination pass rate is 20%–25%). The tests also exposed the shortcomings of AI large models: when faced with non-standardized questions requiring comprehension or logical analysis — such as questions with images, scenario-based questions or "select the incorrect answer" questions — AI large models show relatively high error rates, lacking physicians' deep contextual understanding, flexible adaptability and interpersonal communication skills, particularly in traditional Chinese medicine diagnosis and treatment settings.
Future development and challenges
According to forecasts by EqualOcean, 2023–2027 will be a period of concentrated expansion for healthcare AI large models, for the following reasons:
01 Improved R&D efficiency
AI models can automatically generate content, code and data analysis reports, supporting the rapid development and testing of R&D models;
02 More accessible API services
Improving the accessibility of AI large models and allowing non-technical medical professionals to integrate systems with ease;
03Large pre-trained models ready for direct use
Developers can fine-tune or optimize models for specific application scenarios without training from scratch, lowering the cost of data collection and computing resources.
Market size and growth rate of China's healthcare AI large model products, 2023–2030
Healthcare AI large models show broad promise. They can improve the efficiency and quality of medical services and bring patients better treatment and prevention options. Yet along the path to commercialization, beyond the technical layer, they also face multiple challenges including data privacy, intellectual property, safety and ethics. To overcome these challenges, in addition to strengthening technological R&D, it is necessary to introduce relevant legislation, establish standardized evaluation processes, data management and privacy protection mechanisms, and reinforce ethical norms and regulatory oversight.
