What impact does the big model have on the healthcare segment?

A review of IEEE J-BHI pre-trained large models for medical and health fields

This article provides a comprehensive review of recent large models in healthcare, covering the background of large models and their applications and opportunities in various fields of health sciences. It also focuses on the challenges of developing and deploying large AI models in healthcare settings and provides development recommendations for building trustworthy large models.

In recent years, AI big models have developed rapidly, and their volume and the amount of data used for pre-training have reached hundreds of billions. After pre-training, these big models have shown strong generalization and generalization capabilities in various downstream tasks. ChatGPT is a typical example, and its demonstrated capabilities make people look forward to the transformation and opportunities that big models can bring to the medical and health fields. In the era, researchers in the fields of biomedicine and health have realized the importance of data. In the past 10 years, the scale of multimodal medical and health data has continued to expand, which has laid the foundation for the development, verification, and new breakthroughs of big models in health-related fields.

This article first reviews the background of large AI models. Based on the modality of pre-training data, the article classifies large models into three categories: 1) Large Language Model (LLM); 2) Large Vision Model (LVM); 3) Large Multi-modal Model (LMM). The article also summarizes the four main characteristics of large models.

Figure 1 Key features of large AI models: 1) The number of parameters is huge. For example, for large language models (LLMs), the number of parameters can usually reach hundreds of billions or even more); 2) The scale of training data is huge. For example, for large visual models (LVMs), the amount of training data they use may contain billions of images; 3) Ability to process multimodal information; 4) Ability to adapt to different downstream tasks and demonstrate excellent generalization capabilities in various tasks, such as in zero-shot, one-shot, and few-shot tasks.

The article then discusses seven healthcare sub-sectors where big models may have a significant impact, including 1) bioinformatics; 2) medical diagnosis; 3) vision; 4) medical informatics; 5) medical education; 6) public health and 7) healthcare. It also compares the performance of typical big models and previous specialized models in the above seven fields in terms of model size and specific tasks.

Figure 2 Comparison of the number of parameters and performance on different tasks between the large model and previous specialist models.

The article also lists in detail the large-scale datasets in biomedical and health informatics, and summarizes key information such as the scale, modality, corresponding tasks, and whether they are open source of existing mainstream datasets.

Figure 3. Large-scale datasets in biomedical and health informatics

At present, although large-scale artificial intelligence models have shown very promising application prospects, their development and application in biomedicine, clinical, and medical fields still face many challenges and potential risks. The article also discusses data, credibility, privacy, fairness, transparency, explainability, sustainability, and norms.

Figure 4. Future development direction of artificial intelligence big models in health informatics

Regarding the future direction of big models in the field of healthcare, the article mainly focuses on performance and responsibility. From the perspective of performance, the article suggests that, first, better-performing big AI models should be developed for health informatics, for example, the scale of models and data sets can be expanded, and multi-modal and multi-task pre-training can be performed; second, the potential of existing large AI models can be tapped, such as through prompt engineering, fine-tuning, probing, etc., to further explore their capabilities that are not yet known to people. If these capabilities have been mastered by big models and are only temporarily unknown to people, then tapping them can provide an almost cost-free solution for the health field.

In addition, responsible AI big models are critical to society. The article proposes to develop and deploy complementary strategies to address the challenges of big models in terms of reliability, fairness, transparency, etc. The development strategy focuses on learning responsible big models, while the deployment strategy emphasizes the responsible use of big models.

Finally, the article points out that a paradigm shift is taking place in the AI ​​community, which is gradually affecting the development of AI models in the biomedical and health fields. The goal of the new paradigm is to learn a common basic model on large-scale (multimodal) datasets, covering various data distributions and learning tasks. The boundaries between different tasks and even different data modalities are being broken. As general intelligence and more unknown capabilities are explored and realized, it is believed that large AI models will better assist rather than replace medical professionals and practitioners in the future. Human-machine collaboration will be everywhere. The development of large AI models will require closer cooperation between experts in various fields, as well as the gradual establishment and improvement of corresponding large model supervision mechanisms.

Editor: Huang Fei