List of Artificial Intelligence Models for Medical Landscape (2023)

Given the number of advancements artificial intelligence (AI) has made this year itself, it’s no surprise that it has been a significant point of discussion throughout 2023. AI now finds its use case in almost every realm, and one of its exciting and useful applications is in healthcare and medicine. From drug discovery to transcribing medical documents and even assisting in surgeries, it is transforming medical professionals’ lives and even helps reduce errors and improve their efficiency. This article discusses a few AI models of 2023 that have the capability to transform the medical landscape.

Med-PaLM has been designed by Google Research for the medical domain and is capable of giving high-quality answers to medical questions. The model leverages the power of Google’s LLMs and is one of the first models to achieve a human expert level when answering USMLE-style questions. When evaluated, the model demonstrated the ability to understand symptoms, perform complex reasoning, and choose the appropriate treatment. Moreover, it achieved an 86.5% accuracy on the MedQA medical exam benchmark in research. Although it shows promising capabilities, the researchers want to conduct more rigorous assessments to ensure that the model can be deployed in safety-critical domains.

Bioformer is a compact version of BERT that can be used for biomedical text mining. Although BERT has achieved state-of-the-art performance in NLP applications, its parameters could be reduced with a minor impact on performance to improve its computational efficiency. Bioformer researchers have taken this approach to develop a model whose model size is significantly smaller than that of BERT (60% less). The model was trained on PubMed abstracts and PubMed Central full-text articles and uses a biomedical vocabulary. The researchers have released two versions of the model – Bioformer8L and Bioformer16L-and both performed well even with fewer parameters when evaluated on parameters like named entity recognition, relation extraction, question answering, and document classification.

MedLM is a suite of foundational models developed by Google that have been fine-tuned for healthcare use cases. Two models under MedLM have been designed for dealing with complex tasks and scaling across tasks. The main purpose of these models is to automate tasks to save time, increase efficiency, and improve overall patient health, and the researchers at Google have collaborated with Deloitte to pilot MedLM’s capabilities. MedLM has also been integrated with other AI systems like ASCEND of BenchSci to improve the quality and speed of clinical research and development.

RoseTTAFold is a deep-learning powered software that predicts protein structures just from limited information. It is capable of studying the pattern in protein sequences, the interaction of proteins’ amino acids, and their 3D structure. The model allows researchers to model the way proteins and small-molecule drugs interact with one another, which facilitates drug discovery research. The researchers of the model have also made its code public to benefit the entire community.

AlphaFold is a powerful AI model developed by DeepMind that can predict the 3D structure of the protein from its amino acid sequence. DeepMind has partnered with EMBL’s European Bioinformatics Institute (EMBL-EBI) to release a database containing more than 200M AI-generated protein structure predictions to facilitate scientific research. In CASP14, AlphaFold outperformed the other models by a significant margin, producing results with high accuracy. Additionally, it has the potential to better help researchers understand protein structures and advance biological research.

ChatGLM is a bilingual model (Chinese-English) that has been fine-tuned on a database of medical dialogues in Chinese. The model was fine-tuned in a rather short time (13 hours), making it a very affordable healthcare-purpose LLM. The model also has a longer sequence length, and thus, it supports longer conversations and applications. The model has been trained using techniques like supervised fine-tuning, RLHF, etc., which allows it to understand human instructions better. As a result, the model has excellent dialogue and question-answering capabilities.