Stanford HAI Releases Comprehensive AI Index Report 2025 Highlighting Key Progress in Healthcare AI
- Association Secretary
- Apr 24
- 7 min read
On April 7, the Stanford Human-Centered AI Institute (HAI) published its highly anticipated "AI Index Report 2025," a comprehensive 456-page document that outlines significant advancements in the field of artificial intelligence over the past year. Established in 2019 and co-led by renowned AI expert Fei-Fei Li, the institute's annual reports have become essential resources for understanding cutting-edge developments in AI globally.
This year's report is structured into eight key chapters, covering areas such as research and development, technological performance, responsible AI, economic implications, science and medicine, policy and governance, and education. Notably, the section on science and medicine spans 43 pages, highlighting critical milestones in the AI-driven transformation of the healthcare industry.
Fei-Fei Li expressed her enthusiasm for the growing intersection of AI and healthcare, stating, "This is a vast industry, from fundamental biological research to clinical diagnosis, medical services, and public health. It's exciting to see AI play a crucial role at every stage of this system."
The optimistic outlook for AI in healthcare is echoed by industry leaders. Previously noted by influential figures like Cathie Wood and Oracle's CEO Larry Ellison, the consensus is that healthcare will be one of AI's most profound applications.
AI Index Report: Milestones in AI Healthcare in 2024
At the beginning of the "Science and Medicine" chapter, the report enumerates the milestone achievements of AI in the fields of healthcare and biology over the past year.
Protein Sequence Optimization: Language Models to the Rescue
Although it was not intended, language models have recently acquired a new biological capability: optimizing protein sequences. Traditionally, protein engineering requires extensive laboratory research to refine sequences for improved functionality. However, recent studies have found that un-tuned language models have become unusually effective at this task.
Researchers employed directed evolution methods to demonstrate that language models can generate protein sequences that outperform traditional algorithms in both synthetic and experimental fitness landscapes.
Aviary: Training Agents for Biological Tasks
FutureHouse, a company aimed at automating the scientific discovery process, released Aviary last year, a platform capable of executing multi-step tasks in biological research, with the potential to automate intellectual tasks. Under the Aviary framework, AI agents can already perform challenging scientific tasks such as manipulating DNA structures for molecular cloning, answering research questions through scientific literature review, and constructing protein stability models.
Results show that, with Aviary's support, agents built on open-source large language models can match or even exceed human experts on multiple tasks, paving the way for further automation in scientific exploration.
AlphaProteo: AI for New High-Affinity Protein Binders
AlphaProteo is a model developed by Google DeepMind, focusing on creating new high-affinity protein binders that can attach to specific target molecules.
AlphaProteo designed the first binders for several target proteins, including VEGF-A, associated with cancer and diabetes, demonstrating an effectiveness 300 times greater than existing methods across seven tested target proteins.
Additionally, the binders developed by AlphaProteo are about ten times more robust than those created using current state-of-the-art design methods, marking a significant breakthrough in bioengineering. This model is being utilized in drug development, diagnostics, and biotechnology applications.
Human Brain Atlas: Rebuilding a Part of the Human Brain
A team from Google has reconstructed a small portion of the human brain at the synaptic level—lauded by Wired magazine as "the most detailed brain connectome to date."
The sample was taken from the left anterior temporal lobe of an epilepsy patient during surgery, employing a multi-beam scanning electron microscope. Over 5,000 slices, each about 30 nanometers thick, captured approximately 57,000 cells—comprising neurons, glial cells, and blood vessels—as well as 150 million synapses.
This project represents a significant step in understanding neural circuits and may provide insights for future neural therapies.
Virtual AI Laboratory: Enhancing Biomedical Research
Stanford University has recently introduced a virtual AI laboratory, where multiple "AI scientists" focus on different disciplines and collaborate as autonomous agents.
In experiments, human researchers tasked this AI lab with designing nano-antibody fragments that could bind to viruses, resulting in the generation of 92 nano-antibodies, over 90% of which successfully bound to the virus in validation studies.
The significance of this research lies not in the specific findings but in demonstrating that a fully autonomous, language model-driven lab can produce meaningful scientific discoveries.
GluFormer: Using AI for Continuous Glucose Monitoring
GluFormer is a foundational model developed by Nvidia, the Weizmann Institute, and other institutions, trained on over 10 million glucose measurements from nearly 11,000 individuals, predicting long-term health outcomes by analyzing Continuous Glucose Monitoring (CGM) data.
GluFormer can identify individuals at risk of diabetes or deteriorating blood glucose control before symptoms appear. In a 12-year study involving 580 adults, it accurately flagged 66% of new diabetes cases and 69% of cardiovascular-related deaths. Models like GluFormer will shift diabetes care from passive treatment to proactive prevention, enabling earlier clinical interventions.
ESM3: Simulating Evolution to Generate New Proteins
The ESM3 model from EvolutionaryScale is groundbreaking, trained on 2.78 billion protein sequences and equipped with 98 billion parameters, aimed at generating new protein types by simulating evolutionary processes. One of ESM3's most significant achievements is the design of esmGFP, a new artificial green fluorescent protein estimated to take nature 500 million years to develop.
This was achieved through human-guided thought chain prompts.
ESM3 is also open-source, facilitating collaboration in synthetic biology and protein engineering projects, with applications in drug discovery, materials science, and environmental engineering.
AlphaFold3: Predicting the Structure and Interactions of All Life Molecules
The latest achievement from Google DeepMind and Isomorphic Lab—AlphaFold 3—not only predicts protein structures but also more accurately simulates their interactions with key biological molecules (DNA, RNA, ligands, antibodies).
By modeling the interactions between small molecules and proteins, AlphaFold 3 accelerates drug development, crucial for disease research. Additionally, its open-source nature empowers scientists globally.
AI Sweeping Through Healthcare: From Research and Development to Clinical Applications
Proteins are at the core of AI's impact on life sciences, and over the past year, AI models applied to protein sequences have achieved significant progress.
A clear trend is the increasing scale of model training, with AI-driven methods reducing reliance on costly and time-consuming experimental approaches, thus accelerating the exploration of protein functions and designs.
During this process, the number of entries in various public protein science databases has steadily increased over time. These databases have become indispensable tools for researchers; however, maintaining data quality and preventing model bias remains a continuous challenge.
Currently, the number of AI-driven protein research studies has significantly increased, focusing on key areas such as function prediction (8.4%), protein structure prediction (7.6%), and protein-drug interactions (3.0%).
Moreover, image and multimodal AI applications are also being utilized in scientific discoveries. Advances in techniques such as cryo-electron microscopy, high-throughput fluorescence microscopy, and whole-slide imaging have enabled scientists to accurately examine and analyze atomic, subcellular, and tissue-level structures, revealing new insights into complex biological processes.
In the drug development process, clinical trials are the most time-consuming and costly phase. In recent years, the number of AI-enhanced clinical trials has been steadily increasing.
Last year, the number of AI clinical trials in China reached 105, surpassing the United States (97) to rank first in the world, followed by Italy (42), Turkey (30), and the United Kingdom (24).
In the clinical realm, AI applications in medical imaging are advancing rapidly, extending to new data modalities and addressing increasingly complex clinical issues.
Statistics indicate that over 80% of FDA-approved machine learning software pertains to medical image analysis. Recently, there has been a dramatic increase in large models in the medical imaging field, particularly in pathology, where the number of foundational models has significantly grown.
However, despite some achievements in this area, many AI applications in medical imaging still rely on highly limited training datasets, especially in the field of three-dimensional imaging.
To train more robust medical imaging AI models, larger, more comprehensive, and diverse datasets are required. By enhancing the availability of high-quality, well-annotated training data, it is expected that model performance will improve.
Additionally, the rise of large language models in clinical practice has brought transformations in medical decision-making, auxiliary diagnosis, patient care, and medical record management.
A representative case is the significant improvement in AI performance on the MedQA benchmark test, which includes over 60,000 clinical questions.
Recently, research teams from Microsoft and OpenAI tested the model o1, achieving a new high score of 96.0%, which is a 5.8-point increase over the 2023 record. Since the end of 2022, benchmark performance has improved by 28.4 points.
Industry interest in language models for healthcare tasks has surged, as evidenced by the 1,566 results for the term "large language models" in PubMed, with 1,210 papers published just in 2024.
Currently, large language models have demonstrated robust capabilities in clinical decision-making. A randomized controlled trial involving 92 doctors showed that the decision-making level of GPT-4 operating independently was comparable to that of human doctors using GPT-4 for assisted decision-making.
Clinical documentation has long been a significant burden for physicians, and AI is now addressing this challenge.
Researchers at Stanford University conducted a fully integrated automated AI dictation system test, showing a physician adoption rate of up to 55%. AI dictation saves approximately 30 seconds per record, with physicians' burden and fatigue decreasing by an average of 35% and 26%, respectively.
These findings indicate that AI-driven dictation technology can significantly improve physicians' workflows and well-being, saving time and alleviating administrative burdens.
The report further states that the integration of AI into electronic health record systems can streamline clinical workflows, enhance the experiences of healthcare providers and patients, and reduce the overall burden on healthcare systems. Major vendors—Epic, Oracle Health, Meditech, and TruBridge—have widely adopted AI solutions.
In recent years, the significant impact of AI in healthcare is also reflected in the number of FDA approvals. Statistics show that since 2015, the number of AI medical devices approved by the FDA has exhibited exponential growth, reaching 223 by 2023.
In summary, the report underscores that advancements in AI technology are catalyzing a profound revolution within the healthcare sector, paving the way for innovative therapies and improved patient outcomes. As AI continues to evolve, its capabilities promise to transform the landscape of medicine and healthcare delivery for years to come.