AI in Healthcare & Biology: Empowering the Future of Medicine
AI in medicine involves the application of artificial intelligence techniques to enhance various aspects of healthcare. It encompasses tasks like medical image analysis, disease prediction, personalized treatment recommendations, drug discovery, and more. AI's capacity to analyze vast datasets and recognize patterns holds great promise for improving diagnostics, treatment, and patient outcomes, ultimately advancing the field of medicine.
Courses
-
AI for Medicine Specialization (opens in a new tab) by DeepLearning AI, offers practical experience in applying machine learning to medical challenges. Beyond foundational deep learning, it covers nuanced topics, including treatment effect estimation, diagnostic and prognostic model interpretation, and natural language processing for unstructured medical data. Completing the specialization equips learners with a diverse skill set spanning model interpretation, image segmentation, and more. This specialization comprises three courses. AI for Medical Diagnosis, AI for Medical Prognosis and AI For Medical Treatment.
-
AI in Healthcare Specialization (opens in a new tab) by Stanford University, cover AI basics, medical imaging, working with EHR data, genomics and precision medicine, how to use AI to enhance diagnosis, treatment, delivery of care and examining the impact of AI on society. Introduction to Healthcare, Introduction to Clinical Data, Fundamentals of Machine Learning for Healthcare, Evaluations of AI Applications in Healthcare and AI in Healthcare Capstone.
Explainers
-
DeepMind solves protein folding (opens in a new tab) by Lex Fridman, celebrates DeepMind's AlphaFold 2 breakthrough in protein folding, a long-standing R&D challenge in biochemistry. It elucidates the significance of protein folding, highlighting AlphaFold 2's application of deep learning for precise 3D structure prediction.
-
DeepMind's AlphaFold 2 Explained! AI Breakthrough in Protein Folding! What we know (& what we don't) (opens in a new tab) by Yannic Kilcher, emphasizes the critical role of protein folding in biology. It introduces the AlphaFold algorithm, which employs deep learning to predict protein structures accurately. The video also explores how this breakthrough could benefit R&D in drug discover.
-
LLMs in Healthcare: Benchmarks, Applications, and Compliance (opens in a new tab): This session discusses the progress and challenges of LLMs in healthcare. It highlights benchmarks for healthcare-specific LLMs, a robust architecture for medical chatbots, and comprehensive testing for bias, fairness, robustness, and toxicity.
-
LLMs Encode Clinical Knowledge (opens in a new tab): Researchers are leveraging LLMs for medical applications, introducing MultiMedQA to evaluate their capabilities. Flan-PaLM, a 540-B parameter LLM, exhibits impressive accuracy on multiple-choice datasets, hinting at LLM potential in medicine. (paper)
LLMs in Healthcare
-
Med-PaLM 2 (opens in a new tab) by Google Research, is an AI model tailored for the medical domain, available for limited testing by select Google Cloud customers. It excels in USMLE-style questions, reaching 'expert' performance on the MedQA dataset, and handling diverse biomedical data types. Healthcare organizations like HCA Healthcare, Mayo Clinic, and Meditech have tested it to augment healthcare workflows.
-
ClinicalBERT (opens in a new tab) trained on a vast dataset containing 1.2B words related to various diseases. Fine-tuning utilized EHRs from 3 million patient records. ClinicalBERT, initiated from BERT, employs a masked language model training approach. Text segments with randomly replaced tokens [MASK] challenge the model to predict the original tokens in a contextual context.
-
Gatortron: The Biggest Clinical Language Model (opens in a new tab), a clinical language model by the University of Florida & NVIDIA, uses over 90-B words from clinical notes, PubMed, and Wikipedia. With 8.9-B parameters, it excels in clinical NLP tasks, showing improved performance in concept extraction, relation extraction, textual similarity, inference, and medical question answering. As the largest transformer in clinical domains, GatorTron enhances healthcare by understanding and utilizing patient information from EHRs.
Advancements
-
AlphaFold (opens in a new tab) by DeepMind, revolutionizes protein structure prediction, using deep learning to decipher the 3D shapes of proteins with remarkable accuracy. Its capabilities hold immense promise for understanding diseases and drug development, making it a game-changer in the field of bioinformatics and molecular biology. (code)
-
Evolutionary Scale Modeling (opens in a new tab): ESM uses AI to predict protein structures through evolutionary data embeddings. Trained on millions of protein sequences, ESM aids in COVID-19 evolution prediction and disease genetic discovery. Key features include direct structure prediction, atomic-level precision, and leading performance with ESM-2. (blog)
-
RoseTTAFold (opens in a new tab): By the Baker lab at the University of Washington, utilizes deep learning to swiftly predict protein structures in just ten minutes on a gaming computer. This "three-track" neural network considers protein sequences, amino acid interactions, and potential 3D structures simultaneously. Applied globally, it has solved long-standing biological problems, including understanding poorly known proteins and modeling complex biological assemblies. The latest version, RF Diffusion, enhances functional protein prediction with a diffusion model.
-
OmegaFold (opens in a new tab): A high-res De novo protein structure prediction tool by HeliXonProtein, utilizes deep learning and a LLM for accurate protein structure predictions from primary sequences. Open-source and versatile, it aids in applications like COVID-19 evolution prediction and identifying genetic disease causes. (paper)
-
ESMFold (opens in a new tab): Meta AI's protein prediction tool, employs Evolutionary Scale Modeling (ESM) and a trained LLM to predict a protein's 3D structure from its amino acid sequence. With the ESM-2 model, it swiftly predicts one million biological sequences in under a day. Integrated into ChimeraX, ESMFold is a powerful tool for protein structure forecasting. (blog)
-
AlphaMissense (opens in a new tab): DeepMind's AlphaMissense AI tool accurately categorizes 71 million human DNA 'missense' mutations as likely pathogenic or benign with 89% accuracy. This accelerates molecular biology research, aids rare disorder diagnosis, and contributes to genetic disease treatment development by predicting the impact of genetic mutations. (paper in Science) (code)
Articles & Papers
-
AI in Healthcare (opens in a new tab): Gain profound insights into how AI is revolutionizing the healthcare industry, from diagnosis to treatment, research, and patient care.
-
AI for healthcare (opens in a new tab) at Imperial College London plays a pivotal role in the integration of AI into the healthcare sector. The institution focuses on Perceptual AI and Intervention AI with the goal of enhancing healthcare, improving cost-effectiveness, and ensuring accessibility, all while prioritizing human-machine collaboration.
-
Artificial Intelligence in Biological Sciences (opens in a new tab) explores AI's applications in healthcare and radiology. AI aids efficient billing, promotes NLP for physician notes, and automates pattern recognition in radiology.
-
Strategies for pre-training graph neural networks (opens in a new tab): This publication established the foundation for efficient pre-training techniques applicable to various aspects of drug discovery, including forecasting molecular properties and deducing protein functions. (article)
-
Artificial intelligence in radiology (opens in a new tab): This paper explores the use of deep learning techniques in medical image analysis, including image segmentation, classification, and detection. It discusses the challenges and future directions of deep learning in healthcare.
-
Overview of artificial intelligence in medicine (opens in a new tab) delves into AI's potential in healthcare and emphasizes the significance of collaboration and standardization in AI research. It accentuates AI's transformative impact on healthcare while acknowledging concerns regarding human replacement. Collaboration and data standardization are imperative for efficient and ethical AI deployment.
Books
- Deep Medicine: How Artificial Intelligence Can Make Healthcare Human Again (opens in a new tab) by Dr. Eric Topol explores how AI can revitalize healthcare, offering insights into its role in diagnosis, treatment, and patient care. It envisions AI's potential to humanize and improve healthcare.
Refrence
- AlphaFold Protein Structure Database (opens in a new tab) offers open access to 200+ million protein structure predictions, expediting scientific research. This valuable data is accessible for academic and commercial purposes under the Creative Commons Attribution 4.0 (CC-BY 4.0) license terms.
Medical datasets
Medical datasets are a valuable resource, driving innovations such as medical chatbots for instant assistance and fine-tuning open-source LLMs for enhanced performance in medical tasks. These datasets enable the creation of medical apps, ultimately improving healthcare efficiency and accuracy.