The Google Research team has been on a mission for several years, striving to unravel solutions to what seem like insurmountable challenges in various fields including quantum computing and predicting floods.
Recently, however, a blog post by the company discussed how a significant part of their focus has been directed toward healthcare, aiming to enhance the capabilities of radiologists in detecting breast cancer and aiding the screening process for diabetic retinopathy. They believe that artificial intelligence (AI) holds the potential to spur the development of solutions that are more personalized, accessible, and effective within this domain.
AI Systems for Healthcare
Engaging in collaborative efforts with industry partners, researchers, and healthcare providers, Google Research has been at the forefront of developing AI systems, releasing open-source tools, and publishing research, all with the potential to enhance health outcomes on a global scale. This report provides insights into how AI is revolutionizing healthcare, the journey of turning research into practical applications, and the learning experiences garnered along the way as they work towards creating next-generation solutions.
In their exploration of generative AI’s potential to revolutionize healthcare, the team at Google Research developed a large language model (LLM) last year with the aim of achieving a passing score in U.S. Medical License Exam-style questions. Utilizing Google’s LLM technology, PaLM, they fine-tuned it to create Med-PaLM, achieving state-of-the-art performance in medical question-answering tasks. However, they didn’t stop there. Despite Med-PaLM and Med-PaLM 2 addressing a major grand challenge in AI, the team’s ambition extends to making a bold and responsible impact on healthcare.
ELIXR
Continuing their research and development of models, they have been rigorously evaluating generative AI’s capabilities in medicine, focusing on improving the safety, accuracy, and equity of these tools. The team introduced ELIXR, a novel methodology integrating LLMs with medical imaging models, resulting in specialized multimodal models. Moreover, they are actively involving clinicians in evaluating Med-PaLM’s performance, ensuring transparency by publicizing their findings. Current efforts involve collaborating with partners via Google Cloud to investigate how these medically adapted models can bring transformative changes to healthcare.
The published research from Google’s team demonstrates the potential of AI in aiding medical professionals, from identifying signs of breast cancer in mammograms to contouring organs for radiotherapy. Yet, the full realization of this potential hinges on clinical validation and practical application, without which these findings remain theoretical.
Over recent years, the team has recognized the critical importance of partnering with healthcare organizations to validate their research in real-world clinical settings. Collaborations are in place with entities such as Jacaranda Health in Kenya, aiming to enhance fetal ultrasound AI models, and with Osaka University to refine dermatology classifiers. Additionally, partnerships with a range of healthcare organizations are helping to pinpoint the most effective applications of Med-PaLM 2 technology, ensuring AI supports actual clinical workflows.
Furthermore, the team acknowledges that partnerships are crucial for scaling their initiatives and enhancing global health, particularly in genomics. Here, they are collaborating with PacBio, a company specializing in genome sequencing instruments, working to improve genomic analysis through DeepVariant and DeepConsensus.
In their journey of building and deploying AI-powered health solutions, the team has encountered numerous challenges and complexities, leading to valuable learnings. They emphasize the need for AI systems to immediately demonstrate value and seamlessly integrate with existing health systems, aiming to alleviate rather than exacerbate the technological burdens faced by healthcare organizations. Their commitment extends to sharing their knowledge on effectively and safely deploying health AI solutions, and they have made a conscious effort to open-source technologies such as Open Health Stack and CXR Foundation, empowering others to develop next-generation digital health solutions.
Ultimately, through a blend of foundational and applied research in health AI, Google Research aspires to guide the way toward a future in healthcare that stands out in terms of accessibility, accuracy, and equity.
Featured image: Credit: Surender, PlaygroundAI
