Difference between revisions of "SOCR News GenAI RD March 2023"
(→Overview) |
|||
| Line 1: | Line 1: | ||
== [[SOCR_News | SOCR News & Events]]: SOCR R&D in Generative AI (March 2023) == | == [[SOCR_News | SOCR News & Events]]: SOCR R&D in Generative AI (March 2023) == | ||
| − | == | + | |
| + | ==Present state and future directions of foundational AI models== | ||
[[Image:SOCR_AI_Tools_Schematic.png|150px|thumbnail|right| [https://socr.umich.edu SOCR Resource] ]] | [[Image:SOCR_AI_Tools_Schematic.png|150px|thumbnail|right| [https://socr.umich.edu SOCR Resource] ]] | ||
| − | === | + | ===Current status-Quo=== |
| − | |||
| − | |||
: Indeed, the speed of scientific, technological, and innovative discoveries is increasing (faster than a simple linear trend). This is certainly not new – this trend started 500 years ago during the Renaissance and Enlightenment periods and continuous to this date. The process is likely to move more expeditiously throughout the 21st century. There are many, many AI variants. Their specific biomedicine and healthcare applications represent just one or a wide range of domains impacted by the rapid (evolutionary, but not yet revolutionary) changes. Disruptive innovations are profoundly impacting all aspects of human experiences, far beyond health. In a way, this process of quick technological immersion impacts “everything we do (humans experiences), all the time (longitudinal trajectory), everywhere (geo-spatially), and all at once (simultaneously).” | : Indeed, the speed of scientific, technological, and innovative discoveries is increasing (faster than a simple linear trend). This is certainly not new – this trend started 500 years ago during the Renaissance and Enlightenment periods and continuous to this date. The process is likely to move more expeditiously throughout the 21st century. There are many, many AI variants. Their specific biomedicine and healthcare applications represent just one or a wide range of domains impacted by the rapid (evolutionary, but not yet revolutionary) changes. Disruptive innovations are profoundly impacting all aspects of human experiences, far beyond health. In a way, this process of quick technological immersion impacts “everything we do (humans experiences), all the time (longitudinal trajectory), everywhere (geo-spatially), and all at once (simultaneously).” | ||
: AI is advancing in multiple directions. For instance, in the SOCR lab at UMich, we are driving forward on three complementary fronts. First, utilizing foundational AI models to generate realistic synthetic medical text, e.g., physician/nurse clinical notes. There are many reasons that drive these developments, from an automated holistic interpretation of complex and space-time-varying personal clinical notes, to training future AI models without compromising sensitive data. Second, using generative-AI to simulate 2D medical images, which are very expensive and may unintentionally disclose sensitive personal information. And third, we utilize AI to synthetically generate computer code, computational scripts, algorithm implementations, and end-to-end protocols. This is critical in the cyclical process of efficient software development, testing, validation, analysis, redesign, revalidation, and productization. Ultimately, we may be training AI models to reproduce themselves (AI self-guided evolution), but with desired traits, e.g., enhanced optimality, improved scalability, better performance, and higher efficiency. | : AI is advancing in multiple directions. For instance, in the SOCR lab at UMich, we are driving forward on three complementary fronts. First, utilizing foundational AI models to generate realistic synthetic medical text, e.g., physician/nurse clinical notes. There are many reasons that drive these developments, from an automated holistic interpretation of complex and space-time-varying personal clinical notes, to training future AI models without compromising sensitive data. Second, using generative-AI to simulate 2D medical images, which are very expensive and may unintentionally disclose sensitive personal information. And third, we utilize AI to synthetically generate computer code, computational scripts, algorithm implementations, and end-to-end protocols. This is critical in the cyclical process of efficient software development, testing, validation, analysis, redesign, revalidation, and productization. Ultimately, we may be training AI models to reproduce themselves (AI self-guided evolution), but with desired traits, e.g., enhanced optimality, improved scalability, better performance, and higher efficiency. | ||
| − | + | : ''First generation AI'' including automated monitoring systems and on-demand data processing tools are already in wide-use. ''Second generation AI'' are either FDA approved or in the process of being approved, but not yet widely in clinical use. The ''third-gen AI'' will provide more autonomous features, such as constantly monitoring the electronic health record and any ongoing clinical assessment protocol, diagnostic routine, treatment regiment, or ongoing medical procedure, and provide on-the-fly “holistic” clinician-assistance with derived-inference, which supplements data elements already accounted for by health professionals. Mind that this does not include in any way eliminating or reducing the complete human control over healthcare practices. | |
| − | + | ===Expectations for the future: AI investments & the delicate balance between safety (security) and risk (utility)=== | |
: We have been building electric cars in Michigan since early 1900’s, but they did not really materialize as practical mobility vehicles until just a few years ago. Similarly, self-driving cars have been designed, tested, and deployed with limited scope for decades, yet, these are not yet widely used (at least in their intended full-self-driving mode). The very same risk-benefit principle governs the utilization of AI in healthcare. There is a tug of war between the incredible potential benefits (e.g., cost-effective clinical decision support systems, extra-human clinical abilities, augmentation of medical expert knowledge with almost limitless, quick, and highly effective machine storage and rational judgement) and any potential drawbacks (e.g., risks of misuse, accidental leaks of sensitive information, rare but potentially devastating AI mistakes, etc.) The push-pull dance balancing the strong demand for rapid AI immersion and the trepidation over potential risks of its ubiquitous embedding in clinical practice will continue for some time. | : We have been building electric cars in Michigan since early 1900’s, but they did not really materialize as practical mobility vehicles until just a few years ago. Similarly, self-driving cars have been designed, tested, and deployed with limited scope for decades, yet, these are not yet widely used (at least in their intended full-self-driving mode). The very same risk-benefit principle governs the utilization of AI in healthcare. There is a tug of war between the incredible potential benefits (e.g., cost-effective clinical decision support systems, extra-human clinical abilities, augmentation of medical expert knowledge with almost limitless, quick, and highly effective machine storage and rational judgement) and any potential drawbacks (e.g., risks of misuse, accidental leaks of sensitive information, rare but potentially devastating AI mistakes, etc.) The push-pull dance balancing the strong demand for rapid AI immersion and the trepidation over potential risks of its ubiquitous embedding in clinical practice will continue for some time. | ||
: Two examples of specific cutting-edge AI advances we’ll be working towards in the next 5-10 years include (1) tight integration of effective and efficient AI education, foundational AI research, and ethical and trustworthy AI implementation and immersion in practice; and (2) developing robust generative-AI for simulating 3D medical data (e.g., MRI or CT volumes) and hyper-volumes (e.g., 4D fMRI tensors) representing holistic biological states across space, time, demographic states, and clinical phenotypes. | : Two examples of specific cutting-edge AI advances we’ll be working towards in the next 5-10 years include (1) tight integration of effective and efficient AI education, foundational AI research, and ethical and trustworthy AI implementation and immersion in practice; and (2) developing robust generative-AI for simulating 3D medical data (e.g., MRI or CT volumes) and hyper-volumes (e.g., 4D fMRI tensors) representing holistic biological states across space, time, demographic states, and clinical phenotypes. | ||
| + | |||
| + | == Details== | ||
| + | |||
| + | * [https://rcompute.nursing.umich.edu/SOCR_AI_Bot/ SOCR AI Bot (Live-app)] | ||
| + | * [https://docs.google.com/presentation/d/1HCCWtRaqtPVMLAAcTiO6ankF6yLdLu0NUHxhpSfHIDg/edit#slide=id.p1 Slidedeck] | ||
| + | * Videos: [https://drive.google.com/file/d/1C-HsLraMi_TE0n0o7MlCA2_zv6Kpljqw/view?usp=sharing (5-min) AI-in-Health] & [https://drive.google.com/file/d/1GS3ohrIgy0eJ1qfdTe7L2m0ShNUEmOSQ/view?usp=sharing (15-min) Pressure-Injury Case-Study]. | ||
==Cross-References== | ==Cross-References== | ||
| Line 30: | Line 35: | ||
* [https://www.socr.umich.edu/html/SOCR_UserGoogleMap.html SOCR Global Users] | * [https://www.socr.umich.edu/html/SOCR_UserGoogleMap.html SOCR Global Users] | ||
| − | [[Image:SOCR_Retreat_Fall_2019_Team_P2.png| | + | [[Image:SOCR_Retreat_Fall_2019_Team_P2.png|400px|thumbnail|center| [http://socr.umich.edu/people SOCR Team] ]] |
| − | [[Image:SOCR_Team_2020_Pic.png| | + | [[Image:SOCR_Team_2020_Pic.png|400px|thumbnail|center| [http://socr.umich.edu/people SOCR Team] ]] |
Revision as of 12:58, 26 April 2023
Contents
SOCR News & Events: SOCR R&D in Generative AI (March 2023)
Present state and future directions of foundational AI models
Current status-Quo
- Indeed, the speed of scientific, technological, and innovative discoveries is increasing (faster than a simple linear trend). This is certainly not new – this trend started 500 years ago during the Renaissance and Enlightenment periods and continuous to this date. The process is likely to move more expeditiously throughout the 21st century. There are many, many AI variants. Their specific biomedicine and healthcare applications represent just one or a wide range of domains impacted by the rapid (evolutionary, but not yet revolutionary) changes. Disruptive innovations are profoundly impacting all aspects of human experiences, far beyond health. In a way, this process of quick technological immersion impacts “everything we do (humans experiences), all the time (longitudinal trajectory), everywhere (geo-spatially), and all at once (simultaneously).”
- AI is advancing in multiple directions. For instance, in the SOCR lab at UMich, we are driving forward on three complementary fronts. First, utilizing foundational AI models to generate realistic synthetic medical text, e.g., physician/nurse clinical notes. There are many reasons that drive these developments, from an automated holistic interpretation of complex and space-time-varying personal clinical notes, to training future AI models without compromising sensitive data. Second, using generative-AI to simulate 2D medical images, which are very expensive and may unintentionally disclose sensitive personal information. And third, we utilize AI to synthetically generate computer code, computational scripts, algorithm implementations, and end-to-end protocols. This is critical in the cyclical process of efficient software development, testing, validation, analysis, redesign, revalidation, and productization. Ultimately, we may be training AI models to reproduce themselves (AI self-guided evolution), but with desired traits, e.g., enhanced optimality, improved scalability, better performance, and higher efficiency.
- First generation AI including automated monitoring systems and on-demand data processing tools are already in wide-use. Second generation AI are either FDA approved or in the process of being approved, but not yet widely in clinical use. The third-gen AI will provide more autonomous features, such as constantly monitoring the electronic health record and any ongoing clinical assessment protocol, diagnostic routine, treatment regiment, or ongoing medical procedure, and provide on-the-fly “holistic” clinician-assistance with derived-inference, which supplements data elements already accounted for by health professionals. Mind that this does not include in any way eliminating or reducing the complete human control over healthcare practices.
Expectations for the future: AI investments & the delicate balance between safety (security) and risk (utility)
- We have been building electric cars in Michigan since early 1900’s, but they did not really materialize as practical mobility vehicles until just a few years ago. Similarly, self-driving cars have been designed, tested, and deployed with limited scope for decades, yet, these are not yet widely used (at least in their intended full-self-driving mode). The very same risk-benefit principle governs the utilization of AI in healthcare. There is a tug of war between the incredible potential benefits (e.g., cost-effective clinical decision support systems, extra-human clinical abilities, augmentation of medical expert knowledge with almost limitless, quick, and highly effective machine storage and rational judgement) and any potential drawbacks (e.g., risks of misuse, accidental leaks of sensitive information, rare but potentially devastating AI mistakes, etc.) The push-pull dance balancing the strong demand for rapid AI immersion and the trepidation over potential risks of its ubiquitous embedding in clinical practice will continue for some time.
- Two examples of specific cutting-edge AI advances we’ll be working towards in the next 5-10 years include (1) tight integration of effective and efficient AI education, foundational AI research, and ethical and trustworthy AI implementation and immersion in practice; and (2) developing robust generative-AI for simulating 3D medical data (e.g., MRI or CT volumes) and hyper-volumes (e.g., 4D fMRI tensors) representing holistic biological states across space, time, demographic states, and clinical phenotypes.
Details
- SOCR AI Bot (Live-app)
- Slidedeck
- Videos: (5-min) AI-in-Health & (15-min) Pressure-Injury Case-Study.
Cross-References
- SOCR Website
- MDP and 2023 SOCR-MDP Program
- SOCR R&D Projects and SOCR MDP'23 Projects
- SOCR Navigators
- SOCR Datasets and Challenging Case-studies
- Electronic Textbooks: Probability and Statistics EBook, Scientific Methods for Health Sciences, Data Science and Predictive Analytics (DSPA)
- Hands-on interactive visualization of extremely high-dimensional data (learning module and webapp)
- SOCR News & Events
- SOCR Global Users
- SOCR Home page: https://www.socr.umich.edu
Translate this page:
