Thanks @dollendorf for sharing these insights.
Hi @DidierC, @janansmith, @ClaireM, @mendezra2, @marschenrj, @RickyM, @clairecravero, @acavaco, @care2communities and @C_Castellaz - What do you think on the way AI and ML will be used in the future for population health management and evaluating health programs.
Thanks @dollendorf for sharing these insights.
@Shashi AI and ML, as @synhodo shared, have the potential for huge impact in LMIC’s for disease screening and diagnosis, that leads to early detection and provides access to quality healthcare for rural and underserved communities. One of the biggest challenges to clinician training is the exponential growth of medical knowledge, which currently doubles every few months. AI can help aggregate this knowledge and get it to where it’s needed most. We at Project ECHO see AI playing an increasingly important role in this respect, and in combination with the ECHO Model, helping to democratize knowledge, get it to where it’s needed most, and save lives.
Patients have been swarming to telehealth amid the pandemic—and telehealth vendors have responded by diversifying their offerings to win over customers and keep them happy.
Telemedicine is the provision of medical services to patients remotely when the doctors and patients are physically separated using the two-way voice and visual communication. Modern technology has enabled doctors to consult patients by the HIPAA compliant video conferencing tools powered by satellite tech. Consequently, Telemedicine is already playing a huge role in cost reduction in wellness.
Thanks @SArora for sharing an important perspective of AI helping to democratize knowledge. In project ECHO, have you’ll been able to use AI and ML for data standardization, integration and systems interoperability? if so, what considerations were taken to protects privacy and security of these data?
Hey @Shashi, we’re using new forms of evolutionary ML and adversarial AI-- *not * Generative Adversarial Networks (GANs), but a completely different form of adversarial AI we developed a couple of years before GANs, which has significant operational advantages over the status quo.
Conventional ML is based on artificial neural networks (ANN), which suffer from inherent limitations (the main one being that they never really explicitly ‘model’ the dynamics of a system, but merely come to ‘embody’ those dynamics of the system). In contrast, our form of evolutionary ML takes the best abstract mathematical models of a disease from the medical literature, including its dynamics-- the way it alters the function of the body, and interacts with drugs–and embodies this structure as simulated “chromosomes”. Every “gene” on the chromosome corresponds with some parameter associated with the disease or drug-disease interaction. We then take individual medical histories and use evolutionary pressures to force the chromosomes to evolve-- to select and breed and mutate, across generations-- until organ-scale personalized models have been produced that embody the disease dynamics in the individual subject. (This is an idea that has been around for decades in a form originally known as “genetic algorithms”, GAs, but GAs had fundamental limitations that prevented their real-world use for complex diseases and partial information. By re-thinking models of evolution and adding additional features we’ve solved those limitations, and so now can apply our modified evolutionary process to model complex real-world diseases.)
One of the problems that remains is ambiguity in the evolved computational models. We use a novel adversarial process, whereby the main AI has two subordinate AIs take those models and play a game with them. One subsidiary AI, the “Prime”, takes the models and uses them to try to design an appropriate drug therapy. The other AI, the “Adversary”, uses known ambiguities within the models to cause adverse outcomes, which the Prime then has to counteract by modifying its strategy. (Unlike ANN, our AI “knows what it knows”, including ambiguities among the chromosomes.) The result is ultimately generation of a drug dosage strategy that is personalized, and has been tested for safety and efficacy. This adversarial process was first demonstrated in 2012 and patented in mid-2014, six months prior to GANs. It has major advantages over existing AI-based techniques, in that it is completely explainable, interactive with the users (patient and clinician), able to generate good strategies despite a training history of therapeutic failures, and is able to operate the AI component on an isolated Edge device.
For the IBM Watson AI XPRIZE we demonstrated this technology to generate insulin strategies for highly-unstable forms of type-1 diabetes by data-mining actual medical histories, despite those histories being comprised solely of therapeutic failures to achieve the desired blood glucose outcomes. Last year, as part of AI XPRIZE, we demonstrated our adversarial AI running on an isolated laptop computer, generating 30 hours of insulin strategy over approximately 20 minutes of computing time, that generated far better blood glucose outcomes than the original histories. The generated insulin strategies look nothing like conventional insulin dosing.
This suggests that the medical use of AI is about to look very different from previous use cases.
@Shashi IMHO ML/AI will dominate the clinical decision support domain, simply because the uderlying infrastructure can digest information both vertically and horizonatlly and provide a personalized and integrative perspective. Something only a very few physicians are capable of. This will democratize access to high quality healthcare
Thank you @MachineGenes for providing us with insights on evolutionary ML and adversarial AI. The insulin strategies demo is amazing!
There are two directions. AI will be more and more personal or local. The other is more and more global. Current AI help physicians diagnose or cure diseases. I believe future use cases for AI are more and more for patients and personal, as opposed to physicians or hospitals. The purpose of using AI is becoming from treatment and diagnosis to prevention, to keeping healthy, and to well-being. One AI may monitor health conditions of an individual every day, when one sleeps, when one sits on a chair, when one walks, AI senses data related to his/her health conditions, and suggests him/her his/her behavior of eating, excising, working, etc. to keep him/her healthy.
The other direction is global. By collecting healthcare data from individuals, for example, by using smart watches, big data acquired from the individuals are stored in a cloud server. AI analyzes the big data and monitor if a particular population of individuals provides a sign of a certain disease, like Influenza or a new virus like the COVID-19, and it identifies a possible source of pandemic. The government or an organization like CDC can prevent the new disease from widespread. One caution for us is that AI is not perfect. It is very dangerous to fully believe AI without the knowledge on what AI can do and what AI cannot. Also, AI can facilitate some sort of discrimination or may create a bias. It is important to know what AI can do, and to think how we, individuals or our society, should work with AI when AI is used in a global setting.
@kenjisuzuki - Thank you for sharing these amazing examples of future use cases of AI in healthcare.