A few notes from the morning speakers:
Keynote was a fireside chat with Dr. Eric Topol:
By 2019, most applications of AI were retrospective, afterwards, studies are being designed with the inclusion of AI. The current challenge comes from wearables and continuous monitoring devices - these generate huge amounts of data that managing is a problem, on top of teaching the AI how to interpret what could be minor changes that accumulate over time.
Bias can come from AI assessments when the training data is not inclusive from a racial, socio-economic, and life-style standpoint. The Optum report includes examples where an historic publication or study establishes standards for a disease that are related to a specific population and not applicable to other specialty populations with similar symptoms/disease.
Cell phones are now being used to capture ultrasound imaging for interpretation by local or virtually connected physicians and other health care workers.
Data security privacy computing are a focus currently a number of approaches are in development (e.g., federate and swarm learning).
Emphasized the use of AI to permit more time for the physician to spend with patients to gather more data; allowing the AI to have a richer dataset for assessment and recommendations.
Dr. Alan Edelman (MIT) talk on "Intro to AI & ML"
Talked about 5 areas for AI in healthcare: Scientific AI to look for trends in data to generate new discoveries, Natural Language Processing (patient chatbots, unifying patient records, redacting confidential info), Precision medicine (custom treatments based on individual's characteristics), Computer Vision - detect tumor and lesions, and Physician Guidance (during surgery, recommendations of course of therapy).
Felt that we are not even close to the possibilities of AI & ML">
Dr. James Lu "AI-partnered Dynamic Model Discovery for Precision Medicine"
Wants to have pharmacology and pharmacodynamic models to develop casual models that go on through additional learning, to improve comprehensive disease models that use patient specific data to guide therapy.
Dr. Nadia Terranova (Merck Kga) "Enabling AI learning to support Precision Medicine"
Totality of information (high dimension data) for a patient can be compared to multiple pharmacology models to guide therapy selection and application. Notes inter-tumor heterogeneity for a patient, as well as across patients. Showed data from a study of 6369 individual lesions that applied these concepts.
Next speaker Dr. Ksorbo
Deep PumasAI - bridges the machine learning model which require large data sets with scientific models of known results (e.g., Pk on drug, known biomarkers of disease). Conceptually, this reduces the learning curve while enabling the ability to query the situation for a single patient for the best therapy and dosing model, as well as quality of life and related issues.
Rahul Goyal "ScreenMCM: A machine learning-based product screening tool to accelerate medical countermeasure development"
Due to the nature of the disease or condition (e.g., radiation exposure, toxin exposure) Uses computer models to reduce the number of animals in testing while testing more therapies. Created a model from 3 animal studies that produces 1500 rules against which new compounds can be tested. 60% of study data was used for training, 40% used to verify if the model was accurately predicting outcomes. Accuracy was noted at >70%. Application to additional discovery therapies allowed selection of products using 10 day study data with fewer animals.
Dr. Hao Zhu (FDA) "Application of AI and ML in Drug Developemtn and Precision Medicine"
AI can handle large number of variables vs Traditional Pharmacometric models which typically use fewer variable. However Pharmacometric models can provide good statistical understanding, AI does not provide robust statistics.
No comments:
Post a Comment