AI + Medicine
Real word AI application in Medicine
Lots of data in healthcare
- An overwhelm data not only bio data but physical data generated by our body.
- Several different layer of molecule studies: Genome, transcriptome, proteome, metabolome, microbiome, epigenome, exposome.
- Bad outcomes:
- Very high cost: Data flow is very costly not only in the US but around the world.
- Error galore.
- Profound waste: Extra unnecessary testing.
- Little time.
- Burn out.
How AI will transform Health Professional
|Clinical Category||Data Interpretation|
|Geneticists||Face, BAM file|
|All Doctors||Delete Keyboards|
AI across the health span
Example of applications:
Using hundreds and thousands million examples that human expert can’t
- Predict gender by eyes using red note pictures. (There is better way)
- Chest x-ray: missing information.
- Breasts mamography: to decrease false negative/ false positive.
- High accuracy and higher quality with brain scanner in real time.
- Convenient in smart toilet.
- Common issue high accuracy with skin (no app for consumer to let people know if they need to go to demartologists yet)
- Recognize pan cancer
- Smart phone ultrasound to AI acquisition, with an extra device and phone as a processor.
- Use image detection to detect cancer in stomach and ocular fundus.
- The eyes diagnostic Diabetics using the pre-existing system approved by FDA.
- Train intraperative brain tumour: High accuracy and quick.
- Organ injury.
- Require burden keyboard work for physicians.
Random trials of AI DNN
- AI vs clinical: systematic review, design, report standards and claims of deep learning 81 studies, 10RCTs
The virtual medical assistant
The bottleneck problems.
- Lack of large divers annotated datasets to work on self-supervised learning algorithms.
- Lack of prospective trials.
- Lack of deep collaboration between computer scientist and clinicians.
- Lack of implementation and appreciation for surveillance stage.
- Need for new and hybrid models for multi-dimensional data.
Machines will not replace physicians but physician using AI will soon replace those not using it.
Challenges working in an intersection between AI and Healthcare:
- Communication gap: Context and culture, e.g. language and how to approach problems of two communities.
Digital process of healthcare:
- Accumulation of data. But the problem is medical community is very resistant to change. Want to have change but don’t have. Some time tradition, reimbursement, loss of control, education.
- A similar story of the gap among computer scientists with an example of Google AI pilot project ten years ago.
How does Stanford Healthcare demonstrate AI application?
The use of AI for Chest X-ray. Problem: logistical in working data, regulation, quality control, IT infrastructure to answer basic research question. Control setting and real world data setting, e.g. between different countries. Time of execution, training time. A robust model for clinical physicians to work in real world
How meaningful is a bridge between AI/ tech and medicine?
Different in science culture and disclosure. Example: publication process between two communities, e.g. pre-print and public discussion and then peer-review.
Motivation to close the gap, a willingness to improve conversation.
Why someone would jump into healthcare?
More data is digital. Lots of room in healthcare system, e.g. software and human. There is more eagerness to exchange knowledge between communities.
How AI can help healthcare human again? (We thought the in a converse way)
The key point most of us miss is a degradation of human element in healthcare and medicine. Since 80s, less and less between doctors and patients because of time constraint, shift work and clinical work which eventually lead to a global burn out and depression in doctors and nurses. Increase accuracy, speed and efficiency. The gift of time help us trust and bond between patients and the system.
How to make life and death decision supported by black-box AI? How explainable is helpful?
Similar to the process of two physicians communicating with each other to make decisions. How algorithms work, how data is mapped, lab test there is a rational decision behind the scene.
Doctors need explanation not for day-to-day decision making but faith for something reasonable enough. There is a bag of techniques for decision-making
What are the advantages of AI during this pandemic?
Detection and hospital management. We are in a half year not a year and half. Situation is unpredictable. A smart thermometer could be a way to localize an high potential area of the outbreak, then people can get in, test, isolate, contact tracing – the best prevention practices. Another example is Fitbbit and Apple smart watches to capture data like a resting heart rate. Other ways like digital surveillance, keep people out of the hospital by staying at home with devices but the problem here is that we don’t have proof about its safety and algorithm.
How to stay updated with the community?
- Machine leanring: Deep Learning Specialization.
- Twitter: Eric Topol.
- Dr Penguin newsletter.
Why aren’t these system widely deplyed in hospitals yet?
Practical bottleneck for ML:
- Small (low) data.
- Robustness and generationalization: A model works in a publication (research institution) does not work in a production settings. Potential solution: human augmentation, collect more data.
- Change management (including safety and regulation)
Human vs machine learning: top 10 conditions problems
Given an image and a medical history, a radiologist can diagnose any of a large number of conditions.
- Easy: Diagnose pneumonia from 10K labeled images.
- Hard: Diagnose a rare condition from 10 images of a medical textbook chapter explaining that condition.
- Healthcare worker jobs, e.g. radiologist, other physician, nurses, medical coders, insurance, regulators, etc. It leads to a question of how to get everyone on board in a new workflow.
- High concern about patients safety. How do we make sure an ML algorithm does not cause harm.
ML model vs ML system
- Only 5-10% of AI projects are ML code.
- A huge gape between proof of concepts (and publication) and production. (it’s hard to train a good deep net)
Tagged #ai, #summary.