Module teaching for aft method for data science

Data Science

1. NIH AI-READI and AIM-AHEAD AI Foundation Training (AFT) Module 1: Overview Gordon Gao Johns Hopkins University

2. https://carey.jhu.edu/cdhai ‹#›

3. ‹#› Mission Generate knowledge to enable the digital transformation of healthcare Provide mentorship and education for the next generation of digital health experts Be a thought leader in digital health and artificial intelligence Center for Digital Health and Artificial Intelligence

4. ‹#› 20 years riding technologies Montserrat 2005 EHR, HIE, Telemedicine Digital transformation EHR adoption Routinization Privacy concerns 2010 Patient Empowerment Social media mHealth and wearable devices Nudging Online reviews 2015 AI / Machine Learning Deep learning NLP Trust in AI Algorithmic biases

5. ‹#› 2005 EHR, HIE, Telemedicine Digital transformation EHR adoption Routinization Privacy concerns 2010 Patient Empowerment Social media mHealth and wearable devices Nudging Online reviews 2015 AI / Machine Learning Deep learning NLP Trust in AI Algorithmic biases 2020 Next Generation of AI LLM Generative AI Responsive AI Human-AI Teaming 20 years riding technologies

6. DiaSocial App 27-patient pilot in 2015, with VA Baltimore User comments “I used it to do the input on the medications, the exercise I did, and my food intake for my 3 meals. I got to the place where it was like second nature to want to put it in, and it worked out nice.” “Well previously I wasn’t really managing well, this tool has been helping me. This thing can go on for another year as far as I’m concerned.”

7. Regulatory mode Locomotion (e.g., “I enjoy actively doing things, more than just watching and observing”;) Assessment (e.g., “I spend a great deal of time thinking about my positive and negative characteristics”;) 2018

8. 2020 Trait Description Openness Being curious, original, intellectual, creative, and open to new ideas. Conscientiousness Being organized, systematic, punctual, achievement-oriented, and dependable. Extraversion Being outgoing, talkative, sociable, and enjoying social situations. Agreeableness Being affable, tolerant, sensitive, trusting, kind, and warm. Neuroticism Being anxious, irritable, temperamental, and moody.

9. 2022 Outcome measures knowledge learning task self-tracking task Informational support Emotional support Team Building Team Identity Leadership

10. Part 1: Exciting AI Age

11. Outline Review the Historical Development and Key Milestone in AI/ML Explore AI Applications in Healthcare Dual Impact of AI on Healthcare Knowledge Gap and Research Opportunities Coming modules

12. ‹#› Two-year Anniversary of ChatGPT Launch of ChatGPT 2023 2024 Nov 30 1 Million Users in 5 Days Dec 5 ChatGPT Plus Subscription 100 Million Monthly Users Jan Feb GPT-4 Release Mar iOS App Launch May ChatGPT Enterprise Launch Aug Voice and Image Capabilities Subscription Sep ChatGPT plugins GPT-3.5/4 with browsing Apr iOS App Launch Jul May Advanced Voice Mode Nov GPT-4o O1 Model Launch May 2022

13. Quiz Over the last 18 months, the cost of GPT-4 equivalent intelligence from OpenAI has fallen by about how much? A. 2x B. 10x C. 50x D. 250x ‹#›

14. The dramatic drop in computing using ChatGPT ‹#› $ 180 $ 0.75

15. ‹#›

16. ‹#›

17. The technology could bring about "12 million occupational transitions" by 2030, according to Kweilin Ellingrud, a McKinsey partner. In 2023, Goldman Sachs produced a report speculating that AI could replace the equivalent of up to 300 million jobs — namely in the administrative sector. ‹#› Impact on the labor force

18. AI is GPT

19. ‹#› ‹#›

20. ‹#› Electricity’s Journey as GPT: The Golden Age Light Bulb（1879） First Killer app that wired the world Electric fan(1890) Golden age to appliances starts Vacuum cleaner(1905) Weighted 92 pounds Toaster(1912) a huge success Washing Machine (1910) Everyone wants it But it’s dangerous The Big question: We have electricity now, what will be the light bulb, the electric fan, the vacuum cleaner, the toaster, the washing machine in our time? History doesn't repeat itself, but it often rhymes - Mark Twain

21. What is the nature of AI? What is AI good for? How do we ensure AI use is fair and ethical? How to advance healthcare research using AI? Especially diabetes care? ‹#› Disambiguating AI: Open questions

22. Part 2: What is AI

23. What is AI? Machines that can think How to test the level of intelligence? Turing Test (1950)

24. Why the Turing Test can be biased?

25. “People should stop training radiologists now. It’s just completely obvious that within 5 years, deep learning is going to do better than radiologists […] We’ve got plenty of radiologists already [.]” - Geoffrey Hinton (2016)

26. Key tension: Experience / Human Capital “[a] highly-trained and specialised radiologist may now be in greater danger of being replaced by a machine than his own executive assistant” -Andrew Ng (2016)

27. How to make machines intelligent? AI Mechanical

28. How to make machines intelligent? AI Mechanical Knowledge base: Logic and ontology, with formal rules Think of this approach as the “pure and clean” way. Like playing chess. Problems: the rules exploded! For every rule, there is an exception! Conflicting and hard to maintain.

29. Ambiguity Blog: Definition and Examples of Ambiguity

30. Detecting Fraudulent Doctor Review “Dr Kant and his team is very skillful and hard working, very well recommended to everyone, best dental experience I had at this clinic.” “Dr. explained everything in detail, medication given was really good, we are very happy to consult this doctor, we highly recommend this doctor.” ‹#› Source: Shukla, Wang, Gao and Agarwal 2018

31. How to make machines intelligent? AI Mechanical Knowledge base: Logic and ontology, with formal rules Statistical learning/Machine learning Logistic regression Ex: Predict(Readmission) = f(patient age, gender, diagnosis, medication) Features

32. How to make machines intelligent? AI Mechanical Knowledge base: Logic and ontology, with formal rules Statistical learning/Machine learning Logistic regression Ex: Predict(Readmission) = f(patient age, gender, diagnosis, medication) But some features are hard to quantify: -such as images of the bone fracture Representation learning Features

33. Why is Representation important? Compare: XCIX –XVI =? 99 – 16 = ? 01100011 - 00010000 How to separate the dots by a single straight line?

34. Why is Representation important? How to compute natural language: Word2Vec King – Man = Queen – Woman

35. How to make machines intelligent? AI Statistical learning/Machine learning Representation learning Deep Neural Network (DNN) CNN, RNN, GANS

36. Shaded boxes indicate learning from data (rather than hand-designed) (Raw Features)

37. Credit to Colin Garvey, 2018 AI Booms

38. Milestones in AI 1997 Rule-based system/Symbolic 2011 Machine Learning/ Statistical Learning 2016 Deep Learning/ Neural Network 1997 2011 IBM Deep Blue IBM Watson 2016 Google AlphaGo

39. Part 3: Why AI now?

40. Why AI now (again)? Advances in Algorithm Advances in algorithm Advances in computing/hardware Advances in Big Data

41. Algorithm Breakthroughs

42. Turing Award in 2018 ‹#› Critical Funding from Canada (CIFAR) Geoffrey Hinton: Internalizing the learning and keep the momentum Yoshua Bengio: Generative adversarial networks (GANs) Yann LeCun: Early Convolutional Neural Network (CNN)

43. Modern Deep Learning Development 2002 2010 2011 2013 2014 2016 2017

44. Build the Neural Network Using Keras ‹#›

45. Why AI now? Advances in Computing Power Advances in algorithm Advances in computing power Advances in Big Data

46. Computing power ‹#›

47. Look Mom! I found a Cat!

48. Computing Power Video card computing power - increased 1 million times since the 1990s ‹#› In this course, we will use Google Colab to access GPU resources for enhanced computing power.

49. Why AI now? Advances in Big Data Advances in algorithm Advances in computing power Advances in Big Data Data is the new oil!

50. GPT Models before ChatGPT 45 TB

51. Deep Learning

52. Tipping point? “Most people don’t understand the difference between 95% accuracy and 99% accuracy. 99% is game changing.” Andrew Ng

53. Big Data Timeline ‹#›

54. Here comes Big Data! ‹#›

55. Deep Learning ‹#› Algorithm Computing Big Data AI

56. Part 4: Double-edged sword

57. Diabetes: Increasing Prevalence 7.1% 9.2% 13.4% Data Source $1 out of every $4 in US health care costs is spent on caring for people with diabetes

58. But It Didn't Have to Happen Actions: National Diabetes Prevention Program Hopkins Diabetes Prevention Program

59. Risk Varies among People with Prediabetes Swedish National Study on Aging and Care 36% had prediabetes 22% reverted to normal Conclusion: most of older adults with prediabetes remained stable or reverted to normoglycemia, whereas only one third developed diabetes or died. 44% had prediabetes 13% reverted to normal

60. Risk Assessment

61. AI can exacerbate health disparities A commercial AI that identifies millions of patients with complex health needs exhibits significant racial bias. At a given risk score, black patients are considerably sicker than white patients.

62. AI: a double-edged sword Healthcare organizations and clinicians should be cautious to use AI properly. We can take proactive measures to improve AI fairness. Here are three criteria. 1. Equal outcomes Different groups of populations should benefit equally in terms of patient outcomes. 2. Equal performance AI should perform equally accurate for patients with various backgrounds. 3. Equal allocation AI as resources should be allocated equally. Carey Business School Research

63. Part 5: current research and knowledge gap

64. Current AI in Diabetes: Risk Prediction Prediction for Long-Term Risk Onset of Diabetes. Onset of Diabetes Complications Chronic Kidney Disease (CKD). Diabetic Retinopathy (DR). Aim: Identifying high risk patients as early as possible. Gaps and challenges: Data quality (labeling the ground truth from EHR data) A1C = 6.3 …diabetic or not? A1C > 6.5, versus risk of complications Age adjustment criteria. New Diabetics Types revealed by AI. New data modality (genome, sensors, retina,..)

65. Predicting near-future BG on a minute, hour, or overnight timescale enables better diabetes management. The prediction of BG levels is challenging food and insulin, activity interact with each other. Pretraining large volume Continuous glucose monitoring data will improve the prediction performance. New data modality: Blood Glucose

66. Multi-Modality for AI in Diabetes Large Sensor Model over Continuous Glucose Monitor Developed as a generative foundation model to analyze CGM (continuous glucose monitoring) data for metabolic health insights. Uses autoregressive token prediction to capture nuanced glycemic patterns. A Foundation Model for CGM Data CGM-LSM GluFormer Multi-modal data sources: Use genetic information to predict onset of diabetes. Genetic risk score (GRS) based on 20 T2D-associated SNPs.

67. Current AI in Diabetes: Patient’s Behavior Diet: AI-based diet recommendation Exercise: Physical Therapy The scientific, personalized, and quantitative exercise prescriptions. Medications: Drug Therapy Optimal Dosage Strategies of insulins and anti-diabetic drugs.

68. AI Application in Clinical Practice Automatic retinal screening has been widely integrated into diabetes care. Self-management tools for blood glucose monitoring. 98.5% accuracy for onset of hypoglycemia. Patients can take actions to prevent it.

69. Part 6: future modules

70. Part 6: future sections