The Coming AI Disruption of Healthcare Lawrence Carin




































- Slides: 36
The Coming AI Disruption of Healthcare Lawrence Carin Duke University
So, What is Deep Learning and AI?
Learning a Predictive Model Based on Labeled Data y, associated label 0/1 x, data for a subject
Training Set
Linear Predictive Model
Logistic Regression
Learning Model Parameters
Generalization of Logistic Regression: Learned Features Probability of K latent processes/features
Extended Logistic Regression Probability of particular outcome Probability of K latent processes May be viewed as logistic regression on K latent features, rather than directly on the N components of raw data
Going Deeper Probability of K Layer-1 latent processes Probability of J Layer-2 latent processes
Probability of particular outcome Probability of K Layer-1 latent processes Probability of J Layer-2 latent processes
Increasing Degree of Specificity and Feature Sophistication
Multilayered Perceptron: Neural Network
The Trade: Bias vs. Variance ?
The Seasons of Neural Networks Multi-Layered Perceptron 1960
The Seasons of Neural Networks Multi-Layered Perceptron Back Propagation 1960 1986
The Seasons of Neural Networks Multi-Layered Perceptron Back Propagation Convolutional Neural Network 1960 1986 1989
The Seasons of Neural Networks Multi-Layered Perceptron Back Propagation Convolutional Neural Network Neural Nets in the Wild 1960 1986 1989 1990– 1994
The Seasons of Neural Networks Multi-Layered Perceptron Back Propagation Convolutional Neural Network Neural Nets in the Wild Long Short-Term Memory 1960 1986 1989 1990– 1994 1995
The Seasons of Neural Networks More Neural Nets In the Wild 1998– 2005
The Seasons of Neural Networks More Neural Nets In the Wild Banishment 1998– 2005– 2010
The Seasons of Neural Networks More Neural Nets In the Wild Banishment Rename: Deep Learning 1998– 2005– 2010–
The Seasons of Neural Networks More Neural Nets In the Wild 1998– 2005 Banishment Rename: Deep Learning CNN+GPU+ Image Net 2005– 2010– 2013
The Seasons of Neural Networks More Neural Nets In the Wild 1998– 2005 Banishment Rename: Deep Learning CNN+GPU+ Image Net Alpha Go 2005– 2010– 2013 2015
Those Who Cannot Learn from History Are Doomed to Repeat It ?
Deep Learning + Transfer Learning Image. Net (public image database with 1. 2 M+ images) Convolutional Layers Fully Conn. Layers Categories Weights Confocal Images Convolutional Layers Fully Conn. Layers Scores 0 1 3 2
AI in the Clinic: Medical-Image Analysis Ophthalmology Deep Learning Pathology
AI for the Basic Sciences of Health: Neuroscience K. Dzirasa, L. Carin, et al. , “Brain-wide Electrical Spatiotemporal Dynamics Encode Depression Vulnerability, ” Cell, 2018
Value-Based Care
Enhancing Patient-Doctor Experience with AI q Clinician burn-out and depression significant national issue q The EHR has imposed significant data entry on doctors q Not why people become doctors, also serious degradation of patient experience q Opportunity for AI: § Audio processing to transcribe all MD-patient verbal communications § Use natural language processing to automate data entry § Machine learning to assist MD in diagnosis & treatment
AI to Automate Answering Routine Questions of Clinicians
The Coming AI Disruption of Healthcare q AI has had many seasons, this time spring/summer seems here to stay q Many demonstrations of utility, in the real world q Significant disruption in motion for healthcare, that will be accelerating: - Analysis of all forms of clinical images: radiology, pathology, ophthalmology, etc. - Health-system optimization - Patient-clinician interactions - Basic sciences for health



































