making deep learning powered healthcare a reality: will jack, ceo, remedy
TRANSCRIPT
Making Deep Learning Powered Healthcare a Reality
Will Jack, Co-founder
LARGE CLEAN DATA SETS
DEPLOYING TECHNOLOGY AT THE POINT OF CARE
INTERPRETABLE MODELS
INTERPRETABLE MODELS
Deep models are necessary for accuracy but are hard to interpret
FLU
sneeze
weight
headache
no fatigue
age
sneeze
headache
no fatigue
Model Data and Prediction Explanation Human Decision
Explainer (LIME)
Model Interpretability: LIME
Source: Marco Tulio Ribeiro
Source: Marco Tulio Ribeiro
Model Interpretability: LIME
P ( ) = 0.54 P ( ) = 0.07 P ( ) = 0.05
Source: Marco Tulio Ribeiro
Model Interpretability: LIME
Model Interpretability: Distillation
Target
Output
Input
Deep Learning
Model
Mimic Model
y y y y
x
nn nn
nn
m
x x
Source: Che et al. “Interpretable Deep Models for ICU Outcome Prediction”
Model DistillationBLACK BOXDEEP MODEL
0
1 2
3 4 5 6
OUTPUTS
INPUTS
Model Interpretability: Distillation
Source: Che et al. “Interpretable Deep Models for ICU Outcome Prediction”
LARGE CLEAN DATA SETS
HIGHLY SILOEDLACKS
STANDARDIZATION DIRTY DATA
Interpretation Model: Distillation
Source: Marco Tulio Ribeiro
DEPLOYING TECHNOLOGY AT THE POINT OF CARE
Deploying to health systems is a
NIGHTMARE
VIRTUAL PHYSICIAN’S ASSISTANT
Remy
Remy collects granular, structured data around the diagnostic thought process
INTAKE HISTORY TRIAGE FOLLOW UP
COMPREHENSIVE DATASET
OUTCOMES
Remedy is building a
for automated diagnosis and treatment plan design
based on
Questions?