3. Disclaimer
I don’t own or make claim of the images used in this
slide or it’s copyright.
This work is solely intended as a medium to express
concepts and promote further work in this area.
All copyrights belongs to their respective authors
10. J Clin Oncol. 2006 Aug 10;24(23):3726-34.
N Engl J Med 2004; 351:2817-2826
What is Oncotype Dx ?
• A diagnostic test for breast cancer
• Analyze 21 genes from breast tumor to
determine
– Predicts recurrence
– Predicts the benefit of chemotherapy
11. Predicts recurrence
diagnostics
Predicts benefit of
chemotherapy
較高的議價空間
USD 25-50/test
Margin 5-10%
100% penetration
市場大小 5 million/yr
Biodesign: The Process of Innovating Medical Technologies 1st ed.
化療 15000 USD/人
若減少非必要的治療達50%
約莫 7000 USD/test
市場大小 700 million/yr
42. Interpretable
Trust
- Not just how often it is right
- But also which examples it is right ?
Causality
Transferability
- capacity to generalize to
unfamiliar situations
Being fair
& Ethical
43. Properties of interpretable models
1. Transparency
How does the model work?
2. Post-hoc interpretability
Besides the prediction, what else can
the model tell me?
Learning the model locally around the prediction
52. Construct X
Construct Y
X causes Y by
mechanism F
where F(X) = Y
Measurement x
Measurement y
model f maps x to y
via f(x) = y
53. Construct X
Construct Y
X causes Y by
mechanism F
where F(X) = Y
Measurement x
Measurement y
model f maps x to y
via f(x) = y
Explanation
match f to F
54. Construct X
Construct Y
X causes Y by
mechanism F
where F(X) = Y
Measurement x
Measurement y
model f maps x to y
via f(x) = y
Prediction
use f to generate y
55. + residual
Good fit of known data{X,y} to model
不表示 model 會在 {Xnew, ynew} 表現好
65. It is possible to reduce variance by increasing bias
And still resulting in reduced overall error
Loss of model explanatory power → Increased predictive power
https://arxiv.org/pdf/1101.0891.pdf
66. Predictive power
(How does model perform in out-samples?)
Explanatory power
(Given theory, how does the sample data fit?)
Two dimensions
80. Take home message
• Unmet needs + Market
– 3P: Patient, Payer, Provider
– 預後/風險/安全性 > 方便/使用者體驗
• Model Interpretability
– Transparency vs. Post-hoc interpretability
• Explanation =/= Prediction
• Predictive modeling for observational data
• Software as a Medical Device(SaMD)