[Note] MD vs. AI: Artificial Intelligence in Oncology
Presented by Prof. Yu Kun-Hsing (Havard Medical School) on 2019 TJCC.
AI Growth Factors
Big data
Machine Learning Algorithm: Data analysis
Computing power↑: Cloud computing; hardware upgrade
FDA approved AI tool
Cardiac MRI interpretation
Example of AI application in clinical settings
Diagnosis of skin lesion
Diagnosis with EMR (Electronic Medical Records)
Diagnosis of cancer pathology, gene mutation
Infectious disease consultation
AI hype→AI winter
Over-inflated expectation→subsequent crash
Understand how AI works & its limitations
How does AI work?
Rule-based approach: draw conclusions based on existing rules, which needs to formulate rules first and continues to maintenances and updates.
Machine learning/ Deep learning (neuro network): learn non-obvious associations.
Supervised ML
Can find decision boundary, can be used in the prediction
Non-supervised ML
Without a trained data set, group and interpret based on input data. (Clustering)
Pathology AI system
Pathology images (some features) may predict some clinical factors:
Prognosis
Gene mutation status
Chemotherapy response
Limitations of ML-based AI
GIGO (Garbage-in-garbage-out): Model generalizability depends on the representativeness of the training data (trained data set bias); The labeling of cases could evolve over time
Correlation not causation
Transforming Medicine with AI
Conventional decision support system (運用 EHR, ML, algorithm 去輔助決策)
Turing Test for Medical Utility
“Superhuman fallacy” not to beat physician but to support decision
How does AI benefit down to Pt.?
Decrease co-payment with proper mechanism
Challenges
Misdiagnosis
Regulation update
ML model interpretation into clinical settings
Social-legal issue
Source:
https://www.slideshare.net/petrieflom/kunhsing-yu-ai-vs-md-will-machines-replace-doctors
