Introduction¶
Problems¶
- Misdiagnosis
- Late diagnosis
- Inappropriate management after diagnosis
- Medical errors are pervasive
History¶
| Limitations | ||||
|---|---|---|---|---|
| 1970s | MYSIN expert system | Identifying bacteria | ||
| 1980s | INTERNIST-1/ Quick Medical Reference | Internal medicine | Bayesian network-like | 1. Clinicians entered symptoms manually 2. Difficult to maintain 3. Difficult to generalize (prior probabilities will different across different parts of the world) |
| RX Project | ||||
| 1990s | ANN | 1. Did not fit well into clinical workflow 2. Hard to get enough training data 3. Poor generalization to new places |
What has changed?¶
Data availability¶
- Adoption of Electronics Records
- Lab tests
- Imaging
- Vital signs
- Genomics
- Wearable sensors
### Standardization
- Reports
- Data storage
- APIs
OMOP
Machine Learning¶
- Learning with high-dimensional features
- Semi-supervised and unsupervised learning
- Deep learning
- Democratization of machine learning
- Open-source software
Overview¶
Emergency department
- Limited resources
- Time sensitive
- Critical decisions
Applications¶
- Better triage
- Faster diagnosis
- Early detection of adverse events
- Prevent medical errors
- Recommend treatment pathway
- Anticipating clinicians needs
- Reducing needs for specialist consults
- Automated documentation & billing
- Predicting patient’s future disease progression
- Continuous monitoring
- Discovery of new disease subtypes
- Design of new drugs
- Better targeted clinical trials
What makes ML in healthcare different?¶
- Life/death decisions, similar to Autonomous Driving
- Need robust algorithms
- Checks and balances required for ML deployment
- Need fair & accountable algorithms
- Lot of scope for unsupervised learning
- Causal learning required: just prediction insufficient
- Very little labelled data: need to use semi-supervised algorithms
- Small sample size
- Data quality issues
- Varying time intervals
- Missing data
- Censored labels
- Data sensitivity
- Difficulty of de-identifying
- Difficulty of deploying ML
- Commercial electronic health record software is difficult to modify
- Different standards used
- Careful testing and iteration needed