Concept 1 / 1
Key Concepts to Memorize
Core thesis: ML "often works" in ideal lab conditions but "often fails" in medical reality
- ML works best when: data is uniform, classes balanced, boundaries defined, labelling possible
- ML fails when: data varies, classes imbalanced, boundaries undefined, labelling impossible
- Example failure: 50% of congenital heart disease missed in antenatal ultrasound; 10% false positive rate after a single mammography
"Unknown unknowns" problem (Rumsfeld quote):
- Supervised ML only knows what it has seen
- Key challenge: detecting things the model was never trained on (out-of-distribution / OOD)
Normative Representation Learning:
- A variant of unsupervised single-class learning
- Focus on density estimation and out-of-distribution (OOD) detection
- Learn what is "normal" → anything deviating significantly = anomaly
- Useful when labelling is impossible or rare events need detection
Dense representation models:
- Encode data into dense, continuous, low-dimensional vector spaces
- Allow interpolation between complex samples (→ counterfactual generation)
- Allow detection of low-support areas (→ OOD detection)
- Built with: GANs, VAEs, diffusion models, transformers
Counterfactual analysis:
> "How would the scan of this patient look like if some clinical parameter would be different?"
Safe data sharing:
- Synthetic data must NOT reproduce any training sample (privacy)
- Synthetic data must faithfully represent 100% of training distribution (quality)
- Metric: IRS (Image Retrieval Score) — measures diversity of generated images
Foundation models debate:
- One large foundation model that knows everything vs. many specialized expert models vs. both
- Current challenge: hallucinations and factual inconsistencies in generative models
- Solution explored: Bayesian Decoding Game (Generator + Verifier game to improve consistency)
Open challenges in Germany:
- Retrospective access to patient data almost impossible due to regulatory hurdles
- State-of-the-art healthcare AI requires hundreds of thousands of patient cases
- Most initiatives (NAKO, UK Biobank) involve healthy participants only
- Positive: Bavarian Health Cloud will soon enable secure large-scale data use