Step by Step
S
Successes — matching or exceeding specialist performance
AI has demonstrated genuine successes in specific medical domains: diabetic retinopathy screening matching ophthalmologist-level performance, chest X-ray pneumonia detection (CheXNet), skin cancer detection, and AlphaFold's protein structure prediction.
Example: an AI system for diabetic retinopathy screening achieving accuracy comparable to a trained ophthalmologist reviewing the same retinal images.
C
Challenges — distribution shift and regulatory approval
Real-world deployment faces genuine challenges: distribution shift (a model failing at hospitals whose patient population or equipment differs from its training data), rare conditions being underrepresented in training data, and the need for FDA regulatory approval (510(k) clearance) before deployment.
Example: an AI diagnostic tool performing excellently at the hospital where it was trained and validated, but performing noticeably worse when deployed at a different hospital with a different patient population or imaging equipment.
B
Bias — known disparities across skin tones
Documented bias issues include pulse oximeters being less accurate on dark skin, and dermatology AI models being trained predominantly on light-skinned patients, producing known performance disparities.
Example: a dermatology AI model trained mostly on images of light-skinned patients performing measurably worse when diagnosing skin conditions on darker-skinned patients.
A
Augmentation, not replacement — the core finding
Across the medical AI research literature, the combination of AI plus a physician consistently outperforms either the AI alone or the physician alone — supporting an augmentation model rather than an AI-replaces-doctors model.
Example: a diagnostic accuracy study finding that physicians using AI-assisted tools consistently outperformed both AI-alone and physician-alone approaches.
Applied Walkthrough
1
A hospital is considering deploying an AI diagnostic tool that performed excellently during validation testing at a different research hospital.
2
Ask: does strong performance at the original validation hospital guarantee similarly strong performance once deployed at this new hospital? Not necessarily — distribution shift can cause performance to degrade at hospitals with different patient populations or equipment than the original training data.
3
The hospital also needs to confirm the tool has received appropriate FDA regulatory approval (510(k) clearance) before clinical deployment.
4
Rather than using the AI tool to fully replace physician judgment, the hospital adopts an augmentation model, since AI plus physician has been consistently shown to outperform either one alone across the medical AI research literature.
Exam Application
Exams test whether you can name specific medical AI success stories (diabetic retinopathy, CheXNet, skin cancer detection, AlphaFold) alongside genuine deployment challenges (distribution shift, rare condition underrepresentation, FDA approval requirements) and known bias issues (pulse oximeters, dermatology AI training data). Also expect the core "augmentation, not replacement" finding to be directly tested.
⚠ Common Trap
The most common trap is assuming strong validation performance at one hospital guarantees similarly strong real-world performance everywhere. Distribution shift is a genuine, well-documented challenge — a model's performance can degrade significantly when deployed at a different hospital with a different patient population, equipment, or imaging protocols than its original training data.
✓ Quick Self-Check
1. Name one specific medical AI success story mentioned in this lesson.
Diabetic retinopathy screening, CheXNet (pneumonia detection), skin cancer detection, or AlphaFold (any one).
Tap to reveal / hide
2. What is distribution shift, in the context of medical AI deployment?
A model performing worse at hospitals whose patient population or equipment differs from what it was trained on.
Tap to reveal / hide
3. What regulatory approval is typically required before deploying a medical AI device?
FDA 510(k) clearance.
Tap to reveal / hide
4. Name a documented bias issue affecting medical AI or medical devices.
Pulse oximeters being less accurate on dark skin, or dermatology AI being trained mostly on light-skinned patients.
Tap to reveal / hide
5. What is the core finding regarding AI and physicians working together?
AI plus physician consistently outperforms either alone — supporting augmentation rather than replacement.
Tap to reveal / hide