AI summary by TwinMind
Can AI Catch What Doctors Miss? | Eric Topol | TED
TED
14m 6s
AlphaFold and Protein Structure Prediction
- Speaker: 17 years at Scripps Research (largest nonprofit biomedical institution in US)
- Protein crystal 3D structure determination: previously 2-3 years → now 2-3 minutes
- AlphaFold by DeepMind (Demis Hassabis, John Jumper) - Nobel Prize recipients September
- Transformer model converts amino acid sequence (1D) → 3D protein structure (atomic level)
- Inspired additional models: RNA, antibodies, missense mutation detection
- Can create novel proteins not existing in nature
- Paradox: Team of 30 scientists don't understand how transformer model works
- Question: Should AI get asterisk in Nobel award?
Medical Diagnostic Errors Problem
- National Academy of Medicine: All Americans experience ≥1 diagnostic error in lifetime
- Johns Hopkins study: 800,000 Americans dead/seriously disabled annually from diagnostic errors
- "Precision medicine" critique: Precise repetition of same mistakes isn't helpful
Computer Vision in Medical Imaging
Retinal Imaging Breakthrough
- First major hint: 100,000 images with supervised learning
- Gender identification from retina:
- Human experts: 50% accuracy
- AI: 97% accuracy
- Features not fully defined how AI achieved this
Medical Imaging Applications
- Chest X-rays: AI detects cancerous nodules missed by expert radiologists
- Applies to all medical scans: CT, MRI, ultrasound
- Supervised learning with large labeled datasets
- AI performance: equal or better than expert physicians
Colonoscopy Polyp Detection
- 21 randomized trials: machine vision during colonoscopy
- AI + gastroenterologist > gastroenterologist alone
- Performance degrades later in day for humans, not machines
- Unknown if additional polyp detection changes cancer natural history
Deep Learning Medical Applications
Retinal Analysis Capabilities
- Single retinal image can detect:
- Diabetes/blood pressure control
- Kidney disease
- Heart calcium score
- Alzheimer's disease (before clinical symptoms)
- Heart attacks/strokes prediction
- Hyperlipidemia
- Parkinson's disease (7 years before symptoms)
- Future: Retinal photos at checkups as "gateway to every body system"
- Each study: tens/hundreds of thousands of images, separate investigators
Electrocardiogram (ECG) Analysis
- Speaker: Cardiologist 30+ years, couldn't see these patterns
- AI detects from ECG:
- Patient age/sex
- Heart ejection fraction
- Anemia (hemoglobin to decimal point)
- Atrial fibrillation/stroke prediction
- Diabetes/prediabetes diagnosis
- Heart filling pressure
- Hypothyroidism
- Kidney disease
Chest X-ray Analysis
- Race determination (ethical implications noted)
- Diabetes diagnosis + control level
- Multiple cardiac parameters invisible to radiologists/cardiologists
Pathology Slide Analysis
- Pathologists often disagree on slide interpretation
- AI determines:
- Driver genomic mutations of cancer
- Structural copy number variants in tumors
- Tumor origin (often unknown clinically)
- Patient prognosis
- All using convolutional neural networks (not transformer models)
Transformer Models in Medicine
Technical Foundation
- "Attention is All You Need" - highly cited pre-print
- Ability to contextualize multiple items (language, images)
- GPT-4 example: >1 trillion connections
- Human brain: 100 trillion connections/parameters
- Multimodal: language, images, speech
- Massive GPU requirements
- Self-supervised learning advantage (reduces expert labeling bottleneck)
Clinical Applications
Keyboard Liberation
- Major goal for doctors and patients
- Eliminates data clerk functions for clinicians
- Improves face-to-face patient contact
- Synthetic notes from conversation
- Automates downstream administrative tasks
- US health systems: physicians saving many hours
- Trending toward complete keyboard liberation
Foundation Model Development
- Collaboration with Moorfields Eye Institute (Pearse Keane)
- First medical foundation model from retina
- 1.6 million retinal images
- Single model predicts all 8 different outcomes (vs separate studies previously)
- Open-source for others to build upon
Clinical Case Studies
Andrew (6-year-old patient)
- 3 years: increasing pain, arrested growth, gait issues, severe headaches
- 17 doctors over 3 years missed diagnosis
- Mother used ChatGPT with symptoms
- Diagnosis: Occulta spina bifida (tethered spinal cord)
- Surgery performed, now perfectly healthy
Long COVID Misdiagnosis Case
- Patient diagnosed with long COVID by multiple physicians/neurologists
- No validated treatment available
- Sister entered symptoms into ChatGPT
- Actual diagnosis: Limbic encephalitis (treatable condition)
- Patient treated successfully, doing extremely well
New England Journal Study
- 70 difficult cases from clinical pathologic conferences
- GPT-4 vs expert master clinicians
- Chatbot performed equal or better than experts in diagnosis
Future of Medicine Discussion
Conversation with Fellow
- Andrew Cho: 30 years old, 2nd year cardiology fellow
- Medicine remains apprenticeship model
- Predicted benefits:
- Keyboard liberation era
- Enhanced patient connection (first time in decades)
- Automated administrative tasks: pre-authorization, billing, prescriptions, appointments
- Patient nudges and follow-up automation
- Diagnostic support
- Gift of time: patient data prepared before visits
- Transformed patient-doctor relationship
Validation Requirements
- All applications must be validated
- Benefits must greatly outweigh risks
- Remarkable time for healthcare future
