Much of AI is predict - predict - predict
- using ECG, predict presence of structural heart disease, typically diagnosed with cardiac echo
- predict 10-year heart attack risk using basic medical information
The goal is impact on healthcare
ECG to SHD [@yaoArtificialIntelligenceEnabled2021]
- prediction: structural heart disease
- intervention: refer patient for cardiac echo
- outcome: diagnosis of structural heart disease
- outcome (impact): reduce preventable early cardiac death
10 year heart attack risk [@hippisley-coxDevelopmentValidationNew2024]
- prediction: heart attack in 10 years
- intervention: prescribe cholesterol lowering medication
- outcome: heart attack
- outcome (impact): reduce heart attacks
Process
- data scientists optimize for predictive accuracy, which entails modeling statistical associations in the ‘healthcare system’
- the hope is: better prediction \(\imply\) better impact
- unfortunately, this is not automatically the case
When accurate prediction models yield harmful self-fulfilling prophecies [@vanamsterdamWhenAccuratePrediction2025]
