Transforming Medical Records Into a Heart Attack Prediction Model
summary
Hello Heart combines daily app data (blood pressure, meds, symptoms) with EHRs, builds a daily ETL that extracts and aggregates hundreds of thousands of raw records and codes into tabular features, selects features via clinician input plus statistical tests, and trains tabular ML/gradient boosting models with clinician-validated labels. The deployed model achieves AUC ~80%, beats standard risk calculators, runs in-app to alert users and prompt doctor follow-up, and has real-world impact including at least one reported therapy change after an app alert.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 1:14Predicting heart attack risk
- 2:39Why we built the model
- 4:23Data transformation process
- 7:39Challenges with medical data
- 9:04Raw data complexity
- 10:56High-dimensional data issues
- 12:44Labeling quality requirements
- 14:55From research to production
- 16:39Summary and impact
related talks
23mEngineering AI-Powered Predictive Healthcare Systems
40mBringing mental health data to doctors
16mUtilizing The Power of Machine Learning in Healthcare
04mBlack Box: How AI will Amplify the Best and Worst of Humanity
28mHow Smart is the ECG? Lessons from Advanced ECG Analytics
39mBest Practices for Prototyping ML Models for Healthcare
10mTransforming Population Health with Predictive BI
09mHow AI is changing the healthcare system