ML / Medicine

• AI-Enhanced Handheld ECGs, working paper (with Jenny Wang, Alex Schubert, Esther Duflo, and others)

  You can buy a pocket ECG device online for about $75. That device, with AI layered on top, can bring cardiac screening out of specialty clinics and into low-resource settings.

• Pooled Testing Efficiency Increases With Test Frequency, PNAS 2022 (with Ned Augenblick, Jon Kolstad, and Ao Wang).

  COVID showed why mass testing needs to work better at scale. Pooling, which combines samples across individuals and tests them together, has some interesting properties that can make testing very cost-effective. Something to keep in mind for next time.

• Variation in Common Laboratory Test Results Caused by Ambient Temperature, Med 2021 (with Devin Pope).

  Everyday weather fluctuations can shift common lab results, enough to move some patients across clinical thresholds.

• An Algorithmic Approach to Unexplained Pain Disparities in Underserved Populations, Nature Medicine 2021 (with Emma Pierson, David Cutler, Jure Leskovec, and Sendhil Mullainathan).

  Radiologists grade knee X-rays using criteria developed in studies on English coal miners in the 1950s. AI can find signals those criteria miss, helping explain pain that looks invisible by the old scale.

• Direct Uncertainty Prediction for Medical Second Opinions, ICML 2019 (with Maithra Raghu, Katy Blumer, Rory Sayres, Robert Kleinberg, Sendhil Mullainathan, and Jon Kleinberg).

  Scarce second opinions should go to hard cases, where the first opinion is most likely to be wrong.

• Individuals’ Body Temperatures Vary Meaningfully and Predict Mortality, BMJ 2017 (with Jasmeet Samra and Sendhil Mullainathan).

  Normal temperature is not one number; a person’s own baseline carries information about mortality risk.

ML / Health Policy

• The Health Costs of Cost-Sharing, QJE 2024 (with Amitabh Chandra and Evan Flack).

  Quirks in Medicare policy assign some patients higher or lower copays almost at random. Patients assigned higher copays skip drugs that were keeping them alive, and the highest-risk patients were most likely to skip.

• Diagnosing Physician Error: A Machine Learning Approach to Low-Value Health Care, QJE 2022 (with Sendhil Mullainathan).

  Doctors make mistakes when diagnosing heart attacks in the emergency department. Machine learning can identify both over-testing and under-testing, producing better outcomes at lower cost.

• Predictive Modeling of US Healthcare Spending in Late Life, Science 2018 (with Liran Einav, Amy Finkelstein, and Sendhil Mullainathan).

  If we cannot predict who is near death, targeting end-of-life spending is a weaker savings strategy than it sounds.

Algorithmic Bias

• Race Adjustments in Clinical Algorithms Can Help Correct for Racial Disparities in Data Quality, PNAS 2024 (with Anna Zink and Emma Pierson).

  Race-blind algorithms can be worse than race-adjusted ones if data quality differs across racial groups. Using the example of family history, which is more likely to be accurate for some groups than others, we show that race adjustments can improve equity in cancer screening.

• A Playbook for Identifying and Fixing Algorithmic Bias.

  Bias audits need to find the broken label or proxy, then rebuild the tool around what patients actually need.

• Dissecting Racial Bias in an Algorithm Used to Manage the Health of Populations, Science 2019 (with Brian Powers, Christine Vogeli, and Sendhil Mullainathan).

  A widely used AI model confused health care spending with health needs, so Black patients got less help than White patients despite being just as sick.