MEDamorphosis Podcast · UBC Medicine Learning Network

S6 E2: AI in Medicine Series - Dr. Leo Anthony Celi

·40 min·2 clips

Leo argues that fairness is not a machine-learning problem and calls attempts to define it “AI arrogance and ignorance.”

1. Metamorphosis, presented through the UBC Medicine Learning Network, examines AI in healthcare in S6 E2 with Dr. Leo Anthony-Sely. 2. Monica, a second-year UBC medical student, interviews Leo, an intensive-care physician at Beth Israel DeGeneres Medical Center and researcher at MIT’s Laboratory for Computational Physiology. 3. The episode asks whether data science and AI can improve healthcare without reproducing the inequities embedded in existing medical data. 4. Leo describes MIT Critical Data as a global consortium focused on putting “data and learning at the front and center of healthcare.” 5. MIT Critical Data brings healthcare professionals together with computer scientists, engineers, social workers, sociologists, and anthropologists. 6. Leo says routinely collected electronic health records should produce knowledge that is translated into more equitable patient-care processes. 7. He reports that 80% of clinical trials occur in 20% of countries and that 10% of health research spending concerns diseases affecting 90% of the global population. 8. Leo argues that clinical-trial findings from 1995 may no longer represent current medical truth because “the ground truth continuously shifts.” 9. He proposes complementing expensive, slow clinical trials with data science that learns from natural experimentation in hospitals and clinics. 10. Leo says precision medicine requires large datasets and data-science education from grade school through medical and graduate school. 11. Leo summarizes AI with the phrase “no data, no artificial intelligence” and describes computers learning relationships between decisions and patient outcomes. 12. He distinguishes statistical bias from data problems that create an imperfect representation of the real world. 13. Leo says models trained on problematic data can encode the health disparities already present in healthcare systems. 14. He explains that pulse oximeters use infrared technology and have been known since the 1970s to perform less consistently for people of color. 15. He describes a Veterans Administration COVID-19 study in which smoking and being overweight appeared associated with better outcomes because of sampling selection. 16. Leo warns that a hospital database can exclude people who are not sick enough to seek admission, producing associations that do not represent the wider population. 17. Leo says fairness requires economists, social scientists, patients, and caregivers because “fairness is not a machine learning problem.” 18. He rejects a prediction about radiologists disappearing within five years and says future healthcare should integrate humans and AI rather than place them in separate systems. 19. Medical students who connect with computer scientists and affected communities. 20. Listeners seeking only short-term AI predictions may skip this systems-focused discussion.

Listen to the show on

Apple Podcasts Spotify

Play clips from this episode More from MEDamorphosis Podcast