Ben Lengerich - Machine Learning & AI in Healthcare | Ep. 3
Ben Lengerich — ML Researcher
Ben Lengerich of the University of Wisconsin–Madison on models that adapt to each patient and still explain themselves, and why the odd patterns in hospital data point to better care.
Key Takeaways
Keep the model readable
A context-adaptive model keeps a linear model's coefficients and lets them shift with each patient's history. 9:59
Accuracy is not handed down
Accuracy is a metric people choose and are tempted to flatter. Interpretability lets you check what a model has really learned. 14:07
Treated patients can look healthy
In hospital data the lowest mortality sits with very high creatinine, because those are the patients on dialysis. 36:29
Surprises show where care can improve
Where a treated, sicker group does better than a healthy one, some other group is getting worse care than it could. 40:05
A model does not conserve effort
Doctors found better-grounded anomalies; the LLM checked every feature and found almost twice as many. 43:35
“Accuracy is not a metric given to us by God.”
“I hope in 10 years, we understand how to return medicine to being a form of an art.”
Chapters
- 0:00Intro1 min
- 0:49What computational biology is1 min
- 1:34Advice for students2 min
- 3:18Deep learning in biology since 20184 min
- 6:49The Lengerich lab1 min
- 8:14Context-adaptive models4 min
- 11:50Why interpretability matters5 min
- 16:37Why large models are hard to read3 min
- 19:45Do LLMs reason?5 min
- 24:27Personal without overfitting3 min
- 27:27Industry or academia2 min
- 29:05Starting a lab and teaching5 min
- 34:07When sick patients look healthy6 min
- 40:26Catching paradoxes automatically6 min
- 46:16Where this fits in a hospital3 min
- 49:41Who pays for the research5 min
- 54:34Where the field is strongest5 min
- 59:28Medicine in ten years4 min
Mentioned in This Episode
About this Guest
Ben Lengerich
ML Researcher
Leads the Adaptive Inference Lab, where his work on contextualized machine learning builds models that adapt to each patient instead of averaging across them. PhD from Carnegie Mellon, postdoc at MIT CSAIL and the Broad Institute.