Pancreatic cancer risk prediction using deep sequential modeling of longitudinal diagnostic and medication records
A transformer-based risk model trained on population-scale longitudinal electronic health records.
Machine Learning & AI for Medicine
Asif Khan
Staff Scientist, Harvard Medical School
I am a Staff Scientist at Harvard Medical School, where I work with Chris Sander and Debbie Marks on machine learning and AI methods for cancer research and clinical applications.
My research interests include representation learning and generative models for large-scale data. My current work develops scalable models for longitudinal electronic health records that are robust and well-calibrated for pan-cancer risk stratification. I also work with drug and CRISPR perturbation data to model cellular responses and guide the design of combination therapies.
Before joining Harvard, I completed my PhD in Machine Learning at the University of Edinburgh under the supervision of Amos Storkey. My doctoral research focused on geometry for deep representation learning. I continue to develop geometric methods for understanding representation learning and interpreting large language models.
Selected work
Recent publications
Capacity and Redundancy Trade-offs in Multi-Task Learning
An analysis of how shared representations balance capacity and redundancy across learning tasks.
The Geometrical and Topological Signature of Transformers
A geometric and topological framework for studying how transformer representations evolve across layers.
Research
Current directions
Longitudinal health representations
Scalable representation and generative models that learn patient health states from large-scale longitudinal clinical data.
Pan-cancer risk stratification
Robust and well-calibrated models that identify high-risk patients across cancers, populations, and clinical settings.
Perturbation models
Models of cellular response learned from drug and CRISPR perturbation data to guide combination therapy design.
Geometry and interpretability
Geometric methods for understanding learned representations and interpreting large language models.
Background
Education
PhD in Machine Learning, University of Edinburgh
2018–2023
Bayesian and Neural Systems Group (Prof. Amos Storkey)
MSc in Computer Science, University of Bonn
2016–2018
Machine learning; Knowledge Graphs; Generative models for voice morphism