Interpretability researchPurdue UniversitySoftware Engineer, AI Research2026Research
CLIP-H
A clinical prediction that arrives with its reasoning attached, as named hypotheses a clinician can read and argue with.
Problem
A prediction score tells you what may happen to a patient.
It rarely tells you why, in terms anyone can check.
Approach
CLIP-H is Clinical Latent Interpretation and Prediction via Hypotheses, with Purdue and Harvard Business School faculty.
It is under review at a NeurIPS 2026 workshop, so it is described here only at the level the title and abstract already state.
How it works
The prediction and its justification are the same object, so a clinician can disagree with one named hypothesis instead of with the number.
Find candidate concepts in the patient data.
State each in plain language and drop the ones that fail checking.
Predict from what survives.
Impact
Results are held back until the review is finished.
Built with
Python · PyTorch · MIMIC-IV · LLMs
Written inside a research group’s codebase, so there’s no public source to link. Happy to talk through the design.