I am a fifth year PhD student at Harvard-MIT Health Sciences & Technology working under Dr. Andrew Beam in the Beam Lab at Harvard T.H. Chan School of Public Health . Simultaneously, I am interning part-time as a PhD student researcher at Google, on the AMIE team.
Broadly speaking, I am highly interested in improving and evaluating medical AI with the ultimate goal of de-risking such systems for real-world clinical applications. In particular, I have significant experience developing medically-specialized LLMs (Large Language models) and CLIP (Constrastive Language-Image Pretrained) models:
I have previously received my M.S.E. and B.S. degrees in Biomedical Engineering from Johns Hopkins University. During that time, I collaborated with Johns Hopkins faculty, clinicians, and hospital administrators on a diverse set of Machine Learning for Healthcare projects, including epilepsy localization, traumatic brain injury prognosis, "precision" trans-arterial embolization, and surgical instrument tracking.
Some of my work is available as preprints on arXiv.
Conformal Prediction with Large Language Models for Multi-Choice Question Answering
In this paper, we explore conformal prediction for LLMs in multiple choice question answering, finding a strong correlation between uncertainty estimates and prediction accuracy. We demonstrate that our conformal framework produces valid prediction sets when the calibration and evaluation data come from the same distribution.
Bhawesh Kumar, Charles Lu, Gauri Gupta, Anil Palepu, David Bellamy, Ramesh Raskar, Andrew Beam
TEACH Conversational AI Workshop at ICML 2023
TIER: Text-Image Entropy Regularization for Medical CLIP-style models
We develop a regularization method for CLIP-style models which computes pairwise image-patch and text-similarity scores and uses entropy to penalize the model if these similarities are uniform. This regularization encourages text tokens to only describe a few of image patches and each image patch to only correspond to only a few text tokens.
Anil Palepu, Andrew Beam
Machine Learning for Health Care (MLHC) 2023
Towards Reliable Zero-Shot Classification in Self-Supervised Models with Conformal Prediction
We develop a conformal prediction procedure for CLIP-style models, introducing two non-conformity scores to assess when a given test caption may be reliably used.
Bhawesh Kumar, Anil Palepu, Rudraksh Tuwani, Andrew Beam
Self Supervised Learning: Theory and Practice Workshop at NeurIPS 2022
Self-Supervision on Images and Text Reduces Reliance on Visual Shortcut Features
We demonstrate through synthetically-generated watermark shortcuts on chest x-rays that supervised CNNs are heavily reliant and unable to unlearn shortcuts, and self-supervision with text helps to reduce this reliance.
Anil Palepu, Andrew Beam
2nd Workshop on Spurious Correlations, Invariance, and Stability (SCIS) at ICML 2022
Artificial Intelligence Based on Machine Learning in Pharmacovigilance: A Scoping Review
We present a scoping review to characterize the use of machine learning in pharmacovigilance and suggest opportunities for improvements.
Benjamin Kompa, Joe B Hakim, Anil Palepu, Kathryn Grace Kompa, Michael Smith, Paul A Bain, Stephen Woloszynek, Jeffery L Painter, Andrew Bate, Andrew L Beam
Drug Safety (Vol 45 Issue 5 p. 477-491), The Official Journal of the International Society of Pharmacovigilance
Digital signatures for early traumatic brain injury outcome prediction in the intensive care unit
We leveraged first-day physiology and lab data to develop a model to accurately predict mortality and end-of-stay neurological outcomes for traumatic brain injury patients in the Intensive Care Unit.
Anil Palepu, Aditya Murali, Jenna L Ballard, Robert Li, Samiksha Ramesh, Hieu Nguyen, Hanbiehn Kim, Sridevi Sarma, Jose I Suarez, Robert D Stevens
Scientific Reports, Vol 11 Issue 1 p. 1-9
Establishing a Quantitative Endpoint for Transarterial Embolization From Real-Time Pressure Measurements
We describe the design of a novel pressure-sensing multilumen catheter and present in-vitro experiments demonstrating that 1) occluded vessel pressure served as a targetable embolization endpoint and 2) off-target embolization could be characterized as a function of injection and vessel pressure
Prateek C. Gowda, Victoria X. Chen, Miguel C. Sobral, Taylor L. Bobrow, Tatiana Gelaf Romer, Anil K. Palepu, Joanna Y. Guo, Dohyung J. Kim, Andrew S. Tsai, Steven Chen, Clifford R. Weiss, Nicholas J. Durr
The American Society of Mechanical Engineers (ASME) Journal of Medical Devices, Vol 15 Issue 2
Evaluating Invasive EEG Implantations with Structural Imaging Data and Functional Scalp EEG Recordings from Epilepsy Patients
We demonstrate that concordance between non-invasive scalp EEG and invasive electrode placement was correlated to surgical success for focal-onset epilepsy, suggesting potential for non-invasive epilepsy localization
Anil Palepu, Adam Li, Zachary Fitzgerald, Katherine Hu, Julia Costacurta, Juan Bulacio, Jorge Martinez-Gonzalez, Sridevi V Sarma
41st iEEE Engineering in Medicine & Biology Conference (EMBC) 2019
Automating interictal spike detection: Revisiting a simple threshold rule
Demonstrated that a threshold, when coupled with basic signal processing, was a sufficiently sensitive and specific detector of inter-ictal spikes in intracranial EEG data.
Anil Palepu, S. Premanathan, Feraz Azhar, Martina Vendrame, Tobias Loddenkemper, Claus Reinsberger, Gabriel Kreiman, Kimberly A Parkerson, S Sarma, William S Anderson
40th iEEE Engineering in Medicine & Biology Conference (EMBC) 2018