Grouped by topic; a paper can sit under more than one. This list matches my Google Scholar profile.
* equal contribution† corresponding authorLead-authored: first, co-first or corresponding author
61 papers
Sem-ECE evaluates LLM calibration in open-ended QA by sampling answers, grouping them into semantic classes, and using class frequencies as confidence; two estimators are analyzed and tested on five commercial LLMs.
A native 3D vision-language framework for structured CT and MRI report generation that handles multiple studies per patient, keeps vision-text alignment with a joint generative-contrastive objective, and outperforms 2D and 3D medical VLMs on the Penn Medicine BioBank.
Replaces post-hoc temperature scaling with a bilevel objective that maximizes predictive entropy, solved by a first-order method that scales to LLMs; calibration gains are strongest out of domain.
A recommendation architecture designed for predictable power-law scaling: it raises model FLOPs utilization from 17% to 37% on NVIDIA B200 GPUs, doubles scaling efficiency, and is deployed in major Meta Ads models.
Percival, a CT vision-language model trained on over 400,000 CT-report pairs from the Penn Medicine BioBank, learns representations that track clinical, physiological, and longitudinal variation across patients.
TAP-GPT fine-tunes a tabular LLM for few-shot Alzheimer's disease classification from multimodal biomarker tables, with interpretable outputs, across four ADNI-derived datasets.
IRIS learns patient risk clusters directly from data while estimating feature contributions, trading a modest loss in time-to-event accuracy for interpretable risk stratification.
First- and zeroth-order stochastic online bilevel algorithms that achieve sublinear stochastic bilevel regret without window smoothing, using a new search direction and fewer oracle calls.
Adapts TableGPT2 with few-shot tabular prompts and qLoRA fine-tuning to diagnose Alzheimer's disease from small structured biomarker datasets.
A fair CCA method whose projected features are independent of sensitive attributes, improving fairness in downstream classification on ADNI neuroimaging data while keeping correlation performance.
A benchmark that combines synthetic counseling dialogues with anonymized coach and caregiver session transcripts, for building and evaluating conversational mental health assistants.
Links post-alignment overconfidence to preference collapse and restores calibration with calibration-aware fine-tuning plus an EM-based ECE regularizer.
Fine-tuning only the attention modules improves weight disentanglement for model merging by task arithmetic, without the extra training cost of NTK linearization.
FREEFORM uses chain-of-thought reasoning and ensembling to let LLMs select and engineer genotype features, outperforming data-driven methods especially in low-shot settings.
Mines social determinants of health from literature with LLMs and merges them with Alzheimer's disease entities from PrimeKG, evaluated by link prediction with graph neural networks.
Survival analysis on EHR data from 29,045 older patients examines how medications, laboratory tests, and sensitive attributes affect fall and injury risk in neurodegenerative disorders.
Compares univariate and multivariate machine learning methods on EHR data from 70,420 patients to identify risk features for Alzheimer's, Parkinson's, and other dementias.
A two-stage LLM topic modeling approach for 231,870 dementia caregiver tweets, compared against 11 baselines including LDA, GSDMM, and BERTopic.
Fair multi-modal CCA optimizes correlation and demographic fairness jointly, improving fairness metrics on ADNI neuroimaging data with little loss in correlation.
A multi-stage ensemble feature selection method for vertical federated learning that accounts for feature interactions across clients using conditional feature synthesis.
Builds a knowledge graph of about 92k triplets linking social determinants from literature and EHR to Alzheimer's disease biology, evaluated by link prediction and validated with multimodal biological data.
Adapts 2D medical vision-language models to 3D CT and MRI through slice selection and instruction tuning; selecting about half of the slices gave the best accuracy while cutting training cost by over 40%.
Fuses modality-specific encoders with a mixture of Weibull experts to discover patient subtypes and predict Alzheimer's conversion risk in a cohort of 382 MCI patients.
Adding social determinants of health to multi-modal deep clustering survival models improved injury and fall risk prediction for patients with Alzheimer's and related dementias.
A pipeline to embed brain sub-circuits derived from functional MRI into echo state networks and evaluate how the embedded topology affects processing and memorization tasks.
A review of computational RNA secondary structure prediction, RNA modification identification, how the two interact, and the data sources available.
Combines retrieval with existing detectors to detect LLM-generated text more reliably when it has been paraphrased.
Fair estimation of Gaussian, covariance, and Ising graphical models by jointly minimizing a pairwise graph disparity error and the model loss.
Co-augments an LLM and a knowledge graph built from Alzheimer's disease literature, retrieving knowledge from the graph to improve question answering on a new ADQA benchmark.
Term-weighted GSDMM topic modeling of tweets from dementia caregivers, with topics interpreted using ChatGPT 4 and two human experts.
A fairness-aware method for label-imbalanced MRI classification of mild cognitive impairment, built on FACIMS and evaluated across five sensitive attributes.
Introduces online bilevel optimization, defines bilevel regret, and gives regret bounds for an online alternating time-averaged gradient method.
Classifies long therapist-caregiver interview transcripts and uses Shapley values to find the sentences that drive the predictions.
Fair empirical risk minimization that builds group prevalence ratios into the fairness constraint, balancing accuracy and fairness.
Clusters patients with mild cognitive impairment into two subtypes from 68 cortical amyloid PET features.
A model that derives decision boundaries directly from data by discovering its manifold structure, instead of a parameterized partition of the sample space.
Multi-group tensor CCA with a dual sparsity structure that jointly analyzes subgroups of brain imaging data to find shared and group-specific signals.
Deep Clustering Survival Machines model survival times as a mixture of expert distributions weighted per patient, used to cluster MCI patients into subgroups with different risks of converting to Alzheimer's disease.
An autoencoder that mixes conventional and quadratic neurons, motivated by an approximation result for heterogeneous networks, applied to unsupervised anomaly detection.
A distance-weighted Sinkhorn loss that penalizes misclassified samples by their distance to the true class distribution, applied to Alzheimer's disease stage classification.
Learns global CCA projections that close the gap in correlation across protected groups with little loss of accuracy.
Multi-group tensor CCA that jointly analyzes subgroups of tensor data, such as brain imaging from different sexes or races, to find consistent signals.
VS-Opt-Net adds the vector-scaling loss and Bayesian hyperparameter optimization to the STREAMLINE pipeline to improve balanced accuracy when classifying dementia stages from MRI features.
A tri-level optimization framework with sharpness-aware minimization for fair and class-balanced learning when subgroups are imbalanced and small.
Models survival times as a mixture of expert distributions whose weights are learned per instance, yielding interpretable subgroups alongside predictions.
Measures how a widely used deep learning model underdiagnoses or overdiagnoses glaucoma across demographic groups in the OHTS study.
Online learning from doubly-streaming data, where new features emerge and old ones fade, using deep variational models to relate the two feature spaces.
Builds a weakly labeled chest X-ray database from PubMed Central articles and shows that adding it to training improves detection of several thoracic diseases.
Shows that DETReg pre-training does not help strong DETR-based detectors with full data, and that simple self-training with an improved box predictor and Objects365 does.
OLD³S learns a shared latent space linking old and new features in doubly-streaming data and adapts model depth and size online to the data.
GlaucomaNet uses two CNNs that mimic human grading to diagnose glaucoma from fundus photographs, evaluated on OHTS and LAG.
Feature-evolvable streaming learning with rarely provided labels, using manifold regularization over a buffer that fits the device's storage budget.
Online learning over feature spaces that change with time and mix Boolean, ordinal, and continuous variables, modeled with a Gaussian copula fitted by online EM.
Learning when features vanish or emerge unpredictably: the incomplete overlap period is filled by matrix completion, with a bound on the observations needed.
A proof-of-concept system that uses Clinical BioBERT to retrieve clinical trial reports relevant to long clinical questions, across three medical domains.
Learns finite state automata from gated RNNs as an interpretable structure and studies how the number of gates affects performance.
Learns finite state automata from RNNs when input features change, and studies how feature changes and the number of gates affect performance.
Journal version of feature evolvable streaming learning: recovers vanished features from the overlap period and combines models with theoretical guarantees.
Multiple instance deep forest (MIDF), a deep forest structure adapted to multi-instance learning.
Introduces learning when old features vanish and new ones appear, using an overlap period to keep reusing the old model, with theoretical guarantees.
Graph-based semi-supervised learning that adapts to a storage budget through low-rank approximation of the similarity matrix.