A benchmark that scores digital twins of Type 1 diabetes patients by decision quality: whether a fitted twin ranks candidate insulin treatments the way the real patient would, rather than by trajectory error alone. Built on 30 UVA/Padova virtual patients under 21 basal-bolus treatments, comparing MCMC, simulation-based inference, and Kalman filter twins against linear, neural, and population-prior baselines. The two criteria disagree: the worst-RMSE twin ranked treatments second best, while a twin with half the RMSE ranked at chance. Two papers submitted to the Sim2Science workshop at NeurIPS 2026.
Evaluating Digital Twins for Type 1 Diabetes by Decision Quality
An evaluation protocol that scores a Type 1 diabetes digital twin by whether its glucose trajectories rank candidate insulin treatments as the real patient would, not by trajectory error alone. On 30 UVA/Padova virtual patients under 21 treatments, the two criteria disagree: the worst-RMSE twin ranks second best, and a twin with half the RMSE ranks at chance.
Physiological Identifiability of Type 1 Diabetes Digital Twins Under Behavioral Heterogeneity
A study of when the physiological parameters of a Type 1 diabetes digital twin can actually be recovered from observed data, and how behavioral variation across patients limits that identifiability.
Gumbel-Based Active Sparse Mobile Crowd Sensing with Time Series Transformer
A learned Gumbel-noise sensor selection layer paired with a time series transformer, reducing reconstruction error on missing sensor data by up to 28% on Urban Air and SensorScope St-Bernard datasets.
Ensemble Learning with Early Fusion of Kernel-Transformed and Classical Electrocardiogram Features for Chagas Disease Detection
Ensemble framework over AutoGluon, ECG-FM, FFT, and wavelet features for 12-lead ECG classification. Placed 39th in the 2025 George B. Moody PhysioNet Challenge.
CCQ: A Multi-State Child Care Quality Dataset to Support AI for Children’s Health Research
A de-identified dataset of 64,479 child care providers across 12 U.S. states, curated from state QRIS portals by an LLM pipeline and released on Hugging Face, with within-state and leave-one-state-out benchmarks over tabular models and language models.
Reconciling Set-Valued Policy & Dead-End Discovery in Healthcare Reinforcement Learning: An Empirical Analysis
An empirical study of how consistently Set-Valued Policies and Dead-End Discovery agree on clinician-in-the-loop sepsis treatment, plus a partial ordering over recommended actions. Evaluated on LifeGate and MIMIC-III.
Patched Forecasting with Gumbel-Based Selector for Sparse Mobile Crowd Sensing
A follow-up extended abstract that convolutionally splits each sensing cycle into patches, each with its own selector layer, testing whether intra-cycle structure improves sensor selection.
Softening Overly Demanding Requirements in Recommendation System
Cuts item under-recommendation bias and training cost in Debiased Bayesian Personalized Ranking by replacing the adversarial debiasing network with an autoencoder plus ranking post-processing.
- Apr 2026 — Type 1 diabetes digital-twin benchmark
- Oct 2025 — Child care quality dataset
CCQ, a de-identified dataset of 64,479 child care providers across 12 U.S. states, collected from state QRIS portals and released on Hugging Face as row-aligned text and preprocessed tabular versions. An LLM curation pipeline adapts a hand-built Georgia reference to anonymize and clean the remaining states with open-source Qwen3 agents, and within-state and leave-one-state-out benchmarks pit tabular models against language models: trees win within-state, zero-shot transfer lands near chance, and modest target-state supervision recovers most of the gap.