Research Experience
Research trajectory.
From causal-network methodology to reliability and harm evaluation of multimodal clinical AI — each appointment is documented below with its scientific problem, data, and methods.
Project Co-Investigator and Visiting Scholar
Reliability & Harm Evaluation of Multimodal Clinical AI
MIT Laboratory for Computational Physiology · MIT Critical Data
2026–2027 · Active
Diagnostic-imaging models are usually reported by average accuracy — a number that hides where a model fails. In clinical use, the examinations and patients most vulnerable to model error matter more than aggregate performance. The problem I work on is how to measure, localize, and explain the potential harm of multimodal clinical AI at the level of individual examinations and patients, rather than at the level of a single headline metric.
Why it matters: a model that is accurate on average but silently unreliable for specific subgroups or care contexts is not safe to deploy. Reliable, uncertainty-aware evaluation is a precondition for clinical AI in perioperative and critical-care settings.
Linked imaging and structured clinical data under PhysioNet credentialing, with encounter-level linkage and temporal alignment.
I evaluate ensembles of deep-learning diagnostic-imaging models — on the order of twenty architectures spanning CNNs, vision transformers, and hybrids — for multi-task diagnostic classification, and quantify image-level error accumulation across the ensemble via a consensus-based "harm" score that flags the most vulnerable examinations and patients.
An end-to-end pipeline links imaging-model behavior to ICU care phenotypes, with a causal framing that tests whether AI-vulnerable patients share latent clinical or care-delivery signals.
Evaluation is the heart of the work rather than an afterthought:
- Calibration — agreement between predicted confidence and observed accuracy.
- Uncertainty quantification — explicit uncertainty on predictions and on the harm score.
- Threshold / severity-sensitivity analysis — how decisions shift as operating points move.
- Subgroup & fairness auditing — where performance and harm concentrate across patient groups.
- Instance-level error analysis — patient- and examination-level vulnerability, not just cohort averages.
Illustrative — reliability-diagram and subgroup-harm concepts; not derived from study results.
An A-to-Z breakdown — expand any section.
Retrospective critical-care cohort built from MIMIC-IV (structured EHR) linked to MIMIC-CXR (chest radiographs) at the encounter level, with temporal alignment between imaging studies and clinical events. Access is PhysioNet-credentialed under CITI Human-Subjects, GCP, and HIPAA training.
Image standardization for the diagnostic-imaging models, harmonization of multi-task diagnostic labels, and temporal alignment of EHR variables to each imaging encounter so that model behavior can be tied to the correct clinical context.
An ensemble of roughly twenty architectures — CNNs, vision transformers, and hybrids — trained for multi-task diagnostic classification. Individual architectures serve as baselines; a consensus-based harm score aggregates image-level error across the ensemble to flag the examinations and patients most vulnerable to model error.
Calibration assessment, explicit uncertainty quantification, and threshold / severity-sensitivity analysis that traces how flagged-harm decisions move as operating points shift — so the harm score is reported with, not without, its own uncertainty.
Subgroup and fairness auditing across demographic and clinical strata, with instance-level error analysis to surface where harm concentrates. A causal framing then tests whether AI-vulnerable patients share latent clinical or care-delivery signals, linking imaging-model harm to ICU care phenotypes.
The work is retrospective and built on a single ICU data source; the consensus harm score is a measure of model agreement, not adjudicated clinical harm. These are deliberate scoping choices — the aim is a rigorous evaluation methodology that motivates prospective, perioperative validation rather than a deployed tool.
This work anchors my independent research programme: trustworthy, uncertainty-aware evaluation of multimodal clinical AI, with methods that transfer directly to perioperative monitoring, critical-care physiology, and clinically grounded AI evaluation. It is methodologically distinct from — though complementary to — the mechanistic modeling I contribute to at Duke.
See the research program →Lead Data Scientist & Biostatistician
Surgical Decision Support for Epilepsy Outcomes
Duke Comprehensive Epilepsy Center · Department of Neurology
2026–2027 · Epilepsy Surgery Outcomes Project
I lead the statistical design and reproducible analysis for a temporal-lobe-epilepsy study evaluating whether the SEEG-derived 5-SENSE focality score predicts postoperative seizure freedom.
I assess discrimination, sensitivity and specificity, predictive values, ROC/AUC with bootstrap confidence intervals, calibration, and clinically meaningful thresholds across resection, laser interstitial thermal therapy (LITT), and neuromodulation.
I investigate why some high-score patients do not achieve the expected outcome using histology and imaging-concordance variables, Fisher and penalized-logistic models, structural-missingness handling, and prespecified sensitivity analyses.
The aim: identify when the score may inform patient selection and treatment counseling, and where its limitations should constrain use.
Postdoctoral Associate
ML-Integrated Multiscale Modeling for Biomedical Systems
Duke University · Department of Biostatistics & Bioinformatics
2024–2026 · NIH-funded Duke–Weill Cornell Medicine collaboration (congenital CMV)
Biomedical systems span molecular, cellular, and tissue scales, and neither purely mechanistic simulators nor purely statistical learners capture them well alone. The problem is how to couple mechanistic models with machine learning so that models are fast and data-efficient while still respecting the biology they represent.
Why it matters: hybrid mechanistic–statistical modeling underlies in-silico vaccine-efficacy analysis and maternal–fetal immune analytics on the NIH/NIAID R01 (AI173333) congenital-CMV project, and generalizes to physiologic and pharmacologic systems relevant to perioperative care.
Heterogeneous cytometric and omics measurements integrated across sources, plus simulator-generated data from mechanistic models.
Reproducible frameworks that couple mechanistic simulators (agent-based models, ODEs) with classical ML, LLM workflows, and deep architectures:
Gradient-boosted trees
Supervised emulation with Bayesian hyperparameter optimization.
Gaussian-process surrogates
Uncertainty-aware emulation of expensive simulators.
Physics-informed neural nets
Embedding domain constraints into learned dynamics.
Optimal transport (relaxed marginals)
Distribution alignment on Gaussian-mixture embeddings.
End-to-end cross-dataset harmonization for cytometric / omics measurements, engineered for reproducibility:
Mechanistic simulators are expensive to evaluate; a Gaussian-process surrogate emulates them cheaply while carrying an explicit uncertainty band that widens away from training points.
Illustrative — Gaussian-process surrogate emulating a mechanistic simulator; not derived from study results.
Workflows are containerized and scaled on HPC so that analyses are portable and repeatable — the same reproducibility discipline that clinical-AI governance requires.
An A-to-Z breakdown — expand any section.
Within the NIH/NIAID R01 AI173333 congenital-CMV program, the mechanistic backbone couples maternal viral and immune ODEs, a one-dimensional placental diffusion–decay interface, and stochastic fetal infection. Calibration data include rhesus-macaque transmission studies (Gong et al., Viruses, 2023) and trophoblast-organoid readouts; harmonization work operates on flow-cytometry and omics measurements.
End-to-end pipelines apply quality assurance, batch correction, and feature engineering, then align distributions via optimal transport with relaxed marginals on Gaussian-mixture embeddings, integrating Leiden / igraph graph clustering with Wasserstein-grounded distances.
Gradient-boosted trees with Bayesian hyperparameter optimization, Gaussian-process surrogates for uncertainty-aware emulation, and physics-informed / neural-ODE surrogates that compress the calibrated simulator into a fast, differentiable proxy while preserving its conservation laws and constraint structure.
The calibrated multi-scale system reproduces the transmission rates observed in immunocompetent dams, the elevated transmission under CD4+ T-cell depletion, and the timing-dependent failure of late hyperimmune-globulin administration — behavioral checks that constrain the surrogate before it is used for virtual-trial analysis.
Uncertainty is carried through Gaussian-process residuals over the mechanism gap, so the surrogate reports principled uncertainty under realistic sampling noise and heterogeneous infection timing. Workflows are containerized with Docker and scaled on SLURM / HPC for portable, repeatable analysis.
Validation to date is pre-clinical and in-silico (animal-model and organoid calibration). Perioperative transfer — anesthetic PK/PD, hemodynamic, and ventilator emulation — is the explicit next step, reusing the same constraint-preserving surrogate construction.
The Duke work demonstrates hybrid mechanistic–statistical modeling with explicit uncertainty and reproducible infrastructure — directly transferable to computational pharmacology, physiologic emulation, and integrative modeling of biological and physiologic systems across scales.
Related manuscripts in preparation →NSF NRT Research Fellow · AI & Machine Learning for Digital Health
Scalable Causal-Network Inference & Interpretable ML
National Science Foundation · BridgesDH NRT (Award #2125872) · WVU Cancer Institute
2023–2024 · Two-year, stipend-bearing traineeship
Causal-structure discovery from high-dimensional biomedical signals is statistically hard and computationally expensive. The problem is how to infer reproducible causal networks at scale, with principled control of false discovery and honest assessment of stability.
Why it matters: reliable causal discovery underlies biomolecular and clinical-relation inference — and became the methodological foundation for BaMANI and for uncertainty-aware causal phenotyping in clinical-AI applications.
A platform integrating probabilistic graphical models with both constraint-based and score-based structure learning:
- Conditional-independence testing with multiple-hypothesis correction.
- Bootstrapping for edge-stability assessment.
- Ensemble supervised & unsupervised learning with feature selection.
- Interpretability — permutation importance, partial dependence, and LIME.
- CNN-based signal decomposition separating aggregated molecular / tissue-scale signals into interpretable components.
Structure learning proposes directed relationships; bootstrapping across resamples scores each edge, so high-support edges (solid) are distinguished from uncertain ones (dashed).
Illustrative — causal graph with bootstrap-stability edge weighting; structure is schematic, not study output.
The platform improved computational efficiency for targeted high-dimensional causal-discovery workloads, alongside interpretable ensemble pipelines and deep-learning signal decomposition.
An A-to-Z breakdown — expand any section.
High-dimensional biomedical signals — RNA and tissue gene-expression and other multi-source datasets — with a focus on inferring cellular drivers of disease progression and drug response, including tumor–immune communication relevant to immuno-oncology.
Probabilistic graphical models combining constraint-based and score-based structure learning, with conditional-independence testing under multiple-hypothesis correction and bootstrapping for stability. Single-algorithm structure learning provides the baseline the ensemble improves on in both stability and efficiency.
Ensemble supervised and unsupervised learning with feature selection, made interpretable through permutation importance, partial-dependence analysis, and LIME — so inferred relationships can be inspected rather than taken on faith.
CNN-based decomposition separates aggregated molecular and tissue-scale signals into interpretable components, including computational deconvolution that estimates cell-type abundance and phenotype from RNA and tissue gene-expression profiles — applicable to computational pathology and clinically informed signal decomposition.
The platform improved computational efficiency for targeted high-dimensional causal-discovery workloads by roughly 20% and provided a web-based analytical interface — the methodological foundation later carried into BaMANI.
Causal structure is learned from observational data and scored for stability rather than validated by intervention; edges are hypotheses with calibrated support, and the platform is designed to make that uncertainty explicit.
This fellowship established my through-line in causal discovery, interpretability, and methodological scalability — the basis for causal AI applied to biological and clinical systems.
BaMANI & related work →Machine Learning Researcher · Ph.D. Researcher
Generative & Causal Machine Learning for High-Dimensional Biology
West Virginia University · WVU Cancer Institute · Machine Learning in Systems Biology Lab
Machine Learning Researcher, WVU Cancer Institute · 2020–2023
Ph.D. in Applied Mathematics, West Virginia University, conferred 2024
High-dimensional biological data are scarce, noisy, and hard to reason about causally. The problem is how to generate faithful synthetic biological data and recover robust causal structure under uncertainty, so downstream biomedical predictions generalize.
Why it matters: this foundational work shaped my broader program in machine learning, causal inference, and trustworthy modeling — and produced the Nature Communications analysis of how oncogenic gene expression reshapes heterocellular networks.
VAEs
Representation learning with latent regularization and disentanglement.
WGAN-GP
Wasserstein GANs with gradient penalty for stable synthesis and class-conditional generation.
Ensemble Bayesian networks
MCMC posterior exploration over graph spaces for robust causal discovery.
Deconvolution & graphical workflows
Cell-type abundance and gene-regulatory analysis in cancer systems.
Generative models are judged by how closely synthetic samples reproduce the real data distribution; a tight overlay indicates faithful synthesis for augmentation and domain adaptation.
Illustrative — overlay of real and generated distributions; not derived from study results.
An A-to-Z breakdown — expand any section.
Cancer systems and network immunology: characterizing how CCN4 / WISP1 reorganizes local regulatory context linked to anti-tumor immunity in breast cancer and melanoma (Nature Communications, 2022). Data span bulk expression and single-cell / single-nucleus RNA-seq assembled into in-silico pseudo-bulk references.
VAEs and Wasserstein GANs with gradient penalty (WGAN-GP) for synthesis and representation learning, using latent regularization, disentanglement, and class-conditional generation, with feature-space augmentation for downstream biomedical prediction.
Ensemble Bayesian structure learning with MCMC posterior exploration over graph spaces for robust causal discovery and parameter estimation, deployed across multi-source datasets with explicit uncertainty quantification for downstream reasoning.
Cell-type estimation under platform and biological variability, with reference pipelines evaluated against community-assessment benchmarks for deconvolution methods — grounding the generative augmentation in externally validated tasks.
Synthetic samples generated by the models support domain adaptation in biomedical prediction tasks, improving transfer where labeled data are scarce or heterogeneous.
This is foundational methodological work in high-dimensional biology; it establishes the generative, causal, and uncertainty-quantification capabilities that the later BaMANI, harmonization, and clinical-AI directions build on.
The WVU work is the root of my research identity: generative modeling, ensemble Bayesian causal discovery, and uncertainty quantification for high-dimensional biology — methods I now carry into trustworthy clinical AI.
Nature Communications & foundations →