I help healthcare teams and startups design studies that hold up and read real-world data with clinical judgment. Physician trained at Nishtar, epidemiologist trained at Emory, and a PhD candidate at the University of Michigan, with applied work spanning EHR analytics, clinical AI evaluation, and NIH and BARDA grant science.
My work sits between two rooms that rarely talk to each other: the clinic, where decisions get made under uncertainty, and the analysis, where those decisions are supposed to be defensible. I trained first as a physician in Pakistan, then as an epidemiologist at Emory, and I am now completing a PhD at Michigan on longitudinal recovery after stroke.
For industry and startup teams, that combination does a specific job. I can tell whether a real-world dataset can actually answer the question being asked of it, design the study so the estimate survives scrutiny, and translate the result for a clinical audience without losing what the numbers can and cannot support. I have built EHR and revenue-cycle pipelines, evaluated a clinical AI system for patient-portal message triage, and written the methods sections that NIH and BARDA reviewers read closely.
The through-line is restraint. I care more about an honest estimate with named limitations than a polished one that performs rigor it does not have.
Engagements range from a single methods review to standing advisory work. Each of these is something I have shipped, not a service page.
Design and analysis on EHR, claims, and registry data. I scope what a dataset can support, choose the design that fits the question, and carry it through to a defensible estimate.
Marginal structural models, inverse probability weighting, g-methods, survival and mixed-effects modeling. Delivered in R, Python, SAS, or STATA with reproducible code.
Evaluation of clinical NLP and machine-learning systems, graph analytics on messaging data, and REDCap data infrastructure. First-hand experience assessing model behavior against clinical reality.
GCP-compliant protocols, IRB submissions, data safety monitoring plans, trial registration, and FDA real-world evidence framing built in from the design stage.
Lead and co-author on NIH and BARDA proposals and peer-reviewed publications. Methods sections, study rationale, and the cover-letter case for why the work belongs where it is submitted.
Graduate curriculum development and hands-on mentoring of analysts, students, and clinical residents in study design, critical appraisal, and data management.
Both panels are drawn from my peer-reviewed study of clinical messaging at a large academic health system. Hover the network to trace how message load moves between roles. Every number is from the published paper.
One instinct runs through both of these: keep disagreement and uncertainty in view instead of quietly averaging or automating them away. An R package that keeps a confidence interval honest when its sources disagree, and a Claude Agent Skill that assesses risk of bias by a fixed, checkable rule set and refuses to sign its own name to the result.
popburden pools independent source estimates by inverse-variance weighting with DerSimonian–Laird random effects, so real between-source disagreement widens the interval rather than disappearing into the mean. It standardizes rates against a reference population, with a Poisson bootstrap interval where counts are small; converts a rate into an absolute patient count by Monte Carlo simulation; and projects that count forward under stated demographic assumptions. Every numeric routine returns an uncertainty estimate, and none of them quietly drop the confidence interval. The package ships with a test suite, a worked vignette on a clearly labeled synthetic dataset, and continuous integration that runs on every push.
srx is a deterministic judgement engine for Cochrane's RoB 2 (parallel, cluster, crossover) and ROBINS-I V2 risk-of-bias instruments, built as an Agent Skill for Claude Code. Signalling-question answers are the model's job, and every one must resolve as a verbatim quote in the ingested source text before it can be used; the domain and overall judgements are the algorithm's job, computed by a pure-function rule engine transcribed line-by-line from the official flowcharts and tables — never by a model call. Every rule cites the table or figure it came from, and every place the primary sources were ambiguous or self-contradictory is logged with how it was resolved. The output is a signed-by-nobody draft for a named human assessor, with exports to Covidence, RevMan Web, robvis, and GRADE.
A miniature version of RoB 2 domain 1, “bias arising from the randomization process.” The real engine encodes the full flowchart — five response levels per question, gap-filling rules cited to their source table — as data, not as this toggle. This is here to show the shape of the logic: judgement is computed, not modeled.
Industry, multilateral, academic, and clinical settings across the United States and Pakistan.
Lead research and data strategy at a digital-health startup building AI for clinical workflows and the healthcare revenue cycle. Advise on real-world-data analytics with graph databases (Neo4j) over structured and unstructured EHR data, co-led the technical design and grant writing for a multi-million-dollar BARDA proposal to integrate AI and real-world data into clinical workflows, contribute to NIH proposals, mentor junior staff, and co-author manuscripts on AI in public health.
Developed and taught graduate curriculum in study design, real-world-data analysis, and critical appraisal for more than 30 students. Co-led study design and protocol writing for two successful NIH research-grant acquisitions in cardiovascular epidemiology, and mentored six master's students and clinical residents on real-world-data theses, yielding three peer-reviewed publications.
Supported implementation of WHO Technical Packages and the Regional Framework for Action on Diabetes and Rheumatic Heart Disease across Eastern Mediterranean Region countries, and authored a technical report and policy briefs on non-communicable-disease management in emergency settings. Synthesized IHME GBD, WHO, and national survey data into NCD country profiles for member states including Yemen, Tonga, and Fiji, addressing confounding and bias in the multinational analyses.
Built the evidence base for a database of FDA-approved drugs repurposed for cancer, first assembled at the Morningside Center for Innovative and Affordable Medicine. Designed abstraction protocols, ran structured literature searches, appraised safety and efficacy (quality of life, progression-free and overall survival, response rate) across hundreds of studies, and independently built REDCap instruments for annotation and quality control.
Led disease-prevention work including surveillance, EPI immunization, and polio-eradication activity at the community level. Oversaw frontline health-worker performance and record-keeping, established diagnoses, and managed referral to secondary care.
Supervised development of a prehospital assessment and transport smartphone application for trauma triage, and established a trauma referral network and registry with Pakistani and United States stakeholders.
Managed inpatient and outpatient care across medical and surgical units, with rotations in neurosurgery and psychiatry, reinforcing clinical rigor and diagnostic reasoning.
Peer-reviewed work across clinical AI, EHR analytics, stroke recovery, and cardiovascular epidemiology. Author name in bold indicates first or shared-first authorship.
Full list and citation metrics on Google Scholar.
Dissertation: longitudinal recovery after intracerebral hemorrhage (BASIC study). Methods: marginal structural models, inverse probability weighting, mixed-effects survival analysis.
National conference presenter: EPI Lifestyle 2025, CSTE 2025, SER 2025, Emory Rollins 2021.
Available for consulting, advisory, and collaborative research on real-world evidence, causal inference, and clinical AI. The clearest way to reach me is email.