https://belalzaky.uk Drug-safety data, and what it does and does not represent. en-gb Tue, 28 Jul 2026 00:00:00 GMT https://belalzaky.uk/projects/pv-intake https://belalzaky.uk/projects/pv-intake Tue, 28 Jul 2026 00:00:00 GMT Project The front half of a real pharmacovigilance workflow, end to end: read a medical abstract, decide whether it reports a possible adverse drug event, pull out the drug and the effect, and map that effect to a standard SNOMED CT code — a triaged, coded worksheet a safety team could review. https://belalzaky.uk/projects/rwe-causal https://belalzaky.uk/projects/rwe-causal Thu, 23 Jul 2026 00:00:00 GMT Project Association isn't causation. Using a new-user active-comparator design with propensity-score matching and a self-controlled case series — on data with a sealed true effect — I show how a naive comparison understates a real harm by nearly half, and how two independent causal methods recover it. https://belalzaky.uk/projects/adr-reader https://belalzaky.uk/projects/adr-reader Wed, 15 Jul 2026 00:00:00 GMT Project Four different models — a keyword matcher, a general-purpose embedding model, a domain-tuned biomedical transformer, and a zero-shot LLM — read the same medical text and try to flag adverse drug events. Comparing where each one succeeds and fails is more revealing than any single accuracy number. https://belalzaky.substack.com/p/i-turned-my-data-project-into-a-tool https://belalzaky.substack.com/p/i-turned-my-data-project-into-a-tool Fri, 10 Jul 2026 00:00:00 GMT Writing Taking a personal analysis script and building it into a Streamlit tool anyone can use. https://belalzaky.uk/projects/ehr-survival-r https://belalzaky.uk/projects/ehr-survival-r Mon, 06 Jul 2026 00:00:00 GMT Project A time-to-hypertension survival analysis on Synthea EHR data — Kaplan-Meier curves, a Cox proportional-hazards model, and an honest limitations write-up. Diabetic patients develop hypertension ~2.4× faster; the model's apparently-null age effect turned out to be time-varying. https://belalzaky.substack.com/p/most-reported-isnt-most-dangerous https://belalzaky.substack.com/p/most-reported-isnt-most-dangerous Thu, 02 Jul 2026 00:00:00 GMT Writing Why the most frequently reported drugs in FAERS aren't the ones with the worst safety signals. https://belalzaky.substack.com/p/what-12-million-reports-of-drug-ineffective https://belalzaky.substack.com/p/what-12-million-reports-of-drug-ineffective Thu, 18 Jun 2026 00:00:00 GMT Writing Digging into the most common reaction in FDA adverse-event data and what it actually tells us. https://belalzaky.uk/projects/faers-explorer https://belalzaky.uk/projects/faers-explorer Wed, 01 Oct 2025 00:00:00 GMT Project A live interactive Streamlit dashboard over the FDA Adverse Event Reporting System. Pick any drug and see its top reported reactions, reports per year, and age/sex demographics — with a built-in panel on why these counts measure attention, not harm. https://belalzaky.uk/projects/faers-signal https://belalzaky.uk/projects/faers-signal Mon, 01 Sep 2025 00:00:00 GMT Project A pharmacovigilance disproportionality analysis — PRR, ROR, chi-squared, and Evans criteria — over the full FAERS database via the openFDA API, validated on the known warfarin-haemorrhage signal (PRR ~5) and checked against negative controls. https://belalzaky.uk/projects/faers-sql https://belalzaky.uk/projects/faers-sql Fri, 01 Aug 2025 00:00:00 GMT Project Loaded ~3,000 FAERS reports from the openFDA API into a SQLite database (one row per drug-reaction pair) and queried it with SQL — GROUP BY, WHERE, COUNT(DISTINCT). Key lesson: how sampling choices shape conclusions. https://belalzaky.uk/projects/synthea-ehr https://belalzaky.uk/projects/synthea-ehr Tue, 01 Jul 2025 00:00:00 GMT Project Used Synthea synthetic patient records (CSV) loaded into SQLite to learn the relational structure of EHR data via JOINs and subquery cohorts — including building a hypertension cohort and examining its medications as a real-world-evidence exercise. https://belalzaky.uk/projects/ehr-ml https://belalzaky.uk/projects/ehr-ml Sun, 01 Jun 2025 00:00:00 GMT Project A machine-learning model predicting hypertension from Synthea EHR features — logistic regression and random forest, with honest treatment of data leakage, the precision/recall threshold trade-off, and the limits of synthetic-data performance.