Automation and Algorithmic Quality Assurance in Clinical Genomics

Abstract

Deciphering the clinical impact of genetic mutations is high-stakes work. If incorrect data enters a clinical system, oncologists might receive flawed treatment recommendations. Guided by the 2017 CAP/AMP/ASCO standards, this thesis details my work using Python and data science to fully automate Velsera’s "Delta QA" pipeline. Delta QA is a weekly testing cycle the company runs to validate updates to their clinical genomics database before those updates go live to hospitals. Historically, this was a tedious and exhausting task. Biologists had to manually pick test variants, try to guess which rules should apply based on eight different medical databases, and then stare at massive Excel spreadsheets to verify if the software engine behaved correctly. My primary goal with this research was to see if a math-driven algorithm could completely replace this human guesswork, all while maintaining strict clinical safety standards. To tackle this, I developed a set of four automated Python tools relying on core data science concepts. I wrote greedy algorithms to ensure the selected test data remained entirely unbiased and mathematically diverse. I used regular expressions (Regex) to clean up messy biological text. I also implemented cryptographic hashing (MD5) to track past tests and prevent redundant work, while utilizing mathematical set theory to automatically flag errors. To ensure the clinical team could actually use these tools without needing programming skills, I packaged the entire suite into standalone Windows applications. By taking the human element out of the equation and relying on strict mathematical constraints, this pipeline reduced the total processing time from 10-12 hours down to about five minutes. It eliminated human selection bias entirely and proved to be significantly more accurate than human reviewers, essentially setting a new baseline for how clinical knowledgebases should be maintained.