Research question

Can heterogeneous oncology EHR and authorization records be transformed into analysis-ready evidence suitable for studying documentation readiness, authorization outcomes, and administrative workflows?

My contribution

I personally developed and evaluated research pipelines and predictive models in Python, using data-quality checks, harmonization, structured and text extraction, cohort definition, and model evaluation.

Methods

  • Prepared retrospective records across 24,167 patients, 200,513 authorization case rows, and 148,703 clinical notes.
  • Combined structured clinical data and text-based information extraction, with field-level evaluation.
  • Evaluated rare-outcome denial-risk prediction using patient-grouped validation, PR-AUC and ROC-AUC, calibration, and temporal checks.
  • Examined retrospective workflow timing and simulated potential improvements under defined assumptions.

Selected results

0.97Mean F1 across six extracted fields
0.88Patient-grouped denial ROC-AUC
0.29Denial PR-AUC, retrospective test

Evidence boundary: Metrics reflect retrospective evaluation. Workflow improvements were modeled, not prospectively observed in routine clinical operations. Research data and protected information are not posted here.

Decision relevance

Model evaluation informed what appeared technically feasible, where data quality constrained interpretation, and what prospective instrumentation would be necessary before live clinical claims.

Read the public research documentation ↗