On 26 November 2024, Kisumu County leadership launched a cancer epidemiology dashboard built on multi-year oncology records from Jaramogi Oginga Odinga Teaching and Referral Hospital (JOOTRH). I served as Research Data Manager under Principal Investigator Dr. Thomas Odeny (Washington University in St. Louis), with KEMRI and county partners.
This note describes the data system and my role — not a product pitch.
Why the work mattered
A retrospective review of oncology records at JOOTRH (roughly 2013–2023) made the continuity problem concrete:
- About 59% of patients were lost to follow-up (LTFU) within the first year of care
- 5-year survival in the analysed cohort was on the order of 9%
- Thousands of patient records (on the order of 3,400+ in the core analytic sample; later extracts grew as the registry was extended)
- Advanced-stage presentation was common
Those figures are clinical and operational: if patients disappear from care, treatment plans fail even when diagnosis was possible. A usable surveillance product had to sit on trusted records, not only on pretty charts.
My role in the data lifecycle
I owned research data management across the pipeline. The PI and clinical partners owned clinical interpretation and programme direction; I owned instruments, quality, analysis support, and dashboard data products.
1. Data capture architecture (REDCap)
- Designed REDCap forms for multi-year oncology chart abstraction
- Validation rules and branching for treatment pathways
- Fields spanning demographics, cancer type, staging, treatments, and outcomes
2. Team and protocols
- Recruited and trained research assistants in medical chart abstraction
- Standardised extraction protocols and SOPs
- Supervised abstraction across large volumes of paper and electronic charts
3. Quality assurance
High-frequency and batch quality checks used:
- Python for automated validation scripts
- R for statistical quality checks
- Stata for consistency checks on clinical fields
Discrepancies went back to re-abstraction. Patient confidentiality and access control were treated as non-negotiable.
4. Analysis support
- Survival analysis (including Kaplan–Meier approaches)
- LTFU patterns by cancer type and stage
- Geographic patterns (Kisumu and neighbouring counties)
- Treatment pathway summaries for clinical and county audiences
The first-year LTFU finding was one of the results that most clearly justified a live monitoring product.
5. Dashboard product (Tableau)
Built a Tableau dashboard for programme and county users covering:
- Demographics and case mix
- Cancer type distribution
- Stage at diagnosis
- Treatment and follow-up signals
- Geographic patterns
- Survival / outcome views where data allowed
Patient-level views were restricted and de-identified as required by the study protocol and partner rules.
What the product was for
The dashboard was a decision-support and surveillance surface, not a marketing demo. Practical uses included:
- Seeing where follow-up was failing
- Supporting outreach and re-engagement conversations
- Making case mix and geography legible for planning
- Giving county and facility leaders a shared evidence product
Claims about “first in the region” are easy to overstate; what is defensible is that this was a serious facility–county surveillance product built on multi-year abstracted records and launched with county leadership.
Launch and public record
The November 2024 launch included county leadership (including Governor Anyang’ Nyong’o), health officials, the PI, JOOTRH cancer leadership, and clinical staff. Coverage is public:
Constraints we actually hit
| Constraint | What we did |
|---|---|
| Incomplete older charts | Multi-level validation; re-abstraction for critical fields |
| Inconsistent documentation over a decade | Codebooks, SOPs, peer review of extracts |
| Connectivity and power limits | Batch workflows; offline-friendly capture habits where needed |
| Staff turnover on abstraction | Training modules and continuous supervision |
What this demonstrates
- End-to-end research data management in a clinical registry setting
- HFC and multi-language QC (Python / R / Stata) under real record quality
- Translation from analysis to a county-facing product
- Clear role boundary: data systems lead under a clinical PI