Zimbabwe – Transforming Routine Health Data Use Through AI-Supported Automation
EvalCommunity Academy · Case Study
Zimbabwe – Transforming Routine Health Data Use Through AI-Supported Automation
How a modular, co-designed data platform cut reporting delays, removed critical data errors and helped district teams in Zimbabwe’s HIV programme start using their own data for decisions.
Country
Zimbabwe
Sector
HIV programme M&E
Implementer
OPHID with MoHCC
Period studied
Jan 2023 – Jun 2024
What you will learn
- Why routine health data systems in low- and middle-income countries stall, and where automation helps
- How a modular platform was co-designed with district M&E staff and scaled in phases
- Which reporting outcomes changed, and how strong the evidence behind them is
- How to tell AI components apart from rules-based automation when assessing a digital intervention
The programme at a glance
15
districts (of 64 nationally)
335+
public health facilities supported
345k+
clients on ART, about a quarter of the national total
20+
indicators reported to MoHCC, PEPFAR and the Global Fund
60+
pages per district report, across 15 reports each quarter
4+
full-time M&E officers per district absorbed by reporting
1. Context
Zimbabwe has an adult HIV prevalence of about 11.3% and roughly 1.3 million people living with HIV, so routine programme data drive decisions on resources and epidemic control. The national health information system is built on DHIS2, alongside electronic patient monitoring, laboratory and logistics systems.
In practice, the system was still hybrid. Data started in paper registers at facilities, were summarised onto monthly forms, typed into DHIS2 at district level, and then downloaded, cleaned and re-entered for donor platforms such as PEPFAR’s DATIM. The Organisation for Public Health Interventions and Development (OPHID), a technical partner to the Ministry of Health and Child Care (MoHCC), supports HIV services and data systems in 15 of the country’s districts.
2. The problem
Every manual hand-off added a chance for transcription errors, omissions and delays. Each district produced monthly, quarterly and ad hoc reports covering more than 20 facilities, and formats and disaggregation rules differed by stakeholder. National teams then had to read and synthesise 15 long district reports within tight deadlines, which left little room for analysis or for feedback to facilities.
The cost: skilled M&E staff spent most of their time compiling, cleaning and re-formatting data instead of interpreting it, and reporting flowed one way, up to donors, with little coming back to the districts that produced it.
3. The intervention: a modular data platform
OPHID built the OPHID Modular Data Intelligence Platform (OMDIP). Rather than replacing national systems, it sits on top of them and connects to DHIS2, DATIM, Open Data Kit, Excel, Power BI and electronic health records. It has five modules, each targeting one bottleneck in the data lifecycle.
Report Automation
Pulls data from DHIS2 and generates complete quarterly Word reports, with charts, indicator cascades and narrative outlines already in place, in one click.
Report AID
The AI module. A BERT-based language model, fine-tuned on OPHID’s 2018–2023 reports, extracts key statements from district narratives and drafts summaries that flag trends, anomalies and districts needing follow-up.
Data Diagnostic Module
Runs data through four layers of rules (strict, critical, standard and minor checks) to catch missing values, outliers and logic violations before submission.
Data Analytics Platform
Dashboards comparing expected and actual performance, with maps, trends and early-warning flags, so problems surface during the reporting period rather than long after it.
Data Export Request Listener
Applies each stakeholder’s age-sex disaggregations and format rules automatically, producing validated, submission-ready files for platforms such as DATIM.
Note for evaluators: only RAID uses machine learning. The other four modules are rules-based automation. When you assess an “AI-supported” intervention, separate the outcomes driven by the AI component from those driven by conventional automation and integration.
4. How it was built
Co-designed with users
Nine data clerks, 15 district M&E officers and 10 programme leads shaped the tool through weekly agile sprints. Users asked for automated chart exports and new error categories, and both were built in.
Developers embedded in M&E
The developer sat inside OPHID’s Strategic Information and Evaluation directorate, so problems raised by programme staff could be fixed quickly, following DevOps principles.
Low-cost, offline-ready stack
Open-source tools (Python, Rust, DuckDB, Pandas, Polars) on a hybrid cloud and local set-up that works offline and syncs when connectivity returns.
Data protection followed Zimbabwe’s Cyber and Data Protection Act (2021): AES-256 encryption in transit and at rest, role-based access, secure APIs, audit trails, and data stored on servers inside Zimbabwe.
5. Phased rollout
May 2023
ORA piloted in Bulilima District
July 2023
After peer demonstrations at review meetings, ORA expands to six more districts
October 2023
All 15 districts use ORA, with a user interface and guides
January – May 2024
RAID developed with programme managers and deployed for Q2 reporting; DDM and DAP added in April
June 2024
DERL built to meet changing USAID/PEPFAR export formats
December 2024
Listed in the WHO Digital Health Atlas; MoHCC requests the platform for its own reporting
6. Results: before and after
Timely report submission
Before: 27%
100%
From 4 of 15 districts on time to all 15
Data cleaning per district
Before: 10.2 days
2.9 days
Average time for cleaning and error resolution
Quarterly report preparation
Before: 5–7 days
≤2 days
Narrative compilation time down by over 70%
Report format consistency
Before: 35%
100%
All districts now use one harmonised format
Critical errors in submitted datasets
After rollout
Eliminated
Missing values, indicator mismatches and disaggregation errors removed
Districts running data reviews
After rollout
15 of 15
Mini data reviews on flagged anomalies, Q3 2023 – Q4 2024
The shift went beyond speed. Submissions that once went back and forth with donors up to three times were accepted first time, and PEPFAR cited OPHID as a leader in report completeness and accuracy. Three districts redesigned outreach after dashboards showed weak results among adolescent girls and young women, and at least six began presenting dashboard insights at provincial reviews.
“We now focus on what the data means, not how to compile it.”
District M&E Officer
“For the first time, we feel like we own the data, not just collect it for someone else.”
District Programme Manager
“Now, by the 15th, we’re wrapping up, not still cleaning data.”
District M&E Officer
7. The evaluator’s lens: how strong is the evidence?
Strengths
- Core outcomes come from system records (DATIM logs, DHIS2 dashboards, OMDIP usage logs), not only from self-report
- Qualitative feedback was triangulated with supervision findings and routine performance data
- Reporting followed selected domains of the WHO mERA checklist, which makes the design transparent and comparable
- Clear operational definitions for timeliness, completeness and data quality
Limitations
- A descriptive before-and-after case study with no comparison group, so the authors themselves note that causal claims cannot be made
- Only 18 months of data, from HIV programmes in OPHID-supported districts
- Qualitative evidence came from routine reports and supervision notes, not independent interviews, so positive reporting bias is possible
- The study reports no accuracy measure for RAID’s AI-generated summaries, and the implementer also authored the study
8. Three lessons
Co-creation drives adoption. Weekly feedback from frontline users made the tools fit real workflows, and peer demonstrations spread them from district to district faster than any mandate could.
Build on national systems, not around them. Integrating with DHIS2 and existing tools removed double entry and avoided creating a parallel reporting system.
Visible value creates its own demand. Time savings people could see led the TB and laboratory divisions to ask for the modules unprompted. The platform is being licensed as open source and moving towards MoHCC ownership.
Questions for reflection
- Which of the reported gains would you attribute to the AI module, and which to rules-based automation and integration? How would you test that?
- If you were commissioned to evaluate OMDIP’s impact, what comparison group or design would you propose?
- How would you assess the quality of RAID’s AI-generated summaries before relying on them for national reporting?
- What would need to be in place for a ministry, rather than an implementing partner, to sustain a platform like this after donor funding ends?
Key takeaway
Automation paid off in Zimbabwe because it was modest, modular and built with the people who do the reporting. It freed M&E staff from compiling data so they could use it, and turned donor reporting into something districts now act on themselves.
Source: Dhodho E, Masiye K, Banda F, Bepe T, Nyathi N, Chinyanga TT. “Transforming routine health data use in LMICs through modular, AI-supported automation: insights from Zimbabwe”. Oxford Open Digital Health, 2026; 4: oqag003. Published 30 January 2026 under a CC BY 4.0 licence. Case study adapted by EvalCommunity Academy for learning purposes.
