Data Quality Pipeline for Routine Surveillance Data from HMIS

End-to-end R pipeline for cleaning, harmonizing, and assessing data quality of routine malaria indicators from DHIS2 at health-facility level.

Overview

This project builds a full R pipeline to evaluate the quality of routine malaria data reported through a national DHIS2-based HMIS.

Using several years of health-facility–level data (2016–2022), the workflow quantifies:

• Missing data patterns by indicator, month, and district
  
• Outliers and implausible values (SD-based and Hampel filters)
  
• Internal consistency between key malaria indicators
  
• Health-facility activity and reporting completeness over time

The outputs are designed to support national malaria program teams in monitoring data quality and prioritizing supervision and corrective actions.

This project highlights my expertise in Real World Evidence (RWE) analytics, focusing on routine health data from DHIS2. It demonstrates skills in data wrangling, quality assessment, and visualization—key for roles in pharmacoepidemiology, health analytics, or public health research

---

Data sources and unit of analysis

• Source: DHIS2 routine HMIS exports (malaria indicators and related service statistics)
  
• Unit of analysis: health facility × month × year
  
• Geographic level: national, aggregated into regions and districts (no facility identifiers displayed in the report)

Key malaria indicators used in this project include:

Indicator	Description	Age Groups	Example Use Case
allout	Outpatient visits (all causes)	All, <5y, >5y	Denominator for incidence rates
susp	Suspected malaria cases	All, <5y, >5y	Suspicion surveillance
test	Tested cases (RDT + microscopy)	All, <5y, >5y	Diagnostic coverage
conf	Confirmed malaria cases	All, <5y, >5y	Confirmed incidence
maltreat	Treated uncomplicated cases	All, <5y, >5y	Treatment access
confsev	Confirmed severe cases	All, <5y, >5y	Severity monitoring
maldth	Malaria-related deaths	All, <5y, >5y	Mortality tracking
IPTp/Stocks	Preventive interventions & commodities	Pregnant women	Activity flags

Challenges Addressed: Recurrent missing data, French month names, facility name inconsistencies across years.

---

Data management & indicator construction in R

The R pipeline performs the following steps:

1. Cleaning and recoding

   • Selects and renames geographic and time variables (region, district, health facility, month, year).
   • Recodes French month names to numeric months and splits the period variable into month and year (case_when and separate).
   • Standardizes age groups:
     • <5 years, 5–14 years, ≥15 years, and combined groups (under-5 vs over-5).

2. Construction of aggregate indicators Using dplyr::rowwise() and mutate(), the script creates:

   • Total outpatient visits: allout, allout_u5, allout_ov5
     
   • Admissions: alladm, alladm_u5, alladm_ov5
     
   • Suspected, tested, confirmed cases:
     • susp, test (RDT + microscopy), conf (RDT + microscopy)
       
   • Treatment and severity:
     • maltreat (treated uncomplicated cases),
       
     • pressev, confsev (presumed and confirmed severe cases)
       
   • Malaria deaths: maldth (overall and age-specific)

3. Zero to NA recoding for analytical indicators

   • For malaria indicators, zeros are recoded to NA to distinguish true zeros from non-reporting.

tdata = st

---

Data-quality indicators

1. Missing data by indicator, month, and district

Using group_by() and across():

• Counts the number of missing values per indicator at district × month × year level.
  
• Joins the number of reporting facilities to compute the percentage of missing values.

\[ \text{Missingness }(\%) = \frac{N_{\text{facilities with NA}}}{N_{\text{facilities reporting in district}}} \times 100 \]

The resulting dataset (*_missing_data) includes:

• Region, district, month, year
  
• Indicator type (visits, suspected, tested, confirmed, treated, admissions, deaths)
  
• Age category (under-5, over-5, all ages)
  
• Numerator (facilities with missing data) and denominator (total facilities)
  
• Proportion of missing data (%)

data = t

---

2. Outlier detection and imputation

The pipeline applies two complementary methods:

1. Mean ± 3 standard deviations
   • For each indicator, computes mean and SD across all facility-months.
     
   • Flags values outside [mean − 3×SD, mean + 3×SD] as potential outliers.
2. Hampel filter (median ± 3 MAD)
   • Computes the median and median absolute deviation (MAD).
     
   • Flags values outside [median − 3×MAD, median + 3×MAD] as robust outliers.

Outliers are:

• Flagged in a long-format dataset (facility × month × indicator).
  
• Optionally imputed using local averages from neighboring months for the same facility when justified, while retaining flags for transparency.

data = t

---

3. Internal consistency checks

Programmatic rules derived from malaria surveillance practice are encoded in R; for example:

• Tested cases ≥ confirmed cases
  
• Suspected cases ≥ tested cases
  
• Confirmed cases ≥ treated cases
  
• Malaria deaths ≤ total deaths
  
• Malaria admissions ≤ total admissions

For each facility-month, the script:

• Creates logical flags (“Normal” / “Not normal”) for each rule.
  
• Aggregates results by district and month to compute the proportion of facility-months with incoherent data.

data = t

---

4. Facility activity and reporting completeness

To distinguish inactive facilities from true zero values:

1. For each facility-month, counts the number of non-missing malaria indicators.
   
2. Defines activity status:
   • active if at least one key malaria indicator is reported,
     
   • inactive otherwise.
3. Builds a time series per facility using a unique facility identifier and calendar date.
4. Computes for each district and month:
   • total number of facilities,
   • number of active facilities,
   • reporting completeness (%) = active facilities / total facilities × 100.

This produces district-level completeness trajectories over several years.

data = t

---

Visualizations

Key outputs include:

• Heatmaps of missingness
  • Percentage of missing values per indicator × month, faceted by district.
• Scatter plots for coherence checks
  • Tested vs confirmed cases, colored by coherence status, with regression lines.
• Time series of reporting completeness
  • Monthly completeness by district and at national level.

Detailed interactive visualizations of this project are implemented in a dedicated Visualizations Chapter using Shiny (for dynamic R-based dashboards). These tools enable drill-downs by district, time-series filtering, and exportable insights for stakeholders. Access the Shiny app via [GitHub link] upon request.

---

Results and Insights

• Missingness: 15-25% for confirmed cases (higher in dry season); rural districts > urban.

• Outliers: 5-10% flagged (Hampel more conservative).

• Consistency: 92% “normal” records; issues tied to under-reported tests.

• Completeness: Improved from 65% (2016) to 85% (2024); district trajectories guide interventions.

RWE Insights: Adjustments for reporting bias refine incidence estimates (e.g., +20% under-reported cases), enhancing evidence for intervention efficacy.

Relevance for RWE, and healthcare analytics roles

This project demonstrates:

• Ability to work with large routine health datasets (HMIS) at facility level.
  
• Proficiency in R for data cleaning, reshaping, and quality assessment (dplyr, tidyr, lubridate, ggplot2).
  
• Experience translating programmatic rules from malaria surveillance into automated data-quality checks.
  
• Skills in building reusable analytic datasets and indicators to support national programs and donor reporting.

Category	Skills	R Tools / Examples
Data Wrangling	Cleaning, reshaping, indicator building	dplyr, tidyr, lubridate
Quality Assessment	Missingness, outliers (Hampel/SD), consistency rules	Custom functions, mad(), logical flags
Visualization	Heatmaps, time series, scatterplots	ggplot2, viridis
RWE-Specific	Bias adjustment in routine data, longitudinal trends	DHIS2 handling, facility-level aggregation

Ousmane Diallo, MPH-PhD – Biostatistician, Data Scientist & Epidemiologist based in Chicago, Illinois, USA. Specializing in SAS programming, CDISC standards, and real-world evidence for clinical research.