Co-creating data science solutions for maternal and child health decision-making in tribal primary health centres: an action research using the Three Co's Framework

This action research study demonstrates that applying the Three Co's Framework (Co-Define, Co-Design, Co-Refine) to collaboratively develop data science solutions with frontline health workers in tribal primary health centres in India effectively addresses the mismatch between top-down reporting systems and local decision-making needs, resulting in context-specific tools like an offline-capable dashboard and a measurable improvement in organizational data maturity.

The Gist

Imagine you are trying to navigate a dense, foggy forest to find a hidden village. You have a map, but it's a top-down map drawn by someone who has never walked the forest. It shows big roads and major cities, but it doesn't show the muddy paths, the seasonal floods, or the specific shortcuts the locals know. Worse, the map is designed to help a distant governor count how many people live in the region, not to help a local guide find a sick child who needs medicine right now.

This is exactly the situation in tribal primary health centers (PHCs) in India. They have mountains of data about mothers and children, but it's stuck in a "top-down" system designed for reporting to the government, not for helping doctors make decisions on the ground.

This paper tells the story of how a team of researchers and local health workers decided to stop trying to force the old map to work and instead co-drew a new one together.

Here is the story of how they did it, broken down into simple steps:

1. The Problem: The Wrong Map

In many low-resource areas, digital health tools are built in fancy offices far away. They assume everyone has fast internet, knows how to use complex software, and has time to fill out long forms.

• The Reality: In these tribal areas, internet is spotty (like a radio signal that cuts out during rain), doctors are overwhelmed, and the "official" data is just a list of numbers that doesn't tell them who is sick or where they are.
• The Analogy: It's like giving a fisherman a spreadsheet of fish counts from last year, when what he really needs is a sonar device to show him where the fish are today.

2. The Solution: The "Three Co's" Framework

Instead of the researchers saying, "Here is the tool you need," they used a method called the Three Co's Framework. Think of this as a three-step dance where the local health workers lead the steps, and the researchers follow.

• Step 1: Co-Define (Solving the Right Problem)

  • What happened: The researchers asked the local doctors, "What is the problem?"
  • The Twist: The doctors didn't say, "We need better charts." They said, "We can't see our own data! The system tells us how many babies were born in the whole district, but we need to know which specific village has a mother who missed her check-up."
  • The Lesson: The problem wasn't that the doctors didn't know how to use data; the problem was that the data was locked away in a format they couldn't use. They realized they needed a tool that showed individual names and locations, not just big averages.

• Step 2: Co-Design (Building the Tool Together)

  • What happened: They built a digital dashboard (a screen with charts and maps) together.
  • The Analogy: Imagine building a house. Usually, an architect draws the blueprints and the builder just follows orders. Here, the builder (the doctor) said, "I don't need a fancy front door; I need a kitchen window that looks out onto the garden."
  • The Result: They created five specific tools:
    1. A New List of Questions: Instead of asking "How many births?", they asked "Which mother is overdue for her visit?"
    2. A Better Filing System: A digital model that links a mother, her baby, the nurse, and the village together.
    3. A Quality Check: A "spell-checker" for data that catches errors (like a baby born with a negative weight) before they break the system.
    4. A Custom Map: Since official maps didn't match the real villages, they drew new boundaries based on the doctors' local knowledge. This revealed that anemia (lack of blood) was clustered in specific areas, allowing for targeted help.
    5. The Dashboard: A simple, offline-friendly screen that looks like a WhatsApp chat or a spreadsheet—tools the doctors already knew how to use.

• Step 3: Co-Refine (Polishing the Tool)

  • What happened: They tested the tool, found bugs, and fixed them. Even when some doctors transferred to new jobs, they kept giving feedback via WhatsApp.
  • The Analogy: It's like tuning a guitar. You don't just tune it once; you keep adjusting the strings as you play until the music sounds perfect.

3. The Results: A New Kind of Confidence

The study measured how "data-savvy" the health centers became.

• Before: They had the desire to use data (they wanted to help people), but they lacked the tools and skills to do it.
• After: Their ability to analyze data jumped significantly. They went from being passive recipients of reports to active users of their own information.
• The Catch: The study noted that while the software got better, the hardware (internet, electricity, security) didn't change. You can give a fisherman a perfect sonar, but if the boat has a hole in the bottom, he still can't sail. This is a limitation of the system, not the tool.

Why This Matters

This paper is a blueprint for how to fix technology in places where resources are scarce.

• The Old Way: "We built this app; you must use it." (Often fails).
• The New Way: "You tell us what you need; we help you build it." (Works).

The Big Takeaway:
Technology doesn't have to be complicated to be powerful. By listening to the people on the ground and letting them define the problem, the researchers turned a confusing pile of data into a compass that helps local doctors navigate the forest and find the people who need help the most. It proves that when you treat local health workers as partners rather than just users, you don't just get a better app; you get a better system of care.

Technical Summary

1. Problem Statement

The study addresses a critical gap in India's tribal primary health centres (PHCs): while substantial maternal and child health (MCH) data is generated through national digital systems (e.g., RCH Portal, HMIS), it remains underutilized for local decision-making.

• The Core Issue: Existing digital tools are designed for "upward reporting" (aggregate statistics for district/state supervision) rather than "local decision support."
• The Mismatch: Top-down tools fail in low-resource, tribal settings due to a "design-reality gap." They assume connectivity, digital literacy, and workflows that do not exist in geographically isolated, forested tribal areas.
• Specific Pain Points: Medical officers (MOs) cannot answer basic operational questions (e.g., "Which specific beneficiaries in which villages are overdue for services?") because the data is not disaggregated to the individual or village level, and the logic for "pending" cases is often rigid and clinically irrelevant.

2. Methodology

The study employed an Action Research design conducted over two cycles (August 2023 – August 2024) across three tribal PHCs in the Rampachodavaram Integrated Tribal Development Agency (ITDA), Andhra Pradesh.

Frameworks Used

• The Three Co's Framework: A structured co-creation approach adapted from UK welfare contexts:
  1. Co-Define: Collaboratively identifying and framing the problem.
  2. Co-Design: Collaboratively developing the solution.
  3. Co-Refine: Iterative testing and improvement through field use.
• Keller's Data Science Framework: Used to structure the analytical pipeline (Discovery, Preparation, Quality Assessment, Analysis).
• Data.org Data Maturity Assessment (DMA): Used to measure organizational data maturity (Purpose, Practice, People) pre- and post-intervention.

Participants & Setting

• Co-creators: 5 Medical Officers (MOs), 24 Auxiliary Nurse Midwives (ANMs), and 36 Accredited Social Health Activists (ASHAs).
• Context: Three PHCs with varying connectivity and high Scheduled Tribe (ST) populations (67–93%). Mobile connectivity is inconsistent, especially during monsoons.
• Researcher Positionality: The lead researcher is a member of the local tribal community, facilitating trust and candid feedback.

Technical Implementation

• Language & Tools: R programming language.
• Pipeline: Built using the targets framework for reproducible data pipelines.
• Validation: pointblank R package for automated data quality checks.
• Interface: R Shiny dashboard, designed to be offline-capable with daily synchronization via GitHub Actions.
• Spatial Analysis: Used Voronoi tessellation to construct Health Facility Catchment Area (HFCA) boundaries from scratch, as no official boundaries existed.

3. Key Contributions & Solutions

The study produced five interconnected data science solutions, moving beyond a simple dashboard to a holistic data ecosystem:

A. Co-Defined MCH Indicators (42 Indicators)

Instead of system-defined aggregates, indicators were grounded in clinical workflows:

• Maternal Health: ANC visits tracked against gestational age; Hemoglobin levels categorized by clinical severity (Severe <7, Moderate 7–10, Normal >10 g/dL).
• Child Health: Immunization status tracked against specific windows (e.g., BCG within 15 days, Pentavalent-1 at 6 weeks).
• Logic Shift: "Pending" status was redefined from rigid reporting cutoffs to clinically relevant intervals since the last visit.

B. Entity-Relationship Data Model

A custom data model linking five core entities to enable drill-down navigation:

1. INDIVIDUAL (Mother/Child, keyed by RCHID)
2. HEALTH-SERVICE (Type, date, clinical details)
3. HEALTH-PROVIDER (ANM, ASHA)
4. HEALTH-FACILITY (PHC, Sub-centre)
5. HEALTH-SERVICE-LOCATION (Village, distance)

C. Data Quality Framework

Two validation agents were built using pointblank:

• Maternal Agent: Checks RCHID format, date sequencing (LMP < Registration < Delivery), physiological ranges for Hemoglobin, and age limits.
• Child Agent: Checks birth weight (0.5–6.0 kg), vaccine sequencing, and schedule adherence.
• Note: Rules were iteratively refined by MOs (e.g., relaxing birth weight lower bounds for high-prevalence low-birth-weight communities).

D. Spatial Resources & Analysis

• HFCA Construction: Created 18 PHC-level and 96 Village Secretariat-level catchment boundaries using local knowledge and Voronoi tessellation.
• Findings: Exploratory Spatial Data Analysis (ESDA) revealed significant spatial clustering of ANC coverage and anaemia prevalence (Moran's I = 0.28, p<0.001), validating the need for geographic targeting.

E. Interactive R Shiny Dashboard

• Design Principles: "Names over numbers" (individual records), "Drill-down by default" (PHC → Sub-centre → Village → Individual), and "Familiar over novel" (mimicking WhatsApp/Spreadsheet patterns).
• Functionality: Organized along the lifecycle continuum (pregnancy to immunization) with offline capability.

4. Results

Qualitative Findings

• Problem Reframing: The Co-Define phase shifted the problem from "lack of data analysis skills" to "structural lack of data access." MOs realized they didn't need more data, but existing data restructured for clinical decision points.
• Engagement: Despite staff transfers, two MOs maintained engagement via WhatsApp, demonstrating internalized capacity.
• Design Principles: The shift from aggregate reporting to individual-level tracking was the most significant design change driven by co-creators.

Quantitative Findings (Data Maturity Assessment)

The DMA scores (0–10 scale) showed moderate improvement:

• Overall Score: Increased from 5.04 to 5.75.
• Largest Gains:
  • Analysis: +1.90 (from 4.78 to 6.68).
  • Talent: +0.98.
  • Leadership: +0.94.
• Stagnation/Decline: The Practice dimension remained flat or declined slightly (4.24 to 4.17), specifically in Security (-0.32) and Infrastructure (-0.13). This indicates that while the tools improved, structural barriers (internet, security protocols) remain unsolved by the software alone.

5. Significance and Implications

• Methodological Innovation: This is the first application of the Three Co's Framework in a Low- and Middle-Income Country (LMIC) tribal health setting. It proves that separating "Problem Definition" (Co-Define) from "Solution Design" (Co-Design) is critical in contexts where users are historically excluded from problem framing.
• Technical Feasibility: Demonstrates that complex data science pipelines (spatial analysis, entity modeling, automated validation) can be co-created and deployed in resource-constrained, offline-first environments.
• Policy Impact: Suggests that India's Ayushman Bharat Digital Mission should mandate co-creation processes for tools targeting underserved populations. The study argues that "Co-Define" must be a diagnostic step before any tool deployment.
• Scalability: The five solutions (indicators, model, quality framework, spatial boundaries, dashboard) are individually transferable to other tribal PHCs using the RCH Portal.
• Limitations: The study acknowledges that co-created software cannot fix infrastructure deficits (connectivity, security) which require system-level policy reform, as evidenced by the stagnant "Practice" scores in the DMA.

Conclusion: The study validates that data science for health in tribal settings must be a participatory practice. By combining the Three Co's Framework with reproducible data science methods, the researchers successfully bridged the gap between raw data availability and actionable local decision-making.