Data Readiness Desk — Project Motivation

Git Repo Link

https://github.com/d-padmanabhan/dais-2026-data-readiness-desk/tree/main

The Scale Problem

India has 1.4 billion people, 120+ major languages, and 28 states. Health data collection happens across thousands of facilities, districts, and administrative layers — each with its own reporting cadence, incentive structure, and failure mode. At this scale, data quality is not a one-time problem to solve. It is a continuous, structural challenge.

Why Data Quality Matters Here Specifically

The tools we are building — for resource planning, facility matching, and coverage targeting — are only as good as the data underneath them. A planner who allocates immunisation resources to a district based on unreliable HMIS figures is not making a data-driven decision. They are making a guess with extra steps.

The core issue is that there is currently no signal at the point of use. A planner opens a dashboard, sees a number, and has no way to know whether that number reflects reality or reflects a reporting artefact.

Two Sources, Two Failure Modes

India's primary health data comes from two parallel systems:

HMIS (Health Management Information System)
Self-reported by facilities and states. Subject to completeness gaps (facilities that don't report), over-reporting incentives (targets tied to coverage metrics), and inconsistent aggregation across administrative levels.

NFHS-5 (National Family Health Survey, 2019–21)
Community-based survey, methodologically independent of the facility reporting chain. But it comes with its own constraints: a five-year collection cycle, district-level granularity ceiling, and no sub-district breakdown.

Neither source is authoritative on its own. The gap between what HMIS reports and what NFHS-5 observes is itself diagnostic — it tells you something about the reliability of the facility data.

Our Hypothesis

If the two sources agree, you can trust the data more. If they diverge significantly, something is wrong — either the facility is under-reporting, over-reporting, or the survey didn't capture the population that the facility serves.

Corroboration is not proof of accuracy, but it is the best signal available without sending a field team to every district.

What We Set Out to Build

A tool that answers, for any facility or district: "Can I trust this data?"

The answer takes the form of a trust verdict (green / amber / red), a numeric score, a one-line reason, and a ranked list of fixes — so the person reading it knows not just whether to trust the data, but what to do about it if they can't.

The target user is not a data scientist. It is an on-the-ground process expert — a district health officer, a programme manager, a field coordinator — who needs to make a decision today and needs to know whether the data they are looking at is a reliable basis for that decision.

Why This Is Hard

  • The two sources measure slightly different things, on different timelines, using different denominators. Direct comparison requires understanding the methodology differences and identifying genuinely comparable metrics.
  • India's administrative geography is complex. State names are not standardised across sources. District boundaries changed between NFHS-4 and NFHS-5. Facilities exist in states that have no corresponding NFHS-5 district record.
  • Absence of data is itself a data quality signal, but it is ambiguous — it could mean the facility didn't report, or that the survey didn't cover that area, or that the indicator genuinely doesn't apply.

Our approach encodes these ambiguities explicitly: a suppression cap when the corroborating source is missing, partial confidence tags for AI-extracted attributes, and a weakest-link band rule so a strong completeness score cannot mask a corroboration failure.

Built With

Share this project:

Updates

Submission history