Ranking Madhya Pradesh's Districts on Maternal and Child Health: What a Neural Network Confirms

Data source

The study uses secondary data from the Annual Health Survey (AHS) Factsheet 2012-13 and the AHS Second Updation Bulletin 2012-13 for Madhya Pradesh, published by the Office of the Registrar General & Census Commissioner, India — publicly available government data. The AHS covers all 284 districts across nine "Empowered Action Group" (EAG) states (Bihar, Jharkhand, Madhya Pradesh, Chhattisgarh, Odisha, Rajasthan, Uttar Pradesh, Uttarakhand, and Assam), which together constitute roughly 50% of India's population but a disproportionate 60% of births, 71% of infant deaths, and 72% of maternal deaths nationally. For Madhya Pradesh specifically, the AHS sampled 2,557 units (villages in rural areas, blocks in urban areas), covering about 2.284 million people across an average of 456,800 households.

Indicator selection: 7 sectors, 32 indicators

Following the framework proposed by Ram and Chander Sekhara (2006) and Swain et al. (2010), the study grouped indicators into seven broad sectors of human wellbeing: fertility reduction; health status of women and children; educational status of women; hygiene and sanitation; electrification; maternal and childcare status; and family welfare. Thirty-two specific indicators were drawn from these sectors, including Birth Rate, Death Rate, Maternal Mortality Rate/Ratio, Neo-Natal/Post-Neo-Natal/Infant/Under-Five Mortality Rates, Sex Ratio at Birth and Sex Ratio (0-4), Effective (Female) Literacy, household access to electricity/safe drinking water/toilets/LPG/pucca housing/computer or telephone connectivity, and a range of maternal and child healthcare-utilization measures (antenatal checkups, tetanus injections, IFA supplementation, immunization coverage by vaccine type).

Principal Component Analysis

PCA is a multivariate statistical technique that reduces a set of correlated variables into a smaller number of uncorrelated ("orthogonal") components, each a weighted linear combination of the original variables, ordered so the first component captures the largest share of total variation. Applied to the 32 standardized indicators across 45 districts, six components had eigenvalues at or above 1 (the standard retention threshold) and together explained 81.7% of total variance:

Each district's scores on these six components (d1-d6) were confirmed normally distributed via a Kolmogorov-Smirnov test, then averaged into a single mean score per district and rescaled into a 0-100 index using min-max normalization: I_i = [(d(i) - Min(i)) / (Max(i) - Min(i))] × 100.

Classification thresholds

Backward: less than (mean − 0.6745σ) → index range [-5.43 to -1.71]

Underdeveloped: (mean − 0.6745σ) to mean → index range [-1.71 to 0.0]

Developing: mean to (mean + 0.6745σ) → index range [0.0 to 1.71]

Developed: more than (mean + 0.6745σ) → index range [more than 1.71]

Full district ranking (all 45 districts)

Neural network architecture and validation

The confirmatory model was a three-layer feedforward Multilayer Perceptron (MLP): 32 input units (one per standardized indicator), one hidden layer of 10 units using a hyperbolic tangent activation function, and an output layer of 4 units (one per development category) using softmax activation with a cross-entropy error function. The network's classification accuracy against the PCA-based categories:

Limitations worth flagging

This is a single-state (Madhya Pradesh), single-time-point (AHS 2012-13 data, published 2019) study using secondary government survey data rather than primary data collection, and the 45-district boundaries reflect Madhya Pradesh's administrative geography at that time (the state has since seen some district reorganization). The neural network's "100% accuracy" is validation against the PCA classification's own categories, not against an independent ground truth of actual district development — meaning the two methods agreeing strongly suggests the grouping is statistically robust, but does not independently prove either method captures "true" development correctly. The test-set cell sizes are also very small (as few as one district in a category), which limits how much confidence the accuracy figures alone can support.