PARI is a free and anonymous screening tool that takes about five minutes to complete. It returns a nutrition risk score for each student, a clinical BMI-for-age z-score, and an assessment of which of four barriers is driving that student's risk. The last of these is what no existing tool provides.
Submitted to Union Education Minister Pralhad Joshi · November 2026
The national surveys are methodologically sound. They are also of very little use to the person who actually has to act on them.
The Comprehensive National Nutrition Survey tells us that 24% of Indian adolescents are thin and 27% are stunted, and that children in the poorest households carry roughly 3.2× the odds of stunting. That is a national average.
If you run a school, that figure gives you nothing to work with. It does not tell you which of your students are struggling, and it certainly does not tell you what is causing it.
Four quite different problems can produce the same underweight child. The family may not be able to afford protein. Protein may not be sold anywhere nearby. Nobody may have taught the child what a balanced meal looks like. Or the child may simply not be eating enough meals in a day. Each of these calls for a different response, and buying more dal achieves nothing if the nearest shop selling it is six kilometres away.
No existing tool tells a school which of the four it is dealing with, and that is the gap PARI was built to fill.
What students actually eat, measured across ten food groups.
Whether food runs short, how many meals are eaten, and whether cost is what prevents buying protein.
Whether students know what protein does, and whether they consider balance when they eat.
Whether protein-rich food is sold nearby at all, and whether the school provides a meal.
A student answers 23 questions on a phone, and those answers produce a risk score. Their height and weight produce a clinical z-score through an entirely separate calculation. The two results never share an input, and that independence is what makes the comparison between them worth anything.
Most online health questionnaires assign their points arbitrarily. Every weight in PARI is drawn from published research: the wealth gradient comes from CNNS, the dietary diversity threshold from a multi-country adolescent study, and the protein requirements from ICMR-NIN 2020.
Height, weight, age and sex are run through the WHO 2007 LMS reference tables to produce a BMI-for-age z-score. This is the same statistic a paediatrician would use, rather than a BMI category invented for the application.
Because no measurement of the body feeds into the risk score, the correlation between the two is a genuine external check. Had the tool been built the other way round, any agreement between them would be circular and would prove nothing.
Every item is scored toward risk, so a higher number always means more exposure and the four domains add directly.
Diet carries 35. A diversity count scores how many of the ten food groups a student eats at least twice a week, at 20, 14, 7 or 0 as that count passes two, four and six. A protein term scores the highest frequency among the four groups that carry protein, meaning pulses, dairy, eggs and meat or fish, at 15, 10, 5 or 0. It takes the maximum rather than the sum, because one source eaten reliably is what protects a child.
Economic access carries 30 across food insecurity, meal regularity and cost as a stated barrier. Awareness carries 20 across knowing what protein does, recognising a balanced meal and knowing one’s own requirement. Physical access carries 15 across availability and school meal provision. The four add, cap at 100, and split into bands at 33 and 67.
Thinness rises monotonically across the three bands, and mean BMI-for-age falls as the band rises.
A student in the high band is 4.1 times as likely to be clinically thin as one in the low band, and the trend across the three bands tests at z = 7.64, p = 2 × 10−14. None of that is built in, because the questionnaire never sees height or weight.
Across the full range the score tracks BMI-for-age at ρ = −0.42 and separates clinical thinness with an AUC of 0.72. At the high-band threshold of 67 it catches two thirds of thin students, a sensitivity of 66.7 per cent, while clearing three fifths of the rest, a specificity of 61.8 per cent.
Taken alone, each of the four separates thin students from the rest: economic access at an AUC of 0.708, diet at 0.684, physical access at 0.670 and awareness at 0.635, every one significant beyond p = 10−19. Economic and physical access each hold independently when the other three are controlled for, at −0.19 and −0.11 standard deviations of BMI-for-age per standard deviation of domain score.
Adult BMI thresholds do not apply to a growing body. WHO’s median BMI is 15.27 for a six-year-old girl and 19.00 for a fourteen-year-old boy, so a single cut-off applied across a school is wrong at both ends of it.
The WHO 2007 reference publishes a median, a coefficient of variation and a skew parameter for every age and sex. The skew parameter is there because childhood BMI is not symmetric about its median, and a mean-and-deviation comparison would misread the tails, which is exactly where screening happens. It enters as a power transform applied before the comparison is made.
Those parameters are L = −1.1389, M = 15.2718 and S = 0.10267 for that six-year-old, read from the reference PARI ships. The full table, fourteen ages across two sexes, sits inside this page, so the z-score is computed in the browser and no measurement leaves the device. WHO’s cut-offs then sit on top of the continuous number: severe thinness below −3, thinness to −2, normal to 1 and overweight to 2.
Every response is written against a fixed schema carrying a timestamp and an SGIC, a self-generated code built from facts a student can reproduce months later but which identify nobody. No name, no roll number, no contact detail. It is what lets a school rescreen the same cohort and match rounds record by record, so the change between rounds is measured rather than inferred. The row holds the eighteen items, the measurements, and the derived score, band, BMI and z-score. It never holds a person.
Filtering the cohort recalculates everything below it, including the summary statistics, the distributions and the screening performance.
The response set behind these charts has been anonymised and minimised. It carries no identification code, no school, no class and no height or weight, only the answers themselves, the age band, the sex and the derived scores, held in shuffled order.
This is not a simplified version. It runs PARI's actual scoring code against the real WHO 2007 growth reference, entirely within your browser. Nothing is transmitted and nothing is stored.
This is general nutrition guidance rather than clinical advice. A screening instrument identifies who is worth examining more closely. It does not replace a clinician, and PARI has never claimed otherwise.
The survey never asks how much a student weighs. When its results line up with who is actually underweight, that agreement tells us something.
The test is a simple one. Among the students PARI rated low risk, 7 in 100 turned out to be underweight. Among those rated moderate, the figure was 16 in 100. Among those rated high, it was 30 in 100, which is more than four times the low-risk group. The rate climbs at every step, and it does so on the basis of information the survey was never given.
This makes the tool particularly good at clearing students rather than at flagging them. When PARI rates a student as low risk it is correct roughly nine times out of ten. A school with 600 students and a single afternoon to work with can use that result to set most of them aside and concentrate its attention on the remainder.
It performs less well in the opposite direction. Most of the students it flags as high risk are not currently underweight, which looks at first like a straightforward error but is not one. The survey measures whether adequate food is within reach, and being underweight is one of the later consequences of it not being. Identifying students before the shortfall registers in their body is precisely what a screening tool exists to do.
One finding deserves stating plainly, since it is not a convenient one. Of the four things PARI measures, three independently predict a student's actual weight: what they eat, what they can afford, and what they can physically obtain. The fourth, awareness, does not. Knowing what protein does clearly shapes how a student eats, but in this data it does not register on the scale on its own. This is the main issue the next version of the instrument will need to address, and it is recorded here rather than left out.
PARI is not a one-off questionnaire. It runs on a longitudinal database architecture: every response is written against a fixed schema carrying a timestamp and a self-generated identification code, so a school can screen the same cohort again and match the two rounds record by record.
The code is built from facts a student can reproduce months later but which identify nobody, which is what makes repeat measurement possible without the instrument ever storing a person. Running PARI a second time is a re-run of the same instrument against the same schema, not a rebuild, and the change between rounds is what tells a school whether an intervention worked.
Every round writes the same 32 fields, so rounds are directly comparable without transformation.
Self-generated codes link a student's rounds to each other and to nothing else.
Each record carries its own collection time, so rounds separate cleanly.
An anonymised CSV with a mirrored sheet, so a school keeps its own cohort history.
Phase 1 uses transparent, hand-assigned weights, each traceable to a published source. Phase 2 re-derives them empirically once the response set is large enough to support factor analysis.
The first score counts how many of the ten food groups are eaten at least three days a week and is worth 20 points. The second takes the highest weekly frequency among the four protein-dense groups and is worth 15.
Every option carries a published point value. There are no free parameters, no hidden multipliers and nothing fitted after the fact to make the numbers look better.
The ceilings are 35, 30, 20 and 15 points, so no single domain can push a student into the high-risk band on its own.
The bands sit at 33 and 67. The breakdown is returned alongside the total, since the breakdown is the part a school can actually act on.
The instrument, the scoring engine, the growth-reference tables, the anonymised storage layer and the codebook are all public. Read it, fork it, or tell me where it is wrong.
def _dietary(resp):
ranks = {g: _FREQ_RANK.get(resp.get(g, "Never"), 0)
for g, _ in FOOD_GROUPS}
# Diversity: groups eaten >=3 days/week (rank >= 2)
diversity = sum(1 for g, _ in FOOD_GROUPS if ranks[g] >= 2)
if diversity >= 6: div_pts = 0
elif diversity >= 4: div_pts = 7
elif diversity >= 2: div_pts = 14
else: div_pts = 20
# Protein: best weekly frequency among protein-dense groups
pmax = max(ranks[g] for g in PROTEIN_GROUPS)
prot_pts = {3: 0, 2: 5, 1: 10, 0: 15}[pmax]
return div_pts + prot_pts, {...}
This is the exact function the demo above runs. It reproduces all 1,007 stored scores without a single disagreement between them.
def baz(weight_kg, height_cm, age_years, sex):
"""Return (baz, bmi, category)."""
b = bmi(weight_kg, height_cm)
row = _nearest_row(sex, age_years)
if b is None or row is None:
return None, None, None
_, L, M, S = row
if L == 0:
z = math.log(b / M) / S
else:
z = (((b / M) ** L) - 1) / (L * S)
return round(z, 2), round(b, 1), _category(z)
"""Anonymized response storage.
NO personal identifiers are ever written — only the
SGIC code and survey data.
"""
FIELDS = [
"timestamp", "sgic", "school", "grade", "age", "sex",
"height_cm", "weight_kg", "measured_flag",
# ... the 18 survey items ...
"risk_score", "band", "baz", "bmi", "baz_category",
]
No name, roll number, date of birth or contact detail is collected at any point in the pipeline.
A self-generated code allows the same cohort to be screened again later without the instrument ever learning who anyone is.
Deployment requires the head teacher's written agreement before a single link is shared.
Families are informed before deployment and may withdraw their child without giving a reason.
Results are returned to schools only in aggregate. No individual student's score is ever disclosed to staff, because a screening score in the wrong hands ceases to be a health tool and becomes a label.
PARI is the measurement layer for work that was already running without one.
A free lunch programme at KDO Jain School in Hubli feeds more than 300 students every day. Nutrition screenings with the Welfare Society for Destitute Children in Mumbai covered 150+ children and surfaced 60+ malnutrition cases.
Both programmes serve food and hope for the best, which is the normal condition of charitable feeding. There is no baseline, no targeting and no way of telling whether the meal being served is the meal that was needed. A programme can run for years without ever discovering that its students' real constraint was availability rather than cost.
A screening layer turns the operation into a loop. The cohort is measured, the dominant barrier is identified, the food and the teaching are directed at that specific gap, and the cohort is screened again to establish whether the figure has moved. That is the difference between charity and a programme that can be improved.
Five minutes per student, on any phone, in their own language.
Not simply how many students are at risk, but which barrier dominates at this particular school.
A cost problem calls for subsidised protein, an availability problem for a supply route, and an awareness problem for teaching.
The same codes and the same instrument, at a later date. The change between the two is the evidence.
Founder of SNACC and the developer of PARI.
He came to nutrition through sport. A decade of competitive taekwondo, ending with representing India at the ISF World Games, taught him that what an athlete eats determines what an athlete can do, and that in India the food required to train seriously was either imported, expensive, or full of things that undermined the point. He began formulating an alternative in his kitchen, then took it through food safety licensing and into retail. SNACC now sells across twenty-two outlets in Mumbai.
Working on the product raised a question the product could not answer. A protein bar serves people who can afford a protein bar. It does nothing for the far larger group who cannot, and nothing at all for a school that has no way of knowing which of its students are struggling in the first place.
PARI is his attempt at that second problem. It screens young people across four dimensions, diet quality, affordability, awareness and physical access, and returns both a risk score and, more usefully, the specific barrier driving it. A school with six hundred students and one afternoon can find out not only how many are at risk but which constraint it is actually dealing with. It is free, open-source, available in four languages, and has screened over a thousand students.
What the data showed was not what he expected. Cost and physical availability dominate: nearly two-thirds of students screened said price stops them buying protein, and more than two-thirds said it is either not sold near them or hard to get. Awareness gaps are widespread too, but once cost and access are accounted for they stop explaining which students are underweight. That finding runs against the assumption the instrument was partly built on, and it is published on this site rather than left out.
His intention is to take PARI from a screening tool to nutritional infrastructure, a free layer any school, clinic or programme in India can run to identify risk early and act on it, building a picture of youth nutrition that does not currently exist.
The position underneath all of it is straightforward. Nutritious food is not a privilege earned by income or geography. It is a right, and the distance between that principle and the situation on the ground is a problem worth building against.