This dataset is released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 licence (CC BY-NC-SA 4.0): research, academic, personal and other non-commercial use is permitted with attribution; adaptations must be shared under the same licence; commercial use is not permitted. The original public source data remains under the Government Open Data License - India (GODL-India). Commercial use: contact surya@judicialmind.ai or hello@judicialmind.ai.

IndiaDiseaseWatch

Licence: CC BY-NC-SA 4.0 — research and non-commercial use only; derivatives must be shared under the same licence. Commercial use: contact surya@judicialmind.ai or hello@judicialmind.ai

Access is approved automatically when the email you enter matches your Hugging Face account email and belongs to the organisation you name; other requests are reviewed manually.

District × month counts of 28 disease, morbidity and context items reported by public and private health facilities to India's Health Management Information System (HMIS), April 2009 to March 2022, for 730 districts coded to the Local Government Directory (LGD). Every month with enough history carries an outbreak label (a Farrington-style seasonal exceedance), and two benchmark tasks come with it: forecasting the reported count 1 and 3 months ahead, and early warning of outbreak months. The counts are what facilities report, so they follow testing effort, reporting completeness and the share of private facilities reporting as well as disease incidence.

1,208,817 district-months over 730 districts; 20,385 district × disease series
480,767 labelled district-months, of which 31,736 are outbreaks (6.6%)
Rows by source: HMIS state reports (data.gov.in) 508,510; HMIS item-wise monthly reports 256,720; HMIS sub-district monthly file 2020-22 443,587

Quickstart

import pandas as pd

# after `huggingface-cli download judicialmind/IndiaDiseaseWatch --repo-type dataset --local-dir IndiaDiseaseWatch`
pan = pd.read_parquet("IndiaDiseaseWatch/data/panel.parquet")    # or "hf://datasets/judicialmind/IndiaDiseaseWatch/data/panel.parquet"
dengue = pan[(pan.disease == "dengue_positive") & (pan.outbreak == 1)]
print(len(dengue), "dengue outbreak district-months")
print(dengue.sort_values("baseline_z", ascending=False)
      [["district_name", "state_name", "ts", "value", "baseline_mean", "threshold"]].head(3).round(1).to_string(index=False))

Output: 1,402 dengue outbreak district-months.

Strongest dengue outbreaks (by baseline z)

District	State	Month	Reported	Baseline mean	Threshold
Sambalpur	Odisha	2021-12	3,323	0.0	2.0
Firozabad	Uttar Pradesh	2021-09	3,208	1.0	4.7
Nizamabad	Telangana	2021-10	1,234	0.5	2.9

With the datasets library: load_dataset("judicialmind/IndiaDiseaseWatch", "panel", split="train").

Evaluate a model

evaluate.py scores count forecasts and early-warning scores on the benchmark keys (bench/eval_keys.parquet) and recomputes the baseline rows of leaderboard.json from the shipped files:

cd IndiaDiseaseWatch
pip install -r requirements.txt     # numpy, pandas, pyarrow, scikit-learn
python evaluate.py --reproduce

It checks each baseline row and ends with "reproduced 32 of 32 leaderboard rows (test, test2) from the shipped files".

Your model writes one row per benchmark key: district_lgd, disease, target_month, horizon plus pred (forecast count) and/or score (outbreak score, higher = more likely). A prediction for target_month may use data up to the origin month target_month - horizon only. Example, a naive forecast and the origin month's baseline z-score (these reproduce the naive_last and farrington_last rows):

import pandas as pd

pan = pd.read_parquet("IndiaDiseaseWatch/data/panel.parquet",
                      columns=["district_lgd", "disease", "ts", "value", "baseline_z"])
keys = pd.read_parquet("IndiaDiseaseWatch/bench/eval_keys.parquet")
keys = keys[keys.split.isin(["test", "test2"])]
# origin = target month - horizon: a forecast may only use data up to the origin
keys["ts"] = (pd.PeriodIndex(keys.target_month, freq="M") - keys.horizon.to_numpy()).to_timestamp().date
x = keys.merge(pan, on=["district_lgd", "disease", "ts"], how="left")
x["pred"] = x["value"].fillna(0)                          # naive: the origin month's count (0 if not reported)
x["score"] = x["baseline_z"].clip(-10, 50).fillna(-1e9)   # early warning: origin month's z above its baseline
cols = ["district_lgd", "disease", "target_month", "horizon"]
x[cols + ["pred"]].to_parquet("my_forecasts.parquet")
x[cols + ["score"]].to_parquet("my_scores.parquet")

python IndiaDiseaseWatch/evaluate.py --root IndiaDiseaseWatch --forecasts my_forecasts.parquet --scores my_scores.parquet --model origin_value

Example output (two of the rows):

task	horizon	split	model	n	MAE	WMAPE	MASE (mean)	MASE (median)	AUROC	AUPRC	precision@20
forecast	1	test	origin_value	160,084	17.306	0.511	1.280	0.270	–	–	–
early_warning	1	test	origin_value	116,725	–	–	–	–	0.729	0.174	0.383

Files

file	content
`data/panel.parquet`	one row per district × month × disease (1,208,817 rows)
`data/diseases.parquet`	disease definitions: HMIS item codes in the 2008 and 2017 formats, group, the cross-format check
`bench/eval_keys.parquet`	the benchmark's (district, disease, target month, horizon, split) keys (959,374 rows)
`bench/baseline_predictions_test.parquet`	baseline forecasts on the test years (639,376 rows)
`leaderboard.json`, `leaderboard.csv`	all baseline results (validation and test)
`sources.parquet`	one row per source resource: provider, title, URL, attribution
`evaluate.py`, `requirements.txt`	scoring code and its pinned dependencies

Schema

data/panel.parquet:

column	type	meaning
`state_lgd`, `district_lgd`	int	LGD codes
`state_name`, `district_name`	string	names
`ts`	date	month (first day)
`disease`	string	item key (see Diseases)
`value`	float	reported count
`era`	string	HMIS item format: `old` (2008 format, to March 2017) or `new` (2017 format)
`origin_src`	string	HMIS source of the row: `HMIS sub-district monthly file 2020-22` (2020-22 sub-district file summed to districts), `HMIS item-wise monthly reports` (monthly item-wise district reports), `HMIS state reports (data.gov.in)` (state district-month HMIS sheets, data.gov.in)
`baseline_mean`	float	seasonal baseline mean μ (null when fewer than 6 baseline points)
`threshold`	float	outbreak threshold μ + 2·`sd_eff`
`baseline_z`	float	(value − μ) / `sd_eff`
`n_base`	int	number of baseline points
`outbreak`	int	1 = outbreak month, 0 = not, null = not labelled
`opd_jump`	bool	the district's outpatient volume is above 2× or below ½ of its seasonal baseline (month left unlabelled)
`source_uuid`	string	source resource id (see `sources.parquet`)

bench/eval_keys.parquet: district_lgd, state_lgd, disease, target_month, horizon (1 or 3), split (val, test, test2). bench/baseline_predictions_test.parquet: the test and test2 keys with y (reported count) and pred_naive_last, pred_seasonal_naive, pred_seasonal_mean, pred_lightgbm. leaderboard.json / .csv: task, horizon, split, model, n, mae, wmape, mase_mean, mase_median, series, positives, base_rate, auroc, auprc, precision_at_20. sources.parquet: uuid, source (data.gov.in, NDAP (NITI Aayog) or HMIS state reports (data.gov.in)), provider, title, url, attribution.

Diseases

key	item	group	months	districts	district-months	labelled	outbreaks	2008↔2017 link
`malaria_tests`	Blood smears examined for malaria (testing effort)	denominator	2009-04-01 – 2022-03-01	730	69,992	0	0	ok
`opd_all`	Outpatient attendance (allopathic + AYUSH)	denominator	2009-04-01 – 2022-03-01	730	70,063	0	0	ok
`diarrhoea_children`	Childhood diarrhoea cases (under 5)	enteric	2009-04-01 – 2022-03-01	730	70,179	34,095	3,436	ok
`diarrhoea_children_ip`	Childhood diarrhoea treated as inpatients	enteric	2017-04-01 – 2022-03-01	730	36,554	16,159	1,369	new_only
`ip_diarrhoea`	Inpatients with diarrhoea with dehydration	enteric	2017-04-01 – 2022-03-01	730	36,618	16,183	1,054	new_only
`ip_hepatitis`	Inpatients with hepatitis	enteric	2017-04-01 – 2022-03-01	730	36,344	16,013	1,226	new_only
`ip_typhoid`	Inpatients with typhoid	enteric	2017-04-01 – 2022-03-01	730	36,489	16,131	1,440	new_only
`ip_puo`	Inpatients with pyrexia of unknown origin	other	2017-04-01 – 2022-03-01	730	36,585	16,165	1,436	new_only
`snake_bite`	Emergency attendances for snake bite	other	2017-04-01 – 2022-03-01	730	36,227	15,987	1,522	new_only
`asthma_children`	Childhood asthma cases	respiratory	2017-04-01 – 2022-03-01	729	36,458	16,094	1,305	new_only
`ip_respiratory`	Inpatients with asthma, COPD or respiratory infections	respiratory	2017-04-01 – 2022-03-01	730	36,581	16,177	1,803	new_only
`ip_tb`	Inpatients with tuberculosis	respiratory	2017-04-01 – 2022-03-01	730	36,471	16,126	1,308	new_only
`pneumonia_children`	Childhood pneumonia cases	respiratory	2017-04-01 – 2022-03-01	730	36,539	16,159	1,886	new_only
`resp_children_admitted`	Children admitted with respiratory infections	respiratory	2009-04-01 – 2022-03-01	730	69,817	33,877	3,833	ok
`tb_children`	Childhood tuberculosis cases	respiratory	2017-04-01 – 2022-03-01	729	36,441	16,028	833	new_only
`diphtheria_children`	Childhood diphtheria cases	vaccine_preventable	2017-04-01 – 2022-03-01	729	36,075	15,766	201	fail
`measles_children`	Childhood measles cases	vaccine_preventable	2009-04-01 – 2022-03-01	729	69,693	33,674	1,475	ok
`pertussis_children`	Childhood pertussis cases	vaccine_preventable	2017-04-01 – 2022-03-01	729	36,052	15,754	68	fail
`tetanus_neonatorum`	Tetanus neonatorum cases	vaccine_preventable	2017-04-01 – 2022-03-01	729	36,047	15,742	49	fail
`dengue_positive`	Dengue positive (RDT + ELISA)	vector_borne	2017-04-01 – 2022-03-01	710	32,471	13,604	1,402	new_only
`ip_dengue`	Inpatients with dengue	vector_borne	2017-04-01 – 2022-03-01	730	36,226	15,915	1,358	new_only
`ip_malaria`	Inpatients with malaria	vector_borne	2017-04-01 – 2022-03-01	730	36,250	15,946	832	new_only
`je_positive`	Japanese encephalitis tests positive	vector_borne	2017-04-01 – 2022-03-01	707	31,321	12,781	57	new_only
`kala_azar_positive`	Kala-azar positive cases	vector_borne	2017-04-01 – 2022-03-01	727	35,376	15,328	32	new_only
`malaria_children`	Childhood malaria cases (under 5)	vector_borne	2017-04-01 – 2022-03-01	730	36,518	16,091	802	fail
`malaria_pf`	Malaria Plasmodium falciparum positive, microscopy	vector_borne	2017-04-01 – 2022-03-01	729	36,091	15,780	350	fail
`malaria_positive`	Malaria slides positive, microscopy (Pv + Pf)	vector_borne	2009-04-01 – 2022-03-01	729	69,214	33,411	1,859	ok
`malaria_rdt_positive`	Malaria RDT positive (Pv + Pf)	vector_borne	2017-04-01 – 2022-03-01	729	36,125	15,781	800	new_only

new_only items exist only in the 2017 HMIS format. ok items join the two formats; fail items keep only their 2017-format months. The link check compares each district's FY2016-17 and FY2017-18 totals; it accepts a median new/old ratio between 0.45 and 2.2 with at least 60% of districts within 3× of that median. The band is wide because incidence really moves year to year (reported malaria fell steeply in 2017). The malaria_tests (testing effort) and opd_all (outpatient volume) rows are context series. They carry no outbreak labels.

How the panel is assembled

Three HMIS sources.
- HMIS item-wise monthly reports: the monthly item-wise district reports, FY2009-10 to FY2020-21.
- HMIS state reports (data.gov.in): district-month HMIS sheets for 14 states, 2009-2020, from Ministry of Health and Family Welfare downloads.
- HMIS sub-district monthly file 2020-22: the 2020-22 sub-district file, summed to districts.
Where more than one source reports the same district, month and disease, the 2020-22 file wins, then the item-wise reports, then the state sheets. Agreement where sources overlap: item-wise reports vs 2020-22 file: 12,022 pairs, 72% within 1%; state reports vs 2020-22 file: 34,073 pairs, 67% within 1%; state reports vs item-wise reports: 147,643 pairs, 100% within 1%.
Item matching. Items are matched on their code and their text. Some published files store codes as floats, so 10.10 (childhood malaria) appears as 10.1 (childhood pneumonia); the text tells them apart. A disease that sums several items (e.g. Pv + Pf positives) is kept for a month only when every component was reported.
Outbreak label (Farrington-style seasonal exceedance). For month t the baseline is every reported value in the same calendar month ±1 in the previous 5 years, with at least 6 points. The baseline never crosses the April 2017 format change.
- mean μ, sd_eff = max(sd, √μ, 1), threshold = μ + 2·sd_eff
- outbreak = count > threshold and count ≥ 5
- A month is left unlabelled when the district's total outpatient volume is more than 2× or less than 1/2 of its own seasonal baseline. Such jumps are reporting artefacts (missing or duplicated facility reports, the April 2020 lockdown), not epidemics.
Because a new-format disease needs two years of baseline, labels for the 2017-format items start in FY2019-20.

Benchmark

Time-held-out splits by target month: train to 2019-03, validation 2019-04 to 2020-03 (pre-COVID), test 2020-04 to 2021-03, test2 2021-04 to 2022-03. Horizons 1, 3 months. A series is evaluated if it has at least 18 reported months before April 2020 and at least 6 in the test years (13,815 series). Forecasts and scores may use only data up to the origin month (target − horizon).

Forecasting: predict the reported count. Metrics: MAE, wMAPE, and per-series MASE.
Early warning: score each labelled target month from data up to the origin; the label is outbreak above. Metrics: AUROC, AUPRC, and precision@20 per month.

CPU baselines (a month that was not reported counts as 0 in their inputs):

naive_last, seasonal_naive (same month last year), seasonal_mean (mean of the same month over the last 3 years)
lightgbm: one global model on log counts with lags, rolling means, seasonal lags, calendar month, disease and state
ears_c2: monthly EARS-C2, the z of the origin month against the 7 months ending 2 months earlier
farrington_last: the origin month's own exceedance z
climatology: past outbreak frequency in the target's calendar month
lightgbm_forecast_exceedance: the LightGBM forecast set against the target month's baseline
lightgbm_classifier: the same features, trained on the labels

chronos2_zeroshot: Amazon Chronos-2 (zero-shot, 96-month context, quantile forecasts), scored on the same keys where it returned a forecast. Its early-warning scores are the median forecast against the baseline, and the share of forecast quantiles above the outbreak threshold. chronos2_finetuned (chronos2_ft_* scores): Chronos-2 fine-tuned on the training districts' histories up to March 2019; scored like the zero-shot model.

Results

Headline: fine-tuned Chronos-2 (chronos2_finetuned) is the best forecaster 3 months ahead, with the lowest MASE on both test years (1.369 on test, 1.684 on test2), and its outbreak scores give the highest precision@20 on every test split and horizon (0.725 / 0.792 at 1 month, 0.533 / 0.625 at 3 months).

Forecast, horizon 1 month(s) (MASE: per-series mean absolute error over the mean absolute seasonal difference on training months; lower is better)

split	model	n	MAE	wMAPE	MASE mean	MASE median
test	chronos2_zeroshot	158,539	17.02	0.499	1.076	0.268
test	chronos2_finetuned	158,539	16.15	0.474	1.077	0.247
test	lightgbm	160,084	15.59	0.461	1.225	0.305
test	naive_last	160,084	17.31	0.511	1.280	0.270
test	seasonal_mean	160,084	40.37	1.193	1.605	0.595
test	seasonal_naive	160,084	53.13	1.570	2.495	0.795
test2	lightgbm	159,604	18.94	0.429	1.534	0.322
test2	chronos2_finetuned	156,961	18.76	0.420	1.547	0.267
test2	chronos2_zeroshot	156,961	19.65	0.440	1.678	0.287
test2	seasonal_mean	159,604	29.64	0.672	1.766	0.493
test2	naive_last	159,604	22.28	0.505	2.099	0.322
test2	seasonal_naive	159,604	27.19	0.616	2.397	0.424

Forecast, horizon 3 month(s) (MASE: per-series mean absolute error over the mean absolute seasonal difference on training months; lower is better)

split	model	n	MAE	wMAPE	MASE mean	MASE median
test	chronos2_finetuned	157,948	24.31	0.713	1.369	0.346
test	chronos2_zeroshot	157,948	26.17	0.767	1.394	0.384
test	lightgbm	160,084	23.13	0.683	1.457	0.434
test	seasonal_mean	160,084	40.37	1.193	1.605	0.595
test	naive_last	160,084	25.52	0.754	1.690	0.392
test	seasonal_naive	160,084	53.13	1.570	2.495	0.795
test2	chronos2_finetuned	156,214	22.11	0.497	1.684	0.298
test2	lightgbm	159,604	22.43	0.508	1.696	0.387
test2	seasonal_mean	159,604	29.64	0.672	1.766	0.493
test2	chronos2_zeroshot	156,214	24.25	0.546	1.940	0.345
test2	seasonal_naive	159,604	27.19	0.616	2.397	0.424
test2	naive_last	159,604	30.87	0.700	2.766	0.411

Outbreak early warning, 1 month(s) ahead (precision@20: share of true outbreaks among the 20 highest-scored district-diseases per month)

split	score	n	outbreaks	base rate	AUROC	AUPRC	precision@20
test	lightgbm_classifier	116,725	3,335	0.029	0.815	0.241	0.700
test	chronos2_quantile_prob	115,992	3,309	0.029	0.772	0.221	0.721
test	chronos2_ft_quantile_prob	115,992	3,309	0.029	0.750	0.223	0.725
test	lightgbm_forecast_exceedance	116,725	3,335	0.029	0.742	0.227	0.662
test	chronos2_forecast_exceedance	115,992	3,309	0.029	0.736	0.221	0.633
test	farrington_last	116,725	3,335	0.029	0.729	0.174	0.383
test	chronos2_ft_forecast_exceedance	115,992	3,309	0.029	0.728	0.231	0.675
test	ears_c2	116,725	3,335	0.029	0.624	0.063	0.183
test	climatology	116,725	3,335	0.029	0.521	0.031	0.021
test2	lightgbm_classifier	142,348	6,231	0.044	0.840	0.324	0.679
test2	chronos2_quantile_prob	140,680	6,150	0.044	0.788	0.306	0.762
test2	chronos2_ft_quantile_prob	140,680	6,150	0.044	0.781	0.318	0.792
test2	lightgbm_forecast_exceedance	142,348	6,231	0.044	0.772	0.335	0.762
test2	farrington_last	142,348	6,231	0.044	0.768	0.278	0.508
test2	chronos2_forecast_exceedance	140,680	6,150	0.044	0.757	0.317	0.742
test2	chronos2_ft_forecast_exceedance	140,680	6,150	0.044	0.756	0.330	0.767
test2	ears_c2	142,348	6,231	0.044	0.668	0.113	0.254
test2	climatology	142,348	6,231	0.044	0.539	0.050	0.071

Outbreak early warning, 3 month(s) ahead (precision@20: share of true outbreaks among the 20 highest-scored district-diseases per month)

split	score	n	outbreaks	base rate	AUROC	AUPRC	precision@20
test	lightgbm_classifier	116,725	3,335	0.029	0.747	0.112	0.425
test	chronos2_quantile_prob	115,764	3,309	0.029	0.718	0.108	0.487
test	chronos2_ft_quantile_prob	115,764	3,309	0.029	0.707	0.112	0.512
test	lightgbm_forecast_exceedance	116,725	3,335	0.029	0.669	0.110	0.388
test	chronos2_forecast_exceedance	115,764	3,309	0.029	0.650	0.111	0.471
test	farrington_last	116,725	3,335	0.029	0.641	0.082	0.250
test	chronos2_ft_forecast_exceedance	115,764	3,309	0.029	0.637	0.111	0.533
test	ears_c2	116,725	3,335	0.029	0.595	0.050	0.104
test	climatology	116,725	3,335	0.029	0.521	0.031	0.021
test2	lightgbm_classifier	142,348	6,231	0.044	0.749	0.148	0.371
test2	chronos2_quantile_prob	140,432	6,103	0.043	0.696	0.153	0.592
test2	chronos2_ft_quantile_prob	140,432	6,103	0.043	0.693	0.165	0.625
test2	lightgbm_forecast_exceedance	142,348	6,231	0.044	0.675	0.165	0.521
test2	farrington_last	142,348	6,231	0.044	0.643	0.125	0.321
test2	chronos2_forecast_exceedance	140,432	6,103	0.043	0.627	0.151	0.554
test2	chronos2_ft_forecast_exceedance	140,432	6,103	0.043	0.623	0.159	0.583
test2	ears_c2	142,348	6,231	0.044	0.571	0.068	0.208
test2	climatology	142,348	6,231	0.044	0.539	0.050	0.071

The test years are the COVID-19 years: outpatient volumes fell sharply in April-May 2020 and many diseases were reported less. Validation (FY2019-20, pre-COVID) results for every model are in leaderboard.json.

Coverage

730 districts, April 2009 to March 2022 (156 months), 26 diseases with outbreak labels plus 2 context series. Districts with any data, by financial-year start: 2009: 25, 2010: 108, 2011: 229, 2012: 399, 2013: 555, 2014: 592, 2015: 592, 2016: 618, 2017: 582, 2018: 579, 2019: 595, 2020: 727, 2021: 703. From April 2020 the sub-district file covers nearly every district; origin_src marks the source of each row. Districts are mapped to current LGD districts by name; a district split after 2009 maps to the successor that keeps the parent's name. Counts are district aggregates that HMIS already publishes at sub-district level; there is no person-level data. Rare diseases (tetanus, diphtheria, kala-azar) have small monthly counts.

Source data

Derived from 19,267 resources published on the Open Government Data Platform India (data.gov.in), snapshot May 2026, under the Government Open Data License - India. Per-resource attribution (provider, title, resource URL) is in the attribution column of sources.parquet. GODL-India does not cover personal information; the release holds facility-reported aggregate counts only.

Providers: Ministry of Health and Family Welfare (HMIS).

HMIS state reports (hmis-state-* source ids in sources.parquet) were compiled from Ministry of Health and Family Welfare HMIS downloads published on data.gov.in.

Original data: published by Ministry of Health and Family Welfare (Health Management Information System, HMIS) on the Open Government Data Platform India (data.gov.in) and, where marked, NITI Aayog's National Data and Analytics Platform (NDAP), under the Government Open Data License - India (GODL-India). The original public data remains under GODL-India; the CC BY-NC-SA 4.0 licence of this repository covers JudicialMind's compilation, derived fields, annotations, benchmarks and models.

Citation

@misc{judicialmind2026indiadiseasewatch,
  title={IndiaDiseaseWatch: District-Month Disease Counts, Outbreak Labels and Early-Warning Benchmark from India's HMIS},
  author={JudicialMind},
  year={2026},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/datasets/judicialmind/IndiaDiseaseWatch}}
}

License

CC BY-NC-SA 4.0 (see LICENSE): research, academic, personal and other non-commercial use, with attribution to JudicialMind and to the original data publishers; adaptations must be shared under the same licence. Commercial use requires a separate licence.

Commercial use: contact surya@judicialmind.ai or hello@judicialmind.ai.

Downloads last month: 30

Total file size:

21.6 MB

Collection including judicialmind/IndiaDiseaseWatch

India Open Government Data — Bharat Health Benchmarks

Collection

Research and non-commercial use (CC BY-NC-SA 4.0). Commercial: surya@judicialmind.ai / hello@judicialmind.ai • 4 items • Updated 3 days ago