Extracts India's national antimicrobial resistance surveillance results from the annual reports of both national networks — ICMR's Antimicrobial Resistance Surveillance Network (AMRSN) and NCDC's National Antimicrobial Resistance Surveillance Network (NARS-Net) — and publishes them as structured, fully provenance-annotated data, carried as parallel series.
Author: Yashita Thakur (ORCID 0009-0004-7895-5250)
Code licence: MIT · Data licence: see DATA_LICENSE.md
Derived from the publicly available annual reports of two independent Indian national AMR surveillance networks: ICMR-AMRSN (2017-2024) and NCDC NARS-Net (2017-2024). The two are published here as parallel series and are never pooled. Independent, unofficial analysis - not endorsed by or affiliated with ICMR or NCDC.
India runs two national AMR surveillance networks. They publish separately, on different metrics, from different hospitals. This repository extracts both.
| ICMR-AMRSN | NCDC NARS-Net | |
|---|---|---|
| Publisher | Indian Council of Medical Research | National Centre for Disease Control |
| Metric published | % susceptible | % resistant |
| Feeds WHO GLASS | no | yes |
| Participating sites | central institutes, private and corporate hospitals, a mission hospital, armed forces | state government medical colleges and state/UT institutes |
| Sites in the data | de-identified as RC1–RC21, with data published against the codes |
named in full in an annexure, with no data published against them |
| Organisms extracted here | E. coli, K. pneumoniae, A. baumannii, P. aeruginosa, S. aureus, MRSA | E. coli, S. aureus |
| Years covered | 2017–2024 | 2017–2024 |
| Editions read | 3 — the 6th, 7th and 8th, published 2022–2024 | 8 — every edition published |
| Rows extracted | 1,286 national, plus 1,365 by Regional Centre | 558 — 345 E. coli, 213 S. aureus |
| Files | amr_trends.csv, amr_rc_trends.csv |
narsnet_trends.csv |
The two publish at different depths in opposite directions, and the shape of each dataset follows from that. Every AMRSN edition carries its own 8-year retrospective trend table, so three editions cover the same eight years three times over — which is what makes cross-edition revision detection possible. Every NARS-Net edition reports its own year alone, so eight editions cover eight years once each, and there is nothing to detect.
They share no comparison value. AMRSN publishes % susceptible; NARS-Net publishes % resistant. Converting between them needs the intermediate fraction, and AMRSN publishes no % intermediate for E. coli or S. aureus — so an AMRSN % resistant cannot be computed at all. Going the other way is no better: NARS-Net classifies intermediate isolates separately, under a three-way S/I/R split, so a NARS-Net %R and a %S do not sum to 100.
This is enforced structurally rather than by convention. NarsNetRecord carries
resistant_pct and has no field meaning the same thing as
Record.susceptible_pct, so the two networks cannot be addressed as one series
by accident. The exports keep the same distance — narsnet_trends.csv for
NARS-Net, amr_trends.csv for AMRSN — and nothing joins them on a value. The
two artefacts that do set the networks side by side join on
keys and on
counts respectively, never on
a percentage.
This is a constraint on the comparison rather than a shortcoming of either body.
- Quick start — build the full dataset
- Scope and maintenance
- What this is, and is not
- Outputs — files, figures, row schema, flags
- The ICMR-AMRSN national series — coverage, how tables are located, how values are checked, cross-report revisions
- Regional Centre breakdowns — AMRSN, by centre
- The NCDC NARS-Net series — all eight editions, and what each one lets you check
- Where the two networks meet — coverage map, surveillance volume, structural caveats
- Source data — both registries, and citation form
- Supporting research — the edition-by-edition readings behind the NARS-Net material
- The landing page
- Project history
- Citing
pip install -r requirements.txtThe full dataset is four builders, in this order.
python -m src.build_dataset --fetchDownloads the three ICMR report PDFs to data/raw/, extracts the national trend
tables, validates against known reference values, writes
data/processed/amr_trends.{csv,json}, and prints any places where editions
report the same year differently.
python -m src.build_rc_datasetExtracts the Regional Centre breakdowns from those same three PDFs and writes
amr_rc_trends.{csv,json}, rc_panel.json and rc_revisions.json.
python -m src.build_narsnet_dataset --fetchDownloads the eight NCDC report PDFs and writes narsnet_trends.{csv,json},
narsnet_panel.json and narsnet_revisions.json.
python -m src.build_comparabilityWrites comparability_matrix.json. This one reads no PDFs — it is a
second-order artefact derived from the two datasets built above, so both must
exist first, and it must be rebuilt after either of them changes.
Then the tests, the figures, and the landing page's bibliography:
pytest -vpython viz/trend_charts.py --revisions && python -m src.references --injecttrend_charts.py reads all four builders' outputs and writes all fifteen
figures in docs/figures/, so it runs last.
To re-check the downloaded PDFs against their pinned hashes at any point:
python -m src.fetch --network all --verify-onlyData is current through the ICMR AMRSN 8th edition (2024), and through all eight NCDC NARS-Net editions, 2017–2024, which are extracted in full — every edition NARS-Net has published for the two organisms both networks report at species level.
Built and completed independently before the author began medical school, and not maintained against editions published since. For anything published after 2024, check the ICMR AMRSN site or the NCDC reports page directly.
ICMR and NCDC publish their surveillance results as annual report PDFs. There is no API, no CSV, and no bulk download; isolate-level data is not part of either public release. This repository turns the published national trend tables into a clean, citable dataset, where every single number carries the report edition and table number it came from, so any value can be traced back to a specific printed table and checked by hand.
Running antimicrobial-resistance surveillance across a national network of this many laboratories, sustained over several years and published in a consistent annual form, is a hard operational and data-management problem; the extraction notes in this repository sit inside that context and are not a critique of it. NCDC's NARS-Net published its 2020 report from inside the pandemic, and said plainly why that year's isolate counts were down.
- Not community prevalence. These are isolates from tertiary-care laboratories — a heavily selected, hospital-skewed population in both networks. They do not describe resistance in the general population.
- Not patient-level data. Neither network releases isolate-level records.
- Not a national burden or incidence estimate. Denominators are "isolates tested for this drug", not people, not infections.
- Not a pooled cross-network figure. The two are carried in separate files with separate schemas and are never combined — see The two series are never pooled.
- Not an official ICMR or NCDC product. Where this dataset and the corresponding published report differ, the published report is authoritative and the difference is a limitation of this extraction — please open an issue.
Out of scope throughout: specimen-type splits beyond those the sources print, OPD/ward/ICU splits, non-priority pathogens, resistance-gene data.
| File | Dataset | Contents |
|---|---|---|
data/processed/amr_trends.{csv,json} |
AMRSN national | One row per organism × antibiotic × year × report edition |
data/processed/revisions.json |
AMRSN national | Where editions report the same year differently |
data/processed/extraction_report.json |
AMRSN national | Run metadata: sources, hashes, what parsed |
data/processed/amr_rc_trends.{csv,json} |
AMRSN by Regional Centre | One row per organism × Regional Centre × antibiotic × edition |
data/processed/rc_panel.json |
AMRSN by Regional Centre | The RC set each edition printed, and what changed between editions |
data/processed/rc_revisions.json |
AMRSN by Regional Centre | Cross-edition RC revision check — near-empty by design |
data/processed/rc_extraction_report.json |
AMRSN by Regional Centre | RC run metadata |
data/processed/narsnet_trends.{csv,json} |
NARS-Net | One row per organism × antibiotic × specimen × edition, all eight editions 2017–2024 |
data/processed/narsnet_panel.json |
NARS-Net | The drug panel and specimen columns each edition prints, and what changed between them |
data/processed/narsnet_revisions.json |
NARS-Net | Cross-edition revision check — structurally empty; see its own note |
data/processed/narsnet_extraction_report.json |
NARS-Net | Run metadata, per-edition checks, and which cells no check reaches |
data/processed/comparability_matrix.json |
cross-network | Which network reports each organism × antibiotic × year, on which metric, from which specimen basis. Carries no percentage and no count from either |
docs/data/trends.json |
AMRSN national | 9 KB chart-data extract read by the landing page |
The narsnet_* files deliberately drop the amr_ prefix the AMRSN exports
carry. The two datasets are not concatenable and share no comparison column.
python viz/trend_charts.py --revisions writes all fifteen.
| Figure | Series | Shows |
|---|---|---|
trend_escherichia_coli.png |
AMRSN national | % susceptible by year, 10-drug Enterobacterales panel |
trend_klebsiella_pneumoniae.png |
AMRSN national | % susceptible by year, same panel |
trend_acinetobacter_baumannii.png |
AMRSN national | % susceptible by year, 9-drug panel |
trend_pseudomonas_aeruginosa.png |
AMRSN national | % susceptible by year, 11-drug panel |
trend_staphylococcus_aureus.png |
AMRSN national | % susceptible by year, 11-drug Gram-positive panel |
trend_mrsa.png |
AMRSN national | % susceptible by year, 9-drug panel |
rc_escherichia_coli.png |
AMRSN by Regional Centre | Spread across Regional Centres, per edition |
rc_klebsiella_pneumoniae.png |
AMRSN by Regional Centre | Spread across Regional Centres, per edition |
rc_staphylococcus_aureus.png |
AMRSN by Regional Centre | Spread across Regional Centres, per edition |
narsnet_escherichia_coli.png |
NARS-Net | % resistant by year, blood and urine |
narsnet_staphylococcus_aureus.png |
NARS-Net | % resistant by year, blood only |
narsnet_comparability_escherichia_coli.png |
cross-network | 21 drugs × 8 years, which network reports each cell |
narsnet_comparability_staphylococcus_aureus.png |
cross-network | 13 drugs × 8 years, which network reports each cell |
narsnet_surveillance_volume.png |
cross-network | Isolates tested per year, both networks — the one shared axis |
cross_report_revisions.png |
AMRSN national | The 17 genuine cross-edition revisions |
The AMRSN national trends:
{
"organism": "Klebsiella pneumoniae",
"antibiotic": "meropenem",
"year": 2024,
"susceptible_n": 4276,
"tested_n": 12189,
"susceptible_pct": 35.1,
"source_report_year": 2024,
"source_table": "Table 2.7",
"source_url": "https://www.icmr.gov.in/...",
"extracted_date": "2026-08-26",
"reported_pct": 35.1,
"computed_pct": 35.08,
"flags": ""
}year is the calendar year the measurement describes. source_report_year is
the edition it was read from. Keeping these separate is the point — it is
what lets the same year appear three times, once per edition, and differ.
susceptible_pct is the percentage ICMR printed, and nothing else. Where
the source does not print one — it shows (-) for cells like *0/8, since a
proportion from eight isolates would not be a stable estimate — this field is
null, even though the counts are published. Deriving 0.0% from 0/8 and
presenting it as a susceptibility figure would introduce a number the source
itself does not report. computed_pct carries susceptible_n / tested_n for
anyone who wants it, clearly labelled as derived rather than reported.
reported_pct and computed_pct should agree to within rounding. Any row
where they do not is flagged, not dropped.
The NARS-Net rows use a different schema, described under The metric: % resistant.
| Flag | Meaning |
|---|---|
low_isolate_count_asterisk |
the source marks this cell with * (very few isolates) |
pct_suppressed_in_source |
the source shows (-) in place of a percentage for this cell |
no_isolates_tested |
Denominator is 0 — the drug was not tested at all that year |
pct_mismatch(...) |
printed % and n/N do not fully reconcile (>0.15 pp) — check against the source before use |
label_footnote_asterisk |
The drug's row label carries a * footnote in the source |
colistin_is_intermediate_susceptibility |
Not a susceptibility figure — see below |
antibiotic_assigned_positionally |
Label column failed to extract; drug identity inferred from row order. Audit before use. |
In the current dataset (1,286 rows) no row carries
antibiotic_assigned_positionally — every antibiotic label was read from the
table rather than inferred. Exactly 3 rows carry pct_mismatch, and in all
three the printed figures themselves do not fully reconcile — this is not an
extraction failure (see Reconciling printed values).
Both non-fermenter tables footnote colistin: "*Colistin represents percentage intermediate susceptibility". It is an intermediate figure, not a susceptibility one. At face value it makes colistin appear to be the one drug still working against A. baumannii at ~97% while meropenem sits at 9%. Every such row is flagged, and colistin is excluded from the trend charts for this reason.
| Organisms | E. coli, K. pneumoniae, A. baumannii, P. aeruginosa, S. aureus, MRSA |
| Chapters | Enterobacterales, non-fermenting Gram-negative bacilli, staphylococci |
| Specimens | AMRSN's "all samples" trend tables, with each chapter's own printed exclusions — the E. coli and K. pneumoniae captions exclude faeces and urine, the staphylococcal ones state no exclusion. The separate urine-only tables are never read. |
| Years | 2017–2024 |
| Report editions | 2022 (6th), 2023 (7th), 2024 (8th) |
| Rows | 1,286 |
Each edition carries its own 8-year retrospective trend table, so the same calendar year is covered by up to three independent editions — which is what makes revision detection possible.
Panels are per organism, not per chapter, and are read from each table rather than assumed:
| Organism | Drugs | Notes |
|---|---|---|
| E. coli, K. pneumoniae | 10 | |
| A. baumannii | 9 | minocycline; no gentamicin/tobramycin/ciprofloxacin |
| P. aeruginosa | 11 | gentamicin, tobramycin, ciprofloxacin; no minocycline |
| S. aureus | 11 | Gram-positive panel |
| MRSA | 9 | omits cotrimoxazole and linezolid |
Neither non-fermenter is tested against ertapenem or cefazolin. Daptomycin appears in the specimen-wise staphylococcal tables but in neither yearly trend table, so it is absent here.
Breakdowns by Regional Centre are a separate dataset, covering the three organisms that have an RC-wise susceptibility table — see Regional Centre breakdowns.
Table numbers are not stable across editions, so nothing is hardcoded:
| Organism | 2022 ed. | 2023 ed. | 2024 ed. |
|---|---|---|---|
| E. coli | Table 3.6 | Table 3.6 | Table 2.6 |
| K. pneumoniae | Table 3.7 | Table 3.7 | Table 2.7 |
| P. aeruginosa | Table 5.3 | Table 4.6 | Table 3.3 |
| A. baumannii | Table 5.6 | Table 4.3 | Table 3.6 |
| S. aureus | Table 6.4 | Table 7.4 | Table 6.4 |
| MRSA | Table 6.9 | Table 7.9 | Table 6.9 |
Chapters move between editions (the 2024 edition pulled Enterobacterales from
Chapter 3 to Chapter 2, pushing the non-fermenters from 4 to 3). The parser
therefore searches every page for a caption matching the table's meaning and
reads the table number back out of whatever it finds, so source_table always
cites the number that edition actually printed.
The caption wording is not uniform either. Three grammars occur, and a parser written for the first alone finds nothing at all in the other chapters:
| Caption, as printed | What differs | Where |
|---|---|---|
Yearly susceptibility trend of X isolated from ... |
— | Enterobacterales; non-fermenters in 2023/2024 |
Yearly susceptible trend of X isolated from ... |
susceptible | Non-fermenters in the 2022 edition |
Year-wise susceptibility trends of X from ... |
Year-wise, trends | Staphylococci, all editions |
Three further wrinkles this handles:
- The 2022 edition prints "Klebsiella pneumonia" (without the trailing e), so organism patterns tolerate the shorter spelling.
- Each edition also has urine-only, blood-only and pus/exudate trend tables with near-identical captions. Those are explicitly rejected.
- Table 1.12b is a yearly isolation trend (how many isolates were found), not a susceptibility trend. Requiring "susceptibility"/"susceptible" in the caption keeps it out.
Running pdftotext -layout on these reports produces text that looks
tabular but is not: antibiotic labels are emitted as one contiguous block
before the data, and later year columns are vertically offset from their rows.
Any regex-over-text parser silently attributes values to the wrong antibiotic —
which is the worst possible failure mode, because the output looks plausible.
All extraction therefore goes through pdfplumber's ruling-line table
detection. Three strategies are tried (lines/lines, lines/text, text/lines) and
the best-formed result wins. If all three fail, the parser raises rather
than degrading to text matching.
Five specific failure modes were found in these PDFs and are defended against, because each one produces plausible-looking wrong numbers rather than an error:
-
Column indices are not trustworthy. In the 2023 edition the header row carries an extra leading cell, so
Year-2017sits at column index 4 while its own data sits at index 3. Cells are therefore matched to years by x-coordinate overlap, never by column index. -
Cell text is clipped. In the 2022 edition the ruled box around the final year column is narrower than the digits printed inside it, so pdfplumber's per-cell text silently drops the overhanging characters:
5170 / 14729comes back as170 / 1472, and9980 / 14304as980 / 1430. The percentage survives, so the cell still looks fine while the counts are wrong by an order of magnitude. Cell contents are therefore assembled from whole words assigned to a row and column by their centre point, so no digit can be clipped. -
Merged cells create phantom rows. The wrapped "Piperacillin- / tazobactam" label makes pdfplumber report a nested extra row sitting inside the real one and repeating its percentage. Left in, it is absorbed by the next antibiotic and overwrites that drug's percentage — which is how cefazolin briefly acquired piperacillin-tazobactam's 56.8%. Row bands wholly contained within another band are discarded.
-
Numbers wrap mid-digits. The MRSA table's narrow columns break a number across a line inside its own cell:
4286/4311 (99.4)is emitted as4286/431, then a lone1. Joined on whitespace that reads as4286/431— a susceptibility of 994%. A line ending on a digit followed by a line starting with one is spliced rather than spaced. -
Year headers split across rows. Table 6.9 puts
Year-2020on one header row and the other seven years as bare digits on the next. Taking whichever single row has the most years drops 2020 entirely, and the column boundaries then hand 2020's data to its neighbours. Year positions are accumulated across the whole header band instead.
A sixth trap was a bug in this repo rather than the PDFs, and is recorded
because it was invisible: a fast-path filter skipped any page not containing
the literal "usceptibility trend", which silently excluded every page of the
2022 non-fermenter chapter, where captions read "susceptible trend". The
tables were matched by the regex but never reached it. Pre-filters must stay in
step with the pattern they guard.
Strategy selection scores candidates on whether the grid is internally
consistent — whether each cell's printed percentage reconciles with its own
numerator and denominator — rather than on how many cells it contains. A naive
cell count actively prefers the shredded lines/text grid, and a mis-cut grid
that has lost digits fails arithmetic while still looking well-formed.
Three independent layers:
- Internal consistency. Every cell's printed percentage is recomputed from its own numerator and denominator. A gap over 0.15 pp means the parser paired the wrong numbers, and is flagged.
- Narrative cross-check. Each chapter states key values in prose,
independently of the table ("meropenem susceptibility decreased from 73.2%
to 62.9%"). Those prose values are encoded as fixtures in
src/validate.pyand asserted against table-extracted values. Prose and table are separate renderings of the same underlying figure, so agreement is real corroboration. - Structural assertions. Eight years present, ten antibiotics present, labels read rather than inferred, no susceptibility above 100%, no division by a zero denominator.
All 41 national fixtures pass, plus 10 hand-verified Regional Centre cells and
86 hand-verified NARS-Net cells. pytest runs 418 tests; the ones needing the
PDFs skip cleanly on a fresh clone until python -m src.fetch has been run.
For MRSA there is also a definitional check available nowhere else: MRSA is defined by methicillin/cefoxitin resistance, so cefoxitin susceptibility in that table must be ~0%. It is. Any other value would mean the wrong table or the wrong row had been read.
Correction to the build spec. The spec listed the E. coli / meropenem / 2024 fixture as
62.9% (7594/12061), with the numerator flagged uncertain. Table 2.6 of the 2024 edition prints 7587/12061, which is 62.90%. (7594/12061 would round to 63.0%.) The verified value is used here.
Take any row, e.g. E. coli / meropenem / 2024:
- Read
source_urlandsource_report_year→ the 2024 edition PDF. - Read
source_table→ Table 2.6. - Open that PDF, find the caption "Table 2.6: Yearly susceptibility trend of E. coli isolated from all samples (except faeces and urine)".
- Find the meropenem row and the Year-2024 column.
- The cell reads
7587 / 12061 (62.9)— matchingsusceptible_n,tested_n, andsusceptible_pct.
data/processed/extraction_report.json records the SHA-256 of every PDF used,
so you can confirm you are reading byte-identical source material.
Checking each cell's printed percentage against its own numerator and
denominator turned up three rows where the printed percentage and the printed
counts do not fully reconcile. They are carried exactly as printed and flagged
pct_mismatch — nothing is corrected and nothing is dropped — because any
adjustment would be this project's inference rather than the source's figure,
and the point of this repository is that every number can be traced to a
printed table.
1. A denominator that reads differently across editions.
P. aeruginosa / piperacillin-tazobactam / 2022 is printed as
9017/113156 (68.5) in the 2022 and 2023 editions; 9017/113156 would be
7.97%, not the stated 68.5%. The 2024 edition prints 9017/13156, which is
68.54% and matches the stated percentage. The later edition's denominator has
one fewer digit and reconciles with the printed percentage.
2. A percentage that does not fully reconcile with its counts.
A. baumannii / minocycline / 2022 is printed as 6207/10542 by all three
editions, which is 58.88%. The 2022 edition prints the percentage as 58.5;
the 2023 and 2024 editions print 58.8.
Neither is visible from a single edition's table read on its own. Three rows out of several thousand printed values spanning three editions is an ordinary rate for data work at this scale.
viz/trend_charts.py --revisions renders these, plotting whichever quantity
actually moved. Charts elsewhere in this repo use the most recent edition
reporting each point.
The same calendar year is not always reported with the same number in
successive editions — figures are revised and isolate sets de-duplicated as the
surveillance data matures between publications. Because every row records which
edition it came from, these differences are surfaced in revisions.json
rather than being averaged away or replaced by the newest value.
Crucially, two things that look alike are not conflated:
count_revision— the numerator or denominator itself changed.percentage_revision— counts agree but the printed percentage moved by more than rounding can explain.- Printing precision alone — the 2023 edition prints
14.94%where the 2024 edition prints14.9%for an identical1021/6833— is not a revision and is excluded. Six of the seven raw differences in the national series are of exactly this kind; reporting them as revisions would overstate how much the underlying figures actually move.
Across 432 (organism, antibiotic, year) combinations covered by two or more editions, there are 17 genuine revisions — 16 count revisions and 1 percentage revision.
The clearest single case, and the one that motivated comparing counts rather than percentages:
| 2022 ed. | 2023 ed. | 2024 ed. | |
|---|---|---|---|
| E. coli, piperacillin-tazobactam, 2022 | 5170/14729 | 5170/14729 | 5170/14728 |
One isolate was removed from the denominator on de-duplication. The printed percentage stays 35.1% throughout, so a detector comparing only percentages would report nothing at all.
There is also a recurring pattern worth knowing before quoting any "100%
effective" claim: for the anti-staphylococcal agents (tigecycline,
vancomycin, teicoplanin, in both S. aureus and MRSA), earlier editions
often print a numerator equal to the denominator — a flat 100% — and a later
edition reports a lower numerator. For example S. aureus / tigecycline / 2022
is 2452/2452 (100%) in the 2022 edition but 2314/2452 (94.4%) in the 2023
and 2024 editions.
The ICMR reports also publish, alongside the national trend tables, RC-wise
tables: susceptibility broken down by Regional Centre (RC) — one column per
antibiotic, one row per RC, for the report's own year. These are extracted into
a separate dataset (data/processed/amr_rc_trends.{csv,json}, 1,365
rows), with the same provenance fields as the national series plus a
regional_centre column. Build it with python -m src.build_rc_dataset.
Tables are located by caption meaning, never table number, exactly as the national logic is — the two grammars in use ("… Percentage RC wise of X …" in 2022/2023, "RC-wise … percentages of X …" in 2024) both carry the tokens "RC wise" and "(AMS)", which is what tells them apart from the national trend captions and from the "Regional centre wise distribution" isolate-count tables.
Coverage is narrower than the national set — only three organisms have an RC-wise susceptibility table for the non-urine population:
| Organism | 2022 ed. | 2023 ed. | 2024 ed. | Panel |
|---|---|---|---|---|
| E. coli | — (urine-only that edition) | Table 3.10 | Table 2.10 | 9 drugs (no cefazolin) |
| K. pneumoniae | — (urine-only that edition) | Table 3.11 | Table 2.11 | 9 drugs (no cefazolin) |
| S. aureus | Table 6.3 | Table 7.3 | Table 6.3 | 11 drugs (as national) |
A. baumannii, P. aeruginosa and MRSA have no RC-wise susceptibility table in any edition. The 2022 edition breaks E. coli / K. pneumoniae down by RC for urine isolates only, which is out of scope here exactly as it is for the national series.
The RC-wise tables have no year axis. Each edition's RC table reports that
edition's year and nothing else — there is no 8-year retrospective column like
the national trend tables carry. So no (organism, RC, antibiotic, year) value
is ever reported by more than one edition, and cross-edition revision
detection at RC level has essentially nothing to compare.
rc_revisions.json is therefore near-empty by design. It is produced by the
same detector as the national revisions.json
(find_rc_cross_report_revisions), kept and run so that if a future edition
ever does republish a prior year's RC table the guard fires — but on the
2022–2024 data it correctly returns nothing. At a glance, "no RC revisions
found" can look like a broken feature; it is the opposite. Contrast
revisions.json, where every calendar year is covered up to three times and 17
genuine revisions surface.
RC codes (RC1…RC21) are de-identified in the reports, and the
code-to-institution mapping is not part of the published tables. RC5 in the
2023 edition cannot be assumed to be the same laboratory as RC5 in 2024, and a
change in numbering between editions would not be signposted in the tables. On
top of that, the set itself moves:
RC15is absent from every 2024 table (E. coli, K. pneumoniae, S. aureus), though present in 2022 and 2023.RC18is absent from the S. aureus table in 2023 and 2024; the 2024 S. aureus table also omitsRC1.- The 2024 participating-centres annexure adds two hospitals (Artemis and Fortis, Gurugram) that were not in the 2022/2023 network.
So an RC set is compared against that organism's earliest edition, and every
row from an edition whose set differs carries an
rc_panel_changed(baseline=…,added=[…],dropped=[…]) flag. rc_panel.json
records the full picture. Averaging a metric "across RCs" from one edition to
the next without checking this flag compares different panels.
In the 2023 edition, three RC cells print a susceptibility of 0% where
the cell's own counts would round differently. They are carried exactly as
printed and flagged pct_mismatch — the same policy as the national series: the
dataset reports what the table shows and leaves any reconciliation to the
reader.
| Cell | Printed | n/N as a percentage |
|---|---|---|
| K. pneumoniae / RC7 / levofloxacin (Table 3.11) | 1 / 16 (0) |
6.25% |
| S. aureus / RC2 / tigecycline (Table 7.3) | 2 / 3 (0) |
66.7% |
| S. aureus / RC3 / teicoplanin (Table 7.3) | 1 / 1 (0) |
100% |
NCDC's NARS-Net feeds WHO GLASS and is entirely separate from ICMR's AMRSN. All
eight of its published editions, 2017–2024, are extracted into a separate
dataset (data/processed/narsnet_trends.{csv,json}, 558 rows — 345
E. coli, 213 S. aureus), with its own schema. Build it with
python -m src.build_narsnet_dataset.
What this section summarises is set out edition by edition, with the readings it
rests on, in docs/narsnet_v3_research.md — see
Supporting research.
High is bad in these two figures — they plot % resistant, the opposite reading from every AMRSN chart above.
E. coli and S. aureus only. Klebsiella, Pseudomonas and Acinetobacter are reported at genus level in every NARS-Net edition and at species level by AMRSN, so comparing them would mean comparing a genus against a species. All eight editions were checked for a species-level breakdown in any table, figure or annexure; there is none.
Tables are located by fuzzy caption match, never by table number, because the numbers move between editions — and because several editions print different captions in their List of Tables and in the body:
| 2017 | 2018 | 2019 | 2020 | 2021 | 2022 | 2023 | 2024 | |
|---|---|---|---|---|---|---|---|---|
| E. coli | T5 | T6 | T6 | T8 | T6 | T7 | T8 | T8 |
| S. aureus | T4 | T4 | T4 | T5 | T4 | T5 | T6 | T6 |
Cover-page years are unreliable and are never used as the index. The
edition reporting January–December 2019 has a cover reading
AMR Annual report-2020, and the 2020-data edition's cover reads
Annual Report-2021. source_report_year is always the reporting period; where
a cover year differs it is recorded separately in source_cover_year, so the
discrepancy is carried in the data rather than resolved silently. Those two
editions are the only ones with a non-null source_cover_year.
Every value in this dataset is percent RESISTANT. No NARS-Net edition
prints a susceptibility percentage anywhere, and none is derived here — see
The two series are never pooled for why
%S = 100 − %R does not hold and is not attempted.
What the tables print changes twice inside the series, and the checks follow it rather than assuming a fixed shape:
| Editions | Printed columns | Check that can run |
|---|---|---|
| 2017–2018 | denominator, % | none |
| 2019–2020 | denominator, numerator, % | % against its own counts |
| 2021 | denominator, numerator, % | same, except 15 declared cells |
| 2022–2024 | denominator, %, 95% CI | % against its own interval |
2019–2020 — every cell checkable, 8 of 108 do not agree. Seven are marginal
in the same direction: each computes to between .46 and .49 above a whole number
and is printed as the next integer up. The eighth is different in kind — 2020
S. aureus doxycycline, blood: 24 resistant of 2,638 is 0.91%, printed
beside a %R of 12. All eight are carried exactly as printed and flagged
pct_mismatch; neither figure is corrected.
2021 — the numerator is partly unusable. The E. coli Blood
Number Resistant sub-column is corrupt at source: 11 of its 13 cells do not
follow from the denominator and percentage printed beside them, and meropenem
prints 981 resistant of 854 tested. Reading the column against the printed
denominators and percentages shows the values are the column's own, printed
against the wrong rows — a one-to-one displacement, not thirteen wrong
numbers. Nothing is repaired. Those cells plus two Urine cells — 15 in all —
are declared by hand in CORRUPT_NUMERATORS, carry
numerator_status = corrupt_in_source, and get no computed_pct. The two
Blood cells that happen to agree are declared along with the rest, because which
values in a displaced column came to rest on their own row is not something the
printed table lets a reader establish. The printed %R and Number Tested are
sound throughout, and the chapter restates four of the affected percentages in
prose written independently of the table.
2022–2024 — a different check, because the numerator column is gone. These
258 rows print no numerator, so there is nothing to reconcile a percentage
against. What they do print is a confidence interval, and a percentage and an
interval are two printed statements about one quantity, so they can disagree
with no third figure available. 2 of 258 fall outside their own interval,
and they are not the same kind of thing: the 2023 S. aureus linezolid blood
cell is a difference between how two columns are rounded (the chapter gives the
year's figure as 0.2%, which the interval brackets and which rounds to the
printed 0), while the 2022 E. coli doxycycline OSBF cell prints its interval
as 24.2- 4.02 — an upper bound below the lower one. The bounds are carried
in the order printed and are not swapped, because swapping them would be a
repair.
2017–2018 — neither check reaches a single one of their 108 cells. They print a denominator and a percentage and nothing else. Back-computing a numerator as denominator × %R was considered and rejected: it would be the only invented count in the repository, and a percentage checked against a numerator derived from that same percentage cannot disagree with it, so the check it appears to buy is empty. The rows are published with the gap visible instead, corroborated by 21 specimen-stratified percentages stated in those two editions' chapters, all pinned as fixtures.
There is also a cross-column check a per-cell check cannot see. In the 2019 E. coli table, nitrofurantoin is reported for urine only; the pooled column and the Urine column print the same denominator of 16,741 — necessarily the same isolates — but different numerators, 2,026 against 2,042. Both round to 12%. Composite columns are separately compared against the columns that partition them, and that comparison deliberately raises no flag: the difference is systematic rather than exceptional, so flagging it would mark nearly every composite row and bury the findings that are real. All 50 comparisons are recorded in the extraction report instead.
These are two different facts and they come apart in both directions:
reconcilablesays whether the printed numerator can be trusted as that cell's numerator. It is false on every 2022–2024 row, and those rows are checked — against their own intervals.no_internal_check_possiblesays whether any check had two printed figures to compare. It sits on 125 cells across four editions (48 in 2017, 60 in 2018, 1 in 2020, 16 in 2021), and it is true on one 2021 row that isreconcilable.
A consumer filtering on reconcilable can never reach a count it must not use.
A consumer wanting to know whether anything was verified must read the flag.
narsnet_panel.json records, per organism per edition, the drug panel and the
specimen columns, and what changed between consecutive editions. Both matter,
and a check on one would miss the other:
- 2019 → 2020: the E. coli drug panel is identical while the pooled specimen column disappears. An edition-over-edition comparison of a pooled figure would be comparing a printed column against one no longer printed.
- 2020 → 2021: both axes move at once. The panels grow from 9 to 17 drugs for E. coli and 8 to 9 for S. aureus, and the pooled and PA+OSBF columns give way to pus aspirate and OSBF reported separately — so no 2021 specimen column has the same membership as any 2020 one.
- 2022 → 2023: E. coli prints seventeen drugs in both, and they are not the same seventeen: cefuroxime leaves, ceftriaxone joins. A check on panel size reports nothing across that step, so the check compares membership.
Nothing in the series appears, disappears and returns. Checked across all eight editions on both axes and both organisms: each of the 41 drugs and specimen columns is printed over one unbroken run, which is what makes the consecutive-edition comparison above a complete account of what moves. Two drugs run for a single edition and neither returns — E. coli ceftazidime in 2017 and S. aureus vancomycin in 2018.
Every affected row carries narsnet_panel_changed(from=…) or
narsnet_specimen_columns_changed(from=…).
narsnet_revisions.json is empty by design, and for a different reason than
the RC file's. Each NARS-Net edition reports its own period only, with no
retrospective multi-year table anywhere in the series, so no key is covered by
more than one edition and cross-edition revision detection has nothing to
compare. Contrast revisions.json, where three AMRSN editions cover each year
and 17 genuine revisions surface.
Two artefacts set the networks beside each other, and neither joins them on a value: the comparability matrix joins on keys, and the surveillance-volume figure plots counts.
The two panels overlap without matching. data/processed/comparability_matrix.json
(python -m src.build_comparability) records one cell per organism ×
antibiotic × year — which network reports it, on which metric, from which
specimen basis, out of which printed table:
| Drugs | Cells | Both | NARS-Net only | AMRSN only | Neither | |
|---|---|---|---|---|---|---|
| E. coli | 21 | 168 | 45 | 56 | 35 | 32 |
| S. aureus | 13 | 104 | 55 | 16 | 33 | 0 |
| 272 | 100 | 72 | 68 | 32 |
It carries no percentage and no count from either network, and that is
enforced rather than intended. A matrix carrying values would be exactly the
join that must not exist, so it joins on keys and stops there.
assert_carries_no_values() runs before the file is written, with two checks
that fail differently: a value-bearing field name at any depth, and any float
anywhere — every legitimate number in the file is a year or a tally of cells,
and both are integers. A test derives the forbidden-field set independently from
both row schemas, so a new schema field must be classified rather than silently
allowed through.
Panel-level overlap is not cell-level overlap, and the summary block keeps them apart. E. coli ceftazidime is in the AMRSN panel in all eight years and in a NARS-Net table in 2017 alone, so it counts as a drug both networks report on one of its eight cells; S. aureus vancomycin is the same shape, in the 2018 edition only.
docs/figures/narsnet_surveillance_volume.png is the only figure that puts
both networks on one axis, and it does so because it plots counts. An
isolate tested is the same unit on both sides in a way the two percentages are
not.
The caveat that comes with it: what each network prints as a count is not
the same thing on both sides. Isolates tested against each drug differ widely
inside a single year — E. coli urine in 2024 runs 17,191–41,460 across
sixteen NARS-Net drugs, and the AMRSN 2024 panel runs 94–12,445 across ten.
NARS-Net also prints an organism-level count per specimen in some editions: the
2022–2024 tables carry it in the column header, as Urine (N=44,711) on the
2024 E. coli table, and the 2017 and 2018 pooled columns carry one too.
AMRSN prints no organism-level count anywhere in these trend tables. Each
line is therefore the largest printed denominator in that year's panel —
not because no other count is published, but because it is the one rule that can
be applied identically to both networks and to every year. Where NARS-Net prints
a column-header N as well, it is the larger figure, and the two are not
interchangeable. The all-specimen line is combined rather than printed,
because no edition from 2021 prints a pooled column: the rule takes the
pairwise-disjoint subset of a drug-year's columns covering the most strata,
preferring the one printed as fewest columns, so a printed pooled figure is used
where one exists and the strata are summed where none does. Coverage decides,
never magnitude, and the optimum is unique for all 172 drug-years — asserted
by brute force, because a value settled by iteration order would not be a fact
about the source.
What the figure shows is a divergence in trajectory. NARS-Net grows steadily, falling only in 2020, with S. aureus falling once more in 2023. AMRSN's E. coli denominators do something else: 14,728 tested in 2022, 7,559 in 2023, 11,679 in 2024. That dip is in the source and is not a revision. Every 2023 cell is identical in the 2023 and 2024 editions — both organisms, all 21 drugs, numerator and denominator — and the dip is visible within a single edition's own retrospective column.
Because the two cover different specimen populations, what is comparable is the shape over time, not the size.
NARS-Net publishes national aggregates only. There is no site-level or state-level breakdown in any edition — not named, not coded. So the Regional Centre apparatus has no NARS-Net counterpart, and that is a property of the source rather than an unfinished feature.
Site identification is in fact inverted between the two networks: NARS-Net names its participating institutions in full, in an annexure from 2018 onward, but publishes no data against them; AMRSN de-identifies its centres as RC1–RC21 but publishes data against the codes.
- No institutional overlap, for a structural reason. All 20 publicly named AMRSN institutions were checked against both NCDC network lists. The two recruit from mutually exclusive strata: NARS-Net is exclusively state government medical colleges and state/UT institutes, while AMRSN is dominated by central institutes, private and corporate hospitals, a mission hospital and armed forces. The near-misses are same-city, different-institution pairs — PGIMER against GMCH-32 in Chandigarh, JIPMER against IGMC&RI in Puducherry. This is a plausible partial explanation for any systematic difference between the two, and belongs in any interpretation of one.
- Specimen mix is the largest confound. NARS-Net includes urine, and it is by far its largest stratum — in 2024 the largest printed E. coli urine denominator is 41,460 against 3,254 for blood. AMRSN's E. coli trend table is captioned "all samples (except faeces and urine)", while its S. aureus table says "all samples" with no exclusion stated, so the mismatch is not even the same for the two organisms. Since NARS-Net is specimen-stratified, the defensible construction is specimen-matched — most cleanly blood-only — and never a pooled figure.
- 2018 → 2019 crosses a change in method documentation. The de-duplication rule (first isolate per patient × pathogen × specimen type) is first documented in the 2019 edition. The 2017 and 2018 editions state a "unique patient" total with no stated rule and no raw counts. Any trend crossing that boundary compares numbers produced under an undocumented method against a documented one. That is an observation about method documentation maturing, and is worth stating as one.
- Three things to disclose rather than bury. ICMR names ~16–17 regional centres publicly but its reports reference up to RC21, so ~4–5 participating centres are not publicly named and one could in principle also be on NARS-Net. ICMR de-identifies every data row, so data-level overlap can never be checked — a permanent limitation, not a research gap. And NCDC's state-level AMR networks funnel additional unenumerated medical colleges into its GLASS submission, which is one more reason not to use GLASS figures as a substitute for either network's report: they cover a different population.
Both registries are fetched at run time from the publishers' own servers;
never redistributed here (see DATA_LICENSE.md). src/fetch.py pins a
SHA-256 for every file in both, and a mismatch is a hard failure rather than a
warning — the table locations and source defects recorded above were established
against exactly those bytes.
| Edition | Year covered | URL |
|---|---|---|
| 8th | 2024 | icmr.gov.in/icmrobject/uploads/Report/1763981012_icmramrsnannualreport2024.pdf |
| 7th | 2023 | icmr.gov.in/icmrobject/uploads/Documents/1725536060_annual_report_2023.pdf |
| 6th | 2022 | icmr.gov.in/icmrobject/custom_data/pdf/resource-guidelines/AMRSN_Annual_Report_2022.pdf |
A hash change means ICMR re-uploaded the report — the fetcher reports it loudly instead of overwriting, because extracted numbers may then differ from previously published results.
Fetched with python -m src.fetch --network narsnet. All eight are registered
in NARSNET_SOURCES, verified 2026-09-01.
| Data year | URL |
|---|---|
| 2024 | ncdc.mohfw.gov.in/uploads/pdf/amr30.pdf |
| 2023 | ncdc.mohfw.gov.in/uploads/pdf/amr32.pdf |
| 2022 | ncdc.mohfw.gov.in/uploads/pdf/amr34.pdf |
| 2021 | ncdc.mohfw.gov.in/uploads/pdf/amr35.pdf |
| 2020 | ncdc.mohfw.gov.in/uploads/pdf/amr36.pdf |
| 2019 | ncdc.mohfw.gov.in/uploads/pdf/amr37.pdf |
| 2018 | ncdc.mohfw.gov.in/uploads/pdf/amr38.pdf |
| 2017 | ncdc.mohfw.gov.in/uploads/pdf/amr39.pdf |
Use the /uploads/pdf/ paths. wp-content copies exist but the 2024 one
truncates before its annexure.
Treat these URLs as ephemeral. NCDC's have already migrated twice
(ncdc.gov.in → ncdc.mohfw.gov.in, and /WriteReadData/… →
/wp-content/uploads/… with /uploads/pdf/amrNN.pdf running in parallel), and
citations published in 2022–2024 already point at dead links.
No NARS-Net edition carries a suggested citation, ISBN, DOI, report number, or
named authors — checked in the front matter of all eight. src/references.py
generates a Vancouver entry per edition from the same registry the fetcher uses,
so the bibliography cannot cite a document the pipeline never read. The form is
a corporate author, the reporting period spelled out in the title (there is
no ordinal, so "8th ed." has no counterpart here), doi and isbn left unset,
and an access date recorded because of the URL churn. Only two editions print a
publication date anywhere — July 2022 for the 2021-data report and July 2023 for
the 2022-data one — so the other six carry [date unknown] rather than a year
inferred from those two. The cover year is never substituted for it.
Two documents in docs/ carry the material the NARS-Net sections above
summarise. Both are written to be read on their own.
docs/narsnet_v3_research.md— what each of the eight editions prints and what can be checked against what: the drug panels per edition, the table locations, the column structures and how they change twice mid-series, the cells that do not reconcile and why, and the structural differences that constrain any comparison between the two networks. Its section labels are cited from the code and tests, so they are stable.docs/narsnet_investigation_artifacts.md— the registry of the eight editions with their pinned SHA-256 hashes, and the 2024 edition's full 54-site roster. The hash table is read bytests/test_narsnet_sources.pyand asserted againstsrc/sources.py.
index.html is a self-contained static page, published at
yashita-thakur.github.io/icmr-amr-tracker
from the main branch. It has two rendering paths for the same charts:
- Static PNGs from
viz/trend_charts.py— what you get with no JavaScript. - Interactive SVG charts that load
docs/data/trends.jsonand let you select antibiotics to compare, with a reset control.
Two rendering paths are only safe while they share one source of truth, so
docs/data/trends.json is generated by the same latest_edition_series()
function that draws the PNGs, and carries the same exclusions. It is 9 KB;
serving the full 599 KB dataset to a landing page would be wasteful.
The interactive layer is strictly progressive enhancement — if the fetch fails
(offline, file://, missing file) the PNGs simply stay put. Two deliberate
safeguards, both learned the hard way while building it:
- The reveal-on-scroll styles are scoped behind a class that JavaScript adds.
Unscoped, a script failure would leave the whole page at
opacity: 0— blank, with no error anywhere. - The reveal has a backstop that drops the transition entirely after a short
delay. Content being readable must never depend on an animation running to
completion; an
IntersectionObserverthat never fires (non-compositing tab, prerender, browser quirk) would otherwise hide everything permanently.
No charting library is used. The page has no third-party JavaScript at all — the charts are ~200 lines of vanilla JS emitting inline SVG, which avoids both a CDN dependency and vendoring a library into the repo.
Regenerate everything with:
python viz/trend_charts.py --revisions && python -m src.references --injectTrend charts omit points where the source published no percentage, where the cell is asterisked as low-count, or where fewer than 30 isolates were tested. This is presentational only — every such row is still in the dataset.
The reason is concrete: cefazolin is tested against a handful of E. coli
isolates a year (*0/8, *0/1, *2/6), and charting it draws a line swinging
between 0% and 44% that means nothing at all. Plotting it beside meropenem
invites exactly the misreading this project exists to prevent.
Releases are tagged v0.1.0 through v0.4.0, matching the version numbers in
CHANGELOG.md and CITATION.cff. The changelog additionally refers to them as
V1–V3, which is the vocabulary the commit history uses. The mapping:
| Release | In this README | What it added |
|---|---|---|
| V1 | the AMRSN national series | E. coli and K. pneumoniae, 2017–2024, three editions |
| V1.1 | the AMRSN national series | Four more organisms across two more chapters — A. baumannii, P. aeruginosa, S. aureus and MRSA |
| V2 | the AMRSN Regional Centre series | Regional Centre breakdowns for the three organisms that have an RC-wise susceptibility table, flagging editions whose RC panel changed rather than averaging across it as if it were stable |
| V3 (current) | the NARS-Net series | NCDC NARS-Net carried as a parallel series, all eight editions 2017–2024, E. coli and S. aureus, plus the cross-network coverage map |
Cite both this project (see CITATION.cff) and the underlying annual
report — the network the row came from (ICMR-AMRSN or NCDC NARS-Net), its
report edition, and its table number — for each value used. Every row in both
datasets carries the fields needed to do so, and DATA_LICENSE.md sets out the
required attribution in full.