Use 22 · AI · Reads outside reports and predicts · Health, Life

Forecasting admissions from outbreaks and weather

IIB Board AI Pack, page 24. Second phase. Built here on mock and synthetic data.

Today. The government publishes a weekly report of disease outbreaks by district, and weather and air quality data are public. These sit apart from insurance claims.

What would change. AI reads each weekly outbreak report and pulls out the district, disease and case numbers. Combined with rainfall, heat and air quality, it forecasts hospital admissions two to four weeks ahead, so insurers and hospitals can prepare.

Who would use it. The Health and Life verticals, insurers and, in summary form, public health teams

Data it uses. Weekly outbreak reports (IDSP style) · mockWeather and air quality · mockIIB's health claims · synthetic

When. Second phase

Next three weeks

No rise

No state passes the rise rule for dengue. About 117 admissions forecast in all, against 286 in the last three weeks.

See the answer

Planted surges called ahead

5 of 6

dengue surges built into the test data, called before admissions rose

See each surge

Forecast error, two weeks ahead

23% smaller

than "the same as the last four weeks", over the last 12 weeks

See the measure

Report rows read right

97.3%

by the rules reader, all 1,067 entries. AI reader 99.6% on 11 reports

See what was read

1. Where admissions are likely to rise

A typical question. “Where are dengue admissions likely to rise over the next three weeks?”

Data up to Illness

Over the three weeks from 5 October 2026, no state passes the rise rule for dengue. The largest rise forecast is in Assam: 3.2 admissions in the week of 12 Oct, against 1.2 a week over the last four weeks. Across the country the radar expects about 117 dengue admissions over the three weeks, against 286 in the last three.

State	Last 4 weeks, a week	Forecast, week of 5 October 2026	Forecast, week of 12 October 2026	Forecast, week of 19 October 2026	Call
Assam	1.2	2.9	3.2	1.2	No rise expected
Karnataka	11.8	12.3	1.5	1.7	No rise expected
Tripura	0.0	0.1	0.1	0.1	No rise expected
Arunachal Pradesh	0.0	0.1	0.1	0.1	No rise expected
Lakshadweep	0.0	0.0	0.1	0.0	No rise expected
Jammu and Kashmir	0.0	0.0	0.1	0.1	No rise expected
Manipur	0.0	0.0	0.0	0.0	No rise expected
Mizoram	0.0	0.0	0.0	0.0	No rise expected
Meghalaya	0.0	0.0	0.0	0.0	No rise expected
Sikkim	0.0	0.0	0.0	0.0	No rise expected
Odisha	0.8	0.8	0.7	0.6	No rise expected
Nagaland	0.0	0.0	0.0	0.0	No rise expected

Likely to rise means a forecast week at least 2 times the last four weeks' average and at least 6 admissions more. The table shows the 12 states with the largest forecast rise. Pick an earlier week to see what the radar said then, and what happened.

That week heavy rain fell in Telangana, the start of a dengue surge built into the test data.

The forecast reads admissions, mock rain, heat and air quality, and the outbreak reports as the rules reader read them. No personal data is used.

2. One state: forecast against what happened

State Forecast made

Lines are weekly dengue admissions. Bars are rain in millimetres (mock). The dashed line is the forecast for the next four weeks.

Loading

The chart loads here.

3. The outbreak report, and what was read from itmock

One mock report a week, 104 in all, in the style of the Integrated Disease Surveillance Programme (IDSP), the government's weekly list of outbreaks by district. They hold 1,067 outbreak entries, written the way real reports vary: labelled lines, short lines and prose, older district names, numbers in words. Every outbreak is fictional.

Report for the week ending

WEEKLY OUTBREAK REPORT (MOCK)
In the style of the Integrated Disease Surveillance Programme. Made up for the IIB demo.
Week 40 of 2026: 28 September 2026 to 4 October 2026.
Every outbreak below is fictional. District names are real places used for illustration only.

Summary. 19 outbreaks reported this week. 1 reported late. 3 earlier outbreaks followed up.

PART A. OUTBREAKS REPORTED THIS WEEK
1. Visakhapatnam (Andhra Pradesh). Viral hepatitis A. 45 cases, no deaths. Started 27 Sep 2026. Under surveillance.
   Samples sent to the regional laboratory in Chennai.

2. Assam, Morigaon. ADD. 65 cases, 1 death. Started 26 Sep 2026. Under surveillance.
   Cases treated at the community health centre.

3. Ref AS/BAR/2026/40/003 | State: Assam | District: Barpeta | Disease: Diarrhoea and vomiting | Cases: 61 | Deaths: 1 | Date of start: 24.09.26 | Status: Under surveillance
   Comment: Health education given in schools and markets.

4. Ref AS/KMR/2026/40/004 | State: Assam | District: Kamrup | Disease: Diarrhoea and vomiting | Cases: 59 | Deaths: 2 | Date of start: 26-09-2026 | Status: Under surveillance
   Comment: Chlorine tablets and ORS given to 325 families.

5. Ref AS/KMM/2026/40/005 | State: Assam | District: Guwahati (Kamrup Metro) | Disease: ADD | Cases: 31 | Deaths: 1 | Date of start: 25.09.26 | Status: Under surveillance
   Comment: Water samples from 8 sources found unfit to drink.

6. Barpeta (Assam). Suspected leptospirosis. 17 cases, 2 deaths. Started 27 September 2026. Under surveillance.
   Water samples from 39 sources found unfit to drink.

7. Ref AS/MOR/2026/40/007 | State: Assam | District: Morigaon | Disease: Leptospirosis | Cases: 17 | Deaths: 1 | Date of start: 24/09/2026 | Status: Under surveillance
   Comment: Health education given in schools and markets.

8. Ref AS/MOR/2026/40/008 | State: Assam | District: Morigaon | Disease: Acute diarrhoeal disease, cholera confirmed | Cases: 14 (9 lab confirmed) | Deaths: Nil | Date of start: 25.09.26 | Status: Under surveillance
   Comment: Water samples from 24 sources found unfit to drink.

9. A cluster of leptospirosis was reported from Kamrup, Assam. Since 24 September, eight cases have been found and none died. Surveillance continues. A leaking pipeline near the market was repaired.

10. Ref CG/BSP/2026/40/010 | State: Chhattisgarh | District: Bilaspur | Disease: Acute diarrhoeal disease | Cases: 44 (22 lab confirmed) | Deaths: Nil | Date of start: 27.09.26 | Status: Under surveillance
   Comment: Water samples from 33 sources found unfit to drink.

11. Ref JK/SRI/2026/40/011 | State: Jammu and Kashmir | District: Srinagar | Disease: Food poisoning | Cases: 23 (7 lab confirmed) | Deaths: Nil | Date of start: 23.09.26 | Status: Under surveillance
   Comment: Health education given in schools and markets.

12. Ref JK/SRI/2026/40/012 | State: Jammu and Kashmir | District: Srinagar | Disease: Varicella (chickenpox) | Cases: 21 (10 lab confirmed) | Deaths: 0 | Date of start: 27-09-2026 | Status: Under surveillance
   Comment: Health education given in schools and markets.

13. A cluster of hepatitis A was reported from Belagavi, Karnataka. Since 24 September 2026, 21 cases have been found and none died. The outbreak is under surveillance. Cases treated at the community health centre.

14. Ref KA/BEL/2026/40/014 | State: Karnataka | District: Belagavi | Disease: Leptospirosis | Cases: 7 | Deaths: 1 | Date of start: 22-09-2026 | Status: Under surveillance
   Comment: Active case search in 32 villages and wards.

15. Ref LD/KAV/2026/40/015 | State: Lakshadweep | District: Lakshadweep | Disease: Jaundice (hepatitis A) | Cases: 20 | Deaths: 0 | Date of start: 22-09-2026 | Status: Under surveillance
   Comment: Chlorine tablets and ORS given to 241 families.

16. Ref NL/KOH/2026/40/016 | State: Nagaland | District: Kohima | Disease: Acute diarrhoeal disease (ADD) | Cases: 22 | Deaths: 1 | Date of start: 22-09-2026 | Status: Under surveillance
   Comment: Samples sent to the regional laboratory in Hyderabad.

17. Punjab, Ludhiana. ADD. 44 cases, 1 death. Started 26.09.26. Under surveillance.
   Active case search in 52 villages and wards.

18. Rajasthan, Udaipur. Food poisoning. 48 cases, no deaths. Started 26.09.26. Under surveillance.
   Active case search in 48 villages and wards.

19. Ref UP/AGR/2026/40/019 | State: Uttar Pradesh | District: Agra | Disease: Acute encephalitis syndrome (AES) | Cases: 14 | Deaths: 2 | Date of start: 23.09.26 | Status: Under surveillance
   Comment: Fogging and source reduction done in 423 houses.

PART B. OUTBREAKS REPORTED LATE
20. Delhi, North West Delhi. Diarrhoea and vomiting. 35 cases, 1 death. Started 15 Aug 2026. Under control.
   Water samples from 26 sources found unfit to drink.

PART C. FOLLOW-UP OF EARLIER OUTBREAKS
21. Karnataka / Bangalore Urban / Dengue (NS1 positive): earlier outbreak, started 17.09.26. Total cases now 264, deaths 3. Under surveillance.

22. Karnataka / Bengaluru Urban / Malaria: earlier outbreak, started 20/09/2026. Total cases now 56, deaths 0. Under surveillance.

23. Follow-up. Greater Mumbai, Maharashtra. Falciparum malaria. First reported in week 37. 22 new cases this week, 130 in all since 4 September. No deaths so far. Under surveillance.

End of report.

23 entries, read by the AI reader

The AI read 23 of 23 entries right. The rules read 23 right.

No.	State	Illness	Cases	Deaths	Started	Check
1	Andhra Pradesh	Hepatitis A	45	0	27 Sep 2026	Right
2	Assam	Acute diarrhoeal disease	65	1	26 Sep 2026	Right
3	Assam	Acute diarrhoeal disease	61	1	24 Sep 2026	Right
4	Assam	Acute diarrhoeal disease	59	2	26 Sep 2026	Right
5	Assam	Acute diarrhoeal disease	31	1	25 Sep 2026	Right
6	Assam	Leptospirosis	17	2	27 Sep 2026	Right
7	Assam	Leptospirosis	17	1	24 Sep 2026	Right
8	Assam	Cholera	14	0	25 Sep 2026	Right
9	Assam	Leptospirosis	8	0	24 Sep 2026	Right
10	Chhattisgarh	Acute diarrhoeal disease	44	0	27 Sep 2026	Right
11	Jammu and Kashmir	Food poisoning	23	0	23 Sep 2026	Right
12	Jammu and Kashmir	Chickenpox	21	0	27 Sep 2026	Right
13	Karnataka	Hepatitis A	21	0	24 Sep 2026	Right
14	Karnataka	Leptospirosis	7	1	22 Sep 2026	Right
15	Lakshadweep	Hepatitis A	20	0	22 Sep 2026	Right
16	Nagaland	Acute diarrhoeal disease	22	1	22 Sep 2026	Right
17	Punjab	Acute diarrhoeal disease	44	1	26 Sep 2026	Right
18	Rajasthan	Food poisoning	48	0	26 Sep 2026	Right
19	Uttar Pradesh	Acute encephalitis syndrome	14	2	23 Sep 2026	Right
20	Delhi	Acute diarrhoeal disease	35	1	15 Aug 2026	Right
21	Karnataka	Dengue	264	3	17 Sep 2026	Right
22	Karnataka	Malaria	56	0	20 Sep 2026	Right
23	Maharashtra	Malaria	130	0	4 Sep 2026	Right

The AI reads one report per call and the answer is kept, so a report is never sent twice. The rules reader has read every report, and the forecast always uses the rules reader's rows.

The AI is given only the report text, with the lists of district codes and illnesses to choose from. It returns rows, which are checked before use.

4. How we would measure it

How we would measure it: Forecasts compared with actual admissions, district by district

0.86 against 1.11

Two weeks ahead, the radar's dengue forecast was off by 0.86 admissions a week in each district on average, against 1.11 for the simple rule "the same as the last four weeks". Each of the last 12 weeks with data was taken as today, using only what was known that week, and the forecasts were compared with what happened.

Rise calls that came true

50%

15 of 30 calls that a district's dengue admissions would rise in the next three weeks

Real rises the radar called

50%

15 of 30 times dengue admissions rose by the rule

Planted surges called ahead

5 of 6

2 of them two weeks before admissions rose, the rest one week before

Outbreak entries read right

97.3%

rules reader, 1,038 of 1,067 entries in 104 reports. AI reader 99.6% on the 11 it has read

The surges, the outbreak reports and the rain were built into the test data, so these scores show the method works as designed. The real test is a pilot on real reports and real claims. Rise calls are counted over 12 forecast weeks and every district, so a call is one district in one week. The rule is the same for forecasts and for what happened. Two weeks ahead, the forecasts added up to 1,332 admissions against 1,548 that happened, 14% low. 187 of the 216 missing admissions were in the 6 cities with planted surges, which were larger than their forecasts.

Error by how far ahead

Forecast made	Weeks scored	Radar error	Simple rule error	Radar better by
1 week before	12	0.78	1.03	24%
2 weeks before	11	0.86	1.11	23%
3 weeks before	10	0.90	1.11	19%
4 weeks before	9	0.96	1.13	16%

Error is the average gap between forecast and actual dengue admissions, per district and week. Further ahead the radar knows less: rain that has not fallen and outbreaks not yet reported.

The planted dengue surges

State	Week of heavy rain	Rain, mm	Surge, weeks of	Called
Uttar Pradesh	6 July 2026	267	20 Jul and 27 Jul	Missed
Maharashtra	20 July 2026	643	3 Aug and 10 Aug	Two weeks ahead
Telangana	3 August 2026	243	17 Aug and 24 Aug	Two weeks ahead
Delhi	10 August 2026	232	24 Aug and 31 Aug	One week ahead
West Bengal	24 August 2026	241	7 Sep and 14 Sep	One week ahead
Karnataka	7 September 2026	228	21 Sep and 28 Sep	One week ahead

Story H3: dengue admissions surge two and three weeks after heavy rain in six cities. Called means the radar said "likely to rise" in the week of the rain (two weeks ahead) or the week after (one week ahead), before the first surge week. Uttar Pradesh was missed: its forecast stayed under the bar. The calls one week ahead had that week's outbreak report to go on. In the test data those reports were drawn from the next week's admissions, so real reports would give less warning.

Forecast error state by state

Forecast made

The radar was closer than the simple rule in 21 of the 32 states with any dengue admission in the weeks scored. Largest states first.

State	Admissions	Forecast	Radar error	Simple rule error	Closer
Maharashtra	350	338.2	8.41	19.30	Radar
Delhi	157	80.6	6.95	10.52	Radar
Telangana	152	111.0	5.70	12.16	Radar
Uttar Pradesh	135	121.1	6.39	10.57	Radar
Karnataka	100	71.0	3.09	5.50	Radar
West Bengal	96	66.8	5.21	6.07	Radar
Gujarat	89	70.4	3.01	4.02	Radar
Kerala	77	90.6	3.09	3.09	Simple rule
Madhya Pradesh	70	57.3	1.67	3.39	Radar
Tamil Nadu	56	46.0	1.68	2.23	Radar
Haryana	32	24.3	1.69	2.16	Radar
Rajasthan	31	27.8	1.11	1.41	Radar
Bihar	30	31.5	1.41	1.05	Simple rule
Andhra Pradesh	29	23.8	1.41	1.93	Radar
Chandigarh	22	17.1	1.41	1.59	Radar
Assam	20	28.6	1.06	1.00	Simple rule
Jharkhand	16	18.9	1.01	1.11	Radar
Chhattisgarh	15	17.3	0.85	1.07	Radar
Odisha	15	23.3	1.15	0.95	Simple rule
Uttarakhand	13	20.4	0.98	1.00	Radar
Punjab	12	10.0	0.67	0.86	Radar
Goa	11	7.2	0.87	0.75	Simple rule
Dadra and Nagar Haveli and Daman and Diu	7	4.4	0.73	0.68	Simple rule
Andaman and Nicobar Islands	4	5.7	0.56	0.52	Simple rule
Himachal Pradesh	2	0.6	0.22	0.30	Radar

Reading the reports, field by field

Field	Rules reader, all 104 reports	Rules reader, the 11 the AI read	AI reader, same reports
District	97.3%	95.7%	99.6%
Disease	100.0%	100.0%	100.0%
Cases	100.0%	100.0%	100.0%
Deaths	100.0%	100.0%	100.0%
Start date	100.0%	100.0%	100.0%
Status	100.0%	100.0%	100.0%
Whole entry right	97.3%	95.7%	99.6%

The AI has read 11 of the 104 reports so far. Both readers are scored on the same reports. The rules reader was written for this one report format. All 29 of its mistakes are district names written in a way it was not given: North Twenty Four Parganas, Gautam Budh Nagar, South Twenty Four Parganas, Dehra Dun, Ranga Reddy, Kamrup (Metro). A person writing rules cannot think of every spelling. The AI matches names on meaning, which is why it did better on the reports it read.

Life can use the same radar. The same reports listed 58 deaths from outbreaks reported since 13 July 2026. Life death claims are in the Datahub by district and week too (313 intimated in the same weeks), so the radar could watch for districts where death claims may follow an outbreak. It would need its own test against Life claims before anyone relies on it.

How the radar works, and what to keep in mind

What it reads. Weekly admissions for dengue and malaria by district from IIB's health claims (synthetic here), rain, top temperature and air quality by district and week (mock, standing in for IMD and CPCB), and the weekly outbreak reports (mock).

How it forecasts. Gradient boosted decision trees (scikit-learn HistGradientBoostingRegressor) with a Poisson loss, which suits counts. Every district is in one model, so a lesson from one district's outbreak carries to others. One model for each horizon, one to four weeks ahead, dengue and malaria together.

How it was tested. Each of the last 12 weeks with data was taken in turn as "today". The forecast used only what was known that week. The model was retrained every 2 weeks on everything known by then, 28 models in all. Those forecasts were then compared with the admissions that happened, and with the simple rule "the same as the last four weeks". The whole run took 13.6 seconds and is kept until the data changes.

What moves the forecast most, two weeks ahead. The rise in error when one input is shuffled.

Input	Error added
The district's usual week so far	0.096
Rain this week	0.043
Rain a week earlier	0.041
Average of the last 4 weeks	0.031
Time of year	0.019
Air quality index	0.011
Dengue or malaria	0.009
Top temperature	0.009

Keep in mind. The mock reports were drawn from the synthetic admissions about a week later, so they lead admissions by design. Real reports would lead by less, and less reliably. Real IDSP reports often come out a week or more late, which cuts the warning time. The surges planted in story H3 have no earlier example in the two years of data, so the radar can learn them only from the first surges in the test weeks. 278 of the 320 dengue and malaria entries follow the admissions this way. The rest are small outbreaks that never reach insured patients, as happens in real life. The flood in lower Assam (story P2) adds 15 water-borne outbreak entries to the last two reports.

Personal data is masked before any person or the AI sees it. The AI flags and drafts. A person decides. Mock data stands in for outside sources that need an agreement.