Checking bills against what hospitals can do
IIB Board AI Pack, page 25. Second phase. Built here on mock and synthetic data.
Today. ROHINI lists network hospitals, and the national health facility registry records what each facility declares about itself. Each insurer sees only its own claims from a hospital.
What would change. AI matches hospitals across the two registries, even when names and addresses differ, and maps each procedure to the facilities it needs. Hospitals billing across all insurers for care they do not appear equipped to give are flagged for review.
Two lists, one hospital. ROHINI is the list of hospitals in insurers' networks, with the facilities each declares. The Health Facility Registry (HFR) is the national list where each facility describes itself. Here the registry is mock data built from the network list, with names, addresses and services written the way another registry would write them.
The check. Match each hospital across the two lists. Map each procedure to the facility it needs, for example bypass surgery needs a cardiac surgery unit and angioplasty a cath lab (a catheterisation laboratory). Then flag hospitals that bill, across all insurers, for care that neither list says they can give. A gap in only one list is a record to check, not a reason to review bills.
The typical question, answered
“Which hospitals bill for cardiac surgery without a declared cardiac unit?” Claims admitted 1 October 2024 to 4 October 2026, across all insurers.
5 hospitals in 5 states billed bypass surgery or angioplasty although neither registry lists a cardiac surgery unit or a cath lab for them. Together they made 60 such claims, billed ₹1.8 cr, across 20 insurers. No insurer saw more than 4 of one hospital's claims, so none would notice alone.
| State | Hospitals | Claims | Billed | Approved | Insurers |
|---|---|---|---|---|---|
| Gujarat | 1 | 15 | ₹51.4 lakh | ₹25.3 lakh | 11 |
| Rajasthan | 1 | 10 | ₹37.9 lakh | ₹25.3 lakh | 9 |
| Uttar Pradesh | 1 | 10 | ₹32.2 lakh | ₹22.6 lakh | 8 |
| Maharashtra | 1 | 13 | ₹29.2 lakh | ₹23.3 lakh | 9 |
| Odisha | 1 | 12 | ₹25.8 lakh | ₹12.8 lakh | 8 |
Leadership sees state totals. Hospital detail is for the Vertical Heads, Data Operations and the Analyst.
Rules wrote this answer from the figures in the table. No AI wrote text here.
Where they are
Shaded by what the flagged hospitals billed for bypass surgery and angioplasty. Grey states have none.
One hospital, both lists side by side
Registry records to check
50 hospitals bill for a facility that the network list declares but the matched registry record leaves out. These go to the team that keeps the registries, not to claims review. Of the 48 hospitals with no registry record matched, 47 bill only for facilities the network list declares, so only the network list speaks for them.
| State | Hospitals | Claims | Billed |
|---|---|---|---|
| Maharashtra | 8 | 233 | ₹1.6 cr |
| Telangana | 7 | 237 | ₹1.9 cr |
| Delhi | 6 | 180 | ₹3.9 cr |
| Uttar Pradesh | 6 | 159 | ₹1.0 cr |
| Tamil Nadu | 5 | 32 | ₹14.8 lakh |
| Rajasthan | 4 | 92 | ₹1.4 cr |
| Haryana | 3 | 51 | ₹31.5 lakh |
| Karnataka | 3 | 29 | ₹43.4 lakh |
| Gujarat | 2 | 238 | ₹3.6 cr |
| Jharkhand | 2 | 7 | ₹7.0 lakh |
| Andhra Pradesh | 1 | 35 | ₹22.7 lakh |
| Jammu and Kashmir | 1 | 2 | ₹2.0 lakh |
| Kerala | 1 | 54 | ₹89.3 lakh |
| West Bengal | 1 | 11 | ₹31.7 lakh |
How the two lists were matched
Each hospital is compared only with registry records in the same district. A score from 0 to 1 weighs the name, the address, the PIN code, the beds and the facility type. At 0.74 or above it is a clear match. Between 0.42 and 0.74 it is borderline: 61 pairs for 56 hospitals. Once asked, the AI looks at each of these hospitals with all its borderline records together, as a reviewer would, and picks one or none. Until then an offline threshold of 0.55 decides each pair. Matching never looks at the declared services, so it cannot hide or invent a gap.
Against the mock's ground truth, the offline threshold is right on 55 of the 61 borderline pairs. On the 61 the AI decided, the AI is right on 59 and the threshold on 55.
How the matching works
Names and addresses are cleaned first: lower case, abbreviations written out (Hosp. as hospital, N.H. as nursing home, Rd as road), legal suffixes such as Pvt Ltd dropped, and spellings folded so common transliterations meet (Sanjeevani and Sanjivani, Shree and Sri). The registry's district text is read through the district list and older names (Gurgaon is Gurugram). 70 records use an older or everyday district name.
Name similarity uses TF-IDF on character n-grams with cosine similarity (scikit-learn). TF-IDF counts short runs of letters and gives less weight to runs that are common everywhere, such as "hosp". Address similarity uses the same method on words. Weights: name 0.45, address 0.2, PIN code 0.15, beds 0.1, facility type 0.1. A missing field is left out of the average. Weights and thresholds were set by hand on this mock registry, so the matching scores here are optimistic. One hospital takes at most one record, best score first.
The AI sees a hospital and its borderline records: name, type, beds, address, district and PIN code. It never sees the services, the claims or any id. It picks one record or none, with a reason, and each answer is cached by a fingerprint of what it saw. A reason that quotes a number not in the records, or breaks the house style, is replaced by a rules reason.
The registry's service names are read with a list of synonyms, for example CTVS and Cardiothoracic Surgery both mean a cardiac surgery unit, and Interventional Cardiology means a cath lab. Cardiology OPD is an outpatient clinic and does not count.
How we would measure it
The problems were built into the test data, so these scores show the check finds what was planted. The real test is a pilot, where reviewers judge each flagged hospital. In the test data, 5 hospitals were planted that bill bypass surgery and angioplasty with neither a cardiac surgery unit nor a cath lab in either list.
What the registry adds. In this test data the network list alone would flag the same 5 hospitals, so the precision above is partly built in. The registry is a second, independent witness. A hospital goes for review only when neither list declares the facility. That is why the matching matters. Two planted hospitals each have a sister cardiac centre with the same name in the same district, whose registry record declares a cardiac surgery unit or a cath lab. A link to it would have hidden the flag. Both were turned down.
Links made: 852, of which 851 are right (link precision 99.9%). True links found: 851 of 855 (link recall 99.5%). With the offline threshold alone, matching is right for 892 of 900 hospitals (99.1%).
Which facility each procedure needs
| Procedure | Needs |
|---|---|
| Heart attack, medical treatment | an intensive care unit |
| Coronary artery bypass graft | a cardiac surgery unit |
| Coronary angioplasty with stent | a cath lab |
| Appendix removal | an operating theatre |
| Gall bladder removal | an operating theatre |
| Hernia repair | an operating theatre |
| Hysterectomy | an operating theatre |
| Newborn intensive care | a newborn intensive care unit |
| Haemodialysis, a month of sessions | a dialysis unit |
| Kidney stone removal | an operating theatre |
| Caesarean delivery | a maternity ward |
| Normal delivery | a maternity ward |
| Chemotherapy cycle | cancer care |
| Cancer surgery | cancer care |
| Cataract surgery with lens | an eye theatre |
| Fracture fixation | an orthopaedic theatre |
| Total hip replacement | an orthopaedic theatre |
| Total knee replacement | an orthopaedic theatre |
From the procedures catalogue. Medical treatment for fevers, infections and breathing illness needs no special facility.
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.