Indian insurer IFFCO Tokio General Insurance has revealed it used AI software to uncover and stop fraudulent claims from hospitals under the mass health insurance scheme known colloquially as "Modicare".
Senior executive director of IT Seema Gaur told a H2O World India summit last month that it had observed possible - and actual - fraudulent claims being submitted by a small number of remote hospitals under the scheme.
The scheme is formally known as Pradhan Mantri Jan Arogya Yojana (PMJAY) or Ayushman Bharat, and is a government-sponsored health service delivered in collaboration with insurers that provides free care to India’s poorest citizens and families.
Gaur said IFFCO Tokio received 1200 claims per day under the scheme, which were reviewed and adjudicated by a team of 10 doctors.
Within “four-to-five months”, Gaur said an internal audit team had uncovered claims that contained duplicate support materials, such as pictures of a patient or specimen removed during surgery, pathology reports or prescriptions.
With the adjudication team being so small, “every doctor had to see 120 claims per day and it was practically impossible for them to go through the details and match the pictures,” Gaur said.
Tech support for doctors
IFFCO Tokio started to look at technology support for doctors in the form of artificial intelligence and signed up to use H2O.ai, a startup it had first experimented with back in 2020.
That early experiment had lapsed - it was to predict fraud instances in motor vehicle insurance claims but was stood up just prior to the pandemic starting.
The drop in vehicular traffic meant the model’s efficacy was never properly tested, and it was replaced with a “rules-based engine” later in the year.
H2O.ai’s second chance at IFFCO Tokio took until the end of 2022 to materialise, but a working model was in place quickly.
“We signed an agreement with H2O.ai in October; in November we did the pilot; and since December, it has been working very fine,” Gaur said.
When it was initially deployed against the full history of claims under the scheme, the model found 100,000 duplicated images.
Gaur recalled examples where an image of a gallbladder that was removed was shared by 50 different cases, and a single pathology report was duplicated “over 10 to 20 claims.”
“So there were people and hospitals who were cheating us, and we’d been paying them over the period,” Gaur said.
With an enormous dataset, Gaur said the insurer asked H2O.ai to break down the supporting documents filed for each claim - typically in PDF format - into different document types, such as images of specimen or patient, lab reports or handwritten documents.
Even that was unwieldy, however, and the insurer effectively drew a line and narrowed the scope of the model to provide alerts for every day from a certain point.
“We approached H2O.ai and said, ‘Give us only today’s data … because my doctors don’t have time to go through the back information. What is paid is gone. Tell me only today so I can prevent the [potentially fraudulent] payment for tomorrow.”
That produced a workable rolling dataset and insights, and Gaur said the adjudicating doctors had been “really surprised the tool could give them so much accuracy.”
Additional adjustments had since been made to the system, often at the request of doctors. The work was being performed mostly by H2O directly as well as by a partner, Easy Data Analytics (EDA), which is based out of Noida.
Gaur said she would not share a monetary amount for the amount of fraud attempts - and rejected claims - made possible through the use of the tool.
Unpacking its AI journey
More broadly, Gaur showed IFFCO Tokio’s journey with artificial intelligence, which started back in 2019.
The insurer has “hardly four or five AI engineers” of its own; Gaur said that it did not typically recruit data scientists but instead targeted “computer science freshers from colleges”. Freshers in India refers to graduates.
“We give them a platform and an environment to groom themselves [to become data scientists],” she said.
They had built some models that are used in production, notably one for renewal propensity in motor insurance, which predicted the customers who would pay or leave.
Gaur said that particular model resulted in a five percent increase in retention ratio.
“Five percent is a huge amount because we underwrite around 7 million motor policies in a year,” she said.
The insurer otherwise has leant on technology partners to get AI and machine learning (ML) projects off the ground.
It started out with AI-based damage assessment for motor vehicles, which used IBM technology.
Gaur said customers could take and submit photos of damaged vehicles, find out which parts were repairable or replaceable, get a claim estimate from the insurer, and receive payment to their bank account - within half an hour.
In 2020, the insurer collaborated with a Bangalore-based startup - not named - to apply AI-based optical character recognition (OCR) technology to invoices from workshops that handled motor vehicle repairs.
The technology helped to process invoices that were otherwise barely legible or were in .gif or .tif formats captured via a mobile phone.
This year and into the future, IFFCO Tokio intends to revive fraud analysis on motor vehicle claims, using a revamped model by H2O.ai.
The company will also target other fraud detection around motor vehicle claims, as well as quick settlement of health claims, using H2O.ai’s technology.
