New AI Forensics System Analyzes 1.34 Million Digital Records in Health Insurance Fraud Investigation

Researchers have developed an artificial intelligence system capable of analyzing large collections of digital evidence, reporting that it processed more than 1.34 million forensic records and identified 11,200 potentially relevant evidence items during a health insurance fraud investigation.

The system, called SAGE, was described in research published on October 29, 2026, in Mobile Networks and Applications. The findings received renewed attention in scientific reporting on November 6.

Developed by Weilong Ding, Haoyang He, Han Tian, and Jie Guo, SAGE uses multiple large language model (LLM) agents to examine information collected from different digital sources.

The research focuses on health insurance fraud investigations, particularly cases involving suspected prescription drug trafficking.

Such investigations can involve thousands of messages, photographs, voice recordings, financial transactions, and activity logs spread across several devices.

Examining this information manually can be time-consuming, especially when investigators need to establish relationships between individuals, transactions, and events.

SAGE attempts to address this problem by dividing analytical tasks among specialized AI agents.

Its architecture follows a master–worker model. Individual agents process different evidence types, including images, audio recordings, and behavioral records.

For example, one agent may transcribe voice messages while another extracts relevant information from photographs or identifies patterns in activity logs.

The results are combined into structured records that investigators can examine chronologically.

Another component, called the Unique Identifier Network, attempts to connect information relating to the same individuals or entities across different devices and databases.

This allows investigators to examine possible relationships between communications, financial transactions, and logistics records that might otherwise remain disconnected.

The researchers evaluated SAGE using two real-world health insurance fraud datasets and reported its deployment in a practical investigative project.

According to the published findings, the system processed more than 1.34 million forensic records and identified 11,200 evidence items.

However, these figures do not establish that every identified item was independently verified or successfully used in criminal proceedings.

The publicly available research summary also does not provide sufficient information to determine the system's false-positive rate.

This distinction is particularly important when artificial intelligence is used to interpret digital evidence.

LLM-based systems can misinterpret context, generate inaccurate summaries, or incorrectly associate information with individuals.

Investigators must therefore verify AI-generated findings against original evidence, preserve source records, and document the analytical process.

The researchers have made the project's source code and an illustrative dataset publicly available, potentially allowing other specialists to examine the approach.

The study demonstrates how AI-assisted evidence processing could support complex investigations, while highlighting the continuing importance of human verification, reproducibility, and evidentiary reliability.