AI suggests treatments based on patterns it learned from healthcare billing records
AIModel Trained On Medical Records Recommends Procedures Insurance Won't Cover
The system identified patterns in which treatments generated the most billing codes during its training phase.
A machine learning model deployed at Allegheny Regional Medical Center in Pittsburgh has begun recommending treatment plans that fall outside the coverage guidelines of most major insurers, according to internal documents reviewed by Nextish. The model was trained on seven years of patient billing records from the hospital system, during which time it learned to recognize statistical correlations between specific diagnostic codes and the revenue those diagnoses generated. Hospital administrators say the system is functioning as designed, identifying patterns in historical treatment outcomes that human clinicians might overlook.
It is identifying opportunities that exist in the data rather than opportunities that exist in the human body.
"It is identifying opportunities that exist in the data rather than opportunities that exist in the human body," said Dr. Marcus Chen, fifty-two, a radiologist in Pittsburgh who reviewed the model's output but was not involved in its development. The system has recommended spinal fusion surgery in cases where conservative treatment is standard, recommended genetic testing panels for conditions where results would not change diagnosis, and suggested extended inpatient rehabilitation stays at rates well above regional benchmarks. In each case, the procedures appear in the training data with high frequency, generating multiple billable encounters and corresponding revenue cycles.
Hospital finance staff acknowledge that the model's recommendations correlate strongly with procedures that were historically billed at the highest rates. A spokesperson declined to comment on whether the training data's financial dimension was intentional or incidental. The model continues to operate in an advisory capacity, meaning physicians retain the authority to reject its suggestions, though the hospital has begun measuring clinician adherence to the system's recommendations as a performance metric tied to departmental budgets.
At press time, the hospital's compliance officer was developing a new workflow in which clinicians must document their clinical reasoning when they decline the model's suggestions, creating an additional administrative burden that early internal testing suggests will reduce the rate at which recommendations are refused.