23 - Potential of Large Language Model to Facilitate and Optimize IR Billing in Real-time: A Proof of Concept Analysis using human coders.

Billing for interventional radiology procedures is a complicated and ever-changing task. It is an important facet to sustaining an interventional practice, which places great emphasis on optimizing the number of billable codes. In an ongoing effort to design a machine learning tool that suggests CPT codes from dictations that coders can reference in real time to optimize billable codes, we conducted an outside review to compare in-house coding performance to a gold standard outside consulting group, which would subsequently be used as additional training data for our model.

Materials and Methods:

For the purpose of this study de-identified dictations were randomly selected from the Yale New Haven Hospital Picture Archiving and Communication System from the following procedural categories for review: Pulmonary Thrombectomy (N=10; male (M): 6 , female (F): 4, average age: 54.6, SD (standard deviation) = 9.7), Microwave Ablation (N=10; M: 9, F: 1; average age 69.5, SD: 10.4), TACE (N=11; M: 9, F: 2; average age: 67.2, SD= 5.2), Kyphoplasty (N=3; M , F: 3, average age: 75, SD = 6.2), Gastrointestinal Hemorrhage (N=10; M: 5, F: 5; average age= 63.7, SD= 10.0), TIPS (N=11; M: 3, F: 8, average= 51.2, SD= 10.1), Deep Venous Thrombosis Thrombectomy (N=10; M: 7, F: 3, average age 56.2, SD = 18.5), and Uterine Artery Embolization (N=10; F: 10, average age 44; SD= 5.6). The number of cases reviewed per type procedure ranged from 3-11 (average: 9.44). The outside consulted group was blinded to the codes selected by the in-house coders and were then asked to re-code the provided dictations. Often, multiple coders reviewed each report in order to ensure comprehensive and accurate coding. T-test were performed to determine statistical significance on the in-house results between the in-house and outside consultant group.

Results:

On average the outside group coded for an average of 1.7 (SD .99) more codes than the in-house team. Across all procedures this translated to a statistically significant net additional 8.40 RVU per case (SD: 5.19, t(148) = 8.74, p = 0.0001). Advanced vascular cases made up the largest discrepancy in RVUs. Pulmonary thrombectomy cases produced the greatest increase in billable RVU (average 14.37; SD 5.46) meaning the outside billing group were able to generate an average of 14.37 more RVU per case than the in-house group. TIPS and UFE were second and third to PE with 13.56 (SD 8.50) and 11.01 (SD 2.06) RVUs. The outside consulting group generated a statistically significant increase in RVU for Pulmonary Thrombectomy (t(18) = 8.4, p < 0.0001), TACE (t(20) = 11.7, p < 0.0001), DVT Thrombectomy (t(18) = 2.9, p = 0.01), TIPS (t(20) = 3.3, p = 0.004), UFE (t(18) = 4.3, p = 0.0004). Across all procedures the most frequently added code by the outside group involved catheterization (40.5%).

Conclusion:

In a comparative performance review of in-house billing codes versus a gold standard consulting group the results demonstrated that a consulting group outperformed the in-house team, by an average of 8.40 RVUs across a number of procedures. In a large IR practice, this discrepancy could easily amount to hundreds of thousands of dollars of additional revenue, in professional and facility fees. The results of this study demonstrated that even large tertiary centers may not be coding reports optimally. This study has additional application to the development of algorithms that can suggest required verbiage to be included in the report to ensure proper documentation and then assist coders to generate optimized reports to bill for the most RVU possible.