2025 Poster Presentations
P092: THE EQUITY CHALLENGE: INVESTIGATING AI-BASED PRIOR AUTHORIZATION IN ENDOSCOPIC SINUS AND SKULL BASE SURGERY
Srinidhi Polkampally1; Akash Halagur2; Noel Ayoub3; 1Stanford School of Medicine; 2Geisel School of Medicine at Dartmouth; 3Division of Rhinology and Skull Base Surgery,Department of Otolaryngology-Head & Neck Surgery, Mass Eye and Ear
OBJECTIVES: Insurance companies are increasingly using artificial intelligence (AI) to automate administrative tasks, including claim and prior authorization decisions. There is concern that inherent bias in the large language models (LLMs) used for these decision-making strategies will produce unfair and unethical outcomes. This study investigates potential demographic biases in an AI-based simulation of insurance prior authorization decisions for benign and malignant endoscopic endonasal sinus and skull base surgeries.
METHODS: Using a novel Application Programming Interface (API) that virtually connected to OpenAI’s Generative Pre-Trained Transformer (GPT-4o), a simulated environment was developed. Each environment included a simulated insurance company and prior authorization specialist (PAR) tasked with approving prior authorization for only one patient among a subset of patients with varied demographic backgrounds. Three groups of patients were developed: 1) patients differing by age, race, and gender 2) patients with varied socioeconomic status and race, and 3) patients with unique substance use histories. Prior authorization decisions were made for four surgical cases: two endoscopic endonasal surgeries (resection of clival chondrosarcoma and pituitary macroadenoma) and two endoscopic sinus surgeries (resection of sinonasal squamous cell carcinoma and treatment of chronic rhinosinusitis). All patients had identical tumor staging, imaging findings, prognosis, likelihood of survival, and baseline health status. Each simulation ran 1000 times. Additionally, the GPT-4o simulations included questions to identify rationales for each decision (“Why did you choose this patient?") and self-assessment of potential biases ("Was your choice influenced by bias?). If the model detected bias, a new unbiased selection was requested.
RESULTS: Prior authorization decisions made by simulated PARs consistently demonstrated significant biases based on age, race, socioeconomic status, and substance use history. Most pairwise comparisons revealed statistical significance (p<0.05). For all four surgeries, young Hispanic males and females were most favored over other patients of varied age, gender, and race (all pairwise p<0.001), followed by young White males. Younger patients were generally preferred to older patients for across all four surgeries, with the rationale provided being to maximize the duration of benefit of the procedure. However, for elective sinusitis surgery, older black males and females were chosen over young Asian males and females (all pairwise p<0.001). When socioeconomic status was introduced into the patient profiles, black patients were most preferred, regardless of socioeconomic status, followed by Hispanic patients of low socioeconomic status, across all surgeries (all pairwise p<0.001). Among patients of varied substance use, for all four surgeries, the non-smoker, non-drinker, non-drug user was preferred consistently by simulated PARs for the reasons of optimizing outcomes, minimizing risks of peri/post-operative complications, and minimizing resource utilization and healthcare costs.
CONCLUSION: AI-based simulations of prior authorization decisions of insurance specialists reveal significant biases based on patient age, race, socioeconomic status, and substance use history. The potential for demographic biases in AI-assisted healthcare decision-making must be recognized and mitigated to ensure equitable access to elective and life-saving medical procedures across diverse patient populations.