Efficacy of Artificial Intelligence for Gatekeeping in Referrals to Specialized Care

In Rio Grande do Sul, Brazil, the demand for specialty care referrals has increased sharply with the adoption of the electronic regulatory system, especially in rural areas. In 2023 alone, over 79,000 referrals were submitted monthly, totaling 1.7 million annual gatekeeping decisions. Due to workforce limitations, nearly 70% of referrals are authorized automatically, often without clinical validation. This leads to delays for high-risk patients, unnecessary specialist visits, and a growing backlog, currently over 172,000 pending referrals. To address this, an AI algorithm was developed to triage referrals based on urgency and appropriateness. The investigators propose a prospective controlled study with randomized implementation of the AI tool across selected specialty queues in the electronic referral system. The population will consist of referrals from specialties waitlists from municipalities in Rio Grande do Sul. Specialties to be included will be selected by the State Health Department prospectively according to gatekeeping needs. The intervention will be an AI-based triage algorithm. The control will be a standard gatekeeping process. The primary outcome is the proportion of referrals with a final decision (authorized or redirected to primary care) within six months; secondary outcomes include time to decision and appointment, system-level performance metrics. Referrals will be randomly assigned to algorithmic or human gatekeeping with a 1:1 ratio. The algorithm classifies referrals into two groups: not authorized (pending more data or teleconsultation), authorized. Authorization cases are further divided into routine and high-risk referrals to help the manage demand. Each AI prediction provides a probability from 0 to 1 of authorization (or deferring). The implementation threshold is set at 0.8; cases below this level will be classified as low confidence for decision and will not be included. According to the State Health Department's decisions, several referral lines are expected to be selected for the intervention. A sample size 934 (467 per arm) for each included specialty was calculated to detect a 1.2 relative risk for the primary outcome with 90% power and 5% significance.