#2304 Innovative approaches to nephrology referrals: unlocking the potential of artificial intelligence

Abstract

Abstract Background and Aims The widespread influence of Artificial Intelligence (AI) is evident across various sectors, including the medical field. Hospitals and healthcare systems are increasingly embracing its capabilities in order to improve clinic workflow, to save time and simplify procedures with some degree of complexity. One of those tasks is the referral process to Nephrology appointments by the Primary Care (PC) physicians. As they may face challenges in determining which patients meet the referral criteria, it would be of great help to create a tool that could assist the process. Our work aimed to understand if AI can be useful in the referral to Nephrology appointments in Portugal's National Healthcare System. Method A cross-sectional study was preformed using appointment requisitions made by PC Physicians to a single center Nephrology Department throughout 2023. An AI bot, named RefNef, was developed, incorporating referral criteria from our center, the Portuguese National Healthcare Authority, and the Kidney Disease Improving Global Outcome (KDIGO) guidelines. Stripping personal data, those requisitions were then submitted to the created bot, and the answers were analyzed. Statistical analyses, including T-student, ANOVA, and Chi-squared tests, were employed to compare different groups. Results A total of 408 referrals were analyzed (52% of them males), with a mean age of 74 ± 15 years old. Mean creatinine at referral was 1.87 ± 0.72 mg/dl, with a mean estimated Glomerular Filtration Rate using the Crockford-Gault equation of 33.23 ± 15.12 mL/min. 5.6% of them had hematuria and 15.2% had proteinuria at referral. These four criteria were lacking in the referral information in 19.37%, 56.86%, 92.9% and 61.5% of the requests, respectively, but their absence did not appear to affect the nephrologist's or the bot's rate of acceptance. 86.3% of the referrals were approved by the nephrologists, and the bot accepted 97.1% of them. The rate of acceptance was not related (χ2 (1, N = 208) = 5.038, p = 0.082). The time between the referral and the appointment differed from the PC's requirements, the bot's perspective and the actual time the appointment was scheduled. In 68.6% of the times, the bot provided the user with some useful information, as management tips or other details that should be investigated. There was a need to ask for further information to the bot in 3.18% of the cases. Conclusion While not yet poised to function as a triage system, as the information wasn't consistent with the decision of the nephrologist, the AI tool demonstrated utility in potentially aiding PC physicians in decision-making by providing useful information. Additionally, it exhibited user-friendliness, as it answers with all the needed details at the first question. Further refinement of the AI tool could enhance its alignment with nephrologist decisions, contributing to improved efficiency in the referral process. It is to be noted that the study also highlights the frequent omission of key criteria in Nephrology referrals, emphasizing the need for educational reinforcement in this aspect.