#2782 Clinician usage and perception of an AI-based model predicting peritoneal dialysis dropouts

Abstract

Abstract Background and Aims A machine learning (ML) model was developed to identify patients likely to leave peritoneal dialysis (PD). Weekly reports containing model predictions and relevant information were provided to care teams for clinical decision support. Prior to piloting an updated model in a subset of clinics receiving the report, a survey was conducted to gather feedback on current usage and perceptions of the original model and report. The primary goal of the survey was to collect a baseline for tracking changes in usage after switching to the new model and dashboard. Method 177 clinics were identified to pilot the new model, which would be delivered via a dashboard instead of the weekly emailed report. Home therapy clinic managers (HTCMs) overseeing pilot clinics completed a survey and reviewed training materials with their Interdisciplinary Team (IDT). We received 102 responses covering 96 unique clinics (54% of clinics). After excluding duplicate responses, we were left with valid responses reflecting report use in 91 clinics. Questions covered frequency of use, roles reviewing report, how the report is used, report printing, and agreement with each of five statements assessing the ease of accessibility of the report, understandability of the report, usefulness of the information provided, provision of better patient care, and report satisfaction. Results Of the 91 valid responses, 49.5% used the report 1–3 times per month, 44% used it weekly, and 5.5% used it multiple times per week. In all clinics reporting usage (n = 90), a nurse and/or HTCM reviewed reports at a minimum. Social workers reviewed the report in 42.2% (n = 38) of the clinics, dieticians in 43.3% (n = 39), patient care technicians in 13.3% (n = 12), kidney care advocates in 2.2% (n = 2), physicians in 35.6% (n = 32), area team leads in 13.3% (n = 12), directors of operations in 10% (n = 9), patients in 12.2% (n = 11), and other roles indicated in 3.3% (n = 3). Reports were reviewed during IDT meetings, Quality meetings to review programs at a high level, and physician rounds in 32.2% of clinics (n = 29) each. In clinics reporting usage (n = 90), agreement with statements was endorsed most of the time for all items (Fig. 1). Agreement was endorsed by 76.7% of clinics (n = 69) for ease of access, 82.2% (n = 74) for ease of understanding, 82.2% (n = 74) for the provision of useful information, 72.2% (n = 65) for bettering patient care, and 73.3% (n = 66) for satisfaction with the report. The statement with the most disagreement was regarding the ease of report access; 15.6% (n=14) strongly disagreed or disagreed and 7.8% (n = 7) were neutral. Within clinics using the report 1 or more times per week (49.5%, n = 45), only 4.4% (n = 2) reported disagreement with the statement that the report was easy to access. Clinics reporting less frequent usage of 1-3 times per month (49.5%, n = 45) were much more likely to endorse disagreement regarding ease of access (26.7%, n = 12) (Fig. 2). In clinics that reviewed the report with patients (12.1%, n = 11), reports were also reviewed by physicians, and none of the clinics endorsed disagreement with any of the five statements (Fig. 3). Conclusion Clinician survey responses assessing use and perceptions of a report displaying ML predictions and related information surrounding patients leaving PD were overwhelmingly positive. A relationship was found between frequency of usage and reported ease of access of this weekly emailed report, where those expressing dissatisfaction with the ease of access used the report less frequently. This emphasizes the critical role of workflow integration and ease of use in driving adoption of AI tools. Additionally, clinics reviewing the reports with patients directly also always endorsed reviewing with physicians; these clinics did not express disagreement with any of the statements. This likely reflects the level of buy-in and understanding of the report needed to bring to patients directly.

Ähnliche Arbeiten