Derivation and Validation of a Machine-Learning Decision Model for Incremental Hemodialysis: Emulating Expert Decision-Making

Background: No consensus on the strategy for implementing incremental hemodialysis has been reached. The present study aims to integrate clinical expertise with the power of machine learning to develop and validate a decision model for incremental decision, at the same time, it also makes a preliminary attempt for intelligent hemodialysis. Methods: Hemodialysis patients with RKF between April 2010 and November 2023 in Huashan hospital of Fudan university were included and divided into incremental and conventional group according to the initial dialysis frequency of lower than thrice-weekly or thrice-weekly. All available individual dialysis session records and corresponding periodic laboratory results across the follow-up period was utilized and partitioned into a training-testing set (incremental group) and an internal validation set (conventional group). Incremental event was defined as from once weekly to twice weekly, or twice weekly to conventional thrice-weekly. Positive samples were defined as hemodialysis treatment records obtained during the one-month period preceding an incremental event, matched with the closest biochemical test results. Four machine learning algorithms were employed to construct predictive models, and their performance was evaluated using AUC, sensitivity, and specificity. Results: A total of 175 ESRD patients (91 in incremental group, and 84 in conventional group) were enrolled, yielding 35,858 dialysis session records and 1,987 laboratory measurements. The mean age of the cohort was 61.1 ± 15.7 years; 38% of the patients were female; the initial urine output was 1.1 ± 0.6 L/day; and GFR was 5.2 ± 3.3 mL/min/1.73 m2. Volume overload was the primary reason for transitioning to more frequent dialysis (72.3%), with a median time to increment of 1.2 years. Among the 28 features evaluated, the Extra Trees model demonstrated the most optimal performance with an AUC of 0.99. In the internal validation set, the Extra Trees model also offered precise prediction of adverse outcomes for conventional dialysis patients. Conclusion: The development of an Extra Trees–based incremental hemodialysis decision model was undertaken, and its explainability allows clinicians to understand the key drivers behind each decision, thereby supporting personalized incremental dialysis regimens.