In this retrospective cohort study, we comprehensively assessed the factors associated with successful discontinuation of CRRT in critically ill adults with AKI. Our primary findings indicated that UO before stopping CRRT and non-renal SOFA score were predictors of successful CRRT discontinuation. Also, bicarbonate, SBP, and BUN were predictive of successful CRRT discontinuation. Furthermore, our results suggested that patients who failed CRRT discontinuation had poorer prognosis. More importantly, we assessed the performance of different models for successful short-term discontinuation of CRRT in patients with AKI by modeling based on various evaluation metrics. Our results showed that models based on ensemble methods (RF, XGBoost) exhibited relatively superior performance compared to other prediction algorithms.

There is currently no consensus on the cessation of CRRT treatment for AKI patients. In clinical practice, clinicians typically make decisions based on individual patient circumstances. The timing of CRRT discontinuation for AKI patients depends on various factors [24, 25]. In our study, UO before CRRT discontinuation was identified as one of the most important influencing factors for successful discontinuation. Indeed, UO is one of the most commonly studied variables in research related to CRRT discontinuation. In a recent systematic review and meta-analysis by Li et al., chronic kidney disease, duration of CRRT, and UO at the time of CRRT discontinuation were reported as predictive factors for short-term successful cessation of CRRT [26]. Furthermore, another meta-analysis by Katulka et al. concluded that UO before cessation of RRT is the most commonly described and reliable predictor of successful liberation, with a sensitivity of 66.2% and specificity of 73.6% [27]. Additionally, other studies have reported UO (before initiation of CRRT, during CRRT, and after CRRT discontinuation) as positive predictors of successful CRRT discontinuation [13,14,15, 28, 29]. Mechanistically, higher UO may exert renal protective effects by maintaining adequate renal perfusion pressure. However, there is currently a lack of clear consensus on the optimal UO threshold for predicting successful CRRT discontinuation, which may be influenced by factors such as different time points and the use of diuretics. Further large-scale prospective studies are still needed to provide more evidence.

The significance and application of non-renal SOFA lie in assessing the multi-organ function of patients, aiding clinicians in understanding the severity of other systemic diseases in patients. Our findings suggested that non-renal SOFA was also associated with successful discontinuation of CRRT, suggesting that patients with milder disease severity had a greater chance of being able to undergo successful CRRT discontinuation, which was consistent with the results of several previous studies [30, 31]. Improvement in severity of illness scores such as non-renal SOFA could be considered in the future as a potential indicator of the timing of attempted withdrawal.

In our analysis, BUN and SBP before stopping CRRT, as well as bicarbonate before CRRT initiation were also strongly associated with successful discontinuation of CRRT. BUN is a crucial indicator reflecting renal function. In a recent study, Aniort et al. reported that high daily urea excretion was the optimal marker for renal recovery during IHD withdrawal [32]. The retrospective study by Valle et al. also suggested that serum BUN may play a significant role in determining the timing of RRT discontinuation [31]. Klouche et al. proposed that BUN > 40 mmol/L could be considered as one of the criteria for restarting RRT after discontinuation [33]. Additionally, as BUN are influenced by various factors such as muscle mass, liver function, gender, and race [34, 35], further research is needed to confirm the robustness of this result and determine specific threshold values. Elevated bicarbonate levels may be indicative of compensatory mechanisms against severe metabolic acidosis. In such cases, the body may experience severe acidosis due to impaired renal function, other metabolic abnormalities, or overall systemic deterioration. Our findings align with the broad objectives outlined in the KDIGO guidelines, emphasizing the importance of managing fluid balance, acid-base status, kidney function, and the severity of illness in patients receiving CRRT. Additionally, we observed that most of the patients included in the analysis had concurrent sepsis. We further examined the impact of the primary source of infection (e.g., bloodstream, respiratory tract, urinary tract, abdomen, etc.) on the prediction performance. Although the primary source of infection was an important characteristic of sepsis, the results showed that its impact on predicting successful CRRT discontinuation in the septic patients we studied was not significant (Tables S4, S5, S6, and S7).

In this study, although KNN achieved the optimal AUC when only five independent predictors were considered, the other evaluation metrics did not achieve a significant advantage, and the performance decreased when all available variables were included. Ensemble-based algorithms (RF, XGBoost) showed satisfactory overall performance in both scenarios. The discontinuation of CRRT in AKI patients is influenced by multiple complex variables, including vital signs, laboratory parameters, and urine output. Additionally, nonlinear relationships often exist between clinical data and outcomes. Compared to single models, ensemble models are better at handling complex relationships and non-linear features in the data, reducing overfitting, and performing well when dealing with high-dimensional data and noise. Overall, decrease in the average performance of all models occurred when all variables were included. This phenomenon may be attributed to the presence of redundant features, noise, and multicollinearity issues in the data when considering all variables. These factors can adversely affect the performance of the models. Additionally, simple models such as LR, DT, and KNN are often used as base algorithms or weak learners, and their performance is significantly affected in complex data scenarios. In contrast, RF and XGBoost algorithms belong to ensemble methods, which have improved upon DT, making them more robust in performance. In conclusion, we concluded that employing multivariable analysis to select features was necessary. On one hand, it helps eliminate redundant features, reduces the impact of noise, and addresses multicollinearity issues; on the other hand, fewer features contribute to enhancing the efficiency of model training. These aspects positively impact most machine learning models. Moreover, a more concise set of features allows clinicians to quickly identify which factors are most important for the success of CRRT discontinuation, thereby enhancing the model’s practicality and applicability in clinical practice.

Building upon previous research, our study reaffirmed that machine learning models outperform traditional LR models in predicting CRRT discontinuation. Liang et al.’s single-center retrospective study defined successful discontinuation of RRT as not requiring RRT for one week continuously. Compared to traditional LR models, the RF model demonstrated superior performance with an AUC of 0.95 [36]. Additionally, Pattharanitima conducted analysis on AKI patients in MIMIC-III, defining successful discontinuation as not requiring RRT and being alive for one week before discharge. They evaluated seven methods, including LR, RF, support vector machine, AdaBoost, XGBoost, Multilayer Perceptron (MLP), and Long Short-Term Memory with MLP (MLP + LSTM). MLP + LSTM achieved the highest AUC of 0.70, whereas LR had an AUC of only 0.57 [37]. Although there were methodological differences between these studies and ours, they collectively provided strong evidence for the significant value of machine learning in predicting outcomes in healthcare settings, particularly among patients undergoing CRRT.

This study was subject to several limitations. First, this study adopted a retrospective design, which may limit the generalizability of the results to specific populations. Additionally, while the MIMIC-IV database is broadly representative, it is derived from a single center, which may introduce selection bias and potentially limit the external validity of the study’s findings. The limitations of a retrospective cohort design include uncontrolled confounding factors, and information bias due to the retrospective nature of data collection. Second, long-term outcomes were not analyzed. Third, new biomarkers were not included in our analysis due to the limitations of the retrospective database. Finally, machine learning models may be prone to overfitting when handling complex data. To reduce this risk, we applied five-fold cross-validation to improve the model’s generalization ability. However, despite these measures, the model’s performance on other datasets still requires further validation to ensure its robustness and broad applicability. However, the strength of this study lied in its comprehensive assessment of factors influencing CRRT discontinuation, including multiple physiological and laboratory parameters, UO, fluid balance, and other therapeutic interventions both before CRRT initiation and CRRT discontinuation.