Acute kidney injury (AKI) is associated with increased worsening long-term outcomes and direct medical costs (1–4). The recovery rate at hospital discharge or within 90 days from AKI ranges from 39.3 to 76.2% (5–7) and the recovery rate (free from dialysis therapy) of patients with AKI requiring dialysis at 90 and 365 days after dialysis initiation ranges from 47.7 to 56.6% (8). Timing of kidney function recovery after an AKI episode is associated with an increased risk of onset and long-term progression of chronic kidney disease and can affect survival rates (6–10). There is a need to improve the risk prediction for AKI non-recovery, and the predictive model may serve as a risk stratification tool to assist in the development of an appropriate post-AKI care plan at a time of hospital discharge.

Risk predictive models of AKI non-recovery using traditional statistical methods have achieved a C-statistic or area under the receiver operating characteristic curve (AUROC) of 0.61–0.76 in the validation models (11–13). The variance in AKI non-recovery predictive performance may be associated with heterogeneous AKI cohort, the definition of AKI recovery, and variable length of follow-up. A recent study suggested that clinical experts have selected features that can successfully predict AKI non-recovery within 7 days with an AUROC of 0.879 in the multi-center hospitalized cohort (14). These models tended to focus on patients with AKI undergoing dialysis or with critical illnesses. Machine learning approach using routinely collected health data in practice has demonstrated a promise for predicting AKI development in outpatient (15) or inpatient settings (16, 17). Currently, there is not a single validated machine learning model to identify the risk predictors of AKI non-recovery to establish supporting strategies for post-discharge AKI care. This study aimed to develop and validate a machine learning model for predicting AKI non-recovery at hospital discharge for the development of targeted interventions to prevent long-term adverse kidney outcomes.

The present study results demonstrate an explainable machine learning model for predicting kidney function non-recovery at hospital discharge in patients receiving inpatient AKI care from routinely collected health record data in clinical practice. Timing of AKI occurrence, receipt of dialysis, and SCr measurements; comorbidity of cancer; and pre-hospitalization baseline LPC% and K levels, which are relevant for predicting AKI non-recovery, did not return to a level ≥50% pre-hospitalization baseline SCr level. The SHAP and LIME methods enhance the interpretability of machine learning models and estimate the positive and negative contributions of each feature to the model prediction.

A previous study has identified four selected predictors (estimated glomerular filtration rate ≥ 30 ml/min/1.73 m², preadmission hemoglobin level, platelet level, and diabetes comorbidity) of AKI recovery using stepwise logistic regression in patients with AKI undergoing dialysis therapy, however, failed to achieve acceptable discrimination performance (C-statistic, 0.645) 0.1 (3). Another study identified age, sex, SCr elevation, and urinary output as predictive factors for AKI recovery (without kidney replacement therapy within 5 days and SCr value <1.5 times the pre-ICU admission value) in patients requiring critical care, with an AUROC of 0.73 (65.4–80.8%) 0.1 (1). However, these studies have generally included patients requiring dialysis therapy in ICUs and have been limited to severe forms of AKI. To the best of our knowledge, this is the first study to develop an AKI recovery predictive model that uses machine learning algorithms for patients with AKI who were and were not admitted to ICUs. Importantly, this study included baseline SCr values, which are commonly ignored in previous studies assessing the risk factor for AKI recovery (25, 26). These previous studies have well-recognized limitations in the evaluation of AKI recovery. Based on the same data, our results showed that the XGBoost algorithm model achieved a good predictive performance with the 20 selected features and was reproducible in the validation cohort.

Another major contribution of this study includes its capability to derive reliable important features and validate an accurate AKI non-recovery predictive model based on a large AKI cohort with broad characteristic diversity. Given the poor outcomes among AKI survivors even in less severe AKI cases (3, 7), these study results of AKI non-recovery predictors can facilitate risk assessment for identifying at-risk patients for poor kidney recovery before hospital discharge and promote nephrology referrals for post-discharge care for patients who may benefit most from such interventions (27, 28).

Although the identified features should not be interpreted as etiologic factors or causal relationships, several temporal associations are noteworthy. These findings indicate that HA_AKI, high index SCr, and receipt of dialysis during the hospitalization were the three strongest predictors for non-recovery, a result consistent with those of prior studies (6, 10, 29). Because dialysis (intermittent and continuous hemodialysis) is a complementary therapy for AKI (30), it may be correlated with the severity of AKI. The fact that timing of AKI occurrence (AKI presence at admission [index_AKI] or during hospitalization [HA_AKI]) was also important to AKI non-recovery prediction, indicating that kidney function returned to the pre-hospitalization baseline level was not considered for hospital discharge in current practice. Further research is warranted to focus on long-term adverse outcomes among patients with AKI non-recovery who had a small improvement (e.g., recovery <50% baseline SCr) or worsening kidney function (e.g., increase > peak SCr) at hospital discharge.

The SHAP results showed that the baseline SCr value had bidirectional effects on AKI outcome prediction (Figure 2B). It is possible that patients with a high baseline SCr value (proxy of poor kidney function) or less severe episode of AKI at admission may contribute less to AKI non-recovery prediction (negative SHAP value in Figure 2B). In contrast, patients with a low baseline SCr value (blue color of feature value) may experience a severe episode of AKI at admission or during hospitalization, resulting in positive and significant effects on AKI non-recovery prediction (positive SHAP value in Figure 2B).

Cancer is an important factor for predicting AKI non-recovery. These results are consistent that cancer patients with AKI non-recovery within 7 days had a greater risk of end-stage kidney disease or death (14). We hypothesized that existing cancer influencing AKI non-recovery may reflect higher baseline risk of AKI progression because patients with cancer are at high risk for infections, sepsis, tumor lysis syndrome, and chemotherapy-associated toxicities (31, 32).

Although the relevance is low, the baseline lymphocyte count and low-density lipoprotein cholesterol level increased, positively predicting AKI non-recovery. Lymphocytes are associated with the systemic inflammatory response; the proportion of baseline LPC and its importance to AKI non-recovery prediction are not linear. Moreover, its relevance can be associated with hospital-acquired AKI (Supplementary Figure 2A). These laboratory results suggested that inflammatory response at baseline could mediate the recovery phase of AKI (33). The SHAP dependence plot (Supplementary Figure 2B) shows that baseline K value > 4 mEq/L increased the prediction of AKI non-recovery (Supplementary Figure 2B). In contrast, patients with K > 5 mEq/L and high SCr value at admission had a negative association with AKI non-recovery prediction, which could be due to further medical interventions during the inpatient stay.

Despite the increasing interest in the relative predictive performance of different machine learning approaches, the results of this study are significant as they contribute to a better understanding of the effect (positive or negative correlation) of each feature of the predictive AKI non-recovery outcomes and a possible explanation of model's misclassification in the SHAP and LIME procedures. For instance, the LIME results potentially provided an in-depth explanation for why the scenario B patient (ID 2388), having a similar baseline SCr level with the scenario A patient (ID 2231 with < 1.0 mg/dl), was misclassified as AKI non-recovery (the scenario B patient recovered) in the final predictive model (Supplementary Figures 3A,B). All SHAP and LIME results suggested that to further perform an external validation in different populations, a study investigating the association between time-varying feature values (e.g., SCr trajectory) and mechanisms of AKI non-recovery should be conducted in the future.

This study has some limitations. First, only routinely available clinical data in the electronic health records were used to identify clinical phenotypes. Other data, such as urine output, fluid status, inpatient medication use, or pathogen variables during the hospitalization in the derivation model that could change classification, were not assessed. Additionally, considering the changes to the time window for AKI recovery, the measurable candidate feature inclusions may affect the selected features and outcome predictive performance. However, this does not undermine the current findings but rather highlights the need for further studies to better understand the complex mechanisms correlated with AKI non-recovery. Second, missing data were common for some laboratory results included in the derivation models. Although different approaches of data imputation were examined and the best calibration method was used to preserve at least 70% sensitivity and specificity, the imputed feature's value could bias classification. Finally, the characteristics of patient and clinical data were derived initially from a single integrated health system in Taiwan. Although feature distribution and frequency were found to be generalizable in the derivation and temporal validation cohorts, further external cohort validation is necessary, especially using data from different healthcare systems and more recent clinical data in academic center cohorts.

In this study, XGBoost demonstrated good discrimination performance for AKI non-recovery at hospital discharge. In the temporal validated model, 20 important patients and clinical characteristics reflect the complexity of factors affecting the non-recovery kidney function. The eight characteristics supporting AKI non-recovery prediction should be considered risk predictors of adverse AKI outcome in further risk profiling to drive the development of post-AKI care strategy for inpatients with AKI before hospital discharge.

The original contributions presented in the study are included in the article/Supplementary Material, further inquiries can be directed to the corresponding author/s.

This study was approved by the Institutional Review Board of Chang Gung Memorial Foundation at Taipei, Taiwan. Written informed consent for participation was not required for this study in accordance with the national legislation and the institutional requirements.

C-NH and Y-LT were involved in concept and design. C-NH and C-LL were involved in the drafting of the manuscript. C-LL and Y-CL performed statistical analysis. C-NH obtained funding. All the authors were involved in critical revision of the manuscript for important intellectual content.

This study was funded by Kaohsiung Chang Gung Memorial Hospital (CMRPG8K0101 and CORPG8L0041 to C-NH). The funders had no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; and decision to submit the manuscript for publication.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.