Traditional statistical analysis has been instrumental in identifying risk factors for SA-AKI across a range of comorbidities, infections, medications, and other determinants [4,34–42]. Key comorbid conditions found to significantly predict SA-AKI include hypertension, diabetes mellitus, chronic kidney disease, cardiovascular disease, liver disease, and coronary artery disease [4]. A pooled analysis of 47 observational studies with 55,911 sepsis patients revealed that hypertension, diabetes, and chronic kidney disease increased the odds of AKI, with odds ratios (ORs) of 1.43, 1.59, and 3.49, respectively (Fig. 1) [4]. Furthermore, cardiovascular, liver, and coronary artery diseases emerged as risk factors, with ORs of 1.31, 1.68, and 1.27. Regarding infection sources, pulmonary, abdominal, and undetermined infections were significant SA-AKI predictors, with ORs of 0.77, 1.44, and 2.01, respectively. Medications like vasopressors, angiotensin‐converting enzyme inhibitors/angiotensin receptor blockers, and diuretics also correlated with heightened SA-AKI risk, having ORs of 3.15, 1.61, and 1.40. Other notable risk factors included male sex (OR, 1.22), positive blood culture (OR, 1.60), smoking history (OR, 1.60), septic shock (OR, 1.40), Gram-negative bacteria (OR, 2.19), organ transplantation (OR, 1.96), and mechanical ventilation need (OR, 1.64) [4].

While these findings provide valuable insights, traditional statistical approaches have limitations. Despite identifying these predictors, substantial heterogeneity existed across studies, suggesting potential inconsistencies in outcomes depending on the context of SA-AKI [4].

Traditional statistical approaches to predicting SA-AKI through various studies have been used [34–42]. These studies (Table 1) utilized well-established methods such as logistic regression analysis, least absolute shrinkage and selection operator (LASSO) regression for variable selection, and calibration plots [34–42]. These models identified key predictors like diabetes, chronic kidney disease, cardiovascular disease, and specific lab values such as creatinine and procalcitonin. Performance was evaluated using metrics like area under the receiver operating characteristic curve (AUC), sensitivity, and specificity [42].

For instance, Fan et al. [34] developed an SA-AKI prediction model using logistic regression and LASSO, achieving c-statistics of 0.711 and 0.705 in training and validation cohorts. Xin et al. [42] conducted a retrospective cohort study among elderly sepsis patients and achieved an AUC of 0.852 in the training and 0.858 in the validation cohort using logistic regression. Xie et al.’s prospective study [35] of sepsis patients in the ICU resulted in an impressive AUC of 0.9862. In addition, Zhou et al. [36] utilized a randomized clinical trial approach with 16 predictors and achieved an AUC of 0.857 in the validation cohort. While these examples demonstrate traditional statistical methods can effectively predict SA-AKI, recognizing high-risk patients early to guide treatment, these traditional prediction models for SA-AKI come with several limitations (Fig. 2). These limitations include 1) sensitivity to outliers, which may skew results/predictions. Outliers in medical data could be errors or critical rare events that should not be ignored. 2) Multicollinearity among predictor variables complicates the interpretation of individual predictors’ effects on the outcome. 3) Temporal dynamics, as medical time-series data may not meet assumptions of independent and identically distributed points, impacting predictive accuracy. 4) High-dimensional medical data that traditional models can struggle to handle effectively, limiting the identification of complex relationships and predictive capabilities. 5) Calibration, requiring robust procedures to ensure predicted probabilities align closely with observed outcomes, avoiding suboptimal clinical decisions. 6) Potential human bias in feature selection, as choices rely on existing knowledge and practitioner input, possibly introducing limitations.

In essence, while significant, traditional SA-AKI prediction models have inherent challenges around outliers, multicollinearity, temporal dynamics, high-dimensional data, calibration, and bias in feature selection. Continued model updates and refinements alongside technological and research advancements remain important.

Supervised learning is a type of ML where the algorithm is trained on labeled data, meaning the input data is paired with corresponding output labels [54]. The primary goal is to learn a mapping function that can accurately predict the output labels for new, unseen data. In the context of SA-AKI, supervised learning can be utilized to predict the risk of AKI in sepsis patients [16,23–31,55]. Researchers can gather historical patient data, including clinical parameters such as vital signs, laboratory results, and patient demographics, as well as information about whether AKI developed during their hospital stay. This data is then used to train a supervised learning model, such as random forest, XGBoost, or artificial neural networks, to predict the likelihood of AKI in sepsis patients [16,23–31]. An example application of supervised learning in SA-AKI is the development of a predictive model that uses patient data from the ICU to identify individuals at high risk of developing AKI as a result of sepsis. The model can provide real-time risk scores, allowing clinicians to intervene early with appropriate interventions, such as fluid management or medication adjustments, to mitigate the risk of AKI [16,23–31].

Supervised learning has emerged as a pivotal tool in predicting outcomes and characteristics related to SA-AKI, according to published studies (Table 2) [16,23–31]. Researchers have successfully employed ML models to predict the onset of S-AKI, differentiate between persistent and transient AKI, and anticipate in-hospital mortality and acute kidney disease (AKD) occurrence. Notably, models like XGBoost and recurrent neural network (RNN)-long short-term memory (LSTM) were recurrently highlighted for their exceptional performance [16,23–31]. These algorithms often surpassed traditional risk scores such as SOFA (Sequential Organ Failure Assessment) and SAPS II (Simplified Acute Physiology Score II), achieving commendable discrimination metrics [16,23–31]. For instance, in predicting AKI risk in septic patients, an XGBoost model was found to outperform conventional scoring systems, emphasizing the model’s utility in pinpointing high-risk patients for proactive interventions [30]. Furthermore, in another instance, an RNN-LSTM model showcased exemplary predictive ability, with a remarkable AUC of 1.0 (Fig. 5), emphasizing its potential in guiding early AKD interventions [31]. Moreover, the application of ML was not limited to SA-AKI. Zhou et al. [30] explored sepsis-associated acute respiratory distress syndrome patients, aiming to predict AKI development within a short timeframe post-ICU admission. Here, the XGBoost model once again stood out, indicating its versatility across related medical conditions. A review study by Yu et al. [16] further underlined the importance of ML by summarizing various models tailored for predicting AKI across diverse patient groups and environments. The compiled results suggested that while ML models displayed a range from moderate to superb discrimination for AKI events, there remains room for further refinement and evaluation in real-world clinical settings [16,23–31].

In reflecting upon the strengths of these ML-centered studies, it becomes evident that such methodologies offer several advantages over traditional statistical analysis approaches. ML models, given their ability to handle large data sets and complex interactions, tend to produce more accurate, robust, and generalizable findings [54]. These models are especially skilled at uncovering nonlinear relationships, thereby providing a more sophisticated understanding of the data’s underlying patterns. Moreover, the ability of ML models to surpass traditional risk scores highlights their transformative potential in patient care. This is particularly relevant for early detection and intervention strategies, which can significantly improve patient outcomes.

Unsupervised learning involves training ML algorithms on unlabeled data to discover patterns, structures, or relationships within the data [56,57]. Clustering is a notable technique in unsupervised learning, where algorithms categorize similar data points into groups. In the specialized context of SA-AKI, unsupervised learning provides invaluable insights by delineating distinct subgroups of sepsis patients, each characterized by unique clinical profiles and outcomes. For instance, an unsupervised learning algorithm like k-means clustering can be employed to analyze clinical data from sepsis patients, including vital signs, laboratory results, and comorbidity information. This can reveal different clusters of patients with similar clinical characteristics. Clinicians can then evaluate whether these specific patient clusters have different risks of developing AKI, thereby facilitating the formulation of more personalized treatment plans.

Recently, Lai et al. [32] conducted a prospective observational cohort study of 999 critically ill patients with dialysis-requiring SA-AKI admitted to surgical ICUs in Taiwan between 2009 and 2018. The mean age was 63.9 years, and 71.5% were male. The authors performed unsupervised consensus clustering based on 23 clinical variables upon initializing renal replacement therapy to identify distinct sub-phenotypes (Fig. 6). Three sub-phenotypes that were identified included 1) cluster 1 (n = 352) with favorable baseline conditions but greatest acute illness severity, 2) cluster 2 (n = 396) with intermediate features, and 3) cluster 3 (n = 251) with worst baseline conditions but lowest acute illness severity. Cluster 1 had the highest mortality rate (73.9%) and lowest probability of being dialysis-free at 90 days. Cluster membership and predialysis serum lactate ≥3.3 mmol/L were independent predictors of mortality and dialysis dependence. A clinical prediction model using 11 variables accurately identified cluster 1 as a high-risk sub-phenotype (AUC, 0.99). When applied to an external validation cohort of 898 SA-AKI patients, the model identified a high-risk subgroup with increased mortality. This study demonstrates that ML approaches can identify clinically relevant sub-phenotypes and predictors in heterogeneous syndromes like SA-AKI.

Reinforcement learning is a type of ML where an agent learns to make decisions by interacting with an environment. The agent receives feedback in the form of rewards or penalties based on its actions, and its objective is to maximize cumulative rewards over time. In the context of SA-AKI, reinforcement learning can be used to optimize treatment strategies for sepsis patients to minimize the risk of AKI. An example application of reinforcement learning in SA-AKI is developing a treatment recommendation system for sepsis patients in the ICU (Fig. 7). The reinforcement learning agent can learn from historical patient data and clinical guidelines to recommend actions such as fluid administration, antibiotic choices, and vasopressor usage. The agent continually adapts its recommendations based on patient responses and outcomes, aiming to optimize patient care and reduce the incidence of AKI.

Reinforcement learning has been studied with promising findings for potential utilization among critically ill patients with sepsis [58–64]. From optimizing fluid resuscitation strategies to determining when and which antibiotics should be administered, RL agents have showcased their capability to refine treatment strategies based on patient feedback and outcomes continuously. Moreover, the technology presents the appealing possibility of integrating extensive datasets, facilitating a more comprehensive patient management approach that considers a multitude of variables [58–64]. Notwithstanding these advancements, there remains a notable deficiency in data concerning the specific employment of reinforcement learning for SA-AKI, underscoring the need for dedicated future investigations.

In medical research using ML models, protecting patient data is critical. This encompasses the challenges of ensuring that AI tools neither inadvertently expose nor misuse this confidential information. Additionally, the secure transfer of such data, which frequently traverses between various databases, servers, and institutions, demands the implementation of robust and encrypted protocols to thwart potential breaches. Concurrently, the ethical imperative of obtaining informed consent cannot be understated, emphasizing the necessity for patients to be comprehensively apprised of the utilization and repercussions associated with their data in AI-driven systems. ML models need to be broadly applicable and independently validated. The challenge is making them effective across various healthcare settings. Federated learning is a new approach that improves these models by learning from spread-out data without compromising privacy. This strategy tackles data silos and fosters the development of robust models delivering consistent results across diverse settings. Promoting federated learning and the standardization of ML models is crucial for achieving trustworthy, universal AI healthcare solutions.

Bias can emerge in AI models if the training data lacks representation from the broader population. For example, an algorithm predominately informed by data from a single ethnic group may not be as productive for another, possibly resulting in erroneous predictions or misdiagnoses of SA-AKI risk. Moreover, the opaque nature of “black-box” algorithms, which obscure their decision-making processes, raises ethical quandaries. Healthcare practitioners must discern mechanisms by which these AI models derive their outcomes, especially when these outcomes influence critical medical decisions. Nevertheless, techniques such as SHAP (Shapley Additive exPlanations), Gradient-based Class Activation Mapping (Grad-CAM), and LIME (Local Interpretable Model-agnostic Explanations) can enhance the model’s explainability [20]. These methods highlight the features significantly influencing model predictions, offering more profound insights into its decision-making process. Incorporating these techniques not only improves the interpretability of deep learning models but also bolsters trust and reliability in their application to SA-AKI research and clinical practices.

Integrating AI and ML tools in SA-AKI research and therapeutic interventions necessitates the establishment of standardized protocols and guidelines. Alongside standardization, a complex challenge emerges of ascertaining accountability in AI-induced misdiagnoses or mistreatments, raising questions about whether the onus lies with the software developers, the healthcare establishment, or the treating physician. Furthermore, akin to pharmacological agents or medical apparatuses, AI instruments might be subject to rigorous clinical evaluations and requisite regulatory endorsements before widespread adoption.