In our study, Acute Kidney Injury (AKI) was identified based on the KDIGO (Kidney Disease: Improving Global Outcomes) criteria. A patient was labeled as having AKI if any of the following conditions were met: an increase in serum creatinine (SCr) of ≥ 0.3 mg/dL within 48 h, an increase in SCr to ≥ 1.5 times the baseline within the prior 7 days, or a reduction in urine output to < 0.5 mL/kg/h for more than 6 h. These thresholds align with widely adopted clinical diagnostic standards, ensuring consistency across datasets. For model training and prediction, we focused on clinically relevant time points within a 48-h observation window prior to the first documented AKI event. Time-series features were sampled hourly, resulting in sequences ranging from 6 to 48 time steps depending on data availability. The model was designed to make AKI risk predictions at each hourly time point, enabling dynamic monitoring of renal function. Regarding cohort construction, we included adult ICU patients (age ≥ 18 ) with at least 12 consecutive hours of recorded data prior to AKI diagnosis. Patients were excluded if they had known end-stage renal disease (ESRD), a history of dialysis, missing baseline creatinine, or insufficient data (fewer than 6 recorded time steps). These inclusion and exclusion criteria ensure a robust dataset suitable for temporal modeling and reduce noise introduced by chronic kidney disease or incomplete records.

In real-world clinical environments, time-series data is frequently plagued by missing or irregular entries, which can pose significant challenges to deep learning-based temporal prediction systems. To address this, our workflow incorporates a two-fold strategy. We apply forward-fill and backward-fill imputation for short-term gaps in vital signs and lab measurements, followed by statistical imputation, such as mean or median values within patient-specific context, for persistent missing features. For categorical variables, missing entries are encoded with a designated embedding vector. This hybrid imputation approach is seamlessly integrated into the data preprocessing pipeline prior to model training and inference. Our LSTM-based model is designed to accommodate variable-length sequences without strict requirements on the minimum number of time steps. During training, variable-length sequences are padded with masking to ensure consistent batch sizes, and masked positions are ignored during loss computation.

Kidney injury is a complex and multifactorial condition that can arise from various etiological factors. To formally define and understand the problem of kidney injury, we start by introducing some key concepts and mathematical formulations relevant to the diagnosis and treatment strategies. These preliminaries will lay the foundation for the development of a model to better predict and manage kidney injury.

Let X ∈ R n × d denote the matrix of clinical observations, where n corresponds to the total number of patients and d denotes the number of clinical attributes per individual. Each feature dimension may represent variables such as blood pressure, serum creatinine concentration, urine output, and other relevant physiological indicators. The central objective is to learn a functional mapping from the input space X to a continuous or discrete estimation reflecting the severity of acute kidney injury in each patient (Equation 1). y ̂ i = f x i ; θ , (1) where θ represents the model parameters, which are learned from historical data.

A critical component in predicting kidney injury is the analysis of biomarkers that correlate with kidney function. Let y i ∈ { 0,1 } be a binary indicator of kidney injury for patient i , where y i = 1 indicates the presence of kidney injury, and y i = 0 indicates no injury. The core goal of the model involves optimizing predictions by minimizing their inconsistency with the true labels through the binary cross-entropy criterion (Equation 2). L θ = − 1 n ∑ i = 1 n y i ⁡ log y ̂ i + 1 − y i log 1 − y ̂ i . (2)

Recent studies have identified various biomarkers and physiological variables that might help improve the model’s predictive power. These biomarkers, denoted by b 1 , b 2 , … , b m , are included as additional features in the dataset X , and their relationships with kidney injury outcomes need to be explored. The biomarkers b 1 , b 2 , … , b m refer to clinically validated variables such as blood urea nitrogen (BUN), neutrophil gelatinase-associated lipocalin (NGAL), and cystatin C. These parameters are included in the input feature matrix X after normalization and preprocessing. While some biomarkers correlate with primary features such as serum creatinine, they offer additional diagnostic resolution and temporal sensitivity, improving model robustness in early-stage AKI detection.

Given the imbalance in the data—fewer instances of severe injury compared to mild or no injury—a weighted loss function may be used to improve model performance (Equation 3). L weighted θ = − 1 n ∑ i = 1 n w i y i ⁡ log y ̂ i + w i 1 − y i log 1 − y ̂ i , (3) where w i is the weight assigned to each instance to compensate for class imbalance.

Early detection of kidney injury requires dynamic monitoring of patient data over time. Let X ( t ) represent time-dependent observations for each patient. This motivates the use of temporal models (Equation 4). y ̂ i t = f x i t ; θ , (4)

which capture the evolving nature of the condition and utilize models such as LSTM or RNN to handle sequential dependencies.

Our model is designed to perform multi-task prediction for three clinically relevant outcomes: onset of acute kidney injury (AKI), likelihood of renal function recovery, and requirement for renal replacement therapy (dialysis). Each task is associated with distinct but interrelated clinical endpoints, which are modeled jointly to improve performance and generalizability.

In this section, we propose a novel model to predict kidney injury based on clinical and temporal data. Our approach integrates multiple sources of patient data, including static clinical features and dynamic biomarkers, and employs a deep learning architecture to model the complex dependencies between these variables. The model aims to improve early detection and accurate prediction of kidney injury severity by leveraging temporal patterns in patient data (As shown in Figure 1).

Let X ( t ) ∈ R n × d denote a time-dependent dataset capturing longitudinal clinical measurements, where n represents the cardinality of the patient set, and d defines the number of features measured at each discrete time point t . For an individual patient i , the temporal sequence of observations is expressed as x i = [ x i 1 , x i 2 , … , x i T ] , corresponding to measurements collected across T successive time steps t 1 , t 2 , … , t T . Each element x i t ∈ R d encapsulates the clinical profile at time t , including dynamic biomarkers such as serum creatinine levels, urine output volumes, and blood pressure readings.

We define the problem of kidney injury prediction as a sequence-to-sequence task, where the model learns to predict the probability of kidney injury y i ( t ) ∈ { 0,1 } at each time step given the past observations. The primary challenge lies in capturing the temporal dependencies between the observations, as kidney injury progression depends on prior measurements over time.

In this section, we propose a novel strategy to enhance the prediction of kidney injury by leveraging a combination of advanced techniques in model optimization, feature selection, and dynamic evaluation. Our approach is designed to address the challenges inherent in the prediction task, such as handling class imbalances, incorporating temporal dependencies, and improving the model’s generalization to unseen data (As shown in Figure 3).

In our empirical analysis, we rigorously assessed the performance of the proposed framework across multiple benchmark datasets using a standardized experimental protocol. Each dataset was subjected to consistent preprocessing, model training, and evaluation procedures. To improve model robustness and generalization, we applied augmentation techniques specifically tailored to time-series clinical data. These included introducing small, normally distributed noise to continuous-valued inputs to simulate physiological variability, applying elastic temporal scaling to reflect minor timing inconsistencies, and randomly masking non-essential variables to emulate common patterns of missingness observed in real-world EHR data. No image-based transformations such as rotations or spatial translations were employed, as all datasets used in this study consist entirely of structured, non-visual data. Each model was trained using stochastic gradient descent (SGD) with an initial learning rate of 0.001, which decayed by a factor of 10 every 10 epochs. A mini-batch size of 32 was used uniformly. These settings were sufficient to ensure convergence across all datasets. The model also demonstrated high computational efficiency at inference time, requiring less than 200 milliseconds to generate predictions for an individual patient record, making it viable for clinical deployment.

In our experimental design, the datasets were split using fixed ratios unless otherwise specified. For the MIMIC-III ICU dataset, which includes over 40,000 patient records, we employed an 80% training, 10% validation, and 10% testing split. The size of this dataset ensures statistical robustness, even with a single holdout approach. For smaller datasets such as the metabolomics and CKD cohorts, we adopted a 70-15–15 split to preserve the integrity of the evaluation process. The total sample sizes are as follows: metabolomics dataset contains approximately 2,500 patient entries, and the CKD dataset includes 1,100 samples. All samples available in the public versions of the datasets were used, and no exclusions were made. To evaluate the sensitivity of our results to the data splitting strategy, we performed three independent trials with different random seeds on the CKD dataset. The resulting variation in key performance metrics remained within ± 0.5%, suggesting that our findings are stable and not unduly influenced by the specific partition. Nonetheless, we acknowledge that more robust validation techniques, such as k-fold cross-validation or time-series splitting, are valuable alternatives and may be explored in future iterations of this work for broader generalizability. To provide a comprehensive view of the model’s computational efficiency, we report the average training time across the different datasets. All experiments were conducted on a workstation equipped with an NVIDIA RTX 3090 GPU and 128 GB of RAM. The MIMIC-III ICU dataset required approximately 5.6 h to complete 50 training epochs, while the metabolomics dataset took around 3.1 h. For the Chronic Kidney Disease dataset, the model converged within 2.4 h, and the general ICU dataset training took approximately 4.8 h. These durations include preprocessing, dynamic evaluation updates, and optimization using stochastic gradient descent with scheduled learning rate decay. Given these manageable training times, the method is practical for deployment in research and hospital environments with moderate computing infrastructure. Online inference remains highly efficient, typically requiring less than 200 milliseconds per patient record.

To provide transparency regarding our experimental setup, we summarize the key characteristics of each dataset used in our study, including the number of patients or samples analyzed, the types of variables, and the observation settings. These details are essential for interpreting the scope, temporal structure, and dimensionality of the experimental inputs, as shown in Table 1.

MIMIC-III ICU Dataset Mu et al. (2024) serves as a large, anonymized critical care dataset encompassing detailed medical records from upwards of 40,000 ICU patients, offering a valuable resource for data-driven clinical modeling. It includes detailed data such as demographics, vital signs, laboratory test results, medications, diagnoses, and more, making it a valuable resource for research in clinical decision support, patient outcome prediction, and healthcare analytics. MIMIC-III is particularly notable for its granularity and time-stamped data, which support a wide range of machine learning tasks including sequence modeling and risk stratification in intensive care unit settings. The metabolomics dataset Barupal et al. (2018) is a comprehensive repository of metabolic profiles derived from biological samples through mass spectrometry and nuclear magnetic resonance spectroscopy. It includes quantitative data on metabolite concentrations across different biological states and conditions. This dataset enables detailed investigation into metabolic pathways, disease biomarkers, and physiological changes, and is widely used in systems biology and bioinformatics for tasks such as classification, clustering, and feature selection. The Chronic Kidney Disease Dataset Amirgaliyev et al. (2018) is a clinical dataset that includes data from patients with chronic kidney disease (CKD). It contains attributes such as age, blood pressure, specific gravity, albumin levels, sugar levels, and several other indicators relevant to kidney function and general health. The dataset is frequently used in the development of classification algorithms for early diagnosis of CKD and in decision-support tools for personalized treatment planning. The ICU Dataset Yèche et al. (2021) used in this study refers to a clinical time-series benchmark, known as HiRID-ICU. It comprises high-resolution, multivariate physiological signals and structured patient data collected from intensive care units. The dataset is designed for machine learning research in healthcare and includes detailed temporal records such as heart rate, respiratory rate, blood pressure, and other vital signs. It has been widely adopted for tasks including early warning systems, patient deterioration prediction, and time-series classification in critical care settings.

Although ChronoNet is designed primarily for temporal prediction tasks, we also evaluated its performance on datasets with only static features (single time-point data), such as the metabolomics and CKD cohorts. These datasets were selected for their high-quality, diverse clinical and molecular attributes that offer valuable insight into AKI risk, despite the absence of time-series measurements. For these cohorts, the model configuration omits the temporal encoding modules and instead utilizes the static input processing path to produce predictions. This adjustment maintains architectural consistency while allowing us to assess the model’s versatility across varying data modalities. Including both temporal and static datasets also provides a more comprehensive evaluation of the framework’s clinical applicability, particularly in environments where longitudinal data may be limited or unavailable.

To provide essential context for interpreting model performance and addressing class imbalance, we summarized the frequencies of the two primary clinical outcomes—acute kidney injury (AKI) and dialysis requirement—across all datasets used in this study. These outcome distributions are shown in Table 2. In the MIMIC-III and general ICU (HiRID) datasets, AKI was present in approximately one-third of the patient population, while the need for dialysis occurred in less than 10% of cases. The CKD dataset exhibited a slightly higher prevalence of both outcomes, likely due to its focus on patients with pre-existing renal impairment. The metabolomics dataset provided only AKI outcome labels; dialysis annotations were not available. These statistics highlight the clinical relevance of the selected cohorts and justify the use of class balancing techniques such as weighted loss functions and synthetic oversampling in our model training pipeline.

In order to ensure a fair and comprehensive comparison, we included several baseline models that span diverse architectures and original application domains. Notably, CLIP, BLIP, and Wav2Vec—although primarily developed for computer vision or speech tasks—have demonstrated robust performance in learning complex feature representations across modalities. For the purpose of this study, we adapted their input pipelines to accept structured EHR data, converting tabular variables into appropriate input embeddings or tokenized sequences. These modifications allow for a meaningful evaluation of their transferability and general modeling capacity when applied to clinical time-series prediction tasks such as AKI forecasting. By including these baselines, we aim to highlight the domain-specific advantages of our ChronoNet framework, which integrates sequential modeling and attention mechanisms optimized for medical temporal data. The consistently superior performance of ChronoNet across all benchmark datasets validates the appropriateness and strength of our architectural choices, particularly when compared with models that were not natively designed for healthcare data environments.

This section presents an in-depth empirical evaluation of the proposed model, ChronoNet Model, benchmarked against a collection of leading state-of-the-art (SOTA) methods. The analysis spans four widely used datasets including MIMIC-III ICU, a curated metabolomics dataset, a Chronic Kidney Disease cohort, and a general ICU dataset. To ensure fair and reproducible comparison, several representative baselines are included, namely, CLIP Hafner et al. (2021), ViT Yuan et al. (2021), I3D Peng et al. (2023), BLIP Li et al. (2022), Wav2Vec 2.0 Pepino et al. (2021), and T5 Zhuang et al. (2023). Model effectiveness is assessed using established metrics prevalent in classification and recommendation domains, including accuracy, recall, F1 score, and area under the receiver operating characteristic curve (AUC). In Table 3, ChronoNet Model consistently delivers superior results on the MIMIC-III ICU and metabolomics datasets. On the MIMIC-III ICU dataset, it achieves an accuracy of 93.55 ± 0.02, recall of 91.67 ± 0.03, F1 score of 92.45 ± 0.01, and an AUC of 94.61 ± 0.02, surpassing all compared baselines. Similarly, on the metabolomics dataset, the model records 94.38 ± 0.03 accuracy, 93.24 ± 0.02 recall, 93.67 ± 0.02 F1 score, and 95.30 ± 0.03 AUC—demonstrating robust performance across heterogeneous biomedical domains. Additional comparisons, as reported in Table 4, highlight the model’s effectiveness on the Chronic Kidney Disease and general ICU datasets. For the CKD dataset, ChronoNet Model attains an accuracy of 91.67 ± 0.02, recall of 89.48 ± 0.03, F1 score of 90.31 ± 0.01, and AUC of 93.10 ± 0.03, outperforming the next-best model, ViT, by a notable margin. On the general ICU dataset, the model achieves 92.15 ± 0.02 accuracy, 91.21 ± 0.01 recall, 92.11 ± 0.02 F1 score, and an AUC of 94.35 ± 0.02, further emphasizing its generalizability and effectiveness in diverse clinical prediction settings.

Our proposed method consistently achieves the highest performance across all datasets, demonstrating its effectiveness in recommendation tasks. The improvements can be attributed to the novel architecture and optimization techniques used in ChronoNet Model, which allow it to better capture the underlying patterns in the data compared to existing methods. The detailed comparison in Figures 5, 6 highlights the superior performance of our method and validates its potential for real-world applications in 3D object recognition and recommendation tasks.

In this section, we conduct a structured ablation study to evaluate the contributions of individual components within the ChronoNet Model framework.The effects of these architectural modifications are assessed across four benchmark datasets including MIMIC-III ICU, a metabolomics dataset, a Chronic Kidney Disease cohort, and a general ICU population.The quantitative findings from this analysis are summarized in Tables 5, 6. To better understand the individual contributions of the ChronoNet components, we conducted a structured ablation study in which specific modules were either excluded or replaced with simpler alternatives. In the configuration without sequence-based risk prediction, we removed the LSTM layer entirely and replaced it with a multilayer perceptron (MLP) that processes the same static and temporal input features in a flattened form, without considering time dependencies. For the variant without temporal attention integration, we retained the LSTM backbone but removed the attention layer, relying solely on the final hidden state for prediction. To assess the impact of the generalization enhancement modules—including class imbalance handling, smoothness regularization, and dynamic evaluation—we disabled each of these techniques and trained the model under the original settings without auxiliary components. These controlled modifications allow for a focused evaluation of how each architectural element contributes to predictive performance.

Figure 7 presents the results of the ablation experiments conducted on the MIMIC-III ICU and metabolomics datasets. The experimental evidence highlights that ChronoNet Model consistently outperforms all tested baseline configurations—including Sequence-Based Risk Prediction, Temporal Attention Integration, and Improving Generalization Dynamics—across commonly adopted evaluation metrics such as accuracy, recall, F1 score, and AUC. The model attains accuracies of 93.55 ± 0.02 on the MIMIC-III ICU dataset and 94.38 ± 0.03 on the metabolomics dataset, demonstrating a clear advancement relative to prior methods. In Figure 8, the superior performance of ChronoNet Model also extends to the Chronic Kidney Disease and general ICU datasets. Across all evaluation criteria, the model consistently surpasses competing approaches, reinforcing its robustness and capacity for generalization across varied clinical prediction scenarios. The results further reveal the effectiveness of ChronoNet across various ablation settings. When excluding sequence-based risk prediction, the model’s F1 score on the MIMIC-III dataset dropped from 92.45% to 81.28%, indicating that sequential modeling plays a crucial role in capturing AKI progression. Similarly, removing the temporal attention integration reduced the AUC on the metabolomics dataset from 95.30% to 86.94%, showing that the attention mechanism significantly enhances time-sensitive prediction. Furthermore, the component related to improving generalization dynamics contributed notably to robustness across unseen samples, as reflected by higher AUC and recall values. These findings underscore that each architectural component is instrumental in achieving high-performance AKI prediction.

The findings from the ablation study underscore the significance of key architectural components, the novel feature extraction mechanism and the enhanced optimization strategy—in elevating the overall performance of ChronoNet Model. The integration of these elements contributes substantially to the model’s effectiveness, enabling it to consistently outperform baseline alternatives in both recommendation and object recognition scenarios. This analysis further validates the critical impact of individual design choices within ChronoNet Model and demonstrates its clear advantages over existing state-of-the-art techniques across varied application domains. Our empirical analysis shows that while the model achieves optimal performance with sequences spanning at least 12 h of hourly data, it remains functional and retains over 85% of peak accuracy with as few as 6 time points.

To further assess the explainability of the proposed ChronoNet model, we conducted a post hoc analysis using SHAP (SHapley Additive exPlanations) on the MIMIC-III test set. The goal was to identify which clinical features contributed most significantly to the prediction of AKI. Figure 9 lists the top-10 features with the highest average SHAP values. Notably, serum creatinine, urine output, and systolic blood pressure were the most influential, which is consistent with established clinical knowledge about AKI pathophysiology. This analysis enhances the interpretability of our model and supports its reliability for clinical deployment. Future work may integrate these explanations into a user interface for physicians to improve transparency and decision-making.

The results of our experiments demonstrate the clear potential of AI-based models to enhance early detection of AKI in critically ill patients. The high predictive accuracy observed across four independent datasets suggests that the model is generalizable and robust to different clinical environments. Notably, the model’s temporal attention mechanism enables identification of risk signals several hours before AKI onset, a clinically meaningful lead time that could support preemptive interventions such as fluid resuscitation, medication adjustment, or nephrology consults. The multi-task framework allows simultaneous prediction of dialysis requirement, providing actionable information for resource allocation and patient management. From a clinical implementation standpoint, the model’s compatibility with both time-series and static data broadens its potential use cases, including resource-limited settings where continuous monitoring may not be available. The attention-weighted output enhances interpretability, which is critical for clinician trust. Integration with electronic health records (EHRs) through real-time data streaming could enable automated alerts for impending AKI. However, before clinical deployment, prospective validation and user-interface adaptation will be essential to ensure seamless integration into existing workflows.

To evaluate the role of symbolic AI components in ChronoNet, we performed an ablation experiment where these modules were removed. The symbolic AI mechanisms in our full model include rule-based filters derived from AKI clinical guidelines, ontology-informed attribute priors, and weak supervision from medical knowledge graphs. As shown in Table 7, removing these symbolic components leads to a noticeable drop in predictive performance across both the MIMIC-III and CKD datasets. In particular, AUC drops by over 2% on both datasets, and F1 score drops by more than 1.5%, indicating the symbolic module enhances generalization and improves precision-recall alignment. These results empirically validate that structured clinical knowledge meaningfully complements the deep learning backbone in real-world AKI prediction tasks.

To provide transparency regarding class imbalance, Table 8 presents the distribution of AKI-positive versus negative cases across the datasets used in this study. The original distributions were heavily skewed, with minority class proportions ranging from 13.5% to 24.3%. To address this, we employed SMOTE to generate synthetic AKI-positive instances, achieving a near-balanced distribution in each dataset. We further applied class-weighted binary cross-entropy and focal loss to ensure the model’s learning remained sensitive to rare but clinically critical events. This combination led to an improvement of 3.1% in AKI recall and 2.7% in F1 score compared to the baseline model trained without imbalance handling. These results confirm that ChronoNet’s training pipeline effectively mitigates bias toward the majority class.

To empirically validate the effectiveness of the entropy-based attention regularization introduced in Equation 20, we performed an ablation study by removing this component from the ChronoNet model and retraining it across all benchmark datasets in Table 9. The results reveal a consistent degradation in performance metrics—particularly AUC and F1 score—across both static and dynamic datasets. For example, in the MIMIC-III ICU dataset, the AUC dropped from 94.61% to 91.92%, and the F1 score declined from 92.45% to 89.13%. This performance reduction was most pronounced in cases with irregular or sparse input sequences, which suggests that the entropy regularization plays a critical role in stabilizing the attention distribution. By encouraging a smoother, more diverse allocation of attention weights across time steps, the entropy term reduces the risk of the model overly focusing on a narrow window of temporal data. This is especially important in clinical settings where early indicators of AKI may be distributed across a broader range of time points. The inclusion of this regularization strategy contributes not only to performance gains but also to improved interpretability and reliability of temporal reasoning in high-stakes medical applications.