DCBM-Tri: a dual-channel bilinear mapping triplet model for early recognition of acute kidney injury in imbalanced cohorts Authors: Kai Wang, Ling Lin, Xinye Jin, Han Chen, Xudong Lu, Huilong Duan, Shan Nan Published: 2026 ABSTRACT Acute Kidney Injury (AKI) is a common clinical syndrome with poor prognosis and high mortality in the intensive care unit (ICU). Delayed diagnosis limits timely intervention and worsens outcomes, while early recognition is further challenged by the imbalanced distribution of AKI and non-AKI cases. A Dual-Channel Bilinear Mapping Triplet (DCBM-Tri) model was proposed for early AKI recognition, which used contrastive learning to enhance patient representations by capturing latent clinical features and improving discriminability in high-dimensional space. To identify clinically meaningful risk factors, SHAP-based interpretability analysis was further applied. The 12-hour-ahead prediction setting (AKI: non-AKI = 537: 2339) provided an optimal balance between discriminative performance and positive case identification. DCBM-Tri showed statistically significant improvements over conventional baselines, including LSTM- and resampling-based methods. However, no statistically significant improvement in AUPRC was observed over the feature-channel ablation model. Moreover, decision curve analysis demonstrated that DCBM-Tri provided a broader range of net clinical benefit across relevant risk thresholds. SHAP analysis further identified the top five contributing features as C-reactive protein, ionized calcium, bicarbonate, pH, and sodium. Overall, DCBM-Tri effectively addresses class imbalance in early AKI prediction by learning discriminative patient similarities, leading to improved sensitivity for high-risk patients. Its interpretable outputs further provide clinically meaningful signals to support early recognition and potential individualized prevention. Introduction Acute kidney injury (AKI) is a common clinical syndrome characterized by a rapid decline in kidney function, which is associated with poor prognosis1–3. It is increasingly recognized as a heterogeneous syndrome in which functional decline and structural kidney injury may not occur simultaneously4. In clinical practice, the diagnosis of AKI is based on serum creatinine (SCr) levels and urine output5. AKI may occur prior to detectable changes in SCr or urine output, resulting in AKI being diagnosed only after physiological deterioration has already taken place6. This limitation is especially pronounced in the intensive care unit (ICU), where increased susceptibility to structural kidney injury arises from hemodynamic instability, inflammatory stress, and exposure to nephrotoxic insults7. This increased susceptibility is reflected clinically by the development of subclinical AKI, which precede overt changes in SCr or urine output. In contrast with diagnostic markers, longitudinal clinical data in ICU patients contain rich information about changes in hemodynamics, metabolism, and organ function that precede overt AKI. Data-driven artificial intelligence methods can integrate these high-dimensional time series to capture latent patterns and temporal trajectories, enabling the early prediction of AKI before diagnostic markers become abnormal8–11. However, a major challenge is the data imbalance, as AKI events are relatively rare compared with non-AKI cases. This imbalance can bias model predictions toward the majority class12, reducing the sensitivity for correctly identifying patients at risk. Such data imbalance is a common and intrinsic issue in medical datasets. Facing this issue, increasing the contribution of minority-class samples during training can partially mitigate prediction bias, such as resampling techniques. However, they can’t increase sample diversity and may therefore limit model generalizability. Generative models can augment minority-class samples13,14 and theoretically increase sample diversity. Nevertheless, the generated data is difficult to assess and guarantee the authenticity, and even introduce bias in the feature distribution space15–17. To avoid altering the distribution, reweighting-based methods adaptively adjust the loss function to emphasize minority-class samples, such as dynamic reweighting strategies18. However, these methods are often sensitive to hyperparameter selection, which can adversely affect overall model robustness19. When reweighting methods suffer from limited stability or the authenticity of generated samples is uncertain, contrastive learning offers an alternative strategy. This approach improves minority-class representation by enhancing the model’s ability to distinguish between samples20,21. While contrastive learning has shown promise, most existing approaches rely on Euclidean or cosine distance. These distances are isotropic metrics, implicitly assuming equal importance across all feature dimensions and directions in the embedding space22,23. Moreover, contrastive loss function typically depends on margin-related hyperparameters to control inter-class separation24,25, which are often sensitive to tuning and may compromise representation quality or training stability. In this study, we aimed to leverage contrastive learning to enhance patient feature representations beyond the isotropic properties of conventional distance metrics, enabling a more discriminative characterization of AKI. Accordingly, we proposed a Dual-Channel Bilinear Mapping Triplet (DCBM-Tri) model for AKI early recognition. The proposed method employed a dual-channel feature extraction mechanism to decouple temporal patterns and inter-feature dependencies, enabling more informative representations of patient trajectories. A subsequent bilinear mapping learned anisotropic similarity relationships in the feature space, which were optimized by pairwise similarity classification, reducing reliance on margin-sensitive loss formulations. We evaluated the proposed approach on longitudinal ICU datasets and compared it with baseline methods for early AKI prediction. In addition, the model provided interpretable patterns that offer clinically meaningful insights for early risk assessment. Methods This study protocol was approved by the Ethics Committee of Sir Run Run Shaw Hospital (SRRSH) affiliated with the Zhejiang University School of Medicine (#20210319-30). This retrospective study used data that were fully anonymized and did not involve identifiable personal information and commercial interests; therefore, informed consent was waived by the Ethics Committee according to the “Guidelines for the Establishment of Ethics Review Committees for Clinical Research Involving Humans”. Moreover, the study was conducted in accordance with the principles of the Declaration of Helsinki and all methods were performed in accordance with the relevant guidelines and regulations. The methods followed the Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD) , Minimum Information for Medical AI Reporting (MINIMAR) guidelines, and Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines26–28. Dataset All patients admitted to the ICU of SRRSH between September 2014 and March 2022 were included in the development cohort, while those admitted between April 2022 and December 2022 constituted the temporal external validation cohort. The study outcome was ICU-acquired AKI, including all severity stages (KDIGO stages 1-3), occurring between day 2 and day 9 after ICU admission. AKI was defined according to the “Kidney Disease: Improving Global Outcomes” (KDIGO) criteria as any of the following: (1) Within 48 hours, SCr increases by 0.3 mg/dl (26.5 µmol/L); (2) SCr increases to 1.5 times the basic value (which comes from the lowest SCr value within 7 days); (3) Urine output is less than 0.5 ml/(kg·h) for 6 hours. KDIGO staging was defined as follows: Stage 1 corresponds to a SCr increase of ≥ 0.3 mg/dL (26.5 µmol/L) or 1.5-1.9 times baseline, or urine output < 0.5 mL/(kg·h) for 6-12 hours; Stage 2 to SCr 2.0-2.9 times baseline or urine output < 0.5 mL/(kg·h) for ≥ 12 hours; and Stage 3 to SCr ≥ 3.0 times baseline or ≥ 4.0 mg/dL (353.6 µmol/L), initiation of renal replacement therapy, or urine output < 0.3 mL/(kg·h) for ≥ 24 hours or anuria for ≥ 12 hours. The exclusion criteria were as follows: ICU stay of less than 2 days, development of AKI before ICU admission or within the first 48 hours after ICU admission, pre-existing renal replacement therapy (including hemodialysis or peritoneal dialysis), and history of organ transplantation. In the ICU, arterial blood gas and biochemical measurements are routinely and frequently obtained as part of standard care, reflecting dynamic changes in metabolic status, oxygenation, and organ function. Therefore, these variables were retrospectively extracted from the electronic health record data for all included patients. A total of 12 arterial blood gas parameters were collected, including potential of hydrogen (pH), partial pressure of carbon dioxide (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm {pCO_2}$$\end{document}), partial pressure of oxygen (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm {pO_2}$$\end{document}), base excess in blood (BEB), bicarbonate (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm {HCO_3^-}$$\end{document}), potassium (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm {K^+}$$\end{document}), sodium (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm {Na^+}$$\end{document}), ionized calcium (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm {iCa^{2+}}$$\end{document}), glucose (GLU), lactate (LAC), oxygen saturation (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm {O_2Sat}$$\end{document}), and fraction of inspired oxygen (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm {FiO_2}$$\end{document}). In addition, four biochemical laboratory indicators were included SCr, blood urea (BU), albumin (ALB), and hypersensitive C-reactive protein (hs-CRP). Data preprocessing The observation window was defined as 48 hours, during which patient data were used to predict the risk of AKI occurring within the subsequent K-hour horizon (K = 6, 12, 24, and 48 hours). Prediction was performed until ICU discharge or up to ICU day 9. For patients who did not develop AKI, successive sliding windows exhibit extensive temporal overlap as the prediction horizon K is shorter than the observation window, resulting in highly redundant negative samples. To mitigate this issue, only the first 48-hour observation window after ICU admission was retained for non-AKI patients to predict the subsequent K-hour risk. For patients who developed AKI, the time of first AKI onset was identified, and the preceding 48-hour period was defined as the observation window, ensuring that the model captured the pre-onset AKI trajectory. This sample construction strategy preserves clinically meaningful temporal information while avoiding redundancy and negative-sample inflation induced by sliding windows. To construct time-series inputs, the 48-hour observation window was divided into eight consecutive 6-hour time slices, and the mean value of each variable was computed within each slice. To ensure temporal integrity, samples with complete data absence in any time slice were excluded. If all variables were missing within a given slice, a temporal break was formed; imputing such slices would introduce entirely artificial observations, leading to spurious temporal trends or smoothing effects across both time and feature dimensions. In this condition, the model would be unable to learn meaningful temporal dependencies or feature interactions from that slice. In contrast, random missing values affecting individual variables within a time slice were imputed using mean-based strategies, as such missingness does not disrupt the temporal continuity of the sequence. Finally, positive and negative samples were randomly split into training, validation, and test sets at a ratio of 6:2:2. Triplet sampling Triplets were independently constructed from positive and negative sample sets, where anchor, positive, and negative samples were from different patients. For positive triplets, each positive sample was used as an anchor, another distinct positive sample was selected as the positive instance, and a randomly selected negative sample served as the negative instance. Negative triplets were constructed analogously by using negative samples as anchors. To mitigate class imbalance and enrich contrastive supervision, oversampling was performed at the triplet-construction level rather than by duplicating raw samples. Specifically, minority-class samples were paired with multiple positive and negative to generate additional contrastive relationships. Each positive anchor was sampled 12 times, while each negative anchor was sampled 3 times. Triplets in the test set were constructed according to the original class distribution. Proposed model Early recognition of AKI in the ICU faces two major challenges: (1) severe class imbalance, which biases classification models toward the majority class and limits accurate identification of AKI patients; and (2) the reliance of most contrastive learning frameworks on isotropic distance metrics (Euclidean or cosine distance), which may inadequately capture the characteristics of ICU patients. To address these challenges, we proposed a Dual-Channel Bilinear Mapping Triplet (DCBM-Tri) model for early AKI recognition, as shown in Fig. 1. The model leveraged contrastive learning to enhance class discrimination without altering the original data distribution, and incorporated learnable bilinear mappings and dual-channel feature extraction to relax the limitations of conventional distance metrics.Figure 1Schematic of the Dual-Channel Bilinear Mapping Triplet (DCBM-Tri) model, consisting of the feature extraction module based self-attention mechanism, the low-rank bilinear mapping module, and the convolution-based similarity prediction head. Schematic of the Dual-Channel Bilinear Mapping Triplet (DCBM-Tri) model, consisting of the feature extraction module based self-attention mechanism, the low-rank bilinear mapping module, and the convolution-based similarity prediction head. Feature extraction To support high-dimensional embedding and discriminative representation learning in DCBM-Tri, we first extract information-rich latent features from the input data. DCBM-Tri employed a dual-channel feature extraction framework to model both temporal dynamics and inter-feature relationships. In the temporal channel, a bidirectional long short-term memory (BiLSTM) network was first applied to capture sequential dependencies in patient trajectories. This was followed by the self-attention mechanism that further refines temporal representations by modeling global temporal interactions. In the feature channel, a feed-forward network (FFN) was used to project input features into a higher-level representation space, enabling nonlinear feature transformation. Subsequently, the self-attention mechanism was applied to explicitly capture interactions among clinical variables. Finally, both channels produced feature embeddings with aligned dimensionality, facilitating subsequent bilinear mapping. Bilinear mapping Contrastive learning methods rely on distance computation under an identity matrix, such as Euclidean distance or cosine similarity, which treat all feature dimensions equally. To address this limitation, DCBM-Tri constructed a learnable metric space through the bilinear mapping, as shown in Eq 1. The learnable matrix (M) allows the differences between the anchor and other samples to be represented in a higher-dimensional space by assigning adaptive weights to each feature pair.1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} d(E_{anc}, E_{pos}) = E_{anc} \cdot M \cdot E_{pos} \end{aligned}$$\end{document}However, the learnable matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$M \in {\mathbb {R}}^{d \times d}$$\end{document} (where d denotes the dimensionality of the hidden representations after feature extraction) would introduce a quadratic growth in the number of parameters and computational cost. In medical settings with limited sample sizes, especially when minority classes are scarce, models with a large number of parameters are more prone to overfitting29. Moreover, high-dimensional metrics are prone to the curse of dimensionality, making distance measures unstable and blurring inter- and intra-class boundaries. To address these issues, DCBM-Tri adopted a low-rank factorization, decomposing M to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$U \cdot V^T$$\end{document}, where U, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$V \in {\mathbb {R}}^{d \times r}$$\end{document} and r ≪ d. This low-rank formulation enabled each channel to learn an independent bilinear metric with far fewer parameters. Prediction head After feature projection through the bilinear mapping, DCBM-Tri fused the high-dimensional representations of the positive-anchor and negative-anchor pairs from both the temporal and feature channels into a unified three-dimensional tensor. The fused representation preserves both dynamic temporal patterns and cross-feature couplings, providing rich latent information for contrastive learning. Then, 3D convolution extracted local patterns across time and feature dimensions, capturing higher-order dependencies and enhancing the discriminability of sample pairs in the embedding space. After convolution, average pooling is used to compress the high-dimensional features into a fixed-length vector, which is subsequently mapped to the label space for classification. The AKI prediction was formulated as a mutually exclusive pairwise similarity learning problem, where each anchor is associated with one positive and one negative candidate. The model estimated similarity scores for each pair using sigmoid activations, optimized via binary cross-entropy loss under a structured label constraint. The overall optimization objective is:2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} {\mathcal {L}}=-\sum _{i\in \{ap, an\}}[y_i \log ({\hat{y}}_i)+(1-y_i)\log (1-{\hat{y}}_i)] \end{aligned}$$\end{document}The final prediction vector is [\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\hat{y}}_{ap}$$\end{document},\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\hat{y}}_{an}$$\end{document}]. The label vector is [\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$y_{ap}$$\end{document},\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$y_{an}$$\end{document}], where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$y_{ap}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$y_{an}$$\end{document} denote the binary similarity labels for the anchor-positive and anchor-negative pairs, respectively. Under the mutually exclusive label constraint [1, 0] / [0,1], the loss function is equivalent to a pairwise mutually exclusive binary classification loss.3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} {\mathcal {L}}={\left\{ \begin{array}{ll} -\log ({\hat{y}}_{ap})-\log (1-{\hat{y}}_{an}),& \text {if } [y_{ap},y_{an}]=[1,0] \\ -\log (1-{\hat{y}}_{ap})-\log ({\hat{y}}_{an}),& \text {if } [y_{ap},y_{an}]=[0,1] \end{array}\right. } \end{aligned}$$\end{document} Baseline comparison The performance of DCBM-Tri was compared with the following methods on the test set using the original class distribution: (1) Traditional machine learning models: support vector machine (SVM), random forest (RF), logistic regression (LR), and gradient boosting decision tree (GBDT). (2) Conventional deep learning models (time-channel): long short-term memory (LSTM), one-dimensional convolutional neural network (1D-CNN), and Feedforward Neural Network (FNN). (3) Resampling-based methods: the best-performing model (highest AUROC) among the traditional machine learning and deep learning approaches was combined with the synthetic minority over-sampling technique (SMOTE). Since the dataset is two-dimensional, samples were flattened before applying SMOTE and then reshaped back to their original format. Undersampling was not considered because the number of positive samples in the training and validation sets was small (approximately 100 in the validation set). (4) Reweighting-based methods: the best-performing traditional or deep learning model (with the highest AUROC) was further combined with class weighting to address class imbalance. (5) Single-channel variants of DCBM-Tri: a time-channel (T-channel) model and a feature-channel (F-channel) model. Statistical analysis Within each observation window, summary statistics of all variables were calculated. Differences between the AKI and non-AKI groups were assessed using the Mann-Whitney U test, with a p value < 0.05 considered statistically significant. Model performance was evaluated using the area under the receiver operating characteristic curve (AUROC), the area under the precision–recall curve (AUPRC), precision, sensitivity, specificity, and accuracy. To assess whether differences in AUROC and AUPRC between models were statistically significant, a nonparametric bootstrap (n=1,000) was applied to estimate the empirical distribution of performance differences. Statistical significance was determined by testing whether the 95% confidence interval of the difference excluded zero (p < 0.05). In addition, decision curve analysis (DCA)30 was used to evaluate the net clinical benefit of each model across a range of risk thresholds. Model interpretability was assessed using SHapley Additive exPlanations (SHAP)31 to quantify the contribution of individual features to model predictions. Results Patients enrollment For the development cohort, 774 patients developed AKI between days 2 and 9 after ICU admission, while 2,928 patients did not. Among the AKI cases, 385 patients were classified as Stage 1,222 as Stage 2, and 167 as Stage 3. A total of 192 patients with AKI required hemodialysis and continuous renal replacement therapy. We searched ICU admission records using keyword matching to identify the presence of chronic kidney disease (CKD). No documented CKD cases were identified in the cohort. Clinical variables recorded within the first 48 hours after ICU admission were used to compute the average, and the differences between AKI and non-AKI groups are summarized in Table 1. The results showed that approximately 80% of the features exhibited significant differences between the two groups. In contrast, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm {pCO_2}$$\end{document}, LAC, and ALB did not reach statistical significance. Considering the average may attenuate temporal dynamics, these variables were nevertheless retained as model inputs.Table 1Statistical results of variable characteristics.Development cohortDevelopment cohort (n=3702)Validation cohort (n=744)pAKI (n=774)Non-AKI (n = 2928)ppH7.41 (0.05)7.40 (0.04)< 0.017.41 (0.04)7.40 (0.06)0.041\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm {pCO_2}$$\end{document} (mmHg)39.6 (6.7)39.4 (6.6)0.44339.5 (6.6)38.5 (6.3)0.036\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm {pO_2}$$\end{document} (mmHg)110.8 (25.7)115.2 (26.8)< 0.001114.4 (26.7)109.9 (35.4)< 0.01BEB (mmol/L)0.69 (4.16)− 0.10 (3.76)< 0.0010.05 (3.86)− 0.91 (4.03)< 0.01\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm {HCO_3^-}$$\end{document} (mmol/L)25.3 (4.4)24.6 (4.0)< 0.00124.7 (4.1)23.8 (4.01)< 0.01\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm {K^+}$$\end{document} (mmol/L)3.86 (0.39)3.81 (0.35)0.0113.82 (0.36)3.83 (0.41)0.358\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm {Na^+}$$\end{document} (mmol/L)141.4 (7.1)140.1 (5.0)0.047140.4 (5.5)139.6 (5.3)< 0.01\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm {iCa^{2+}}$$\end{document} (mmol/L)1.10 (0.08)1.10 (0.06)0.0051.10 (0.07)1.12 (0.08)< 0.01GLU (mmol/L)9.30 (2.66)8.75 (2.33)< 0.0018.85 (2.40)8.67 (2.83)< 0.01LAC (mmol/L)1.98 (1.35)1.95 (1.20)0.6301.95 (1.23)1.97 (2.72)< 0.01\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm {O_2Sat}$$\end{document} (%)97.5 (3.0)98.0 (1.7)< 0.00197.9 (2.1)97.9 (2.6)0.425\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm {FiO_2}$$\end{document} (%)44.0 (12.5)40.8 (9.5)< 0.00141.4 (10.2)36.1 (11.8)< 0.01BU (μmol/L)8.13 (5.60)6.60 (4.03)< 0.0016.89 (4.4)7.83 (5.77)< 0.01SCr (mmol/L)105.3 (112.9)78.3 (60.0)< 0.00183.5 (72.6)105.7 (104.2)< 0.01ALB (g/L)31.0 (5.2)31.1 (5.3)0.56631.1 (5.3)34.0 (5.9)< 0.01hs-CRP (mg/L)86.2 (78.9)77.1 (72.7)0.02278.8 (74.0)48.1 (68.3)< 0.01Data are presented as mean (standard deviation). The differences (development cohort vs temporal external validation cohort; AKI vs. non-AKI) were assessed using the Mann–Whitney U test; p < 0.05 indicates statistical significance Table 1Statistical results of variable characteristics.Development cohortDevelopment cohort (n=3702)Validation cohort (n=744)pAKI (n=774)Non-AKI (n = 2928)ppH7.41 (0.05)7.40 (0.04)< 0.017.41 (0.04)7.40 (0.06)0.041\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm {pCO_2}$$\end{document} (mmHg)39.6 (6.7)39.4 (6.6)0.44339.5 (6.6)38.5 (6.3)0.036\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm {pO_2}$$\end{document} (mmHg)110.8 (25.7)115.2 (26.8)< 0.001114.4 (26.7)109.9 (35.4)< 0.01BEB (mmol/L)0.69 (4.16)− 0.10 (3.76)< 0.0010.05 (3.86)− 0.91 (4.03)< 0.01\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm {HCO_3^-}$$\end{document} (mmol/L)25.3 (4.4)24.6 (4.0)< 0.00124.7 (4.1)23.8 (4.01)< 0.01\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm {K^+}$$\end{document} (mmol/L)3.86 (0.39)3.81 (0.35)0.0113.82 (0.36)3.83 (0.41)0.358\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm {Na^+}$$\end{document} (mmol/L)141.4 (7.1)140.1 (5.0)0.047140.4 (5.5)139.6 (5.3)< 0.01\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm {iCa^{2+}}$$\end{document} (mmol/L)1.10 (0.08)1.10 (0.06)0.0051.10 (0.07)1.12 (0.08)< 0.01GLU (mmol/L)9.30 (2.66)8.75 (2.33)< 0.0018.85 (2.40)8.67 (2.83)< 0.01LAC (mmol/L)1.98 (1.35)1.95 (1.20)0.6301.95 (1.23)1.97 (2.72)< 0.01\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm {O_2Sat}$$\end{document} (%)97.5 (3.0)98.0 (1.7)< 0.00197.9 (2.1)97.9 (2.6)0.425\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm {FiO_2}$$\end{document} (%)44.0 (12.5)40.8 (9.5)< 0.00141.4 (10.2)36.1 (11.8)< 0.01BU (μmol/L)8.13 (5.60)6.60 (4.03)< 0.0016.89 (4.4)7.83 (5.77)< 0.01SCr (mmol/L)105.3 (112.9)78.3 (60.0)< 0.00183.5 (72.6)105.7 (104.2)< 0.01ALB (g/L)31.0 (5.2)31.1 (5.3)0.56631.1 (5.3)34.0 (5.9)< 0.01hs-CRP (mg/L)86.2 (78.9)77.1 (72.7)0.02278.8 (74.0)48.1 (68.3)< 0.01Data are presented as mean (standard deviation). The differences (development cohort vs temporal external validation cohort; AKI vs. non-AKI) were assessed using the Mann–Whitney U test; p < 0.05 indicates statistical significance Statistical results of variable characteristics. Data are presented as mean (standard deviation). The differences (development cohort vs temporal external validation cohort; AKI vs. non-AKI) were assessed using the Mann–Whitney U test; p < 0.05 indicates statistical significance The temporal range of feature values extracted within the observation window shifted according to the prediction window. After excluding patients with temporal discontinuities in their longitudinal records, the number of AKI patients varied across different prediction horizons, resulting in 532 (K = 6 h), 537 (K = 12 h), 519 (K = 24 h), and 520 (K = 48 h) eligible AKI cases, while 2,339 non-AKI patients were included for all horizons, as shown in Fig. 2.Figure 2Flowchart of patient selection. The development cohort was collected between September 2014 and March 2022, while the temporal external validation cohort was collected from the ICU of SRRSH between April 2022 and December 2022. Flowchart of patient selection. The development cohort was collected between September 2014 and March 2022, while the temporal external validation cohort was collected from the ICU of SRRSH between April 2022 and December 2022. Identification of the prediction windows To evaluate the model’s ability to predict AKI at different lead times, four prediction horizons (6 h, 12 h, 24 h, and 48 h) were evaluated to select the optimal prediction window. Table 2 shows that DCBM-Tri achieved the highest AUROC at the 6-hour prediction horizon and the highest AUPRC at the 12-hour horizon. When the prediction horizon was further extended to 24 h and 48 h, both metrics declined markedly.Table 2The performance of the model on the test set under different prediction windows K.KAUROCAUPRC6 h0.87300.624212 h0.85950.663324 h0.72220.383148 h0.72440.3785 Table 2The performance of the model on the test set under different prediction windows K.KAUROCAUPRC6 h0.87300.624212 h0.85950.663324 h0.72220.383148 h0.72440.3785 The performance of the model on the test set under different prediction windows K. When the prediction horizon is short (6 h), the observation window is temporally close to AKI onset, allowing early physiological abnormalities to be more pronounced and resulting in the highest AUROC. However, such a short horizon provides limited lead time for clinical intervention in ICU practice. At the 12-hour horizon, although the observation window is farther from AKI onset, more AKI patients share temporally aligned observation windows. For example, patients who develop AKI at 50 h and 56 h correspond to the temporally aligned observation window (0-48 h). These aligned observation windows reduce temporal misalignment between positive samples, which alleviates similarity distortion in time-series32. This improves the ranking of positive samples, potentially contributing to improved AUPRC. With the longer prediction horizons (24 h and 48 h), early AKI-related physiological signals are weakened and masked by temporal variability. This reduces the separability between AKI and non-AKI samples, leading to declines in both AUROC and AUPRC. Therefore, although the 6-hour horizon achieved the highest AUROC, the 12-hour horizon was selected for subsequent experiments. It offers a longer clinically actionable lead time while preserving robust discrimination and improved positive-case identification performance. Model performance of the early AKI recognition For imbalanced binary prediction of clinical events, metrics such as accuracy and AUROC alone are insufficient to reflect performance on the minority class. In contrast, AUPRC is more informative for evaluating the ability to correctly identify AKI patients in real-world clinical settings where the prevalence of AKI is low. Table 3 shows the performance of DCBM-Tri on the training, validation, and test sets. On the test set, the model achieved an AUROC of 0.8595, indicating strong overall discrimination. The model achieved an AUPRC of 0.6633, with a precision of 0.6068 and a sensitivity of 0.6574, indicating a balanced ability to identify high-risk patients while maintaining moderate false-positive rates. The remaining false positives may be attributed to the limited and heterogeneous positive samples. The specificity reached 0.9017, indicating accurate identification of non-AKI patients and effective control of false alarms in this imbalanced setting.Table 3The performance of the model at the 12-hours prediction window.AUROCAUPRCPrecisionSensitivitySpecificityAccuracyTrain0.93710.93450.89590.80640.90170.8625Valid0.82810.80410.79420.57710.86320.7266Test0.85950.66330.60680.65740.90170.8559 Table 3The performance of the model at the 12-hours prediction window.AUROCAUPRCPrecisionSensitivitySpecificityAccuracyTrain0.93710.93450.89590.80640.90170.8625Valid0.82810.80410.79420.57710.86320.7266Test0.85950.66330.60680.65740.90170.8559 The performance of the model at the 12-hours prediction window. Model comparison The ROC and PR curves of all models are shown in Fig. 3a,b. DCBM-Tri achieved the best performance on both curves. Among the machine learning and deep learning baseline models, LSTM achieved the highest performance (AUROC of 0.7548, AUPRC of 0.4556). We further applied SMOTE-based resampling and sample reweighting to the LSTM model. Both strategies improved model performance, with SMOTE achieving an AUROC of 0.8082 and an AUPRC of 0.5571, and sample reweighting achieving an AUROC of 0.8199 and an AUPRC of 0.5387. These results indicate that imbalance-aware strategies can partially mitigate performance degradation due to class imbalance, but they remain inferior to DCBM-Tri (AUROC of 0.8595, AUPRC of 0.6633). Moreover, the single-channel variants (F-channel and T-channel) both showed reduced performance compared with DCBM-Tri, indicating that the dual-channel design effectively integrates temporal and feature-level information to achieve more comprehensive representations and improve early AKI recognition in the ICU.Figure 3Performance evaluation curves. (a) Receiver operating characteristic curves (ROC). (b) Precision-recall curves (PRC). (c) Bootstrap-based statistical comparison of AUROC (upper triangle) and AUPRC (lower triangle) between models, where p-values < 0.05 indicate statistically significant differences. (d) Decision curve analysis (DCA). Two extreme reference curves are shown in the DCA: one assumes that all samples are negative and therefore receive no treatment, yielding a net benefit of zero; the other assumes that all samples are positive and all receive treatment, resulting in a reference line with a negative slope. Performance evaluation curves. (a) Receiver operating characteristic curves (ROC). (b) Precision-recall curves (PRC). (c) Bootstrap-based statistical comparison of AUROC (upper triangle) and AUPRC (lower triangle) between models, where p-values < 0.05 indicate statistically significant differences. (d) Decision curve analysis (DCA). Two extreme reference curves are shown in the DCA: one assumes that all samples are negative and therefore receive no treatment, yielding a net benefit of zero; the other assumes that all samples are positive and all receive treatment, resulting in a reference line with a negative slope. To assess the statistical significance of differences in AUROC and AUPRC among models, a bootstrap resampling procedure (n = 1000) was used to estimate the distributions of pairwise metric differences. The results are shown in Fig. 3c, where the upper triangular matrix reports p-values for AUROC and the lower triangular matrix reports p-values for AUPRC. A p-value < 0.05 indicates statistical significance. For AUROC, DCBM-Tri showed statistically significant improvements over most baseline models. In contrast, DCBM-Tri showed no statistically significant differences compared with SMOTE or reweighting methods (p > 0.05). For AUPRC, DCBM-Tri achieved statistically significant improvements compared with SMOTE and reweighting methods. However, no statistically significant difference was observed between DCBM-Tri and the F-channel model (p > 0.05). This may be attributed to the strong representational capacity of the feature-channel component in capturing discriminative patterns. Although AUROC and AUPRC can evaluate classification performance, they do not directly reflect clinical usefulness. DCA was used to assess the net clinical benefit of each model across different risk thresholds by balancing the benefit of identifying high-risk patients (true positives) against the harm of unnecessary interventions in low-risk patients (false positives). As shown in Fig. 3d, two reference strategies are included: treating all patients (solid line) and treating none (dashed line). At low thresholds, models achieve high net benefit but are close to the “treat-all” strategy with many false positives. As the threshold increases, false positives decrease, while false negatives increase. By selecting appropriate risk thresholds, clinicians can balance sensitivity and false-positive rates to maximize clinical benefit. Although the F-channel achieved high AUROC and AUPRC, it did not demonstrate a clearly robust decision threshold, indicating a limited balance between false positives and sensitivity. In contrast, DCBM-Tri showed a wider range of useful thresholds (0.1–0.6) with higher net benefit and greater separation from the extreme strategies, demonstrating superior clinical applicability. Temporal external validation The same inclusion and exclusion criteria, as well as the same AKI definition strategy, were consistently applied to both cohorts. The external validation cohort included 194 patients with AKI and 550 patients without AKI. We performed a comparative analysis of clinical variables between the development and temporal external validation cohorts (Table 1). The results demonstrated statistically significant differences in more than 80% of variables, which may be due to temporal distribution shifts. After excluding patients with temporal discontinuities in their longitudinal records, the final external validation cohort (K = 12) consisted of 82 AKI patients and 192 non-AKI patients. Despite temporal heterogeneity between cohorts, the proposed model maintained stable predictive performance in the temporal external validation cohort, achieving an AUROC of 0.8449 and an AUPRC of 0.7223. These results indicate robust discriminative ability and generalizability. Model interpretability Feature importance quantifies the contribution of individual input variables to model performance, thereby improving interpretability and helping identify key risk factors33. SHAP analysis uses a game-theoretic framework to estimate each feature’s contribution by perturbing the input data. Using the training set as background data, SHAP analysis was performed on the test set, and the results are shown in Fig. 4.Figure 4SHAP-based interpretability visualizations. (a) Mean SHAP values of the anchor samples in the test set, where the x-axis represents features and the y-axis represents time slices. (b) Feature contributions aggregated over the temporal dimension for each feature. SHAP-based interpretability visualizations. (a) Mean SHAP values of the anchor samples in the test set, where the x-axis represents features and the y-axis represents time slices. (b) Feature contributions aggregated over the temporal dimension for each feature. Figure 4a shows the contribution of each feature at different time steps, where higher SHAP values are observed at the first time point (T1). SHAP assumes feature independence, meaning that the distributions of other features remain unchanged when a subset of features is sampled or removed. However, time-series features are highly autocorrelated in practice. Under the temporal dependencies introduced by LSTM and the temporal interactions modeled by the attention mechanism, this independence assumption is violated. As a result, highly correlated time steps tend to be treated as interchangeable, leading SHAP to attribute their shared contributions preferentially to earlier time point. The top five contributing features are hs-CRP, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm {iCa^{2+}}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm {HCO_3^-}$$\end{document}, pH, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm {Na^+}$$\end{document}, as shown in Fig. 4b. Interestingly, the kidney function markers SCr and BUN show the lowest contributions. This observation is consistent with previously reported physiological mechanisms. Due to the long half-life of creatinine and the presence of renal functional reserve (renal reserve capacity that can be recruited before basal glomerular filtration rate (GFR) starts to decline34), SCr requires time to accurately reflect GFR resulting in delayed recognition of AKI7. Although SCr is used to diagnose AKI, it exhibits a delayed response to underlying pathophysiological changes and therefore provides limited early-warning information for the LSTM-attention model. Consequently, its limited temporal variability provides little informative signal for the LSTM-attention model. Discussion In the ICU, timely early recognition of AKI is crucial for improving patient outcomes. However, widely used diagnostic markers often exhibit delayed responses to underlying physiological changes, which can lead to missed opportunities for intervention. To address this limitation, several AKI biomarkers, such as neutrophil gelatinase-associated lipocalin (NGAL)35, kidney injury molecule-1 (KIM-1)36, interleukin-18 (IL-18)37, and and heat shock proteins (HSP)38, have been proposed and shown in prior studies to enable earlier detection. Nevertheless, the diagnostic performance of these biomarkers remains influenced by threshold effects, heterogeneous definitions of AKI, and variability in patient populations and clinical settings39. In addition, these biomarkers are relatively costly40 and are not routinely or continuously available in the ICU. In our retrospective ICU cohort, NGAL measurements were largely unavailable, with a missing rate exceeding 90%, while KIM-1, IL-18, and HSP were not routinely collected. These limitations highlight the need for scalable strategy that rely on routinely available, continuously recorded clinical variables to enable early predict AKI risk. To address this challenge, we proposed DCBM-Tri, which early recognized the AKI for a target patient by measuring the similarity to patients who are known to develop AKI. In conventional classification frameworks, class imbalance leads to the majority class dominating loss optimization and gradient updates, resulting in suboptimal representation learning for minority cases and reduced sensitivity to high-risk patients41. Contrastive learning enhances patient representations by focusing on similarity among patients with the same outcome and dissimilarity across different outcomes42. To further improve the distance metric and margin-based loss in contrastive learning, we introduced a bilinear mapping. The learnable matrix projected differences between anchor and paired samples into a higher-dimensional space, enabling adaptive weighting of feature interactions and more discriminative similarity estimation. In addition, we adopt a dual-channel feature extraction to better capture latent patient representations from both temporal and feature perspectives. The results demonstrate that DCBM-Tri significantly improves the sensitivity and early identification of AKI in the ICU under highly imbalanced data conditions. In comparison with biomarker-based approaches for early AKI detection, a prior meta-analysis39 reported a highest diagnostic odds ratio (DOR) of 16.2 for NGAL/Cr in predicting AKI. In our study, the proposed model achieved a higher DOR of 17.6 in the test set. However, this comparison should be interpreted with caution, as DOR is highly sensitive to disease prevalence, patient spectrum, clinical setting, and threshold selection. In addition, differences in study design, case, and AKI definition across studies may further limit the validity of direct performance comparisons. Therefore, the observed numerical difference does not imply superior generalizable diagnostic performance, but rather suggests comparable discriminative ability within the analyzed cohort. Beyond predictive performance, identifying the key features driving the model’s decisions is essential for improving its interpretability43. Feature importance analysis reveals which clinical variables the model primarily relies on when early recognizing AKI, thereby reflecting its different physiological states. Using SHAP to rank feature contributions allowed the identification of variables with the greatest influence on the prediction. These high-impact features corresponded to well-established pathophysiological processes preceding AKI, such as systemic inflammation, metabolic and acid-based disturbances, hemodynamic instability, or impaired renal perfusion. However, SHAP in longitudinal clinical data should be interpreted with caution. SHAP assumes feature independence, whereas temporal clinical variables exhibit autocorrelation and interdependence across adjacent time steps. Consequently, SHAP should be interpreted cautiously in longitudinal clinical settings, as its attributions may be biased toward earlier time steps. Inflammation is an important driver of the onset of AKI. The 13th Acute Dialysis Quality Initiative (ADQI) Consensus Conference highlighted inflammatory injury as one of the key mechanisms in AKI pathogenesis44. The high SHAP importance of hs-CRP suggested that the model was sensitive to systemic inflammatory activity. CRP actively contributes to kidney injury by amplifying local inflammation, inhibiting tubular epithelial repair, and promoting fibrotic processes45. Therefore, the strong contribution of CRP in the model was consistent with its known role in driving early kidney damage. On the other hand, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm {iCa^{2+}}$$\end{document} plays a critical role in cellular signaling, myocardial contractility, and vascular regulation, all of which influence renal perfusion and inflammatory responses. The model’s reliance on \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm {iCa^{2+}}$$\end{document} is supported by clinical evidence showing that both hypocalcemia and hypercalcemia at hospital admission are associated with AKI46. Hypercalcemia can impair urinary concentrating ability and lead to intravascular volume depletion, thereby worsening renal hypoperfusion47, whereas low \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm {iCa^{2+}}$$\end{document} levels often reflect severe systemic illness or sepsis48,49, both of which predispose patients to kidney injury. Similarly, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm {HCO_3^-}$$\end{document}, a key indicator of acid-based status, showed a strong contribution to AKI risk in the model. Reduced \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm {HCO_3^-}$$\end{document} levels reflect metabolic acidosis, which has been linked to adverse renal outcomes and disease progression50. Mechanistically, metabolic acidosis increases renal ammoniagenesis and activates the alternative complement pathway, thereby promoting tubulointerstitial fibrosis. Consistent with this, observational studies have shown that lower \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm {HCO_3^-}$$\end{document} levels are associated with a higher risk of AKI, potentially in a nonlinear manner51. The importance of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm {HCO_3^-}$$\end{document} in the model suggests that disturbances in acid-base balance are a prominent early signal of renal vulnerability. This alignment between model behavior and known biological mechanisms supports the clinical interpretability of the predictions and provides physiologically meaningful signals for early risk stratification and potential intervention. In addition, early recognition of AKI combined with interpretable risk factor analysis can provide the potential preventive strategies, such as correcting calcium abnormalities or acid-based imbalance to enable early intervention. Previous clinical studies52–54 have suggested that sodium bicarbonate, as an acid-based regulating intervention, may exert a protective effect against AKI in specific patient populations. As a simple and low-cost therapy, sodium bicarbonate may represent a feasible option for preventing AKI in low-risk patients. Although the existing evidence remains heterogeneous across studies55, these findings offer important insights for auxiliary intervention, indicating that individualized management targeting modifiable risk factors may constitute a practical strategy for the early prevention and mitigation of AKI. From a nephrology and critical care perspective, the proposed framework provides an opportunity for earlier identification of high-risk ICU patients before overt AKI onset. The value of AKI recognition lies not merely in earlier labeling, but in enabling time-sensitive clinical decisions during a potentially reversible window of kidney injury progression56. Through the actionable time window, clinicians can reassess hemodynamic status, avoid nephrotoxic exposure, optimize fluid management, and adjust vasopressor or diuretic strategies before overt creatinine-defined AKI occurs. This temporal advantage is particularly relevant in ICU settings where renal insults are often multifactorial and dynamically evolving. Moreover, the interpretable feature attribution provided by SHAP may improve clinician understanding of individualized risk patterns and facilitate model acceptance in real-world critical care workflows57. Future prospective studies are warranted to determine whether model-guided early recognition can translate into improved renal and clinical outcomes. Such a study would integrate the model into ICU monitoring systems and generate time-stamped risk alerts. It would also evaluate not only predictive discrimination but also downstream clinical outcomes. These outcomes include time to intervention, incidence of stage progression, renal recovery rate, and need for renal replacement therapy. Importantly, evaluation should also consider clinician response and alert fatigue to ensure feasibility in real-world workflows. Limitation Although DCBM-Tri demonstrates promising performance in the early identification of AKI, further randomized controlled trials and prospective studies are required to comprehensively evaluate its clinical utility in terms of safety, effectiveness, and cost-effectiveness. Moreover, AKI is a multifactorial condition driven by the combined effects of multiple risk factors. Future high-quality clinical studies are needed to disentangle the independent contributions of these factors and to determine whether targeted interventions on specific risk factors can effectively reduce the incidence of AKI in the ICU. Although class imbalance reflects an inherent characteristic of real-world clinical data, its degree may vary depending on cohort definition and patient selection. The exclusion of community-acquired AKI and differences in specific clinical subgroups, such as cardiac surgery patients with a higher incidence of AKI, may influence the observed class distribution. These subgroup-specific imbalance patterns may affect model performance and generalizability. However, the investigation of such subgroup-specific settings is beyond the scope of the present study. Conclusions To address the severe class imbalance in early AKI risk prediction, we propose a dual-channel bilinear mapping-based contrastive learning framework, termed DCBM-Tri. By casting AKI risk prediction as a similarity learning problem, the model exploits dual-channel feature mappings to better capture patient relationships and improve the representation of high-risk minority cases. This framework enhances feature interaction and multi-view representation learning, while significantly improving sensitivity to high-risk AKI patients within the minority class. Furthermore, interpretability analysis identifies clinically meaningful and potentially modifiable risk factors, providing valuable guidance for early recognition and individualized prevention strategies.