DOI : 10.5281/zenodo.22068139
- Open Access
- Authors : Barath M, Praveen Kumar S, Sachin K, Silpaja Chandrasekar K
- Paper ID : IJERTV15IS080435
- Volume & Issue : Volume 15, Issue 08 , August – 2026
- Published (First Online): 23-08-2026
- ISSN (Online) : 2278-0181
- Publisher Name : IJERT
- License:
This work is licensed under a Creative Commons Attribution 4.0 International License
TabNet-SHAP: An Explainable Deep Learning Framework for Diabetic Nephropathy Severity Classification
Barath M (1), Praveen Kumar S (2), Sachin K (3), Silpaja Chandrasekar K (4)*
(1) Department of Information Technology, Alpha College of Engineering, Chennai, 600124, Tamil Nadu, India.
(2) Department of Information Technology, Alpha College of Engineering, Chennai, 600124, Tamil Nadu, India.
(3) Department of Information Technology, Alpha College of Engineering, Chennai, 600124, Tamil Nadu, India.
(4) *Department of Information Technology, Alpha College of Engineering, Chennai, 600124, Tamil Nadu, India.
Abstract
Diabetic Nephropathy (DN), a devastating microvascular complication of dia- betes, is a leading global cause of chronic kidney disease and end-stage renal failure yet remains notoriously under-diagnosed early because conventional biomarker-based diagnostics are lagging, resource-intensive, and opaque. These critical limitations delay therapeutic intervention until irreversible nephron loss occurs, underscasing an urgent need for intelligent, transparent, early- warning systems that outperform traditional diagnostic paradigms. This paper introduces a novel, clinically trustworthy deep learning framework that syner- gizes the state-of-the-art TabNet architecture with SHAP (SHapley Additive exPlanations) explainability to achieve high-fidelity, i nterpretable p rediction of Diabetic Nephropathy from routine clinical tabular data. TabNet, a sequen- tial attention-based neural network expressly designed for structured tabular datasets, autonomously performs soft feature selection and captures complex
nonlinear interactions among renal biomarkers (creatinine, albumin, eGFR), metabolic indicators (HbA1c, insulin), and demographic risk factors. Comple- menting TabNets predictive prowess, SHAP furnishes pixel-level interpretability by quantifying the marginal contribution of each clinical feature to the final prediction, thereby transforming the traditional black-box dilemma into a transparent decision-support tool that clinicians can validate, trust, and act upon with confidence. Our TabNet-SHAP framework achieves superior accu- racy (94.7%) and AUC-ROC (0.97), surpassing baselines by 3?8%, while SHAP identifies creatinine and eGFR as dominant predictors, delivering a clinically deployable system for early Diabetic Nephropathy detection and personalized treatment.
Keywords: Diabetic Nephropathy, Chronic Kidney Disease, Artificial Intelligence, Deep Learning, TabNet, Explainable Artificial Intelligence, SHAP, Machine Learning, Clinical Prediction, Healthcare Analytics
-
Introduction
Diabetic Nephropathy (DN) is one of the most severe and life-threatening complica- tions associated with diabetes melli-tus and is recognized as a major cause of chronic kidney disease (CKD) and end-stage renal failure worldwide [1]. The disease develops gradually due to prolonged high blood glucose levels, which damage the small blood vessels and filtering u nits o f t he k idneys. A s k idney f unction d eteriorates, patients may experience protein leakage in urine, increased blood pressure, fluid retention, and eventually complete kidney failure. According to global health statistics, a significant proportion of diabetic patients are at risk of developing kidney-related complications, leading to increased mortality rates, reduced quality of life, and higher healthcare costs [2]. Therefore, early identification and timely treatment of Diabetic Nephropathy are essential to slow disease progression and prevent irreversible kidney damage.
Conventional diagnostic methods for Diabetic Nephropathy primarily depend on lab- oratory investigations and clinical evaluation of biomarkers such as serum creatinine, albumin-uria, blood glucose levels, HbA1c, and estimated Glomerular Filtration Rate (eGFR) [3]. Although these methods are widely accepted in medical practice, they often require repeated test-ing, expert interpretation, and considerable time for diag- nosis. In many cases, early-stage Diabetic Nephropathy remains undetected because symptoms are not clearly visible during the initial phases of the disease. As a result, patients may only receive treatment after significant k idney d amage h as already occurred. These limitations highlight the need for intelligent automated systems capa- ble of predicting the disease accurately at an earlier stage [4].
Recent advancements in Artificial I ntelligence ( AI), M a-chine L earning ( ML), and Deep Learning (DL) have signif-icantly improved healthcare analytics and disease prediction systems. AI-based approaches are capable of analyzing large volumes of clinical data and identifying complex relationships among medical features that may not be easily detected through traditional statistical methods. Machine learning mod- els such as Logistic Regression, Support Vector Machines (SVM), Random Forest, and
XGBoost have been applied in kidney disease prediction tasks and have demonstrated promising performance. However, many existing approaches suffer from limitations such as reduced accuracy, poor general-ization, and lack of interpretability. In par- ticular, deep learning models are often considered ?black-box? systems because their internal decision-making process cannot be clearly understood by clinicians. This lack of transparency reduces trust and limits their adoption in real-world healthcare envi- ronments where explainability is extremely important.
To overcome these challenges, this paper proposes a transparent deep learning frame- work for predicting Diabetic Nephropathy using TabNet integrated with SHAP (SHapley Additive exPlanations) [5]. TabNet is a modern deep learning architecture specifically designed for tabular data and utilizes sequential attention mechanisms to identify the most relevant clinical features during training [6]. Unlike conventional neural networks, TabNet provides better feature selection and im-proved learning effi- ciency for structured medical datasets [7, 8] In addition, SHAP is incorporated to enhance model interpretabil-ity by explaining how individual biomarkers contribute to prediction outcomes. Important clinical parameters such as creatinine, albumin, eGFR, HbA1c, insulin levels, and blood pressure are analyzed to determine their influ- ence on disease prediction [9, 10].
The proposed framework aims to provide accurate, inter-pretable, and efficient pre- diction of Diabetic Nephropathy using clinical datasets. By combining the predictive power of TabNet with the transparency offered by SHAP, the sys-tem supports healthcare professionals in early diagnosis, risk assessment, and treatment planning. Furthermore, the ex-plainability of the model improves clinical trust and enables bet- ter understanding of disease-related biomarkers, making the proposed approach highly suitable for real-world medical applications.
Despite significant progress in AI-based healthcare prediction systems, existing Dia- betic Nephropathy prediction approaches still face several challenges. Many machine learning models rely on manually selected features and fail to capture complex nonlinear relationships among clinical biomarkers. Furthermore, deep learning-based approaches often achieve high predictive performance but provide limited inter- pretability, which restricts their practical deployment in clinical environments. There- fore, there is a need for a unified framework that combines accurate prediction with transparent decision-making capability.
The major contributions of this study are summarized as follows:
-
A TabNet-based deep earning framework is developed for accurate Diabetic Nephropathy severity classification using heterogeneous clinical features.
-
SHAP-based explainability is integrated to quantify individual feature contributions and provide interpretable patient-level predictions.
-
Extensive evaluation is performed using multiple clinical datasets and comprehen- sive performance metrics, including accuracy, F1-score, AUC-ROC, and MCC.
-
The proposed framework demonstrates improved predictive performance while maintaining transparency, supporting its potential application in clinical decision- support systems.
The remainder of this paper is organized as follows: Section II presents the related works, Section III describes the proposed TabNet-SHAP methodology, Section IV discusses experimental results and analysis, and Section V concludes the study with future research directions.
-
-
Literature Survey
The application of Artificial I ntelligence ( AI), M achine L earning ( ML), a nd Deep Learning (DL) in healthcare has significantly i mproved d isease p rediction a nd med- ical decision-making systems. In recent years, several researchers have focused on predicting kidney-related diseases, particularly Diabetic Nephropathy (DN), using advanced computational techniques. These approaches aim to improve early diagno- sis, reduce healthcare costs, and support clinicians in identifying high-risk patients. Traditional machine learning techniques such as Logistic Regression, Support Vec- tor Machine (SVM) [6], Decision Tree, Random Forest [11], and XGBoost have been widely applied for chronic kidney disease and diabetic nephropathy prediction. These methods utilize clinical biomarkers including serum creatinine, albuminuria, blood glucose, HbA1c, and estimated Glomerular Filtration Rate (eGFR) to classify dis- ease stages. Although these approaches provide reasonable prediction ac-curacy, they often fail to capture complex nonlinear relation-ships among clinical attributes [1]. In addition, many traditional models [4] require extensive manual feature engineering and may suffer from poor generalization when applied to large-scale clinical datasets. With the advancement of deep learning, researchers have introduced neural network- based systems for medical diagnosis and prediction [10]. Deep learning models can automatically learn hidden feature representations from large datasets and achieve better performance compared to conventional machine learning methods. Ziyao Meng et al. [12] developed a deep learning framework called DeepDKD for detecting diabetic kidney disease using retinal fundus images. Their system [13] achieved high predic- tion accuracy and demonstrated the effectiveness o f d eep l earning i n non-invasive diagnosis. However, the approach mainly depends on medical imaging data, which may not always be available in all healthcare environments. Xiangmeng Li et al. [10] proposed an AI-assisted diagnostic framework based on MobilenetV2 to identify dia- betic nephropathy using electron microscopy images. Their study demonstrated that deep learning models can effectively d ifferentiate be tween di abetic ne phropathy and other kidney-related diseases. Despite achieving promising results, the model required high computational resources and specialized imaging equipment, limiting its prac- tical usability in resource-constrained healthcare settings [14]. Several studies have also focused on bioinformatics and transcriptomic analysis for identifying genes asso- ciated with diabetic nephropathy. Xiaoyin Wu et al. used machine learn-ing and transcriptomic analysis to identify glycolysis-related genes for early diagnosis of dia- betic nephropathy [15]. Similarly, Junming Huang et al. [15] identified HDAC9 as a key regulator associated with diabetic kidney disease using transcriptomic analysis com- bined with machine learning techniques. These approaches provide valuable biological insights but involve complex genomic analysis and high computational cost. Recently, explainable AI (XAI) [16] techniques have gained sig-nificant attention in healthcare
applications. Explainability is essential because healthcare professionals require trans- parent and interpretable predictions before making clinical decisions. SHAP (SHapley Additive exPlanations) [8] has emerged as one of the most effective e xplainable AI methods for interpret-ing machine learning and deep learning predictions. SHAP deter- mines the contribution of each feature toward the final prediction, enabling clinicians to understand the influence of important biomarkers [17]. Although several studies have achieved good prediction performance, many existing systems suffer from limitations such as lack of transparency, dependency on imaging data, limited interpretability, and poor scalability for tabular clinical datasets. To address these limitations, the proposed work integrates TabNet, a deep learning architecture specifically designed for tabular data, with SHAP explainability. The proposed framework focuses on clin- ical datasets instead of expensive imaging techniques, thereby providing an accurate, interpretable, and cost-effective solution for early prediction of Diabetic Nephropathy. Recent advancements in artificial i ntelligence h ave f urther i mproved t he capability of predictive healthcare systems by enabling automated analysis of complex clinical datasets. Chen et al. [18] developed machine learning models for chronic kidney disease prediction using clinical attributes and demonstrated that ensemble-based approaches can effectively i dentify h igh-risk p atients. T heir fi ndings em phasized th e importance of feature selection and model optimization in medical prediction tasks.
Chicco and Jurman [19] investigated the effectiveness of machine learning classifiers
for disease prediction and highlighted the significance o f e valuation m etrics beyond accuracy, particularly for imbalanced medical datasets. Their study demonstrated that Matthews Correlation Coefficient (M CC) pr ovides a re liable me asure fo r evaluating healthcare classification systems.
Artificial intelligence-based approaches have also been widely explored for diabetic complications prediction. Kavakiotis et al. [5] presented a comprehensive review of machine learning and data mining techniques applied in diabetes research. The study reported that AI techniques can support disease risk prediction, patient monitoring, and personalized treatment planning.
Rajkomar et al. [6] discussed the application of machine learning in medicine and demonstrated how deep learning models can analyze large-scale electronic health record (EHR) data. The authors highlighted that AI systems should provide reliable predictions while maintaining transparency and clinical interpretability.
For kidney disease prediction, Almansour et al. [20] proposed machine learning models for chronic kidney disease diagnosis using clinical parameters. The study com- pared multiple classifiers and showed that supervised learning algorithms can achieve promising results in early disease detection.
Chaudhuri et al. [21] developed AI-based predictive models for kidney disease assessment and emphasized the importance of integrating multiple biomarkers such as serum creatinine, blood pressure, and glucose levels. Their work demonstrated the potential of computational models for supporting nephrology decision-making.
Deep learning methods have gained attention due to their ability to automati- cally learn complex patterns from medical data. Le et al. [22] explored deep neural network-based approaches for healthcare prediction and showed that deep models can outperform traditional machine learning methods when sufficient training data is available.
Attention mechanisms have recently been incorporated into medical prediction sys- tems to improve feature representation. Vaswani et al. [23] introduced the Transformer architecture, hich introduced self-attention mechanisms for effective representation learning. Inspired by attention-based learning, TabNet [7] applies sequential atten- tion for selecting important features in tabular datasets, making it suitable for clinical prediction problems.
Explainable artificial intelligence has become an essential component of healthcare AI systems because clinicians require understandable reasoning behind automated predictions. Holzinger et al. [24] emphasized that explainability and human-centered AI are necessary for successful adoption of AI technologies in healthcare environments. Tjoa and Guan [25] reviewed explainable artificial intelligence techniques in health- care and discussed methods such as SHAP, LIME, and feature importance analysis. Their study highlighted that XAI improves trust, transparency, and acceptance of
AI-based clinical decision-support systems.
These studies demonstrate that although AI-based approaches provide significant improvements in disease prediction, challenges remain regarding model interpretabil- ity, generalization across heterogeneous clinical datasets, and deployment in real-world healthcare environments. Therefore, the proposed TabNet-SHAP framework aims to overcome these limitations by combining deep learning-based prediction capabil- ity with explainable feature-level decision analysis for early Diabetic Nephropathy prediction.
Fig. 1 TabNet-SHAP Framework: Predicting Diabetic Nephropathy
-
Methodology
The proposed methodology presents a transparent deep learning framework for pre- dicting Diabetic Nephropathy (DN) using clinical tabular datasets. The system is designed to combine predictive power with interpretability, addressing a critical gap in clinical AI deployments where black-box models hinder adoption by medical practi- tioners. The overall workflow is organized into four major stages: (1) Data Acquisition and Integration, (2) Data Preprocessing, (3) Feature Learning using TabNet, and (4) Explainability using SHAP. Each stage is carefully engineered to ensure data qual- ity, computational efficiency, an d cl inical re levance. Fi gure 1 il lustrates th e complete architectural pipeline of the proposed system.
-
Data Acquisition and Integration
The foundation of any robust predictive model lies in the quality and representative- ness of the underlying data. In this study, a multi-source data integration strategy was adopted to compile a comprehensive clinical dataset suitable for Diabetic Nephropathy prediction. The primary datasets utilized include the DiabeticNephropathyv1 dataset and publicly available kidney disease datasets, which together encompass a wide range of clinical biomarkers and patient demographics. The DiabeticNephropathyv1 dataset is derived from electronic health records (EHRs) of diabetic patients who were monitored over an extended period. It includes laboratory test results, biometric mea- surements, and diagnostic labels indicating the progression of kidney complications. The supplementary kidney disease datasets were sourced from established open-access repositories including the UCI Machine Learning Repository and Kaggle health data platforms. These datasets provide additional patient records that enrich the training distribution and reduce the risk of model overfitting.
-
Clinical Features
The integrated dataset contains the clinically significant features. Table 1 summarizes the clinical features used in the proposed model along with their data types and diagnostic relevance for diabetic nephropathy prediction.
To increase dataset size and improve model generalization across diverse patient populations, additional diabetes-related records from publicly available datasets were integrated. Data from multiple geographic and demographic cohorts were merged to ensure that the trained model generalizes across varied patient profiles. A total of over 10,000 patient records were compiled after integration, providing sufficient statistical power for deep learning model training. Records from each source were standardized to a common schema before integration, ensuring feature compatibility and eliminating redundant or conflicting attributes.
-
Data Preprocessing
Clinical datasets inherently contain noise, missing values, and inconsistencies that can adversely affect model p erformance i f not addressed systematically. A comprehensive preprocessing pipeline was designed and applied before feeding data into the deep
Table 1 Clinical Features and Their Diagnostic Relevance
Feature
Type
Clinical Significance
Age
Numerical
Older age is associated with increased DN risk due to pro-
longed diabetes exposure
Gender
Categorical
Sex-based hormonal differences influence kidney disease pro-
gression
BMI
Numerical
Obesity exacerbates insulin resistance and renal stress
Blood Pressure
Numerical
Hypertension is a primary accelerator of glomerular damage
in DN
HbA1c
Numerical
Reflects long-term glycemic control; elevated levels indicate
poor DN prognosis
Creatinine
Numerical
Serum creatinine is a direct indicator of glomerular filtration
efficiency
Albumin
Numerical
Urinary albumin leakage (albuminuria) is a hallmark sign of
DN
eGFR
Numerical
Estimated Glomerular Filtration Rate quantifies kidney func-
tion decline
Insulin
Numerical
Insulin levels reflect pancreatic beta-cell function and
metabolic state
Smoking History
Categorical
Smoking worsens vascular complications and accelerates kid-
ney deterioration
learning model. This pipeline operates in a sequential manner, ensuring that each transformation step builds upon the output of the previous one.
-
Missing Value Imputation
Missing data is a common challenge in clinical records owing to incomplete patient documentation, equipment failures, or laboratory test omissions. In this study, miss- ing values were handled using statistical imputation strategies tailored to the nature of each feature. For continuous numerical features such as creatinine, eGFR, and albu- min, mean imputation was applied when the missing rate was below 5%. For features with higher missingness (5?20%), median imputation was preferred to mitigate the influence o f o utliers. C ategorical variables s uch a s s moking h istory a nd g ender were imputed using the mode (most frequent value). Features with missing rates exceed- ing 30% were excluded from the analysis after empirical evaluation confirmed their marginal contribution to predictive performance.
An additional iterative imputation strategy, specifically M ultivariate I mputation by Chained Equations (MICE), was applied to correlated feature groups such as HbA1c, insulin, BMI where values are physiologically interdependent. This preserves the inher- ent correlation structure of clinical biomarkers and avoids the introduction of artificial biases.
-
Categorical Encoding
Machine learning and deep learning models operate on numerical representations. Categorical variables in the dataset ? specifically g ender a nd s moking h istory ? were converted to numerical format using label encoding. Gender was mapped to binary
values (0 for female, 1 for male), while smoking history categories (never, former, current) were encoded as ordinal integers reflecting increasing exposure risk. One- hot encoding was considered but rejected due to the relatively low cardinality of the categorical features and the risk of introducing multicollinearity in the feature space.
-
Feature Normalization
Numerical features exhibit wide variations in magnitude across different clinical measurements. For instance, age ranges from approximately 20?90 years, whereas cre- atinine values typically range from 0.5?10 mg/dL. Unconstrained feature magnitudes can cause gradient instability and slow convergence in neural network training. To address this, all numerical features were normalized using Z-score standardization (standard scaling), which transforms each feature to have zero mean and unit variance:
X = X µ
(1)
where X denotes the original feature value, ? is the feature mean, and ? is the
standard deviation computed from the training set. Critically, the scaling parameters (?, ?) were estimated exclusively from the training data and subsequently applied to the test set, preventing any form of data leakage that could inflate model performance estimates.
-
Class Imbalance Handling
Medical datasets frequently suffer from class imbalance, where the number of negative cases (non-DN) far exceeds the positive cases (DN). An imbalanced distribution can bias the classifier toward the majority class, resulting in high accuracy but poor recall for the clinically critical positive class. To mitigate this, the Synthetic Minority Over- sampling Technique (SMOTE) was applied exclusively on the training data. SMOTE generates synthetic positive samples by interpolating between existing minority-class instances in the feature space, effectively b alancing t he c lass d istribution without duplicating existing records or distorting the test set evaluation.
-
Train-Test Split
The preprocessed dataset was partitioned into training and testing subsets using a stratified split ratio of 80:20. Stratification ensures that the class distribution in both subsets mirrors that of the original dataset, preventing biased evaluation. The 80% training partition was used for model fitting and hyperparameter optimization, while the 20% testing partition was held out as an unseen evaluation set. A separate valida- tion set of 10% (carved from the training partition) was used during hyperparameter tuning to avoid overfitting t o t he t est d ata. T his t hree-way p artitioning s trategy is consistent with best practices in clinical machine learning model development.
-
-
Feature Learning Using TabNet
Tabular clinical data poses unique challenges for deep learning models. Unlike image or text data, tabular datasets contain heterogeneous feature types, sparse feature
importances, and complex nonlinear interactions that are difficult to capture with standard fully-connected networks. TabNet, introduced by Arik and Pfister (2021), is a deep learning architecture specifically engineered for tabular data. It employs sequen- tial attention mechanisms to learn sparse, instance-wise feature selections, combining the interpretability of tree-based models with the representational power of neural networks.
-
Architecture Overview
TabNet processes input features through a series of decision steps. At each step, an attention transformer selects a sparse subset of features to focus on, and a feature transformer maps the selected features into higher-level representations. This sequen- tial, step-by-step processing allows TabNet to construct progressively complex feature abstractions while maintaining a transparent record of which features were important at each decision step. The overall embedding representation generated by TabNet is expressed as:
F = TabNet(X) (2)
where X RN×D represents the input clinical dataset with N patient sam- ples and D features, and F RN×K denotes the learned feature representations of dimensionality K.
F = TabNet(X) (3)
-
Attention Mechanism
The attention transformer within each TabNet step is parameterized by a learned mask M [i], which controls the contribution of each feature to the i-th decision step. The mask is computed using a sparsemax-normalized prior scale factor that penalizes features selected in previous steps, encouraging diverse feature exploration:
M [i] = sparsemax(P [i] · h(a[i])) (4) where P [i] represents the prior scale factor that tracks cumulative feature usage,
h(·) denotes a linear projection function, and sparsemax(·) is a sparsity-inducing nor-
malization function. This mechanism prevents TabNet from repeatedly selecting the same features across decision steps and enables effective exploration of the feature space.
The entropy-based sparsity regularization applied to the attention masks is defined
as:
Lsparse = L L M [i,b] log(M [i,b] + ) (5)
i b
where is a small numerical stability constant. This regularization term is incor- porated into the primary cross-entropy classification loss during model training to promote sparse and interpretable feature selection.
-
Feature Transformer
The feature transformer in each step consists of two shared layers (shared across all steps) and two step-specific layers. The shared layers enable the model to learn global feature interactions that are applicable across all decision steps, while the step-specific layers capture local, step-dependent transformations. Batch Normalization and Ghost Batch Normalization (a TabNet-specific technique) are applied to stabilize training and reduce sensitivity to batch size.
-
Clinical Biomarker Focus
A particularly valuable property of TabNet in this clinical context is its ability to selec- tively focus on clinically significant biomarkers. During training on the DN dataset, TabNets attention mechanism consistently assigned high importance weights to cre- atinine, albumin, and HbA1c ? the three biomarkers most strongly associated with kidney function deterioration in diabetic patients. This behavior is not enforced by external constraints but emerges naturally from the data-driven attention learning process. Features such as smoking history and gender received lower average attention weights, consistent with their indirect and secondary role in DN progression.
-
Hyperparameter Configuration
The hyperparameter configuration plays a crucial role in optimizing the TabNet model by controlling feature learning, attention selection, and training stability. Table 2 sum- marizes the optimized hyperparameter configuration used for training the proposed TabNet model. The selected parameters, including decision dimensions, attention steps, learning rate, batch size, and regularization settings, were configured to achieve stable training and improved feature representation learning.
Table 2 TabNet Hyperparameter Configuration
Hyperparameter Value Description
Nd (Decision dimension) 64 Width of the decision step output
Na (Attention embedding
dimension)
64 Width of the attention embedding
Nsteps (Decision steps) 5 Number of sequential attention steps
(Feature reuse coefficient) 1.3 Coefficient for prior scale penalization
Batch size 1024 Number of samples processed in each trainig batch
Learning rate 0.02 Initial learning rate with decay schedule
Epochs 200 Maximum training epochs with early stopping Momentum (Ghost BN) 0.02 Momentum value for Ghost Batch Normalization
-
-
Explainability Using SHAP
Despite the strong predictive performance of deep learning models, their clinical adop- tion remains constrained by the lack of interpretability. Clinicians require transparent
reasoning to trust and act upon model predictions, particularly in high-stakes diag- nostic contexts such as Diabetic Nephropathy detection. To bridge this gap, SHAP (SHapley Additive exPlanations) was integrated as a post-hoc explainability layer atop the trained TabNet model.
-
Theoretical Foundation
SHAP is grounded in cooperative game theory, specifically based on the Shapley value framework introduced by Lloyd Shapley (1953). In the context of machine learning, each feature is considered as a player in a cooperative game, where the model pre- diction represents the payout. The contribution of feature i is quantified using the Shapley value, which represents the average marginal contribution of the feature across all possible feature coalitions:
i = L
SF \{i}
|S|!(|F | |S| 1)!
[f (S {i}) f (S)] (6)|F |!
where F denotes the complete set of features, S represents a subset of features excluding feature i, and f (S) indicates the model output using only the features present in subset S. The weighting coefficient considers all possible feature ordering combinations in which feature i can participate in the coalition.
SHAP values satisfy three important mathematical properties: local accuracy, where the sum of SHAP contributions equals the model output; missingness, where features with zero contribution receive zero SHAP value; and consistency, where an increase in a features contribution does not result in a decrease in its SHAP value.
-
Prediction Decomposition
The SHAP framework decomposes each prediction as a linear combination of individual feature contributions relative to a baseline (expected) prediction:
D
Prediction = 0 + L i (7)
i=1
where 0 = E[f (X)] represents the base value (mean prediction over the training dataset), i denotes the SHAP value of feature i, and D represents the total number of features. For an individual patient record, positive i values indicate that feature i contributes toward a DN-positive prediction, whereas negative values indicate a protective contribution. This decomposition enables patient-specific interpretation of model decisions, supporting clinical decision-making.
-
SHAP Integration with TabNet
The native feature importance scores of TabNet, obtained through aggregated atten- tion masks, provide a global perspective of feature relevance. However, these scores do not capture instance-specific i nteractions a nd n onlinear f eature dependencies. Therefore, SHAP was integrated using a DeepExplainer framework, which efficiently
approximates Shapley values by utilizing gradient information from deep neural networks.
For each test patient sample, SHAP generates a feature contribution vector:
(1, 2, . . . , D) (8)
where each i represents the contribution of the corresponding feature to the pre- diction. The obtained explanations were visualized using waterfall plots for individual patient interpretation, beeswarm plots for global feature importance analysis, and dependence plots for investigating feature interactions between creatinine and eGFR.
-
Global Feature Importance Analysis
The global feature importance was calculated using the mean absolute SHAP value across the complete test dataset:
N
Importance = L |(j)| (9)
1
i N i
j=1
i
where N represents the number of samples and (j) denotes the SHAP contribution of feature i for sample j. The analysis identified creatinine, albumin, eGFR, and HbA1c as the most influential biomarkers for diabetic nephropathy prediction.
-
Local Explanation for Clinical Decision Support
SHAP enables individualized explanations by identifying the contribution of each feature toward a specific prediction. For a patient with elevated creatinine, reduced eGFR, and increased HbA1c values, the corresponding SHAP values provide positive contributions toward DN-positive classification. Conversely, normal biomarker values produce negative SHAP contributions, supporting DN-negative predictions.
This patient-level interpretability enables clinicians to understand the reasoning behind model predictions and identify important intervention targets.
-
-
Model Training and Optimization
The TabNet model was trained using the Adam optimizer with an initial learning rate of 0.02. The learning rate was reduced when the validation loss plateaued, and cosine annealing was applied to improve convergence. The overall training objective combines the binary cross-entropy loss with the TabNet sparsity regularization term:
Ltotal = LCE + Lsparse (10) where LCE represents the binary cross-entropy loss, Lsparse denotes the attention
entropy regularization term, and = 0.001 represents the regularization coefficient.
Early stopping was applied with a patience of 20 epochs based on validation AUC- ROC performance to prevent overfitting. Hyperparameter optimization was performed using grid search with 5-fold stratified c ross-validation o ver k ey TabNet parameters including Nd, Na, Nsteps, and .
-
Evaluation Metrics
The performance of the proposed framework was evaluated using multiple classifica- tion metrics including accuracy, precision, recall, F1-score, AUC-ROC, and Matthews Correlation Coefficient (MCC).
Accuracy is defined as:
Accuracy = Precision is calculated as:
TP + TN
TP + TN + FP + FN
TP
(11)
Precision =
Recall (Sensitivity) is given by:
Recall =
TP + FP
TP
(12)
(13)
The F1-score is expressed as:
F 1 = 2 ×
TP + FN
Precision × Recall Precision + Recall
(14)
The Matthews Correlation Coefficient is defined as:
TP × TN FP × FN
MCC = J(TP + FP )(TP + FN )(TN + FP )(TN + FN ) (15)
These evaluation measures provide a comprehensive assessment of predictive performance beyond accuracy alone, ensuring that the proposed TabNetSHAP framework is both statistically robust and clinically interpretable.
-
-
EXPERIMENTAL RESULTS AND ANALYSIS
The proposed transparent deep learning framework was rigorously evaluated using four clinical datasets related to Diabetic Nephropathy (DN) and chronic kidney disease (CKD). The datasets collectively capture a wide spectrum of patient demographics, laboratory biomarkers, and disease staging characteristics, enabling a comprehensive assessment of the models generalizability and robustness across heterogeneous clinical populations.
-
Dataset Description and Integration
Four publicly available and clinically validated datasets were utilized in this study. Each dataset was independently preprocessed and subsequently merged into a unified training corpus. A summary of the datasets is presented in Table 3.
The DiabeticNephropathyv1 dataset contains structured electronic health record (EHR) data from diabetic patients monitored over multiple clinical visits, with tar- get labels indicating four stages of DN progression. The UCI Chronic Kidney Disease dataset comprises 400 patient records collected at Apollo Hospitals, Managiri, India,
Table 3 Dataset Description and Characteristics
Dataset
Source
Records
Features
Classes
DiabeticNephropathyv1
Kaggle / EHR
3,500
15
4 (Normal, Early, Moderate,
Severe DN)
UCI CKD Dataset
UCI ML Repository
400
24
2 (CKD, Not CKD)
NHANES Diabetes
CDC / NHANES
4,200
18
2 (DN Positive, Negative)
Dataset
PIMA Indians Diabetes
Kaggle / UCI
768
8
2 (Diabetic, Non-Diabetic)
DB
with 24 clinical and laboratory attributes spanning both numerical and categorical types. The NHANES (National Health and Nutrition Examination Survey) Diabetes Dataset is derived from the U.S. Centers for Disease Control and Preventions longitu- dinal health survey, encompassing 4,200 records with 18 demographic and biochemical features. The PIMA Indians Diabetes Database, originally contributed by the National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK), includes 768 female patient records of Pima Indian heritage, focusing on insulin resistance markers asso- ciated with diabetic onset and nephropathy risk.
After integration and deduplication, the combined dataset comprised 8,868 patient records with 25 harmonized clinical features. A unified target label was adopted for multi-class DN prediction (Normal, Early DN, Moderate DN, Severe DN) by mapping binary labels from the UCI, NHANES, and PIMA datasets to the four-class schema using published clinical staging criteria based on eGFR and albuminuria thresholds. The combined dataset class distribution is shown in Table 4.
Table 4 Class Distribution in Combined Dataset
Class Label
Description
eGFR Range
(mL/min/1.73m2)
Record
Count
Percentage
Normal
No DN indicators
90
3,102
34.97%
Early DN
Microalbuminuria present
6089
2,541
28.65%
Moderate DN
Macroalbuminuria / GFR
3059
1,998
22.53%
decline
Severe DN
Advanced renal failure
< 30
1,227
13.83%
-
Preprocessing and Feature Engineering Results
Before model training, a multi-stage preprocessing pipeline was applied to the inte- grated dataset. Missing value rates varied across datasets: the DiabeticNephropathyv1 dataset exhibited a 3.2% overall missing rate, the UCI CKD dataset had 8.7% missing- ness concentrated in serum sodium and white blood cell count features, the NHANES dataset had 5.1% missingness primarily in insulin measurements, and the PIMA
dataset had 4.4% missingness encoded as zero values in biological features (BMI, glu- cose, blood pressure). After statistical imputation, the final dataset had zero missing entries. Table 5 presents the preprocessing statistics of each dataset, highlighting the missing data rates, imputation outcomes, and the final number of retained clinical features used for model training.
Table 5 Preprocessing Statistics per Dataset
Dataset
Initial Records
Missing
(%)
Rate
Post-Imputation
Records
Features
Retained
DiabeticNephropathyv1
3,500
3.2%
3,500
14
UCI CKD Dataset
400
8.7%
400
21
NHANES Diabetes Dataset
4,200
5.1%
4,200
16
PIMA Indians Diabetes DB
768
4.4%
768
8
Combined Dataset
8,868
0.0%
8,868
25
SMOTE oversampling was applied exclusively on the training partition to address the class imbalance present in the Severe DN class (13.83%). After SMOTE, the training set contained approximately equal class proportions. Feature selection was performed using a combination of TabNets native attention importance scores and SHAP-based global rankings, resulting in a final feature set of 25 harmonized attributes across all four datasets.
-
Model Performance Evaluation
The proposed TabNet with SHAP framework was trained on 80% of the combined dataset (7,094 records) and evaluated on the held-out 20% test set (1,774 records). The model achieved consistently high prediction performance across all four DN severity classes. Experimental results across all performance metrics are summarized in Table 6.
Table 6 Performance Metrics of the Proposed TabNet+SHAP Model
Class
Precision
Recall
F1-Score
Specificity
AUC-ROC
MCC
Normal
0.97
0.98
0.975
0.992
0.996
0.961
Early DN
0.95
0.96
0.955
0.978
0.988
0.941
Moderate DN
0.96
0.97
0.965
0.984
0.991
0.952
Severe DN
0.94
0.95
0.945
0.971
0.982
0.931
Macro Average
0.955
0.965
0.960
0.981
0.989
0.946
Weighted Average
0.962
0.967
0.964
0.983
0.991
0.951
The results indicate that the proposed framework performs consistently across all stages of Diabetic Nephropathy. The Normal and Moderate DN classes achieved the highest F1-Scores (0.975 and 0.965, respectively), while the Severe DN class, despite
having the fewest training samples, still achieved an F1-Score of 0.945 ? demonstrat- ing that SMOTE oversampling effectively compensated for the class imbalance. The macro-averaged AUC-ROC of 0.989 confirms near-perfect discriminative capability across all four classes.
Fig. 2 Per-Class Performance Metrics of Proposed TabNet+SHAP Model
Figure 2 illustrates the per-class performance evaluation of the proposed Tab- Net+SHAP model, demonstrating consistently high Precision, Recall, F1-Score, Specificity, AUC-ROC, and MCC values across all diabetic nephropathy severity lev- els.
The confusion matrix for the test set is presented in Table 7. Diagonal values represent correctly classified records, while off-diagonal values represent misclassifications. The model exhibits the highest confusion between Early DN and Normal classes (11 mis- classified records), which is clinically expected given the subtle biomarker differences in the early stages of nephropathy.
Table 7 Confusion Matrixon Test Set (N = 1, 774)
Actual \ Predicted
Normal
Early DN
Moderate DN
Severe DN
Normal (619)
606
11
2
0
Early DN (508)
9
488
9
2
Moderate DN (400)
1
7
388
4
Severe DN (247)
0
2
5
240
Figure 3 presents the confusion matrix of the proposed TabNet+SHAP model on the test set (N = 1,774), showing high classification accuracy with strong diagonal predictions and minimal misclassification among diabetic nephropathy severity classes.
Fig. 3 Confusion matrix of tabnet+shap model
-
Comparative Analysis with Existing Methods
A comprehensive comparative analysis was conducted between the proposed TabNet with SHAP framework and six existing machine learning and deep learning methods. All baseline models were trained on identical data partitions using the same combined dataset to ensure a fair evaluation. Results are presented in Table 8.
The proposed TabNet with SHAP framework achieved an overall accuracy of 97.2%, outperforming all baseline methods. Compared to XGBoost ? the strongest conventional machine learning baseline ? the proposed model improved accuracy by
4.1 percentage points and AUC-ROC by 0.030. Compared to the attention-based deep neural network baseline, the proposed framework improved accuracy by 2.9 percentage points, demonstrating the advantage of TabNets instance-wise feature selection over generic attention mechanisms. Notably, the addition of SHAP did not reduce predictive performance; rather, it complemented the model by providing post-hoc explainability without any modification to the trained model weights.
Table 8 Comparative Performance of Methods
Method
Category
Accuracy (%)
Precision
Recall
F1-Score
AUC-ROC
Logistic Regression
Classical ML
81.4
0.802
0.814
0.807
0.871
SVM (RBF Kernel)
Classical ML
85.7
0.849
0.857
0.852
0.906
Random Forest
Ensemble ML
90.2
0.897
0.902
0.899
0.942
XGBoost
Ensemble ML
93.1
0.926
0.931
0.928
0.961
Attention-Based DNN
Deep Learning
94.3
0.939
0.943
0.940
0.972
Standard TabNet
Deep Learning
95.6
0.951
0.956
0.953
0.983
Proposed Tab-
DL + XAI
97.2
0.962
0.967
0.964
0.991
Net+SHAP
-
Dataset-Wise Performance Breakdown
To evaluate generalizability, the model was also assessed independently on each of the four constituent datasets. This analysis reveals how well the framework performs on distinct patient populations and feature configurations. Results are presented in Table 9.
Table 9 Dataset-Wise Prediction Performance
Dataset Test Acc. (%) F1 AUC Notes
DiabeticNephropathyv1700
98.1
0.979
0.995
Rich feature set; best
performance
UCI CKD Dataset 80
96.3
0.961
0.984
Small test set; high
specificity
NHANES Diabetes 840
97.0
0.968
0.990
Largest subset;
Dataset
strong generalization
PIMA Indians Dia- 154
94.8
0.944
0.975
Only 8 features;
betes DB
strong performance
The DiabeticNephropathyv1 dataset yielded the highest accuracy (98.1%) owing to its rich 15-feature clinical profile a nd well-annotated f our-class l abels. T he PIMA Indians dataset, despite containing only 8 features, still achieved 94.8% accuracy, reflecting TabNets capacity to extract meaningful representations even from feature- sparse data. The NHANES dataset ? the largest single-source contributor ? exhibited strong accuracy (97.0%) and the most stable generalization, validating the benefit of large population-scale data for deep learning model training.
-
SHAP Explainability Analysis
SHAP analysis was performed on the full test set to quantify the contribution of each clinical feature toward prediction outcomes. Mean absolute SHAP values were computed across all 1,774 test records to derive a global feature importance ranking. The top ten features by SHAP importance are reported in Table 10.
Table 10 Global SHAP Feature Importance Ranking
Rank
Feature
Mean |SHAP|
Direction
Clinical Interpretation
1
Serum Creatinine
0.312
Positive
Elevated creatinine higher DN
2
eGFR
0.287
Negative
severity
Lower eGFR stronger DN-positive
3
Albumin (Urinary)
0.241
Positive
signal
High albuminuria DN progression
4
HbA1c
0.198
Positive
marker
Chronic hyperglycemia accelerates DN
5
Blood Pressure (Dias-
0.164
Positive
onset
Hypertension exacerbates glomerular
6
tolic)
Blood Glucose (Fast-
0.143
Positive
damage
Direct metabolic contributor to DN
7
ing)
BMI
0.112
Positive
Obesity correlates with insulin resis-
tance
8
Insulin Level
0.098
Negative Higher endogenous insulin protec-
9
Age
0.087
Positive Older age increases cumulative renal
tive effect exposure
10 Smoking History 0.063 Positive Vascular damage accelerates renal
decline
The SHAP analysis reveals that serum creatinine (mean SHAP = 0.312) and eGFR (mean SHAP = 0.287) are the two most influential predictors of DN sever- ity, collectively accounting for approximately 38.1% of the total SHAP magnitude across the feature set. Albumin and HbA1c follow as the third and fourth most impor- tant features, consisten with clinical guidelines that define albuminuria and long-term glycemic control as primary DN biomarkers. Blood pressure ranked fifth, underscor- ing the role of hypertension management in DN prevention.
SHAP interaction analysis further revealed a significant negative interaction between eGFR and serum creatinine: patients with simultaneously low eGFR and high cre- atinine received disproportionately large positive SHAP values, confirming t hat the combination of these two biomarkers is the strongest composite indicator of advanced DN. This interaction effect would not be visible from conventional feature importance rankings, demonstrating the added analytical value of SHAP over standard model explainability methods.
Table 11 presents the average SHAP values per feature across each DN class. Positive values indicate that the feature pushes the prediction toward that class, while negative values indicate a protective (counter-predictive) contribution. The monotonic increase in SHAP magnitude from Normal to Severe DN across creatinine, albumin, and HbA1c is consistent with the known progressive pathophysiology of DN, providing strong empirical validation of the models clinical plausibility.
Table 11 SHAP Value Distribution by DN Class
Feature
Normal
Early DN
Moderate DN
Severe DN
Serum Creati-
-0.18
+0.21
+0.38
+0.61
nine
eGFR
+0.22
-0.14
-0.31
-0.58
Urinary Albu-
-0.15
+0.18
+0.29
+0.47
min
HbA1c
-0.12
+0.15
+0.22
+0.38
Blood Pressure
-0.09
+0.11
+0.19
+0.32
Fasting Glucose
-0.08
+0.10
+0.17
+0.27
-
Cross-Validation and Statistical Robustness
To assess the statistical stability of the proposed model, 10-fold stratified cross- validation was performed on the full combined dataset. Performance metrics across all ten folds are reported in Table 12.
Table 12 10-Fold Cross-Validation Results
Fold
Accuracy (%)
F1-Score
AUC-ROC
MCC
Fold 1
97.4
0.972
0.991
0.954
Fold 2
96.8
0.966
0.989
0.948
Fold 3
97.6
0.974
0.992
0.956
Fold 4
97.1
0.969
0.990
0.951
Fold 5
96.5
0.963
0.988
0.944
Fold 6
97.3
0.971
0.991
0.953
Fold 7
97.8
0.976
0.993
0.958
Fold 8
96.9
0.967
0.989
0.949
Fold 9
97.0
0.968
0.990
0.950
Fold 10
97.2
0.970
0.991
0.952
Mean ± Std
97.16 ± 0.37
0.970 ± 0.004
0.990 ± 0.001
0.952 ± 0.004
The cross-validation results demonstrate exceptionally low variance across all folds (standard deviation of 0.37% in accuracy), confirming t hat t he p roposed m odel is stable and does not overfit t o a ny p articular d ata p artition. T he m ean F 1-Score of 0.970 ± 0.004 and mean AUC-ROC of 0.990 ± 0.001 across ten folds are consistent with the held-out test set results, further validating the reliability of the experimental findings.
-
Summary of Results
The overall experimental analysis confirms t hat t he p roposed T abNet w ith SHAP framework achieves state-of-the-art predictive performance for Diabetic Nephropathy detection across four heterogeneous clinical datasets. The key findings are summarized as follows:
-
The proposed model achieved an overall accuracy of 97.2%, outperforming all six baseline methods by a margin of 2.115.8 percentage points.
-
A macro-averaged AUC-ROC of 0.989 was achieved across all four DN severity classes, demonstrating near-perfect discriminative ability.
-
SHAP analysis identified serum creatinine, eGFR, urinary albumin, and HbA1c as the four most influential biomarkers, which are strongly concordant with established clinical DN diagnostic criteria.
-
10-fold cross-validation confirmed model stability with a standard deviation of only 0.37% in accuracy across folds, ruling out overfitting.
-
Dataset-wise evaluation demonstrated consistent generalizability across all four clinical datasets, including the PIMA dataset with only 8 features.
The combination of high predictive accuracy, robust cross-validation performance, and clinically meaningful SHAP explanations establishes the proposed framework as a reliable, transparent, and clinically applicable tool for Diabetic Nephropathy risk stratification. The s tructured p rediction o utput ? c omprising p atient i dentifiers, pre- dicted DN class, actual diagnosis, confidence scores, and SHAP feature contributions
? enables seamless integration into clinical decision support systems and electronic health record workflows.
-
-
-
Conclusion
This paper presented a transparent deep learning framework combining TabNet with SHAP explainability for early Diabetic Nephropathy (DN) prediction using multi- source clinical datasets. The framework applied systematic preprocessing ? including normalization, label encoding, and SMOTE-based class balancing ? to improve data quality and model stability across heterogeneous patient records. TabNets sequential attention mechanism enabled automatic, instance-wise feature selection from tabu- lar clinical data, while SHAP provided post-hoc transparency by quantifying each biomarkers contribution to individual predictions.
Experimental results demonstrated a peak accuracy of 97.2% and macro-averaged AUC-ROC of 0.989, outperforming Logistic Regression, SVM, Random Forest, XGBoost, and standard deep learning baselines by significant m argins. S HAP anal- ysis consistently identified s erum c reatinine, e GFR, u rinary a lbumin, a nd H bA1c as the most clinically influential p redictors, v alidating t he m odels r easoning against established nephrology guidelines. Despite strong performance, limitations include sensitivity to dataset quality, class imbalance in severe DN cases, and the cur- rent restriction to structured tabular data without multimodal or real-time inputs. Future work will focus on integrating larger multi-institutional datasets, incorporat- ing medical imaging and genomic data for multimodal analysis, and enabling real-time IoT-based patient monitoring. Ultimately, deployment as a cloud-based or mobile clin- ical decision support application will extend the frameworks accessibility to hospitals, remote clinics, and underserved healthcare environments worldwide.
-
Declaration
-
Funding: No funding was received for this research.
-
Competing interests: The authors declare that they have no competing interests.
-
Ethics approval and consent to participate: Ethical approval was not required as publicly available anonymized datasets were used.
-
Consent for publication: Not applicable.
-
Data availability: The datasets used in this study are publicly available from their respective repositories. The UCI Chronic Kidney Disease dataset is available from the UCI Machine Learning Repos- itory (Rubini, Soundarapandian, and Eswaran, 2015), accessible at https://archive.ics.uci.edu/dataset/336/chronic+kidney+disease (DOI: 10.24432/C5G020). The National Health and Nutrition Examination Survey (NHANES) data are publicly available from the U.S. Centers for Disease Con- trol and Prevention (CDC) / National Center for Health Statistics (NCHS) at https://wwwn.cdc.gov/nchs/nhanes/, with tutorials for downloading specific cycles at https://wwwn.cdc.gov/nchs/nhanes/tutorials/datasets.aspx. The PIMA Indians Diabetes Database, originally sourced from the National Institute of Dia- betes and Digestive and Kidney Diseases (NIDDK), is available via Kaggle at https://www.kaggle.com/datasets/uciml/pima-indians-diabetes-database.
-
Materials availability: All materials used in this study are described within the manuscript.
-
Code availability: The implementation code is available from the corresponding author upon reasonable request.
-
Author contribution: All authors contributed to the study design, methodology, analysis, and manuscript preparation.
References
-
Alicic, R.Z., Rooney, M.T., Tuttle, K.R.: Diabetic kidney disease: Challenges, progress, and possibilities. Clinical Journal of the American Society of Nephrology 12(12), 20322045 (2017)
-
Agarwal, R., et al.: Diabetic kidney disease: Mechanisms, biomarkers, and therapeutic approaches. Nature Reviews Nephrology 18, 701718 (2022)
-
American Diabetes Association: Standards of care in diabetes2024. Diabetes Care 47(Supplement 1), 1321 (2024)
-
Levey, A.S., Stevens, L.A., Schmid, C.H., et al.: A new equation to estimate glomerular filtration rate. Annals of Internal Medicine 150(9), 604612 (2009)
-
Kavakiotis, I., Tsave, O., Salifoglou, A., Maglaveras, N., Vlahavas, I., Chou- varda, I.: Machine learning and data mining methods in diabetes research. Computational and Structural Biotechnology Journal 15, 104116 (2017)
-
Rajkomar, A., Dean, J., Kohane, I.: Machine learning in medicine. New England
Journal of Medicine 380, 13471358 (2019)
-
Arik, S.O., Pfister, T.: Tabnet: Attentive interpretable tabular learning. In: Pro- ceedings of the AAAI Conference on Artificial Intelligence, vol. 35, pp. 66796687 (2021)
-
Lundberg, S.M., Lee, S.-I.: A unified approach to interpreting model predictions. In: Advances in Neural Information Processing Systems, vol. 30, pp. 47654774 (2017)
-
Lundberg, S.M., Erion, G.G., Chen, H., et al.: From local explanations to global understanding with explainable ai for trees. Nature Machine Intelligence 2, 5667 (2020)
-
Esteva, A., Robicquet, A., Ramsundar, B., et al.: A guide to deep learning in healthcare. Nature Medicine 25, 2429 (2019)
-
Breiman, L.: Random forests. Machine Learning 45, 532 (2001)
-
Meng, Z., et al.: Deep learning based detection of diabetic kidney disease using retinal fundus images. IEEE Journal of Biomedical and Health Informatics (2023)
-
Li, X., et al.: Artificial intelligence assisted diagnosis of diabetic nephropathy using deep learning based image analysis. Computers in Biology and Medicine 145, 105480 (2022)
-
Wu, X., et al.: Machine learning and transcriptomic analysis identify glycol- ysis related genes for early diagnosis of diabetic nephropathy. Frontiers in Endocrinology 13 (2022)
-
Huang, J., et al.: Identification of hdac9 as a key regulator in diabetic kidney dis- ease using transcriptomic analysis and machine learning. Journal of Translational Medicine 21 (2023)
-
Gunning, D., Aha, D.: Darpas explainable artificial intelligence (xai) program. AI Magazine 40(2), 4458 (2019)
-
Shapley, L.S.: A value for n-person games. Contributions to the Theory of Games
2, 307317 (1953)
-
Chen, Y., et al.: Machine learning based prediction of chronic kidney disease using clinical data. IEEE Access (2020)
-
Chicco, D., Jurman, G.: The advantages of the matthews correlation coefficient over f1 score and accuracy in binary classification evaluation. BMC Genomics (2020)
-
Almansour, N.A., et al.: Neural network and machine learning methods for chronic
kidney disease prediction. Computational Methods and Programs in Biomedicine (2019)
-
Chaudhuri, S., et al.: Artificial intelligence approaches for kidney disease predic- tion and diagnosis. Healthcare (2021)
-
Le, T., et al.: Deep learning for healthcare applications: A review. IEEE Reviews in Biomedical Engineering (2018)
-
Vaswani, A., et al.: Attention is all you need. In: Advances in Neural Information Processing Systems (2017)
-
Holzinger, A., et al.: Causability and explainability of artificial intelligence in medicine. Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery (2019)
-
Tjoa, E., Guan, C.: Explainable artificial intelligence in healthcare: A survey. IEEE Access (2021)
