Randomized Explainable Machine Learning Models for Efficient Medical Diagnosis

New Breakthrough in Intelligent Medical Diagnosis: Randomized Explainable Machine Learning Models Drive Efficient Medical Diagnostics

I. Academic Background and Research Motivation

In recent years, Deep Learning (DL) models have played a crucial role in the field of healthcare. By processing vast amounts of medical data, DL significantly improves the accuracy of disease diagnosis and the level of clinical decision-making. In domains such as medical image analysis, genomics data processing, and clinical disease prediction, DL models have demonstrated powerful automatic feature extraction and complex pattern recognition capabilities. However, the “black box” nature of deep models (i.e., the difficulty in interpreting their decision processes), huge consumption of computational resources, and lengthy training times have become significant obstacles to their practical application in clinical settings.

The decision-making process in the medical field requires not only high accuracy, but also speed and transparency—ensuring rapid diagnosis for emergency medical situations on the one hand, and meeting legal and regulatory requirements for the explainability of automated decisions, such as the GDPR, on the other. In addition, the energy consumption problems brought by DL models, especially large-scale neural networks, are becoming more prominent, prompting calls for Green AI and energy-saving algorithms. Faced with these challenges, the development of intelligent diagnostic models that are highly efficient in training, fast in execution, and interpretable in decision-making has become a hot research topic in AI for healthcare.

This paper focuses on randomized machine learning models, namely Extreme Learning Machines (ELMs) and Random Vector Functional Link (RVFL) networks, which aim to greatly reduce the training complexity and computational cost of deep learning models by introducing randomization of model parameters, while maintaining or even improving diagnostic accuracy. To break the “black box” dilemma, the study integrates two mainstream Explainable AI (XAI) technologies—LIME (Local Interpretable Model-agnostic Explanations) and SHAP (Shapley Additive Explanations)—to uncover the mechanisms behind model decisions and increase the trust of doctors and patients. Overall, this research focuses on efficiency and interpretability, dedicated to bringing practical new tools to medical AI and promoting the popularization and development of intelligent medical diagnosis.

II. Paper Source and Author Information

This paper, titled “Randomized Explainable Machine Learning Models for Efficient Medical Diagnosis,” was published in the IEEE Journal of Biomedical and Health Informatics, Volume 29, Issue 9, in September 2025. The main authors are Dost Muhammad (University of Galway, Ireland), Iftikhar Ahmed (University of Europe for Applied Sciences, Germany), Muhammad Ovais Ahmad (Karlstad University, Sweden), and Malika Bendechache (ADAPT Centre, University of Galway). The research was supported by multiple European research programs and employed generative AI tools for language improvement.

III. Detailed Research Process

1. Dataset Selection and Preprocessing

The paper focuses on two medical diagnosis case studies:

(1) Genitourinary Cancer Dataset
- Source: Mackay Memorial Hospital, data collected from 2017-2021
- Sample Size: 1,337 patients
- Feature Dimensions: 39 features in total, including various biochemical markers (such as A/G ratio, albumin, alkaline phosphatase, ALT/GPT, AST/GOT, BUN, calcium, chloride, creatinine, direct bilirubin, estimated GFR, fasting glucose, uric acid, etc.), urinalysis indicators (nitrite, occult blood, urine pH, potassium, sodium, specific gravity, strip WBC, total bilirubin, total protein, triglyceride, urine epithelial cell count, etc.), as well as demographic and lifestyle information (age, gender, hypertension, diabetes, smoking, alcohol consumption, betel nut consumption, and family history).
- Diseases Studied: Kidney cancer, prostate cancer, bladder cancer, cystitis, uterine cancer, etc. All diagnostic results were confirmed by experts and pathology reports.

(2) Coronary Artery Disease Dataset
- Source: UCI Machine Learning Repository
- Sample Size: 303 cases
- Feature Dimensions: 13 features, including typical clinical indicators (such as age, gender, ECG, exercise-induced angina, fasting glucose, chest pain type, serum cholesterol, etc.)

In order to ensure the quality of model inputs, both datasets underwent comprehensive preprocessing:
- Missing values were filled with the mean of their respective columns to avoid losing information
- Categorical labels were converted using one-hot encoding
- Data were split into training (70%) and test (30%) sets
- All features were standardized (StandardScaler) so that the mean is zero and the variance is one, eliminating differences in feature scales

2. Randomized Learning Model Design

(1) Extreme Learning Machines (ELMs)
ELM is a feedforward neural network known for its extremely fast training speed and low complexity. Its main innovation points are:
- Input layer and hidden layer weights and biases are all randomly initialized and remain fixed;
- Only the weights between the hidden layer and the output layer are linearly optimized (least-squares optimization), which greatly reduces parameter adjustment and computational costs;
- Supports multi-layer hidden layers (1-4 layers randomly selected in this study), with ReLU as activation function and Softmax in the output layer;
- To tackle overfitting, L2 regularization is embedded

The mathematical foundation of ELM is as follows: suppose the input is x and the output is y, the model randomly fixes w and b, obtains hidden layer output h via activation function g(·), and finally calculates output weights β directly through an analytic solution, aiming for the model output to be as close as possible to the true y. The optimization is achieved through the formula: $$ \min_{\beta} ||h\beta - y||^2 + \lambda||\beta||^2 $$

(2) Random Vector Functional Link Network (RVFL)
RVFL builds upon ELM with further innovations:
- In addition to random hidden layer weights, it establishes direct connections (Direct Links) from the input layer to the output layer, so that input data can also directly influence outputs in a linear fashion; - Applies Tikhonov regularization to reduce model complexity;
- The hidden layer uses tanh activation function and applies second-degree polynomial expansion to the hidden output, improving non-linear expression ability;
- Only the output layer weights are adjusted, while hidden and direct connection weights are random and fixed
- Analytic (non-iterative) training optimization for speed improvement

The model output for RVFL is: o = hwhidden + xwinput, and w* is optimized to minimize prediction error and improve generalization.

3. Benchmark Model Configuration

To highlight the advantages of randomized models, the study designed three traditional neural network models for comparison:

• Deep Neural Network (DNN): Contains four hidden layers with ReLU activation, Softmax at the output end (Binary Cross-Entropy for binary classification, Categorical Cross-Entropy for multi-class problems).
• Recurrent Neural Network (RNN): Combines three fully connected hidden layers (with different numbers of neurons), formerly adept at processing sequential data but used for static data testing in this study.
• Simple Neural Network (SNN): Has only one hidden layer, using ReLU and Sigmoid/Softmax structure, focusing on feature representation and structural simplicity.

4. Integration of Explainable AI Technologies

To reveal the decision mechanisms of the models, the following XAI methods are introduced:

• LIME (Local Interpretable Model-agnostic Explanations): Perturbs original inputs to generate neighboring samples, and uses a local linear model g to approximate the complex model f’s reasoning for a specific prediction, thus providing causal explanation for a single data point in feature space.
• SHAP (Shapley Additive Explanations): Based on cooperative game theory, assigns each feature a contribution score φ by calculating the marginal contribution of all feature combinations, giving a comprehensive quantification of how each feature affects the final decision.

5. Evaluation Metrics and Operating Environment

Main evaluation metrics: Accuracy, Precision, Recall, F1-Score
Experimental hardware: AMD Ryzen 7 5700X processor, 16GB NVIDIA GeForce RTX 4080 graphics card
All models are implemented in Python to ensure reproducibility of results.

IV. Detailed Main Research Findings

1. Genitourinary Cancer Diagnosis Results

• DNN: Accuracy 82.98%, training time 13.9 seconds;
• SNN: Accuracy 83.76%, training time 14.25 seconds;
• RNN: Accuracy 83.42%, training time 18.82 seconds;
• ELM: Accuracy increased to 87.26%, training time only 5.31 seconds, and all other metrics outperformed traditional models;
• RVFL: Highest accuracy at 88.29%, training time only 6.22 seconds, demonstrating outstanding speed and classification performance.

Through LIME and SHAP visualization analysis, the models found:
- Betel nut consumption, urinary epithelial cells, gender, etc., have significant positive contributions to kidney cancer prediction; - Diabetes, fasting blood glucose, etc., also substantially increase the risk of kidney cancer; - Urine glucose, protein, etc., tend to indicate “no cancer”; - For cystitis diagnosis specifically, nitrite in urine, decreased albumin level, alcohol consumption, cholesterol, etc., are major positive indicators. - The combined use of SHAP and LIME quantitatively explains the positive and negative contributions of different lab indices to the model’s diagnostic results.

2. Coronary Artery Disease Diagnosis Results

• DNN: Accuracy 75.82%, training time 6.65 seconds;
• SNN: Accuracy 73.63%, training time 4.15 seconds;
• RNN: Accuracy 73.63%, training time 7.86 seconds;
• ELM: Accuracy 76.45%, training time only 0.021 seconds;
• RVFL: Best accuracy at 81.64%, training time only 0.0308 seconds, highlighting computational efficiency.

Results from LIME and SHAP analyses show:
- thal (myocardial perfusion scan results), exang (exercise-induced angina), oldpeak (ST segment depression induced by exercise), ca (number of major vessels), cp (type of chest pain), trestbps (resting blood pressure), fbs (fasting blood sugar), etc. contribute positively to coronary artery disease diagnosis; - slope (ST segment slope), restecg (resting ECG), age, etc., tend to indicate “no disease”; - SHAP further quantitatively demonstrates the specific contribution values of cp, slope, ca, oldpeak, fbs, etc., comprehensively revealing the model’s decision logic.

3. Discussion of Performance Differences and Model Mechanisms

The outstanding performance of ELM and RVFL arises from a streamlined parameter space and the elimination of the high complexity of iterative training. By relying on random initialization and analytic solutions, they greatly enhance learning speed and generalization ability. RVFL’s direct input-to-output connections enable the model to more easily learn multi-level non-linear relationships among complex features, particularly suitable for medical laboratory data. In contrast, DNN, though powerful in representational capacity, struggles with balancing speed and resource consumption. RNNs, specialized for sequential data, have limited efficiency with static medical records, explaining their subpar performance here. SNNs, while structurally simple, are constrained in their ability to express complex nonlinearity, making it difficult to further improve diagnostic performance.

With the combined analysis of LIME and SHAP, the transparency of model decisions is greatly enhanced, supporting clinicians in tracing and questioning model recommendations, and thereby significantly increasing trust and decision support in actual clinical applications.

V. Conclusion and Scientific Application Value

This study innovatively proposes a medical intelligent diagnosis framework based on randomized extreme learning machines and random vector functional link networks, not only surpassing traditional deep learning models in terms of accuracy and computational efficiency, but also effectively solving the “black box” problem. With the incorporation of LIME and SHAP interpretability technologies, the model can provide clear and quantitative explanations of feature contributions for every diagnostic result, truly achieving a trinity of “high accuracy + interpretability + high efficiency” in medical AI.

The scientific value is reflected in:
- Opening up a new path for the field of intelligent medical diagnosis outside of traditional deep models, promoting the broad, rapid, and energy-saving adoption of AI medical tools; - Improving model transparency and clinical acceptability, in line with the latest international regulations on AI explainability in healthcare; - Greatly reducing resource consumption, making large-scale medical imaging and laboratory data processing economically feasible

The application value lies in:
- In crucial clinical diagnoses (such as genitourinary cancers and coronary artery disease), providing physicians with precise and easy-to-understand decision support; - Broad applicability to other disease diagnosis scenarios, offering practical help for real-time medical decisions and edge computing in medical devices; - Fast scalability and deployment capability, suitable for clinical scenarios requiring rapid response, such as emergency medicine and epidemic screening.

VI. Research Highlights and Future Prospects

• The first systematic integration of randomized neural networks and AI explainability frameworks, with comprehensive multi-model comparative analysis on large clinical cases;
• Demonstrated exceptional advantages of ELM and RVFL in terms of speed and accuracy for medical diagnosis, providing a solution path for the future implementation of efficient AI healthcare;
• Created a new paradigm linking model mechanisms and decision causality visualization, greatly improving the trust of clinical staff and patients;
• The paper proposes future expansion directions, including applying ELM and RVFL to neurological, respiratory, and infectious diseases, as well as integrating with cutting-edge AI techniques like reinforcement learning or generative adversarial networks to further enhance model capabilities.