Prediction and analysis of risk factors for diabetic retinopathy based on machine learning and interpretable models

Convolutional Neural Network-Based ECG Signal Classification Model: A Study on Addressing Class Imbalance and Enhancing Model Interpretability

“…SHAP values, by quantifying the contribution of each feature to the model's prediction outcome, provide a tool for doctors and researchers to identify the most critical features in ECG signals for model predictions and to gain a deep understanding of the model's decision-making logic [32,34,38,39]. The SHAP values, based on the Shapley values from cooperative game theory which are used for fair resource allocation, are applied in machine learning to measure the marginal contribution of each feature to the model's prediction [11,40]. The SHAP value for each feature represents its average contribution across all possible combinations of features, with the sign indicating the direction of the feature's impact on the prediction outcome: positive correlation or negative correlation [39,41].…”

“…Despite the computational intensity required to calculate SHAP values, considering all possible feature combinations, the algorithms provided by the SHAP library can efficiently compute these values, especially suitable for decision tree and deep learning models [11,14]. SHAP values not only help to reveal the internal workings of the model but also visually display each feature's contribution to the prediction outcome through visualization, which is particularly useful for non-technical users to understand model predictions [32,38].…”

“…To enhance the model's transparency and interpretability, this study introduces Uniform Manifold Approximation and Projection (UMAP) dimensionality reduction technology for visualizing the model's classification boundaries and employs Shapley Additive exPlanations (SHAP) value analysis to quantitatively assess the specific contribution of each ECG signal feature to the model's prediction outcome [8][9][10][11][12]. UMAP technology provides us with a macro perspective, helping us to observe the structure of the data as a whole [13], while SHAP value analysis reveals the feature contributions behind each prediction from a microscopic perspective [14].…”

Optimizing Ocular Pathology Classification with CNNs and OCT Imaging: A Systematic and Performance Review

“…Other related conditions include Diabetic Macular Edema (DME), Drusen formation, and Choroidal neovascularization (CNV). This study focuses on analyzing classification methods to detect these three pathologies using medical imaging [10] [11]. Recent advances have shown that deep learning algorithms, particularly using optical coherence tomography (OCT) [12], [13], [14] and fundus images [15], [16], can automatically extract pathological features.…”

“…It also evaluates the impact of training data volume, model complexity, and their influence on performance in scenarios including random weight initialization, fine-tuning through transfer learning, and retraining only the classification layers. Despite most studies achieving over 99% accuracy in classifying retinal pathologies with OCT images [11], [14], [19], [20], [21], [22], [23], [24] the relationship between model learning capacity and data volume has been underexplored. This study aims to address this gap, potentially guiding the data requirements for clinical trials to effectively classify ocular pathologies.…”