Explainable Prediction of Dacryocystitis From Noninvasive Ocular Indicators Using Deep Stacked Network and The Shapley Additive Explanations Approach.

Dacryocystitis diagnosis is important for preventing rapid blurring and vision loss. Existing state-of-the-art methods focus on routine clinical examinations and objective scattering index-based statistical analysis. Such approaches are invasive operations or lack quantitative indicators, and their application is limited. in addition, little attention has been paid to the explainability and clinical utility of models. This paper proposes an explainable dacryocystitis prediction model from noninvasive ocular indicators. The proposed model is based on an deep stacked network with 4 improvements: a multivariable feature extraction module, obtaining comprehensive predictive factors including the quantitative ocular indictors, conventional texture features, and deep learning features from shallow to deep convolutional layers; a multifeature fusion and attribute selection module based on the ReliefF method, guiding the network to focus on useful information at variables; Decision curve analysis the model is introduced into the model to evaluates the risks and benefits; and appending a SHapley Additive exPlanations (SHAP) module to the framework to automatically and efficiently interpret the prediction of the models. By integrating the above improvements in series, the models' performances are gradually enhanced. Real labeled data samples are used to train and test the model, and our model achieves high accuracy and reliability.