DEEP LEARNING CLASSIFICATION OF OTOSCOPIC IMAGES FOR AUTOMATED DIAGNOSIS OF MIDDLE EAR DISEASE IN LOW-RESOURCE CLINICAL SETTINGS

Abstract

This study aims to propose a deep learning-based predictive model for early recognition of the risk of difficult extubation and transfer to the intensive care unit (ICU) for high-risk patients receiving anesthesia. The aim is to optimize decision making and patient safety in the perioperative period by using advanced artificial intelligence-based risk stratification. High-risk surgical patients were collected a comprehensive database of preoperative, intraoperative and postoperative clinical parameters. The data preprocessing process comprised normalization, handling missing values, feature engineering, and class balance. A novel deep neural network architecture combining the multiple hidden layers, attention mechanism, and temporal feature learning was designed to model the rich clinical relationship. The accuracy, precision, recall, F1-score, area under the receiver operating characteristic curve (AUC-ROC) and calibration metrics were used to assess the model performance. Comparative analyses were conducted with traditional machine learning algorithms and existing clinical risk assessment tools. The deep learning framework proposed showed better performance in predicting the difficult extubation and interaction with the ICU outcomes. The model had high discrimination and was able to successfully identify patients at high risk perioperatively. The feature importance analysis showed that the most important features were respiratory parameters, airway characteristics, hemodynamic instability indicators, surgical complexity, and the presence of comorbidities. The deep learning approach increased the sensitivity and offered earlier risk detection than the traditional approach, with high specificity.