Deep Learning-Based Anomaly Detection and Multi-Label Classification for the Analysis of Chest Radiographs

In clinical decision support systems, artificial intelligence (AI) and natural language processing (NLP) techniques are widely employed to enhance the efficiency of healthcare services. In this study, chest radiograph reports from the MIMIC-CXR dataset are analyzed to perform anomaly detection and multi-label disease classification. During the pre-processing stage, abbreviations in medical texts were expanded, spelling errors corrected, and irrelevant words removed. To generate robust text representations, domain-specific language models such as PubMedBERT and ClinicalBERT were employed. For anomaly detection, deep learning architectures—including the LSTM Autoencoder, Informer Autoencoder, and Transformer Autoencoder—were implemented, while in the classification stage, disease prediction was achieved using Binary Relevance combined with SMOTEN techniques. On retrospective MIMIC-CXR data, the model attained ROC-AUC = 0.986 and F<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</inf> = 0.961 for pneumonia, while sustaining F<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</inf> ≥ 0.89 across the four thoracic disease classes evaluated. Moreover, the developed AiXRisk score, which fuses the outputs from anomaly detection and classification, enables the early identification of high-risk patients. Overall, the findings demonstrate that the combined use of deep learning and NLP techniques offers a powerful approach for extracting meaningful clinical insights from medical texts.