Machine Learning gestützte Informationsextraktion aus radiologischen Befundtexten der oberen Extremität

Background: The usage of artificial intelligence for classification and information extraction is becoming increasingly widespread in the medical field. Despite this, algorithms often still lack a deeper medical text comprehension, needed to perform tasks such as the complex classification of injuries of the upper extremity. To address this, a new Deep Learning model of the BERT architecture was developed and trained in two different forms (medBERT.de and medBERT.dededup). Methods: After completing self-supervised training, the models were fine-tuned by the manual annotation of 1295 medical reports, to enable the recognition and complex classification based on condition and entity name of 35 individual pathologies. Subsequently, the classification performance was compared with four state-of-the-art BERT models using the metrics of precision, recall, F1 score, F1-Macro, F1-Micro and the AUROC. Results: Through supervised training, all six models were able to complete the classification task. The base version of the medBERT.de model achieved the best performance and best results in the metrics of precision (65,66 %), F1-value (59,11 %) and recall (60,59 %). Discussion: With the help of the dataset, a model was developed that acquired a deep understanding and comprehension of medical language, exceeding the ones of all other state-of-the-art models. Therefore, the medical classification of injuries of the upper extremity is enabled in a more comprehensive way than previously possible.