The classification of medical statements in German doctor-patient interactions presents significant challenges for automated medical information extraction, particularly due to complex domain-specific terminology and the limited availability of specialized training data. To address this, we introduce a manually annotated dataset specifically designed for distinguishing medical from non-medical statements. This dataset incorporates the nuances of German medical terminology and provides a valuable foundation for further research in this domain. We systematically evaluate Transformer-based models and multimodal embedding techniques, comparing them against traditional embedding-based machine learning (ML) approaches and domain-specific models such as medBERT.de. Our empirical results show that Transformer-based architectures, such as the Sentence-BERT model combined with a support vector machine (SVM), achieve the highest accuracy of 79.58% and a weighted F1-Score of 78.81%, demonstrating an average performance improvement of up to 10% over domain-specific counterparts. Additionally, we highlight the potential of lightweight ML-models for resource-efficient deployment on mobile devices, enabling real-time medical information processing in practical settings. These findings emphasize the importance of embedding selection for optimizing classification performance in the medical domain and establish a robust foundation for the development of advanced, domain-adapted German language models.