Leveraging Transformer-Based Models for Post-Market Surveillance: A Case Study on Turkish E-Commerce Reviews of Diabetes Management Devices

Background: Sentiment analysis in Turkish is particularly challenging due to the language’s agglutinative structure and rich morphological variation. At the same time, consumer-generated data on healthcare products—especially self-monitoring blood glucose (SMBG) systems—represent an underutilized source of real-world evidence for understanding device performance and user experience. Objective: This study aims to develop a Bidirectional Encoder Representations from Transformers (BERT) based sentiment analysis model for Turkish e-commerce reviews and to evaluate its applicability to SMBG devices as a potential complementary tool for post-market surveillance. This study utilizes an infoveillance approach to bridge the gap between traditional vigilance systems and real-world patient experiences, providing a scalable model for digital health surveillance in non-English speaking contexts. Furthermore, we demonstrate how this infoveillance framework can support health policy decision-making by providing real-time oversight of medical device safety. Methods: A pre-trained Turkish BERT model was fine-tuned to classify sentiments into positive, negative, and neutral categories. A corpus of 75,000 Turkish e-commerce reviews, one of the largest manually annotated datasets for Turkish sentiment analysis, was collected and processed through tokenization, cleaning, and manual labeling. Model performance was compared with traditional machine-learning baselines to assess the benefits of contextual embeddings. Results: The fine-tuned BERT model outperformed all baseline classifiers, achieving higher overall accuracy and class-wise F1-scores. In the SMBG subset, sentiment distribution was 89.44% positive, 7.22% negative, and 3.33% neutral, indicating that the model successfully captured fine-grained linguistic cues and provided robust insights into user perspectives. Conclusion: The findings demonstrate that BERT-based models can effectively address the linguistic complexity of Turkish and extract meaningful sentiment patterns from consumer reviews of SMBG devices. These insights show potential value for regulatory and vigilance contexts, suggesting that consumer-generated data could serve as a complementary early-signal source in post-market device surveillance. Future work may integrate multimodal inputs and explainability techniques to improve domain adaptability and transparency.