A Medical Time Series Classifier with False Negative and Positive Mitigation via UNITS Transformer Learning

Effective biosignal interpretation is critical for intelligent healthcare. Addressing the global challenge of diabetes mellitus, this study proposes a fully data-driven framework to classify type I and type II diabetes using solely 96-point continuous glucose monitoring (CGM) traces. The model leverages a Transformer backbone to capture global temporal dependencies in glucose dynamics, augmented by a lightweight Unified Multi-Task Time-Series (UNITS) subnetwork for local feature extraction. A tunable weighted binary cross-entropy loss function is employed to address the asymmetric clinical risks of misdiagnosis, effectively mitigating the impact of class imbalance while prioritizing safety. Evaluated on real-world CGM datasets, the model achieves 88.85% accuracy and demonstrates robust daylevel stability. This approach bridges AI with medical electronics, offering a lightweight solution for scalable, data-driven diagnostic systems.