Medical time series (MedTS) data, including Electroencephalography (EEG) and Electrocardiography (ECG), are widely utilized in clinical diagnosis and physiological monitoring. With the rapid advancement of deep learning, various models have been applied to MedTS classification. However, existing approaches still have significant limitations. From a temporal perspective, many methods fail to capture both local fluctuations and long-term trends simultaneously, which are essential for identifying disease-related patterns. In terms of spatial modeling, most approaches overlook redundant and noisy information across channels, leading to suboptimal performance and reduced generalization ability. To address these issues, we propose MedDPA, a multi-scale framework based on MLP for MedTS classification. MedDPA explicitly separates short-term fluctuations and long-term trends through a decomposition module. We also introduce a prototype-based channel aggregation module to suppress noise and reduce redundancy while preserving essential information. Finally, we integrate multi-scale features through a dual-direction fusion strategy and dynamically adjust the contribution of each scale. Our method is evaluated on multiple real-world EEG and ECG datasets. Results demonstrate that MedDPA outperforms 10 baselines across different metrics, validating its effectiveness, robustness, and potential for real-world applications.