Diabetes imposes a substantial burden on global healthcare systems. Worldwide, nearly half of individuals with diabetes remain undiagnosed, while conventional diagnostic techniques are often invasive, painful, and expensive. In this study, we propose a noninvasive approach for diabetes detection using photoplethysmography (PPG), which is widely integrated into modern wearable devices. First, we derived velocity plethysmography (VPG) and acceleration plethysmography (APG) signals from PPG to construct multi-channel waveform representations. Second, we introduced a novel multiset spatiotemporal feature fusion framework that integrates hand-crafted temporal, statistical, and nonlinear features with recursive feature elimination and deep feature extraction using a one-dimensional statistical convolutional neural network (1DSCNN). Finally, we developed an interpretable diabetes detection method based on XGBoost, with explainable artificial intelligence (XAI) techniques. Specifically, SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME) were employed to identify and interpret potential digital biomarkers associated with diabetes. To validate the proposed method, we extended the publicly available Guilin People’s Hospital dataset by incorporating in-house clinical data from ten subjects, thereby enhancing data diversity. A subject-independent cross-validation strategy was applied to ensure that the testing subjects remained independent of the training data for robust generalization. Compared with existing state-of-the-art methods, our approach achieved superior performance, with an area under the curve (AUC) of $80.5\pm 15.9\mathrm{\%}$, sensitivity of $77.2\pm 7.5\mathrm{\%}$, and specificity of $64.3\pm 18.2\mathrm{\%}$. These results demonstrate that the proposed approach provides a noninvasive, interpretable, and accessible solution for diabetes detection using PPG signals.