Diabetes mellitus is a disorder that disrupts blood glucose regulation and can lead to health complications if it is left uncontrolled. Continuous glucose monitoring provides dynamic, real-time insights into glucose patterns, enabling deep learning models for glucose forecasting. While models bespoke to populations with diabetes are common, their effectiveness for prediabetics remains uncertain. This study applied Long Short-Term Memory models for type 1 diabetes (T1D), type 2 (T2D), and prediabetes (PRED), which were evaluated with internal and external validity. Notably, the PRED model achieved normalized RMSEs of 0.21 mg/dL, 0.11 mg/dL, and 0.25 mg/dL on PRED, T1D, and T2D datasets, respectively.