The prevalence of type 2 diabetes mellitus (T2DM) is increasing in recent decades. T2DM can wreak havoc on human beings that affects patients’ everyday life and may even cause death. The long-term monitoring of blood glucose level of patients usually relies on invasive blood glucometer. It has various disadvantages such as potential infection, skin changes, costly, non-compliance, and painful. Attention is drawn to a non-invasive approach for blood glucose monitoring using electrocardiogram (ECG) signals. A feasibility is thus conducted to evaluate the performance of T2DM prediction using ECG signals. A modified long short-term memory network is proposed for the early prediction of blood glucose level in a time period of 6-, 12-, 18-, and 24 month performance evaluation is carried out using benchmark dataset, Cerebromicrovascular Disease in Elderly with Diabetes Dataset. Results show the feasibility of blood glucose level and thus potential T2DM prediction using ECG signals. Blood glucose level, Electrocardiogram, Long short-term memory, Machine learning, Type 2 diabetes mellitus