Motor imagery (MI)-based brain-computer interface (BCI) is a system that allows users to control computer devices by imaging body part movements or MI tasks. In BCI applications, the classification of MI using electroencephalogram (EEG) is challenging because EEG is highly susceptible to noise and artifacts. The latency and length of MI period also vary between subjects and sessions; however, many conventional applications tend to empirically define time windows for feature extraction. This can lead to lower MI-BCI performance. This paper proposes two mutual information-based time window adaptation (MT) algorithms; sliding window MT (SWMT) and genetic algorithm MT (GAMT). Both algorithms used optimized reference signals and mutual information analysis to constantly adjust the time window starting point and length. Reference signals were optimized based on mutual information analysis and performance evaluation. Feature extraction and classification algorithms were finally applied to evaluate SWMT and GAMT performance. The results indicate that SWMT and GAMT were able to improve the conventional approach by increasing the classification accuracy by 6.00% and 6.37%, respectively.