The world’s population is aging. By 2030, the number of U.S. adults aged 65+ will be approximately 71 million. Increase in older adult population is resulting in an epidemical growth of diseases related to cognitive impairment, such as dementia. Therefore, innovative healthcare options need to be considered in order to provide quality care to our aging population and to reduce caregiver burden. Smart home technologies can play a pivotal role in disrupting conventional “caregiving”. A primary intervention that is valuable for individuals with cognitive impairment is automated prompts that aid with initiation and completion of daily activity. We postulate that prompt timing can be automated by incorporating contextual information of activities gathered from sensors located in a smart home. To determine the ability of machine learning algorithms to generate appropriate activity-aware prompts, we performed a study in an on-campus smart home with 300 volunteer participants, aged 50+, who are healthy older adults or individuals with mild cognitive impairment. The sensor data collected from the smart home were used to generate various contextual features of an individual’s daily activities. Machine learning algorithms were trained on the data to classify a “prompt” situation from a “no-prompt” situation. However, lack of training samples representing prompt situations raises a fundamental machine learning problem known as imbalanced class distribution. We proposed a probabilistic oversampling technique that helps in better learning of the “prompt” class. While existing approaches achieve 0-40% accuracy on predicting prompt situations correctly, our approach achieved >80% accuracy.