Assisting older adults with medication reminders through an audio-based activity recognition system

Poor adherence to prescribed drug treatments is one of the leading causes of illness and treatment failure, which increases re-hospitalizations. In Mexico, the factors that most contribute to the non-adherence problem are age, polypharmacy, and education. For instance, elderly patients are prescribed with an average of seven medications after they are discharged from hospitals, and 25% of them face problems managing medications at home. A strategy that older adults use for medication adherence is to link their medication regimens to daily activities. We propose a system based in machine learning for audio-based activity recognition using Hidden Markov Models over Mel Frequency Cepstral Coefficients. The system triggers an assistive conversational agent that adapts its interaction model to the context detected. We report on two studies that provide evidence of the feasibility of our approach to assist older adults to develop consistent medication behaviors by associating them to daily routines. We first conducted an observational study with two older adults to understand the role of daily activities to develop consistent medication behaviors. Afterwards, we conducted an in situ assessment of the audio-based activity recognition system with the two study subjects. Our results showed that anchor activities with an audible manifestation were recognized with an accuracy of 79% for subject 1, and 97.6% for subject 2. Additionally, we validated how the integration of conversational agents into the system may support the mental association among activities and medication regimens that older adults fail to realize when, for instance, their intention plans involve multiple behaviors associated to an activity. The deployment of the proposed approach requires only a smart speaker, which increases its feasibility of adoption in Latin American and other developing countries.