Predicting Mild Cognitive Impairment through Ambient Sensing and Artificial Intelligence

This paper reports an emerging application leveraging ambient and artificial intelligence techniques for in-home sensing and cognitive health assessment. The application involves a prospective longitudinal study, wherein non-pervasive sensing devices are installed in homes of over 63 real users undergoing clinical cognitive assessment, and digital signals of the users’ activities and behaviour are transmitted to a central cloud-based data server for further processing and analysis. Based on the sensor readings, we identify a set of digital biomarkers covering four key aspects of daily living, namely physical, activity, cognitive, and sleep, and develop a suite of customized feature extraction methods for deriving them from the sensor readings. As sensor data captured from real world are inherently sparse and noisy, we build predictive models using various machine learning techniques and evaluate their sensitivity to missing and noisy data. Validated with findings of clinical assessment, our experiments show that machine learning-based predictive models are able to identify mild cognitive impairment (MCI) cases based on the extracted digital biomarkers with reasonably high F1 scores of more than 0.85. This shows that the sensor-based digital biomarkers are indicative of the users’ cognitive health status and could be further exploited for more general health assessment applications. With a vision of massively deploying such sensor-based AI systems, the paper discusses the challenges we encountered and shares our lessons learned.