Mobility limitations are associated with poor clinical outcomes, including higher mortality and disability rates, especially in older adults. Early detection of decline in mobility is of great importance for clinical practice, as it can still be stabilized or even reversed in the early stages. Mobile sensing offers a great range of sources, such as GPS, accelerometer, gyroscope, that can be used to implement mobility measures. Unlike traditional assessment tools, these technologies allow passively observing patients’ functions in real-life settings. Therefore, the purpose of this study is to develop a machine learning-based model to passively follow up on patients’ mobility over time from passively sensed mobility descriptor biomarkers and socio-demographic data of patients. The WHODAS 2.0 Questionnaire is used as a mobility measurement tool, which queries whether the individual had difficulty performing a set of tasks over the past 30 days. Using these scores as target outcomes, we define a pipeline that performs feature encoding for the daily information by applying Time2Vec, followed by an LSTM encoder for the 30-day embedded input sequence. A feed-forward layer on top of the LSTM outputs concatenated with demographic data is then used to get the predictions. Moreover, since the temporal data is regularly sampled but frequently missing, probabilistic generative models will be used to perform data imputation.