Generation of patient-focused medical insights using predicted time-series of post-surgery recovery

Founded in 2015, moveUP is a Belgian company that specialises in telerehabilitation, i.e. digital post-surgery follow-up solutions. This results in a more efficient rehabilitation process for both patients and physiotherapists alike thanks to the simultaneous use of data-based insights and expert medical supervision. However, as with in-person rehabilitation, some patients report feeling lost, struggling to assess their progress, and being unable to set realistic expectations for the rest of the rehabilitation. Consequences include a reduced motivation to comply with the therapy, negatively impacting the outcome. This thesis aims to address this issue using custom messages sent to the patient through the chat feature available in moveUP's smartphone app. These messages provide patients with both feedback and realistic expectations about their pain and knee range of motion levels based on data collected by moveUP as well as a predictive time-series model. Such predictive insights may also be used to trigger alerts that warn physiotherapists when one of their patients is not evolving as they should. Additionally, they also allow physiotherapists to better assess the amount of time required for the follow-up of each patient, thereby enabling them to more accurately plan their workload ahead of time. Through this research, multiple approaches to such a predictive model were investigated, including time-series clustering, dimensionality reduction techniques such as splines and Fourier transforms, and multidimensional scaling (MDS) and t-distributed stochastic neighbour (t-SNE) embeddings using a distance metric based on correlation between the time-series. The retained method is Functional Principal Component Analysis through Conditional Expectation (FPCA-PACE) as it allows to effectively deal with time-series data in which few samples are available. This is achieved by identifying principal functional modes of variation in the provided time-series to decompose them into a linear combination of these "principal functions". This work also focuses on the usability of the system by physiotherapists, as the intent is currently that any generated message be validated by a physiotherapist before being sent to patients. To this end, a user-friendly browser-based dashboard has also been developed. The performance of the predictive model was assessed using the Mean Absolute Error (MAE), the Root-Mean-Square Error (RMSE), and the percentage of predictions that fell within a reasonable threshold of the true values. Simple benchmark models were used as comparison points, which highlighted the superiority of the proposed model. The system as a whole was also assessed through a testing phase on five patients with the help of a physical therapist working at moveUP. Although encouraging feedback was given by the PT and one of the patients, no significant conclusion could be attained due to the small sample size.