Analysis of Patient Evolution on Time Series of Different Lengths

The recent explosion in amount of data and the increase in the capabilities of computer systems to cheaply store and process this information has led to what has been called the `big data revolution'. One of the industries that has most benefited from the big data revolution is medicine, where data is increasingly used not just to monitor patient state but also to proactively intervene when necessary. The focus of this study is the data provided by a mobile crowd sensing platform called TrackYourTinnitus, which collects ecologically valid longitudinal assessments of patients' disease states at different points in time. Tinnitus is a neuropsychiatric disorder characterised by the perception of phantom sound. The causes for tinnitus are quite varied, and many studies have remarked that more basic research needs to be done to understand tinnitus, its causes, and effective treatments. Collecting real-world data is of paramount importance to better understand tinnitus. It is the goal of this work to analyse longitudinal data from the TrackYourTinnitus platform using time series analysis techniques to discover patients that show similar patterns of evolution. It was found that the DTW algorithm can be used to compute time series distances and find subgroups of similar patients. However, data quality issues like very short time series, a large range in the time series lengths and missing values necessitate many preprocessing steps. We propose a visualisation-driven solution for subgroup discovery among time series of similar lengths by using DTW distances (computed over both univariate and multivariate data) in conjunction with multiple visualisation tools (force directed graphs, time series visualisation and dendrograms). Evolution within the discovered subgroups is computed as the average relative difference between the beginning and end of the time series. We found that visualisation driven subgroup discovery is possible in the case of highly noisy data. It was also found that subgroup trajectories were markedly different from cluster averages. In addition, we were also able to show that the early interactions of the user with the system do not necessarily re
ect his actual disease state. We conclude with a list of topics and suggestions for future work.