Assessment of Bipolar Disorder using Heterogeneous Data of Smartphone-based Digital Phenotyping

Abstract

In mental health disorder Bipolar Disorder (BD) is one of the most common mental illness Using scales for assessment is one of the approaches for diagnosing and tracking BD patients However the requirements for manpower and time is heavy in the process of evaluation In order to reduce the cost of social and medical resources this thesis collect the data remotely by using smartphones and build an automatic system to predict the scale score for a more convenient way to diagnosing and evaluating in clinical application This thesis designed an android application (App) on smartphones to collect the user’s digital phenotyping data with various categories This thesis uses these heterogeneous digital phenotyping data to predict the score of Hamilton Depression Rating Scale (HAM-D) and Young Mania Rating Scale (YMRS) as a reference for the evaluation of BD This thesis collect the user’s data by the App on smartphones consisting of location data (GPS) self-report scales daily mood sleeping time and records of multi-media (text、speech、video) to build a database containing these heterogeneous digital phenotyping data First the features of various digital phenotyping data are extracted individually and then fed into models for training and predicting the score of scales As there wasn’t a universal model in previous studies this thesis picks 7 models for experimental test and comparison Moreover combinations of different numbers of feature categories are used to observe the improvement of performance in heterogeneous data In order to the complete lack of certain categories in heterogeneous data this thesis builds a model pool and uses ensemble method to predict and generate the score of scales for a more flexible system This study collected the heterogeneous digital phenotyping data from 84 BDs and 11 health controls Five-fold cross validation scheme is employed for evaluation Experimental results show that the performance of Lasso Regression and ElasticNet Regression are outstanding and the heterogeneous data has better performance than homogeneous data The prediction error (MAE) of HAM-D is 1 36 and the error of YMRS is 0 55 In the future more and long-term data should be collected to make the model more robust and hopefully obtain more feasibility of models and pre-processing In addition the tracking of the user’s historical data can be applied to build a more personalized and long-tracking system

Date of Award	2020
Original language	English
Supervisor	Chung-Hsien Wu (Supervisor)