PhD Public Defense, Nida Saddaf Khan

Title: Face Your Fear: Human Activity Recognition for Anxiety and Obsessive-Compulsive Disorders
by Nida Saddaf Khan
PhD Public Defense
Advisor: Dr. Sayeed Ghani
External Examiners: Dr. Jawwad Shamsi (FAST-NU), Dr. Junaid Qadir (ITU)
Date: Friday, Dec 23 at 9:30am
Venue: Conference Room, Tabba Academic Block, Main Campus, IBA Karachi

Abstract
Artificial Intelligence (AI) has transformed traditional healthcare systems using its techniques to continuously learn and serve in various medical situations, ranging from support in clinical decisions to the assistance in the operation theatre. Specifically, Machine Learning (ML), Computer Vision (CV), and Deep Learning (DL) are some of the most prominent branches of AI which have shown tremendous success in healthcare. Healthcare professionals use AI-equipped tools to understand the disease better, allowing them to make informed and intelligent decisions promptly.

Human Activity Recognition (HAR) is a subfield of ML which recognizes human activities automatically based on streaming data obtained from various sensors, such as inertial sensors, physiological sensors, location sensors, camera, time, and many other sensors. HAR is beneficial in multiple healthcare applications such as patient care, aged care, personal care, rehabilitation engineering, and ambient living. Its contribution has made it possible to remotely monitor the patients and ensure their wellbeing 24 hours a day, even from distant places. Hence, minimizing the need for the physical presence of a health professional all the time.

HAR can also be beneficial in psychiatry, where people suffer from mental, emotional, and behavioral disorders. Some of the most common conditions are Anxiety Disorder (AD), Autism Spectrum Disorder (ASD), Bipolar Disorder (BD), Eating Disorder (ED), and Obsessive-Compulsive Disorder (OCD). According to American Psychiatry Association (APA), "OCD is an anxiety disorder in which people have recurring, unwanted thoughts, ideas or sensations (obsessions) that make them feel driven to do something repetitively (compulsion)." The repetitive behavior of hand washing, cleaning routines, list-making, and checking-on things are the most common OCD disorders. There is a dire need for an automatic system that can identify the human behaviors related to OCD to understand the intensity and severity of the disorder so that timely action/interruption can be taken.

Generally, human activities are divided into two major categories: simple activities and complex activities. Simple activities have small repetitive patterns that can be easily recognized, such as, sitting, running, and standing. The complex activities lack this repetitive nature; hence require complex algorithms for their recognition. Some complex activities are eating, making coffee, and smoking. Various machine learning algorithms have been used to recognize simple activities such as random forest, support vector machine, and k-nearest neighbor. However, deep learning algorithms have made remarkable success in this field. Various researchers have extensively used deep networks for HAR in multiple domains. Most research is conducted to recognize simple human activities, and fewer efforts have been reported to recognize complex activities. In OCD, the sufferers perform certain physical behaviors repeatedly. These behaviors may serve as an indicator to detect an anxiety attack/situation. In these situations, complex behavioral patterns are exhibited frequently by the sufferers. Recognizing such behaviors may lead to a better understanding of the disease and help devise an effective treatment plan. Furthermore, it can also monitor the progress of the treatment so that the psychologist can stay up to date with the patients' condition.

In this thesis, a detailed understanding of anxiety and OCD is established, and AI/ML techniques are researched in maintaining mental health. DL structures are explored to recognize the simple and complex activities related to anxiety and OCD behaviors. Wearable devices equipped with motion sensors of accelerometer, gyroscope, and magnetometer are used for recording the physical movements of the human body. Two datasets are created for such behaviors; one is for anxiety-displaying behaviors, and the other is related to OCD behaviors. These datasets are novel and unique in the following aspects: a) there exist no other datasets for similar activities or purposes, b) these are comprised of both simple and complex activities, c) the behaviors are recorded in three different body postures of standing, sitting, and lying down, d) they are created in a semi-natural environment. These features have brought greater complexity in terms of intra-class variation, making it hard to recognize the correct behavior. A novel algorithm is proposed to recognize these behaviors for which various deep learning-based models are evaluated.

In this thesis, the dataset and algorithm for anxiety-related activities and OCD-related activities are called ADAM-Sense and OCD-Sense, respectively. After extensive experimental analysis, a hybrid model of convolution and recurrent neural network structures named CNN-LSTM was found to be the most suitable model. The performance of this model on ADAM-Sense and OCD-Sense was 93.37% and 92.16%, respectively. The algorithms are further evaluated on real-world test cases to assess their performance further. The results strengthen the belief regarding the applicability of motion sensors and DL algorithms for the behavioral analysis of mental disorders such as anxiety and OCD. This approach may further be applied to other areas of mental care such as ASD, Alzheimer's, and stress disorder.