A Novel Bio-inspired Signal Processing Framework and its Applications in Data Analysis and Digital Phenotyping

Abstract

Neuropsychiatric disease healthcare and Machine learning speak two different languages, entailing a translational effort that fills the explanatory gap between the two domains, thereby reinforcing the generation of new knowledge-based systems and pivoting the partnership between clinical practice in neuropsychiatry and empirical research in Artificial Intelligence (AI). In particular, the surge in the use of digital technology that coincided with advancement in computational power and AI algorithms, opened a new path in medical diagnostics that explores a new class of digital biomarkers for the early detection of neuropsychiatric symptoms that are otherwise hard to detect by the conventional, questionnaire-based clinical techniques. This Ph.D. Dissertation mainly focuses on a special class of digital biomarkers, referred to as keystroke dynamics, that quantify the fine motor skills of subjects while they type on their smartphone keyboards/touchscreens. This is because the decline in fine motor skills, also referred to as psychomotor retardation, which is one of the early symptoms of depression. For this endeavor, this Ph.D. Dissertation introduces the basis of a new research trend, that of Explainable Digital Phenotyping, with the aim of establishing a paradigm shift from the well-known data-driven methods to interpretable machine learning pipelines that besides diagnosis, generate interpretable explanations of neuropsychiatric disease manifestations deciphered from data collected in-the-wild. The main gap/problem that is targeted herein is the lack of bio-inspired signal processing frameworks that could provide human-like perceptual capabilities resulting in efficient and fine-grained analysis of bio-signals. To reach this aim, this PhD Dissertation contributes (1) novel bio-inspired signal processing framework, (2) data representation methods based on keystroke dynamics sonification and (3) new auditory music-based dataset to solve biomedical data scarcity issues. Because the main objective is to detect neuropsychiatric perturbation from keystroke dynamics, which are time-sensitive, we believe that auditory representation outperforms visual representation, a motivation behind the auditory inspired data-representation of keystroke dynamics proposed in this Ph.D. Dissertation. Moreover, this motivated the development of bio-inspired signal processing framework, namely the Cochlear Transform (CT), Cochlear Cepstrum and its higher order extensions, the Cochlear Bispectrum. Finally, we demonstrate applications of our contributions in sleep spindle detection from EEG signals and depression detection from the auditory representation of keystroke dynamics. While we witnessed progress in some specialized tasks in machine audition research and auditory problem domains are now enormous, we still see that artificial auditory systems lack the generalizability and the robustness of the human auditory cortex. As we are gaining more and more insight into the functional and the structural mechanisms of the human auditory cortex, mainly through neuroscience and fMRI studies, we believe that time is ripe now for a new endeavor to bring the biological capabilities to computer audition, primarily in terms of robustness and generality. i

Date of Award	22 Jul 2024
Original language	American English
Supervisor	AHSAN Khandoker (Supervisor)