Beyond Psychiatric Scales: A Novel Approach to Conceptualize Mental Disorders Using Networks, Reinforcement Learning, and NLP.
MetadataShow full item record
CitationCampbell, Stephanie. 2019. Beyond Psychiatric Scales: A Novel Approach to Conceptualize Mental Disorders Using Networks, Reinforcement Learning, and NLP.. Bachelor's thesis, Harvard College.
AbstractEmerging techniques in data analysis, machine learning, and artificial intelligence have the potential to redefine the study and treatment of mental disorders. The field of psychiatry relies heavily on self-reported and observed symptoms since there is insufficient understanding of the complex models of the biological circuitry affected by mental illness. The arbitrary nature of the default practice of checklist diagnostics in psychiatry is problematic (Rutter et al., 2011). Applying advanced computational approaches to psychiatric data has the potential to determine underlying patterns across behaviors and experiences of individuals with mental disorders (Grisanzio et al., 2018). We collected behavior task and transdiagnostic self-report measures data sets from nonclinical subjects using a large-scale data collection on Amazon Mechanical Turk. A reinforcement learning model fit the sequential decision-making task data to determine the arbitration between model-based and model-free control for each participant. We used network analysis to visualize the interplay between self-reported symptoms, traits, and factors of psychiatric disorders and adaptive, goal-directed learning. We used standard dimensionality reduction techniques—like factor analysis and principal component analysis—as well as graphical models for regularized partial correlation networks. We compared our estimations from a regularized partial correlation network, a relative importance network, and other graphical methods to estimate the functional relations between the self-reported variables. We transformed the psychiatric questionnaire text items into sentence embedding vectors via Google's pre-trained Universal Sentence Encoder, a variation of the Transformer model, to better understand the effects of the semantic similarity between measures (Cer et al., 2018). We used these semantic embeddings in linear models and symptom networks, as a new method of finding redundant measurements and as a means of exploring the overlap between various psychiatric scales. This research explores how different psychiatric symptoms interact, which—in turn—can help improve measure interpretation. We hope that these computational perspectives on psychiatric disorders may offer insights towards more targeted diagnoses and interventions in psychopathology.
Citable link to this pagehttps://nrs.harvard.edu/URN-3:HUL.INSTREPOS:37364586
- FAS Theses and Dissertations