Banca de QUALIFICAÇÃO: MARINA TATIANE RIBEIRO DA SILVA

Uma banca de QUALIFICAÇÃO de DOUTORADO foi cadastrada pelo programa.
STUDENT : MARINA TATIANE RIBEIRO DA SILVA
DATE: 02/09/2024
TIME: 18:00
LOCAL: Google Meet
TITLE:

A METHODOLOGY FOR EXPLAINING COMPUTATIONAL PSYCHIATRIC DIAGNOSES WITH LARGE LANGUAGE MODELS, INTEGRATED GRADIENTS, AND LINGUISTIC ANALYSIS

KEY WORDS:

Computational psychiatry; Natural Language Processing; Large Language Models; Integrated Gradients; Bidirectional Encoder Representations from Transformers - BERT; Alzheimer's Disease.

PAGES: 10
BIG AREA: Ciências Biológicas
AREA: Biologia Geral
SUMMARY:

Brain disorders, such as Alzheimer's Disease (AD) and Schizophrenia, impact the language output of those affected by the disease in numerous ways that are still not fully known. Recent advances in computational psychiatry, a fresh discipline in neuroinformatics, have approached this subject by providing multiple strategies for automatically identifying one's eventual disorder through analyzing the discourse. Natural Language Processing (NLP) methods have achieved extraordinary accuracy in this task. However, for actual clinical suitability, there needs to be more clarity on which aspects of the speech were determinant for the automatic diagnostic decision. Here, we describe a methodology for achieving syntax-related explainability on Large Language Models (LLMs) in text classification tasks of interest to the computational psychiatric community. The method uses Integrated Gradients (IG) attribution to identify the segments of the text more relevant to the decision process and the Linguistic Inquiry and Word Count (LIWC) toolkit to annotate these segments with appropriate syntactic and linguistic descriptors. On a study level, the methodology can pinpoint which descriptors are statistically pertinent for the diagnoses, whereas on an individual analysis, it can describe the relevant segments for the decision. We demonstrate the use of the methods with an English dataset of audio recordings and transcripts from the Cookie Theft picture description task and a fine-tuned BERT (Bidirectional Encoder Representations from Transformers) model that achieved an accuracy of 87% in a 5-fold cross-validation method. We discuss how to apply the methodology in scientific and clinical settings and its limitations.

COMMITTEE MEMBERS:
Externa à Instituição - ALINE VILLAVICENCIO
Externa à Instituição - LILIAN CRISTINE HÜBNER - PUCRS
Presidente - 3083298 - RENAN CIPRIANO MOIOLI