Organisers: Alberto Parola and Riccardo Fusaroli

Workshop Summary
Machine learning applications to voice and language analysis have great potential for developing digital phenotyping of mental disorders. Indeed, quantitative analysis of voice and speech samples can both scaffold the current evaluation of mental disorders and provide a finer perspective on clinical, social, and cognitive dimensions that characterize these conditions. 

However, current applications of ML to the analysis of speech and language disorders in mental illness are subject to important limitations. The main limitations concern: a) the low generalizability of machine learning (ML) and natural language processing (NLP) models to new samples, tasks, and languages; b) the presence of potential biases, which are especially harmful for minorities that are underrepresented in training data c) the limited clinical applicability and interpretability of ML model.  

The aim of the workshop is to bring together experts in clinical applications of machine learning and NLP to discuss what knowledge, resources and tools we need to transform speech and voice markers into clinical useful psychiatric applications.  In particular, the focus is to discuss how to provide a solid foundation for developing robust and generalizable ML data analysis pipelines, for assessing the presence of potential bias and account for heterogeneous performance, and, more generally, for developing fruitful recommendations for greater accountability in ML, speech signal processing (SSP), and NLP clinical research.

Location: Interacting Mind Center, Aarhus University
Jens Chr. Skous Vej 4, 8000 Aarhus C, building 1483, room 312 and online

Time: November 25, at 11:00 - 16.30

Hybrid event

Join Zoom Meeting: https://aarhusuniversity.zoom.us/j/3843004426
Meeting ID: 384 300 4426

Program
11.00-11.15: Introduction to the workshop. Riccardo Fusaroli, Aarhus University.
11.15-12.00: “Acoustic markers of schizophrenia: a meta-analysis based assessment of cross-linguistic generalizability of ML and univariate models”. Alberto Parola, Aarhus University.
12.00-13.00: Lunch break
13.00-13.45: “Machine learning for medical imaging: methodological failures and recommendations for the future”. Veronika Cheplygina, IT University of Copenhagen
13.45-14.15: “Language as a fingerprint: modeling individual traits using large language models”. Roberta Rocca, Aarhus University. 
14.15-14.30: “How to evaluate clinical word embeddings”. Kenneth Enevoldsen, Aarhus University
14.30-15:00: "Integrating feature importance across models" - Daniel Low, Harvard University & MIT
15.00-16.30: Discussion and project ideas (Guest: Konstantinos Sechidis, Machine Learning Researcher. Advanced Methodology and Data Science Group, Novartis)

Please sign up, and if you want a lunch sandwich send an email to alberto.parola@cc.au.dk before Thursday 24.11.2022.