Repository for the exam project in the course Social and Cultural Dynamics

---

⚥ Gender differences in vocal markers of autism; Investigating generalizability of ML classifiers across gender groups

Individuals with Autism Spectrum Disorder (ASD) tend to exhibit atypical speech and prosody patterns, which have been researched as potential markers of ASD. Machine learning (ML) has been increasingly employed to identify such markers, given its ability to capture complex signals, however, many studies have primarily focused on male participants, leaving it unclear how such models would generalize to females who remain underrepresented due to diagnostic and systemic biases.

This exam project investigated gender differences in vocal markers of ASD by examining cross-gender generalizability of ML models. Using gender-balanced data from a single study, three SVM models were trained on male-only, female-only and mixed-gender data, and their performance on both a female and a male test sample was assessed. The male-only model showed moderate generalization to both test sets, the female-only model failed to learn any meaningful patterns, and the mixed-gender model reached the most robust results, particularly on the female test set.

While these findings are limited by dataset size and methodological constraints, they highlight some of the gender disparities, the importance of more heterogeneous training data, and the need for more fine-grained approaches in studying the ‘autistic voice’. Future research should focus on developing robust generalizability procedures, implementing inclusive recruitment strategies and collaborating with clinical experts to study clinically relevant research questions.

---

📁 Repository Structure

r_code/

• grand.Rmd
  main script that handles data partitioning, min-max scaling, and feature selection via Elastic Net
• grand_function.R
  contains helper functions used in grand.Rmd for preprocessing steps

python_code/

• main.py
  main script for training SVM models, performing hyperparameter tuning, testing on held-out data, and evaluating performance
• functions.py
  contains functions for model training, evaluation, and utility operations used in main.py
• get_roc.py
  script to calculate ROC AUC scores for the trained models

data/

• contains key data outputs generated by the project:
  • selected stable features per gender subset (after Elastic Net selection)
  • trained model files per gender subset and fold