Niloofar Momeni presents her licentiate thesis

Reliable AI for Parkinson’s Disease Detection Using Voice Data

Abstract:

Parkinson’s disease (PD) is a neurodegenerative disorder that affects speech and voice, making them promising non-invasive biomarkers for early detection. However, building trustworthy detection systems remains challenging due to data bias and limited interpretability. This thesis addresses these challenges by proposing machine learning methods for anomaly detection, bias mitigation, and interpretability to enable fair and generalizable PD detection.  

In Paper I, we introduced Group-Wise Scaling (GWS), a feature normalization strategy to reduce demographic bias by accounting for age- and sex-related variability in vocal features. This approach improved representation and model performance compared to standard scaling.  

In Paper II, we combined GWS with anomaly detection using isolation forests to improve predictive performance by discarding corrupted recordings and enhancing model reliability. The approach demonstrated strong generalizability across multiple datasets, resulting in a more trustworthy PD detection pipeline. Moreover, model interpretability was examined using SHAP analysis, which highlighted the vocal features influencing the PD predictions. The interpretability results are consistent with clinically validated evidence.

In Paper III, we proposed the Age-Aligned Validation (AAV) protocol, a model-agnostic strategy that balances age distributions during validation. AAV mitigates age bias, improving fairness and reliability, and was shown to be effective across diverse models and datasets, enhancing generalization to unseen cohorts.  

In this thesis, experiments on multiple internal and external datasets demonstrated that the proposed methods collectively improve accuracy, fairness, and generalizability. Overall, the thesis contributes practical, interpretable, and bias-aware machine learning approaches that advance the development of trustworthy voice-based monitoring tools for Parkinson’s disease. 

A licentiate thesis in Mathematical statistics, 60 hp.

Opponent: Dr Ina Kodrasi, Idiap Research Institute.

Examiner: Professor Carl Olsson, Centre for Mathematical Sciences.

Supervisors: Andreas Jakobsson, Susanna Whitling, and Maria Sandsten

The thesis is available in the library, Centre for Mathematical Sciences.