AIPD-DC4-KIE

• Host Institution: ki:elements
  
• PhD Enrolment: University of Bonn
• Start Date: October 2025
  
• Duration: 36 months
• Official PhD Supervisor: Holger Fröhlich

Research Objectives

KIE has developed techniques to detect deviations in speech from PD patients compared to healthy controls. FH has developed a novel AI method (VaDER) to cluster multivariate clinical disease trajectories in PD patients, which may potentially contain missing values. VaDER revealed significant differences in L-DOPA treatment response between identified progression clusters in PPMI. The objective of this project is to further explore the potential of speech biomarkers for early diagnosis, early detection of fatigue, response to L-DOPA treatment, and monitoring of progression. We will employ data from mPower (English speakers), PC-GITA (Spanish speakers), and digital voice recordings within the LuxPARK study (Luxembourgish speakers). In addition to digital voice recordings, all datasets comprise questionnairebased clinical assessments (MDS-UPDRS), which also cover a question related to fatigue. Different AI/ML classifiers will be trained and evaluated on digital voice to discriminate between a) patients in a prodromal phase (iRBD) and PD (early diagnosis) using data from LuxPARK; b) patients with and without a diagnosis of fatigue (early fatigue detection) using the newly acquired data from LuxPARK, mPower and, PC-GITA; c) treatment ON and OFF states (treatment response detection) using separate data from PC-GITA and mPower; d) different PD progression subtypes (progression biomarker) identified in earlier work by FH in the LuxPARK study. Following voice feature extraction, we will train and evaluate a panel of ML algorithms, such as XGBoost, Random Forests, SVM, and penalized logistic regression. We will investigate the impact of individual features on model predictions using XAI techniques such as (causal) SHAP and we will explore how voice features relevant for the same medical endpoint differ across languages. To better understand the principal value of digital voice as a predictor of fatigue and treatment response we will also make a comparison against digital gait features, which are jointly available in mPower. Finally, we will also explore a multimodal combination of voice and gait. Due to the relatedness of projects, DC4 will work closely together with DC2 and DC12.

Expected Results

• Innovative AI/ML models for early detection of PD symptoms (speech impairments, fatigue) and treatment response on an individual patient level.
• Better understanding of the relationship between digital speech biomarkers and clinical disease progression.
• Better understanding of the generalizability of speech biomarkers across different languages.
   

Planned Secondment(s)

• Host 1: Fraunhofer SCAI
  
  • Duration: 9 months
    
  • Purpose: Learning about AI/ML progression models
• Host 2: Novo Nordisk
  • Duration: 14 months
    
  • Purpose: Learning about digital biomarkers in pharma

This project is part of the "Digital Health" work package.

• Johannes Tröger, Ebru Baykara, Jian Zhao, Daphne ter Huurne, Nina Possemis, Elisa Mallick, Simona Schäfer, Louisa Schwed, Mario Mina, Nicklas Linz, Inez Ramakers, Craig Ritchie; Validation of the Remote Automated ki:e Speech Biomarker for Cognition in Mild Cognitive Impairment: Verification and Validation following DiME V3 Framework. Digit Biomark 30 November 2022; 6 (3): 107–116. https://doi.org/10.1159/000526471.