Department of Telecommunications, Faculty of Electrical Engineering and Communication, Brno
Brno University of Technology, Brno
International Clinical Research Center, St. Anne’s University Hospital Brno
Research objectives:
– To identify digital speech/voice biomarkers facilitating the prodromal diagnosis of DLB
Regarding the speech/voice biomarkers, the state of the art covers mainly the discrimination of healthy controls (HC), MCI and AD. In the frame of this WP, we will move further and bridge the knowledge gap associated with the acoustic analysis of DLB and its early diagnosis. We will find an optimal combination of speech/voice tasks and features that will provide good discrimination power, and that will also provide good clinical interpretability. The acoustic measures will be correlated with scores of neuropsychological assessments.
Hypothesis: Based on the state of the art in the acoustic analysis of dementia, we assume that mainly speech features quantifying prosody (more specifically speech rate and pausing – temporal features) will play a significant role in the prodromal diagnosis of DLB.
– To identify digital handwriting/drawing biomarkers facilitating prodromal diagnosis of DLB
We are going to explore the impact of computerised drawing/handwriting processing on the prodromal diagnosis of DLB. Similarly, to Aim 1, we will find an optimal combination of handwriting/drawing tasks and features that will provide good discrimination power. The outcomes of this part will be discussed in relation to the neuropsychological profile of a patient.
Hypothesis: We assume that cognitive tasks such as the pentagon copy test will enable us to effectively quantify visuospatial deficits. Analysis of in-air movement recorded during a sentence copy task will probably bring some valuable information as well.
-To train machine-learning models enabling supportive prodromal diagnosis of DLB
We are going to train and evaluate state-of-the-art machine learning models supporting the prodromal diagnosis of DLB. The models will be evaluated in terms of sensitivity and specificity and will be interpreted using feature importances or Shapley values. The models will be trained for each modality separately, as well as for both modalities together.
Hypothesis: We hypothesise that a model based on a gradient boosting algorithm will reach more than 70% sensitivity/specificity when fed by features from both modalities.