Strengths and Limitations of Computerized PD Diagnosis from Voice

Towards a Corpus (and Language)-Independent Screening of Parkinson’s Disease from Voice and Speech through Domain Adaptation

“…The precision of these experiments is also summarised in Figure 4 compared to those obtained for the baseline architectures using a single corpus, as reported in Table 3 using TL. The bar graphs in Figure 4 reveal a noticeable trend: a decrease in accuracy for the baseline architectures trained and tested, mixing the different corpora available, which is also consistent with the results reported in previous studies [28,54]. However, the DA network effectively mitigates this trend.…”

“…This has led researchers focused on applying DL methods to combine data from several sources, which were recorded in different conditions and from speakers with different demographic characteristics (including their mother tongue). To this respect, and in order to address the generalisation capabilities of trained models, the authors in [21,27,28] presented crossdataset experiments, reporting significant drops in precision of more than 20 absolute points when using different corpora for testing and training. As shown in [29,30], although the combination of multiple datasets is intended to model the representation space better and to avoid overfitting due to data scarcity, it can also induce certain unwanted behaviours, since DL approaches typically make use of shortcut learning strategies capable of reducing the training loss function by learning characteristics associated with the dataset (e.g., the language, microphone, recording equipment, acoustic environment, etc.)…”