Objective: The aim of the clinical study was the assessment of efficiency of the screening diagnositics method for Parkinson’s disease (PD) by voice using AI-based technology (BRAINPHONE project). The method used is convolutional neural network (CNN), trained with more than 1000 unique audiorecords of voices of PD patients vs healthy volunteers (50/50%).

Background: Recently, a large number of publications have appeared on the testing of digital tools for the diagnosis of PD [1-4]. Here we represent our clinical experience.

Method: In the cohort clinical study 50 patients were included in a continuous manner. Every patient was assessed by the experienced neurologist and the diagnosis for “PD” with Hoehn and Yahr staging (group P) or “non-PD” (control group C) was made in every case according to MDS clinical criteria for PD [5]. The clinical trial was performed on the base of Republican Consultative and Diagnostics Center for Movement Disorders (Kazan, Russia). After the clinician’s assessment the digital diagnostics tool one was made. The trained CNN output was “possible PD” (counted probability of 0.5 and higher) and “non-PD” (counted probability less than 0.5). From the sample were excluded patients with atypical and other types of parkinsonism and with disorders associated with speech disturbances. The statistical analysis was performed with DATAtab [6].

Results: General characteristics of the evaluated population are presented in the Table 1.

The two-factor ANOVA showed that there were no significant difference between the groups related to age (p = 0,475). There are results of the AI-based diagnostics shown on the Figure 1.

Logistic regression analysis shows that the model as a whole is significant (Chi²(1) = 30,28, p < 0.001, n = 50).

Calculated area under curve (AUC) counted by ROC analysis for the represented binary classifier (threshold 0.5) was 0.902.

Conclusion: AI-based voice diagnostics via trained NN shows good results in primary clinical trial with high AUC result. Potentially, diagnostics using trained CNN can be applied as a screening tool.

The limitations of the study are almost balanced testing sample (vs 99,5-98/0.5-2% of PD and healthy persons in real-life depending on age and area); people with speech disturbances and other types of parkinsonism were excluded.

ROC curve for the AI-based classifier

Results of the AI-based diagnostics of PD

General characteristics of evaluated population

References: 1. Ahmed I. et al. Classification of Parkinson disease based on patient’s voice signal using machine learning //Intelligent Automation and Soft Computing. – 2022. – Т. 32. – №. 2. – С. 705.
2. Nagasubramanian G., Sankayya M. Multi-variate vocal data analysis for detection of Parkinson disease using deep learning //Neural Computing and Applications. – 2021. – Т. 33. – №. 10. – С. 4849-4864.
3. Amato F. et al. Machine learning-and statistical-based voice analysis of Parkinson’s disease patients: A survey //Expert Systems with Applications. – 2023. – Т. 219. – С. 119651.
4. Costantini G. et al. Artificial intelligence-based voice assessment of patients with Parkinson’s disease off and on treatment: machine vs. deep-learning comparison //Sensors. – 2023. – Т. 23. – №. 4. – С. 2293.
5. Postuma R. B. et al. MDS clinical diagnostic criteria for Parkinson’s disease //Movement disorders. – 2015. – Т. 30. – №. 12. – С. 1591-1601.
6. DATAtab Team (2024). DATAtab: Online Statistics Calculator. DATAtab e.U. Graz, Austria. URL https://datatab.net

« Back to 2024 International Congress

MDS Abstracts - https://www.mdsabstracts.org/abstract/efficiency-of-ai-based-service-for-screening-diagnostics-of-parkinsons-disease-brainphone-project/