Dataset Information

Analysis of Smartphone Recordings in Time, Frequency, and Cepstral Domains to Classify Parkinson's Disease.

ABSTRACT: OBJECTIVES:Parkinson's disease (PD) is the second most common neurodegenerative disorder; it affects more than 10 million people worldwide. Detecting PD usually requires a professional assessment by an expert, and investigation of the voice as a biomarker of the disease could be effective in speeding up the diagnostic process. METHODS:We present our methodology in which we distinguish PD patients from healthy controls (HC) using a large sample of 18,210 smartphone recordings. Those recordings were processed by an audio processing technique to create a final dataset of 80,594 instances and 138 features from the time, frequency, and cepstral domains. This dataset was preprocessed and normalized to create baseline machinelearning models using four classifiers, namely, linear support vector machine, K-nearest neighbor, random forest, and extreme gradient boosting (XGBoost). We divided our dataset into training and held-out test sets. Then we used stratified 5-fold cross-validation and four performance measures: accuracy, sensitivity, specificity, and F1-score to assess the performance of the models. We applied two feature selection methods, analysis of variance (ANOVA) and least absolute shrinkage and selection operator (LASSO), to reduce the dimensionality of the dataset by selecting the best subset of features that maximizes the performance of the classifiers. RESULTS:LASSO outperformed ANOVA with almost the same number of features. With 33 features, XGBoost achieved a maximum accuracy of 95.31% on training data, and 95.78% by predicting unseen data. CONCLUSIONS:Developing a smartphone-based system that implements machine-learning techniques is an effective way to diagnose PD using the voice as a biomarker.

SUBMITTER: Tougui I

PROVIDER: S-EPMC7674819 | biostudies-literature | 2020 Oct

REPOSITORIES: biostudies-literature

ACCESS DATA

Publications

Analysis of Smartphone Recordings in Time, Frequency, and Cepstral Domains to Classify Parkinson's Disease.

Tougui Ilias I Jilbab Abdelilah A Mhamdi Jamal El JE

Healthcare informatics research 20201031 4

<h4>Objectives</h4>Parkinson's disease (PD) is the second most common neurodegenerative disorder; it affects more than 10 million people worldwide. Detecting PD usually requires a professional assessment by an expert, and investigation of the voice as a biomarker of the disease could be effective in speeding up the diagnostic process.<h4>Methods</h4>We present our methodology in which we distinguish PD patients from healthy controls (HC) using a large sample of 18,210 smartphone recordings. Thos ...[more]

PMID: 33190461

Similar Datasets

Project description:Background: Galvanic vestibular stimulation (GVS) is being increasingly explored as a non-invasive brain stimulation technique to treat symptoms in Parkinson's disease (PD). To date, behavioral GVS effects in PD have been explored with only two stimulus types, direct current and random noise (RN). The interaction between GVS effects and anti-parkinsonian medication is unknown. In the present study, we designed multisine (ms) stimuli and investigated the effects of ms and RN GVS on motor response time. In comparison to the RN stimulus, the ms stimuli contained sinusoidal components only at a set of desired frequencies and the phases were optimized to improve participants' comfort. We hypothesized GVS motor effects were a function of stimulation frequency, and specifically, that band-limited ms-GVS would result in better motor performance than conventionally used broadband RN-GVS. Materials and Methods: Eighteen PD patients (PDMOFF/PDMON: off-/on-levodopa medication) and 20 healthy controls (HC) performed a simple reaction time task while receiving sub-threshold GVS. Each participant underwent nine stimulation conditions: off-stimulation, RN (4-200 Hz), ms-θ (4-8 Hz), ms-α (8-13 Hz), ms-β (13-30 Hz), ms-γ (30-50 Hz), ms-h1 (50-100 Hz), ms-h2 (100-150 Hz), and ms-h3 (150-200 Hz). Results: The ms-γ resulted in shorter response time (RPT) in both PDMOFF and HC groups compared with the RN. In addition, the RPT of the PDMOFF group decreased during the ms-β while the RPT of the HC group decreased during the ms-α, ms-h1, ms-h2, and ms-h3. There was considerable inter-subject variability in the optimum stimulus type, although the frequency range tended to fall within 8-100 Hz. Levodopa medication significantly reduced the baseline RPT of the PD patients. In contrast to the off-medication state, GVS did not significantly change RPT of the PD patients in the on-medication state. Conclusions: Using band-limited ms-GVS, we demonstrated that the GVS frequency for the best RPT varied considerably across participants and was >30 Hz for half of the PDMOFF patients. Moreover, dopaminergic medication was found to influence GVS effects in PD patients. Our results indicate the common "one-size-fits-all" RN approach is suboptimal for PD, and therefore personalized stimuli aiming to address this variability is warranted to improve GVS effects.

Project description:To compare measurements of the full-field photopic negative response (PhNR), as well as intra-subject variation in the PhNR, using time and time-frequency domain analyses.Full-field ERGs were recorded from 20 normally sighted subjects (aged 24-65 years) elicited by a long-wavelength pulse (3 cd s m-2) presented against a short-wavelength adapting field (12.5 cd m-2). Three to 10 waveforms were obtained from each subject, and each waveform was analyzed using standard time domain analyses of the PhNR, as well as a discrete wavelet transform (DWT) to extract time-frequency components that correspond to the PhNR. Three different measures of the PhNR were derived and compared: (1) amplitude at the PhNR trough; (2) amplitude at 72 ms following stimulus onset; (3) energy in the 11 Hz, 60-120 ms DWT frequency bin that corresponds to the PhNR. In addition, the effect of normalizing the PhNR by the b-wave was evaluated for each of the measures. Coefficients of variation (CVs) were computed for each definition to evaluate intra-subject variation.PhNR amplitudes measured at the trough and at 72 ms were significantly correlated (r = 0.88, p < 0.001). Additionally, PhNR energy derived by DWT was significantly correlated with the amplitude measured at the trough (r = 0.64, p = 0.002) and at 72 ms (r = 0.60, p = 0.005). Mean (±SD) intra-subject CVs were 26 % (15 %), 49 % (26 %), and 30 % (15 %), for measures at the trough, 72 ms, and DWT, respectively. Normalization by the b-wave amplitude (i.e., PhNR/b) had minimal effect on the intra-subject CVs, whereas normalization by the sum of the b-wave and PhNR amplitudes (i.e., PhNR/[b + PhNR]) substantially reduced the CVs for all three measures (mean CVs were less than 17 % for all conditions).Although each PhNR definition has advantages and disadvantages, all three metrics provide similar estimates of the PhNR. Intra-subject CVs, however, were relatively high for measurements made at 72 ms, indicating that definitions based on a fixed time point may introduce variability. The substantial decrease in intra-subject variation after normalization by the sum of the PhNR and b-wave amplitudes may be advantageous under some conditions.

Dataset Information

Analysis of Smartphone Recordings in Time, Frequency, and Cepstral Domains to Classify Parkinson's Disease.

Publications

Analysis of Smartphone Recordings in Time, Frequency, and Cepstral Domains to Classify Parkinson's Disease.

Similar Datasets

OmicsDI is part of the ELIXIR infrastructure