Extracting digital biomarkers for classifying Parkinson's Disease with voice recordings from mobile phones

INTRODUCTION Due to varying symptoms, early diagnosis of Parkinson's disease (PD) and objective measurement of medication effect is challenging. We investigated which features of mobile phone voice recordings can differentiate between participants with and without PD (PDCLASS) and identify pre- and post-medication conditions (PRE/POST). METHODS We formed 4 classes ‘With PD’ (3758), ‘Without PD’ (4001), ‘BeforeMedication’ (1324) and ‘AfterMedication’ (1602) from 10-second recordings of /a/-vowel phonation from the mPower dataset [2]. We extracted 60 features and grouped them into jitter, shimmer, non-linear dysphonia, mel-cepstral-coefficients, amplitude, frequency-power, temporal, wavelet, pitch, and tremor. To identify the most salient features, we performed a cyclical analysis using a multilayer perceptron classifier for each classification. Groups were added one-by-one during a 5-fold cross-validation, updating the feature vector if accuracy increased (Figure 1). The most informative groups were those that the most often increased accuracy. Features were then removed one-by-one and, if the average accuracy remained unchanged or increased, removed definitively. Then all remaining unused features were randomly added one-by-one and, if the average accuracy increased, included. Finally, we build two classifiers with the selected features and demographics (age, sex). RESULTS The number of features reduced from 60 to 7 (PDCLASS) from 3 groups (amplitude, temporal and pitch) and to 9 (PRE/POST) from 4 groups (tremor, mel cepstral-coefficients, frequency power and pitch). The AUC of PDCLASS was 95.1%, which was slightly higher than the 82.6% achieved with 5 features plus demographics in [3]. The AUC for PRE/POST was 59.1%. CONCLUSION PDCLASS requires different voice features than PRE/POST medication detection.