**11. Conclusions**

Parkinson's disease (PD) detection is an active area of research. These efforts are oriented to assist the provision of a better quality of life for PD patients. Vocal-based detection of PD is a non-invasive and inexpensive alternative for the early detection of the disease. According to neurology studies, the female brain and the male brain are functionally different and this is the motivation to conduct separate studies, according to gender. Fortunately, the availability of large datasets allows such research efforts. The work by Solana-Lavalle et al. [8, 19] is based on the largest publicly available dataset to train and test different classifier so that separate studies for male and female patients are carried out. Experiment results show that the most relevant features for accurate classification are highly dependent on gender. In the case of male patients, low-frequency voice content is the most significant, while for female patients, high-frequencies give better results. Most features selected in the feature selection process are extracted by using the *Tunnable Q-factor Wavelet Transform* (TQWT) and the *Mel Frequency Cepstral Coefficients* (MFCC). Both groups of features are obtained through the use of banks of filters, where these extraction mechanisms operate in a similar way the human auditory system does. The accuracy obtained by the classifying algorithms reaches up to 95.9%, showing the best results with the male population. Also, a statistical analysis of the variability of the most significant features, from each gender, is done to assist the clinical interpretation of the classification result (PD positive and PD negative).

Another method to detect neurological alterations is through medical images such as *Magnetic Resonance Imaging*, *DaTscan*, and *Diffusion Tensor Imaging*. Physicians have used these images modalities to help diagnose PD. However, they rely on visual inspection, which is prone to misdiagnosis due to human error. For this reason, a quantitative analysis of these images is suggested. Solana et al. [6] proposed a method for using structural MRI combined with signal processing and machine learning classifiers to assist the diagnosis of PD. This method achieves

competitive results and insights. The classification results deliver an accuracy of 99.01% in male patients and 96.97% in female patients.

*Voxel-Based Morphometry* is a statistical study that has been used to identify brain regions that show differences between PD patients and controls. Features, based on first-order (histogram) and second-order statistics (co-occurrence matrix), have been extracted from the regions of interest identified by VBM. Since the number of features, extracted from multiple regions of interest, is very large, feature selection techniques have to be used such as *wrapper for feature selection*. The aim of using *feature subset selection* is to identify the most important features for discrimination and to reduce computational complexity. Regions of interest for PD detection usually include the striatum. However, by using feature subset selection it has been possible to identify several regions, outside the striatum, suggesting an affectation in those areas of the brain. These regions include the *somatosensory cortex*, *temporal gyrus*, and *cerebellum*.

Future work on the detection of PD could make use of other imaging techniques such as *Functional Magnetic Resonance Imaging* and *Diffusion Tensor Imaging*. These imaging techniques provide information about the activity within the brain, and about the connectivity of the brain respectively. Thus, these modalities are good candidates to provide new information about PD and an alternative to assist the physicians with early detection of the disease. On the other hand, some of the best classification results in voice recordings are obtained using *deep learning* techniques which demand the availability of a larger dataset. To the best of our knowledge, deep learning has not been applied to the largest dataset from Sakar et al. [7] and could be an opportunity to compare these new learning techniques with classical approaches.
