Fractal dimensions and machine learning for detection of Parkinson’s disease in resting-state electroencephalography

Parkinson’s disease (PD) is a long-term progressive neurodegenerative disorder that affects a part of the brain called the substantia nigra, which lies in the midbrain region [1]. The neurons in the midbrain start dying, leading to dopamine deficit. PD affects more than 10 million people globally [2]. The main symptoms of PD are rhythmic shaking (tremors) that originates from the limbs, slowed movement (bradykinesia), stiffness (co-wheel rigidity), loss of balance, postural instability, and slurred speech [3, 4]. Although PD is an incurable condition, treatments include dopamine boosters, antidepressants, and other medications that help alleviate the symptoms.

PD can manifest in multiple forms, which may remain undetected for several years and gradually progress till the symptoms become discernible [5]. Unified Parkinson’s disease rating scale (UDPRS) [6] and the Hoehn and Yahr [7] (HY) rating are the two most widely used measures to assess PD. The UDPRS consists of the following scales: These scales are assessed by interrogation and medical observation of the patient. The HY system is similar to UDPRS and has been essentially supplanted by UDPRS.

The UDPRS diagnosis is non-automated and is subject to human error [6]. Consequently, there is a growing need to automate the methods for effective PD detection. Machine learning (ML) and deep learning (DL) are immensely useful in building automated pipelines for the detection and diagnosis of neurological conditions like Parkinson’s disease (PD), especially when there is a need to process enormous data volumes, which is manually impossible to analyze [8].

However, neurological data exhibit high dimensionality with noise, which impedes the performance of ML and DL models [9‐11]. In such cases, a suitable feature extraction technique becomes essential to extract the underlying salient features from the data to give better results with higher efficiency.

As observed from the current research, two types of feature extraction and analysis methods are broadly used for capturing features from EEG, namely, linear and nonlinear. Plenty of research has been done on Parkinson’s disease (PD) by examining EEG signals, using linear analysis methods such as wavelet transforms [12, 13], Fourier transforms [14, 15], and autoregressive methods [16, 17] to extract features. Among nonlinear methods, numerous studies based on entropy [18, 19], Lyapunov exponent [20], Hurst exponent [21, 22], and fractal dimensions (FD) have been proposed to understand crucial aspects of Parkinson’s disease (PD) by deriving insights from the EEG signal. However, fractal dimensions have several key advantages over previously mentioned measures as mentioned below: -Though fractal dimensions provide an effective way for feature extraction from the EEG signal, the use of various fractal dimensional methods for detecting Parkinson’s disease has not been deeply explored in the previous studies. A discernible gap exists in the literature concerning the comparative efficacy of diverse fractal dimensional methods in accurately capturing the essence of Parkinson’s disease from resting-state EEG signals. Our study aims to fill the gap and throw light upon the efficiency of fractal dimensional methods in detecting Parkinson’s disease. Thus, this study also shows that fractal dimensions (FD) can be used for feature extraction for all EEG-based studies.

The objectives designed for this study are as follows:The study’s objectives are structured to address the following research question (RQ):

RQ: How effectively can fractal dimensions, as feature extraction measures, discriminate Parkinson’s disease (PD) patients who are ON or OFF medication from healthy controls (HC)?

The rest of this paper is organized as follows. Section 2 comprehensively reviews the state-of-the-art Parkinson’s disease (PD) detection techniques. The overall methodology and solution design to construct an optimal pipeline by experimenting with various pre-processing, segmentation, feature extraction, and machine learning techniques are presented in Sect. 3. Section 4 unveils the study’s findings. These findings are then elaborated upon in Sect. 5. Finally, Sect. 6 derives inferences from the observations while throwing light upon the contribution of this study to the existing literature.

Numerous works have employed machine learning (ML) and deep learning (DL) for PD detection [27‐29]. The data retrieved utilizing feature extraction techniques, including partial directed coherence (PDC), have been used to train the machine learning (ML) models [30]. PDC is a Granger causality measure applied to EEG data in the frequency domain to infer the magnitude of information flow over the brain [31]. PDC and random forest (RF) classifier have been implemented for detecting early-stage PD [30]. A novel method combining PDC with the analysis of blood oxygenation level-dependent (BOLD) signal variations has been developed to comprehend the regional peculiarities of oscillatory processes in the human brain and to examine functional connectivity in resting-state EEG-fMRI [31]. The same approach can be applied to detect Parkinson’s disease (PD). Similar to the previous studies that make use of machine learning (ML), deep learning (DL) models have also been utilized for detecting early-stage PD with bio-indicators such as rapid eye movement, olfactory loss, cerebrospinal fluid data, and dopaminergic imaging markers [32]. Detection accuracy of \(96.45\%\) was achieved from 183 healthy individuals and 401 early PD patients. Furthermore, a sensor-based mobility analysis was conducted using wearable sensors to detect bradykinesia, one of the PD’s main symptoms, and a convolutional neural network to achieve high classification accuracy [33]. Some other studies utilize intertial wearable sensors and employ a combination of supervised and unsupervised learning for an in-depth analysis of mobility-based disorders and symptoms [34]. Without the use of ML and DL models in the studies mentioned in this section, it would not be possible to process the large volume of data, let alone understand the biomarkers that indicate the symptoms of Parkinson’s disease (PD). Therefore, machine learning and deep learning are essential in understanding and automating the diagnosis of neurological disorders, especially Parkinson’s disease (PD).

ML models can be applied to all kinds of neurological data acquired from MRI (magnetic resonance imaging), fMRI (functional MRI), BOLD (blood oxygen level-dependent) signals, PET (positron emission tomography), MEG (magnetoencephalography), and EEG (electroencephalography), which are physiological bio-signals that throw light upon reliable information about the onset of a neurological disorder/condition. EEG is a direct measure of electrical brain activity that captures the voltage fluctuations of the brain in a span of milliseconds [35, 36]. EEG’s main advantage over other methods is its high temporal resolution and precision, which can match cognition response. Though MEG has a few advantages over EEG due to its higher sensitivity to superficial cortical activity [37], capturing MEG signals requires expensive apparatus and a magnetically shielded room, making it less feasible than EEG. EEG analysis has been used as an approach to successfully detect many other neurological conditions such as Alzheimer’s [38, 39], epilepsy [40‐43], brain stroke [44], schizophrenia [45‐47], depression [48, 49], insomnia [50], autism [51], and attention-deficit hyperactivity disorder (ADHD) [52, 53]. From the previous studies, it is reasonably evident that EEG captures the nuances and salient features of the brain on a millisecond scale, which can lead to the detection of neurological disorders. Therefore, EEG has been chosen as the modality for this study that aims to find the salient biomarkers to accurately and efficiently detect Parkinson’s disease (PD) using the resting-state EEG signal.

Multiple approaches have been designed in the past for the analysis of motor and cognitive impairments in PD by utilizing the EEG signals of patients. The previous research focuses on the analysis of the deterioration of motor functions of the patient, including the study of gait signals [54‐56] and handwriting signals [57‐59]. Assessing gait characteristics such as speed, asymmetry, and variability during natural and demanding conditions such as twisting-turning, multi-tasking, and day-to-day life routines enhances the sensitivity of gait quantification and throws light upon exactly how gait gets impacted in PD and also how it progresses through various stages of PD [60]. Freezing of gait (FOG) is one of the most severe and disabling conditions of PD, wherein the patient is unable to move their feet, feeling a sensation as if their feet are frozen. Analysis of EEG signal has proved to be useful in identifying the onset of FOG, using wavelet decomposition [61‐63]. Therefore, by analyzing and detecting the gait signals (from EEG) or any such motor function impairment at its onset, the progression stage of PD can be detected early and diagnosed before it exacerbates.

Besides the deterioration of motor functions, PD also causes multiple cognitive impairments. These non-motor impairments can be analyzed very efficiently using EEG. In multiple studies, EEG was captured while performing tasks such as listening to music [64] or watching a movie [65], and the emotional response to these stimuli was analyzed through the EEG signal. Numerous clips of the movie expressing various emotions such as tragedy, happiness, romance, and many others were shown to the patient, and their response was observed. Some studies also used odors as stimuli to induce an emotional response to the smell [66‐68]. The previous studies designed on the same approach show that PD patients emote less spontaneously to video clips and bad odors [67]. Therefore, it can be inferred that there is an underlying link between Parkinson’s disease (PD) and the emotional response of the patient, which can be attributed to the deficit of dopamine in the brain, as dopamine influences the emotional response of the brain [69].

Degeneration of dopamine-producing neurons in the part of the brain called substantia nigra disrupts effective communication between the cortex and basal ganglia. These effects are captured by the EEG signal [70, 71], but an effective feature extraction method is often required to bring these effects to the foreground. EEG is characterized by nonlinear dynamics; hence, nonlinear methods such as fractal dimensions (FD) and entropies provide reliable and accurate information that is instrumental in distinguishing between a PD patient and healthy control (HC).

Fractal dimensions (FD) have been employed in some previous studies for the analysis of Parkinson’s disease (PD). Few studies, such as [72], utilized optical coherence tomography angiography to obtain images of superficial and deep capillary plexuses and retinal anatomy to investigate the capillary complexity of the retina. Fractal dimensional analysis of retinal capillary complexity was performed to differentiate between PD patients and healthy controls (HC). The experiment was conducted on 25 PD patients and 25 HC. It was observed that PD patients had a lower value of fractal dimensions (FD), indicating lower retinal capillary complexity. In another study [22], fractal dimensions of EEG were extracted to analyze the deterioration of emotional response in PD patients. Twenty PD patients consisting of 10 left side-affected and 10 right side-affected were recruited along with 10 HC for this experiment. Nonlinear EEG signals were analyzed for identifying emotions using fractal dimensions (FD), entropy, Hurst exponent, and multiple other feature extraction measures. Classification of PD was performed using fuzzy KNN and support vector machine (SVM). The results revealed that neuronal degeneration might contribute to the decline of emotional responsiveness in PD patients.

As investigated in the studies mentioned above, fractal dimensions have been explored as a good feature extraction measure for the analysis of PD using EEG. However, the previous studies are limited, and there is scope to conduct an exploratory analysis of the various fractal dimensional techniques. Therefore, it is essential to investigate: A) What is the optimal fractal dimensional measure to extract salient PD features from EEG? B) What ML model yields the best performance in detecting PD in both ON Meds and OFF Meds PD patients? and C) Can fractal dimensions accurately identify the parts of the brain severely impacted by PD?

The following research hypothesis was set by considering the EEG data to investigate the robustness of fractal dimensions (FD) as biomarkers for detecting Parkinson’s disease (PD).

IF EEG features are extracted using fractal dimensions (FD), which are generated by the sliding window technique that segments multichannel EEG signals along with overlapping and is given as input to the classifier, THEN, there exists an optimal FD and classifier combination that discriminates Parkinson’s disease (PD) patients who are ON and OFF medications from healthy controls with greater than 99% classification accuracy.

We hypothesize finding the optimal feature extraction and machine learning method to detect and analyze Parkinson’s disease (PD) in resting-state EEG. The study also aims to analyze the EEG signal to understand which brain parts are the most affected by Parkinson’s disease (PD), and if the significance of these regions can be highlighted using fractal dimensions (FD). For this purpose, different methods of segmentation, feature extraction, and machine learning models have been implemented and analyzed extensively, as shown in the previous sections. The results shed light upon the efficiency of fractal dimensions (FD) as an effective way of feature extraction from EEG signals. Higuchi FD outperforms other methods considerably across all employed machine learning models. Among all the methods implemented in this study, the most effective method for detecting Parkinson’s disease (PD) in resting-state EEG is the following:The accuracies of almost all feature extraction methods improved drastically with \(90\%\) overlapping window compared to \(50\%\) (Fig. 10). The increase in accuracy can be attributed to the number of training instances available for \(90\%\) overlap. However, the boost in performance when shifting from \(50\%\) overlapping window to \(90\%\) is less significant for Higuchi FD than for other methods. There is an increase in only \(2.63\%\) in Higuchi FD’s performance, while for other methods, it is around \(6.5\%\) (Fig. 10). Higuchi FD gives accuracies between \(90-95\%\) even in \(50\%\) overlapping window, while others give much lower accuracies. Thus, Higuchi FD can capture relevant information from the EEG signal even with relatively few training instances. Furthermore, Higuchi’s method creates multiple time series by repeatedly sub-sampling the signal in the window. The underlying procedure enables Higuchi FD to accurately extract relevant information even in the case of larger windows with less overlap. Petrosian FD and permutation entropy also perform well, especially when using a sliding window with \(90\%\) overlap. However, their performance decreases when using a sliding window with \(50\%\) overlap. On the other end of the spectrum, Katz FD and spectral entropy consistently perform poorly, with Katz FD giving the lowest accuracies by a considerable margin. Also, spectral entropy does not perform well and does not capture the complexities of the EEG signal in the temporal domain.

The topographic plots exemplify the workings of explainable AI, illuminating those brain regions that possess elevated feature importance compared to other areas, particularly in the context of Parkinson’s disease classification through the lens of the extreme gradient boosting (XGBoost) classifier. The topo plots in Fig. 9 show similar patterns to those in Fig. 5. Higuchi FD (Fig. 5: a1 and a2), Petrosian FD (Fig. 5: c1 and c2), and permutation entropy (Fig. 5: e1 and e2) show high feature importance in the midbrain and the occipital region of the brain, which are the regions adversely affected by PD. Prominent symptoms of PD include rhythmic shaking (tremors), loss of balance, and body stiffness. These motor functions of the body are controlled by the motor cortex, which lies in the midbrain region. It is the same region where the topo plots of Higuchi FD, Petrosian FD, and permutation entropy highlight as higher feature importance. Therefore, it can be inferred that features extracted from the channels of the EEG signal derived from the motor cortex were key contributors to the classification of PD, and these features were very accurately captured using fractal dimensions. This reinforces that Higuchi FD, Petrosian FD, and permutation entropy extract relevant features and accurately capture the complexities of the EEG signal that is instrumental in differentiating between a PD patient and a healthy person. These feature extraction methods can discern that the information captured from the electrodes around the motor cortex plays a more decisive role in the classification of PD than other parts of the brain, except the occipital cortex, which is also highlighted in the topo plots as a region of high feature importance. The high feature importance of occipital cortex in the classification of PD suggests that PD might be related to visual cortical activity, as is also investigated astutely in [96]. Studies like [97] show that PD also results in sensory disorders such as loss of sharpness, slower eye movement (Saccadic eye movement), reduced color perception, and reduced blink reflex and movement. The link between ocular impairments and Parkinson’s disease is significant and will impact the occipital cortex. This relation between PD and the occipital cortex has also been captured in the topo plots of this study. On the other hand, Katz FD and spectral entropy give topo plots very different from other methods. Both these methods give much lower accuracy in classification, which is clear from the topo plots since the brain areas with higher feature importance are not around the motor cortex or the occipital cortex. Therefore, Katz FD and spectral entropy cannot accurately capture the complexities of the EEG from the time domain.

This study delivers an architectural pipeline for detecting PD in resting-state EEG while emphasizing fractal dimensions (FD) as an effective way of extracting salient features from EEG signals. The final model built in this study is robust to the impact of medication on the symptoms of Parkinson’s disease (PD) and accurately identifies the signs of underlying disease in resting-state EEG with almost equal accuracy in patients OFF medication and ON medication, both. There is scope to strengthen this study further and understand the efficiency of fractal dimensional methods and their applications on time-series-based datasets. For example, varied-sized windows can be implemented to test the optimal length of the window for calculating fractal dimensions. Also, further research can be conducted to uncover the ocular symptoms of Parkinson’s disease (PD), and how they manifest in the EEG signal captured from the occipital cortex.

The limitation of the proposed model is the manual removal of artifacts from the EEG signal using independent component analysis. The whole pipeline can be effectively executed without human intervention if this artifact removal process can be automated using deep learning procedures [9]. While EEG offers a remarkable temporal resolution capable of capturing nuanced neuronal fluctuations, its spatial resolution is somewhat limited, making it challenging to pinpoint precise brain regions. This limitation presents obstacles in determining the specific areas of the brain associated with Parkinson’s disease through EEG. Nevertheless, our adoption of explainable AI, manifested in topographic brain plots, demonstrates the potential to extract pertinent spatial information from EEG data via machine learning. The topographic plots from our study emphasize the midbrain region’s pronounced feature importance in Parkinson’s disease detection, consistent with current research, underscoring its efficacy in accurately discerning spatial information.

Moreover, our study demonstrates that EEG signals, when processed using Higuchi fractal dimension and combined with machine learning models such as KNN, can effectively differentiate between healthy individuals and Parkinson’s disease patients. Moreover, our approach distinguishes between patients based on their medication status, an aspect that holds considerable clinical importance. Some of the potential clinical implications are as follows: -

Parkinson’s disease (PD) is a long-term progressive neurodegenerative disorder that causes motor and cognitive impairments. EEG analysis can provide reliable information about the condition of neurological impairments. Machine learning (ML) is immensely useful in building automated pipelines for diagnosing Parkinson’s disease (PD) using EEG signals. However, noise and artifacts from EEG signals can decrease the efficiency of such pipelines. Traditional feature extraction methods, such as partial directed coherence (PDC) works only in the EEG signal’s frequency domain. Recently, few studies have employed fractal dimension (FD) and entropy as feature extraction measures in the ML pipeline for detecting Parkinson’s disease (PD). However, to the best of the authors’ knowledge, no exploratory study exists on the various FD and entropy measures used in the feature extraction phase for neural signal analysis. We aimed to conduct a comparative analysis of the FD and entropy measures that, as feature extraction methods, can discriminate Parkinson’s disease (PD) patients who are ON and OFF medication from healthy controls (HC) using the ML pipeline. For this purpose, different segmentation, feature extraction, and machine learning models have been implemented and analyzed extensively to provide the best architectural pipeline. The results show that the larger overlap among the segmented windows that produced more input instances effectively improved the discriminating capacity. Higuchi FD outperforms other feature extraction methods across all employed machine learning models. K-nearest neighbors (KNN) classifier performs consistently well and gives the highest accuracies for both the PD cases discussed in this study, yielding a mean accuracy of \(99.65\pm 0.15\%\) for PD patients ON medication, and \(99.45\pm 0.18\%\) for PD patients OFF medication, respectively. The final model built in this study is robust to the impact of medication on the symptoms of Parkinson’s disease (PD) and accurately identifies the signs of the disease in resting-state EEG with almost equal accuracy in patients of both OFF and ON medication.

This study contributes to the body of knowledge with an architectural pipeline for detecting PD in resting-state EEG while emphasizing fractal dimensions (FD) as an effective way of extracting salient features from EEG signals. The limitation of the proposed model is the manual removal of artifacts from the EEG signal using independent component analysis (ICA). If the artifact removal process is automated using deep learning (DL) procedures as future work, the whole pipeline can be effectively executed without human intervention. Furthermore, the study was conducted on 31 subjects, and the data were collected using 32 EEG channels. Therefore, there is a need to further test this model on a larger dataset comprising more subjects and EEG channels to amply understand its applicability and usability in the real world.