Deep learning architectures for Parkinson's disease detection by using multi-modal features

https://doi.org/10.1016/j.compbiomed.2022.105610Get rights and content

Highlights

  • Feature-level framework and modal-level framework are presented for classifying the subjects into PD and healthy.

  • These frameworks are based on deep learning, and use neuroimaging and biological features.

  • The results prove that the approach based on multi-modal features is useful for classifying the subjects into PD and healthy.

  • A maximum accuracy of 93.33% and 92.38% is observed, for CNN, in the feature-level framework and modal-level framework, respectively.

Abstract

Background

The use of multi-modal features for improving the diagnosing accuracy of Parkinson's disease (PD) is still under consideration.

Method

Early diagnosis of PD is very crucial for better management and treatment planning of PD, as the delay in the diagnosis may even lead to death of the patient. Two frameworks, feature-level and modal-level, both of which are based on deep learning, are presented to classify the given subjects into PD and healthy by using neuroimaging (T1 weighted MRI scans and SPECT) and biological (CSF) features as the dataset. In the feature-level framework, all these features are integrated to form a heterogeneous dataset which is then supplied to two deep learning models to diagnose PD. In the modal-level framework, the number of features from T1 weighted MRI scans is reduced first, by using the filter feature selection method ReliefF. Those reduced number of features from the MRI scans are integrated with SPECT and CSF features to form another heterogeneous dataset, which is then fed to a deep learning model.

Results

Due to imbalanced nature of the dataset (consists of 73 PD and 59 healthy subjects), F1-score, geometric-mean, sensitivity, and specificity are measured, in addition to measuring the accuracy, to evaluate the performance of the developed models. A maximum accuracy of 93.33% and 92.38% is observed, for CNN, in the feature-level framework and modal-level framework, respectively.

Conclusions

Though the complexity of the approach based on multi-modal features is high as compared to an approach that uses only one type of feature, i.e., either neuroimaging or biological; the results prove that the approach based on multi-modal features is useful for classifying the given subjects into PD and healthy and can help the clinicians in accurately diagnosing the PD.

Introduction

Parkinson's disease (PD) is a neurological degenerative disorder that is affecting millions of people globally. It is very important to diagnose PD in its early stages as the delay in its diagnosis may lead to harmful results including death of the patient. Hence, early diagnosis of PD is important for better management and treatment planning of PD. A list of acronyms used in this research is provided in Table 1.

In recent years, researchers [1,2] have provided models for PD classification (i.e., for classifying the given subjects into PD and healthy) based on machine learning and statistical approaches and by integrating the features available from different modalities. The current research focuses on integrating the features from two different modalities (a) neuroimaging (T1 weighted MRI scans and SPECT) and (b) biological (CSF), and applying deep learning architectures on these integrated features. Since the features from these multiple modalities are combined or integrated together to form a heterogeneous dataset, the terms multi-modal and heterogeneous are used interchangeably in this research.

CAD systems, based on the deep learning architectures AE and CNN, are used not only for diagnosing PD [3,4] and tracking its progression but also for diagnosing other neurodegenerative diseases such as Alzheimer's disease [5,6] and Creutzfeldt-Jakob disease [7]. Morabito et al. [7] achieved an average accuracy of 89%, an average sensitivity of 92%, and an average specificity of 89% in differentiating early-stage Creutzfeldt-Jakob disease from rapidly progressive dementia by using deep learning concept. Zhang et al. [8] reviewed deep learning methods for two neurodegenerative disorders (Alzheimer's disease and PD), one neurodevelopmental (Autism spectrum disorder), and one psychiatric disorder (Schizophrenia) and discussed the limitations of the existing studies. Tufail et al. [9] applied 3D CNN for multiclass classification of Alzheimer's and PD by using PET and SPECT neuroimaging modalities.

Deep learning architectures have shown better results than traditional methods that are based on AI [10,11] because of their ability to learn the latent representation of data automatically. Also, deep learning-based architectures have the potential to model the non-linear relationships between data, thus making the deep learning-based architectures more appropriate for classification tasks. Inspired by earlier researches [3,[12], [13], [14], [15]], in the current research, we used deep learning architectures and developed a model that can perform PD classification more accurately.

For PD classification, in the current research, we presented a FL framework and a ML framework. The FL framework is based on the deep learning models SSAE (i.e., SAE1 as well as SAE1 + SAE2) and CNN, while the ML framework is based on CNN alone. An overview of these deep learning models is provided in Fig. 1.

Deep learning architectures provide more accurate results, for early detection of PD, than conventional AI [15], because deep learning architectures can extract features from the input set without a need for preprocessing the input. The terms deep architectures, deep learning architectures, deep learning models, and deep neural networks are used interchangeably in the current research. Deep architectures consist of simple structures that can do nonlinear operations and these simple structures are arranged to form a stack.

AE is an unsupervised learning-based ANN that uses backpropagation algorithm to provide the output value closest to the input value. That is, the goal of AE is to reduce the deformation between input and output data. The encoder and decoder are two parts of AE [5,16]. One of the variations of AE is SAE that comprises of a hidden layer connected to the input values by a weight matrix known as encoder. The hidden layer provides the output to a reconstruction vector by using a weight matrix to form the decoder [4].

Depending on the complexity of the problem to be solved, any number of SAEs can be layered or stacked, where the output of each hidden layer is connected to the input of the successive hidden layer to form a SSAE [4]. To enhance the accuracy of PD classification, the assumed deep learning architecture is optimized in a supervised manner by stacking another layer (as the last layer in SSAE) called softmax classifier (Fig. 2) that can predict the output labels efficiently. Softmax classifier is a generalization of logistic regression and this classifier provides a different probability for each of the labeled data or category, where the sum of all the probabilities is equal to one [4].

In the current research, both SAE and SSAE architectures are used for PD classification.

A CNN comprises three types of layers namely convolution layers, pooling layers, and fully-connected layers that can be stacked according to their functionality [17]. The convolution layer is connected to the input layer and determines the output from the neurons by convoluting the scalar product between the weights and the inputs, by applying different sized filters. The output of the convolution layer is then passed through the activation function followed by the pooling layer that further reduces the dimensions of the input data within that activation, thus reducing the computational complexity of the model. The last layer, i.e., the fully connected layer, produces the scores of each class from the activations that are to be used for classification. There are various hyper parameters that are employed by CNN. They are filter size, stride, and pooling type. More details about CNN can be found in Liu et al. [18]. In the current research, CNN is also used for PD classification.

The contributions of the current research include the following:

  • (1)

    Based on the authors' knowledge, this is the first model, for PD classification, that employs deep learning architectures on a heterogeneous dataset made up of neuroimaging (features from MRI scans; SBR values from SPECT) and biological (CSF) markers.

  • (2)

    Only those subjects for which the values of the above specified features (i.e., neuroimaging as well as CSF markers) are available are used in the current research. If any of those features' values are not available for a subject then that subject is not used in the current research.

The rest of the paper is outlined as follows: Section 2 describes an overview of the existing PD classification studies. Section 3 provides the details of the dataset used in the current research. Section 4 discusses FL and ML frameworks, for PD classification, by using heterogeneous datasets. Section 5 discusses the current research results and their comparison with existing research. Finally, Section 6 provides conclusion.

Section snippets

Related work

This section provides an overview of some recent studies that used deep learning architectures, for PD classification, by using neuroimaging or other biomarkers.

Dataset details

The values of (a) MRI scans, (b) SBR features of four striatal regions i.e., LC, LP, RC, RP, and (c) four CSF markers i.e., α-synuclein, Total-tau, P-tau181P, and Aβ42 for the 132 subjects (consist of 59 healthy and 73 PD) are downloaded from PPMI database (www.ppmi-info.org/data) in January 2021. In the current research, only those subjects for which the data from both the modalities i.e., neuroimaging (MRI scans; and SBR values from SPECT) and biological (CSF) markers is available are used.

FL and ML frameworks, for PD classification, by using heterogeneous dataset

In the current research, two frameworks, FL and ML, are presented for PD classification. These frameworks are based on deep learning architectures, as explained later on in 4.1 FL framework, for PD classification, by using heterogeneous dataset, 4.2 ML framework, for PD classification, by using heterogeneous dataset. These two frameworks are different from each other in terms of combining and using the features from two modalities i.e., from neuroimaging (MRI scans and SBR) and biological

Performance results and analysis

Since the dataset used for both FL and ML frameworks is imbalanced (i.e., the number of PD subjects (the majority class) is not same as the number of healthy subjects (the minority class) considered in this study), accuracy parameter alone is not enough to properly measure the overall performance [31] as the accuracy parameter tends to bias towards the majority class. To overcome this issue, for both FL and ML frameworks, other parameters, sensitivity, specificity, F1-Score, and geometric-mean [

Conclusion

In the current research, we presented FL and ML frameworks, both are based on deep learning architectures and heterogeneous dataset, for PD classification. ML is based on CNN alone, while FL is based on SSAE and CNN. The way the heterogeneous dataset is formed for FL is different from that for ML although both the heterogeneous datasets are made up of neuroimaging and biological markers.

The value of the accuracy parameter obtained in the FL framework (by using CNN) is better than the value of

Funding

The current research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.

Declaration of competing interest

The authors do not have any conflict of interest.

Acknowledgement

Data used in the preparation of this article were obtained from the Parkinson's Progression Markers Initiative (PPMI) database (http://www.ppmi-info.org/data). For up-to-date information on the study, visit http://www.ppmi-info.org. Data biospecimens used in the analyses presented in this article were obtained from the Parkinson's Progression Markers Initiative (PPMI) (http://www.ppmi-info.org/specimens). As such, the investigators within PPMI contributed to the design and implementation of

References (31)

  • V.J. Kadam et al.

    Feature ensemble learning based on sparse autoencoders for diagnosis of Parkinson's disease

  • H.I. Suk et al.

    Deep learning-based feature representation for AD/MCI classification

  • A. Ortiz et al.

    Ensembles of deep learning architectures for the early diagnosis of the Alzheimer's disease

    Int. J. Neural Syst.

    (2016)
  • F.C. Morabito et al.

    Deep learning representation from electroencephalography of early-stage Creutzfeldt-Jakob disease and features for differentiation from rapidly progressive dementia

    Int. J. Neural Syst.

    (2017)
  • L. Zhang et al.

    A survey on deep learning for neuroimaging-based brain disorder analysis

    Front. Neurosci.

    (2020)
  • Cited by (0)

    View full text