Parkinson’s disease (PD) is one of the most common clinically progressive neurodegenerative diseases worldwide, with prevalence second only to Alzheimer’s disease, and it affects more than 10 million people worldwide (Kim et al., 2017; Srivastav et al., 2017). Early diagnosis and treatment of PD are crucial to stop its progression in the initial stages (Chen et al., 2014; Adeli et al., 2016; Heim et al., 2017). In the early stage of PD, the main manifestations are non-motor symptoms, which are nonspecific and difficult to diagnose (Peng et al., 2017; Cigdem et al., 2018; Tuovinen et al., 2018; Rubbert et al., 2019). However, advancements in neuroimaging and machine learning technologies have led to an increasing role of such technologies in the accurate diagnosis of PD (Chen et al., 2014; Szewczyk-Krolikowski et al., 2014; Peng et al., 2017; Amoroso et al., 2018).

Resting-state functional magnetic resonance (rs-fMRI; Qin et al., 2019; Zhou Z. W. et al., 2019) is one of the most commonly used techniques for neuroimaging. The amplitude of low-frequency fluctuation (ALFF; Zhang et al., 2017; Xu et al., 2019), which can detect the amplitude of spontaneous brain fluctuations, is one of the most commonly used fMRI measurements.

Radiomics has been widely used in clinical oncology studies (Ji et al., 2020; Zhao K. et al., 2020); the medical images can be converted into feature sets that can be used to characterize tumor characteristics by a series of algorithms (Lambin et al., 2012; Ji et al., 2020). Radiomics has now been widely used in the study of neuropsychological diseases, their diagnosis, and their neurological mechanism (Sun et al., 2018; Huang K. et al., 2019; Mo et al., 2019; Wang et al., 2020). Histogram analysis is the most commonly used radiomic feature extraction method, which is widely used in neuroimaging research (Cui et al., 2016; Sun et al., 2018; Huang K. et al., 2019; Zhou et al., 2020). To our knowledge, there is no existing study that has used histogram analysis to diagnose PD.

Therefore, this study aimed to explore the value of using ALFF-based histogram analysis in the diagnosis of PD and to investigate the regions of the most important discriminative features and their contribution to classification discrimination in order to explore its potential pathological mechanism.

In this study, we used the ALFF measurement of rs-fMRI data to perform feature selection and data dimensionality reduction by two-way independent samples t-test and LASSO logistic regression based on 10-fold cross-validation. We calculated the participants’ Rag-score and performed ROC analysis based on the selected features. The classification model had a high diagnostic efficiency, with an AUC of 0.966, a sensitivity of 92.9%, a specificity of 97.6%, and an accuracy of 95.2%. These results demonstrate that histogram analysis based on rs-fMRI can be used for the diagnosis of PD.

Radiomics was first proposed by Lambin et al. (2012). It can extract high-throughput quantitative features of medical images and can provide greater features than conventional image analysis methods, thus improving disease diagnosis and prognostic evaluation (Nie et al., 2019; Tang et al., 2019; Ji et al., 2020). It was first used for tumor heterogeneity evaluation, tumor diagnosis and differential diagnosis (Nie et al., 2019), prognostic evaluation (Zhao L. et al., 2020), and tumor recurrence (Tang et al., 2019; Ji et al., 2020), among other uses. Recent studies have also indicated that it has a great application value in neuropsychological diseases (Tang et al., 2017; Péran et al., 2018; Sun et al., 2018; Huang K. et al., 2019). Different from the previous radiomic studies in the field of oncology, they often have only one ROI. However, in our study, there are 116 brain ROIs. We therefore only extracted the first-order intensity statistical features—histogram features; otherwise, our model will have too many features. It is easy for a model to fall into a “curse of dimension” and model overfitting.

Rs-fMRI is widely used in the study of PD (Gu et al., 2016; Amoroso et al., 2018; Hohenfeld et al., 2018; Rispoli et al., 2018), and ALFF is one of its commonly used metrics (Hu et al., 2015; Gu et al., 2016). In this study, the ROI-based method (Tang et al., 2017; Sun et al., 2018) was used to extract the histogram features of the ALFF images in specific brain areas using the AAL 116 atlas, and the machine learning method was used for data dimensionality reduction and modeling to explore the diagnostic value of the rs-fMRI-based machine learning in PD.

In this study, we standardized the data before data analysis. This procedure is widely used in the preprocessing of machine learning, which can effectively reduce the influence of the different units between features or reduce the influence of signal changes in a larger image signal range, which is beneficial to the improvement of model performance (Pereira et al., 2009; Tang et al., 2019). Among them, the Z-standardization (Tang et al., 2019; Wang et al., 2020) used in this research is one of the most commonly used methods.

Most previous studies have extracted features based on the brain atlas, which only extracted the mean values of the metrics (such as ALFF, fractional anisotropy, mean diffusivity, regional homogeneity, functional connectivity, voxel-mirrored homotopic connectivity, etc.) in the ROI defined by the brain atlas (Dai et al., 2012; Cui et al., 2016; Ding et al., 2017; Tang et al., 2017; Sun et al., 2018; Zhou et al., 2020). In our study, we extracted not only the mean ALFF values in the predefined ROIs but also other histogram features, including the minimum, maximum, range, standard deviation, variance, median, skewness, kurtosis, 10th percentile, and the 90th percentile, which could more comprehensively reflect ALFF information in the predefined ROIs. Our method achieved great classification performance on both the primary set and the validation set. We found that these features could be used to distinguish PD patients and HCs, which indicated that these features have physiological significance. Tang et al. (2017) combined the mean diffusion tensor imaging metrics to detect HIV patients. The accuracy and AUC of the model were 83.08% and 0.911, respectively. A machine learning study using ALFF metrics to detect PD used the AAL 116 atlas to extract the mean ALFF features. The accuracy and AUC were 80.75% and 0.8109, respectively. Another study used the Harvard–Oxford atlas to extract the mean ReHo, ALFF, VMHC, gray matter volume, and FC features. The authors found that both random forest (accuracy = 0.8261, AUC = 0.9015) and support vector machine (accuracy = 0.8483, AUC = 0.9697) achieved the perfect accuracy and AUC for distinguishing between PD and HC subjects (Cao et al., 2020). Our results are better than those of these studies, which may indicate that our histogram analysis method can extract more information in the ROIs (Lambin et al., 2012; Gillies et al., 2016).

Radiomics can extract a large number of quantitative features; however, its data dimensionality is too high when compared with the sample size of most studies, making it is easy to fall into a “curse of dimensionality,” thus causing the model to overfit. Hence, the features must be selected for dimensionality reduction to obtain the most valuable features in order to improve the reliability and accuracy of the model (Gu et al., 2016; Péran et al., 2018; Rubbert et al., 2019; Wang et al., 2020). In our study, we performed a two-way independent samples t-test on the standardized data to select the features that were significantly different between groups for subsequent analysis. Then, we applied previous research data processing methods (Mo et al., 2019; Tang et al., 2019; Wang et al., 2020) and used correlation analysis to remove the features with high correlation coefficients in order to reduce the multicollinearity of variables. We chose a correlation coefficient threshold of 0.9 to remove highly correlated variables (Tang et al., 2019). The LASSO logistic regression is very suitable for high-dimensional data processing. It can select the most predictive radiomic features and compress the non-important feature coefficients to zero in order to achieve the purpose of data dimensionality reduction and feature selection (Huang K. et al., 2019; Wang et al., 2020). The linear combination of the selected features weighted by their respective coefficients was used to calculate the Rag-score of each subject, and the resulting Rag-score was analyzed by ROC analysis to evaluate the diagnostic efficacy of the model. In this study, LASSO based on 10-fold cross-validation was used for further dimensionality reduction and feature selection of the data; at the same time, the Rag-score of each subject was calculated and ROC analysis was performed to evaluate the performance of the model.

Previous studies (Varoquaux, 2018; Gorriz et al., 2021) have demonstrated that cross-validation in a small sample size leads to large error bars. Varoquaux (2018) also pointed out that the best way to solving the problem was to test the model performance across several datasets. Therefore, we divided the dataset into a primary set and an independent validation set, and we tested the model performance using the independent validation set. In our study, we found that our models achieved high classification performances with different cross-validation methods both in the primary and validation sets. However, the error bars were relatively large, especially in the validation set. Our findings are consistent with those of previous studies (Varoquaux, 2018; Gorriz et al., 2021). In spite of the arguable power of the model, cross-validation is the best tool available because it is the only non-parametric method to test for model generalization and it can measure the ability of the model to predict new data (Varoquaux, 2018). Currently, cross-validation is still the most popular method in use.

In this study, the discriminative brain areas that can be used for PD diagnosis are located in the frontal, temporal, parietal, occipital, and limbic lobes, as well as in the cerebellum and the thalamus, which are consistent with the results of previous PD studies (Szewczyk-Krolikowski et al., 2014; O’Callaghan et al., 2016; Chen et al., 2017; Tuovinen et al., 2018). Many studies have indicated that damage to the corticothalamus–striatum–cortical pathway and the reciprocal striatum–cerebellar loop is the basis of many clinical symptoms of PD, and the thalamus plays an important role in function regulation and conduction (Szewczyk-Krolikowski et al., 2014; O’Callaghan et al., 2016; Tuovinen et al., 2018). In this study, we found that the radiomic features of the bilateral thalamus are helpful in the diagnosis of PD. A previous study (O’Callaghan et al., 2016) showed that multiple brain areas of the cerebellum of PD patients were atrophied, including the bilateral lobules I–IV, VI, VII (crus I and crus II), VIIb, VIIIa, VIIIb, right V, and the cerebellar vermis. The study also indicated that the modulatory relationship of the subthalamic nucleus on intracerebellar connectivity was lost in PD patients and that there were extensive cerebellar–cortical network abnormalities. Tuovinen et al. (2018) found that the connections within the cerebellum and between the cerebellum and the sensorimotor network in PD patients increased, the connections between the cerebellum and the caudate nucleus, thalamus, and amygdala increased, and the connection between the supplementary motor area and the cingulate gyrus decreased. Chen et al. (2017) found that the bilateral cerebellar gray matter volume was reduced and that the functional connections among the bilateral cerebellum, angular gyrus, hippocampus, middle occipital gyrus, and posterior cingulate gyrus were abnormal. Hu et al. (2015) also showed that PD patients had cerebellum ALFF abnormalities and abnormal functional connections between the cerebellum and the left middle cingulate gyrus. In our study, we also found extensive brain dysfunction in the cerebellum and multiple gray matter areas, which is consistent with the findings of a previous study. Besides, previous studies (Szewczyk-Krolikowski et al., 2014; Hu et al., 2015; Lucas-Jiménez et al., 2015; Hou et al., 2016; Li et al., 2016; Huang L. C. et al., 2019) also found structural and functional abnormalities in the insula, lingual gyrus, paracentral lobule, right inferior temporal gyrus, parahippocampal gyrus, precuneus, middle frontal cortex, dorsolateral prefrontal lobe, and other brain areas, which were similar to the results of the present study. This study showed that radiomics based on rs-fMRI can effectively identify abnormal brain activity areas in PD patients and provide support for the interpretation of PD neural mechanisms.

This study has some limitations. Firstly, as obtaining a large dataset is difficult in neuroimaging, where data acquisition is costly, the sample size of this study is relatively limited, but compared with the sample sizes of previous intergroup comparison studies (Lei et al., 2017; Zhou M. et al., 2019) and some machine learning articles (Hou et al., 2016; Tang et al., 2017), the sample size of this study is relatively large. Previous studies (Sun et al., 2018; Huang K. et al., 2019) have shown that multimodal imaging and clinical indicators can improve the performance of machine learning models. This study only selected the ALFF of rs-fMRI as input variables. Subsequent work can incorporate more rs-fMRI metrics, different functional imaging sequences, and clinical measurements to explore its effect on the classification model. In this study, both the primary and validation sets confirmed that the model has good classification accuracy, but no external validation was performed to further verify the generalization ability of the model. Other machine learning models or methods, such as support vector machine (Zhou et al., 2020), convolution neural network (Zhang J. et al., 2019), deep neural network (Liu et al., 2020), and transfer learning (Wang S. et al., 2019), have also been applied in previous neuroimaging studies, and they can obtain great classification performance. A previous study (Wang S. H. et al., 2019) used AlexNet as the basic transfer learning model to identify alcoholism. The method yielded a sensitivity of 97.44 ± 1.15%, a specificity of 97.41 ± 1.51%, and an accuracy of 97.42 ± 0.95% on the test set. In another study (Wang S. H. et al., 2019) employing densely connected neural network as the basic algorithm for transfer learning to detect cerebral micro-bleedings, the model achieved 97.71% classification accuracy. In the future, we will try other machine learning methods to detect PD.

In conclusion, this study used rs-fMRI data to extract the histogram features of brain regions based on brain atlases and used machine learning analysis to build a PD classification model. The study showed that this method can classify PD patients and HCs well and effectively identify PD patients’ regions of abnormal brain function; furthermore, it can assist in the early diagnosis of PD and provide a means for PD mechanism research and clinical therapeutic efficacy evaluation.

The original contributions presented in the study are included in the article, further inquiries can be directed to the corresponding author.

The studies involving human participants were reviewed and approved by Medical Research Ethical Committee of Nanjing Brain Hospital. The ethics committee waived the requirement of written informed consent for participation.

DS conducted the experiment, performed the data processing and analysis, and wrote the manuscript. HZ, GW, and SW collected the data and performed the data processing and analysis. KR supervised the whole study, including experiments and manuscript writing. All authors contributed to the article and approved the submitted version.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.