Firstly, the study began with selecting dysregulated miRNAs significantly related to the development and pathogenesis of PD. Only circulating miRNAs (CSF, serum, plasma, peripheral blood mononuclear cells (PBMCs), and saliva) are of PD patients were extracted from a study done by Nies and coworkers [16]. Along with this list of miRNAs that are dysregulated in PD, a set of controls was collected randomly from the miRPathDB database. The inclusion criteria for the PD set involved selecting miRNAs that have been identified as significantly dysregulated in peer-reviewed studies focused on PD and are implicated in pathways known to be involved in PD pathogenesis. The exclusion criteria for the PD set ruled out miRNAs with conflicting evidence across multiple studies and those that are also significantly dysregulated in other neurodegenerative diseases. For the control set (not related to PD), the inclusion criteria consisted of miRNAs that have not been implicated in any neurodegenerative diseases and are considered to be stably expressed across multiple tissue types. The exclusion criteria for the control set eliminated miRNAs with known roles in other neurological disorders and those showing significant variability in expression across different tissue types. All included miRNAs are presented in Table 1.

The choice to exclusively use miRPathDB 2.0 as our sole database for control miRNAs is due to its focus on high-quality, experimentally verified data. This aligned with our goal of ensuring data reliability. Primarily, miRPathDB is acknowledged for its robust data quality, as it curates miRNA targets from a variety of sources including peer-reviewed publications, thus ensuring a high level of data reliability. For our analysis, we included all predicted target genes for our selected miRNAs, using both intersection and union prediction methods.

Gene-target data was collected for each miRNA associated with miRPathDB [17], as well as the miRNAs in the control group. miRPathDB is a comprehensive database that consolidates miRNA-target interactions and pathway annotations from multiple pathway databases, providing valuable insights into miRNA regulation and signaling pathways. It utilizes various sources such as TargetScan, miRTarBase, and DIANA-TarBase to compile miRNA–target interactions.

To systematize each miRNA’s unique features, we implemented an algorithm that employed a singular gene list (derived from the aggregate of each miRNA’s target genes), assigning a binary value (0 or 1) to each miRNA for each gene and pathway. Consequently, we built a dataset column-by-column, each representing a gene target attribute (e.g., TP53), and rows representing individual miRNAs with their binary descriptors. A “1” signifies the miRNA holding the gene target or pathway in the column, whereas a “0” denotes the absence of that specific genomic target or pathway.

While we acknowledge the binary coding system simplifies the intricacies of miRNA–target links, such as site binding or conservation factors, it serves the purpose of facilitating a general first-level analysis. Future work could delve into these finer details for a more nuanced understanding.

After the preparation, the dataset included 16,299 attributes (descriptors) for 56 associated miRNA and 56 control miRNA (Fig. 2). The process of attribute selection involved utilizing WEKA’s InfoGainAttributeEval module that ranks descriptors according to their capacity to differentiate between classes in a classification problem. By calculating the mutual information between each feature and the class variable, this algorithm quantifies the extent to which the feature provides information about the class [18].

After employing WEKA’s InfoGainAttributeEval module, the number of descriptors was substantially reduced. We validated the efficacy of this reduced attribute set through additional analyses. Specifically, the reduced set showed higher classification accuracy of 93.3% on the independent testing set after reducing the number of descriptors, confirming that it retains sufficient statistical power for reliable classification.

On the miRNA training and testing datasets, ML analysis was performed using the WEKA program environment. This open-source workbench includes several tools for data cleaning and filtering, classification and pattern recognition [18]. In addition to the ML analysis performed using WEKA, the TensorFlow Python library Keras was used to generate a neural network for making predictions from the data. Keras is a high-level neural networks application programming interface (API) that provides an interface to build, train, and evaluate neural networks. It is built on top of lower-level libraries such as TensorFlow (v. 2.10, Google LLC, Mountain View, CA, USA), which handle the computations underlying the neural network [19]. The neural network utilizes a Sequential model with 5 layers. The first layer is a 32-neuron Dense layer with Rectified Linear Unit (ReLU) activation, which use a threshold function to introduce non-linearity. The second layer is also a 32-neuron Dense layer using ReLU activation but includes Dropout, which randomly deactivates 20% of neurons during training to prevent overfitting. This is followed by two hidden Dense layers, also using ReLU activation, of 128 and 64 neurons, respectively, allowing for complex pattern recognition. The final layer is a single-neuron Dense output layer with sigmoid activation. Sigmoid activation transforms the output into a probability value between 0 and 1, representing the likelihood of the binary classification. The architecture is illustrated in Fig. 3.

The objective of our model was to identify patterns of prospective miRNA blood-based biomarkers specific to PD for diagnostic applications. Each discrete dataset underwent individual analysis through various ML classification algorithms. We subsequently evaluated these models’ validity by applying them to test data, ensuring no overlap with the respective training datasets. At the end of this stage the ML models were developed and cross-validated.

In a second—validation phase, we introduced the model to an independent and distinct set of miRNAs significantly associated with PD, as well as another set tied to a different pathology, specifically breast cancer. The model’s effectiveness was gauged by its accuracy, which is defined as the proportion of correctly classified instances to the total instances.

The cross-validation test sets yielded above 90% accuracy for the Hoeffding Tree algorithm, the Naïve Bayes, Multilayer Perceptron, and the Sequential Model that was trained using Keras. The training sets for the model contained 112 miRNAs (with 56 miRNAs being PD biomarkers and 56 not), which was then filtered from the dataset reducing the number of attributes from 16,299 to 61. The main parameters of models prepared with different algorithms are presented in Table 2. The highest accuracy model was the Sequential Model, which was then used to make predictions on independent data (Fig. 4).

The Hoeffding Tree model outperformed Random Forest in our analysis, a result that may seem counterintuitive given Random Forest’s ensemble approach. This outcome is likely due to unique dataset characteristics that favor the Hoeffding Tree’s online learning method. It also suggests that our Random Forest model may have been overfitting as its the parameters was not optimized.

The Sequential Model that was created using Keras was able to reach accuracies of 95.65% (Fig. 5). To further eliminate potential bias and prevent overfitting, we created an independent test set comprising miRNA dysregulated in PD (Table 3). The independent test set used for the additional was extracted from a study done by Ming-Che Kuo and coauthors [20]. We selected eight miRNAs from this source that have been significantly associated with PD. There was no overlap between these miRNAs and those utilized in the training set. This independent set allowed us to test the model’s capacity to accurately classify new, unseen data. Upon analysis of this independent test set, the Sequential Model continued to show strong performance, achieving an accuracy of 93.3% (Fig. 6).

Following the validation of our independent set, we initiated additional testing to ascertain that our models are selective for PD. We aimed to confirm that the models could specifically detect PD and not misclassify other conditions as PD. To verify this specificity, we challenged our classifier with miRNA datasets of a different disease—Breast Cancer (BC, Table 4). The BC miRNA dataset was drawn from the comprehensive research conducted by van Schooneveld and coworkers [21]. Our classifier model, trained on PD-specific miRNA data, was then run against the BC miRNA dataset. The expectation was a low number of false positives, indicating the model’s capability to discern between the miRNA profiles of different diseases.

The results were in line with our expectations. While the model achieved an accuracy of 95.65% on the PD-specific miRNA test set, its accuracy was significantly lower at 40% on the BC miRNA test set (Fig. 7). Although these findings suggest a degree of specificity for PD, further studies with multiple control groups and additional statistical tests are required for robust confirmation.

The Area Under the Receiver Operating Characteristic Curve (AUROC) values for each model are as follows: Random Forest scored 0.916, Hoeffding Tree scored 0.966, Naïve Bayes scored 0.9692, Multilayer Perceptron (MLP) scored 0.9676, and the Sequential model scored a 0.992. These values indicate high classifying power for all models, with Naïve Bayes, MLP, and the Sequential models showing the highest performance (Fig. 8).

From the confusion matrices of each model, the Matthews Correlation Coefficient (MCC) was computed. The MCC values, calculated using the formula from Fig. 8, for the Random Forest, Hoeffding Tree, Naïve Bayes, and MLP models, and the Sequential model were 0.754, 0.822, 0.840, and 0.805, 0.914, respectively. These values indicate that the Sequential model achieved the best balance between sensitivity and specificity among the tested models.

For the Random Forest confusion matrix as shown in Fig. 9A, the top left cell, representing True Positives, contains 52 instances, indicating that the model correctly identified these cases as positive. The bottom right cell, representing True Negatives, contains 46 instances, showing that these cases were correctly identified as negative. However, the model was not perfect. The top right cell contains 4 instances, representing False Positives, where the model incorrectly classified negative cases as positive. Similarly, the bottom left cell contains 10 instances, representing False Negatives, where the model incorrectly classified positive cases as negative.

The confusion matrix for the Hoeffding Tree model (Fig. 9B) shows a slightly different accuracy. The top left cell, representing True Positives, contains 52 instances, showing that these cases were correctly identified as positive. The bottom right cell, representing True Negatives, contains 50 instances, indicating that these cases were correctly identified as negative. However, the model did make some errors. The top right cell contains 4 instances, representing False Positives, where the model incorrectly classified negative cases as positive. The bottom left cell contains 6 instances, representing False Negatives, where the model incorrectly classified positive cases as negative.

The confusion matrix for the Naïve Bayes (Fig. 9C) model presents yet another pattern. The top left cell, representing True Positives, contains 53 instances, indicating that these cases were correctly identified as positive. The bottom right cell, representing True Negatives, contains 50 instances, showing that these cases were correctly identified as negative. However, the model was not without errors. The top right cell contains 3 instances, representing False Positives, where the model incorrectly classified negative cases as positive. The bottom left cell contains 6 instances, representing False Negatives, where the model incorrectly classified positive cases as negative.

The confusion matrix for the MLP (Fig. 9D) model shows another pattern. The top left cell, representing True Positives, contains 52 instances, indicating that these cases were correctly identified as positive. The bottom right cell, representing True Negatives, contains 49 instances, showing that these cases were correctly identified as negative. However, the model did make some mistakes. The top right cell contains 4 instances, representing False Positives, where the model incorrectly classified negative cases as positive. The bottom left cell contains 7 instances, representing False Negatives, where the model incorrectly classified positive cases as negative.