The role of MER processing pipelines for STN functional identification during DBS surgery: a feature based machine learning approach.

The role of MER processing pipelines for STN functional identification during DBS surgery: a feature based machine learning approach. | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Method Article The role of MER processing pipelines for STN functional identification during DBS surgery: a feature based machine learning approach. Vincenzo Levi, Stefania Coelli, Chiara Gorlini, Federica Forzanini, and 5 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6915773/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Background: Micro-electrode recording (MER) is one of the modalities used to confirm pre-operative planning during Deep Brain Stimulation (DBS) surgery of the subthalamic nucleus (STN) for the symptomatic treatment of Parkinson’s Disease. MER signals have been widely used in combination with machine learning (ML) techniques to improve STN functional localization. However, the impact of data processing and preparation has mostly been overlooked. Methods: A total of twenty-four combinations of processing approaches have been implemented with the aim of exploring the impact of data processing pipelines on the performance of feature-based ML classifiers. These comprise four signal artefact treatments, three outlier management procedures, and an option to standardize or not the feature sets. The effects of the implemented pipeline on the classification results were evaluated by training and testing three classifiers, both with and without feature selection. A final fundamental step to explore the feature importance using SHAP approach has also been implemented. Results: Improvements in performance metrics have been noticed after implementing approaches to artefact rejection and optimal outlier management, while the preliminary features standardization based on single patient and brain hemisphere data reduce all the performance metrics (accuracy, F1-score, recall, precision and area under the curve (AUC)). Interestingly, feature importance analysis through SHAP approach highlighted a good agreement between features contributing to classification across most of the implemented pipelines. Conclusions: Proper identification and rejection of artefacts combined with appropriate outlier management are crucial steps during MER processing pipelines for STN identification, while pre-normalization of features based on data from single patient and brain hemisphere may lead to overall performance degradation. In addition, the SHAP approach may represent an adjunctive useful tool to guide and improve the implementation of future algorithms. Biotechnology and Bioengineering Deep Brain Stimulation (DBS) Microelectrode recording (MER) Subthalamic Nucleus (STN) machine learning feature-based Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 1. Introduction Deep brain stimulation (DBS) is a safe and effective surgical treatment used to control the symptoms of Parkinson’s disease[ 1 ]. Through DBS surgery a permanent electrode is implanted inside the brain to chronically deliver high-frequency electrical pulses to the subthalamic nucleus (STN). Although STN is easily visible on preoperative Magnetic Resonance Imaging (MRI), sources of error arising from brain shift and small uncontrollable deviations from the pre-planned electrode trajectory often require intra-operative real-time data to verify the correct placement of the electrode in the planned target [ 2 ]. One of the modalities used to confirm pre-operative planning is micro-electrode recording (MER) [ 3 ]. MER explores electrophysiological properties of the brain tissue surrounding the STN and within the STN itself. In a typical DBS surgery, the microelectrodes record electrophysiological activity along a planned track as they are sequentially advanced into the brain by the neurosurgeon. Since each part of the brain has its own characteristic neural activity (such as spike firing counts and patterns), that of the STN can be recognized over the background noise level [ 4 ]. As a result, based on monitoring of this electrophysiological activity, the neurosurgeon decides when the microelectrode has entered the STN. However, MER intra-operative interpretation may be challenging and time-consuming requiring a high level of proficiency and expertise. Recently, several machine learning algorithms have been proposed as decision-making support tools with the aim of minimizing subjectivity and improving patient outcomes [ 5 , 6 ]. These algorithms are based on a variety of different machine learning paradigms. Most methods use traditional approaches which classify signals based on specific features that are known to be of interest, such as the power in particular frequency bands or various spike-dependent or spike-independent features [ 7 , 8 ]. Others exploit a more modern, deep learning approach in which the entire signal is provided to the algorithm rather than reducing it to a smaller vector of features [ 9 – 11 ]. One common limitation of these approaches does not lie specifically in the paradigm or machine learning architecture chosen, but in methodological design and dataset processing [ 12 , 13 ]. Few studies indeed report in detail how the signals have been pre-processed and how the dataset has been managed [ 14 ]. Therefore, the goal of this study was to assess the impact of different MER data processing and management pipelines on the classification of the STN using feature-based ML. A total of twenty-four combinations of processing approaches have been implemented, comprising four signal artefact treatments and three outlier management procedures, as well as the option to standardize or not the feature sets. Three classifiers were then trained and tested, both with and without feature selection, to evaluate the effects of the implemented pipeline on the classification results. The paper is organized as follows. In section 2, first the analyzed dataset is described, then in section 2.2 the implemented pipelines and the evaluation framework are explained in detail. Specifically in 2.2.1 the artefact rejection approaches are described, in 2.2.2 all the extracted features are listed, in 2.2.3 and 2.2.4 the outlier management and the feature standardization methods are respectively presented. Section 2.3 provides a description of the implemented classifiers and of performance evaluation. Results are reported in section 3 and discussed in section 4, also addressing the study limitations. Finally, the study is summarized in section 5. 2. Materials and Methods 2.1. Dataset Retrospective MER data from 28 Parkinson’s disease patients undergoing bilateral DBS surgical procedures targeting the STN from 2020 to 2023 at the Neurosurgery Department of the Fondazione IRCCS Istituto Neurologico Carlo Besta (Milan, Italy) were reviewed. Patients signed an informed consent regarding the surgery procedure and data exploitation for research aims. In awake patients, MERs were collected employing a Medtronic LeadpointTM system (Medtronic Inc. Minneapolis, MN, USA). Stereotactic planning was performed using the StealthStation Surgical Navigation System (Medtronic, Minneapolis, MN, USA). Direct targeting on T2 pre-operative MRI was always performed for stereotactic targeting of the STN. All patients had a preoperative stereotactic imaging with MRI or CT and immediate intra-operative verification of electrode positioning with intra-operative CT scan (O-Arm, Medtronic). Surgical procedures were performed using the Vantage frame (Elekta, Stockholm, Sweden). Glass-coated platinum/iridium micro-electrodes with impedance of 0.4–1.0 mΩ (FHC Inc., Bowdoinham ME, USA) were used. Each hemisphere was explored using three simultaneous parallel microelectrodes (anterior, central, posterior) spaced 2 mm apart one from another. Recordings lasted at least 10 s for each single depth. MER recordings started at an estimated distance from the target (EDT) of 10 mm and extended to 5 mm EDT after the planned target, advancing with 1 mm steps. Traces were high pass filtered at 200 Hz to allow the visualization of the firing activity during surgery and digitalized with a sampling rate of 24 kHz. At each EDT, traces were recorded for a duration of at least 10 seconds. Along the electrode path, various functional regions were identified, including white matter, thalamus (TH), zona incerta (ZI), substantia nigra (SNr), and STN. For each patient’s surgical procedure, the intraoperative annotations were reviewed by an expert neurosurgeon and the MER traces assigned to two classes: ‘STN’ if recognized as within the STN, and ‘NOT STN’ if belonging to other structures. 2.2. Dataset Preparation Pipelines In this section, the different steps adopted to compose the pipelines tested in the current work are described. The full framework is displayed in Fig. 1 . 2.2.1. Data preparation and Pre- processing Data preparation steps were performed offline in MATLAB (version R2023b) environment. First, we decided to consider further MER signals recorded within an EDT range [-5 mm: +2 mm] to improve dataset balance between classes (i.e., within and outside the STN). The selected raw signal underwent three artifact detection procedures: i) visual inspection by and expert (EXP); ii) an unsupervised variance-based algorithm (COV), and iii) an algorithm based on thresholds estimated on the signal noise characteristics (BCK). Additionally, we considered the RAW dataset, unprocessed for artefacts, for comparative purposes. The EXP approach consisted of an expert (V. L.) visually screening all the traces to identify evident artifacts (e.g., electromagnetic interference, mechanical electrode shift). A trace marked with artifacts was fully rejected. The remaining traces comprise the dataset ‘EXP’. The second artifact detection approach was based on the algorithm presented in [ 13 , 14 ] for the identification of stationary signal segments. This was preferred over other approaches to exploit its unsupervised nature. The algorithm implementation steps were the following: Segmentation of the signal \(\:x\left(n\right)\) into m 0.5-second segments \(\:\{\:{x}_{k};\:k\:=\:1:m\};\) Compute the autocorrelation of each segment \(\:\:\{R\left({x}_{k}\right);\:k\:=\:1:m\}\) Compute the variance of the transformed segment \(\:\{{v}_{k}=var(R\left({x}_{k}\right);\:k\:=\:1:m\}\) ; Comparison of the variances of neighboring segments by computing their distance as \(\:{d}_{kl}=\frac{\text{m}\text{a}\text{x}({v}_{k},{v}_{l})}{\text{m}\text{i}\text{n}({v}_{k},{v}_{l})},\:with\:k=\:1:m-1;l=k+1\) ; Creation of a distance matrix \(\:D\) of all possible distances between segment pairs. The matrix elements \(\:{d}_{kl}\) exceeding an experimentally identified threshold (Th = 1.8) are replaced with ones, other with zeros and an adjacency matrix \(\:A\) is obtained. The resulting matrix is scanned for the longest uninterrupted segment (sequence of zeros) using a greedy algorithm. The algorithm identifies segments that are different from the rest of the signals, marks them as artifacts, and removes them from the analyzed signal. The last approach identifies artifacts signal segments based on specific amplitude and frequency criteria. Therefore, two assessment steps are performed: Amplitude check: By exploiting the definition of background noise proposed by [ 15 , 16 ], the amplitude back-ground noise level is computed for the whole segment ( \(\:BCKtot\) ). Then the estimation is repeated for each 0.5-second segments ( \(\:{BCK}_{k}\) ) separately and compared to the noise level of the complete signal. If \(\:{BCK}_{k}>20*BCKtot\) , the k-th segment is marked as artifact. Frequency check: the maximum amplitude of the Fourier Transform is computed for each 0.5-second segment ( \(\:{\text{m}\text{a}\text{x}(FFT}_{k})\) ). If for the k-th segment the \(\:{\text{m}\text{a}\text{x}(FFT}_{k})\) is higher than 2.5 times the median across all the segments in the signal, the epoch is marked as artifact. In the current study, both the thresholds were experimentally set, as suggested in the related literature. Signal segments marked as noisy according to one of the two criteria were removed from the signal. After applying the two automatic artifact rejection approaches to the raw dataset, in both cases, only traces with a residual length of at least 4 seconds[ 5 ] were further analyzed leading to the definition of the COV dataset and the BCK dataset respectively in the first and second case. Further pre-processing steps were common for the four obtained datasets, including the selection of unique recordings distinguished by depth, electrode, and hemisphere, that is, if multiple traces shared the same depth, electrode, and hemisphere, the latest in time was selected as the one with sufficient quality satisfying the surgical equipe during the DBS exploration process. Signals were finally band-passed between 300 and 3000 Hz using a second-order elliptic (zero-phase) filter. 2.2.2. Feature extraction In line with our previous works [ 7 , 8 ] and the literature [ 17 , 18 ], 22 features were extracted from each MER trace, belonging to the time and frequency domain. Given the possible different lengths of the traces, when opportune, features were normalized with respect to the number of samples N composing the analyzed signal \(\:x\left(n\right);n=1:N\) . The list of the extracted features and the relative definitions are reported in Table I. Such features have been proposed in several previous literature studies [ 8 , 17 – 19 ]. Frequency domain features were extracted from the power spectral density (PSD) computed for each trace using the Welch method, employing 1-second windows (50% overlap), resulting in a spectral resolution of 1Hz. The PSD for frequency bands below 300Hz was derived from the mean-subtracted absolute value of the signal since the original data were bandpass filtered within 300-3000Hz [ 20 ]. Subsequently, power bands were extracted from these PSD values. Table I List of extracted features and their formulation Feature Definition WL- Wave or Curve length Represents the unfolded waveform’s distance \(\:WL=\frac{1}{N}\sum\:_{n=1}^{N-1}\left|x\left(n\right)+x\left(n+1\right)\right|\) ZC - Zero crossing The number of times the signal crosses the threshold calculated by estimating the noise level of the signal PKS - Peaks Number of positive peaks identified in a signal segment normalized for the segment length. MAV - Mean value of the absolute amplitude \(\:MAV=\:\frac{1}{N}\sum\:_{n=1}^{N}\left|x\left(n\right)\right|\) MED - Median value of absolute amplitude. Middle value separating the greater and lower halves of the ordered absolute amplitude \(\:\left|x\left(n\right)\right|\) of the trace TH - Signal threshold \(\:TH=\:\frac{3}{N-1}\sqrt{\sum\:_{n=1}^{N}{\left(x\left(n\right)-\stackrel{-}{x}\right)}^{2}}\) Root mean square (RMS) of the signal \(\:RMS=\sqrt{\:\frac{1}{N}\sum\:_{n=1}^{N}{x\left(n\right)}^{2}}\) AKUR - Amplitude distribution kurtosis \(\:AKUR=\:\frac{1}{(N-1){\sigma\:}^{4}}\sum\:_{n=1}^{N}\left|x\left(n\right)-{\stackrel{-}{x}}^{4}\right|\) ASKW - Amplitude distribution skewness \(\:ASKW=\:\frac{1}{(N-1){\sigma\:}^{3}}\sum\:_{n=1}^{N}\left|x\left(n\right)-{\stackrel{-}{x}}^{3}\right|\) NL - Noise level Derived from the signal’s analytic envelope [ 15 , 16 ] PWRA- Averaged Power \(\:PWRA=\:\frac{1}{N}\sum\:_{n=1}^{N}{x\left(n\right)}^{2}\) ANE - Average non-linear energy \(\:ANE=\frac{1}{N-2}\sum\:_{n=2}^{N-1}{\left[x\right(n)}^{2}-x\left(n-1\right)x(n+1)]\) [ 21 ] powVHFrel_1 Relative power in the 300-1000Hz frequency range powVHFrel_2 Relative power in the 1000-2000Hz frequency range powVHFrel_3 Relative power in the 2000-3000Hz frequency range powHFrel_1 Relative power in the 70-220Hz frequency range powHFrel_2 Relative power in the 220-320Hz frequency range powLFrel_1 Relative power in the 1-4Hz frequency range powLFrel_2 Relative power in the 4-8Hz frequency range powLFrel_3 Relative power in the 8-13Hz frequency range powLFrel_4 Relative power in the 13-30Hz frequency range powLFrel_5 Relative power in the 30-70Hz frequency range 2.2.3. Outlier detection and management Once features were extracted for all the prepared datasets, they were analyzed to assess their distributions and the presence of outliers. The impact of different treatment methods on the presence of outliers was considered another processing step worth investigating. Thus, we decided to test three different approaches: NONE - The first simple possibility is to not apply any outlier detection. Outlier Rejection for Hemisphere (ORH) of each patient set, that is the classic approach based on feature distribution to remove samples according to lower and upper bound (inter quartile range - IQR) identification with tolerance of 3 [ 22 ]. Outlier Rejection Model (ORM) based on machine learning methodologies, i.e. the local outlier factor algorithm (LOF), an unsupervised based algorithm which computes the local density deviation of a given data point with respect to its neighbors, applied on single patient’s data [ 23 ]. 2.2.4. Dataset Normalization As the last preparation step, we considered the possibility of normalizing the extracted features to reduce patients’ variability. Indeed, we applied feature standardization to all the generated datasets based on single patient cerebral hemisphere through Min-Max scaler [ 17 ]. As a result, a total of 24 processing pipelines were defined, each of them generating a dataset with different characteristics. 2.3. Classification models The impact of the different pipelines described above was evaluated through the performance of supervised machine learning (ML) classification models developed in Python (version 3.11.0) using the Scikit-learn library [ 24 ]. Specifically, we employed and compared three binary classifiers: a Support Vector Machine Classification (SVC) model, an Elastic Net (EN) model, and a Random Forest (RF) model. We selected model hyperparameters a priori based on domain knowledge and default library recommendations. Specifically, the SVC model adopted a Radial Basis Function (RBF) kernel, which enables non-linear decision boundaries. Other hyperparameters, including the regularization parameter (C) and kernel coefficient (gamma), were left at default values (C = 1.0, gamma= 'scale'). A fixed random state ensured reproducibility. The EN model is a logistic regression classifier employing elastic net regularization to combine L1 and L2 penalties, with an L1 ratio of 0.5 controlling the mix. The SAGA solver was used for efficient optimization. To ensure convergence, the maximum number of iterations was increased to 10000. A fixed random state was also used. The RF model used default hyperparameters. Specifically, the number of estimators was set to 100, Gini criterion was used as a measure of the split quality, the maximum depth of each tree was unrestricted, the minimum number of samples required to split an internal node was 2, and the minimum number of samples needed to be at a leaf node was 1. A fixed random state was used to ensure reproducibility. Strategies were incorporated into the modelling pipeline to address class imbalance, which can adversely affect the performance of classification algorithms. The dataset was partitioned into training and test subsets and preprocessing steps were applied. These included cross-validation and normalization. Cross-validation was employed to reduce overfitting and enhance the generalization ability of the models. Specifically, we applied stratified k-fold cross-validation with shuffling, maintaining the original class distribution within each fold. The dataset was divided into k = 5 non-overlapping subsets, each approximately equal in size. Each iteration used one-fold as the test set while the remaining k − 1 folds served as the training set. This process was repeated k times, ensuring that every fold was used once as the test set. The final model performance was assessed by averaging the results across all folds. Before training, the features in the training set that did not previously undergo normalization (see paragraph 2.2.4.) were scaled using MinMax normalization, implemented via the `MinMaxScaler` function from the Scikit-learn library (Python). Normalization parameters were derived exclusively from the training set and subsequently applied to the test data. This transformation scaled each feature to the [0, 1] range across the whole dataset. The described classifiers were applied to the full set of features and after an automatic feature selection approach based on Recursive Feature Elimination (RFE) with cross-validation [ 25 ], leading to six classification models applied at the end of each processing pipeline. 2.3.1. Performance Evaluation The classification of each model was evaluated by computing performance metrics based on the confusion matrix and the Receiver Operating Characteristic (ROC) curve at each classification split. The computed metrics were the area under the ROC curve (AUC), classification accuracy, precision, recall and F1-score, as reported in Eq. (1–4) \(\:Accuracy=\frac{True\:Positive+True\:Negative}{Total\:samples}\) (1) \(\:Precision=\frac{True\:Positive}{True\:Positive+False\:Positive}\) (2) \(\:Recall=\frac{True\:Positive}{True\:Positive+False\:Negative}\) (3) \(\:F1score=2*\frac{Precision*Recall}{Precision+Recall}\) (4) Finally, to better understand the importance of the extracted features for and the impact of their contribution to the classification results, the Shapley additive explanations (SHAP) method was applied and results explored [ 26 , 27 ]. Specifically, features were ranked by means of their SHAP value (probability) and we counted how many times each feature was ranked in the top ten positions across the pipelines and folds in order to define the percentage of presence across all the pipelines. The counting was kept separated for the three classifiers. 3. Results In this section, we describe the results of the compared classification pipelines, starting from the effect of the early preparation steps or dataset pre-processing. These steps, comprising the management of MER signal artefacts, feature extraction, and subsequent outlier management, determine twelve different datasets regarding sample numerosity on which the classification models are trained. Then, a feature normalization procedure is applied to all the resulting datasets, leading to the generation of twelve additional datasets. The classification performance of the 24 tested pipelines, given by the different combinations, is described and commented on. Finally, an exploration of the importance of the features performed through the SHAP approach is also reported. 3.1. Pre-processing results: datasets composition Table II reports the number of samples in each dataset specifying the observations for each class (STN and NOT STN). Datasets are the results of the different combinations of early steps procedures that were tested for dataset preparation. Thus, the original 'RAW' dataset is composed of 1228 observations (804 NOT STN, 424 STN); the MER signals are then screened for artefact rejection by an expert (V. L.), and the whole 10-sec segments are removed if marked as 'with artefacts'. This procedure reduced the number of total samples to 1115 in the 'EXP' dataset. The RAW dataset was also processed for artefact rejection using two automatic approaches described in 2.2.1. Both methods did not automatically reject the whole 10-sec signal segment, but only the portion with the artefact, and the remaining segment was further considered if with a length > 4 seconds. This approach allowed for less drastic dataset pruning. Indeed, the COV approach determined the rejection of 12.82% of the signal, and, in terms of whole segments, only 11 were removed. As for the BCK method, 12.47% of the signal was rejected, and 21 whole segments were removed. The overall agreement between the two algorithms reached 88.5% with a statistically significant (p < 0.05) Cohen's Kappa [ 1 ] with a mean across subjects equal to 0.49 (where (0.2 ≤ k ≤ 0.4 is 'Fair agreement', 0.41 ≤ k ≤ 0.6 is 'Moderate agreement', and 0.61 ≤ k ≤ 0.8 is 'Substantial agreement'). The different formulation of the rejection criteria justifies this moderate agreement. Specifically, COV identified more 'short' artefacts, leading to fewer entirely rejected segments concerning BCK. After this first step, all the features described were extracted from the remaining MER segments. At this point, the dataset was tested for outliers applying two approaches, ORH and ORM, and the effect of this additional procedure was tested against the possibility of not controlling for the presence of outliers (none). Indeed, applying the ORH methods determined the exclusion of 20.1% of samples from the RAW dataset, 19.9% from EXP, 15.5% from COV and 13.5% from the BCK dataset. Applying the ORM method led to the following dataset reductions: RAW = 10.9%, EXP = 11.0%, COV = 10.8%, and BCK = 10.9%. Table II Numerosity of each dataset according to the preparation pipeline. The total number of samples and the number of samples in each class (STN and NOT STN) are shown. Outlier treatment none ORH ORM Dataset total NOT STN STN total NOT STN STN total NOT STN STN RAW 1228 804 424 981 633 348 1094 757 337 EXP 1115 726 389 893 582 311 992 697 295 COV 1217 794 423 1028 693 335 1085 775 310 BCK 1207 793 414 1044 699 345 1076 769 307 Finally, as the last possible dataset preparation step, the hemisphere-based features normalization was applied to all the already mentioned combinations, leading to 24 prepared datasets. 3.2. Effect of processing pipelines on performance evaluation The described classifiers, with and without RFE feature selection step, led to six classification models being applied to the 24 datasets. Performance metrics are reported as mean and standard deviation values computed across the five cross-folds. Figures 2 , 3 and 4 show the performance metrics of Accuracy, F1-score, AUC, precision and recall values across all the pipelines directly comparing the six classification models. Overall, the most critical step seems to be the application of the feature normalization based on data from the same hemisphere of each patient. This step has the aim to prevent model overfitting, but in our case, it reduces the performance of all the classifiers on all possible datasets. Indeed, accuracy values pass from a range (min-max across all the pipelines) of 0.91–0.945 for the not pre-normalized datasets, to an accuracy range of 0.846–0.88 when the hemisphere-based normalization is applied. Similarly, the F1-score range of values is reduced from 0.867–0.922 to 0.742–0.835 across the twelve pipelines applied to non–normalized and normalized datasets. As for the effect of the other applied procedures, it can be observed that the artefacts management approaches alone provide a weak performance improvement, particularly for the EN and RF classification models. Table III reports these specific results for the not normalized dataset as the mean and standard deviation of Accuracy and F1-score values. A further slight improvement due to the ORH procedure was obtained for the RF classification model on COV treated dataset (Accuracy = 0.945 (0.029), F1-score = 0.915 (0.044)), and the SVC on the RAW dataset (Accuracy = 0.943 (0.015), F1-score = 0.92 (0.02)). ORM, however, appears to have no discernible impact on the results. Table III: Mean and standard deviation of accuracy and F1-score values obtained applying the three classifiers considering the full set of features on datasets only processed with the artefact rejection approaches. EN RF SVC ACC F1-score ACC F1-score ACC F1-score RAW 0.912 (0.025) 0.876 (0.032) 0.932 (0.013) 0.899 (0.019) 0.931 (0.021) 0.902 (0.028) EXP 0.916 (0.016) 0.881 (0.023) 0.936 (0.015) 0.906 (0.022) 0.931 (0.02) 0.902 (0.028) COV 0.929 (0.009) 0.899 (0.013) 0.938 (0.015) 0.909 (0.022) 0.93 (0.011) 0.902 (0.015) BCK 0.926 (0.013) 0.893 (0.016) 0.934 (0.017) 0.903 (0.025) 0.932 (0.016) 0.904 (0.021) Similar trends are observed for the AUC metrics represented in Fig. 3 . Looking at more detailed metrics, such as the precision and recall shown in Fig. 4 , it is possible to appreciate the different behavior of the classification models adopted that is not clearly visible considering the previous three metrics. Specifically, while EN and SVC, both with or without the automatic feature selection, performed similarly with a good balance between precision and recall, the RF classifier showed a higher precision and a lower recall across all the datasets, both with and without feature selection. As for applying a feature selection algorithm before the classification, we did not find a particular performance improvement since in most cases only a few features were not considered, while in some cases no features were removed (maximum number of removed features = 8). This result is probably due to the already quite small proposed set of independent features, since a preliminary analysis has been performed based on literature and previous experience to avoid the extraction of highly correlated features [ 8 ]. 3.3. Analysis of feature importance This section presents an in-depth analysis of the importance of features across different classification models. In particular, the aim was to assess the robustness of the feature importance in function of the dataset preparations by the quantification of their contribution to the STN classification using the SHAP approach. Specifically, features were ranked by means of their SHAP value (based on probability), and we counted how many times each feature was ranked in the first ten positions across the pipelines (five folds each). Based on this count, we defined the percentage of presence across all the pipelines and considering the not normalized and the normalized features separately. The counting was kept separated for the three classifiers, since results were model-specific. Figure 5A shows the percentage of presence for each feature in the top-ranked positions, meaning that they contributed to the classification with high importance. First of all, it is possible to notice that some features seem to not contribute to classification of the STN. This observation is coherent across classifiers and with low impact due to the application of the dataset normalization. These low-ranked features are (percentage of presence 50%), such as RMS, powHFrel 1 − 2 , powVHFrel 2 − 3 , MAV, MED, PKS and WL. Interestingly, RF showed different feature importance attributes with respect to SVC and EN classifiers that were, instead, quite similar to each other. Moreover, it is possible to observe an effect due to the dataset normalization. Specifically, PKS, RMS and TH seem to reduce their importance for classification, while WL, NL and powHFrel 2 increased their contribution after the hemisphere-based normalization was applied. Figure 5B shows an example of a SHAP representation obtained when the pipeline reaches the highest accuracy (COV-ORH). From this representation, it is also possible to interpret how the features impact the result. For example, a high value of RMS and MAV is associated with the Fig. 5A) Features percentage of presence in the top ranked set of important features across all the pipelines, the pipeline without normalization and with normalization for the EN model (red), RF model (green) and SVC model (blue). B) SHAP Beeswarm plot for the models applied to the pipeline reaching the highest accuracy (COV-ORH). STN (class 1) while a high value of PKS and powVHFrel 3 indicates class 0 (NOT STN). To confirm further the repeatability of feature contributions to the classification results, the same approach was repeated to analyze the features importance for the models trained on a selected set of features (of variable number given the automatic approach employed). Figure 6 reports the results in terms of percentage of presence in the top 10 ranked features. It is possible to confirm the previous results with an overall agreement and with few changes in the rankings. 4. Discussion Accurate detection of the STN as the target for DBS electrode implantation is fundamental to efficient neuromodulation treatment and sufficient improvement of symptoms in patients with Parkinson's disease. To this end, imaging-guided stereotactic trajectories are carefully planned for electrode implantation. However, to cope with potential brain shifts and discriminate between close brain structures, explorative MER signals are routinely recorded to provide surgeons with additional functional information. Indeed, information from MER can be digitally extracted and analyzed, leading to the exploitation of machine learning (ML) and deep learning (DL) approaches widely applied in the literature for target detection and to provide support for the medical team [ 6 , 28 ]. In this context, much attention has been paid to improving the accuracy of feature-based ML and DL algorithms by increasing the complexity of the methods [ 9 , 18 ]. However, few studies report in detail how the signals have been pre-processed and how the dataset has been managed. In very rare cases, the impact of these procedures has also been discussed [ 13 , 29 ]. Therefore, the goal of this study was to analyze the impact of different MER data processing and management pipelines on the classification of the STN using feature-based ML. To this end, we implemented an evaluation framework to assess the impact of different choices for managing specific crucial procedural steps. Combined, these choices produced a total of 24 pipelines. Three ML models were then used for binary classification, both with and without a further feature selection step. Finally, through the application of SHAP analysis, we explored the impact of the different pipelines on feature importance and model explainability. Specifically, we first analyzed the impact of different artefacts rejection procedures to clean MER signals before feature extraction. Our results suggest that slight improvements in accuracy can be observed (see Table III), and that different artefact rejection approaches (i.e. EXP, COV [ 13 , 14 ] and BCK [ 15 , 16 ]) can similarly improve MER classification compared to not removing them at all. Interestingly, the effect of this varies depending on the final classification model applied: the SVC model produced the same result for all four datasets. Following features extraction, three outlier management procedures were considered and their effect analyzed. Results suggested that the classical ORH approach generally increased both accuracy and F1-score when applied to three out of four datasets (i.e., RAW, EXP and COV), while no improvement was observed if applied to BCK dataset. Instead, the ORM approach did not improve performance, and, in some cases, lower accuracy was obtained compared to the possibility of not removing outliers. These considerations also apply to pipelines, including the hemisphere-based normalization step, albeit to a lesser extent. As last manipulation, indeed, we performed a standardization procedure based on data from single hemisphere as suggested in [ 17 ]. While the obvious aim of this method is to improve generalizability and avoid overfitting, in our case it led to a general performance decrease, suggesting that the intrinsic normalization applied within model training may be enough for single center MER classification study. Our final analysis regarded the explainability of features contribution to classification outcomes, by ranking their SHAP values across pipelines and validation folds. From this evaluation, it was possible to observe that some features (i.e., RMS, powHFrel 1 − 2 , powVHFrel 2 − 3 , MAV, MED, PKS and WL) greatly contribute to MER classification accuracy, consistently across pipelines and classification models, while some others (i.e., ZC, AKUR, ASKW, powLFre 2 − 5 , powVHFrel 1 ) resulted in a not significantly contribution to STN detection. Interestingly, if we compare feature contribution between normalized and not normalized datasets, slightly different importance among the top ranked features can be observed, while few changes are present among the less contributing ones. These considerations are still valid considering results obtained after feature selection. In summary, in line with previous literature [ 8 , 17 , 18 , 30 , 31 ], our study demonstrated that, besides their simplicity, feature-based ML approaches are effective in classifying MER traces for STN detection with good accuracy (best result: COV-ORH Accuracy = 0.945 (0.029)). Even so, pre-processing and data preparation approaches should be carefully selected and applied since they may differently impact classification performance, both depending on the original dataset and the classification model applied. 4.1. Limitations Although our study provides an in-depth and valuable analysis of the impact of processing pipelines on STN identification using MER traces and ML approaches, it is affected by some limitations that may have influenced results interpretation. A first limitation regards the artefact detection approaches. In our study, in fact, we only evaluated the effects in terms of dataset reduction (Table II) and final impact on classification performance, while we could not perform an accurate analysis since systematic and precise labeling of artefacts present in our dataset was not performed. Indeed, the EXP dataset resulted from a manual dataset pruning in the presence of evident artefacts (i.e., the whole 10-second segment is discarded without identification of artefact timing and type). Further study, in line with [ 9 , 13 ] should be performed for a complete characterization of potential artefact affecting MER recordings. Second, in our study data from a single center (28 Parkinsonian patients) have been included, limiting methods and results generalizability, which is a fundamental step to further improve both ML and DL actual applicability in clinical practice [ 11 ]. Furthermore, it should be noted that the classification models were implemented using predefined and default hyperparameter settings to avoid introducing factors that could impact STN detection performance, as our main aim was to analyse the effect of different data preparation pipelines. Indeed, our reported accuracy aligns with the literature implementing similar approaches [ 6 ]. 5. Conclusion The presented study provides a comprehensive comparison of MER processing and dataset management procedures in different pipelines for the functional identification of the STN. The work focused on examining and describing how different choices might affect classification performance. It also provided a potential framework for comparison and useful guidelines for designing ML algorithms for STN localization based on MER traces. Specifically, results pointed out the need of properly identifying and rejecting artefacts paired with appropriate outlier management. The highest accuracy values were obtained for the COV-ORH pipeline, while a pre-normalization of features based on data from single patient and brain hemisphere led to performance degradation in our study. Finally, we used the SHAP approach to further explore feature importance, which is a fundamental step that could guide and improve the implementation of future algorithms. Declarations Given that MER data are routinely acquired during standard surgical procedures, when we started the study we had a EC waiver. We were asked to make patients sign an informed consent regarding both the surgery procedure and data exploitation for research aims. Of course, the research was conducted in accordance with the principles embodied in the Declaration of Helsinki and in accordance with local statutory requirements. References Kremer NI, van Laar T, Lange SF, Statius Muller S, la, Bastide-van Gemert S, Oterdoom DM et al (2023) STN-DBS electrode placement accuracy and motor improvement in Parkinson’s disease: systematic review and individual patient meta-analysis. J Neurol Neurosurg Psychiatry. ;94:236–44 van den Munckhof P, Bot M, Schuurman PR (2021) Targeting of the Subthalamic Nucleus in Patients with Parkinson’s Disease Undergoing Deep Brain Stimulation Surgery. Neurol Ther 10:61–73 Vinke RS, Geerlings M, Selvaraj AK, Georgiev D, Bloem BR, Esselink RAJ et al (2022) The Role of Microelectrode Recording in Deep Brain Stimulation Surgery for Parkinson’s Disease: A Systematic Review and Meta-Analysis. J Parkinsons Dis 12:2059–2069 Zakharov N, Belova E, Gamaleya A, Tomskiy A, Sedov A (2024) Neuronal activity features of the subthalamic nucleus associated with optimal deep brain stimulation electrode insertion path in Parkinson’s disease. Eur J Neurosci Wan KR, Maszczyk T, See AAQ, Dauwels J, King NKK (2019) A review on microelectrode recording selection of features for machine learning in deep brain stimulation surgery for Parkinson’s disease. Clin Neurophysiol 130:145–154 Inggas MAM, Coyne T, Taira T, Karsten JA, Patel U, Kataria S et al (2024) Machine learning for the localization of Subthalamic Nucleus during deep brain stimulation surgery: a systematic review and Meta-analysis. Neurosurg Rev 47:774 Coelli S, Levi V, Del Vecchio V, Mailland J, Rinaldo E, Eleopra S et al (2020) R,. Characterization of Microelectrode Recordings for the Subthalamic Nucleus identification in Parkinson’s disease. 42nd Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC). IEEE; 2020. pp. 3485–8 Coelli S, Levi V, Del Vecchio V, Mailland J, Rinaldo E, Eleopra S (2021) An intra-operative feature-based classification of microelectrode recordings to support the subthalamic nucleus functional identification during deep brain stimulation surgery. J Neural Eng 18:016003 Hosny M, Zhu M, Gao W, Fu Y (2020) A novel deep LSTM network for artifacts detection in microelectrode recordings. Biocybern Biomed Eng 40:1052–1063 Maged A, Zhu M, Gao W, Hosny M (2024) Lightweight deep learning model for automated STN localization using MER in Parkinson’s disease. Biomed Signal Process Control. ;96 Martin T, Jannin P, Baxter JSH (2024) Generalisation capabilities of machine-learning algorithms for the detection of the subthalamic nucleus in micro-electrode recordings. Int J Comput Assist Radiol Surg Gorlini C, Forzanini F, Coelli S, Rinaldo S, Eleopra R, Bianchi AM et al (2024) Impact of Microelectrode Recording Artefacts on Subthalamic Nucleus Functional Identification via Features-Based Machine Learning Classifiers. 2024 IEEE International Conference on Metrology for eXtended Reality, Artificial Intelligence and Neural Engineering (MetroXRAINE). IEEE; pp. 13–8 Bakštein E, Sieger T, Wild J, Novák D, Schneider J, Vostatek P et al (2017) Methods for automatic detection of artifacts in microelectrode recordings. J Neurosci Methods 290:39–51 Aboy M, Falkenberg JH (2006) An Automatic Algorithm for Stationary Segmentation of Extracellular Microelectrode Recordings. Med Biol Eng Comput 44:511–515 Cagnan H, Dolan K, He X, Contarino MF, Schuurman R, van den Munckhof P et al (2011) Automatic subthalamic nucleus detection from microelectrode recordings based on noise level and neuronal activity. J Neural Eng [Internet]. ;8:046006. Available from: https://iopscience.iop.org/article/10.1088/1741-2560/8/4/046006 Dolan K, Martens HCF, Schuurman PR, Bour LJ (2009) Automatic noise-level detection for extra-cellular micro-electrode recordings. Med Biol Eng Comput 47:791–800 Bellino GM, Schiaffino L, Battisti M, Guerrero J, Rosado-Muñoz A (2019) Optimization of the KNN supervised classification algorithm as a support tool for the implantation of deep brain stimulators in patients with Parkinson’S Disease. Entropy. ;21 Benouis M, Rosado-Muñoz A (2024) Using Ensemble of Hand-Feature Engineering and Machine Learning Classifiers for Refining the Subthalamic Nucleus Location from Micro-Electrode Recordings in Parkinson’s Disease. Appl Sci (Switzerland). ;14 Cao L, Li J, Zhou Y, Liu Y, Liu H (2020) Automatic feature group combination selection method based on GA for the functional regions clustering in DBS. Comput Methods Programs Biomed 183:105091 Moran A, Bar-Gad I (2010) Revealing neuronal functional organization through the relation between multi-scale oscillatory extracellular signals. J Neurosci Methods 186:116–129 Wong S, Baltuch GH, Jaggi JL, Danish SF (2009) Functional localization and visualization of the subthalamic nucleus from microelectrode recordings acquired during DBS surgery with unsupervised machine learning. J Neural Eng 6:026006 Smiti A (2020) A critical overview of outlier detection methods. Comput Sci Rev 38:100306 Breunig MM, Kriegel H-P, Ng RT, Sander J (2000) LOF. Proceedings of the 2000 ACM SIGMOD international conference on Management of data. New York, NY, USA: ACM; pp. 93–104 Pedregosa FABIANPEDREGOSAF, Michel V, Grisel OLIVIERGRISELO, Blondel M, Prettenhofer P, Weiss R et al (2011) Scikit-learn: Machine Learning in Python Gaël Varoquaux Bertrand Thirion Vincent Dubourg Alexandre Passos PEDREGOSA, VAROQUAUX, GRAMFORT ET AL. Matthieu Perrot [Internet]. Journal of Machine Learning Research. Available from: http://scikit-learn.sourceforge.net Guyon I, Weston J, Barnhill S, Vapnik V (2002) Gene Selection for Cancer Classification using Support Vector Machines. Mach Learn 46:389–422 Ponce-Bobadilla AV, Schmitt V, Maier CS, Mensing S, Stodtmann S (2024) Practical guide to SHAP analysis: Explaining supervised machine learning model predictions in drug development. Clin Transl Sci. ;17 Shapley L (1953) 7. A Value for n-Person Games. Contributions to the Theory of Games II 307–317. Classics in Game Theory. Princeton University Press; 1997. pp. 69–79 Chao-Chia Lu D, Boulay C, Chan ADC, Sachs AJ (2024) A Systematic Review of Neurophysiology-Based Localization Techniques Used in Deep Brain Stimulation Surgery of the Subthalamic Nucleus. Neuromodulation: Technology at the Neural Interface. ;27:409–21 Varga I, Bakstein E, Gilmore G, May J, Novak D (2024) Statistical segmentation model for accurate electrode positioning in Parkinson’s deep brain stimulation based on clinical low-resolution image data and electrophysiology. PLoS ONE 19:e0298320 Rajpurohit V, Danish SF, Hargreaves EL, Wong S (2015) Optimizing computational feature sets for subthalamic nucleus localization in DBS surgery with feature selection. Clin Neurophysiol 126:975–982 Khosravi M, Atashzar SF, Gilmore G, Jog MS, Patel RV (2020) Intraoperative Localization of STN during DBS Surgery Using a Data-Driven Model. IEEE J Transl Eng Health Med. ;8 Additional Declarations The authors declare no competing interests. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-6915773","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Method Article","associatedPublications":[],"authors":[{"id":472648357,"identity":"f7a25d16-f576-4780-9fd2-9c6eb63062ce","order_by":0,"name":"Vincenzo Levi","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABJ0lEQVRIie3Rv0vDQBQH8BceXJbUWy8I+RuuHJRs+VfuCDiK4JJBQqFwU6Wrgvg3dBI3GwJxCXTNmOI/EHGpUNFLtaI26exwXzju8eDDux8ANjb/MbgrXLNkwr478nPjB0hbyJKBh6ZabEnb4h3mJ3E0gAe/yP4Y/jjImuf7NADEgq1uwyhyaXHcJHBK6TJr4GyzR/Kj2L8uiQAkJ0zdMTVFJGxRwrl/FSPrOJg/8UY40J4aoye4IdJrSabf1bxC6CVvmhlCX7i6YdEXAfWwzHHdQSga4mjeTnFqNWbOdEfmEJOuKYYI/1JLQZAMa1mYu+QowtLchVXxKJRc/CWElsPmVacBdSd1tr5II3eWraqkfbFZ9lQ1m6D3W0h3W/YCGxsbG5sD+QC0U1i2nforuQAAAABJRU5ErkJggg==","orcid":"https://orcid.org/0000-0003-2713-6357","institution":"Functional Neurosurgery Unit, Neurosurgery Department, Fondazione IRCCS Istituto Neurologico Carlo Besta, Milano, Italy","correspondingAuthor":true,"prefix":"","firstName":"Vincenzo","middleName":"","lastName":"Levi","suffix":""},{"id":472648358,"identity":"b4baf5d6-c189-4c06-ba8b-611b8fb6ae5d","order_by":1,"name":"Stefania Coelli","email":"","orcid":"https://orcid.org/0000-0003-4104-4790","institution":"Department of Electronic, Information and Bioengineering, Politecnico di Milano, Milano (Italy)","correspondingAuthor":false,"prefix":"","firstName":"Stefania","middleName":"","lastName":"Coelli","suffix":""},{"id":472648359,"identity":"672dafce-b1ba-4ea9-a28c-39f47019f88d","order_by":2,"name":"Chiara Gorlini","email":"","orcid":"","institution":"Functional Neurosurgery Unit, Neurosurgery Department, Fondazione IRCCS Istituto Neurologico Carlo Besta, Milano, Italy","correspondingAuthor":false,"prefix":"","firstName":"Chiara","middleName":"","lastName":"Gorlini","suffix":""},{"id":472648360,"identity":"3a437280-fd14-42c3-ae31-c30aa92d80d8","order_by":3,"name":"Federica Forzanini","email":"","orcid":"","institution":"Functional Neurosurgery Unit, Neurosurgery Department, Fondazione IRCCS Istituto Neurologico Carlo Besta, Milano, Italy","correspondingAuthor":false,"prefix":"","firstName":"Federica","middleName":"","lastName":"Forzanini","suffix":""},{"id":472648361,"identity":"506ed232-d9e0-4daa-96ac-bc42580e6508","order_by":4,"name":"Sara Rinaldo","email":"","orcid":"https://orcid.org/0000-0002-5945-132X","institution":"Parkinson and Movement Disorders Unit, Department of Clinical Neurosciences, Fondazione IRCCS Istituto Neurologico Carlo Besta, Milano, Italy","correspondingAuthor":false,"prefix":"","firstName":"Sara","middleName":"","lastName":"Rinaldo","suffix":""},{"id":472648362,"identity":"bb867848-c068-4fc1-8634-4f821993c357","order_by":5,"name":"Nico Golfre Andreasi","email":"","orcid":"https://orcid.org/0000-0003-3555-5367","institution":"Parkinson and Movement Disorders Unit, Department of Clinical Neurosciences, Fondazione IRCCS Istituto Neurologico Carlo Besta, Milano, Italy","correspondingAuthor":false,"prefix":"","firstName":"Nico","middleName":"Golfre","lastName":"Andreasi","suffix":""},{"id":472648363,"identity":"4815a80f-3c39-49db-a09f-7d985154a70c","order_by":6,"name":"Luigi Romito","email":"","orcid":"https://orcid.org/0000-0002-6772-1035","institution":"Parkinson and Movement Disorders Unit, Department of Clinical Neurosciences, Fondazione IRCCS Istituto Neurologico Carlo Besta, Milano, Italy","correspondingAuthor":false,"prefix":"","firstName":"Luigi","middleName":"","lastName":"Romito","suffix":""},{"id":472648364,"identity":"5be0ea50-26ab-49d1-9e5d-4a84888620a7","order_by":7,"name":"Roberto Eleopra","email":"","orcid":"https://orcid.org/0000-0002-6035-9715","institution":"Parkinson and Movement Disorders Unit, Department of Clinical Neurosciences, Fondazione IRCCS Istituto Neurologico Carlo Besta, Milano, Italy","correspondingAuthor":false,"prefix":"","firstName":"Roberto","middleName":"","lastName":"Eleopra","suffix":""},{"id":472648365,"identity":"b9332c0d-4cab-4049-978d-716417f81046","order_by":8,"name":"Anna Maria Bianchi","email":"","orcid":"https://orcid.org/0000-0002-8290-7460","institution":"Department of Electronic, Information and Bioengineering, Politecnico di Milano, Milano (Italy)","correspondingAuthor":false,"prefix":"","firstName":"Anna","middleName":"Maria","lastName":"Bianchi","suffix":""}],"badges":[],"createdAt":"2025-06-17 15:18:57","currentVersionCode":1,"declarations":{"humanSubjects":true,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":true,"humanSubjectConsent":true,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-6915773/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6915773/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":85319171,"identity":"f509af71-ed14-48e7-aab7-aff0b9fe08b0","added_by":"auto","created_at":"2025-06-24 15:01:12","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":245914,"visible":true,"origin":"","legend":"\u003cp\u003eSchematic framework for comparing the implemented pipelines. Four pre-processing approaches for managing artefacts were implemented prior to feature extraction. The extracted feature sets underwent two procedures to identify and correct outliers. The twelve datasets obtained from the combinations of methods are then either passed directly to the three classification models (with or without a feature selection step) or normalised first.\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-6915773/v1/551baee540ae058eec2a6e75.png"},{"id":85318319,"identity":"4603d786-dd1c-4f84-9b3a-c4372a776b05","added_by":"auto","created_at":"2025-06-24 14:53:12","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":372495,"visible":true,"origin":"","legend":"\u003cp\u003ePerformance evaluation metrics in function of the dataset processing pipeline. Mean values (and standard deviation) of F1-score and Accuracy.\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-6915773/v1/4c4e27cc7b0bff10a2a693bf.png"},{"id":85318337,"identity":"e5177865-7e6f-4014-bf33-e71e9a038e5b","added_by":"auto","created_at":"2025-06-24 14:53:13","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":75205,"visible":true,"origin":"","legend":"\u003cp\u003eMean values (and standard deviation) of the Area under the curve (AUC).\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-6915773/v1/25aa227ecb23b6dedf3c96d3.png"},{"id":85319172,"identity":"a79591e8-4e8a-4e37-aee3-ef63da79e0d3","added_by":"auto","created_at":"2025-06-24 15:01:12","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":265452,"visible":true,"origin":"","legend":"\u003cp\u003ePerformance evaluation metrics in function of the dataset processing pipeline. Mean values (and standard deviation) of Recall and Precision.\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-6915773/v1/b7dc5804954fe316c934a031.png"},{"id":85318311,"identity":"27a29cf9-79a1-42a9-a8dc-f1abfb01f970","added_by":"auto","created_at":"2025-06-24 14:53:12","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":1133689,"visible":true,"origin":"","legend":"\u003cp\u003eA) Features percentage of presence in the top ranked set of important features across all the pipelines, the pipeline without normalization and with normalization for the EN model (red), RF model (green) and SVC model (blue). B) SHAP Beeswarm plot for the models applied to the pipeline reaching the highest accuracy (COV-ORH).\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-6915773/v1/e92198250bb4de8383d7f255.png"},{"id":85318327,"identity":"6c155910-de33-4b8a-8eb4-a83a93e4660a","added_by":"auto","created_at":"2025-06-24 14:53:12","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":245272,"visible":true,"origin":"","legend":"\u003cp\u003eFeatures percentage of presence in the top ranked set of important features across all the pipelines, the pipeline without normalization and with normalization for the EN model (red), RF model (green) and SVM model (blue) in the case of feature selection.\u003c/p\u003e","description":"","filename":"6.png","url":"https://assets-eu.researchsquare.com/files/rs-6915773/v1/904366ee92401a75ece513a3.png"},{"id":85320045,"identity":"9a8c6fd8-a819-4eed-a8aa-a3e4c1c94b4b","added_by":"auto","created_at":"2025-06-24 15:09:15","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":3087483,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6915773/v1/ca7bb6e2-1e8a-4736-8eba-ddc9bf4e34ef.pdf"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"\u003cp\u003e\u003cstrong\u003eThe role of MER processing pipelines for STN functional identification during DBS surgery: a feature based machine learning approach.\u003c/strong\u003e\u003c/p\u003e","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eDeep brain stimulation (DBS) is a safe and effective surgical treatment used to control the symptoms of Parkinson\u0026rsquo;s disease[\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. Through DBS surgery a permanent electrode is implanted inside the brain to chronically deliver high-frequency electrical pulses to the subthalamic nucleus (STN). Although STN is easily visible on preoperative Magnetic Resonance Imaging (MRI), sources of error arising from brain shift and small uncontrollable deviations from the pre-planned electrode trajectory often require intra-operative real-time data to verify the correct placement of the electrode in the planned target [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. One of the modalities used to confirm pre-operative planning is micro-electrode recording (MER) [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. MER explores electrophysiological properties of the brain tissue surrounding the STN and within the STN itself. In a typical DBS surgery, the microelectrodes record electrophysiological activity along a planned track as they are sequentially advanced into the brain by the neurosurgeon. Since each part of the brain has its own characteristic neural activity (such as spike firing counts and patterns), that of the STN can be recognized over the background noise level [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. As a result, based on monitoring of this electrophysiological activity, the neurosurgeon decides when the microelectrode has entered the STN. However, MER intra-operative interpretation may be challenging and time-consuming requiring a high level of proficiency and expertise. Recently, several machine learning algorithms have been proposed as decision-making support tools with the aim of minimizing subjectivity and improving patient outcomes [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e, \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]. These algorithms are based on a variety of different machine learning paradigms. Most methods use traditional approaches which classify signals based on specific features that are known to be of interest, such as the power in particular frequency bands or various spike-dependent or spike-independent features [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e, \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. Others exploit a more modern, deep learning approach in which the entire signal is provided to the algorithm rather than reducing it to a smaller vector of features [\u003cspan additionalcitationids=\"CR10\" citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. One common limitation of these approaches does not lie specifically in the paradigm or machine learning architecture chosen, but in methodological design and dataset processing [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e, \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. Few studies indeed report in detail how the signals have been pre-processed and how the dataset has been managed [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. Therefore, the goal of this study was to assess the impact of different MER data processing and management pipelines on the classification of the STN using feature-based ML. A total of twenty-four combinations of processing approaches have been implemented, comprising four signal artefact treatments and three outlier management procedures, as well as the option to standardize or not the feature sets. Three classifiers were then trained and tested, both with and without feature selection, to evaluate the effects of the implemented pipeline on the classification results. The paper is organized as follows. In section 2, first the analyzed dataset is described, then in section 2.2 the implemented pipelines and the evaluation framework are explained in detail. Specifically in 2.2.1 the artefact rejection approaches are described, in 2.2.2 all the extracted features are listed, in 2.2.3 and 2.2.4 the outlier management and the feature standardization methods are respectively presented. Section 2.3 provides a description of the implemented classifiers and of performance evaluation. Results are reported in section 3 and discussed in section 4, also addressing the study limitations. Finally, the study is summarized in section 5.\u003c/p\u003e"},{"header":"2. Materials and Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1. Dataset\u003c/h2\u003e \u003cp\u003eRetrospective MER data from 28 Parkinson\u0026rsquo;s disease patients undergoing bilateral DBS surgical procedures targeting the STN from 2020 to 2023 at the Neurosurgery Department of the Fondazione IRCCS Istituto Neurologico Carlo Besta (Milan, Italy) were reviewed. Patients signed an informed consent regarding the surgery procedure and data exploitation for research aims. In awake patients, MERs were collected employing a Medtronic LeadpointTM system (Medtronic Inc. Minneapolis, MN, USA). Stereotactic planning was performed using the StealthStation Surgical Navigation System (Medtronic, Minneapolis, MN, USA). Direct targeting on T2 pre-operative MRI was always performed for stereotactic targeting of the STN. All patients had a preoperative stereotactic imaging with MRI or CT and immediate intra-operative verification of electrode positioning with intra-operative CT scan (O-Arm, Medtronic). Surgical procedures were performed using the Vantage frame (Elekta, Stockholm, Sweden). Glass-coated platinum/iridium micro-electrodes with impedance of 0.4\u0026ndash;1.0 mΩ (FHC Inc., Bowdoinham ME, USA) were used. Each hemisphere was explored using three simultaneous parallel microelectrodes (anterior, central, posterior) spaced 2 mm apart one from another. Recordings lasted at least 10 s for each single depth. MER recordings started at an estimated distance from the target (EDT) of 10 mm and extended to 5 mm EDT after the planned target, advancing with 1 mm steps. Traces were high pass filtered at 200 Hz to allow the visualization of the firing activity during surgery and digitalized with a sampling rate of 24 kHz. At each EDT, traces were recorded for a duration of at least 10 seconds. Along the electrode path, various functional regions were identified, including white matter, thalamus (TH), zona incerta (ZI), substantia nigra (SNr), and STN. For each patient\u0026rsquo;s surgical procedure, the intraoperative annotations were reviewed by an expert neurosurgeon and the MER traces assigned to two classes: \u0026lsquo;STN\u0026rsquo; if recognized as within the STN, and \u0026lsquo;NOT STN\u0026rsquo; if belonging to other structures.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2. Dataset Preparation Pipelines\u003c/h2\u003e \u003cp\u003eIn this section, the different steps adopted to compose the pipelines tested in the current work are described. The full framework is displayed in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cdiv id=\"Sec5\" class=\"Section3\"\u003e \u003ch2\u003e2.2.1. Data preparation and Pre- processing\u003c/h2\u003e \u003cp\u003eData preparation steps were performed offline in MATLAB (version R2023b) environment. First, we decided to consider further MER signals recorded within an EDT range [-5 mm: +2 mm] to improve dataset balance between classes (i.e., within and outside the STN). The selected raw signal underwent three artifact detection procedures: i) visual inspection by and expert (EXP); ii) an unsupervised variance-based algorithm (COV), and iii) an algorithm based on thresholds estimated on the signal noise characteristics (BCK). Additionally, we considered the RAW dataset, unprocessed for artefacts, for comparative purposes. The EXP approach consisted of an expert (V. L.) visually screening all the traces to identify evident artifacts (e.g., electromagnetic interference, mechanical electrode shift). A trace marked with artifacts was fully rejected. The remaining traces comprise the dataset \u0026lsquo;EXP\u0026rsquo;.\u003c/p\u003e \u003cp\u003eThe second artifact detection approach was based on the algorithm presented in [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e, \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e] for the identification of stationary signal segments. This was preferred over other approaches to exploit its unsupervised nature. The algorithm implementation steps were the following:\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eSegmentation of the signal \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:x\\left(n\\right)\\)\u003c/span\u003e\u003c/span\u003e into m 0.5-second segments \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\{\\:{x}_{k};\\:k\\:=\\:1:m\\};\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eCompute the autocorrelation of each segment\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\:\\{R\\left({x}_{k}\\right);\\:k\\:=\\:1:m\\}\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eCompute the variance of the transformed segment \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\{{v}_{k}=var(R\\left({x}_{k}\\right);\\:k\\:=\\:1:m\\}\\)\u003c/span\u003e\u003c/span\u003e;\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eComparison of the variances of neighboring segments by computing their distance as \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{d}_{kl}=\\frac{\\text{m}\\text{a}\\text{x}({v}_{k},{v}_{l})}{\\text{m}\\text{i}\\text{n}({v}_{k},{v}_{l})},\\:with\\:k=\\:1:m-1;l=k+1\\)\u003c/span\u003e\u003c/span\u003e;\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eCreation of a distance matrix \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:D\\)\u003c/span\u003e\u003c/span\u003e of all possible distances between segment pairs.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eThe matrix elements \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{d}_{kl}\\)\u003c/span\u003e\u003c/span\u003e exceeding an experimentally identified threshold (Th\u0026thinsp;=\u0026thinsp;1.8) are replaced with ones, other with zeros and an adjacency matrix \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:A\\)\u003c/span\u003e\u003c/span\u003e is obtained.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eThe resulting matrix is scanned for the longest uninterrupted segment (sequence of zeros) using a greedy algorithm.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003cp\u003eThe algorithm identifies segments that are different from the rest of the signals, marks them as artifacts, and removes them from the analyzed signal. The last approach identifies artifacts signal segments based on specific amplitude and frequency criteria. Therefore, two assessment steps are performed:\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eAmplitude check: By exploiting the definition of background noise proposed by [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e, \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e], the amplitude back-ground noise level is computed for the whole segment (\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:BCKtot\\)\u003c/span\u003e\u003c/span\u003e). Then the estimation is repeated for each 0.5-second segments (\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{BCK}_{k}\\)\u003c/span\u003e\u003c/span\u003e) separately and compared to the noise level of the complete signal. If \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{BCK}_{k}\u0026gt;20*BCKtot\\)\u003c/span\u003e\u003c/span\u003e, the k-th segment is marked as artifact.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eFrequency check: the maximum amplitude of the Fourier Transform is computed for each 0.5-second segment (\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{m}\\text{a}\\text{x}(FFT}_{k})\\)\u003c/span\u003e\u003c/span\u003e). If for the k-th segment the \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{m}\\text{a}\\text{x}(FFT}_{k})\\)\u003c/span\u003e\u003c/span\u003e is higher than 2.5 times the median across all the segments in the signal, the epoch is marked as artifact.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003cp\u003eIn the current study, both the thresholds were experimentally set, as suggested in the related literature. Signal segments marked as noisy according to one of the two criteria were removed from the signal. After applying the two automatic artifact rejection approaches to the raw dataset, in both cases, only traces with a residual length of at least 4 seconds[\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e] were further analyzed leading to the definition of the COV dataset and the BCK dataset respectively in the first and second case.\u003c/p\u003e \u003cp\u003eFurther pre-processing steps were common for the four obtained datasets, including the selection of unique recordings distinguished by depth, electrode, and hemisphere, that is, if multiple traces shared the same depth, electrode, and hemisphere, the latest in time was selected as the one with sufficient quality satisfying the surgical equipe during the DBS exploration process. Signals were finally band-passed between 300 and 3000 Hz using a second-order elliptic (zero-phase) filter.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section3\"\u003e \u003ch2\u003e2.2.2. Feature extraction\u003c/h2\u003e \u003cp\u003eIn line with our previous works [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e, \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e] and the literature [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e, \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e], 22 features were extracted from each MER trace, belonging to the time and frequency domain. Given the possible different lengths of the traces, when opportune, features were normalized with respect to the number of samples N composing the analyzed signal \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:x\\left(n\\right);n=1:N\\)\u003c/span\u003e\u003c/span\u003e. The list of the extracted features and the relative definitions are reported in Table I. Such features have been proposed in several previous literature studies [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e, \u003cspan additionalcitationids=\"CR18\" citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]. Frequency domain features were extracted from the power spectral density (PSD) computed for each trace using the Welch method, employing 1-second windows (50% overlap), resulting in a spectral resolution of 1Hz. The PSD for frequency bands below 300Hz was derived from the mean-subtracted absolute value of the signal since the original data were bandpass filtered within 300-3000Hz [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]. Subsequently, power bands were extracted from these PSD values.\u003c/p\u003e \u003cp\u003e \u003cem\u003eTable I List of extracted features and their formulation\u003c/em\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Taba\" border=\"1\"\u003e \u003ccolgroup cols=\"3\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFeature\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDefinition\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eWL- Wave or Curve length\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRepresents the unfolded waveform\u0026rsquo;s distance\u003c/p\u003e \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:WL=\\frac{1}{N}\\sum\\:_{n=1}^{N-1}\\left|x\\left(n\\right)+x\\left(n+1\\right)\\right|\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eZC - Zero crossing\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eThe number of times the signal crosses the threshold calculated by estimating the noise level of the signal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003ePKS - Peaks\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eNumber of positive peaks identified in a signal segment normalized for the segment length.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eMAV - Mean value of the absolute amplitude\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:MAV=\\:\\frac{1}{N}\\sum\\:_{n=1}^{N}\\left|x\\left(n\\right)\\right|\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eMED - Median value of absolute amplitude.\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMiddle value separating the greater and lower halves of the ordered absolute amplitude \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\left|x\\left(n\\right)\\right|\\)\u003c/span\u003e\u003c/span\u003e of the trace\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eTH - Signal threshold\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:TH=\\:\\frac{3}{N-1}\\sqrt{\\sum\\:_{n=1}^{N}{\\left(x\\left(n\\right)-\\stackrel{-}{x}\\right)}^{2}}\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eRoot mean square (RMS) of the signal\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:RMS=\\sqrt{\\:\\frac{1}{N}\\sum\\:_{n=1}^{N}{x\\left(n\\right)}^{2}}\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eAKUR - Amplitude distribution kurtosis\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:AKUR=\\:\\frac{1}{(N-1){\\sigma\\:}^{4}}\\sum\\:_{n=1}^{N}\\left|x\\left(n\\right)-{\\stackrel{-}{x}}^{4}\\right|\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eASKW - Amplitude distribution skewness\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:ASKW=\\:\\frac{1}{(N-1){\\sigma\\:}^{3}}\\sum\\:_{n=1}^{N}\\left|x\\left(n\\right)-{\\stackrel{-}{x}}^{3}\\right|\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eNL - Noise level\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDerived from the signal\u0026rsquo;s analytic envelope\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e[\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e, \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003ePWRA- Averaged Power\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:PWRA=\\:\\frac{1}{N}\\sum\\:_{n=1}^{N}{x\\left(n\\right)}^{2}\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eANE - Average non-linear energy\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:ANE=\\frac{1}{N-2}\\sum\\:_{n=2}^{N-1}{\\left[x\\right(n)}^{2}-x\\left(n-1\\right)x(n+1)]\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e[\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003epowVHFrel_1\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRelative power in the 300-1000Hz frequency range\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003epowVHFrel_2\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRelative power in the 1000-2000Hz frequency range\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003epowVHFrel_3\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRelative power in the 2000-3000Hz frequency range\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003epowHFrel_1\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRelative power in the 70-220Hz frequency range\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003epowHFrel_2\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRelative power in the 220-320Hz frequency range\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003epowLFrel_1\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRelative power in the 1-4Hz frequency range\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003epowLFrel_2\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRelative power in the 4-8Hz frequency range\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003epowLFrel_3\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRelative power in the 8-13Hz frequency range\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003epowLFrel_4\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRelative power in the 13-30Hz frequency range\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003epowLFrel_5\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRelative power in the 30-70Hz frequency range\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section3\"\u003e \u003ch2\u003e2.2.3. Outlier detection and management\u003c/h2\u003e \u003cp\u003eOnce features were extracted for all the prepared datasets, they were analyzed to assess their distributions and the presence of outliers. The impact of different treatment methods on the presence of outliers was considered another processing step worth investigating. Thus, we decided to test three different approaches:\u003c/p\u003e \u003cp\u003e \u003col\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eNONE - The first simple possibility is to not apply any outlier detection.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eOutlier Rejection for Hemisphere (ORH) of each patient set, that is the classic approach based on feature distribution to remove samples according to lower and upper bound (inter quartile range - IQR) identification with tolerance of 3 [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e].\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eOutlier Rejection Model (ORM) based on machine learning methodologies, i.e. the local outlier factor algorithm (LOF), an unsupervised based algorithm which computes the local density deviation of a given data point with respect to its neighbors, applied on single patient\u0026rsquo;s data [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e].\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003c/ol\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section3\"\u003e \u003ch2\u003e2.2.4. Dataset Normalization\u003c/h2\u003e \u003cp\u003eAs the last preparation step, we considered the possibility of normalizing the extracted features to reduce patients\u0026rsquo; variability. Indeed, we applied feature standardization to all the generated datasets based on single patient cerebral hemisphere through Min-Max scaler [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eAs a result, a total of 24 processing pipelines were defined, each of them generating a dataset with different characteristics.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003e2.3. Classification models\u003c/h2\u003e \u003cp\u003eThe impact of the different pipelines described above was evaluated through the performance of supervised machine learning (ML) classification models developed in Python (version 3.11.0) using the Scikit-learn library [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e]. Specifically, we employed and compared three binary classifiers: a Support Vector Machine Classification (SVC) model, an Elastic Net (EN) model, and a Random Forest (RF) model. We selected model hyperparameters a priori based on domain knowledge and default library recommendations. Specifically, the SVC model adopted a Radial Basis Function (RBF) kernel, which enables non-linear decision boundaries. Other hyperparameters, including the regularization parameter (C) and kernel coefficient (gamma), were left at default values (C\u0026thinsp;=\u0026thinsp;1.0, gamma= 'scale'). A fixed random state ensured reproducibility. The EN model is a logistic regression classifier employing elastic net regularization to combine L1 and L2 penalties, with an L1 ratio of 0.5 controlling the mix. The SAGA solver was used for efficient optimization. To ensure convergence, the maximum number of iterations was increased to 10000. A fixed random state was also used. The RF model used default hyperparameters. Specifically, the number of estimators was set to 100, Gini criterion was used as a measure of the split quality, the maximum depth of each tree was unrestricted, the minimum number of samples required to split an internal node was 2, and the minimum number of samples needed to be at a leaf node was 1. A fixed random state was used to ensure reproducibility. Strategies were incorporated into the modelling pipeline to address class imbalance, which can adversely affect the performance of classification algorithms. The dataset was partitioned into training and test subsets and preprocessing steps were applied. These included cross-validation and normalization. Cross-validation was employed to reduce overfitting and enhance the generalization ability of the models. Specifically, we applied stratified k-fold cross-validation with shuffling, maintaining the original class distribution within each fold. The dataset was divided into k\u0026thinsp;=\u0026thinsp;5 non-overlapping subsets, each approximately equal in size. Each iteration used one-fold as the test set while the remaining k \u0026minus;\u0026thinsp;1 folds served as the training set. This process was repeated k times, ensuring that every fold was used once as the test set. The final model performance was assessed by averaging the results across all folds. Before training, the features in the training set that did not previously undergo normalization (see paragraph 2.2.4.) were scaled using MinMax normalization, implemented via the `MinMaxScaler` function from the Scikit-learn library (Python). Normalization parameters were derived exclusively from the training set and subsequently applied to the test data. This transformation scaled each feature to the [0, 1] range across the whole dataset. The described classifiers were applied to the full set of features and after an automatic feature selection approach based on Recursive Feature Elimination (RFE) with cross-validation [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e], leading to six classification models applied at the end of each processing pipeline.\u003c/p\u003e \u003cdiv id=\"Sec10\" class=\"Section3\"\u003e \u003ch2\u003e2.3.1. Performance Evaluation\u003c/h2\u003e \u003cp\u003eThe classification of each model was evaluated by computing performance metrics based on the confusion matrix and the Receiver Operating Characteristic (ROC) curve at each classification split. The computed metrics were the area under the ROC curve (AUC), classification accuracy, precision, recall and F1-score, as reported in Eq.\u0026nbsp;(1\u0026ndash;4)\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Tabb\" border=\"1\"\u003e \u003ccolgroup cols=\"2\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:Accuracy=\\frac{True\\:Positive+True\\:Negative}{Total\\:samples}\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003e(1)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:Precision=\\frac{True\\:Positive}{True\\:Positive+False\\:Positive}\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e(2)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:Recall=\\frac{True\\:Positive}{True\\:Positive+False\\:Negative}\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e(3)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:F1score=2*\\frac{Precision*Recall}{Precision+Recall}\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e(4)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eFinally, to better understand the importance of the extracted features for and the impact of their contribution to the classification results, the Shapley additive explanations (SHAP) method was applied and results explored [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e]. Specifically, features were ranked by means of their SHAP value (probability) and we counted how many times each feature was ranked in the top ten positions across the pipelines and folds in order to define the percentage of presence across all the pipelines. The counting was kept separated for the three classifiers.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e"},{"header":"3. Results","content":"\u003cp\u003eIn this section, we describe the results of the compared classification pipelines, starting from the effect of the early preparation steps or dataset pre-processing. These steps, comprising the management of MER signal artefacts, feature extraction, and subsequent outlier management, determine twelve different datasets regarding sample numerosity on which the classification models are trained. Then, a feature normalization procedure is applied to all the resulting datasets, leading to the generation of twelve additional datasets. The classification performance of the 24 tested pipelines, given by the different combinations, is described and commented on. Finally, an exploration of the importance of the features performed through the SHAP approach is also reported.\u003c/p\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003e3.1. Pre-processing results: datasets composition\u003c/h2\u003e \u003cp\u003eTable II reports the number of samples in each dataset specifying the observations for each class (STN and NOT STN). Datasets are the results of the different combinations of early steps procedures that were tested for dataset preparation. Thus, the original 'RAW' dataset is composed of 1228 observations (804 NOT STN, 424 STN); the MER signals are then screened for artefact rejection by an expert (V. L.), and the whole 10-sec segments are removed if marked as 'with artefacts'. This procedure reduced the number of total samples to 1115 in the 'EXP' dataset. The RAW dataset was also processed for artefact rejection using two automatic approaches described in 2.2.1. Both methods did not automatically reject the whole 10-sec signal segment, but only the portion with the artefact, and the remaining segment was further considered if with a length\u0026thinsp;\u0026gt;\u0026thinsp;4 seconds. This approach allowed for less drastic dataset pruning. Indeed, the COV approach determined the rejection of 12.82% of the signal, and, in terms of whole segments, only 11 were removed. As for the BCK method, 12.47% of the signal was rejected, and 21 whole segments were removed. The overall agreement between the two algorithms reached 88.5% with a statistically significant (p\u0026thinsp;\u0026lt;\u0026thinsp;0.05) Cohen's Kappa [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e] with a mean across subjects equal to 0.49 (where (0.2\u0026thinsp;\u0026le;\u0026thinsp;k\u0026thinsp;\u0026le;\u0026thinsp;0.4 is 'Fair agreement', 0.41\u0026thinsp;\u0026le;\u0026thinsp;k\u0026thinsp;\u0026le;\u0026thinsp;0.6 is 'Moderate agreement', and 0.61\u0026thinsp;\u0026le;\u0026thinsp;k\u0026thinsp;\u0026le;\u0026thinsp;0.8 is 'Substantial agreement'). The different formulation of the rejection criteria justifies this moderate agreement. Specifically, COV identified more 'short' artefacts, leading to fewer entirely rejected segments concerning BCK. After this first step, all the features described were extracted from the remaining MER segments. At this point, the dataset was tested for outliers applying two approaches, ORH and ORM, and the effect of this additional procedure was tested against the possibility of not controlling for the presence of outliers (none). Indeed, applying the ORH methods determined the exclusion of 20.1% of samples from the RAW dataset, 19.9% from EXP, 15.5% from COV and 13.5% from the BCK dataset. Applying the ORM method led to the following dataset reductions: RAW\u0026thinsp;=\u0026thinsp;10.9%, EXP\u0026thinsp;=\u0026thinsp;11.0%, COV\u0026thinsp;=\u0026thinsp;10.8%, and BCK\u0026thinsp;=\u0026thinsp;10.9%.\u003c/p\u003e \u003cp\u003e \u003cem\u003eTable II Numerosity of each dataset according to the preparation pipeline. The total number of samples and the number of samples in each class (STN and NOT STN) are shown.\u003c/em\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Tabc\" border=\"1\"\u003e \u003ccolgroup cols=\"10\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c9\" colnum=\"9\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c10\" colnum=\"10\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOutlier treatment\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colspan=\"3\" nameend=\"c4\" namest=\"c2\"\u003e \u003cp\u003enone\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colspan=\"3\" nameend=\"c7\" namest=\"c5\"\u003e \u003cp\u003eORH\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colspan=\"3\" nameend=\"c10\" namest=\"c8\"\u003e \u003cp\u003eORM\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eDataset\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cem\u003etotal\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u003cem\u003eNOT STN\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u003cem\u003eSTN\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cem\u003etotal\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u003cem\u003eNOT STN\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e\u003cem\u003eSTN\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e\u003cem\u003etotal\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e\u003cem\u003eNOT STN\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e\u003cem\u003eSTN\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eRAW\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1228\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e804\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e424\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e981\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e633\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e348\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e1094\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e757\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e337\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eEXP\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1115\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e726\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e389\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e893\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e582\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e311\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e992\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e697\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e295\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eCOV\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1217\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e794\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e423\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1028\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e693\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e335\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e1085\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e775\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e310\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eBCK\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1207\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e793\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e414\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1044\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e699\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e345\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e1076\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e769\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e307\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eFinally, as the last possible dataset preparation step, the hemisphere-based features normalization was applied to all the already mentioned combinations, leading to 24 prepared datasets.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003e3.2. Effect of processing pipelines on performance evaluation\u003c/h2\u003e \u003cp\u003eThe described classifiers, with and without RFE feature selection step, led to six classification models being applied to the 24 datasets. Performance metrics are reported as mean and standard deviation values computed across the five cross-folds. Figures\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, \u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e and \u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e show the performance metrics of Accuracy, F1-score, AUC, precision and recall values across all the pipelines directly comparing the six classification models. Overall, the most critical step seems to be the application of the feature normalization based on data from the same hemisphere of each patient. This step has the aim to prevent model overfitting, but in our case, it reduces the performance of all the classifiers on all possible datasets. Indeed, accuracy values pass from a range (min-max across all the pipelines) of 0.91\u0026ndash;0.945 for the not pre-normalized datasets, to an accuracy range of 0.846\u0026ndash;0.88 when the hemisphere-based normalization is applied. Similarly, the F1-score range of values is reduced from 0.867\u0026ndash;0.922 to 0.742\u0026ndash;0.835 across the twelve pipelines applied to non\u0026ndash;normalized and normalized datasets. As for the effect of the other applied procedures, it can be observed that the artefacts management approaches alone provide a weak performance improvement, particularly for the EN and RF classification models. Table III reports these specific results for the not normalized dataset as the mean and standard deviation of Accuracy and F1-score values. A further slight improvement due to the ORH procedure was obtained for the RF classification model on COV treated dataset (Accuracy\u0026thinsp;=\u0026thinsp;0.945 (0.029), F1-score\u0026thinsp;=\u0026thinsp;0.915 (0.044)), and the SVC on the RAW dataset (Accuracy\u0026thinsp;=\u0026thinsp;0.943 (0.015), F1-score\u0026thinsp;=\u0026thinsp;0.92 (0.02)). ORM, however, appears to have no discernible impact on the results.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003cem\u003eTable III: Mean and standard deviation of accuracy and F1-score values obtained applying the three classifiers considering the full set of features on datasets only processed with the artefact rejection approaches.\u003c/em\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Tabd\" border=\"1\"\u003e \u003ccolgroup cols=\"7\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colspan=\"2\" nameend=\"c3\" namest=\"c2\"\u003e \u003cp\u003e\u003cem\u003eEN\u003c/em\u003e\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colspan=\"2\" nameend=\"c5\" namest=\"c4\"\u003e \u003cp\u003e\u003cem\u003eRF\u003c/em\u003e\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colspan=\"2\" nameend=\"c7\" namest=\"c6\"\u003e \u003cp\u003e\u003cem\u003eSVC\u003c/em\u003e\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cem\u003eACC\u003c/em\u003e\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u003cem\u003eF1-score\u003c/em\u003e\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u003cem\u003eACC\u003c/em\u003e\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cem\u003eF1-score\u003c/em\u003e\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u003cem\u003eACC\u003c/em\u003e\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003e\u003cem\u003eF1-score\u003c/em\u003e\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eRAW\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.912 (0.025)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.876 (0.032)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.932 (0.013)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.899 (0.019)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.931 (0.021)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.902 (0.028)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eEXP\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.916 (0.016)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.881 (0.023)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.936 (0.015)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.906 (0.022)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.931 (0.02)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.902 (0.028)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eCOV\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.929 (0.009)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.899 (0.013)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.938 (0.015)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.909 (0.022)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.93 (0.011)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.902 (0.015)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eBCK\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.926 (0.013)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.893 (0.016)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.934 (0.017)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.903 (0.025)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.932 (0.016)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.904 (0.021)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eSimilar trends are observed for the AUC metrics represented in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e. Looking at more detailed metrics, such as the precision and recall shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e, it is possible to appreciate the different behavior of the classification models adopted that is not clearly visible considering the previous three metrics. Specifically, while EN and SVC, both with or without the automatic feature selection, performed similarly with a good balance between precision and recall, the RF classifier showed a higher precision and a lower recall across all the datasets, both with and without feature selection.\u003c/p\u003e \u003cp\u003eAs for applying a feature selection algorithm before the classification, we did not find a particular performance improvement since in most cases only a few features were not considered, while in some cases no features were removed (maximum number of removed features\u0026thinsp;=\u0026thinsp;8). This result is probably due to the already quite small proposed set of independent features, since a preliminary analysis has been performed based on literature and previous experience to avoid the extraction of highly correlated features [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e].\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003e3.3. Analysis of feature importance\u003c/h2\u003e \u003cp\u003eThis section presents an in-depth analysis of the importance of features across different classification models. In particular, the aim was to assess the robustness of the feature importance in function of the dataset preparations by the quantification of their contribution to the STN classification using the SHAP approach. Specifically, features were ranked by means of their SHAP value (based on probability), and we counted how many times each feature was ranked in the first ten positions across the pipelines (five folds each). Based on this count, we defined the percentage of presence across all the pipelines and considering the not normalized and the normalized features separately. The counting was kept separated for the three classifiers, since results were model-specific.\u003c/p\u003e \u003cp\u003eFigure 5A shows the percentage of presence for each feature in the top-ranked positions, meaning that they contributed to the classification with high importance. First of all, it is possible to notice that some features seem to not contribute to classification of the STN. This observation is coherent across classifiers and with low impact due to the application of the dataset normalization. These low-ranked features are (percentage of presence\u0026thinsp;\u0026lt;\u0026thinsp;15%): ZC, AKUR, ASKW, powLFre\u003csub\u003e2\u0026thinsp;\u0026minus;\u0026thinsp;5\u003c/sub\u003e, powVHFrel\u003csub\u003e1\u003c/sub\u003e. On the other hand, some features are among the top-ranked with good repeatability (percentage of presence\u0026thinsp;\u0026gt;\u0026thinsp;50%), such as RMS, powHFrel\u003csub\u003e1\u0026thinsp;\u0026minus;\u0026thinsp;2\u003c/sub\u003e, powVHFrel\u003csub\u003e2\u0026thinsp;\u0026minus;\u0026thinsp;3\u003c/sub\u003e, MAV, MED, PKS and WL.\u003c/p\u003e \u003cp\u003e \u003cem\u003eInterestingly, RF showed different feature importance attributes with respect to SVC and EN classifiers that were, instead, quite similar to each other. Moreover, it is possible to observe an effect due to the dataset normalization. Specifically, PKS, RMS and TH seem to reduce their importance for classification, while WL, NL and powHFrel\u003c/em\u003e \u003csub\u003e \u003cem\u003e2\u003c/em\u003e \u003c/sub\u003e \u003cem\u003eincreased their contribution after the hemisphere-based normalization was applied. Figure\u0026nbsp;5B shows an example of a SHAP representation obtained when the pipeline reaches the highest accuracy (COV-ORH). From this representation, it is also possible to interpret how the features impact the result. For example, a high value of RMS and MAV is associated with the Fig.\u0026nbsp;5A) Features percentage of presence in the top ranked set of important features across all the pipelines, the pipeline without normalization and with normalization for the EN model (red), RF model (green) and SVC model (blue). B) SHAP Beeswarm plot for the models applied to the pipeline reaching the highest accuracy (COV-ORH).\u003c/em\u003e\u003c/p\u003e \u003cp\u003eSTN (class 1) while a high value of PKS and powVHFrel\u003csub\u003e3\u003c/sub\u003e indicates class 0 (NOT STN).\u003c/p\u003e \u003cp\u003eTo confirm further the repeatability of feature contributions to the classification results, the same approach was repeated to analyze the features importance for the models trained on a selected set of features (of variable number given the automatic approach employed). Figure\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e6\u003c/span\u003e reports the results in terms of percentage of presence in the top 10 ranked features. It is possible to confirm the previous results with an overall agreement and with few changes in the rankings.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"4. Discussion","content":"\u003cp\u003eAccurate detection of the STN as the target for DBS electrode implantation is fundamental to efficient neuromodulation treatment and sufficient improvement of symptoms in patients with Parkinson's disease. To this end, imaging-guided stereotactic trajectories are carefully planned for electrode implantation. However, to cope with potential brain shifts and discriminate between close brain structures, explorative MER signals are routinely recorded to provide surgeons with additional functional information. Indeed, information from MER can be digitally extracted and analyzed, leading to the exploitation of machine learning (ML) and deep learning (DL) approaches widely applied in the literature for target detection and to provide support for the medical team [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e, \u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e]. In this context, much attention has been paid to improving the accuracy of feature-based ML and DL algorithms by increasing the complexity of the methods [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e, \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]. However, few studies report in detail how the signals have been pre-processed and how the dataset has been managed. In very rare cases, the impact of these procedures has also been discussed [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e, \u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e]. Therefore, the goal of this study was to analyze the impact of different MER data processing and management pipelines on the classification of the STN using feature-based ML. To this end, we implemented an evaluation framework to assess the impact of different choices for managing specific crucial procedural steps. Combined, these choices produced a total of 24 pipelines. Three ML models were then used for binary classification, both with and without a further feature selection step. Finally, through the application of SHAP analysis, we explored the impact of the different pipelines on feature importance and model explainability. Specifically, we first analyzed the impact of different artefacts rejection procedures to clean MER signals before feature extraction. Our results suggest that slight improvements in accuracy can be observed (see Table III), and that different artefact rejection approaches (i.e. EXP, COV [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e, \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e] and BCK [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e, \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]) can similarly improve MER classification compared to not removing them at all. Interestingly, the effect of this varies depending on the final classification model applied: the SVC model produced the same result for all four datasets. Following features extraction, three outlier management procedures were considered and their effect analyzed. Results suggested that the classical ORH approach generally increased both accuracy and F1-score when applied to three out of four datasets (i.e., RAW, EXP and COV), while no improvement was observed if applied to BCK dataset. Instead, the ORM approach did not improve performance, and, in some cases, lower accuracy was obtained compared to the possibility of not removing outliers. These considerations also apply to pipelines, including the hemisphere-based normalization step, albeit to a lesser extent. As last manipulation, indeed, we performed a standardization procedure based on data from single hemisphere as suggested in [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eWhile the obvious aim of this method is to improve generalizability and avoid overfitting, in our case it led to a general performance decrease, suggesting that the intrinsic normalization applied within model training may be enough for single center MER classification study. Our final analysis regarded the explainability of features contribution to classification outcomes, by ranking their SHAP values across pipelines and validation folds. From this evaluation, it was possible to observe that some features (i.e., RMS, powHFrel\u003csub\u003e1\u0026thinsp;\u0026minus;\u0026thinsp;2\u003c/sub\u003e, powVHFrel\u003csub\u003e2\u0026thinsp;\u0026minus;\u0026thinsp;3\u003c/sub\u003e, MAV, MED, PKS and WL) greatly contribute to MER classification accuracy, consistently across pipelines and classification models, while some others (i.e., ZC, AKUR, ASKW, powLFre\u003csub\u003e2\u0026thinsp;\u0026minus;\u0026thinsp;5\u003c/sub\u003e, powVHFrel\u003csub\u003e1\u003c/sub\u003e) resulted in a not significantly contribution to STN detection. Interestingly, if we compare feature contribution between normalized and not normalized datasets, slightly different importance among the top ranked features can be observed, while few changes are present among the less contributing ones. These considerations are still valid considering results obtained after feature selection. In summary, in line with previous literature [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e, \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e, \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e, \u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e, \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e], our study demonstrated that, besides their simplicity, feature-based ML approaches are effective in classifying MER traces for STN detection with good accuracy (best result: COV-ORH Accuracy\u0026thinsp;=\u0026thinsp;0.945 (0.029)). Even so, pre-processing and data preparation approaches should be carefully selected and applied since they may differently impact classification performance, both depending on the original dataset and the classification model applied.\u003c/p\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003e4.1. Limitations\u003c/h2\u003e \u003cp\u003eAlthough our study provides an in-depth and valuable analysis of the impact of processing pipelines on STN identification using MER traces and ML approaches, it is affected by some limitations that may have influenced results interpretation. A first limitation regards the artefact detection approaches. In our study, in fact, we only evaluated the effects in terms of dataset reduction (Table II) and final impact on classification performance, while we could not perform an accurate analysis since systematic and precise labeling of artefacts present in our dataset was not performed. Indeed, the EXP dataset resulted from a manual dataset pruning in the presence of evident artefacts (i.e., the whole 10-second segment is discarded without identification of artefact timing and type).\u003c/p\u003e \u003cp\u003eFurther study, in line with [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e, \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e] should be performed for a complete characterization of potential artefact affecting MER recordings. Second, in our study data from a single center (28 Parkinsonian patients) have been included, limiting methods and results generalizability, which is a fundamental step to further improve both ML and DL actual applicability in clinical practice [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. Furthermore, it should be noted that the classification models were implemented using predefined and default hyperparameter settings to avoid introducing factors that could impact STN detection performance, as our main aim was to analyse the effect of different data preparation pipelines. Indeed, our reported accuracy aligns with the literature implementing similar approaches [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e].\u003c/p\u003e \u003c/div\u003e"},{"header":"5. Conclusion","content":"\u003cp\u003eThe presented study provides a comprehensive comparison of MER processing and dataset management procedures in different pipelines for the functional identification of the STN. The work focused on examining and describing how different choices might affect classification performance. It also provided a potential framework for comparison and useful guidelines for designing ML algorithms for STN localization based on MER traces. Specifically, results pointed out the need of properly identifying and rejecting artefacts paired with appropriate outlier management. The highest accuracy values were obtained for the COV-ORH pipeline, while a pre-normalization of features based on data from single patient and brain hemisphere led to performance degradation in our study. Finally, we used the SHAP approach to further explore feature importance, which is a fundamental step that could guide and improve the implementation of future algorithms.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003eGiven that MER data are routinely acquired during standard surgical procedures, when we started the study we had a EC waiver. We were asked to make patients sign an informed consent regarding both the surgery procedure and data exploitation for research aims. Of course, the research was conducted in accordance with the principles embodied in the Declaration of Helsinki and in accordance with local statutory requirements.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eKremer NI, van Laar T, Lange SF, Statius Muller S, la, Bastide-van Gemert S, Oterdoom DM et al (2023) STN-DBS electrode placement accuracy and motor improvement in Parkinson\u0026rsquo;s disease: systematic review and individual patient meta-analysis. J Neurol Neurosurg Psychiatry. ;94:236\u0026ndash;44\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003evan den Munckhof P, Bot M, Schuurman PR (2021) Targeting of the Subthalamic Nucleus in Patients with Parkinson\u0026rsquo;s Disease Undergoing Deep Brain Stimulation Surgery. Neurol Ther 10:61\u0026ndash;73\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVinke RS, Geerlings M, Selvaraj AK, Georgiev D, Bloem BR, Esselink RAJ et al (2022) The Role of Microelectrode Recording in Deep Brain Stimulation Surgery for Parkinson\u0026rsquo;s Disease: A Systematic Review and Meta-Analysis. J Parkinsons Dis 12:2059\u0026ndash;2069\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZakharov N, Belova E, Gamaleya A, Tomskiy A, Sedov A (2024) Neuronal activity features of the subthalamic nucleus associated with optimal deep brain stimulation electrode insertion path in Parkinson\u0026rsquo;s disease. Eur J Neurosci\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWan KR, Maszczyk T, See AAQ, Dauwels J, King NKK (2019) A review on microelectrode recording selection of features for machine learning in deep brain stimulation surgery for Parkinson\u0026rsquo;s disease. Clin Neurophysiol 130:145\u0026ndash;154\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eInggas MAM, Coyne T, Taira T, Karsten JA, Patel U, Kataria S et al (2024) Machine learning for the localization of Subthalamic Nucleus during deep brain stimulation surgery: a systematic review and Meta-analysis. Neurosurg Rev 47:774\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCoelli S, Levi V, Del Vecchio V, Mailland J, Rinaldo E, Eleopra S et al (2020) R,. Characterization of Microelectrode Recordings for the Subthalamic Nucleus identification in Parkinson\u0026rsquo;s disease. 42nd Annual International Conference of the IEEE Engineering in Medicine \u0026amp; Biology Society (EMBC). IEEE; 2020. pp. 3485\u0026ndash;8\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCoelli S, Levi V, Del Vecchio V, Mailland J, Rinaldo E, Eleopra S (2021) An intra-operative feature-based classification of microelectrode recordings to support the subthalamic nucleus functional identification during deep brain stimulation surgery. J Neural Eng 18:016003\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHosny M, Zhu M, Gao W, Fu Y (2020) A novel deep LSTM network for artifacts detection in microelectrode recordings. Biocybern Biomed Eng 40:1052\u0026ndash;1063\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMaged A, Zhu M, Gao W, Hosny M (2024) Lightweight deep learning model for automated STN localization using MER in Parkinson\u0026rsquo;s disease. Biomed Signal Process Control. ;96\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMartin T, Jannin P, Baxter JSH (2024) Generalisation capabilities of machine-learning algorithms for the detection of the subthalamic nucleus in micro-electrode recordings. Int J Comput Assist Radiol Surg\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGorlini C, Forzanini F, Coelli S, Rinaldo S, Eleopra R, Bianchi AM et al (2024) Impact of Microelectrode Recording Artefacts on Subthalamic Nucleus Functional Identification via Features-Based Machine Learning Classifiers. 2024 IEEE International Conference on Metrology for eXtended Reality, Artificial Intelligence and Neural Engineering (MetroXRAINE). IEEE; pp. 13\u0026ndash;8\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBakštein E, Sieger T, Wild J, Nov\u0026aacute;k D, Schneider J, Vostatek P et al (2017) Methods for automatic detection of artifacts in microelectrode recordings. J Neurosci Methods 290:39\u0026ndash;51\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAboy M, Falkenberg JH (2006) An Automatic Algorithm for Stationary Segmentation of Extracellular Microelectrode Recordings. Med Biol Eng Comput 44:511\u0026ndash;515\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCagnan H, Dolan K, He X, Contarino MF, Schuurman R, van den Munckhof P et al (2011) Automatic subthalamic nucleus detection from microelectrode recordings based on noise level and neuronal activity. J Neural Eng [Internet]. ;8:046006. Available from: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://iopscience.iop.org/article/10.1088/1741-2560/8/4/046006\u003c/span\u003e\u003cspan address=\"https://iopscience.iop.article/10.1088/1741-2560/8/4/046006\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDolan K, Martens HCF, Schuurman PR, Bour LJ (2009) Automatic noise-level detection for extra-cellular micro-electrode recordings. Med Biol Eng Comput 47:791\u0026ndash;800\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBellino GM, Schiaffino L, Battisti M, Guerrero J, Rosado-Mu\u0026ntilde;oz A (2019) Optimization of the KNN supervised classification algorithm as a support tool for the implantation of deep brain stimulators in patients with Parkinson\u0026rsquo;S Disease. Entropy. ;21\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBenouis M, Rosado-Mu\u0026ntilde;oz A (2024) Using Ensemble of Hand-Feature Engineering and Machine Learning Classifiers for Refining the Subthalamic Nucleus Location from Micro-Electrode Recordings in Parkinson\u0026rsquo;s Disease. Appl Sci (Switzerland). ;14\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCao L, Li J, Zhou Y, Liu Y, Liu H (2020) Automatic feature group combination selection method based on GA for the functional regions clustering in DBS. Comput Methods Programs Biomed 183:105091\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMoran A, Bar-Gad I (2010) Revealing neuronal functional organization through the relation between multi-scale oscillatory extracellular signals. J Neurosci Methods 186:116\u0026ndash;129\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWong S, Baltuch GH, Jaggi JL, Danish SF (2009) Functional localization and visualization of the subthalamic nucleus from microelectrode recordings acquired during DBS surgery with unsupervised machine learning. J Neural Eng 6:026006\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSmiti A (2020) A critical overview of outlier detection methods. Comput Sci Rev 38:100306\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBreunig MM, Kriegel H-P, Ng RT, Sander J (2000) LOF. Proceedings of the 2000 ACM SIGMOD international conference on Management of data. New York, NY, USA: ACM; pp. 93\u0026ndash;104\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePedregosa FABIANPEDREGOSAF, Michel V, Grisel OLIVIERGRISELO, Blondel M, Prettenhofer P, Weiss R et al (2011) Scikit-learn: Machine Learning in Python Ga\u0026euml;l Varoquaux Bertrand Thirion Vincent Dubourg Alexandre Passos PEDREGOSA, VAROQUAUX, GRAMFORT ET AL. Matthieu Perrot [Internet]. Journal of Machine Learning Research. Available from: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttp://scikit-learn.sourceforge.net\u003c/span\u003e\u003cspan address=\"http://scikit-learn.sourceforge.net\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGuyon I, Weston J, Barnhill S, Vapnik V (2002) Gene Selection for Cancer Classification using Support Vector Machines. Mach Learn 46:389\u0026ndash;422\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePonce-Bobadilla AV, Schmitt V, Maier CS, Mensing S, Stodtmann S (2024) Practical guide to SHAP analysis: Explaining supervised machine learning model predictions in drug development. Clin Transl Sci. ;17\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShapley L (1953) 7. A Value for n-Person Games. Contributions to the Theory of Games II 307\u0026ndash;317. Classics in Game Theory. Princeton University Press; 1997. pp. 69\u0026ndash;79\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChao-Chia Lu D, Boulay C, Chan ADC, Sachs AJ (2024) A Systematic Review of Neurophysiology-Based Localization Techniques Used in Deep Brain Stimulation Surgery of the Subthalamic Nucleus. Neuromodulation: Technology at the Neural Interface. ;27:409\u0026ndash;21\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVarga I, Bakstein E, Gilmore G, May J, Novak D (2024) Statistical segmentation model for accurate electrode positioning in Parkinson\u0026rsquo;s deep brain stimulation based on clinical low-resolution image data and electrophysiology. PLoS ONE 19:e0298320\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRajpurohit V, Danish SF, Hargreaves EL, Wong S (2015) Optimizing computational feature sets for subthalamic nucleus localization in DBS surgery with feature selection. Clin Neurophysiol 126:975\u0026ndash;982\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKhosravi M, Atashzar SF, Gilmore G, Jog MS, Patel RV (2020) Intraoperative Localization of STN during DBS Surgery Using a Data-Driven Model. IEEE J Transl Eng Health Med. ;8\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"Fondazione IRCCS Istituto Neurologico Carlo Besta","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Deep Brain Stimulation (DBS), Microelectrode recording (MER), Subthalamic Nucleus (STN), machine learning, feature-based","lastPublishedDoi":"10.21203/rs.3.rs-6915773/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6915773/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003eBackground:\u003c/strong\u003e Micro-electrode recording (MER) is one of the modalities used to confirm pre-operative planning during Deep Brain Stimulation (DBS) surgery of the subthalamic nucleus (STN) for the symptomatic treatment of Parkinson’s Disease. \u0026nbsp;MER signals have been widely used in combination with machine learning (ML) techniques to improve STN functional localization. However, the impact of data processing and preparation has mostly been overlooked.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMethods:\u003c/strong\u003e A total of twenty-four combinations of processing approaches have been implemented with the aim of exploring the impact of data processing pipelines on the performance of feature-based ML classifiers. \u0026nbsp;These comprise four signal artefact treatments, three outlier management procedures, and an option to standardize or not the feature sets. The effects of the implemented pipeline on the classification results were evaluated by training and testing three classifiers, both with and without feature selection. A final fundamental step to explore the feature importance using SHAP approach has also been implemented.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResults:\u003c/strong\u003e Improvements in performance metrics have been noticed after implementing approaches to artefact rejection and optimal outlier management, while the preliminary features standardization based on single patient and brain hemisphere data reduce all the performance metrics (accuracy, F1-score, recall, precision and area under the curve (AUC)). Interestingly, feature importance analysis through SHAP approach highlighted a good agreement between features contributing to classification across most of the implemented pipelines.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConclusions:\u003c/strong\u003e Proper identification and rejection of artefacts combined with appropriate outlier management are crucial steps during MER processing pipelines for STN identification, while pre-normalization of features based on data from single patient and brain hemisphere may lead to overall performance degradation. In addition, the SHAP approach may represent an adjunctive useful tool to guide and improve the implementation of future algorithms.\u003c/p\u003e","manuscriptTitle":"The role of MER processing pipelines for STN functional identification during DBS surgery: a feature based machine learning approach.","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-06-24 14:53:07","doi":"10.21203/rs.3.rs-6915773/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"f7a7e078-3b70-40c4-84fd-e2aa39aee167","owner":[],"postedDate":"June 24th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":50196034,"name":"Biotechnology and Bioengineering"}],"tags":[],"updatedAt":"2025-06-24T14:53:07+00:00","versionOfRecord":[],"versionCreatedAt":"2025-06-24 14:53:07","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-6915773","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6915773","identity":"rs-6915773","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}