Machine Learning-Driven Radiomic Profiling of Thalamus-Amygdala Nuclei for Prediction of Postoperative Delirium After STN-DBS in Parkinson's Disease Patients: A Pilot Study.

Machine Learning-Driven Radiomic Profiling of Thalamus-Amygdala Nuclei for Prediction of Postoperative Delirium After STN-DBS in Parkinson’s Disease Patients: A Pilot Study - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Search Log in Primary site navigation Search Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice J Imaging Inform Med . 2025 Aug 13;39(3):2440–2455. doi: 10.1007/s10278-025-01616-1 Machine Learning-Driven Radiomic Profiling of Thalamus-Amygdala Nuclei for Prediction of Postoperative Delirium After STN-DBS in Parkinson’s Disease Patients: A Pilot Study Andrius Radziunas 1 , Gustavas Davidavicius 2 , Karolina Reinyte 2, 3 , Aiste Pranckeviciene 3 , Augustinas Fedaravicius 1, 2 , Vytautas Kucinskas 2 , Ovidijus Laucius 4 , Arimantas Tamasauskas 1, 2 , Vytenis Deltuva 1, 2 , Ausra Saudargiene 2, 3, ✉ Author information Article notes Copyright and License information 1 Department of Neurosurgery, Medical Academy, Lithuanian University of Health Sciences, Eiveniu str. 2, Kaunas, LT-50161 Lithuania 2 Neuroscience Institute, Medical Academy, Lithuanian University of Health Sciences, Eiveniu str. 4, Kaunas, LT-50161 Lithuania 3 Department of Health Psychology, Faculty of Public Health, Medical Academy, Lithuanian University of Health Sciences, Tilzes str. 18, Kaunas, LT-47181 Lithuania 4 Department of Neurology, Medical Academy, Lithuanian University of Health Sciences, Eiveniu str. 2, Kaunas, LT-50161 Lithuania ✉ Corresponding author. Received 2025 Mar 27; Revised 2025 Jun 30; Accepted 2025 Jul 10; Collection date 2026 Jun. © The Author(s) under exclusive licence to Society for Imaging Informatics in Medicine 2025. Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law. PMC Copyright notice PMCID: PMC13230431  PMID: 40804595 Abstract Postoperative delirium is a common complication following sub-thalamic nucleus deep brain stimulation surgery in Parkinson’s disease patients. Postoperative delirium has been shown to prolong hospital stays, harm cognitive function, and negatively impact outcomes. Utilizing radiomics as a predictive tool for identifying patients at risk of delirium is a novel and personalized approach. This pilot study analyzed preoperative T1-weighted and T2-weighted magnetic resonance images from 34 Parkinson’s disease patients, which were used to segment the thalamus, amygdala, and hippocampus, resulting in 10,680 extracted radiomic features. Feature selection using the minimum redundancy maximal relevance method identified the 20 most informative features, which were input into eight different machine learning algorithms. A high predictive accuracy of postoperative delirium was achieved by applying regularized binary logistic regression and linear discriminant analysis and using 10 most informative radiomic features. Regularized logistic regression resulted in 96.97% (±6.20) balanced accuracy, 99.5% (±4.97) sensitivity, 94.43% (±10.70) specificity, and area under the receiver operating characteristic curve of 0.97 (±0.06). Linear discriminant analysis showed 98.42% (±6.57) balanced accuracy, 98.00% (±9.80) sensitivity, 98.83% (±4.63) specificity, and area under the receiver operating characteristic curve of 0.98 (±0.07). The feed-forward neural network also demonstrated strong predictive capacity, achieving 96.17% (±10.40) balanced accuracy, 94.5% (±19.87) sensitivity, 97.83% (±7.87) specificity, and an area under the receiver operating characteristic curve of 0.96 (±0.10). However, when the feature set was extended to 20 features, both logistic regression and linear discriminant analysis showed reduced performance, while the feed-forward neural network achieved the highest predictive accuracy of 99.28% (±2.71), with 100.0% (±0.00) sensitivity, 98.57% (±5.42) specificity, and an area under the receiver operating characteristic curve of 0.99 (±0.03). Selected radiomic features might indicate network dysfunction between thalamic laterodorsal, reuniens medial ventral, and amygdala basal nuclei with hippocampus cornu ammonis 4 in these patients. This finding expands previous research suggesting the importance of the thalamic-hippocampal-amygdala network for postoperative delirium due to alterations in neuronal activity. Supplementary Information The online version contains supplementary material available at 10.1007/s10278-025-01616-1. Keywords: Parkinson’s disease, Deep brain stimulation, Postoperative delirium, Radiomics, Machine learning Introduction Parkinson’s Disease (PD) is a progressive neurodegenerative disorder characterized by motor symptoms, including rigidity, resting tremor, and bradykinesia. It is recognized as the fastest-growing neurological disorder worldwide and a paramount source of disability with increasing incidence in recent years [ 1 – 3 ]. The prevalence of PD doubled from 1990 to 2015, accounting for over six million global cases. This trend, primarily attributed to aging, is projected to sustain its momentum, anticipating a rise to twelve million cases by 2040 [ 4 ]. Subthalamic nucleus deep brain stimulation (STN-DBS) has emerged as an efficacious treatment modality for PD, improving motor function, reducing medication dependency, and enhancing the overall quality of life [ 5 – 7 ]. Nevertheless, DBS is linked with the emergence or exacerbation of psychiatric symptoms across the spectrum [ 8 – 11 ]. Postoperative delirium (POD) is among the most frequent psychiatric complications associated with DBS, reported in up to 42.6% of patients [ 12 , 13 ]. Generally manifesting between 3 to 5 days post-surgery, POD can present as hypoactive, hyperactive, or in mixed forms [ 14 ]. Despite its transitory nature, POD has been shown to prolong hospital stays, inflict a detrimental effect on cognitive function, and negatively impact long-term outcomes [ 15 – 17 ]. Considering the increasing number of patients who could potentially benefit from DBS and the high probability of postoperative psychiatric complications, identifying PD patients at heightened preoperative risk for these complications becomes a pivotal focus for research. Postoperative delirium is detected by 4AT scale [ 18 ]. This method has been extensively validated on patients with neurological disorders by Lees et al. [ 19 ], and Infante et al. [ 20 ]. Postoperative psychiatric complications, such as impulsivity, mood elevation, anxiety, apathy, and cognitive changes, prevalent in PD patients following DBS, potentially arise from disruptions in cortico-striatal network function or changes in dopaminergic medication [ 21 ]. Postoperative delirium, characterized by acute confusion and altered consciousness, has been linked to reductions in brain white and gray matter, suggesting the role of brain morphometry in its development [ 22 ]. Postoperative delirium has been linked to reductions in brain white and gray matter, particularly in the prefrontal cortex, anterior cingulate cortex, and hippocampal regions. Cavallari et al. (2015) [ 23 ] demonstrated that patients who developed postoperative delirium showed pre-existing lower gray matter volumes in the left middle temporal gyrus and right cerebellum. Additionally, reduced white matter integrity in the corpus callosum and internal capsule has been associated with increased delirium risk [ 24 ]. A systematic review further emphasized the involvement of several brain areas in the generation of postoperative delirium, highlighting the significance of brain morphometry data in predicting postoperative neuropsychiatric complications [ 25 ]. Radiomics has emerged as a quantitative approach to medical imaging and employs advanced mathematical analysis to extract textural information from images [ 26 ]. This technique quantifies the spatial distribution of signal intensities and pixel interrelationships, providing additional data that complement existing diagnostic processes. Radiomics has demonstrated potential in various imaging fields to enhance clinical decision-making and improve patient outcomes [ 27 ]. Radiomics, with the application of machine learning (ML) methods using extracted features, can potentially provide additional information that could help in the preoperative identification of patients with elevated risk of psychiatric complications after STN-DBS. Radiomics combined with ML and deep learning (DL) are widely used to extract neuroimaging patterns in PD, aiming at improving the diagnosis and prognosis of PD patients. Radiomic features combined with machine learning methods predicted global motor outcomes of STN-DBS in PD with 96.67% accuracy (AUC 0.98) [ 28 ] and rigidity outcomes with 80 % accuracy (AUC 0.85) [ 29 ]. These features also distinguish PD patients from healthy controls [ 30 , 31 ]. Studies using radiomics achieve high precision in differentiating PD from healthy individuals and other Parkinsonian syndromes. For example, texture and metabolic radiomic characteristics from brain scans, paired with support vector machines (SVM), yield 88–91% accuracy using 18F-fluorodeoxyglucose positron emission tomography [ 32 ] and about 83% accuracy using magnetic resonance image (MRI)-based texture biomarkers [ 33 ]. ML-driven radiomic models can also differentiate PD from atypical Parkinsonian syndromes: routine magnetic resonance imaging radiomic analysis has separated PD from multiple system atrophy (MSA) and progressive supranuclear palsy with up to 92 - 96 % precision [ 33 ]. Radiomic signatures from multi-sequence MRI (e.g., T1-weighted plus susceptibility-weighted imaging) have further demonstrated value in discriminating idiopathic PD from MSA and healthy subjects [ 34 ]. Bu et al. (2023) [ 34 ] successfully employed multi-parametric radiomics combining T1-weighted and susceptibility-weighted imaging to differentiate idiopathic PD from multiple system atrophy, achieving an accuracy of 81.4% (AUC 0.90). Radiomic features from specific brain regions have been used to differentiate PD motor subtypes, specifically tremor-dominant (TD) and postural instability gait difficulty (PIGD), in early disease stages [ 35 ]. The light gradient boosting machine (LGBM) classifier achieved an accuracy of 85 % (AUC 0.94), and the most informative radiomic features came from the thalamus, basal ganglia, hippocampus, amygdala, and substantia nigra using T1-weighted and diffusion tensor imaging (DTI) data [ 35 ]. These approaches show that radiomics can improve early diagnostic precision when combined with ML methods. Recent research has highlighted the predictive value of preoperative brain structural connectivity data in assessing the risk of depression following subthalamic DBS, with distinct connectivity patterns between left prefrontal areas and the DBS electrode associated with worsening depressive symptoms [ 36 ]. Using radiomics as a predictive tool to identify patients at risk of postoperative delirium and other neuropsychiatric complications presents a novel and personalized approach. This technique can potentially optimize treatment outcomes and mitigate psychiatric complications following STN-DBS. By elucidating the interrelationships between brain morphometry, structural connectivity, and neuropsychiatric complications in PD patients undergoing DBS, a comprehensive understanding can be achieved to guide preventive strategies and improve patient outcomes. Our aim in this study is to use radiomics as a predictive tool to identify PD patients at risk for postoperative delirium and other neuropsychiatric complications prior to surgery. By examining the relationships between brain structure and connectivity, we hope to enhance treatment outcomes and develop preventive strategies for those undergoing STN-DBS. Materials and Methods We built a workflow for postoperative STN-DBS delirium prediction in PD patients based on the MRI radiomic features and machine learning models (Fig. 1 ). The methodology can be outlined in the following steps: 1. Pre-operative MRI data acquisition; 2. Image segmentation and radiomic feature extraction; 3. Most informative feature selection using minimum Redundancy Maximum Relevance (mRMR) method, and formation of the feature subsets consisting of 5, 10, and 20 top-ranked features; 4. Prediction of postoperative STN-DBS delirium using eight machine learning methods. Fig. 1. Open in a new tab Workflow illustration for preoperative prediction of STN-DBS delirium in Parkinson’s disease patients. ( 1 ) A) Image data: preoperative MRI images were acquired using the 1.5 T Siemens Avanto scanner. ( 2 ) Image segmentation and radiomic feature extraction: B) amygdala, hippocampus, and thalamus were segmented employing FreeSurfer; 21 segments of the amygdala and hippocampus × 2 hemispheres, 47 segments of the thalamus; C) 10,680 radiomic features were extracted using 3D Slicer. ( 3 ) D) 20 most informative features were selected by the minimum Redundancy Maximum Relevance (mRMR) method; subsets of 5, 10, and 20 features were formed. ( 4 ) Predictive machine learning models: E) regularized binary logistic regression (LR), decision tree classifier (DT), linear discriminant analysis (LDA), naive Bayes classifier (NB), kernel support vector machine (SVM), feed-forward neural network (FNN), one-class support vector machine (OC-SVM), feed-forward neural network-based autoencoder for anomaly detection (FNN-A); F) prediction of postoperative STN-DBS delirium First, preoperative MRI scans of selected PD patients were obtained using a 1.5T Siemens Avanto scanner (Fig. 1 A). For the T1-weighted (T1W) images, we used a 3D magnetization prepared rapid gradient echo (MPRAGE) sequence with TR = 1900 ms, TE = 3.42 ms, TI = 1100 ms, flip angle = 15 ∘ , matrix size = 256 × 256, field of view = 256 × 256 mm 2 , slice thickness = 1.0 mm with no gap, resulting in isotropic voxel dimensions of 1.0 × 1.0 × 1.0 mm 3 . The T2-weighted (T2W) images were acquired using a turbo spin echo sequence with the following parameters: TR = 3200 ms, TE = 85 ms, flip angle = 150 ∘ , echo train length = 17, matrix = 256 × 230, FOV = 230 × 230 mm 2 , slice thickness = 3.0 mm with 10% gap, yielding voxel dimensions of 0.9 × 0.9 × 3.0 mm 3 . Next, the amygdala, hippocampus, and thalamus were segmented using FreeSurfer, resulting in 21 segments for the amygdala and hippocampus (per hemisphere) and 47 segments for the thalamus. Radiomic features were extracted using 3D Slicer, yielding a total of 10,680 features (Fig. 1 B-C). Afterwards, the mRMR method was employed to identify the 20 most informative features. Feature selection was performed once for all models. Subsets of features comprising 5, 10, and 20 features were created for further analysis (Fig. 1 D). Finally, to predict postoperative STN-DBS delirium, eight machine learning models were built: regularized binary logistic regression (LR), decision tree classifier (DT), linear discriminant analysis (LDA), Naive Bayes classifier (NB), kernel support vector machine (SVM), artificial feed-forward neural network (FNN), one-class support vector machine (OC-SVM), and autoencoder-based anomaly detection using a feed-forward artificial neural network (FNN-A) (Fig. 1 E-F). The workflow integrates advanced imaging techniques, radiomic feature selection methods, and machine learning models to enhance predictive accuracy for STN-DBS outcomes. Procedures This single-center prospective observational cohort study was conducted at the Lithuanian University of Health Sciences Hospital following approval by the Kaunas Regional Biomedical Research Ethics Committee (No. BE-2-48). Written informed consent was obtained from all participants. The dataset is available upon reasonable request to the corresponding author, subject to ethical approval and data sharing agreements. The patient selection followed a systematic approach. All PD patients scheduled for routine visits at the Movement Disorders Clinic of the Departments of Neurology at the Lithuanian University of Health Sciences Hospital between April 2019 and August 2021 were screened for potential DBS candidacy. The initial screening was performed by the treating neurologist (O.L.) during regular clinical consultations. The patients were adults, diagnosed with levodopa-responsive idiopathic PD with normal brain MRI scans, showed a mini-mental state examination (MMSE) score greater than 24 points, and did not have untreated or active depression or comorbid psychiatric disorders. To assess levodopa (L-DOPA) responsiveness, the study used the L-DOPA challenge test and defined it as a 30% or more improvement of the Unified Parkinson Disease Rating Scale (UPDRS) motor (III) part scores between the off-medication state and after administration of a 1.5-fold higher than usual L-DOPA dose. The 30% improvement threshold for L-DOPA responsiveness aligns with the Core Assessment Program for Surgical Interventional Therapies in Parkinson’s Disease (CAPSIT-PD) guidelines [ 37 ] and has been validated as a reliable predictor of DBS outcomes in multiple studies [ 38 , 39 ]. We excluded patients with atypical Parkinsonism, dementia, or other current or past psychiatric disorders and clinical comorbidities that precluded DBS implantation. We excluded patients with active psychosis, bipolar disorder, current substance use disorders, or severe anxiety disorders requiring benzodiazepine treatment. However, patients with remitted psychiatric conditions (symptom-free for >12 months) or well-controlled mild depression were not excluded. We also included patients with apathy or mild cognitive symptoms typical of PD, as long as their MMSE score exceeded 24. The PD-related symptoms were assessed by Parkinson‘s Disease Composite Scale (PDCS) pre- and post-operatively. Patients who met preliminary criteria for DBS based on disease duration, motor symptoms, and L-DOPA responsiveness were then referred to our multidisciplinary DBS team. This team consisted of a movement disorders neurologist (O.L.), a functional neurosurgeon (A.R.), and a neuropsychologist (A.P). Each potential candidate underwent a comprehensive evaluation following the CAPSIT-PD protocol, including detailed motor assessment in both ON and OFF medication states, neuropsychological testing, and brain MRI review. The neurosurgeon (A.R.) who performed all procedures was involved in the selection process but did not make unilateral decisions about patient eligibility. This team-based approach ensured standardized selection criteria and reduced potential bias in patient enrollment. The DBS electrode implantation protocol was the same for all patients while under general anesthesia. Preoperative stereotactic MRI scans were taken to guide safe lead placement and ensure accurate targeting within the STN. The target was selected 2 mm from the medial border of STN at the maximal rubral diameter described by Bejjani et al. [ 36 ]. Lead placement was guided solely by stereotactic MRI without intraoperative microelectrode recording (MER). This approach was chosen based on our center’s protocol, which has demonstrated comparable clinical outcomes to MER-guided procedures while reducing operative time and potential complications [ 40 ]. Accurate placement was confirmed by postoperative stereotactic MRI showing all electrodes within 1 mm of the intended target. The neurosurgeon (AR) selected the target for DBS implantation on a case-by-case basis. The medical staff utilized the 4AT [ 18 ] scale to detect postoperative delirium. This scale comprises four elements that assess alertness, mental state (by asking name, birth date, age, and hospital name), attention (by reciting months in reverse order), and fluctuations over time. For every patient with implanted DBS electrodes, a postoperative stereotactical MRI was conducted to confirm the position of the electrodes. The results showed that no participants needed reimplantation, as they were placed within 1 mm of the target. This indicates that the clinical outcomes were not affected by the electrode position in STN. The STN-DBS implantation surgery was performed for 34 PD patients at the Department of Neurosurgery. Seven patients (21%, 5 males and 2 females) developed STN-DBS delirium, and the remaining 27 patients (79%, 11 males and 16 females) had no delirium post-STN-DBS. The median age of the patients was 62 years (interquartile range 57–62.5 years) in the postoperative delirium group, and 60 years (interquartile range 57–64 years) in the postoperative no-delirium group. Demographic characteristics did not differ statistically significantly between the two groups (Table 1 ). Table 1. Demographic data Postoperative No postoperative Significance p STN-DBS delirium STN-DBS delirium Number of patients 7 27 Age 1 62 (57–62.5) years 60 (57–64) years 0.814 Gender 2 Male: 5 (71%) Male: 11 (41%) 0.214 Female: 2 (29%) Female: 16 (59%) PDCS preoperative 3 25 (23–27.5) 22 (18–27.5) 0.348 PDCS postoperative 4 14 (11.5–14.5) 11 (7–18) 0.898 LEDD preoperativee (mg) 5 800 (715–820) 680 (610–800) 0.207 LEDD postoperativef (mg) 6 200 (200–350) 400 (250–500) 0.081 Displacement of the DBS electrode, mm 7 0.3 (0.0–0.55) 0.3 (0.1–0.5) 0.828 Open in a new tab Age, scores of the preoperative Parkinson’s Disease Composite Scale (PDCS) and postoperative PDCS, pre-operative levodopa equivalent daily dose (LEED) and postoperative LEED (mg), DBS electrode displacement (mm) are presented as median and interquartile (1st and 3rd). Abbreviations: STN-DBS , subthalamic nucleus deep brain stimulation; PDCS , Parkinson’s Disease Composite Scale; LEDD , levodopa equivalent daily dose 1 Kruskal-Wallis H test χ 2 (1)= 0.055 2 Fisher’s exact test statistics 3.636 3 Kruskal-Wallis H test χ 2 (1)= 0.882 4 Kruskal-Wallis H test χ 2 (1)= 0.016 5 Kruskal-Wallis H test χ 2 (1)= 1.589 6 Kruskal-Wallis H test χ 2 (1)= 3.038 7 Kruskal-Wallis H test χ 2 (1)= 0.047 The preoperative and postoperative Parkinson’s Disease Composite Scale scores did not show statistically significantly differences between the groups of STN-DBS delirium and no delirium ( p = 0.348 and p = 0.898, respectively). Likewise, no significant differences were observed between the groups in doses of preoperative L-DOPA equivalent daily dose (LEDD) and doses of postoperative LEDD ( p = 0.207 and p = 0.081, respectively). The difference in displacement of the DBS electrode between the groups did not differ statistically significantly ( p = 0.828). Image Segmentation and Radiomic Feature Extraction and Selection Prior to segmentation, all MRI images underwent standardized preprocessing: (1) Bias field correction using the N4 bias field correction algorithm from the Insight Toolkit (N4ITK), (2) Intensity normalization tothe Montreal Neurological Institute 152-subject average template (MNI152) space, (3) Skull-stripping using the high definition brain extraction Tool (HD-BET) [ 41 ], and (4) Resampling to isotropic 1 mm 3 resolution using spline interpolation. We employed the FreeSurfer 7.0 module to extract morphometric data from preoperative T1W and T2W brain MRI images and to segment the subregions of the hippocampus, amygdala, and thalamus [ 42 ] (Fig. 1 A). FreeSurfer 7.2 segmentation was performed using the ’recon-all’ pipeline with default parameters, followed by hippocampal and amygdalar subfield segmentation using dedicated algorithms [ 43 ]. Quality control involved visual inspection of all segmentations by two independent raters (A.R. and V.K.), with manual correction performed for 6/34 cases showing minor boundary errors. For each patient, we performed radiomic feature extraction using the amygdalar-hippocampal, thalamic FreeSurfer segmentation files and applying PyRadiomics, an open-source platform. PyRadiomics v3.0.1 settings included: fixed bin width of 25 HU, isotropic resampling to 2 × 2 × 2 mm 3 using B-spline interpolation, and extraction of features from original images plus wavelet-filtered images (coif1 wavelet, all possible combinations of high and low-pass filters). Gray level discretization used 32 bins to balance feature stability and information preservation. Specifically, we estimated 26 shape features, 19 first-order features, 24 Gray level co-occurrence matrix (GLCM) features, 16 gray level run length matrix (GLRLM) features, 16 gray level size zone matrix (GLSZM) features, 14 gray level dependence matrix (GLDM) features, and 5 neighboring gray tone difference matrix (NGTDM) features. We used the Radiomics extension for 3D Slicer and extracted 120 radiomic features for each segment. We performed analysis of 21 segments of the amygdala and hippocampus in both hemispheres and 47 segments of the thalamus, and the procedure resulted in 10,680 radiomic features (Fig. 1 B-C). The 21 amygdala-hippocampus segments comprised bilateral subdivisions based on the FreeSurfer atlas, while the 47 thalamic segments followed the probabilistic atlas of Krauth et al. (2010) [ 44 ]. We selected 20 most informative features using the mRMR method [ 45 ] and formed three feature subsets that included 5, 10, and 20 top-ranked radiomic features. Machine Learning Methods for Postoperative STN-DBS Delirium Prediction We analyzed the predictive power of selected feature subsets using eight ML models of increasing complexity: regularized binary logistic regression, decision tree classifier, linear discriminant analysis, naive Bayes classifier, kernel support vector machine, feed-forward neural network (FNN), one-class support vector machine, and feed-forward neural network-based autoencoder for anomaly detection (Fig. 1 E-F). The architecture and details of the ML algorithms are provided in Supplement. The dataset comprising 34 patients was evaluated using a repeated stratified K-Fold cross-validation approach with 5 folds and 20 repetitions, to ensure robust model generalization despite the small sample size. This method preserved the proportion of patients with and without STN-DBS delirium across folds. Performance metrics, including balanced accuracy, specificity, sensitivity, and the area under the receiver operating characteristic curve (ROC AUC) and F1 score, were calculated as the average across all folds and repetitions. The performance of all models was compared using non-parametric Friedman and Nemenyi post hoc tests. Statistical significance level of α = 0.05 was chosen. We used the smazzanti mRMR package ( https://github.com/smazzanti/mrmr ), the sci-kit-learn Python package, and the Keras Python framework [ 46 ] to select radiomic features and build machine learning models. Results Top-Ranked Radiomic Features Identified by mRMR Method Segmentation of amygdala-hippocampus and thalamus regions , radiomic feature extraction, and feature selection resulted in 20 most informative features that represented the thalamic laterodorsal nucleus (LD), reuniens nucleus (RE), and nucleus basalis of Meynert (NBM) (Table 5 ). The ranks of the selected 20 most important features are presented in (Table 6 ). Table 5. Radiomic features derived from FreeSurfer’s segmentation of the amygdala, hippocampus, and thalamus, including associated nuclei Nuclei/subfields Amygdala Hippocampus Thalamus Nuclei Radiomic Feature Subdivisions Radiomic Feature Nuclei Radiomic Feature Basal nucleus Shape Major axis length Parasubiculum - Laterodorsal Shape Elongation, Major axis length, Minor axis length Flatness Lateral nucleus – Presubiculum GLSZM SAHGLE Reuniens (medial ventral) GLCM Cluster Prominence Accessory basal nucleus – Subiculum - Central medial Shape Maximum 3D diameter Anterior amygdaloid nucleus – CA1 GLCM MCC, IMC1 Pulvinar lateral GLDM DV, DNN GLCM MCC, IMC1 Pulvinar lateral GLRLM SRLGLE Central nucleus – CA3 – Lateral geniculate GLSZM GLN Medial nucleus – CA4 GLCM IMC1 Medial geniculate GLSZM GLNN Cortical nucleus – GC-ML-DG – Ventral anterior First-order Kurtosis Paralaminar nucleus GLDM SDE Molecular layer – – – LDLGLE, SAHGLE 77/70 Open in a new tab Radiomic features used to predict STN-DBS postoperative delirium. Abbreviations: CA1-4 , Cornu Amonis; DNN , dependence non-uniformity normalized; DV , dependence variance; GLCM , gray level co-occurrence matrix; GLDM , gray level dependence matrix; GLN , gray level non-uniformity; GLNN , gray level non-uniformity normalized; GLRLM , gray level run length matrix; GLSZM , gray level size zone matrix; IMC1 , informational measure of correlation (IMC) 1; LDLGLE , large dependence low gray level emphasis; MCC , maximal correlation coefficient; SAHGLE , small area high gray level emphasis; SDE , small dependence emphasis; SRLGLE , short run low gray level emphasis; STN-DBS , subthalamic nucleus deep brain stimulation Table 6. Ranks of the selected 20 most informative radiomic features used in 8 machine learning (ML) algorithms Feature rank Feature 1 Th - Right-LD laterodorsal Shape Elongation 2 Th - Left-MV(Re) reuniens (medial ventral) GLCM Cluster Prominence 3 Lh - Left CA4 GLCM IMC1 4 Rh - Basal-nucleus Shape Major Axis Length 5 Th - Left-CeM central medial Shape Maximum 3D Diameter 6 Rh - Paralaminar-nucleus GLDM SDE 7 Rh - CA1 GLCM MCC 8 Th - Left-PuL pulvinar lateral GLDM DV 9 Th - Right-PuL pulvinar lateral GLRLM SRLGLE 10 Lh - Presubiculum GLSZM SAHGLE 11 Rh - Paralaminar-nucleus GLSZM SAHGLE 12 Th - Right-LD laterodorsal Shape Major Axis Length 13 Rh - Paralaminar-nucleus GLDM LDLGLE 14 Th - Right-LGN lateral geniculate GLSZM GLN 15 Rh - CA1 GLCM IMC1 16 Th - Right-LD laterodorsal Shape Minor Axis Length 17 Th - Left-PuL pulvinar lateral GLDM DNN 18 Th - Right-MGN medial geniculate GLSZM GLNN 19 Th - Right-VA ventral anterior First-order Kurtosis 20 Th - Right-LD laterodorsal Shape Flatness Open in a new tab Selected radiomic features for STN-DBS postoperative delirium prediction. A 5-feature data set includes features ranked 1–5; a 10-feature data set includes features ranked 1–10; a 20-feature data set includes features ranked 1–20. Abbreviations: CA1-4 , Cornu Amonis; DNN , dependence non-uniformity normalized; DV , dependence variance; GLCM , gray level co-occurrence matrix; GLDM , gray level dependence matrix; GLN , gray level non-uniformity; GLNN , gray level non-uniformity normalized; GLRLM , gray level run length matrix; GLSZM , gray level size zone matrix; IMC1 informational measure of correlation (IMC) 1; LDLGLE , large dependence low gray level emphasis; MCC , maximal correlation coefficient; SAHGLE , small area high gray level emphasis; SDE , small dependence emphasis; SRLGLE , short run low gray level emphasis; STN-DBS , subthalamic nucleus deep brain stimulation Accuracy of the Machine Learning Models for Postoperative STN-DBS Delirium Prediction We trained eight ML algorithms using three data sets that included 5, 10, and 20 most informative features based on their ranking. We present the balanced accuracy, specificity, sensitivity, AUC and F1 score in Tables 2 , 3 , and 4 . Table 2. Balanced accuracy, specificity, sensitivity, AUC, and F1 score of the ML methods applied in post-operative STN-DBS motor outcome prediction using 5 selected features No ML method Balanced accuracy % Sensitivity % Specificity % AUC F1 score 1 LR 96.08 ±6.44 98.50 ±8.53 93.67 ±9.61 0.96 ± 0.06 0.80 ± 0.45 2 DT 75.25 ±19.87 59.50 ±42.83 91.00 ±12.92 0.75 ± 0.20 0.80 ± 0.45 3 LDA 92.90 ±15.43 86.00 ±30.89 99.80 ±1.99 0.93 ± 0.15 0.80 ± 0.45 4 NB 81.15 ±22.58 67.50 ±42.65 94.80 ±11.16 0.81 ± 0.23 0.80 ± 0.45 5 SVM 50.00 ±0.00 0.00 ±0.00 100.00 ±0.00 0.50 ± 0.00 0.80 ± 0.45 6 FNN 81.53 ±21.80 64.50 ±43.76 98.57 ±5.91 0.82 ± 0.22 0.80 ± 0.45 7 OC-SVM 56.60 ±21.83 63.50 ±41.14 49.70 ±18.93 0.57 ± 0.22 0.80 ± 0.45 8 FNN-A 67.90 ±21.63 57.00 ±43.02 78.80 ±8.02 0.68 ± 0.22 0.80 ± 0.45 Open in a new tab Abbreviations: ML , machine learning; AUC , area under ROC; LR , regularized binary logistic regression; DT , decision tree classifier; LDA , linear discriminant analysis; NB , Naive Bayes classifier; SVM , support vector machine; FNN , artificial feed-forward neural network; OC-SVM , one-class support vector machine; FNN-A , feed-forward neural network-based autoencoder Table 3. Balanced accuracy, specificity, sensitivity, AUC, and F1 score of the ML methods applied in post-operative STN-DBS motor outcome prediction using 10 selected features No ML method Balanced accuracy % Sensitivity % Specificity % AUC F1 score 1 LR 96.97 ±6.20 99.50 ±4.97 94.43 ±10.70 0.97 ± 0.06 0.80 ± 0.45 2 DT 77.18 ±22.15 61.50 ±43.51 92.87 ±13.60 0.77 ± 0.22 0.80 ± 0.45 3 LDA 98.42 ±6.57 98.00 ±9.80 98.83 ±4.63 0.98 ± 0.07 0.80 ± 0.45 4 NB 79.08 ±20.75 61.00 ±40.98 97.17 ±7.17 0.79 ± 0.21 0.80 ± 0.45 5 SVM 50.00 ±0.00 0.00 ±0.00 100.00 ±0.00 0.50 ± 0.00 0.80 ± 0.45 6 FNN 96.17 ±10.40 94.50 ±19.87 97.83 ±7.87 0.96 ± 0.10 0.80 ± 0.45 7 OC-SVM 57.35 ±23.26 70.00 ±40.62 44.70 ±24.60 0.57 ± 0.23 0.80 ± 0.45 8 FNN-A 77.62 ±20.32 75.50 ±39.68 79.73 ±7.10 0.78 ± 0.20 0.80 ± 0.45 Open in a new tab Abbreviations: ML , machine learning; AUC , area under ROC; LR , regularized binary logistic regression; DT , decision tree classifier; LDA , linear discriminant analysis; NB , Naive Bayes classifier; SVM , support vector machine; FNN , artificial feed-forward neural network; OC-SVM , one-class support vector machine; FNN-A , feed-forward neural network-based autoencoder Table 4. Balanced accuracy, specificity, sensitivity, AUC, and F1 score of the ML methods applied in post-operative STN-DBS motor outcome prediction using 20 selected features No ML method Balanced accuracy % Sensitivity % Specificity % AUC F1 score 1 LR 91.70 ±15.03 91.50 ±25.55 91.90 ±12.98 0.92 ± 0.15 0.80 ± 0.45 2 DT 70.60 ±21.98 54.50 ±45.33 86.70 ±14.19 0.71 ± 0.22 0.80 ± 0.45 3 LDA 93.32 ±12.24 95.50 ±18.83 91.13 ±13.58 0.93 ± 0.12 0.80 ± 0.45 4 NB 90.08 ±17.16 81.50 ±34.39 98.67 ±4.88 0.90 ± 0.17 0.80 ± 0.45 5 SVM 50.00 ±0.00 0.00 ±0.00 100.00 ±0.00 0.50 ± 0.00 0.80 ± 0.45 6 FNN 99.28 ±2.71 100.00 ±0.00 98.57 ±5.42 0.99 ± 0.03 0.80 ± 0.45 7 OC-SVM 61.17 ±20.29 89.00 ±30.48 33.33 ±22.31 0.61 ± 0.20 0.80 ± 0.45 8 FNN-A 86.88 ±13.76 92.00 ±27.13 81.77 ±5.11 0.87 ± 0.14 0.80 ± 0.45 Open in a new tab Abbreviations: ML , machine learning; AUC , area under ROC; LR , regularized binary logistic regression; DT , decision tree classifier; LDA , linear discriminant analysis; NB , Naive Bayes classifier; SVM , support vector machine; FNN , artificial feed-forward neural network; OC-SVM , one-class support vector machine; FNN-A , feed-forward neural network-based autoencoder The set of 5 most informative radiomic features allowed to achieve high predictive power in postoperative STN-DBS delirium identification: LDA resulted in 92.90% (±15.43) accuracy, 86.00% (±30.89) sensitivity, 99.80% (±1.99) specificity, and AUC of 0.93 (±0.15); LR showed 96.08% (±6.44) accuracy, 98.50% (±8.53) sensitivity, 93.67% (±9.61) specificity, and AUC of 0.96 (0.06) (Table 2 ). The remaining algorithms, such as DT, NB, SVM, FNN, OC-SVM, and FNN-A, had a lower performance with the accuracy varying from 50.00% for SVM to 81.53% for FNN. However, the majority of the methods suffered from large variance, showing high sensitivity in predicting postoperative STN-DBS. A Friedman test, conducted to compare the performance of eight ML models, showed significant differences among the models ( χ 2 ( 7 ) = 373.35 , p < 0.0001 ). Post-hoc analysis using the Nemenyi test revealed that LDA and LR both performed significantly better than the other models (p < 0.05); however, the difference between LDA and LR was not statistically significant, indicating comparable performance between the two. We extended our analysis by including the 10 most informative features in the data set (Table 3 ). The LDA model performed exceptionally well, leading to accuracy of 98.42% (±6.57), sensitivity of 98.00% (±9.80), specificity of 98.83% (±4.63), and AUC of 0.98 (±0.07). LR obtained 96.97% (±6.20) accuracy, 99.50% (±4.97) sensitivity, 94.43% (±10.70) specificity, and AUC of 0.97 (±0.06). The feedforward neural network achieved slightly lower but comparable performance, with 96.17% (±10.40) accuracy, 94.5% (±19.87) sensitivity, 97.83% (±7.87) specificity, and AUC of 0.96 (±0.10) Again, Friedman test indicated significant differences among the models ( χ 2 ( 7 ) = 427.79 , p < 0.0001 ), and the post-hoc Nemenyi test showed that LDA, LR, and FNN all outperformed other models ( p = 0.001), not being statistically significantly different between each other ( p = 0.9) For a full data set of 20 selected features, LDA got an accuracy of 93.32% (±12.24), sensitivity of 95.50% (±18.83), specificity of 91.13% (±13.58), and AUC of 0.93 (±0.12) (Table 4 ), displaying a slight decrease in performance compared to the 10-feature set. LR got similar results of 91.70% (±15.03) accuracy, 91.50% (±25.55) sensitivity, 91.90% (±12.98) specificity, and AUC of 0.92 (±0.15), also showing reduced performance relative to the 10-feature configuration. In contrast, the feedforward neural network achieved the best overall results with 99.28% (±2.71) accuracy, 100.0% (±0.00) sensitivity, 98.57% (±5.42) specificity, and AUC of 0.99 (±0.03), indicating strong generalization when using the full mRMR-selected feature set. The models showed statistically significant differences in performance (Friedman test χ 2 ( 7 ) = 434.13 , p < 0.0001 ), FNN outperformed other models, including LR (post hoc Nemenyj test p = 0.001), but not LDA, showing no significant difference (post hoc Nemenyj test p = 0.06) The highest precision in predicting postoperative STN-DBS delirium was obtained by regularized LR and LDA models for 5 and 10 feature sets, and by FNN for 20 feture set. Increasing the number of features from 5 to 10 led to the improved prediction of all algorithms; however, only the simple models, such as LR, LDA, maintained robust performance with low variance in accuracy, sensitivity, specificity, and the resulting AUC. Further increasing the feature set to 20 slightly degraded the performance of LR and LDA, but increased the performance of more complex models, such as FNN, SVM, and FNN-A. Discussion DBS surgery in PD patients carries some risks for postoperative complications. Psychiatric complications such as depression, suicidality, apathy, fatigue, or personality changes as well as neurocognitive changes, such as decreased memory function, verbal fluency, or increased risk of dementia, may have long-term effects on patients’ functioning and quality of life [ 47 – 49 ]. Other complications are of neuropsychological origin, such as impairments in verbal fluency, memory, executive functions, and dementia [ 49 , 50 ]. Postoperative delirium is a common complication following DBS surgery in PD patients. POD is characterized by acute cognitive impairment, including attention, memory, and perception disturbances. The pathophysiology of POD in PD is complex and multifactorial [ 12 ]. Neurotransmitters such as dopamine and acetylcholine are most important in developing delirium in elderly patients [38]. Moreover, POD after PD DBS surgery is common and reaches up to 42% of operated patients [ 13 ]. A recent review of research studies of PD patients after DBS surgery highlighted the importance of patient age, disease severity, regional brain atrophy, and surgical lesion for developing POD [ 51 ]. Our previous voxel-based morphometry study with PD DBS patients and POD suggested that brain regional atrophy might be responsible for network dysfunction. Employing radiomic and statistical analysis tools lets us analyze our increased patient cohort from another perspective. Results showed that the thalamic laterodorsal nucleus, reuniens nucleus, central medial nucleus (CeM), and amygdalar basal and paralaminar nucleus, and with hippocampal cornu ammonis (CA1-4) might play a role in POD development from a network dysfunction perspective. The prominence of shape-based features (elongation, flatness, major axis length) from the laterodorsal thalamic nucleus suggests structural alterations that may reflect neuronal loss or atrophy predisposing to network dysfunction. The texture features (GLCM Cluster Prominence, GLDM Dependence Variance) from the reuniens nucleus and hippocampal cornu ammonis 4 (CA4) region indicate heterogeneous tissue composition, potentially representing microstructural changes invisible to conventional visual analysis. These findings align with histopathological studies showing that patients with delirium vulnerability exhibit decreased neuronal density and altered cytoarchitecture in thalamic relay nuclei (Trzepacz et al., 2012). In Parkinson’s disease, several brain regions critical for cognition and emotion become dysfunctional. The laterodorsal nucleus of the thalamus (LD), which is important for spatial memory and attention, displays reduced activity, structural changes, and disrupted connections. These issues result from the loss of cholinergic input and the effects of dopamine depletion [ 52 ]. Moreover, the medial ventral nucleus of the thalamus (MV) is associated with emotional processing and memory, and changes in its function lead to emotional and memory disturbances. Meanwhile, the basal nucleus of the amygdala (BA), responsible for processing emotions such as fear and anxiety, also shows dysfunction due to Lewy body pathology and disrupted connections, and may contribute to emotional issues. [ 53 ]. The hippocampus CA4 in PD exhibits synaptic dysfunction and is indirectly impacted by dopamine depletion. While current animal studies do not provide specific details about the laterodorsal nucleus of the thalamus, medial ventral nucleus, basal nucleus of the amygdala or the CA4 region of the hippocampus, they do indicate that widespread neuronal dysfunction beyond the nigrostriatal pathway is a feature of PD, and this is reproduced to some extent in animal models [ 54 ]. Overall, these findings suggest that dysfunction in the LD, MV, BA, and hippocampus CA4 significantly contributes to the cognitive and emotional deficits observed in PD [ 55 ]. This is due to a combination of factors, including loss of cholinergic input, disrupted circuitry, neurotransmitter imbalances, Lewy body pathology, and synaptic dysfunction (Table 5 and 6 ). The thalamus is a key node in the brain’s sensory and cognitive processing network. For now, four parallel circuits are identified: the sensorimotor circuit, the dorsal cognitive circuit, the ventral cognitive circuit, and the affective circuit [ 14 ]. These circuits presumably involve habitual stimulus-response behavior, response inhibition, executive functioning, and affective and reward processing [ 56 ]. The LD and RE nuclei are two subregions of the thalamus that are particularly important for cognitive function [ 57 ]. The cortex receives information from anterior thalamic and LD “higher-order” thalamic nuclei, enabling them to control how different parts of the cortex can coordinate with each other [ 58 ]. The LD is also involved in attention and working memory, while the RE mediates communication between the prefrontal cortex and other brain regions [ 59 ]. Dysfunction of these nuclei has been implicated in various cognitive disorders, including dementia and schizophrenia. The CeM also plays a key role in cognitive function in PD patients [ 60 ]. Cholinergic innervation is provided by this nucleus to many cortical regions and is critical for attention, learning, and memory. PD patients have abnormalities in dopamine, acetylcholine, and 5-hydroxytryptophan concentrations, which make this patient more vulnerable to POD incidence [ 61 ]. Dysfunction of the CeM with its neurotransmitter activation changes in the cortex has been implicated in the development of cognitive disorders, including Alzheimer’s disease and PD dementia [ 60 ]. For decades, it has been accepted that the amygdala is the main hub responsible for emotion control [ 62 ]. Its direct connection with the hippocampus and prefrontal cortex, and amygdala involvement in different circuits may influence changes in cognitive and emotional processes after surgery. The basal nucleus of the amygdala is a major structure in this part of the brain and is a key node of all amygdala [ 63 ]. We have found that basal and paralaminar nuclei had a significant importance in developing POD in PD patients undergoing DBS surgery. Moreover, a recent ML-based research study identified functional specialization of the amygdala-hippocampal circuits, which proves this connection’s importance for confusion seen in delirium [ 64 ]. Based on existing literature, there appears to be a link between delirium and circadian rhythms, which may play a role in developing delirium [ 65 ]. The studies found that a disturbance in core clock genes such as E4bp4, Bmal1, Rev-erb α , Cry1, and Per2 in the mouse hippocampus caused delirium. A recent study with a mouse-delirium model with the deactivation of E4BP4 protein showed that mice lacking in microglial E4bp4 are delirium prone [ 66 ], whereas mice with E4bp4 specifically deleted in hippocampal CA1 neurons had no delirium. These results suggest the importance of hippocampal CA1 microglia and possible radiomic changes, which might be found in our study. We chose conventional algorithms such as regularized binary logistic regression, decision tree classifier, linear discriminant analysis, naive Bayes classifier, kernel support vector machine, and advanced machine learning methods such as artificial neural network, one-class support vector machine, feed-forward neural network-based autoencoder for anomaly detection, to predict postoperative STN-DBS delirium using the sets of selected most informative radiomic features. Our group’s previous study aimed to investigate whether cortical thickness and volume of subcortical structures play a role in neuropsychiatric complications following deep brain stimulation (DBS) in PD patients [ 22 ]. Similarly, we showed that radiomics might be even more helpful in predicting the risk of POD in PD patients undergoing DBS surgery. It has already been demonstrated that radiomics and ML methods can assist in diagnosing PD [ 30 , 31 ], assessing motor outcomes [ 28 ], and identifying the best PD candidates for surgery. Recent studies have combined handcrafted radiomic features with deep features in hybrid models for differential diagnosis of PD Hoehn-Yahr stages [ 67 ]. Jiang et al. (2024) [ 67 ] used deep learning and handcrafted features from baseline SPECT scans and achieved 83.5 % accuracy (AUC 0.87) in distinguishing slow vs. fast PD progression. Additionally, DL-based radiomic methods continue to evolve. Majhi et al. (2024) [ 68 ] proposed a metaheuristic-enhanced deep model that improved PD classification from MRI, achieving accuracies above 90 % . These studies demonstrate the potential of radiomic approaches in capturing subtle morphometric changes that may not be visually apparent. However, the application of radiomics to predict postoperative complications in PD patients undergoing DBS remains unexplored, representing a significant gap in the literature. Our results, showing high prediction accuracy ( 99.94 % ), align well with the performance of machine learning methods reported in other studies of PD and radiomic features. However, radiomic features have not yet been used to predict delirium following DBS. The study strongly suggests its importance and potential for clinical use to predict POD in PD patients in the preoperative timeframe. The main limitations of this study are a small sample of patients that might not sufficiently reflect the diversity of the PD patient population and STN-DBS surgery outcomes. In addition, the limited data set prevented the use of advanced deep learning methods. The study has limitations that are discussed in the following section. Potential bias in feature selection: the mRMR method was used to select the most relevant features, and we recognize that this approach may still introduce bias due to the usage of the full dataset. To mitigate this, we repeated the analysis across three different feature set sizes (20, 10, and 5 top features), allowing us to assess model robustness to dimensionality reduction. Overfitting risk: we now emphasize the risk of overfitting, especially in models showing near-perfect performance. To address this, the revised analysis implements a repeated stratified K-Fold cross-validation scheme (5 folds, 20 repetitions), which ensures that each model is retrained on different data partitions and evaluated in a more generalized manner. Lack of external validation: we acknowledge that no independent dataset was available for external validation. This limitation is now clearly stated. However, the rigorous cross-validation procedure was designed to approximate generalization performance as closely as possible within the constraints of our dataset. The validation in larger cohorts is essential before clinical implementation. Future validation should follow a structured approach: (1) External validation using independent cohorts from multiple centers with varying imaging protocols, (2) Prospective validation incorporating standardized delirium assessment protocols and longer follow-up periods, (3) Integration with clinical risk factors to develop composite prediction models, (4) Investigation of the identified radiomic signatures in relation to neurotransmitter imaging (positron emission tomography) to elucidate underlying mechanisms, and (5) Development of automated software tools for clinical implementation with real-time risk stratification. Conclusion We have shown that ML tools with high accuracy can help predict PD patients with high POD risk after STN-DBS. Selected radiomic features might indicate dysfunction of the thalamic laterodorsal, reuniens medial ventral, and amygdala basal nuclei with hippocampus CA4 in these patients. This finding expands previous research, suggesting the importance of the thalamic-hippocampal-amygdala network for cognitive dysfunction due to alterations in patterns of neuronal activity and neurotransmitter systems. Radiomics may be a helpful tool for predicting the risk of POD in PD patients undergoing DBS surgery, and further research is needed to explore the potential of this approach. Future studies should further explore the thalamic-hippocampal-amygdala network, refine radiomic features, and investigate the role of neurotransmitters, circadian rhythms, and microglial changes in POD. The study also opens opportunities for developing clinical tools based on ML for preoperative risk assessment and for validation in larger patient cohorts. While our results are promising, validation in larger multicenter cohorts is essential before clinical implementation. Materials and Methods Procedures The prospective observational cohort study included PD patients treated at the Departments of Neurosurgery and Neurology, Lithuanian University of Health Sciences Hospital in Kaunas, Lithuania, from April 2019 to August 2021. Supplementary Information Below is the link to the electronic supplementary material. Supplementary file 1 (pdf 6432 KB) (6.3MB, pdf) Acknowledgements We are grateful to the participants in this study who generously and enthusiastically attended study visits. Abbreviations Abbreviation Description AUC Area under the curve CA4 Cornu ammonis 4 CAPSIT-PD Core Assessment Program for Surgical Interventional Therapies in PD CeM Central medial nucleus DT Decision tree classifier DTI Diffusion tensor imaging FNN Feed-forward neural network FNN-A Feed-forward neural network-based autoencoder GLCM Gray level co-occurrence matrix GLDM Gray level dependence matrix GLRLM Gray level run length matrix GLSZM Gray level size zone matrix HD-BET High definition brain extraction tool LD Laterodorsal nucleus L-DOPA Levodopa LDA Linear discriminant analysis LEDD L-DOPA equivalent daily dose LGBM Light gradient boosting machine LR Logistic regression MER Microelectrode recording ML Machine learning MMSE Mini-mental state examination MNI152 Montreal Neurological Institute’s 152-subject average template MPRAGE Magnetization prepared rapid gradient echo MRI Magnetic resonance imaging mRMR Minimum redundancy maximal relevance MSA Multiple system atrophy NB Naive Bayes classifier N4ITK N4 bias field correction algorithm from the Insight Toolkit NBM Nucleus basalis of Meynert NGTDM Neighboring gray tone difference matrix OC-SVM One-class support vector machine PD Parkinson’s disease PDCS Parkinson’s Disease Composite Scale PIGD Postural instability gait difficulty POD Postoperative delirium RE Reuniens nucleus ROC Receiver operating characteristic SPECT Single-photon emission computed tomography STN-DBS Subthalamic nucleus deep brain stimulation SVM Kernel support vector machine T1W T1-weighted T2W T2-weighted TD Tremor-dominant Author Contributions AR: conception, radiomic feature extraction, and writing the first draft. AS: statistical data analysis and machine learning methods, writing methods, results section, introduction and discussion. VK and GD: statistical data analysis and machine learning methods. KR: literature review. OL: neurological patient assessment and critical revision. VK, AP, AF, VD, and AT: review and critical revision. All authors read and approved the final manuscript. Funding This research was funded by the Research Council of Lithuania, grant No. S-SEN-20-15. The authors acknowledge support from the grant “Personalised brain models of Parkinson’s disease patatients. Modèle de cerveau personnalisé pour les patients atteints de la maladie de Parkinson.” Lithuanian-French programme Gilibert, project No 2024-PRO-00148/ P-LZ-24-6, 2025-2026. Availability of Data and Materials The dataset is available upon reasonable request to the corresponding author, subject to ethical approval and data sharing agreements. Declarations Ethical Approval The study protocol was approved by the Kaunas Regional Biomedical Research Ethics Committee (No. BE-2-48). Consent to Participate Informed consent was obtained from all individual participants included in the study. Consent for Publication Not applicable Conflict of Interest The authors declare no competing interests. Footnotes Publisher's Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References 1. Savica R, Grossardt BR, Bower JH, Ahlskog JE, Rocca WA. Time Trends in the Incidence of Parkinson’s Disease: a 30-year Study. JAMA neurology. 2016 Aug;73(8):981. 10.1001/jamaneurol.2016.0947. [ DOI ] [ PMC free article ] [ PubMed ] 2. Willis AW, Roberts E, Beck JC, Fiske B, Ross W, Savica R, et al. Incidence of Parkinson disease in North America. NPJ Parkinson’s disease. 2022 Dec;8(1):170. 10.1038/s41531-022-00410-y. [ DOI ] [ PMC free article ] [ PubMed ] 3. Yang W, Hamilton JL, Kopil C, Beck JC, Tanner CM, Albin RL, et al. Current and projected future economic burden of Parkinson’s disease in the U.S. NPJ Parkinson’s disease. 2020;6:15. 10.1038/s41531-020-0117-1. [ DOI ] [ PMC free article ] [ PubMed ] 4. Dorsey ER, Sherer T, Okun MS, Bloem BR. The Emerging Evidence of the Parkinson Pandemic. Journal of Parkinson’s Disease. 2018 Dec;8(Suppl 1):S3. 10.3233/JPD-181474. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 5. Rughani A, Schwalb JM, Sidiropoulos C, Pilitsis J, Ramirez-Zamora A, Sweet JA, et al. Congress of Neurological Surgeons Systematic Review and Evidence-Based Guideline on Subthalamic Nucleus and Globus Pallidus Internus Deep Brain Stimulation for the Treatment of Patients With Parkinson’s Disease: Executive Summary. Neurosurgery. 2018 Jun;82(6):753–756. 10.1093/neuros/nyy037. [ DOI ] [ PMC free article ] [ PubMed ] 6. Hacker ML, Turchan M, Heusinkveld LE, Currie AD, Millan SH, Molinari AL, et al. Deep brain stimulation in early-stage Parkinson disease: Five-year outcomes. Neurology. 2020 Jul;95(4):e393–e401. 10.1212/WNL.0000000000009946. [ DOI ] [ PMC free article ] [ PubMed ] 7. Limousin P, Foltynie T. Long-term outcomes of deep brain stimulation in Parkinson disease. Nature Reviews Neurology. 2019 Apr;15(4):234–242. 10.1038/s41582-019-0145-9. [ DOI ] [ PubMed ] 8. Thobois S, Ardouin C, Lhommée E, Klinger H, Lagrange C, Xie J, et al. Non-motor dopamine withdrawal syndrome after surgery for Parkinson’s disease: Predictors and underlying mesolimbic denervation. Brain : a journal of neurology. 2010 Mar;133:1111–27. 10.1093/brain/awq032. [ DOI ] [ PubMed ] 9. Somma T, Esposito F, Scala MR, Scelzo A, Baiano C, Patti S, et al. Psychiatric Symptoms in Parkinson’s Disease Patients before and One Year after Subthalamic Nucleus Deep Brain Stimulation Therapy: Role of Lead Positioning and Not of Total Electrical Energy Delivered. Journal of Personalized Medicine. 2022 Oct;12(10):1643. 10.3390/jpm12101643. [ DOI ] [ PMC free article ] [ PubMed ] 10. Accolla EA, Pollo C. Mood Effects After Deep Brain Stimulation for Parkinson’s Disease: An Update. Frontiers in Neurology. 2019 Jun;10. 10.3389/fneur.2019.00617. [ DOI ] [ PMC free article ] [ PubMed ] 11. Nassery A, Palmese CA, Sarva H, Groves M, Miravite J, Kopell BH. Psychiatric and Cognitive Effects of Deep Brain Stimulation for Parkinson’s Disease. Current Neurology and Neuroscience Reports. 2016 Oct;16(10):87. 10.1007/s11910-016-0690-1. [ DOI ] [ PubMed ] 12. Carlson JD, Neumiller JJ, Swain LDW, Mark J, McLeod P, Hirschauer J. Postoperative delirium in Parkinson’s disease patients following deep brain stimulation surgery. Journal of Clinical Neuroscience: Official Journal of the Neurosurgical Society of Australasia. 2014 Jul;21(7):1192–1195. 10.1016/j.jocn.2013.12.007. [ DOI ] [ PubMed ] 13. Tanaka M, Tani N, Maruo T, Oshino S, Hosomi K, Saitoh Y, et al. Risk Factors for Postoperative Delirium After Deep Brain Stimulation Surgery for Parkinson Disease. World Neurosurgery. 2018 Jun;114:e518–e523. 10.1016/j.wneu.2018.03.021. [ DOI ] [ PubMed ] 14. Li H, Han S, Feng J. Delirium after Deep Brain Stimulation in Parkinson’s Disease. Parkinson’s Disease. 2021 Feb;2021:8885386. 10.1155/2021/8885386. [ DOI ] [ PMC free article ] [ PubMed ] 15. Goldberg TE, Chen C, Wang Y, Jung E, Swanson A, Ing C, et al. Association of Delirium With Long-term Cognitive Decline: A Meta-analysis. JAMA neurology. 2020 Nov;77(11):1373–1381. 10.1001/jamaneurol.2020.2273. [ DOI ] [ PMC free article ] [ PubMed ] 16. Umemura A, Oeda T, Tomita S, Hayashi R, Kohsaka M, Park K, et al. Delirium and high fever are associated with subacute motor deterioration in Parkinson disease: a nested case-control study. PloS One. 2014;9(6):e94944. 10.1371/journal.pone.0094944. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 17. Serrano-Dueñas M, Bleda MJ. Delirium in Parkinson’s disease patients. a five-year follow-up study. Parkinsonism & Related Disorders. 2005 Sep;11(6):387–392. 10.1016/j.parkreldis.2005.05.002. [ DOI ] [ PubMed ] 18. Bellelli G, Morandi A, Davis DHJ, Mazzola P, Turco R, Gentile S, et al. Validation of the 4AT, a new instrument for rapid delirium screening: a study in 234 hospitalised older people. Age and Ageing. 2014 Jul;43(4):496–502. 10.1093/ageing/afu021. [ DOI ] [ PMC free article ] [ PubMed ] 19. Lees R, Corbet S, Johnston C, Moffitt E, Shaw G, Quinn TJ. Test accuracy of short screening tests for diagnosis of delirium or cognitive impairment in an acute stroke unit setting. Stroke. 2013 Nov;44(11):3078–3083. 10.1161/STROKEAHA.113.001724. [ DOI ] [ PubMed ] 20. Infante MT, Pardini M, Balestrino M, Finocchi C, Malfatto L, Bellelli G, et al. Delirium in the acute phase after stroke: comparison between methods of detection. Neurological Sciences: Official Journal of the Italian Neurological Society and of the Italian Society of Clinical Neurophysiology. 2017 Jun;38(6):1101–1104. 10.1007/s10072-017-2832-x. [ DOI ] [ PubMed ] 21. Irmen F, Horn A, Mosley P, Perry A, Petry-Schmelzer JN, Dafsari HS, et al. Left Prefrontal Connectivity Links Subthalamic Stimulation with Depressive Symptoms. Annals of Neurology. 2020 Jun;87(6):962–975. 10.1002/ana.25734. [ DOI ] [ PubMed ] 22. Radziunas A, Deltuva VP, Tamasauskas A, Gleizniene R, Pranckeviciene A, Surkiene D, et al. Neuropsychiatric complications and neuroimaging characteristics after deep brain stimulation surgery for Parkinson’s disease. Brain Imaging and Behavior. 2020 Feb;14(1):62–71. 10.1007/s11682-018-9971-4. [ DOI ] [ PubMed ] 23. Cavallari M, Hshieh TT, Guttmann CRG, Ngo LH, Meier DS, Schmitt EM, et al. Brain atrophy and white-matter hyperintensities are not significantly associated with incidence and severity of postoperative delirium in older persons without dementia. Neurobiology of Aging. 2015 Jun;36(6):2122–2129. 10.1016/j.neurobiolaging.2015.02.024. [ DOI ] [ PMC free article ] [ PubMed ] 24. Shioiri A, Kurumaji A, Takeuchi T, Matsuda H, Arai H, Nishikawa T. White matter abnormalities as a risk factor for postoperative delirium revealed by diffusion tensor imaging. The American Journal of Geriatric Psychiatry: Official Journal of the American Association for Geriatric Psychiatry. 2010 Aug;18(8):743–753. 10.1097/JGP.0b013e3181d145c5. [ DOI ] [ PubMed ] 25. Wang F, Lai Y, Pan Y, Li H, Liu Q, Sun B. A systematic review of brain morphometry related to deep brain stimulation outcome in Parkinson’s disease. NPJ Parkinson’s disease. 2022 Oct;8:130. 10.1038/s41531-022-00403-x. [ DOI ] [ PMC free article ] [ PubMed ] 26. Avanzo M, Wei L, Stancanello J, Vallières M, Rao A, Morin O, et al. Machine and deep learning methods for radiomics. Medical Physics. 2020 Jun;47(5):e185–e202. 10.1002/mp.13678. [ DOI ] [ PMC free article ] [ PubMed ] 27. Timmeren JEv, Cester D, Tanadini-Lang S, Alkadhi H, Baessler B. Radiomics in medical imaging—“how-to” guide and critical reflection. Insights into Imaging. 2020 Aug;11:91. 10.1186/s13244-020-00887-2. [ DOI ] [ PMC free article ] [ PubMed ] 28. Saudargiene A, Radziunas A, Dainauskas JJ, Kucinskas V, Vaitkiene P, Pranckeviciene A, et al. Radiomic features of amygdala nuclei and hippocampus subfields help to predict subthalamic deep brain stimulation motor outcomes for Parkinson’s disease patients. Frontiers in Neuroscience. 2022;16:1028996. 10.3389/fnins.2022.1028996. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 29. Liu Y, Xiao B, Zhang C, Li J, Lai Y, Shi F, et al. Predicting Motor Outcome of Subthalamic Nucleus Deep Brain Stimulation for Parkinson’s Disease Using Quantitative Susceptibility Mapping and Radiomics: A Pilot Study. Frontiers in Neuroscience. 2021 Sep;15:731109. 10.3389/fnins.2021.731109. [ DOI ] [ PMC free article ] [ PubMed ] 30. Ren Q, Wang Y, Leng S, Nan X, Zhang B, Shuai X, et al. Substantia Nigra Radiomics Feature Extraction of Parkinson’s Disease Based on Magnitude Images of Susceptibility-Weighted Imaging. Frontiers in Neuroscience. 2021;15:646617. 10.3389/fnins.2021.646617. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 31. Cao X, Wang X, Xue C, Zhang S, Huang Q, Liu W. A Radiomics Approach to Predicting Parkinson’s Disease by Incorporating Whole-Brain Functional Activity and Gray Matter Structure. Frontiers in Neuroscience. 2020 Jul;14. 10.3389/fnins.2020.00751. [ DOI ] [ PMC free article ] [ PubMed ] 32. Wu Y, Jiang JH, Chen L, Lu JY, Ge JJ, Liu FT, et al. Use of radiomic features and support vector machine to distinguish Parkinson’s disease cases from normal controls. Annals of Translational Medicine. 2019 Dec;7(23):773–773. Publisher: AME Publishing Company. 10.21037/atm.2019.11.26. [ DOI ] [ PMC free article ] [ PubMed ] 33. Gore S, Dhole A, Kumbhar S, Jagtap J, Thakare A, Gudadhe S. Radiomics for Parkinson’s disease classification using advanced texture-based biomarkers. MethodsX. 2023 Dec;11:102359. 10.1016/j.mex.2023.102359. [ DOI ] [ PMC free article ] [ PubMed ] 34. Bu S, Pang H, Li X, Zhao M, Wang J, Liu Y, et al. Multi-parametric radiomics of conventional T1 weighted and susceptibility-weighted imaging for differential diagnosis of idiopathic Parkinson’s disease and multiple system atrophy. BMC Medical Imaging. 2023 Dec;23(1):204. 10.1186/s12880-023-01169-1. [ DOI ] [ PMC free article ] [ PubMed ] 35. Panahi M, Hosseini MS. Multi-modality radiomics of conventional T1 weighted and diffusion tensor imaging for differentiating Parkinson’s disease motor subtypes in early-stages. Scientific Reports. 2024 Sep;14:20708. 10.1038/s41598-024-71860-y. [ DOI ] [ PMC free article ] [ PubMed ] 36. Bejjani BP, Dormont D, Pidoux B, Yelnik J, Damier P, Arnulf I, et al. Bilateral subthalamic stimulation for Parkinson’s disease by using three-dimensional stereotactic magnetic resonance imaging and electrophysiological guidance. Journal of Neurosurgery. 2000 Apr;92(4):615–625. 10.3171/jns.2000.92.4.0615. [ DOI ] [ PubMed ] 37. Defer GL, Widner H, Marié RM, Rémy P, Levivier M. Core assessment program for surgical interventional therapies in Parkinson’s disease (CAPSIT-PD). Movement Disorders: Official Journal of the Movement Disorder Society. 1999 Jul;14(4):572–584. 10.1002/1531-8257(199907)14:4<572::aid-mds1005>3.0.co;2-c. [ DOI ] [ PubMed ] 38. Charles PD, Gill CE, Davis TL, Konrad PE, Benabid AL. Is deep brain stimulation neuroprotective if applied early in the course of PD? Nature Clinical Practice Neurology. 2008 Aug;4(8):424–426. Publisher: Nature Publishing Group. 10.1038/ncpneuro0848. [ DOI ] [ PubMed ] 39. Kleiner-Fisman G, Fisman DN, Sime E, Saint-Cyr JA, Lozano AM, Lang AE. Long-term follow up of bilateral deep brain stimulation of the subthalamic nucleus in patients with advanced Parkinson disease. 2003 Sep;Section: Journal of Neurosurgery. 10.3171/jns.2003.99.3.0489. [ DOI ] [ PubMed ] 40. Blasberg F, Wojtecki L, Elben S, Slotty PJ, Vesper J, Schnitzler A, et al. Comparison of Awake vs. Asleep Surgery for Subthalamic Deep Brain Stimulation in Parkinson’s Disease. Neuromodulation: Technology at the Neural Interface. 2018 Aug;21(6):541–547. 10.1111/ner.12766. [ DOI ] [ PubMed ] 41. Isensee F, Schell M, Pflueger I, Brugnara G, Bonekamp D, Neuberger U, et al. Automated brain extraction of multisequence MRI using artificial neural networks. Human Brain Mapping. 2019;40(17):4952–4964. https://onlinelibrary.wiley.com/doi/pdf/10.1002/hbm.24750 . 10.1002/hbm.24750. [ DOI ] [ PMC free article ] [ PubMed ] 42. Fischl B. FreeSurfer. NeuroImage. 2012 Aug;62(2):774–781. 10.1016/j.neuroimage.2012.01.021. [ DOI ] [ PMC free article ] [ PubMed ] 43. Iglesias JE, Augustinack JC, Nguyen K, Player CM, Player A, Wright M, et al. A computational atlas of the hippocampal formation using ex vivo , ultra-high resolution MRI: Application to adaptive segmentation of in vivo MRI. NeuroImage. 2015 Jul;115:117–137. 10.1016/j.neuroimage.2015.04.042. [ DOI ] [ PMC free article ] [ PubMed ] 44. Krauth A, Blanc R, Poveda A, Jeanmonod D, Morel A, Székely G. A mean three-dimensional atlas of the human thalamus: Generation from multiple histological data. NeuroImage. 2010 Feb;49(3):2053–2062. 10.1016/j.neuroimage.2009.10.042. [ DOI ] [ PubMed ] 45. Zhao Z, Anand R, Wang M. Maximum Relevance and Minimum Redundancy Feature Selection Methods for a Marketing Machine Learning Platform; 2019. . 46. Pedregosa F, Varoquaux G, Gramfort A, Michel V, Thirion B, Grisel O, et al. Scikit-learn: Machine Learning in Python. Journal of Machine Learning Research. 2011;12(85):2825–2830. [ Google Scholar ] 47. Cartmill T, Skvarc D, Bittar R, McGillivray J, Berk M, Byrne LK. Deep Brain Stimulation of the Subthalamic Nucleus in Parkinson’s Disease: A Meta-Analysis of Mood Effects. Neuropsychology Review. 2021 Sep;31(3):385–401. 10.1007/s11065-020-09467-z. [ DOI ] [ PubMed ] 48. Giannini G, Francois M, Lhommée E, Polosan M, Schmitt E, Fraix V, et al. Suicide and suicide attempts after subthalamic nucleus stimulation in Parkinson disease. Neurology. 2019 Jul;93(1):e97–e105. 10.1212/wnl.0000000000007665. [ DOI ] [ PubMed ] 49. Lilleeng B, Gjerstad M, Baardsen R, Dalen I, Larsen JP. The long-term development of non-motor problems after STN-DBS. Acta Neurologica Scandinavica. 2015 Oct;132(4):251–258. 10.1111/ane.12391. [ DOI ] [ PubMed ] 50. Bucur M, Papagno C. Deep Brain Stimulation in Parkinson Disease: A Meta-analysis of the Long-term Neuropsychological Outcomes. Neuropsychology Review. 2023 Jun;33(2):307–346. 10.1007/s11065-022-09540-9. [ DOI ] [ PMC free article ] [ PubMed ] 51. Inouye SK, Westendorp RGJ, Saczynski JS. Delirium in elderly people. Lancet (London, England). 2014 Mar;383(9920):911–922. 10.1016/S0140-6736(13)60688-1. [ DOI ] [ PMC free article ] [ PubMed ] 52. Bezdudnaya T, Keller A. Laterodorsal nucleus of the thalamus: A processor of somatosensory inputs. The Journal of Comparative Neurology. 2008 Apr;507(6):1979–1989. 10.1002/cne.21664. [ DOI ] [ PMC free article ] [ PubMed ] 53. Yang Y, Wang JZ. From Structure to Behavior in Basolateral Amygdala-Hippocampus Circuits. Frontiers in Neural Circuits. 2017;11:86. 10.3389/fncir.2017.00086. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 54. Lama J, Buhidma Y, Fletcher EJR, Duty S. Animal models of Parkinson’s disease: a guide to selecting the optimal model for your research. Neuronal Signaling. 2021 Dec;5(4):NS20210026. 10.1042/NS20210026. [ DOI ] [ PMC free article ] [ PubMed ] 55. Pazinatto Aguiar R. The contribution of 5-HT1A receptors in improving plasticity and function of the brain. Maastricht: Maastricht University; 2020. [ Google Scholar ] 56. van den Heuvel OA, van Wingen G, Soriano-Mas C, Alonso P, Chamberlain SR, Nakamae T, et al. Brain circuitry of compulsivity. European Neuropsychopharmacology: The Journal of the European College of Neuropsychopharmacology. 2016;26(5):810–827. 10.1016/j.euroneuro.2015.12.005. [ DOI ] [ PubMed ] [ Google Scholar ] 57. Usrey WM, Sherman SM. Corticofugal circuits: Communication lines from the cortex to the rest of the brain. The Journal of Comparative Neurology. 2019 Feb;527(3):640–650. 10.1002/cne.24423. [ DOI ] [ PMC free article ] [ PubMed ] 58. Perry BAL, Mitchell AS. Considering the Evidence for Anterior and Laterodorsal Thalamic Nuclei as Higher Order Relays to Cortex. Frontiers in Molecular Neuroscience. 2019 Jul;12. 10.3389/fnmol.2019.00167. [ DOI ] [ PMC free article ] [ PubMed ] 59. Wolff M, Vann SD. The Cognitive Thalamus as a Gateway to Mental Representations. The Journal of Neuroscience. 2019 Jan;39(1):3. 10.1523/JNEUROSCI.0479-18.2018. [ DOI ] [ PMC free article ] [ PubMed ] 60. Gratwicke J, Oswal A, Akram H, Jahanshahi M, Hariz M, Zrinzo L, et al. Resting state activity and connectivity of the nucleus basalis of Meynert and globus pallidus in Lewy body dementia and Parkinson’s disease dementia. NeuroImage. 2020 Nov;221:117184. 10.1016/j.neuroimage.2020.117184. [ DOI ] [ PMC free article ] [ PubMed ] 61. Pasina L, Colzani L, Cortesi L, Tettamanti M, Zambon A, Nobili A, et al. Relation Between Delirium and Anticholinergic Drug Burden in a Cohort of Hospitalized Older Patients: An Observational Study. Drugs & Aging. 2019 Jan;36(1):85–91. 10.1007/s40266-018-0612-9. [ DOI ] [ PubMed ] 62. Salzman CD, Fusi S. Emotion, cognition, and mental state representation in amygdala and prefrontal cortex. Annual Review of Neuroscience. 2010;33:173–202. 10.1146/annurev.neuro.051508.135256. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 63. Asami T, Nakamura R, Takaishi M, Yoshida H, Yoshimi A, Whitford TJ, et al. Smaller volumes in the lateral and basal nuclei of the amygdala in patients with panic disorder. PLoS ONE. 2018 Nov;13(11):e0207163. 10.1371/journal.pone.0207163. [ DOI ] [ PMC free article ] [ PubMed ] 64. Li J, Cao D, Yu S, Xiao X, Imbach L, Stieglitz L, et al. Functional specialization and interaction in the amygdala-hippocampus circuit during working memory processing. Nature Communications. 2023 May;14:2921. 10.1038/s41467-023-38571-w. [ DOI ] [ PMC free article ] [ PubMed ] 65. Fitzgerald JM, Adamis D, Trzepacz PT, O’Regan N, Timmons S, Dunne C, et al. Delirium: a disturbance of circadian integrity? Medical Hypotheses. 2013 Oct;81(4):568–576. 10.1016/j.mehy.2013.06.032. [ DOI ] [ PubMed ] 66. Chen M, Zhang L, Shao M, Du J, Xiao Y, Zhang F, et al. E4BP4 Coordinates Circadian Control of Cognition in Delirium. Advanced Science. 2022;9(23):2200559. 10.1002/advs.202200559. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 67. Jiang H, Du Y, Lu Z, Wang B, Zhao Y, Wang R, et al. Radiomics incorporating deep features for predicting Parkinson’s disease in 123I-Ioflupane SPECT. EJNMMI Physics. 2024 Jul;11(1):60. 10.1186/s40658-024-00651-1. [ DOI ] [ PMC free article ] [ PubMed ] 68. Majhi B, Kashyap A, Mohanty SS, Dash S, Mallik S, Li A, et al. An improved method for diagnosis of Parkinson’s disease using deep learning models enhanced with metaheuristic algorithm. BMC Medical Imaging. 2024 Jun;24(1):156. 10.1186/s12880-024-01335-z. [ DOI ] [ PMC free article ] [ PubMed ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Supplementary Materials Supplementary file 1 (pdf 6432 KB) (6.3MB, pdf) Data Availability Statement The dataset is available upon reasonable request to the corresponding author, subject to ethical approval and data sharing agreements. Articles from Journal of Imaging Informatics in Medicine are provided here courtesy of Springer ACTIONS View on publisher site PDF (703.1 KB) Cite Collections Permalink RESOURCES Similar articles Cited by other articles Links to NCBI Databases Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top

Related documents

• Changes of Cerebral Glucose Metabolism After 12 Weeks of Paroxetine Treatment in Panic Disorder#1173204
• GPNMB and glycosphingolipid measurements in cerebrospinal fluid and plasma from Parkinson's disease patients in the BioFind cohort#176512
• A Clinical Trial to Evaluate NB-4746 in Participants With Amyotrophic Lateral Sclerosis.#906365
• Deciphering Effects of Thalidomide on Red Blood Cells in Transfusion Dependents Beta Thalassemia Patients#876550
• Study of AROSOD-1 in Adult Participants With Amyotrophic Lateral Sclerosis (ALS)#850951