SPSidharth Pancholi

Sidharth Pancholi

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article

Previous Organizations / Institutes

The organizations / institutes from the scholar's other articles in the past ten years


Sidharth Pancholi has a total of 24 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • NeuroKinect: A Deep Learning Framework for Hand Kinematics Reconstruction from EEG Signals

    by Sidharth Pancholi, Amita Giri

    Brain-Computer Interface (BCI) technology facilitates direct brain-to-device communication, empowering individuals to control their surroundings through brain signals. However, prevalent BCI methods encounter challenges: a) time-intensive preprocessing, b) improper loss function use, and c) non-intuitive hyperparameter settings. To overcome these challenges, we propose NeuroKinect, a novel deep learning framework designed to reconstruct hand kinematics from electroencephalography (EEG) signals. Trained on Grasp and Lift (GAL) task data with streamlined preprocessing, NeuroKinect enhances computational efficiency. NeuroKinect introduces a task-specific loss function, LStat, which balances correlation and error metrics to better align predicted and actual motion. Driven by scientific intuition, we analyze motor task dynamics via event-related potentials and brain source localization to optimize hyperparameters. Extensive evaluation on EEG data collected during grasp-and-lift tasks demonstrates NeuroKinect’s ability to accurately decode 3D hand movement trajectories. The model achieves strong correlation between predicted and actual motion across all spatial dimensions (mean Pearson coefficients: 0.92±0.015 for X, 0.93±0.019 for Y, and 0.83±0.018 for Z dimensions), supported by low mean squared error values (MSE: 0.016±0.001 for X, 0.015±0.002 for Y, and 0.017±0.005 for Z). These results highlight NeuroKinect’s potential for enhancing real-time BCI systems for motor control and rehabilitation.

  • Open Access

    SURVEY ON DEEP LEARNING AND EXPLAINABLE AI FOR MAGNETIC RESONANCE IMAGING BASED ALZHEIMER'S DISEASE CLASSIFICATION

    by Nikhileswara Rao Sulake, Sidharth Pancholi

    Alzheimer's disease (AD), a progressive neurodegenerative disorder, demands early detection to mitigate irreversible cognitive decline. Structural magnetic resonance imaging (MRI) serves as a critical diagnostic tool, capturing 3D brain volumes to identify atrophy in the hippocampus, cortex, and ventricles-key biomarkers of AD. Traditional methods relying on manual MRI interpretation or invasive biomarkers like CSF analysis lack sensitivity and scalability for early-stage diagnosis. Recent advancements in artificial intelligence (AI), particularly 3D convolutional neural networks (CNNs) and vision transformers (ViTs), automate AD detection by extracting subtle pathological patterns from MRI data. Hybrid models integrating CNNs with attention mechanisms achieve over 95% accuracy in classifying AD, mild cognitive impairment (MCI), and cognitively normal subjects, while unsupervised frameworks address data scarcity through anomaly detection. However, challenges such as patient-level data leakage, computational complexity, and ethical concerns around AI interpretability persist. Future research must prioritize multimodal integration (e.g., genetic, cognitive, and gait data), lightweight architectures for clinical deployment, and rigorous validation across diverse cohorts. This review synthesizes advancements in AI-driven MRI analysis, underscoring its potential to transform AD diagnostics into a precise, accessible, and patient-centric framework.

  • Explainable AI-Guided Optimization of EMG Channels and Features for Precise Hand Gesture Classification: A SHAP-Based Study

    by Parul Rani, Sidharth Pancholi, Vikash Shaw, Suraj Pandey, Manfredo Atzori, Sanjeev Kumar

    Extraction of the correct and efficient descriptors of muscular activity plays a vital role in tackling the challenging problem of myoelectric control of powered prostheses. This work presents a feature extraction framework that aims to enhance the representation of muscular activities by increasing the amount of information that can be extracted from individual and combined electromyogram (EMG) channels. The proposed method for feature selection is based on Shapley Additive explanations (SHAP). The SHAP value is used to reduce the feature dimension. The proposed approach has been evaluated on two datasets obtained at a sampling rate of 1 kHz through a band consisting of seven EMG channels. The Standard deviation (SD) and Integrated EMG (IEMG) of electrodes 3, 5, 6, and 7 recognized four motions with a classification accuracy of 98.42%± 1.16% and six gestures with a classification accuracy of 96.6%± 0.91%, respectively. In the present work, an ensemble technique called bagging in the random forest algorithm has been used to obtain the optimum results.


  • Enhanced EMG-Based Hand Gesture Classification in Real-World Scenarios: Mitigating Dynamic Factors With Tempo-Spatial Wavelet Transform and Deep Learning

    by Parul Rani, Sidharth Pancholi, Vikash Shaw, Manfredo Atzori, Sanjeev Kumar

    Dynamic factors, like limb position changes and electrode shifting, significantly impact the performance of EMG-based hand gesture classification as the transition is made from a laboratory-controlled environment to real-life scenarios. Traditionally, researchers have employed conventional wavelet transform methods to improve classification performance. This study compares a tempo-spatial technique that utilizes the wavelet multiresolution method and compares it with the conventional wavelet transform using eight machine learning algorithms. Two public datasets are utilized. DB1 comprising ideal conditions with a range of limb positions, while DB2 incorporates dynamic factors like electrode shifting and muscle fatigue. The training/testing involves two cases: one using single-position data and other with multiple positions. Results demonstrate that the Deep Neural Network (DNN) classifier outperforms others in classification accuracy. Proposed technique achieves mean accuracies of 84.07% (DB1) and 68.15% (DB2), while conventional wavelet transform methods achieve 79.39% (DB1) and 53.48% (DB2) for single-position DNN training. For multiple positions, particularly two limb positions, the proposed technique achieves mean accuracies of 94.43% (DB1) and 73.79% (DB2), compared to conventional wavelet transform, which achieves 84.38% (DB1) and 51.98% (DB2) with DNN. Paired t-tests (p-value<0.001) show significant improvement over conventional wavelet transformation, indicating the proposed technique’s potential to enhance gesture classification in real-world scenarios.


  • Open Access

    Novel Time Domain Based Upper-Limb Prosthesis Control using Incremental Learning Approach

    by Sidharth Pancholi, Amit M. Joshi Deepak Joshi, Bradly S. Duerstock

    The upper limb of the body is a vital for various kind of activities for human. The complete or partial loss of the upper limb would lead to a significant impact on daily activities of the amputees. EMG carries important information of human physique which helps to decode the various functionalities of human arm. EMG signal based bionics and prosthesis have gained huge research attention over the past decade. Conventional EMG-PR based prosthesis struggles to give accurate performance due to off-line training used and incapability to compensate for electrode position shift and change in arm position. This work proposes online training and incremental learning based system for upper limb prosthetic application. This system consists of ADS1298 as AFE (analog front end) and a 32 bit arm cortex-m4 processor for DSP (digital signal processing). The system has been tested for both intact and amputated subjects. Time derivative moment based features have been implemented and utilized for effective pattern classification. Initially, system have been trained for four classes using the on-line training process later on the number of classes have been incremented on user demand till eleven, and system performance has been evaluated. The system yielded a completion rate of 100% for healthy and amputated subjects when four motions have been considered. Further 94.33% and 92% completion rate have been showcased by the system when the number of classes increased to eleven for healthy and amputees respectively. The motion efficacy test is also evaluated for all the subjects. The highest efficacy rate of 91.23% and 88.64% are observed for intact and amputated subjects respectively.

  • Enhancing Gesture Classification Using Active EMG Band and Advanced Feature Extraction Technique

    by Parul Rani, Sidharth Pancholi, Vikash Shaw, Manfredo Atzori, Sanjeev Kumar

    This study addresses two critical aspects of gesture classification: data acquisition and feature extraction. It introduces an efficient surface electromyogram (sEMG) acquisition band with seven active electrodes and conducts experiments with eight healthy subjects for six gestures. A novel feature extraction method, the Hjorth secant line (HSL), is also introduced. The self-generated dataset (DB1) and two publicly available datasets (DB2 and DB3), containing data from both healthy subjects and amputees, are analyzed to assess the proposed feature set’s performance. The study evaluates the sEMG acquisition system’s effectiveness by measuring the signal-to-noise ratio (SNR), and the performance of the proposed feature set was evaluated in terms of gesture classification accuracy. The feature set’s performance is compared to other existing feature sets in the literature. Additionally, a time complexity analysis is performed for the proposed feature set. The SNR for the sEMG acquisition system is 49.36±5.50 dB, demonstrating its efficiency. When using the proposed feature set with a Random Forest (RF) classifier, the study achieves classification accuracy of 97.94%±0.47%, 98.94%±0.46%, and 82.16%±2.07% for DB1, DB2, and DB3, respectively. Paired t-tests indicate that the proposed feature set significantly improves gesture classification accuracy across all three datasets (p-value < 0.05). The time required to calculate the proposed features was 0.207 msec, which is shorter than the computational time for other feature extraction methods reported in the literature. This study highlights the enhanced SNR from the proposed acquisition system and the feature set’s potential to improve gesture classification. These findings have substantial implications for advancing intelligent prosthetics.


  • Optimizing Bio-Sensor Design With Support Vector Regression Technique for AlGaN/GaN MOS–HEMT

    by Ashish Kumar, Arathy Varghese, Dheeraj Kalra, Sidharth Pancholi, Gaurav Kumar Sharma

    This study introduces a novel approach using Support Vector Regression (SVR) for sensitivity modeling of GaN MOS-HEMTs (Metal Oxide Semiconductor- High Electron Mobility Transistors). By combining experimental and simulation results, the SVR-based model is developed to predict sensitivities. The fabricated AlGaN/GaN HEMTs incorporate a graded transition scheme, a 1nm AlN spacer, 2nm GaN cap layer, and 10nm Al2O3 as the gate dielectric/sensing layer. To train the model, feature matrices are prepared using pH sensing results from 32 device dimensional variants. The trained model is then used to predict sensitivities for other device dimensions, allowing for device design optimization and exploration of the design space. Among the five considered kernels (Linear, Cubic, Fine Gaussian, Medium Gaussian, and Coarse Gaussian), the quadratic kernel-based SVR demonstrates the best performance, yielding a root mean square (RMS) error of 0.1767 and a standard deviation of 0.0654.


  • Open Access

    Advancing Brain-Computer Interface System Performance in Hand Trajectory Estimation with NeuroKinect

    by Sidharth Pancholi, Amita Giri

    Brain-computer interface (BCI) technology enables direct communication between the brain and external devices, allowing individuals to control their environment using brain signals. However, existing BCI approaches face three critical challenges that hinder their practicality and effectiveness: a) time-consuming preprocessing algorithms, b) inappropriate loss function utilization, and c) less intuitive hyperparameter settings. To address these limitations, we present NeuroKinect, an innovative deep-learning model for accurate reconstruction of hand kinematics using electroencephalography (EEG) signals. NeuroKinect model is trained on the Grasp and Lift (GAL) tasks data with minimal preprocessing pipelines, subsequently improving the computational efficiency. A notable improvement introduced by NeuroKinect is the utilization of a novel loss function, denoted as LStat. This loss function addresses the discrepancy between correlation and mean square error in hand kinematics prediction. Furthermore, our study emphasizes the scientific intuition behind parameter selection to enhance accuracy. We analyze the spatial and temporal dynamics of the motor movement task by employing event-related potential and brain source localization (BSL) results. This approach provides valuable insights into the optimal parameter selection, improving the overall performance and accuracy of the NeuroKinect model. Our model demonstrates strong correlations between predicted and actual hand movements, with mean Pearson correlation coefficients of 0.92 (±0.015), 0.93 (±0.019), and 0.83 (±0.018) for the X, Y, and Z dimensions. The precision of NeuroKinect is evidenced by low mean squared errors (MSE) of 0.016 (±0.001), 0.015 (±0.002), and 0.017 (±0.005) for the X, Y, and Z dimensions, respectively. Overall, the results demonstrate unprecedented accuracy and real-time translation capability, making NeuroKinect a significant advancement in the field of BCI for predicting hand kinematics from brain signals.

  • Open Access

    Active Surface Electromyogram (sEMG) Electrodes band with Enhanced Features based technique for Hand Gesture Classification

    by Parul Rani, Sidharth Pancholi, Vikash Shaw, Manfredo Atzori, Sanjeev Kumar

    The proposed EMG data acquisition band, which utilizes seven electrodes, has the potential to revolutionize gesture recognition. Unlike conventional EMG data acquisition methods that necessitate skin preparation and can cause skin irritation or adverse reactions, this band is inexpensive, user-friendly, and capable of capturing high-quality EMG signals. Our study evaluated the performance of the proposed EMG band by comparing it with established techniques on eight healthy subjects, and the results revealed that it achieved comparable or significantly better performance in terms of classification accuracy. This suggests that our EMG data acquisition band has significant practical applications, particularly in prosthetic hand control, where real-time processing is critical. Moreover, The proposed feature extraction method based on spectral moments also exhibited significant improvements in terms of classification accuracy, class separability, and processing time, making it a promising alternative for gesture recognition applications. Additionally, the proposed feature set was computationally efficient, with a maximum computation time of 250.209 milliseconds, making it suitable for real-time prosthetic hand control.

  • DLPR: Deep Learning-Based Enhanced Pattern Recognition Frame-Work for Improved Myoelectric Prosthesis Control

    by Sidharth Pancholi, Amit M. Joshi, Deepak Joshi

    In EMG based pattern recognition (EMG-PR), deep learning-based techniques have become more prominent for their self-regulating capability to extract discriminant features from large data-sets. Moreover, the performance of traditional machine learning-based methods show limitation to categorize over a certain number of classes and degrades over a period of time. In this paper, an accurate, robust, and fast convolutional neural network-based framework for EMG pattern identification is presented. To assess the performance of the proposed system, five publicly available and benchmark data-sets of upper limb activities were used. This data-set contains 49 to 52 upper limb motions (NinaPro DB1, NinaPro DB2, and NinaPro DB3), Data with force variation, and data with arm position variation for intact and amputated subjects. The classification accuracies of 92.18% (53 classes), 91.56% (49 classes), 84.98% (49 classes of amputees), 95.67% (6 classes with force variation), and 99.12% (8 classes with arm position variation) have been observed during the testing. The performance of the proposed system is compared with the state of art techniques in the literature. The findings demonstrate that classification accuracy and time complexity have been improved significantly. For signal pre-processing and deep learning techniques, Keras which is a high-level API for TensorFlow is utilised to build profound learning models. The proposed method has been tested on the Intel Core i7 3.5GHz, 7th Gen CPU with 8GB DDR4 RAM.