Arafat Rahman

University of Virginia

University of Dhaka

CGPA: 3.67 out of 4.00

Objective Motor Function Index for Neuromuscular Disorders

Clinical assessments for neuromuscular disorders, such as Spinal Muscular Atrophy (SMA) and Duchenne Muscular Dystrophy (DMD), continue to rely on subjective measures to monitor treatment response and disease progression. We introduce a novel method using wearable sensors to objectively assess motor function during daily activities in 19 patients with DMD, 9 with SMA, and 13 age-matched controls. Pediatric movement data is complex due to confounding factors such as limb length variations in growing children and variability in movement speed. Our approach uses Shape-based Principal Component Analysis to align movement trajectories and identify distinct kinematic patterns, including variations in motion speed and asymmetry. Both DMD and SMA cohorts have individuals with motor function on par with healthy controls. Notably, patients with SMA showed greater activation of the motion asymmetry pattern. We further combined projections on these principal components with partial least squares (PLS) to identify a covariation mode with a canonical correlation of r = 0.78 (95% CI: [0.34, 0.94]) with muscle fat infiltration, the Brooke score (a motor function score) and age-related degenerative changes, proposing a novel motor function index. This data-driven method has the potential to inform future home deployments with wearable devices, allowing better longitudinal tracking of treatment efficacy for children with neuromuscular disorders.

Fair and Adaptive Framework for Work-Related Musculoskeletal Disorders

Predicting external hand load from sensor data is essential for ergonomic exposure assessments, as obtaining this information typically requires direct observation or supplementary data. While machine learning can estimate hand load from posture or force data, we found systematic bias tied to biological sex, with predictive disparities worsening in imbalanced training datasets. To address this, we developed a fair predictive model using a Variational Autoencoder with feature disentanglement, which separates sex-agnostic from sex-specific motion features. This enables predictions based only on sex-agnostic patterns. Our proposed algorithm outperformed conventional machine learning models, including -Nearest Neighbors, Support Vector Machine, and Random Forest, achieving a mean absolute error of 3.42 and improving fairness metrics like statistical parity and positive and negative residual differences, even when trained on imbalanced sex datasets. These results underscore the importance of fairness-aware algorithms in avoiding health and safety disadvantages for specific worker groups in the workplace.

EEG and Keystroke Based Multimodal Biometric System

Electroencephalography (EEG) based biometric systems are gaining attention for their anti-spoofing capability but lack accuracy due to signal variability at different psychological and physiological conditions. On the other hand, keystroke dynamics-based systems achieve very high accuracy but have low anti-spoofing capability. To address these issues, a novel multimodal biometric system combining EEG and keystroke dynamics is proposed in this project. A dataset was created by acquiring both keystroke dynamics and EEG signals simultaneously from 10 users. Each user participated in 500 trials at 10 different sessions (days) to replicate real-life signal variability. A machine learning classification pipeline is developed using multi-domain feature extraction (time, frequency, time-frequency), feature selection (Gini impurity), classifier design, and score level fusion. Different classifiers were trained, validated, and tested for two different classification experiments – personalized and generalized. For identification and authentication, 99.9% and 99.6% accuracies are achieved, respectively for the Random Forest classifier in 5 fold cross-validation. These results outperform the individual modalities with a significant margin (~5%). We also developed a binary template matching-based algorithm, which gives 93.64% accuracy 6X faster. The proposed method can be considered secure and reliable for any kind of biometric identification and authentication.

Nurse Care Activity recognition

Nurse care activity recognition is a new challenging research field in human activity recognition (HAR) because unlike other activity recognition, it has severe class imbalance problem and intra-class variability depending on both the subject and the receiver. In this project, we applied the Random Forest-based resampling method to solve the class imbalance problem in the Heiseikai data, nurse care activity dataset. This method consists of resampling, feature selection based on Gini impurity, and model training and validation with Stratified KFold cross-validation. By implementing the Random Forest classifier, we achieved 65.9% average cross-validation accuracy in classifying 12 activities conducted by nurses in both lab and real-life settings. Our team, "Britter Baire" developed this algorithmic pipeline for "The 2nd Nurse Care Activity Recognition Challenge Using Lab and Field Data".

Deep Learning Based EMG Hand Gesture Classification

In this project, a comparative study of classifying different hand gestures of two well-known surface Electromyogram (sEMG) data sets, Rami Khusaba EMG repository, and UCI Machine Learning Repository, is shown. Applying transfer learning and CNN-LSTM neural network architectures, we find out a suitable control scheme for a myoelectric prosthetic hand (we mention it as DUFAB Hand). At first, the continuous wavelet transform (CWT) is exploited to create images from the sEMG signal, which serves as a powerful feature for the classification of different hand gestures. Then, we transferred the learning of various neural nets of image classification, e.g., AlexNet, and ResNet-18 to the sEMG image classification. Application of these deep neural networks outperformed general machine learning techniques with higher accuracy and performance. For example, the combination of CNN and LSTM has achieved the state of the art accuracies for these data sets, of 99.72% for UCI Machine Learning Repository and 99.83% for Rami Khusaba EMG repository respectively. The main contribution of this paper is, establishing an algorithmic pipeline using continuous wavelet transform (CWT) and CNN-LSTM deep neural networks to achieve high accuracy in two sEMG datasets.

Design and Implementation of a Low-Cost prosthetic hand

In this project, the design and implementation of a low-cost myoelectric prosthetic hand prototype are done. Our main contributions are - (i) proposing a new Computer Aided Design (CAD) suitable for laser cutter for prosthetic hand and (ii) a new feature, XCORR that calculates cross-correlation between two channels of EMG signals for same gesture, is introduced to classify Electromyography (EMG) signals. Individual parts are designed in a 2D plane and assembled to make a 3D structure in CAD software. Acrylic is used as the main material. The prosthetic hand can exhibit 6-degree of freedom (DOF), i.e., six gestures which are required to perform essential daily tasks and grabbing objects. Commonly used features, such as SMAT, HJORTH parameters are extracted from raw EMG signals and combined with newly introduced feature XCORR to classify 6 types of individual and combined finger movements. Using the classified EMG signal as input we have controlled the designed prosthetic hand in both simulation and hardware level. The controller is designed in such a way that the angular displacement of each finger meets a predefined angle value for each classified gesture. For classification of EMG signals, the highest accuracy is observed as 93.9% with a low computational load. The proposed hand will be affordable and suitable for amputees and help them regain their ability to work.

Research Assistant

Qatar University Machine Learning Group, Remote

Research Assistant

Fab Lab DU, University of Dhaka, Bangladesh