Umair Mohammad
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
Research Subjects
Umair Mohammad has a total of 12 co-authors. Here are the top co-authors by number of shared publications:
- Sameh Sorour
School of Computing, Queens University, Kingston, ON, Canada
5 shared publications - Fahad Saeed
Florida International University (FIU)
4 shared publications - Mohamed Hefeida
Lane Department of Computer Science and Electrical Engineering, West Virginia University, Morgantown, WV, USA
3 shared publications Sameh Sorour
School of Computing, Queen’s University, Kingston, ON, Canada
3 shared publicationsFahad Saeed
Florida International University,Knight Foundation School of Computing and Information Sciences,Miami,FL,USA
2 shared publicationsMohamed Hefeida
Department of Electrical and Computer Engineering, University of Idaho, Moscow, ID
2 shared publications
Latest Publications
- Open Access
Utilizing Pretrained Vision Transformers and Large Language Models for Epileptic Seizure Prediction
Repeated unprovoked seizures is a major source of concern for patients suffering from epilepsy. Predicting seizures before they occur is of interest to both machine-learning scientists as well as clinicians, and is an active area of research. The variability of EEG sensors, type of seizures, and specialized knowledge required for annotating the data complicates the large-scale annotation process essential for supervised predictive models. To address these challenges, we propose the use of Vision Transformers (ViTs) and Large Language Models (LLMs) that were originally trained on publicly available image or text data. Our work leverages these pre-trained models by refining the input, embedding, and classification layers in a minimalistic fashion to predict seizures. Our results demonstrate that LLM's outperforms the ViT's in patient-independent seizure prediction achieving a sensitivity of 79.02% which is 8% higher compared to ViT's and about 12% higher compared to a custom-designed ResNet-based model. Our work demonstrates the successful feasibility of pre-trained models for seizure prediction with its potential for improving the quality of life of people with epilepsy. Our code and related materials are available open-source at: https://github.com/pcdslab/UtilLLM_EPS
Open AccessUtilizing Pretrained Vision Transformers and Large Language Models for Epileptic Seizure Prediction
Repeated unprovoked seizures is a major source of concern for patients suffering from epilepsy. Predicting seizures before they occur is of interest to both machine-learning scientists as well as clinicians, and is an active area of research. The variability of EEG sensors, type of seizures, and specialized knowledge required for annotating the data complicates the large-scale annotation process essential for supervised predictive models. To address these challenges, we propose the use of Vision Transformers (ViTs) and Large Language Models (LLMs) that were originally trained on publicly available image or text data. Our work leverages these pre-trained models by refining the input, embedding, and classification layers in a minimalistic fashion to predict seizures. Our results demonstrate that LLM's outperforms the ViT's in patient-independent seizure prediction achieving a sensitivity of 79.02% which is 8% higher compared to ViT's and about 12% higher compared to a custom-designed ResNet-based model. Our work demonstrates the successful feasibility of pre-trained models for seizure prediction with its potential for improving the quality of life of people with epilepsy. Our code and related materials are available open-source at: https://github.com/pcdslab/UtilLLM_EPS
- Open Access
MLSPred-Bench: ML-Ready Benchmark Leveraging Seizure Detection EEG data for Predictive Models
Predicting epileptic seizures is a significantly challenging task as compared to detection. While electroen-cephalography (EEG) data annotated for detection is available from multiple repositories, they cannot readily be used for predictive modeling. In this paper, we designed and developed a strategy that can be used for converting any EEG big data annotated for detection into ML-ready data suitable for prediction. The generalizability of our strategy is demonstrated by executing it on Temple University Seizure (TUSZ) corpus which is annotated for seizure detection. This execution results in 12 ML-ready datasets, collectively called MLSPred-Bench benchmark, which constitutes data for training, validating and testing seizure prediction models. Our strategy uses different variations of seizure prediction horizon (SPH) and the seizure occurrence period (SOP) to make more than 150GB of ML-ready data. To illustrate that the generated data can be used for predictive modeling, we executed an ML model on all the benchmarks which resulted in variable performances when compared with the original model and its performance. We expect that our strategy can be used as a general method to transform seizure detection EEG big data into ML-ready datasets useful for seizure prediction. Our code and related materials will be made available at https://github.com/pcdslab/MLSPred-Bench .
Open AccessMLSPred-Bench: ML-Ready Benchmark Leveraging Seizure Detection EEG data for Predictive Models
Predicting epileptic seizures is a significantly challenging task as compared to detection. While electroen-cephalography (EEG) data annotated for detection is available from multiple repositories, they cannot readily be used for predictive modeling. In this paper, we designed and developed a strategy that can be used for converting any EEG big data annotated for detection into ML-ready data suitable for prediction. The generalizability of our strategy is demonstrated by executing it on Temple University Seizure (TUSZ) corpus which is annotated for seizure detection. This execution results in 12 ML-ready datasets, collectively called MLSPred-Bench benchmark, which constitutes data for training, validating and testing seizure prediction models. Our strategy uses different variations of seizure prediction horizon (SPH) and the seizure occurrence period (SOP) to make more than 150GB of ML-ready data. To illustrate that the generated data can be used for predictive modeling, we executed an ML model on all the benchmarks which resulted in variable performances when compared with the original model and its performance. We expect that our strategy can be used as a general method to transform seizure detection EEG big data into ML-ready datasets useful for seizure prediction. Our code and related materials will be made available at https://github.com/pcdslab/MLSPred-Bench .
- 29 April 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- Vol. 2024, 620-626
Robustness of ML-Based Seizure Prediction Using Noisy EEG Data From Limited Channels
Seizures pose a significant health hazard for over 50 million individuals with epilepsy worldwide, with approximately 56% experiencing uncontrollable seizures according to the CDC. Predicting seizures is challenging even with the availability of various sensors (gyroscopes, pulse rate sensors, heart rate monitors, etc). Electroencephalography (EEG) data can directly measure the activity of the brain and has been the choice of leveraging deep learning (DL) models for seizure prediction. Despite DL models achieving over 95% accuracy on retroactive clinical-grade EEG data, this performance fails to translate in real-world settings where the accuracy goes down to 66% which warrants further investigation. Moreover, consumer-grade wearable EEG headsets, characterized by lower data quality and a varying number of channels across brands, present additional challenges. In this paper, we estimate the robustness of DL models which are trained on clinical-grade EEG data but tested on the type of data expected from consumer-grade wearable EEG headsets. We select the previously published model SPERTL to estimate its robustness when: (1) predicting with data from less leads/channels, (2) predicting when faced with streaming data, (3) evaluating performance on imbalanced data with more interictal segments. Our results are compared against baseline results from the SPERTL model which we have re-configured to operate independently of the number of channels with an average baseline area under the curve (AUC) score of 98.56%. Our results demonstrate that though the model is surprisingly resilient to streaming and noisy data, reducing the number of channels and a higher class imbalance have a more severe degradation. The AUC across all cross-validation sets degrades only by 2% and 3% on average for noisy and streaming data, respectively. However, a performance reduction, on average, is observed by 32% when imbalance is increased with higher percentage of interictal samples, and up to 16% when using lower number of channels.
- 29 April 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- Vol. 2024, 620-626
Robustness of ML-Based Seizure Prediction Using Noisy EEG Data From Limited Channels
Seizures pose a significant health hazard for over 50 million individuals with epilepsy worldwide, with approximately 56% experiencing uncontrollable seizures according to the CDC. Predicting seizures is challenging even with the availability of various sensors (gyroscopes, pulse rate sensors, heart rate monitors, etc). Electroencephalography (EEG) data can directly measure the activity of the brain and has been the choice of leveraging deep learning (DL) models for seizure prediction. Despite DL models achieving over 95% accuracy on retroactive clinical-grade EEG data, this performance fails to translate in real-world settings where the accuracy goes down to 66% which warrants further investigation. Moreover, consumer-grade wearable EEG headsets, characterized by lower data quality and a varying number of channels across brands, present additional challenges. In this paper, we estimate the robustness of DL models which are trained on clinical-grade EEG data but tested on the type of data expected from consumer-grade wearable EEG headsets. We select the previously published model SPERTL to estimate its robustness when: (1) predicting with data from less leads/channels, (2) predicting when faced with streaming data, (3) evaluating performance on imbalanced data with more interictal segments. Our results are compared against baseline results from the SPERTL model which we have re-configured to operate independently of the number of channels with an average baseline area under the curve (AUC) score of 98.56%. Our results demonstrate that though the model is surprisingly resilient to streaming and noisy data, reducing the number of channels and a higher class imbalance have a more severe degradation. The AUC across all cross-validation sets degrades only by 2% and 3% on average for noisy and streaming data, respectively. However, a performance reduction, on average, is observed by 32% when imbalance is increased with higher percentage of interictal samples, and up to 16% when using lower number of channels.
- 5 December 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 3313-3320
Energy Efficient AI/ML based Continuous Monitoring at the Edge: ECG and EEG Case Study
In this paper, we propose an energy-efficient approach for machine-learning based continuous testing and monitoring of long-term patients at the wireless edge. The approach is applicable for any wearable sensors that generate time-series data. Our scheme simultaneously performs sensor-server clustering while ensuring the delay requirements of every user are met. In contrast to previous works on task offloading for generic edge computing/machine learning, our proposed model considers application specific parameters including the sampling rate, measurement duration and number of input channels/leads. We formulate the problem as a mixed integer nonlinear program (MINLP) and propose a heuristic solution. Two applications, cardiac event prediction from a wearable electrocardiograms (ECG), and epileptic seizure prediction from wearable scalp electroencephalography (EEG) are used to demonstrate the superiority of the proposed approach. Results indicate that our proposed algorithm can provide up-to 70% energy savings compared to the case when maximum settings are used.
- 5 December 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 3313-3320
Energy Efficient AI/ML based Continuous Monitoring at the Edge: ECG and EEG Case Study
In this paper, we propose an energy-efficient approach for machine-learning based continuous testing and monitoring of long-term patients at the wireless edge. The approach is applicable for any wearable sensors that generate time-series data. Our scheme simultaneously performs sensor-server clustering while ensuring the delay requirements of every user are met. In contrast to previous works on task offloading for generic edge computing/machine learning, our proposed model considers application specific parameters including the sampling rate, measurement duration and number of input channels/leads. We formulate the problem as a mixed integer nonlinear program (MINLP) and propose a heuristic solution. Two applications, cardiac event prediction from a wearable electrocardiograms (ECG), and epileptic seizure prediction from wearable scalp electroencephalography (EEG) are used to demonstrate the superiority of the proposed approach. Results indicate that our proposed algorithm can provide up-to 70% energy savings compared to the case when maximum settings are used.
- 14 February 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Green Communications and Networking
- Vol. 7 (4) , 1766-1777
Energy Aware Task Allocation for Semi-Asynchronous Mobile Edge Learning
This paper extends the paradigm of "mobile edge learning (MEL)" by designing an energy-aware optimal task allocation scheme for training a machine learning (ML) model in a semi-asynchronous manner across multiple learners connected via the resource-constrained wireless edge network. The tasks are allocated such that the local dataset size selected at each learner ensures completion within a given global delay constraint and a local maximum energy consumption limit. Hence, the designed method is heterogeneity aware (HA) because it offers a trade-off between resource consumption and MEL performance by directly relating the time and energy consumption to the heterogeneous communication/computational capabilities of learners. Because the resulting optimization is an NP-hard quadratically-constrained integer linear program (QCILP), a two-step suggest-and-improve (SAI) solution is proposed. The proposed HA semi-asynchronous (HA-Asyn) approach is compared against the HA synchronous (HA-Sync) scheme and the heterogeneity unaware (HU) synchronous/asynchronous (HU-Sync/Asyn) equal batch allocation schemes. Results from a system of 20 learners tested for various completion time and energy consumption constraints show that the proposed HA-Asyn method works better than the HU-Sync/Asyn approaches and can even provide gains of up-to 25% compared to the HA-Sync scheme.
- 14 February 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Green Communications and Networking
- Vol. 7 (4) , 1766-1777
Energy Aware Task Allocation for Semi-Asynchronous Mobile Edge Learning
This paper extends the paradigm of "mobile edge learning (MEL)" by designing an energy-aware optimal task allocation scheme for training a machine learning (ML) model in a semi-asynchronous manner across multiple learners connected via the resource-constrained wireless edge network. The tasks are allocated such that the local dataset size selected at each learner ensures completion within a given global delay constraint and a local maximum energy consumption limit. Hence, the designed method is heterogeneity aware (HA) because it offers a trade-off between resource consumption and MEL performance by directly relating the time and energy consumption to the heterogeneous communication/computational capabilities of learners. Because the resulting optimization is an NP-hard quadratically-constrained integer linear program (QCILP), a two-step suggest-and-improve (SAI) solution is proposed. The proposed HA semi-asynchronous (HA-Asyn) approach is compared against the HA synchronous (HA-Sync) scheme and the heterogeneity unaware (HU) synchronous/asynchronous (HU-Sync/Asyn) equal batch allocation schemes. Results from a system of 20 learners tested for various completion time and energy consumption constraints show that the proposed HA-Asyn method works better than the HU-Sync/Asyn approaches and can even provide gains of up-to 25% compared to the HA-Sync scheme.
- 21 December 2021
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Mobile Computing
- Vol. PP (15361233) , 1
Dynamic Task Allocation for Mobile Edge Learning
This paper introduces the new paradigm of Mobile Edge Learning "MEL" that enables the implementation of realistic distributed machine learning (DML) tasks on wireless edge nodes while taking into consideration the heterogeneous computing and networking environments. Therefore, a heterogeneity aware (HA) scheme is designed to solve the problem of dynamic task allocation for MEL in a way that maximizes the DML accuracy over wireless heterogeneous nodes or 'learners' while respecting the time constraints. The problem is first formulated as a quadratically-constrained integer linear program (QCILP). Being NP-hard, it is relaxed into a non-convex problem over real variables which can be solved using commercially available numerical solvers. The relaxation also allows us to propose a solution based on deriving the analytical upper bounds of the optimal solution using Lagrangian analysis and Karush-Kuhn-Tucker (KKT) conditions. The merits of the proposed analytical solution are demonstrated by comparing its performance to the numerical approaches and comparing the validation accuracy of the proposed HA scheme to the baseline heterogeneity unaware (HU) equal task allocation approach. Simulation results show that the HA schemes decrease convergence time up-to 56% and increase the final validation accuracy up-to 8%.
- 21 December 2021
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Mobile Computing
- Vol. PP (15361233) , 1
Dynamic Task Allocation for Mobile Edge Learning
This paper introduces the new paradigm of Mobile Edge Learning "MEL" that enables the implementation of realistic distributed machine learning (DML) tasks on wireless edge nodes while taking into consideration the heterogeneous computing and networking environments. Therefore, a heterogeneity aware (HA) scheme is designed to solve the problem of dynamic task allocation for MEL in a way that maximizes the DML accuracy over wireless heterogeneous nodes or 'learners' while respecting the time constraints. The problem is first formulated as a quadratically-constrained integer linear program (QCILP). Being NP-hard, it is relaxed into a non-convex problem over real variables which can be solved using commercially available numerical solvers. The relaxation also allows us to propose a solution based on deriving the analytical upper bounds of the optimal solution using Lagrangian analysis and Karush-Kuhn-Tucker (KKT) conditions. The merits of the proposed analytical solution are demonstrated by comparing its performance to the numerical approaches and comparing the validation accuracy of the proposed HA scheme to the baseline heterogeneity unaware (HU) equal task allocation approach. Simulation results show that the HA schemes decrease convergence time up-to 56% and increase the final validation accuracy up-to 8%.
- 1 August 2021
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- Vol. 2021, 137-142
Simulation Testbed for Evaluating Distributed Querying and Searching of Mass Spectrometry Big Data in a Network-based Infrastructure
Advance access and reuse mechanisms for large-scale Mass Spectrometry (MS) data are essential for democratizing data for the omics research community and making it adhere to FAIR (Findable, Accessible, Interoperable, Reusable) principles. Although a number of centralized data repositories have been established, they have been limited to search mechanisms that depend on the meta-data associated with these MS datasets. Furthermore, they require constant influx of resources for maintenance. In this paper, we proposed an alternative novel distributed infrastructure for direct MS/MS spectral search. We designed and developed a simulation testbed using concepts from computer networks, queuing theory, and stochastic simulation methods. Results show that a distributed MS search based on raw MS/MS spectra can scale gracefully for up-to 2000 participating nodes, while simultaneously processing queries using the proposed networked infrastructure on the order of milliseconds to a few seconds for up-to a total of fifty billion MS/MS spectra.
- 1 August 2021
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- Vol. 2021, 137-142
Simulation Testbed for Evaluating Distributed Querying and Searching of Mass Spectrometry Big Data in a Network-based Infrastructure
Advance access and reuse mechanisms for large-scale Mass Spectrometry (MS) data are essential for democratizing data for the omics research community and making it adhere to FAIR (Findable, Accessible, Interoperable, Reusable) principles. Although a number of centralized data repositories have been established, they have been limited to search mechanisms that depend on the meta-data associated with these MS datasets. Furthermore, they require constant influx of resources for maintenance. In this paper, we proposed an alternative novel distributed infrastructure for direct MS/MS spectral search. We designed and developed a simulation testbed using concepts from computer networks, queuing theory, and stochastic simulation methods. Results show that a distributed MS search based on raw MS/MS spectra can scale gracefully for up-to 2000 participating nodes, while simultaneously processing queries using the proposed networked infrastructure on the order of milliseconds to a few seconds for up-to a total of fifty billion MS/MS spectra.
- 9 January 2021
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 1-4
Optimal Task Allocation for Mobile Edge Learning with Global Training Time Constraints
This paper proposes to maximize the accuracy of a distributed machine learning (ML) model trained on learners connected via the resource-constrained wireless edge. We jointly optimize the number of local/global updates and the task size allocation to minimize the loss while taking into account heterogeneous communication and computation capabilities of each learner. By leveraging existing bounds on the difference between the optimal and actual training loss, we derive an expression for the objective function in terms of the local updates. The resulting convex program is solved to obtain the optimal number of local updates which is used to obtain the total updates and batch sizes for each learner. The merits of the proposed solution, which is heterogeneity aware (HA), are exhibited by comparing its performance to the heterogeneity unaware (HU) approach.
- 9 January 2021
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 1-4
Optimal Task Allocation for Mobile Edge Learning with Global Training Time Constraints
This paper proposes to maximize the accuracy of a distributed machine learning (ML) model trained on learners connected via the resource-constrained wireless edge. We jointly optimize the number of local/global updates and the task size allocation to minimize the loss while taking into account heterogeneous communication and computation capabilities of each learner. By leveraging existing bounds on the difference between the optimal and actual training loss, we derive an expression for the objective function in terms of the local updates. The resulting convex program is solved to obtain the optimal number of local updates which is used to obtain the total updates and batch sizes for each learner. The merits of the proposed solution, which is heterogeneity aware (HA), are exhibited by comparing its performance to the heterogeneity unaware (HU) approach.
- 7 July 2020
- Published by SAGE Publications in Transportation Research Record: Journal of the Transportation Research Board
- Vol. 2674 (9) , 898-912
Vendor-Independent Reliability Testing Model for Vehicle-to-Infrastructure Communications
This paper describes a vendor-independent reliability testing approach for vehicle-to-infrastructure (V2I) communications in connected vehicle traffic signal system applications. It provides an alternative to using the communication data reported by proprietary vendor-supplied interfaces. This approach was based on building a rigorously tested translation model that uses measured received signal strength indicator (RSSI) from any V2I communication equipment to predict the corresponding packet delivery ratio (PDR). This was achieved by correlating the signal strength, measured using a generic power meter, to PDR values reported in the communication interface of the equipment of different vendors. Both stationary and in-motion (10–40 mph) field data collection tests were conducted at three intersections. These tests were performed over distances of up to 500 m between the road-side units (RSUs) and the on-board units (OBUs). In each test, the RSSI values for line-of-sight packet exchange between various RSUs and OBUs was collected in the field, using both a generic power meter and vendor-specific tools. Next, the results were statistically analyzed and logistic and linear regression models that predict PDR values were developed. A case study to test and validate this new PDR prediction model was conducted at two intersections in Boise, Idaho. This prediction model will enable transportation system operators to test and validate the efficiency of connected vehicle RSU/OBU communications at signalized intersection approaches under different traffic conditions, independent of vendor-provided tools.
- 7 July 2020
- Published by SAGE Publications in Transportation Research Record: Journal of the Transportation Research Board
- Vol. 2674 (9) , 898-912
Vendor-Independent Reliability Testing Model for Vehicle-to-Infrastructure Communications
This paper describes a vendor-independent reliability testing approach for vehicle-to-infrastructure (V2I) communications in connected vehicle traffic signal system applications. It provides an alternative to using the communication data reported by proprietary vendor-supplied interfaces. This approach was based on building a rigorously tested translation model that uses measured received signal strength indicator (RSSI) from any V2I communication equipment to predict the corresponding packet delivery ratio (PDR). This was achieved by correlating the signal strength, measured using a generic power meter, to PDR values reported in the communication interface of the equipment of different vendors. Both stationary and in-motion (10–40 mph) field data collection tests were conducted at three intersections. These tests were performed over distances of up to 500 m between the road-side units (RSUs) and the on-board units (OBUs). In each test, the RSSI values for line-of-sight packet exchange between various RSUs and OBUs was collected in the field, using both a generic power meter and vendor-specific tools. Next, the results were statistically analyzed and logistic and linear regression models that predict PDR values were developed. A case study to test and validate this new PDR prediction model was conducted at two intersections in Boise, Idaho. This prediction model will enable transportation system operators to test and validate the efficiency of connected vehicle RSU/OBU communications at signalized intersection approaches under different traffic conditions, independent of vendor-provided tools.