@article{das_multimodal_2024,
  title = {Multimodal speech recognition using {EEG} and audio signals: {A} novel approach for enhancing {ASR} systems},
  volume = {32},
  copyright = {All rights reserved},
  issn = {2352-6483},
  shorttitle = {Multimodal speech recognition using {EEG} and audio signals},
  url = {https://www.sciencedirect.com/science/article/pii/S2352648324000333},
  doi = {10.1016/j.smhl.2024.100477},
  urldate = {2024-05-29},
  journal = {Smart Health},
  author = {Das, Anarghya and Soni, Puru and Huang, Ming-Chun and Lin, Feng and Xu, Wenyao},
  month = jun,
  year = {2024},
  keywords = {Electroencephalography (EEG), Overt speech, Deep learning, Automatic Speech Recognition (ASR), Brain–Computer Interface (BCI), Multimodal analysis, Multimodal speech recognition, Neural signal processing},
  pages = {100477},
  file = {Multimodal_Speech_Recognition:/Users/anarghya/Zotero/storage/5FMFL8ZQ/Multimodal_Speech_Recognition.pdf:application/pdf}
}

Speech recognition using EEG signals captured during covert (imagined) speech has garnered substantial interest in Brain–Computer Interface (BCI) research. While the concept holds promise, current implementations must improve performance compared to established Automatic Speech Recognition (ASR) methods using audio. An area often underestimated in previous studies is the potential of EEG utilization during overt speech. Integrating overt EEG signals with speech data by leveraging advancements in deep learning presents significant potential to enhance the efficacy of these systems. This integration proves particularly advantageous in noisy environments and for individuals with speech impairments—challenges even conventional ASR techniques struggle to address effectively. Our investigation delves into this relationship by introducing a novel multimodal model that merges EEG and speech inputs. Our model achieves a multiclass classification accuracy of 95.39%. When subjected to artificial white noise added to the input audio, our model exhibits a notable level of resilience, surpassing the capabilities of models reliant solely on single EEG or audio modalities. The validation process, leveraging the robust techniques of t-SNE and silhouette coefficient, corroborates and solidifies these advancements.

@article{chen_shortvanet_2023,
  title = {Short:{VANet}: {An} {Intuitive} {Light}-{Weight} {Deep} {Learning} {Solution} {Towards} {Ventricular} {Arrhythmia} {Detection}},
  volume = {28},
  copyright = {All rights reserved},
  issn = {2352-6483},
  shorttitle = {Short},
  doi = {10.1016/j.smhl.2023.100388},
  urldate = {2024-01-30},
  journal = {Smart Health},
  author = {Chen, Tianyu and Gherardi, Alexander and Das, Anarghya and Li, Huining and Xu, Chenhan and Xu, Wenyao},
  month = jun,
  year = {2023},
  keywords = {Artificial intelligence, High accuracy, Neural network compression, Real-time heart monitoring, Sudden cardiac death},
  pages = {100388},
  file = {Chen et al. - 2023 - ShortVANet An Intuitive Light-Weight Deep Learni.pdf:/Users/anarghya/Library/CloudStorage/GoogleDrive-anarghya@buffalo.edu/My Drive/Papers/Chen et al. - 2023 - ShortVANet An Intuitive Light-Weight Deep Learni.pdf:application/pdf}
}

Ventricular Arrhythmia (VA) is a leading cause of sudden cardiac death (SCD), which kills an average of 180,000 to 350,000 people annually, accounting for 15%–20% of all deaths. Furthermore, fewer than 6% of those who experience sudden cardiac arrest outside the hospital survive, compared to 24% of those who experience SCD inside a hospital. To aid in earlier detection and improve outcomes for out-of-hospital cardiac events, an automated passive detection system for these events could be used. Such automated detection would allow users to raise their self-awareness of potential cardiac risks in life-threatening situations. Diagnosis and detection of heart dysfunctions at early stages can help to prevent complications of a patient’s condition. In this work, we propose VANet and design framework for ECG-related application, a small-scale deep learning-based real-time inference solution for VA detection. VANet achieves milliseconds scale inference speed on various platforms, including desktop CPUs, mobile devices, micro-controllers, and devices with constrained computation resources. It only requires a minimum of 13 kb of storage space and 34 kb of available run-time, making it small enough to be integrated into portable devices such as smartwatches and other Internet of Things (IoT) medical monitoring devices. VANet can trigger an alarm whenever it is necessary to alert someone with cardiac dysfunction. VANet leverages optimization techniques, such as residual connections, and architecture designs, such as transformers and RNNs, to maximize neural network performance and minimize computational and storage costs. Our architecture achieved a 96.89% accuracy using multiple different ECG collection devices.

@article{das_prediction_2023,
  title = {Prediction, {Risk} {Assessment} and {Comparison} of {Selected} {Emerging} {Markets}' {Stock} {Indices} {During} {COVID}-19 {Pandemic} {Using} the {Coherent} {Measure}: {Comparing} {Selected} {Emerging} {Markets}’ {Stock} {Indices}},
  volume = {16},
  copyright = {All rights reserved},
  issn = {1750-676X, 1750-6751},
  shorttitle = {Prediction, {Risk} {Assessment} and {Comparison} of {Selected} {Emerging} {Markets}' {Stock} {Indices} {During} {COVID}-19 {Pandemic} {Using} the {Coherent} {Measure}},
  doi = {10.5750/jpm.v16i3.1974},
  number = {3},
  urldate = {2024-01-30},
  journal = {The Journal of Prediction Markets},
  author = {Das, Prabir Kumar and Das, Anarghya},
  month = feb,
  year = {2023},
  pages = {81--97}
}

In this study, we modeled the log-return of three emerging markets’ stock indices, namely, Shanghai SSE, Russia MOEX, and Bombay Stock Exchange Sensex using the generalized hyperbolic family of distributions. We found the generalized hyperbolic family of distributions as the best fit for describing the probability density based on AIC and likelihood ratio test. The coherent risk measure, i.e., the expected shortfall, predicted using the best fit probability distribution, was used as a market risk quantification metric. During the COVID-19 period, the Indian stock market showed maximum market risk, followed by the Russian. The Chinese market showed the least market risk. Our experiment demonstrated a significant (p = 0.000) difference in the three markets concerning the coherent risk at different probability levels from 0.001 to 0.05 in the COVID-19 period using the Jonckheere-Terpstra test. The coherent market risk increased substantially in the Indian and Russian markets during the COVID-19 pandemic compared to the pre-COVID-19 period. However, in the Chinese market, we found that the coherent risk decreased during the COVID-19 period compared to the pre-COVID-19 period. We carried out the empirical study using the adjusted daily closing values of SSE, MOEX, and Sensex from July 2018 to July 2021 and dividing the data sets into pre-COVID-19 and COVID-19 periods based on the first emergence of the COVID-19 case.

@article{li_privacy_2022,
  title = {Privacy computing using deep compression learning techniques for neural decoding},
  volume = {23},
  copyright = {All rights reserved},
  issn = {2352-6483},
  doi = {10.1016/j.smhl.2021.100229},
  urldate = {2024-01-30},
  journal = {Smart Health},
  author = {Li, Huining and Chen, Huan and Xu, Chenhan and Das, Anarghya and Chen, Xingyu and Li, Zhengxiong and Xiao, Jian and Huang, Ming-Chun and Xu, Wenyao},
  month = mar,
  year = {2022},
  keywords = {Brain–computer interface, Neural decoding, Privacy computing, Unsupervised deep learning},
  pages = {100229},
  file = {Li et al. - 2022 - Privacy computing using deep compression learning .pdf:/Users/anarghya/Library/CloudStorage/GoogleDrive-anarghya@buffalo.edu/My Drive/Papers/Li et al. - 2022 - Privacy computing using deep compression learning .pdf:application/pdf}
}

The brain–computer interface supports a variety of applications with the help of machine learning technology. However, existing edge-cloud infrastructure requires subjects to send their sensitive neural signals to the cloud for training the model which brings privacy concerns. Although the recent distributed learning technology is used to help protect subjects’ privacy, it brings high communication costs and cannot avoid privacy reconstruction attacks. In this paper, we propose deep compression learning techniques in the privacy computing infrastructure that can be used for neural decoding while preserving privacy and largely reducing the communication cost. Specifically, we first perform heterogeneous neural signals processing and convert them to resized functional brain connectivity images. Then, a semantics structure-based unsupervised deep compression learning network is trained and generates a neural hash locally for each image. Each hash value is irreversible that cannot be used to reconstruct the user’s original neural signal. After that, the cloud end receives the uploaded neural hashes and corresponding labels and filters the abnormal ones. Finally, these neural hashes are used for training neural decoding models. Since a single hash value can correspond to different types of labels, it only needs to be uploaded once with a very small size and then reused for different training tasks, which largely reduce communication cost. Our experiment results show that the proposed privacy computing techniques can be applied to heterogeneous neural signals for training different neural decoding models where the relative accuracy can achieve above 83%.

@article{das_application_2020,
  title = {Application of nonlinear stochastic single source of error state space models in the forecasting of mobile subscribers in {India}},
  volume = {5},
  copyright = {All rights reserved},
  issn = {2053-0811, 2053-082X},
  doi = {10.1504/IJDS.2020.115874},
  language = {en},
  number = {4},
  urldate = {2024-01-30},
  journal = {International Journal of Data Science},
  author = {Das, Prabir Kumar and Das, Anarghya},
  year = {2020},
  pages = {333}
}

The nonlinear stochastic single source of error state space model with error, trend, and seasonality (ETS) was employed and found appropriate for modelling mobile subscriber time series data for individual metro cities, total mobile subscribers in all metro cities, and subscribers in all of India using monthly data from March 1997 to December 2018. Out of the different ETS models, the multiplicative error, additive trend, and no seasonality (M, A, N) models were appropriate for all series. These models were compared to the autoregressive integrated moving average model. The final model was identified based on the DieboldMariano test and time series cross-validation. The performance of the final model was compared to the long short-term memory (LSTM) model. The mean absolute error and root mean squared error showed that the ETS (M, A, N) performed superior over the standard LSTM. The ETS (M, A, N) model was used for computing the point forecast and 95% confidence intervals of the forecast values for the next 24 months. The subscribers of Delhi, Mumbai, Kolkata, and India are projected, at 95% probability, to have a high of 100 million, 70 million, 60 million, and 2000 million subscribers, respectively, by December 2020.

@inproceedings{das_towards_2023,
  title = {Towards {Analysis}-aware {EEG} {Compression} in {Wearable} {Computing}},
  copyright = {All rights reserved},
  doi = {10.1109/BSN58485.2023.10331060},
  urldate = {2024-01-30},
  booktitle = {2023 {IEEE} 19th {International} {Conference} on {Body} {Sensor} {Networks} ({BSN})},
  author = {Das, Anarghya and Xu, Wenyao},
  month = oct,
  year = {2023},
  note = {ISSN: 2376-8894},
  keywords = {Electroencephalography, Real-time systems, Analyzability, Brain modeling, Compression Domain Analysis, EEG Compression, Image coding, Image reconstruction, Training, Wearable Computing, Wearable devices},
  pages = {1--4},
  file = {Das and Xu - 2023 - Towards Analysis-aware EEG Compression in Wearable.pdf:/Users/anarghya/Zotero/storage/8LGCKQ86/Das and Xu - 2023 - Towards Analysis-aware EEG Compression in Wearable.pdf:application/pdf}
}

EEG signals are known for their high temporal resolution and data volume, posing challenges in storage, transmission, and analysis. With the increasing prevalence of wearable mobile BCI devices replacing large medical-grade recording devices, the need for efficient on-device processing and wireless transmission has grown. We propose a novel analyzability-aware EEG compression scheme optimized for machine consumption, facilitating efficient, real-time wearable processing while allowing reconstruction for detailed analysis and human understanding. Our method employs a wavelet-based compression technique that considers the Compression Factor, reconstruction error, and the impact of compression on on-device processing and data transmission. Experimental results using publicly available datasets demonstrate that machine learning models trained on compressed data using our method achieve performance comparable to the winning solution of BCI Competition II. Moreover, our approach offers multiple compression configurations suitable for various wearable computing scenarios.