Soumyadeep Bhattacharjee
  Grade 9, High School Student
  Williamsville East High School
  151 Paradise Rd, East Amherst, NY 14051
  Google Scholar Link

Soumyadeep Bhattacharjee is a High School student, who dreams of building a humanoid robot that will be smart enough to learn any repititive tasks to support humans, while, the exclusive focus of the human race will be innovation and exploration within as well as beyond our universe. Soumyadeep is bestowed with an enormous passion about Mathematics and Computer Science. He has been homeschooled most of his elementary, middle school years and joined traditional schooling system only after shifting to New York in 2019, when he joined grade 7 at Transit Middle School. His present research intrests lie in building intelligent Internet-Of-Things (IOT) systems for various real world applications using innovative applications of Artificial Intelligence(AI), Machine Learning(ML), and Deep Learning(DL) Techniques. He is advised by Professor Wenyao Xu from the Computer Science & Engineering Dept., University at Buffalo.

Recent Updates :


  • VocalScop: A Knowledge Preserving Multi-class Disease Recognition Model through Daily-Conversation Data; Soumyadeep B, Wenyao Xu; Under Review at ACM Transactions on Computing for Healthcare Journal
  • VoiceLens: A Multi-class Disease Classification Model through Daily-Life Speech Data; Soumyadeep B, Wenyao Xu; IEEE/ACM Conference on Connected Health Applications, Systems, and Engineering Technologies (IEEE CHASE 2021) Slides PDF
  • Anomalous Pattern Recognition in Vital Health Signals via Multimodal Fusion; Soumyadeep B, Huining Li, Wenyao Xu; Body Area Networks. Smart IoT and Big Data for Intelligent Health, Volume 420, Lecture Notes of the Institute for Computer Sciences, Social Informatics and Telecommunications Engineering, January 27, 2022. Video Presentation.
  • Projects:

    VoiceLens: A Multi-class Disease Classification Model through Daily-Life Speech Data


    In this project, we present a disease-specific classification of pathology signals based on voice samples. The proposed method enables a proactive, deep learning based sequential audio analysis and classification framework that combines the effectiveness of the powerful Mel-Frequency-Cepstral-Coefficients (MFCC) within a two-phase classification framework to build an accurate disease prediction. The first phase of pathology detection analysis captures the fine-grained details of these disorders and their sequential variation patterns within a stacked Long Short-Term Memory (LSTM) network to make the baseline binary detection (Healthy Vs. Pathology). In the second phase, the potential pathology samples are then investigated by a deep multi-layer learned descriptor to accurately identify the disease condition. The present research has evaluated the performance of the system using the large scale Saarbruecken Voice Database comprising of samples from 2000 individuals with 71 disease patterns including Laryngel Cancers, Dish-Syndrome and Parkinson's disease. Unlike most of the existing methods, which simplify the original multi-class voice pathology detection problem by formulating a binary classifier that considers all pathology classes as a single non-healthy category, the main advantage of the proposed model is that it is sufficiently generic, real time, and can effectively distinguish multiple pathology classes within an integrated multi-class framework. The experimental results show remarkable improvement in the system performance by reporting an accuracy up to 97.5% in binary classification experiment setting, where the model also obtains 98.00% and 97.13% for F1-Score and Recall. The proposed system reports around 15% (and 12%) average gain in the accuracy (and F1-score) in a multi-class scenario with as many as 7 categories across 6 different pathology classes.

    Anomalous Pattern Recognition in Vital Health Signals via Multimodal Fusion


    Increasingly, care-giving to senior citizens and patients requires monitoring of vital signs of heartbeat, respiration and blood pressure for an extended period of time.In this paper, we propose an unobtrusive multimodal machine learning based synchronized biological signal monitoring process deployed into a chair setup that may execute a continual health observation task without interrupting the seat occupant's daily activities. A cepstral based peak fusion technique is introduced to obtain a robust characterization of each biological signal that combines the list of dominant peaks in the input signal and its corresponding cepstrum. This works as an input to the following multimodal anomaly detection process that not only enables an accurate identification and localization of aberrant signal patterns, but also facilitates the proposed model to adopt an individual's unique health characteristics over time. In this work, we use ECG (Electrocardiogram), Femoral Pulse, PPG (Photoplethysmogram), and Body Temperature to monitor an individual's health condition. An extensive analysis that demonstrates performance both in the publicly available datasets as well as our real-life lab experimental settings with 10 participants over a wide range of ages.

    Social Engagement Interactive Prototype


    In this project, we develop an interactive prototype between users and Alexa chatbot that integrates mental health assessment and mindfulness intervention in a closed-loop fashion to reduce loneliness and depression and improve quality of life in community-dwelling older adults. Specifically, we employ Alexa to daily query some mental health related questions. The question poll consists of CAMS-R, PHQ-2, GAD-2, PSS-4, Sleep Disturbance short form, fatigue, UCLA Loneliness Scale-8, Sleep Hygiene Index, and negative mood survey. These questions are designed to identify mental health related high-risk symptoms, including fatigue, loneliness, sleep quality, mindfulness, depression, stress, and negative mood. Once the user launches a request for interaction, Alexa will randomly select the questionnaire set to ask the user corresponding questions. The definite answers are recorded via Alexa interface, and are stored in the local database. Based on the user provided answers, we will develop a personalized mindfulness activity recommendation algorithm to provide just-in-time mindfulness coaching service for users with mental health problems. After receiving the recommended list of mindfulness activities through Alexa, the user responds to execute daily mindfulness practice. The type, amount, and quality of daily completed mindfulness will affect the user’s behavior patterns and life habits. The changes of the user's behavior patterns and lifestyle will alter symptom assessment. Then, Alexa will update the recommendation list of mindfulness practices based on the user’s new symptoms and deliver it to the user. After that, the user will respond and continue another round of intervention. Therefore, such a closed-loop mindfulness intervention scheme can iteratively improve user’s high-risk symptoms and solve mental health problems.
    Youtube Link: