Nazmun Nahar

Research Interests

Digital Health
Human Computer Interaction
Social Computing
Mental Health
Natural Language Processing
Data Science

Publications

Preprint
Automatic Speech Recognition for Biomedical Data in Bengali Language.
Shariar Kabir*1; Nazmun Nahar*1; Mamunur Rashid*2; Shyamasree Saha*1. arXiv preprint arXiv:2406.12931 (2024)(pdf)

1. MedAi Limited 2. Birmingham Univeristy, UK

Working on

AmarDoctor: First Multilingual Digital Platform For AI-Driven Primary Care Triage And Patient Management System For Bengali Speakers. Nazmun Nahar*1; Shariar Kabir*1; Sumaiya Tasnia Khan*4; Suparna Das*3; Shyamasree Saha*1; Mamunur Rashid*2. (pdf)
1. MedAi Limited 2. Birmingham University, UK 3. National Health Services, England 4. Central Police Hospital, Bangladesh

Award and Achievements

MedAi's solution AmarDoctor has been selected as one of the six solvers for the MIT Solve Global Health Equity Challenge Award 2024 out of nearly 2200+ companies worldwide.[Solution]

Being a major contributor to this project since its inception, seeing it flourish makes me feel proud. Therefore, this achievement is very personal to me.

Education

2015-2019

BSc in Computer Science and Engineering

Bangladesh University of Engineering & Technology(BUET)

Thesis Project: Bengali Text Recognition Using Deep Learning, under the supervision of Professor Dr. Md. Monirul Islam. For this project, I created a word image dataset from printed documents, annotated it, then trained deep neural networks on it using a variety of methods, including CNN, RNN, LRU, and others.

Coursework: Artificial intelligence, Structured programming language, Object oriented programming language, Data Structures, Algorithms, Database, Computer architecture, Software engineering and information system design, Software development, Basic graph theory and others.

Experience

Nov, 2019-Mar, 2020

Teaching

Bangladesh Institute of Science & Technology (Part-Time)

520223, Microprocessor and Assembly Languages
520224, Microprocessor and Assembly Languages Lab
540206, Computer Graphics Lab
540219, Network and Information Security

2021-Present

Research Engineer | Data Science | NLP | Cloud

MedAi Pvt. Limited

Develoment of Clinical Decision Support System (CDSS) and Symptom Checker Module for Bengali Speakers
Creation of Knowledge Graph Database (TypeDB) to represent biomedical entities and their relationships such as patients’ disease, symptom (phenotypes), drug and diagnostic tests.
Curating symptoms across multiple languages, exploring symptom variations, and examining relationships between symptoms.
Organizing disease data alongside their associations with symptoms.
Medical Assistant Chatbot directing patients to the appropriate specialization.
(NER) pipeline for extracting relevant information from patient complaints.
Normalizing symptoms using Bio-Bert, mapping user-provided symptoms to existing symptom entries.
Patient scenarios used to validate the accuracy of our symptom checker module. Deriving significant insights from these scenarios.
Extensive expertise in creating intricate CSV files for data science purposes and visualization tasks utilizing the Pandas library.
Implementation of a backend API service using Django REST Framework
Utilizing SQLite and PostgreSQL for managing credentials and client data
Authentication based on JSON Web Tokens (JWT)
Integration of payment gateway service
Integration of SMS service
Deploying backend services on EC2 instances.
Routing all APIs via an API gateway.
Using an S3 bucket for data storage.
Utilizing AWS Simple Email Service (SES) for sending emails.

2021-2022

Assistant Porgrammer

Janata Bank PLC.

I secured a position in a government bank through a rigorous and highly competitive selection process. However, I found that the role lacked the technical challenges I seek in my career. Consequently, I made a bold decision to leave the position in pursuit of opportunities that allow me to continuously learn and apply my technical skills to real-world problems.

Tech Stack

Programming Languages

Python

C, C++

Java

LaTeX

SQL, TypeQL, PostgreSQL

Bash

Libraries/Farmeworks

Django Rest Framework

Pandas

Keras, Tensorflow

Huggingface Diffusers

Huggingface Transformers

Matplotlib, NumPy

OpenCV

Sphinx

Beautifulsoup

AWS

AWS EC2 Instances

AWS API Gateway

Route53, DNS

S3 Bucket, Elsastic IPs

AMI, SES

NLP

NER

Rasa

NLU

LLM

NLTK

Tools & OS

Gunicorn, Nginx

Git

Jira, Confluence

Linux, Windows

Projects

Multilingual Symptom Checker & Clinical Decision Support System

This symptom checker module serves as an intelligent tool that facilitates the symptom selection process by suggesting relevant symptoms based on the input provided by the patient.
Once the patient has completed entering their symptoms, the module prompts additional questions tailored to the patient's responses (yes, no, or don't know). Subsequently, it presents provisional diagnoses along with pertinent specialization recommendations, guiding the patient towards the appropriate healthcare professional.

Fig: Flow of Symptom Checker and Disease Prognosis Tool

Medical Assistant Chatbot

The medical assistant chatbot discerns the user's mood and offers general illness options if the mood is suboptimal. It then tailors additional questions based on the identified intent of the chosen option and ultimately suggests whether the user should consult a physical health specialist or a mental health specialist.

Fig: Medical Assistant Chatbot Flow-Diagram

Stable Diffusion Model Fine Tuning

To enhance user experience and provide more visually appealing symptom icons, I fine-tuned a Stable Diffusion model to generate custom vector arts for image data. This process involved preparing images, generating preliminary captions with image-to-text model, curating the captions, and training a Lora model.

Fig: Stable-diffusion model fine-tuning steps

Symptom Normalization

While there are variations of the same symptoms in both Bengali and English, it's impractical to encompass all potential variations. Therefore, this module aligns user-provided symptoms with our existing database.

We employ a BERT model for sentence embeddings, specifically leveraging Bio-BERT for superior performance with medical data. To ensure accuracy, we utilize a combination of sentence similarity scores and organ matching. This prevents symptoms specified for one organ from being erroneously matched to another organ solely based on similarities in sentence structure.

Fig: Symptom Normalization Flow-Diagram

Client Management System

Created a client management system allowing authorized clients to access our AI-powered services. Implemented JWT-based authentication, enabling clients to purchase services for customized durations according to available packages.

Fig: Client Management System

Clinical NER for Bengali (In progress)

Due to the limited availability of conversational medical data in Bengali, our goal is to gather medical data, annotate it, and fine-tune a Large Language Model (LLM) for Named Entity Recognition (NER) of entities in the medical domain.
This project is currently in progress, with the objective of extracting medical data from conversations between doctors and patients. The data includes vital signs such as blood pressure, pulse rate, and temperature, as well as existing diseases, current symptoms, and diagnostic test information.