Uma Subbiah

Uma Subbiah

Sofware Engineer | University of Oxford | Machine Learning & Data Science Enthusiast | Passionate about Tech Communities

Projects

Automated Classification of Automated Classification of Cardiac Action Potential Phenotypes for Prediction of Drug-Induced Pro-Arrhythmic Risk

Apr ‘21 - Oct ‘21

I worked on my Master’s thesis in the Computational Cardiovascular Science group, Dept. of Computer Science, University of Oxford.

The research conducted involves the use of machine learning techniques to enhance cardiac drug safety.

Our contribution to the field has been selected for presentation at the 2021 Safety Pharmacology Society’s (SPS) Virtual Annual Meeting (Oct 2021) and the 2021 Cardiac Physiome Meeting (Nov 2021).

Our accepted abstract was awarded the 2021 SPS Student Award. We also won the 1st place in the SPS Junior Investigator Poster Contest in October 2021.

Prototype - A Case Based Reasoning System to determine resource allocation in COVID wards

June ‘20

Case Based Reasoning is an Artificial Intelligence method that shows which of the known situations is the most similar to a new (previously unseen) case. By way of a literature survey, I’ve created a prototype of a CBR system to determine which previously admitted patient is most similar to an incoming patient using various medical readings. Using this, we can determine the probability of how severely COVID will affect the incoming patient. For instance, if the patient with the most similar case died of COVID, then we know that the incoming patient is at high risk. Limited resources in a hospital can therefore be allocated to the most severe cases, in times of resource suffocation/lack of adequate resources.

Link to the project

An Enhanced Deep Learning Architecture for the Classification of Cancerous Lymph Node Images

January ‘20 - July ‘20

Abstract: The use of deep learning techniques to diagnose medical disorders has gained increasing popularity in recent times. The unbeatable accuracy of deep learning algorithms often surpasses the performance of human doctors in the diagnosis of diseases like cancer. In line with this trend, this paper proposes a novel architecture aimed at classifying sections of lymph node scans. Analysis of methods to enhance the existing architecture and produce an algorithm that can be used to identify tumor cells is discussed. An enhancing combination of kernel initialization and image preprocessing such as stain channel extraction, color inversion, and hybridization of these has been obtained by contextual and empirical analysis. Incorporating the successful enhancement features with a ResNet-50 leads to a new strategy to identify the presence of metastatic cancer in lymph node patch images. This system can be used to aid human pathologists or independently (with human supervision) for the diagnosis of metastatic cancer from lymph node images.


This project draws heavily on my learnings from deeplearning.ai’s Deep Learning Specialisation, which I completed in Jan 2020.

Link to the paper

Analysis of Deep Learning Architecture for Non-Uniformly Illuminated Images

August ‘19 - December ‘20

Abstract: The use of deep learning to hone image processing techniques has become increasingly popular. Following the success of Convolutional Neural Networks (CNNs) for image classification, they have been tested for various applications. By training CNNs on a dataset with ground truth (light) images and the corresponding darkened version of the images, neural networks can be used for enhancement. This must account for the non-uniform illumination seen in night-time images. A novel method of training a neural network to enhance non-uniformly illuminated images is proposed. Further, the visualization of convolutional features extracted at each layer of the neural network is discussed, to understand which parts of an image helps the neural network identify the object, thereby enhancing its recognition power. The potential application of this system lies in detecting animals in the non-uniformly lit surveillance video, useful to settlements near forest regions, where wild animals pose a threat to the living areas.

Link to the paper

Hospital Location Predictor in London

August ‘19
I used the skills I gained from IBM’s Data Science Professional Certificate program to design a system that predicts the ideal location of constructing a new hospital in London using data science techniques. This project employs data collection (web scraping), data cleaning, data visualisation and K-NN prediction techniques. I’ve written a post on this, which can be found here: https://thegoldencode.wordpress.com/2019/08/31/finding-the-optimal-location-to-establish-a-new-hospital-in-london/

Link to the Github repo

Fuzzy Logic based Heart Condition Predictor

May ‘19
Inspired by the scope of fuzzy logic and its ability to predict the outcome of vague, uncertain scenarios, I worked on the use of fuzzy logic to predict heart disease. I used three variables - age, blood pressure and cholestrol levels. These variables were obtained from a literature review of medical articles, but can easily be modified/augmented, depending on the requirements of the medial professional using the system. The output is the level of risk of suffering a heart attack.

Link to the project

Simple Neural Network Algorithm Implementation

April ‘19
This month I’ve been working on the implementation of neural network algorithms, either from their original papers or from simplified articles. In this endeavour, I’ve implemented activation functions, the perceptron learning algorithm and boolean gates using single / mulit- layer neural network structures.

Link to the project

IDC Cancer Prediction

Feb ’19
Worked in a team of 2 members to refine and outperform the existing neural networks for predicting the presence or absence of breast cancer (IDC) in histology images. Achieved an accuracy of 95.28% using a combination of convolutional layers obtained by trial and error. While I can’t share the code, I’ve linked the project page on my portfolio below:

Link to the project

Tamil Character Recognition System

Jan ‘19
Tamil is one of the oldest languages in the world. I built an extendable prototype neural network that was able to classify 5 characters in Tamil. Achieved an accuracy of > 98%. I used a Tamil character dataset that has been generously open-sourced for the development of technology in Tamil by the University of Jaffna.

Link to the project on my portfolio

Other Projects

Amrita Lab Management System

May ’17 – June ’17
Worked on the PHP backend of a University Lab Management System aimed to integrate the laboratories on campus

Other Coursework projects (on Github)