| USA | linkedin.com/in/suresh-akshay/ | github.com/sureshaks |
|---|
About me
My name is Akshay Suresh and my educational background is in Data Science and Computer Engineering. I have always been interested in math and have been coding since I was 16.
Having worked at a highly dynamic startup like Unotech as well as a top consulting firm like BCG, I believe that impact revolves around great teams who understand collaboration, respect and are collectively willing to go the "extra mile".
As of 2025, I work at CVS Health (Aetna) as a Lead Data Scientist. My specialty lies in Collaborative problem-solving.
I am always looking forward to learn from people - people who either want to change the world or who already are...
Education and Experience
2023 - Current: Lead Data Scientist, Aetna, a CVS Health Company
2021 - 2023: Senior Data Scientist, Aetna, a CVS Health Company
2019 - 2021: Data Scientist, Aetna, a CVS Health Company
2018 : Data Scientist Intern, The Boston Consulting Group
2017 - 2019 : M.S. in Data Science, Northeastern University
2015 - 2017 : Software Developer, Unotech Software Pvt. Ltd.
2011 - 2015 : B.E. in Computer Science, University of Mumbai
Projects
X-ray anomaly detection
Overview
MURA (musculoskeletal radiographs) is a large dataset of bone X-rays. Algorithms are tasked with determining whether an X-ray study is normal or abnormal.
Musculoskeletal conditions affect more than 1.7 billion people worldwide, and are the most common cause of severe, long-term pain and disability, with 30 million emergency department visits annually and increasing.
The goal of this project is to develop a model that will accurately identify if there is an anomaly in an X-ray image.
Methodology
The dataset contains more than 40,000 images. There are anomalies across 7 body parts namely: Elbow, finger, forearm, hand, humerus, shoulder and wrist. The preprocessing steps involve:
- Resizing and rescaling.
- Converting 3-channel RGB images to 1-channel grayscale.
- Performing Adaptive thresholding for contrast adjustment.
The model used is a hierarchical convolutional neural network with 2 stages:
- Classify the body part
- Look for abnormalities within the body part
AlexNet and DenseNet are used for stages 1 and 2 respectively. The overall architecture looks like this:
Architecture
Results
Our final results looked like this:
Accuracy
Precision
A technique called Gradient Class Activation Mapping (GradCAM) is used to add interpretability to the model. GradCAM encodes the last layer of activation to understand how the neural network learns from the images.
Gradient Class-Activation mapping
Multi-armed bandit with unobserved confounders
Overview
Consider n-slot machines. Given that the payouts of each slot machine are unknown, the goal is to come up with a sequence of actions that would maximize the reward. The multi-armed bandit problem revolves around developing a strategy in learning the true payout distributions of each of these slot machines. Causality in statistical inference is an upcoming topic in research and it is being widely used to construct a model-based machine learning approach.
This project takes a causal approach to understand how unobserved confounders can affect the performance of these bandit algorithms. A modified version of Thompson sampling (Causal Thompson sampling) takes unobserved confounders into account while learning the payout rates of each arm.
Methodology
Our approach is based on this paper. Consider this scenario, a greedy casino owner observes gamblers and their behavior so as to maximize the profits of the casino. On observation, the casino owner learns that the choice of pulling an arm (of the slot machine) is influenced by 2 factors
- The gambler being drunk
- The machine blinking
Suppose also that a new gambling law requires that casinos maintain a minimum attainable payout rate for slots of 30%. Cognizant of this new law, while still wanting to maximize profits by exploiting gamblers' natural arm choices, the casino executives modify their new slots with the payout rates depicted in this table.
(a) D=0 D=1 B=0 B=1 B=0 B=1 X=M1 0.1* 0.5 0.4 0.2* X=M2 0.5 0.1* 0.2* 0.4 D=1: gambler is drunk P(D=1) = 0.5
B=1: blinking machine P(B=1) = 0.5
d ⊕ b: choice of arm
The value in each cell is the payout probability of each arm. For example, 0.1 means that the machine has a 10% chance of payout.
We've implemented a Causal Thompson Sampling which takes into account these unobserved confounders (B and D), compare it with standard Thompson sampling and observational data.
Results
Cumulative regret and probability of optimal action are used to evaluate these algorithms. Causal Thompson sampling visibly outperforms observational and standard Thompson sampling.
Blindness detection
Overview
Imagine being able to detect blindness before it happened! Diabetic retinopathy is a condition that affects the blood vessels in the light-sensitive tissues called the retina. People with all types of diabetes are at risk of this condition. Millions of people suffer from diabetic retinopathy, the leading cause of blindness among working-aged adults. Our goal is to build a machine learning model to speed up disease detection.
Methodology
Our training data consisted of a little more than 3500 images of various resolutions and sizes. The preprocessing steps involve:
- Rescaling and standardization
- A technique called Contrasted Limited Adaptive Histogram Equalization for Image Enhancement
3 candidate models are fitted to the dataset and the best one is used in the test set.
Architecture
Results
F1-score is used for model evaluation. All the models perform well but Deep FishNet outperforms AlexNet and VggNet in both training and validation sets.
AlexNet Confusion Matrix
Training set
Predicted
F1 Score
0
1
Actual
0
1404
50
0.93
1
156
1319
Validation set
Predicted
0
1
Actual
0
336
15
0.92
1
44
338
VggNet Confusion Matrix
Training set
Predicted
F1 Score
0
1
Actual
0
1444
10
0.95
1
140
1335
Validation set
Predicted
0
1
Actual
0
336
15
0.93
1
39
343
Deep FishNet Confusion Matrix
Training set
Predicted
F1 Score
0
1
Actual
0
1434
20
0.96
1
100
1375
Validation set
Predicted
0
1
Actual
0
340
11
0.96
1
22
360
Confusion matrix for AlexNet, VggNet and Deep FishNet
Libraries
Rcane
Rcane which is an implementation of different gradient descent algorithms used in regression models.
d3-charts
D3-charts is a higher abstraction to embed basic graphs and charts in web pages using d3.js.
CaraML
CaraML is a machine learning library containing simple and intuitive implementations of various machine learning algorithms.