1 of 28

Predicting Miscarriage Using Vaginal Microbiomes

By: Mariam Elsharkawy, Faith Egbeni, Joann Tran, Kacee Chan, Yuyan Huang, and Gayathri Govind

Lead TAs: Leen Arnaout & Oluwaseun Adegbite

Assistant TA: Anjali Singh

2 of 28

01

Overview

03

02

04

Vaginal Microbiome and Miscarraige

Introduction

Supervised vs. Unsupervised

Data & Algorithms

Results & Analysis

Impact

Real World

Applications

3 of 28

Introduction

4 of 28

What is a Miscarriage?

Miscarriage, or the failure of a pregnancy to continue beyond the 20th week, mostly occur due to the fetus not developing as expected.
10-15% of all pregnancies end in miscarriage
It can be physically challenging for the pregnant person due to a number of factors:

* Inability to clear miscarriage on its own

* Could lead to infertility

Introduction

5 of 28

Causes of Miscarriage

Chromosomal abnormalities cause about 50% of all miscarriages in the first trimester(13 weeks) of pregnancy.
In recent data, research has also shown that one of the most common causes of miscarriages can be dysbiosis, or imbalances of the vaginal microbiome.

Introduction

6 of 28

Vaginal Microbiome

Imbalance of healthy and unhealthy bacteria in vagina increases likelihood of miscarriage

Introduction

7 of 28

Objective

Develop different types of machine learning models to predict the likelihood of a miscarriage based on a patient’s vaginal microbiome

Introduction

8 of 28

Data & Algorithms

9 of 28

Data Used

Data from 2019 Imperial College London study

The data contained DNA sequences from microbiome samples and whether a miscarriage occured
Needed to clean data so we could predict miscarriage

Data & Algorithms

10 of 28

Data Prep

Data & Algorithms

11 of 28

Algorithm:

decision tree

Data & Algorithms

12 of 28

Algorithm:

random forest

Data & Algorithms

13 of 28

Algorithm:

logistic regression

Data & Algorithms

14 of 28

Unsupervised Learning Algorithm:

k means clustering

Data & Algorithms

15 of 28

Results & Analysis

16 of 28

Principal Components Analysis (PCA):

Results & Analysis

Dimensionality-reduction method that is used to reduce large data sets
Transforms a large set of variables into a smaller one that still contains most of the information in the large set.
PCA is only a reliable method if the data are at least interval scaled and approximately normally distributed.

17 of 28

Accuracy Scores

Code: classifier.score(X_test, y_test)
The closer the score is to 1, the more accurately the algorithm can predict the likelihood of miscarriage from microbiome data (true positive / number of samples)
Used to objectively compare different algorithms

Results & Analysis

18 of 28

Accuracy Scores

Results & Analysis

19 of 28

Comparing Algorithms

Results & Analysis

20 of 28

Final Algorithm Choice

Results & Analysis

More Reliable / Consistent

21 of 28

Quantifiable Results

Results & Analysis

22 of 28

Our Best Algorithm

Results & Analysis

Logistic Regression

Most reliable / consistent algorithm
S-curve is good at approximating datasets with binary classifications

23 of 28

Limitations

24 of 28

Limitations & Resolutions

Limitations:

Data bias

-large enough sample size

-information about the patients

Unknown sampling method
Misrepresentation of the population

Resolutions:

Larger sample data set
1000 samples per class
Randomly sampled

Some of the limitations we encountered were a small sample size, insufficient info about the patient demographic, and unknown sampling method which could lead to data bias and misrepresentation.

To combat these issues in future projects, we could use a larger sample data set, at least 1000 samples per class to ensure fair representation and make sure the data used is randomly sampled to properly represent the target population and eliminate sampling bias.

This study contain data bias. It didn’t reflect the diversity of people who can get pregnant well, because we didn’t have a large enough sample size and enough information about the patients. We don’t know the patient ages, demographic information. Which could have lead to bias and misrepresentation of the population. (For example, miscarriages that may be due to genetic, anatomic, or other confounding variable, and not due to bacterial infection.) The study also didn’t mention how they collected the data, whether it’s voluntary, convenience or other sampling method.

25 of 28

Applications

26 of 28

Applications

Understand harmful bacteria in microbiome

Prevent miscarriages

-after running our data through multiple algorithms and now we can use this data to prevent miscarriages as we now know how to predict them because we now have a better understanding on the certain bacterias that are more likely to cause a miscarriage

-We could use this understanding of our data to find out how to avoid these bacterias as a first step of understanding how to prevent miscarriages

-We could also potentially use this data to assist doctors while treating pregnant patients in order to prevent miscarriages

-The eventual goal of all this would be to actually use all this data to prevent miscarriages from happening at all which would have a large impact on society

-This project gave us an opportunity to put our developing skills in machine learning and coding in python to a real world application as we analyzed real data and tested them through various algorithms to find our best result

-due to this exposure we learned more about coding and troubleshooting and what to do in such situations

-we also got exposure to classifying and cleaning data, using the train-test split and stratification, all tools that can help us in further experiences in AI or machine learning

-We ran our data through multiple algorithms and now we can use this data to prevent miscarriages as we now know how to predict them because we now have a better understanding on the certain bacterias that are more likely to cause a miscarriage

-We could use this understanding of our data to find out how to avoid these bacterias as a first step of understanding how to prevent miscarriages

-We could also potentially use this data to assist doctors while treating pregnant patients in order to prevent miscarriages

-The eventual goal of all this would be to actually use all this data to prevent miscarriages from happening at all which would have a large impact on society

27 of 28

Works Cited

PCA: https://builtin.com/data-science/step-step-explanation-principal-component-analysis

https://starship-knowledge.com/pca-vs-linear-regression

Intro miscarriage: https://my.clevelandclinic.org/health/diseases/9688-miscarriage

Logistic regression image:

Source: https://www.tibco.com/reference-center/what-is-logistic-regression

Random forest image:

https://en.wikipedia.org/wiki/Random_forest

Decision tree image:

https://towardsai.net/p/machine-learning/fully-explained-decision-tree-classification-with-python

28 of 28

Thanks!

Do you have any questions?

CREDITS: This presentation template was created by Slidesgo, including icons by Flaticon, and infographics & images by Freepik