WombWatch

Inspiration

The United Nations' Sustainable Development Goal (SDG) 3 aims to "ensure healthy lives and promote well-being for all at all ages." A key target under this goal is to reduce neonatal and maternal mortality
The devastating reality is that preventable deaths of newborns and pregnant women still occur at alarming rates.
Globally, a woman dies during pregnancy or childbirth every two minutes, and fetal mortality remains a major challenge in many countries, including the U.S., where the fetal mortality rate was 5.73 per 1,000 live births in 2021
Inspired by the urgency to address this global health crisis, our project set out to predict fetal mortality, with the hope that earlier detection and better medical interventions could improve outcomes for at-risk pregnancies

Our fetal mortality prediction tool uses cardiotocography (CTG) data to assess fetal well-being and predict the risk of adverse pregnancy outcomes
The tool processes real-time CTG recordings, which include fetal heart rate, uterine contractions, and fetal movements, to detect patterns that might indicate fetal distress or a higher risk of mortality.

We obtained a dataset of cardiotocography (CTG) recordings that had been classified by expert obstetricians into three fetal health categories: normal, suspect, and pathological.
From the raw CTG data, we extracted important features, such as fetal heart rate variability, contraction frequency, and acceleration patterns.
After testing various algorithms, we decided to use K-Nearest Neighbors (KNN) due to its simplicity and efficiency in classifying health conditions based on proximity in feature space. KNN performed well in identifying high-risk pregnancies based on the historical CTG patterns of past cases.
The model was trained on the labeled CTG data, with the goal of classifying new CTG readings as either indicative of a normal pregnancy, a suspect case, or a pathological condition that could lead to fetal mortality.
We used cross-validation techniques to ensure the robustness of our model and tested its predictive accuracy against an unseen test dataset.