COVID-19 Mortality Prediction using Shallow Neural Network

This repository contains Assignment 0 Part II from the CSE 676 – Deep Learning course, University at Buffalo.
The project builds, trains, and evaluates a simple Multilayer Perceptron (MLP) model to predict mortality rates from the official CDC COVID-19 Provisional Death Counts dataset.

---

📘 Project Overview

The objective of this assignment was to implement an end-to-end deep learning pipeline from raw data to evaluation, without using high-level frameworks such as Keras.
The dataset includes demographic and geographic statistics related to COVID-19 deaths by jurisdiction in the U.S.

Core Steps

1. Data Import & Preprocessing

   • Read the CDC dataset (.csv file).
   • Handle missing values, normalization, and categorical encoding.
   • Apply quantile binning (qcut) to categorize mortality rates (optional flag --use_qcut).

2. Model Definition – Base MLP

   • Input layer → Hidden layer (ReLU activation) → Output layer (Softmax).
   • Implemented manually in PyTorch with custom forward and backward logic.
   • Optimization via Adam optimizer and Cross-Entropy Loss.

3. Training & Validation Loop

   • Train for a fixed number of epochs (--epochs, default = 40).
   • Evaluate after each epoch to monitor overfitting.
   • Early stopping if validation loss plateaus.

4. Evaluation & Visualization

   • Compute Test Accuracy, Confusion Matrix, and Correlation Heatmap between features.
   • Save training artifacts (model.pt, plots) in the artifacts/ directory.

---

📊 Dataset

File: Provisional_COVID-19_death_counts__rates__and_percent_of_total_deaths__by_jurisdiction_of_residence.csv
Source: CDC Open Data Portal

Key Columns Used

Category	Example Features
Jurisdiction	State / Territory
COVID-19 Deaths	Number of reported deaths
Total Deaths	All-cause deaths
Percent of Total	Share of COVID deaths
Crude Rate	Deaths per 100k population

---

🧮 Model Performance

Metric	Value
Train Accuracy (Epoch 0)	51.42 %
Train Accuracy (Epoch 10)	81.34 %
Train Accuracy (Epoch 20)	89.72 %
Train Accuracy (Epoch 30)	94.72 %
Final Test Accuracy	87.58 % ✅

Confusion Matrix – Test Set

---

⚙️ Usage

1️⃣ Install Dependencies

pip install -r requirements.txt

2️⃣ Run Training

   python3 covid-mlp.py \
  --csv data/Provisional_COVID-19_death_counts__rates__and_percent_of_total_deaths__by_jurisdiction_of_residence.csv \
  --outdir artifacts \
  --epochs 40 \
  --use_qcut

🧠 Insights

• Shallow neural networks can achieve > 85 % accuracy for structured data tasks with limited features.
• Quantile binning helps stabilize training by reducing outliers.
• Proper correlation analysis is crucial to select non-redundant variables.