This is a course project for CIS545 Big Data Analytics(Spring 2023) at University of Pennsylvania.

Data Source and Project Objectives

Our machine learning project utilized the “Diabetes 130-US hospitals for years 1999-2008” dataset from the UCI Machine Learning repository. The dataset includes 49 ranging from basic information like gender, age, and weight to specific medical details such as prescribed drugs and diagnoses.

• Target variable is ‘readmitted’, indicating the number of days until a patient’s readmission to the hospital.
• Our object is to predict the readmission status of patients using various machine learning and deep learning model, aiming to identify factors that influence readmission and improve the model accuracy.

Data Preprocessing and Exploratory Data Analysis(EDA)

We started with extensive data preprocessing to handle missing values, normalize different scales of features, and convert categorical variables into a suitable format for modeling.

• Drop some unnecessary columns, like encounter_id, and patient_nbr, which are more like the identifier rather than features.
• We have a lot of categorical features and some categorical data has more than 700 categories, which is not good to be represented by one-hot encoding or ordinary encoding. Therefore, feature selection and composition are vital.
• We use data augmentation, imputation and feature engineering to enhance the dataset’s quality and relevance.
• Utilized the Seaborn to visualize data distribution and feature importance.

Modeling and Comparison

Our team implemented several ML and DL models, including Gradient Boosted Decision Trees, AdaBoost, and neural networks.

• Training technique: Utilized the cross-validation to ensure robustness and avoid overfitting. Additionally, we also implemented the Ensemble method to combine predictions from multiple models, amining to boost overall prediction performance.
• Metrics: Compared different models based on the accuracy, precision, recall, and F1-score.
• Monitor: The training process is monitor by TensorBoard.

Challenges and Solutions

During the project, we encountered several challenges, including imbalanced data, sparse features, and varying feature scales.

• Imbalanced data/Long-tailed distribution: Utilized the SMOTE for balancing the dataset.
• Sparse feature: Implemented L1/L2 regularization to handle sparity.
• Varying feature scales: Utilized the standardization and normalization to solve the varying feature scaling. Dispite these challenges, our models demonstrated significant potential in predicting patient readmission with some features.

Results

Our project successfully developed pridictive models and get some results:

• A Gradient Boosted Decision Tree, achieved an accuracy of 85% and an F1-score of 83%, outperforming traditional classification algorithms.
• The fully connected Neural Network also showed promising results with an accuracy of 82% and an F1-score of 80%.