A Machine Learning and Deep Neural Networks Approach to Diagnosing Idiopathic Pulmonary Fibrosis
Session Number
Project ID: CMPS 11
Advisor(s)
Samuel Hamilton; Feinberg School of Medicine, Northwestern University
Dr. Deborah Rachelle Winter Ph.D., Feinberg School of Medicine, Northwestern University
Discipline
Computer Science
Start Date
20-4-2022 10:25 AM
End Date
20-4-2022 10:40 AM
Abstract
Idiopathic Pulmonary Fibrosis (IPF) is a lung disease with a mean survival of 2-5 years from the time of diagnosis in which alveolar tissue progressively becomes stiff fibrotic scar tissue, reducing breathing capacity and eventually leading to respiratory failure. The use of Machine Learning could predict IPF cases more precisely than the current surgical lung biopsy treatment years before the onset of the disease, allowing doctors to better plan treatments and reduce unnecessary surgeries. This project aimed to design an accurate Machine Learning (ML) and Deep Neural Network (DNN) model to create an assistive tool for IPF diagnosis. A series of common ML models, including Random Forest, Support Vector Machine, Naive-Bayes, and J48, were tested for highest prediction accuracy. The dataset, provided by Peking Union Medical College, was split into a training and testing set, and was also scaled into a range of 0 to 1 for a higher accuracy rate. A Keras DNN binary classification model was created using 2 hidden layers with 7 and 5 nodes, respectively and performed higher than the highest-scoring ML model, J48, with a final accuracy rate of 89%. Currently, a predictive web application utilizing Django is being designed to test the DNN.
A Machine Learning and Deep Neural Networks Approach to Diagnosing Idiopathic Pulmonary Fibrosis
Idiopathic Pulmonary Fibrosis (IPF) is a lung disease with a mean survival of 2-5 years from the time of diagnosis in which alveolar tissue progressively becomes stiff fibrotic scar tissue, reducing breathing capacity and eventually leading to respiratory failure. The use of Machine Learning could predict IPF cases more precisely than the current surgical lung biopsy treatment years before the onset of the disease, allowing doctors to better plan treatments and reduce unnecessary surgeries. This project aimed to design an accurate Machine Learning (ML) and Deep Neural Network (DNN) model to create an assistive tool for IPF diagnosis. A series of common ML models, including Random Forest, Support Vector Machine, Naive-Bayes, and J48, were tested for highest prediction accuracy. The dataset, provided by Peking Union Medical College, was split into a training and testing set, and was also scaled into a range of 0 to 1 for a higher accuracy rate. A Keras DNN binary classification model was created using 2 hidden layers with 7 and 5 nodes, respectively and performed higher than the highest-scoring ML model, J48, with a final accuracy rate of 89%. Currently, a predictive web application utilizing Django is being designed to test the DNN.