Loading Events

« All Events

  • This event has passed.

Michael Lynn’s MS in Statistics Project Presentation “Predicting COPD Severity: A Deep Learning Approach”

May. 2 @ 12 p.m. - 2 p.m.

Join us for Michael Lynn’s MS in Statistics Project Presentation “Predicting COPD Severity: A Deep Learning Approach” on Thursday May 2 from 12-2 pm in Student Commons Building 4017 or on Zoom. Email mathstats-staff@ucdenver.edu for the link.


Title: Predicting COPD Severity: A Deep Learning Approach

Abstract: Numerous studies have shown that Chronic Obstructive Pulmonary Disease (COPD) patients have a higher risk of dementia (Wang et al., 2022), leading to a need for more research into how pulmonary function affects cognition, especially for smokers. The Mini-Cog test, consisting of a clock drawing component and a word recall component, is a validated tool to identify individuals at high risk of cognitive impairment. The clock test portion involves the patient being asked to draw 11:10 on an analog clock face. The drawn clock is then manually scored on a pass or fail basis by a clinical psychologist, with fail indicating the individual is at high risk of having cognitive impairment. The COPDGene project is a 15-year longitudinal study of nearly10,000 heavy smokers with COPD. Approximately 2,500 of the study participants have taken the Mini-Cog and also have detailed measurements for four specific forms (known as subtypes) of COPD: emphysema, airway disease, gas trapping, and hyperinflation. Previous research demonstrated that neural network models can score these COPDGene participants’ Mini-Cog clocks similarly to the manual evaluation done by the clinical psychologists, suggesting that these types of machine learning models can leverage hidden information in the clocks related to cognition. Subsequent work, which focused only on the failed clocks, used neural networks and simple, unsupervised algorithms to discover clusters of clocks that were later found to be associated with the severity of the four COPD subtypes. However, the work did not directly study if and how the clocks predict the individual subtypes, did not utilize the passing clocks, and did not account for known potential confounders of any link between cognitive impairment and pulmonary function. In the present work, I use neural networks, including convolutional neural networks, to build deep learning models that simultaneously predict the severity of four different types of COPD from the Mini-Cog clocks, after accounting for known confounders such as age. In this way, I identify stronger associations between the clocks (theoretically cognitive impairment) and specific COPD subytypes. More, I compare the performance of models that use the pass/fail variable with models that use the full clock images to establish that there is more nuanced information about COPD in the clocks than a broad pass/fail label. I found that models using the raw images do perform better based on reduced Mean Squared Error (MSE), with a stronger effect on predicting emphysema than airway disease. As a result, there is reason to believe that the raw clock images contain more information about COPD severity than just the pass/fail score, and, importantly, this information is evidence that emphysema is a solid candidate as the mechanism between pulmonary function and cognitive impairment.


May. 2
12 p.m. - 2 p.m.
Event Type:
College of Liberal Arts and Sciences
Mathematical and Statistical Sciences