ATS Exclusive Interview: How A Pulmonologist Uses Machine Learning to Predict COPD Mortality

Dr. Matthew Moll, MD, wants to disrupt the lung disease landscape with machine learning and public health. His research abstract titled Improved COPD mortality prediction using machine learning in the COPDGene and ECLIPSE studies,” will be presented at the 2019 American Thoracic Society (ATS) International Conference in Dallas, Texas. Doximity had the pleasure of speaking with Dr. Moll on his research and his thoughts on future clinical and research implications for lung diseases.

Doximity: Thank you so much for taking the time to speak with me about your work. Can you please tell me about what you are currently doing in Boston?

Matthew Moll: Of course. I am a third-year fellow in the Pulmonary and Critical Care Medicine Program at Brigham & Women’s Hospital and also pursuing my Master of Public Health at Harvard T.H. Chan. I’m interested in using machine learning and computational methods to improve our understanding of disease.

Dox: Excellent. And have you been to the ATS conference before? What sessions/topics were you most interested in this year?

MM: Yes, twice before. It’s a huge conference with a lot to offer for everyone depending on career stage. When I was a resident and early fellow, I loved the “Clinical Year in Review” sessions and case presentations discussed by master clinicians. At this point in my career, I’m very interested in risk prediction, the use of genomics to understand the pathophysiology of major lung diseases, and new therapeutic targets. At this year’s ATS conference, I am most interested about endobronchial valve treatments for patients with advanced emphysema, “Lung-omics,” and COPD risk prediction.

Dox: What do you think are the most pressing topics right now in pulmonary and critical care medicine?

MM: We need to better educate the public about lung diseases. Conditions like heart attack, stroke, diabetes are household words. Unfortunately, COPD, pulmonary fibrosis, and pulmonary hypertension are not as well-known, leading to less research funding efforts. And there is still stigma against people with smoking-related lung diseases. COPD is the third leading cause of death globally due to smoking, air pollution, and exposure to biofuels, such as from cooking.

Dox: How long have you been researching COPD and machine learning?

MM: My interest in machine learning came first, which I’ve been studying for about seven to eight years. I’ve been researching COPD for about a year now. The disease interests me because of its high global burden and how it affects all people, regardless of socioeconomic status. There has been many years of research, but few great therapies developed. We just don’t know who will progress with the disease, and who will stay the same.

Dox: How is COPD currently defined and diagnosed?

MM: Currently, COPD is defined by the GOLD criteria, which is based entirely on the FEV1 and the presence of obstruction as determined by the ratio of FEV1/FVC. There is the potential for new definitions, which would include patients who are earlier in the disease process, so that we can start investigating who progresses, and how we may be able to help them. This would be groundbreaking since there isn’t much success treating late in the disease. However, as of now, these patients are not included in current definitions.

Dox: What do you think are the shortcomings of the BODE index and how COPD mortality is currently predicted? 

MM: First of all, for the simplicity of the BODE index, it is actually quite remarkable – there have been many models since BODE, which have all failed to outperform it. However, all tools have limitations; one issue is that the BODE index was developed for predicting four year mortality, while clinically, we may be more interested in five year mortality for transplant referral or six month mortality for palliative care consideration. Additionally, BODE was developed before the availability of quantitative imaging features, which can capture aspects of the disease such as pulmonary vascular pathology, airway wall disease, and degree of emphysema.

Dox: What can other pulmonary or critical care clinicians apply from your research? 

MM: In our study, we were interested in whether Machine learning methods applied to an expanded set of clinical and quantitative CT imaging features will 1) identify the most important predictors of mortality, and 2) improve mortality prediction in moderate-to-severe COPD subjects (GOLD 2-4) compared to BODE, updated BODE, e-BODE, and ADO. Our model outperformed all of these other mortality prediction indices in both the COPDGene and ECLIPSE cohorts. In terms of what this means for clinicians, we have built an online web app for clinicians and researchers, in which they can plug in a hypothetical patient’s values into the calculator, and the calculated survival and survival function will be plotted. This allows clinicians to see how different clinical variables can impact the survival of such a patient. In the future, I think we can improve the model by modelling cause-specific mortality, including more imaging features, and possibly -omics or biomarker data. 

Dox: How do you see machine learning being used in treating other pulmonary/critical care diseases?

MM: I think machine learning will become increasingly important in diagnosing and treating pulmonary diseases. It can help us identify disease endotypes, discover mechanistic link for drug targets, and personalize our therapeutic approaches. My hope is that the tools we develop will be linked directly to the electronic medical record so that our new tools improve, rather than diminish, clinical efficiency.

Dox: How can clinicians with an interest in machine learning acquire this skill to enhance their practice?

MM: I think that for clinicians, the goal would be to understand the advantages and limitations of machine learning methods. This would allow them to understand how to use the tools developed with machine learning methods for their patients – and maybe more importantly – how not to use the tools. These days, there are machine learning articles in top-tier clinical journals, and I would suggest clinicians read them. It’s important to look up the authors of these articles and put their commentary in context; for example, some of these articles have been authored by some of the pioneers in machine learning methods like Andrew Beam and Isaac Kohane, and I would seek out those articles.

Dox: Great, I am sure there is much to learn and uncover in the opportunities of machine learning applications in Pulmonology. Thank you for your thoughtful insights.

This interview was conducted by Angelica Recierdo, Op-Med Editor

Image: LuckyStep / shutterstock

More from Op-Med