Hybrid Neural Networks for Mortality Prediction from LDCT Images

Deep learning predicts all-cause mortality from longitudinal total-body DXA imaging

Background

Mortality research has identified biomarkers predictive of all-cause mortality risk. Most of these markers, such as body mass index, are predictive cross-sectionally, while for others the longitudinal change has been shown to be predictive, for instance greater-than-average muscle and weight loss in older adults. And while sometimes markers are derived from imaging modalities such as DXA, full scans are rarely used. This study builds on that knowledge and tests two hypotheses to improve all-cause mortality prediction. The first hypothesis is that features derived from raw total-body DXA imaging using deep learning are predictive of all-cause mortality with and without clinical risk factors, meanwhile, the second hypothesis states that sequential total-body DXA scans and recurrent neural network models outperform comparable models using only one observation with and without clinical risk factors.

Methods

Multiple deep neural network architectures were designed to test theses hypotheses. The models were trained and evaluated on data from the 16-year-long Health, Aging, and Body Composition Study including over 15,000 scans from over 3000 older, multi-race male and female adults. This study further used explainable AI techniques to interpret the predictions and evaluate the contribution of different inputs.

Results

The results demonstrate that longitudinal total-body DXA scans are predictive of all-cause mortality and improve performance of traditional mortality prediction models. On a held-out test set, the strongest model achieves an area under the receiver operator characteristic curve of 0.79.

Conclusion

This study demonstrates the efficacy of deep learning for the analysis of DXA medical imaging in a cross-sectional and longitudinal setting. By analyzing the trained deep learning models, this work also sheds light on what constitutes healthy aging in a diverse cohort.

Body composition – the overall proportion of fat, muscle, and bone in one’s body – has been associated with mortality. It is important to better understand the relationship between body composition and mortality as changing body composition is an important goal of many drug and lifestyle interventions. Here, we combine medical images used for body composition measurement directly with information from the medical history of a large number of people to predict mortality. We use machine learning, which relies on mathematical models that extract useful features from images and use these to predict an outcome. Our findings show that combining body composition imaging with traditional mortality risk factors improves the prediction of mortality. This may help clinicians to more accurately predict who is at risk of dying in the future and target these patients with appropriate interventions.

Glaser et al. develop a deep learning system to predict all-cause mortality from total-body DXA scans. Their best predictive model integrates longitudinal body composition data with traditional mortality risk factors.

The Future of Concurrent Automated Coronary Artery Calcium Scoring on Screening Low-Dose Computed Tomography

Low-dose computed tomography (LDCT) has been extensively validated for lung cancer screening in selected patient populations. Additionally, the use of gated cardiac CT to assess coronary artery calcium (CAC) burden has been validated to determine a patient's risk for major cardiovascular adverse events. This is typically performed by calculating an Agatston score based on density and overall burden of calcified plaque within the coronary arteries. Patients that qualify for LDCT for lung cancer screening commonly share major risk factors for coronary artery disease and would frequently benefit from an additional gated cardiac CT for the assessment of CAC. Given the widespread use of LDCT for lung cancer screening, we evaluated current literature regarding the use of non-gated chest CT, specifically LDCT, for the detection and grading of coronary artery calcifications. Additionally, given the evolving and increasing use of artificial intelligence (AI) in the interpretation of radiologic studies, current literature for the use of AI in CAC assessment was reviewed.  We reviewed primary scientific literature dating up to April 2020 using Pubmed and Google Scholar, with the search terms low dose CT, lung cancer screening, coronary artery calcium, EKG/cardiac gated CT, deep learning, machine learning, and AI. These publications were then independently evaluated by each member of our team. Overall, there was a consensus within these papers that LDCT for lung cancer screening plays a role in the evaluation of CAC. Most studies note the inherent problems with the evaluation of the density of coronary calcifications on LDCT to give an accurate numeric calcium or Agatston score. The current method of evaluating CAC on LDCT involves using a qualitative categorical system (none, mild, moderate, or severe). When performed by cardiac imaging experts, this method broadly correlates with traditional CAC score groups (0, 1 to 100, 101 to 400, and > 400). Furthermore, given the high sensitivity of a properly protocolled LDCT for coronary calcium, a negative study for CAC precludes the need for a dedicated gated CT assessment. However, qualitative methods are not as accurate or reproducible when performed by general radiologists. The implementation of AI in the LDCT screening process has the potential to give a quantifiable and reproducible numeric value to the calcium score, based on whole heart volume scoring of calcium. This more closely aligns with the Agatston score and serves as a better guide for treatment and risk assessment using current guidelines. We conclude that CAC should be assessed on all LDCT performed for lung cancer screening and that a qualitative categorical scoring system should be provided in the impression for each patient. Early studies involving AI for the assessment of CAC are promising, but more extensive studies are needed before a final recommendation for its use can be given. The implementation of an accurate, automated AI CAC assessment tool would improve radiologist compliance and ease of overall workflow. Ultimately, the potential end result would be improved turnaround time, better patient outcomes, and reduced healthcare costs by maximizing preventative care in this high-risk population.