AI beats standard regression models at predicting lung cancer risk
When it comes to predicting an individual’s risk of developing lung cancer, artificial intelligence models outperform standard risk assessments, according to a new meta-analysis.
Published in the Journal of the American College of Radiology, the new paper compares the utility of AI-enabled predictions and traditional regression models for identifying individuals who are most at risk of being diagnosed with lung cancer.
Low dose CT lung screenings significantly increase the likelihood that cancer is identified early, which inevitably results in improved outcomes. However, the screening is not necessary for all patients, even if they have a history of smoking. Lung cancer screening referrals are based on a number of factors, including how much an individual has smoked, age, medical conditions and sometimes demographic factors. Risk models that predict the likelihood of someone developing lung cancer also can be used to determine who should be screened for the disease.
But not all models are the same, nor do they all incorporate the same data when assessing patient risk. This is what led researchers to delve into prior studies on risk prediction models—to determine whether certain methods are superior in terms of accuracy.
“Optimization of screening efforts may be achieved through using accurate prediction models to identify individuals most at risk of a future lung cancer diagnosis. Several multivariable risk prediction models have been developed and validated in the literature,” Scott J. Adams, MD, PhD, with the Department of Medical Imaging at Royal University Hospital in Canada, and colleagues noted. “These models seek to accurately stratify individuals into risk categories based on demographic and clinical information to inform screening eligibility and optimal screening intervals.”
The team conducted a systematic review of studies reporting the performance of AI or traditional regression models for predicting lung cancer risk. This resulted in 140 studies that met inclusion criteria, which included performance metrics on 185 traditional and 64 AI-based models.
Out of all the models evaluated, 16 AI models and 65 traditional models had been externally validated. When considering the externally validated models alone, the AI group achieved a pooled AUC of 0.82, while the traditional cohort yielded an AUC of 0.73, respectively.
A subgroup analysis focused on specific factors included in the models’ predictions that may have helped to improve their performance. The team found that AI models incorporating CT imaging yielded the best results, with a pooled AUC of 0.85.
“The results of this analysis may support future integration of AI in lung cancer screening to identify high-risk individuals for screening and to personalize screening intervals based on lung cancer risk,” the authors wrote.
While the prospect of AI improving risk prediction in lung cancer settings is promising, the team acknowledged that issues with bias remain (this also is the case with traditional models). In the future, the group suggested that AI model development and validation focus on incorporating more diverse datasets to enhance performance.
Learn more about the findings here.
In addition to her background in journalism, Hannah also has patient-facing experience in clinical settings, having spent more than 12 years working as a registered rad tech. She began covering the medical imaging industry for Innovate Healthcare in 2021.