Deep Learning-Based Long Term Mortality Prediction in the National Lung Screening Trial

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Public Health

Background:

• Long-term mortality prediction is crucial for preventative healthcare.
• Traditional methods for mortality risk assessment have limitations.
• National Lung Screening Trial (NLST) data offers a valuable resource for research.

Purpose of the Study:

• To investigate long-term mortality using deep learning.
• To develop models for predicting non-lung-cancer mortality (cardiovascular and respiratory).
• To identify key features in CT scans associated with mortality risk.

Main Methods:

• Utilized a deep learning approach with neural network models (3D-ResNet).
• Trained models on a cohort matched for age, gender, and smoking history from the NLST.
• Integrated 3D CT scan data and clinical information for prediction.
• Employed 3D saliency maps for model interpretation.

Main Results:

• Achieved an Area Under the Curve (AUC) of 0.73 for mortality prediction.
• Outperformed human performance in cardiovascular mortality prediction.
• Obtained an F1 score of 0.60 and a Matthews Correlation Coefficient of 0.38.
• Identified specific thoracic regions on CT scans indicative of mortality risk.

Conclusions:

• Deep learning models can effectively predict non-lung-cancer mortality.
• AI-driven analysis of CT scans can reveal subtle mortality-related features.
• This approach can enhance early detection and guide preventative interventions, reducing patient morbidity.