Predicting 5-Year Mortality in Non-Small-Cell Lung Cancer Using the Korean Central Cancer Registry: Model Development...

Area of Science:

• Oncology
• Medical Informatics
• Machine Learning

Background:

• Non-small-cell lung cancer (NSCLC) is a major cause of cancer mortality, necessitating accurate prognostic prediction.
• Machine learning (ML) shows promise for prognosis assessment, but integrating high discrimination with clinical interpretability remains a challenge.

Purpose of the Study:

• To develop deep learning (DL) models for predicting 5-year mortality in NSCLC patients.
• To quantify feature importance using permutation testing for clinical interpretability.

Main Methods:

• Utilized data from 3144 NSCLC patients including clinical, pulmonary function, histological, genomic, and staging details.
• Developed and tuned five DL models using Hyperband, evaluating performance with AUC, accuracy, F1-score, and Brier score.
• Assessed feature importance via groupwise permutation testing and concordance using the Friedman test.

Main Results:

• All five DL models achieved comparable discrimination (AUC=0.875-0.879), with the primary model reaching an AUC of 0.879.
• Tumor stage significantly impacted prediction (AUC decrease of 0.217), followed by pulmonary function tests (0.016).
• Gene mutation importance varied, being higher in adenocarcinoma subsets; feature importance rankings were consistent across models (P=.93).

Conclusions:

• A grouped-input DL framework demonstrated performance comparable to Cox models for NSCLC 5-year mortality prediction.
• Group-level permutation importance offers stable, reproducible insights into risk factors, aiding clinical decision-making and model refinement.