Stage Migration, Treatment Patterns, and Machine Learning-Based Prediction of 12-Month Mortality in Lung Adenocarcinoma Across the COVID-19 Era: A SEER Population-Based Study
Related Experiment Video
Updated: Oct 28, 2026
Comparison of Predictive Performance of Three Lymph Node Staging Systems in Colorectal Signet Ring Cell Carcinoma Based on Machine Learning Model
Published on: April 18, 2025
Ömer Emin Çınar1,2, Yasemin Adalı3,4
1School of Electronics, Electrical Engineering and Computer Science, Queen's University Belfast, Belfast BT9 5BN, UK.
Abstract:
Background/Objectives: The COVID-19 pandemic coincided with widespread disruption of cancer care. We examined changes in lung adenocarcinoma presentation and outcomes across the COVID-19 era and whether diagnosis-time information identifies patients at risk of early death. Methods: Adults with microscopically confirmed lung adenocarcinoma diagnosed during 2018-2023 were identified from the Surveillance, Epidemiology, and End Results 17 Registries database and grouped into three diagnostic periods. Logistic regression, time-varying Cox, and Fine-Gray models evaluated stage, metastasis, surgery, delay, and mortality. Four machine-learning algorithms using only diagnosis-time variables were developed (2018-2021; n = 87,281) and temporally validated (2022; n = 22,012). Results: Among 133,306 patients, COVID-19 versus pre-COVID-19 differences were significant but small in absolute terms: stage IV disease 47.9% versus 47.4% (adjusted odds ratio [OR], 1.03; 95% CI, 1.00-1.06), organ-specific metastasis 45.8% versus 44.4% (OR, 1.06; 1.04-1.09), and treatment delay exceeding 90 days 11.1% versus 10.2% (OR, 1.11; 1.06-1.17). Post-COVID-19 (2022-2023) diagnosis showed less advanced disease but substantially longer delay (14.5% versus 10.2%; OR, 1.45; 1.39-1.52). Gradient boosting achieved a 2022 ROC AUC of 0.840 (0.834-0.845); it flagged 46.8% of patients and correctly flagged 81.6% of those who died, against 44.0% and 75.4% for stage IV classification. Conclusions: COVID-era changes in stage and metastatic presentation were small in absolute terms, whereas the increase in treatment delay was larger and persisted. A diagnosis-time model modestly improved identification of early death beyond stage alone, flagging 28 more patients per 1000, of whom 21 died within the year. These findings support further evaluation; external validation is required before clinical use.