Integrating Machine Learning and Dynamic Bayesian Networks to Identify the Factors Associated with Subsequent...

Background/Objectives: Intrapulmonary metastasis (IPM) after an initial single primary lung cancer (SPLC) is an adverse follow-up pattern; however, when studying population-based longitudinal records, the determinants remain unclear. We aimed to identify factors associated with subsequent IPM after initial SPLC using artificial intelligence (AI)-driven analytical approaches. Methods: We used Surveillance, Epidemiology, and End Results (SEER) lung cancer records from 2000 to 2019. Adults with at least two records were restricted to those with SPLC at the first record. Outcome at the second record was registry-classified IPM versus persistent SPLC. A machine learning framework based on random forest models was developed using baseline variables, first record characteristics, and the interval between records. Temporal validation was performed by training on cases from 2000 to 2013 and testing on cases from 2014 to 2019. A dynamic Bayesian network (DBN) supported simulated intervention (SI) analyses to estimate model-implied risk ratios (RRs) with 95% confidence intervals (CIs). Results: Among 3450 patients, 361 had registry-classified IPM at the second record. The random forest model achieved an area under the curve (AUC) of 0.852 in internal validation and 0.929 in temporal validation. Surgery and record timing were the leading predictors. The DBN retained surgery as the only direct parent and achieved an AUC of 0.779. SI analyses showed higher IPM probability for pleural invasion level (PL) 3 versus PL 0, RR 1.378 (95% CI, 1.080-1.657). Lobectomy with mediastinal lymph node dissection versus wedge resection lowered the IPM probability, RR 0.378 (95% CI, 0.219-0.636). Conclusions: AI-based time-sequence modeling integrating machine learning and a DBN allowed for the identification of surgery, pleural invasion, and record timing as key factors associated with subsequent IPM classification after initial SPLC. This framework demonstrates the potential of combining predictive and probabilistic dependency modeling to investigate registry-based disease classification patterns, and may support hypothesis generation for future prospective studies.