A Growing Public Health Challenge. Lung cancer remains one of the most lethal cancers worldwide, with mortality rates continuing to rise. The disease is especially prevalent among middle-aged and older adults - more than half of all non-small cell lung cancer diagnoses occur in people aged 65 or older.
Insidious Onset. Lung cancer tends to develop without obvious symptoms, making early detection difficult and leading to frequent misdiagnosis. This insidious nature makes timely, accurate risk assessment particularly valuable for aging populations who face elevated baseline risk.
Beyond Static Risk Scores. Most existing tools assign a single risk score at one point in time, without modeling how an individual's risk evolves over years or what specific behaviors could shift someone from a high-risk to a lower-risk state. This study addresses that gap by simulating entire risk trajectories and identifying the most effective behavioral interventions to redirect those trajectories.
Health and Retirement Study (HRS). The primary dataset came from the HRS, a large US longitudinal cohort that has tracked over 20,000 adults since 1992 with biennial follow-up surveys. A total of 8,780 samples were used for model training and 3,763 for internal testing.
China Health and Retirement Longitudinal Study (CHARLS). For external validation, the team used CHARLS, which covers 17,000+ adults aged 45 and older from China, followed from 2010 to 2020. This cross-national test of 16,442 samples was critical for demonstrating that the model generalizes beyond its training population.
Key Risk Features. Both datasets supplied demographic and behavioral variables including age, sex, BMI, marital status, education, smoking status and duration, physical activity, alcohol consumption, preventive behaviors, cancer history, and pre-existing lung disease. Careful variable mapping between the two datasets ensured comparable inputs.
Data Quality Steps. Missing values were handled through multiple imputation using iterative linear regression, and the synthetic minority oversampling technique (SMOTE) was applied to correct for the rarity of lung cancer cases in the dataset, preventing the model from simply predicting the majority class.
Four-Tier Risk Classification. A deep neural network (DNN) was first trained to assign each individual to one of four risk states: non-risk, low risk, medium risk, or high risk. These tiers were defined by weight thresholds computed through the DNN's iterative training, not by arbitrary cutoffs.
Architecture and Training. The DNN featured one input layer with 56 nodes, three hidden layers with 22 nodes each, and a single output node. It used the ReLU (rectified linear unit) activation function to capture nonlinear relationships among risk factors. The model was trained with a learning rate of 0.01 and batch size of 32 over 30 episodes.
Validating the Stratification. Kaplan-Meier survival curves confirmed that the four risk groups showed meaningfully different outcomes: the high-risk group had approximately 25% lung cancer incidence at last follow-up, far exceeding the low-risk and non-risk groups. High-risk individuals also had worse overall survival and lung cancer-specific survival, validating that the DNN's classifications captured genuine biological differences.
From Classification to Action. Once risk tiers were established, the team deployed a deep Q-network (DQN) - a form of deep reinforcement learning - to simulate how individuals can move between risk states over multiple time cycles. Unlike a passive predictor, DQN actively identifies the best sequence of behavioral changes to reduce risk.
How DQN Works in This Context. The model's state space consisted of the four risk levels (high, medium, low, non-risk), and the action space comprised modifiable behaviors such as quitting smoking or reducing smoking frequency. The DQN learned which actions generate the greatest cumulative reduction in lung cancer incidence across many simulated cycles.
Technical Components. The DQN combined Q-learning with convolutional neural networks to approximate the optimal action-value function. Experience replay - reusing historical data from past cycles - stabilized training. The model used a discount factor between 0.9 and 1, a learning rate of 0.01, a batch size of 256, and ran for 500 episodes. The framework was built using OpenAI Gym and PyTorch.
Three Transfer Directions. The DQN modeled three possible risk movement directions: risk improvement (moving to lower risk states), risk deterioration (moving to higher risk states or developing lung cancer), and risk maintenance (staying in the same state). This comprehensive mapping reveals not just how to improve, but also which pathways lead to worsening outcomes.
Distinct Survival Curves. Kaplan-Meier analysis of the HRS cohort clearly separated the four risk groups on lung cancer incidence survival, overall survival, and lung cancer-specific survival. The high-risk group had the steepest decline in all three measures, while the non-risk group remained nearly event-free throughout follow-up.
International Consistency. When the same DNN model was applied to the CHARLS cohort from China, the survival curve separations were maintained. This cross-national consistency is important evidence that the model's risk stratification captures universal biological and behavioral drivers of lung cancer rather than population-specific noise.
Smoking as the Dominant Factor. A forest plot of lung cancer incidence across subgroups revealed that smoking had the strongest relationship with risk in the high-risk group, more so than in medium, low, or non-risk groups. This gradient underscores that smoking cessation will have the greatest impact precisely where the stakes are highest.
Optimal Pathways Identified. The DQN model identified the most effective transition sequence for each risk group. For high-risk individuals, the optimal path was: high risk to medium risk, then to low risk, then to non-risk. For medium-risk individuals: medium risk to low risk to non-risk. For low-risk individuals: low risk to non-risk directly.
Dramatic Incidence Reductions. Following the optimal pathway, simulated lung cancer incidence in the high-risk group dropped from 1,060 per 100,000 to 337 per 100,000 - a reduction of 68.2%. In the medium-risk group, incidence fell from 671 to 289 per 100,000, a 56.9% reduction. The low-risk group achieved a remarkable 99.1% reduction by transitioning to non-risk status.
Key Behavioral Levers. The interventions driving these improvements were extending smoking cessation duration (for example, from less than 1 month to more than 5 years) and reducing smoking frequency (from daily smoking to complete quitting). These straightforward, practical strategies were consistently identified as optimal across risk groups.
Continued Improvement After Transition. After applying the simulated optimal pathway, the high-risk group's incidence fell further to 107 per 100,000 and the medium-risk group's to 29 per 100,000 in the follow-on analysis, validating the cumulative benefit of sustained risk-state transition.
Head-to-Head Comparison. The researchers compared their DQN approach against random forest models - a widely used machine learning method - using the same datasets. While random forest identified the same optimal intervention strategies (smoking cessation and frequency reduction), its predictive accuracy was consistently lower.
Performance Gap. DQN models achieved accuracies of 0.917 to 0.949 and AUROC values of 0.906 to 0.927 on external validation. Random forest models achieved accuracies of 0.823 to 0.851 and AUROC values of 0.778 to 0.836 - roughly 10 percentage points lower across all risk groups.
Why the Difference Matters. In a clinical setting, a 10% gap in model accuracy can translate to meaningfully different guidance for individual patients. Higher accuracy means fewer individuals are misclassified into risk groups that don't match their true biology, ensuring the right interventions reach the people most likely to benefit from them.
A New Tool for Prevention Planning. This model shifts the focus from prediction alone to actionable guidance. Rather than simply telling a clinician that a patient is at high risk, the DQN framework specifies which behavioral changes, undertaken in what order, will most efficiently move that patient toward lower risk states.
Smoking Cessation at the Center. Both the DQN and random forest models converged on the same conclusion: extending the duration of smoking cessation - particularly quitting for more than 5 years - and reducing smoking frequency to zero are the most powerful levers available. This strengthens the case for intensive smoking cessation support programs targeted specifically at middle-aged and older adults.
Stratified Intervention Efficiency. Because the model stratifies risk, interventions can be prioritized. High-risk individuals will see the largest absolute reductions in incidence from the same behavioral changes, justifying more intensive outreach and support for this subgroup while still offering meaningful benefits to medium and low-risk individuals.
Observational Data Without Causal Framework. The study is based on retrospective cohort data and does not establish causality between behavioral changes and risk reduction. The authors acknowledge that constructing directed acyclic graphs in future work would strengthen causal claims and rule out confounding explanations.
Limited Biomedical Variables. The current datasets rely primarily on demographic and self-reported questionnaire data. Future models could incorporate genomic mutation data, low-dose CT scan images, laboratory test results, and electronic health records to improve biological plausibility and real-world applicability.
Simulated, Not Observed, Pathways. The risk transfer pathways and their incidence-reduction effects were derived from simulation rather than observed longitudinal transitions. Prospective validation in a real-world cohort, following individuals through actual behavioral changes and risk-state transitions, is a critical next step before clinical deployment.
Broader Applications. The authors suggest that this DQN-based risk pathway methodology is not limited to lung cancer. The same framework could be adapted to other chronic diseases with identifiable risk states and modifiable behavioral factors, potentially offering a generalizable approach to AI-guided prevention medicine.
A Novel Framework for Cancer Prevention. This study demonstrates that deep reinforcement learning - specifically the DQN architecture - can move lung cancer AI beyond passive risk prediction into active prevention planning. By simulating the consequences of behavioral interventions across multiple time cycles, the model provides a roadmap for reducing lung cancer incidence in aging populations.
Validated Cross-Nationally. The model's strong performance on both the US-based HRS cohort and the Chinese CHARLS cohort suggests that its core findings are not artifacts of a single population. The consistency of both the risk stratification and the optimal intervention strategies across two very different demographic contexts adds credibility to the approach.
Precision Prevention Medicine. By combining granular risk stratification with pathway simulation, this research charts a course toward precision prevention - where interventions are matched not just to a disease but to an individual's specific risk level, behavioral profile, and most efficient path to risk reduction. For the millions of middle-aged and older adults at elevated lung cancer risk, this kind of targeted guidance could translate into thousands of preventable diagnoses.