Lung cancer screening faces a cost-benefit tension. Lung cancer is the leading cause of cancer-related deaths worldwide, with most cases detected at advanced stages where treatment is primarily palliative. In South Korea, the National Lung Cancer Screening Program targets individuals aged 54-74 with at least 30 pack-years of smoking history, providing biennial low-dose CT scans.
Chest X-ray is the most commonly performed radiologic examination globally and is widely used for health checkups in South Korea and other countries. Although large randomized trials have shown that conventional chest X-ray alone does not reduce lung cancer mortality, AI-based computer-aided detection (CAD) systems have shown promise in improving detection accuracy for this widely available and low-cost modality.
AI integration has significantly improved CAD performance in chest radiography through deep learning, with multiple studies showing that AI-based systems markedly enhance radiologist performance as secondary readers. However, the economic value of integrating AI into lung cancer screening programs remained uncertain, motivating this first formal cost-effectiveness evaluation.
The study compared five mutually exclusive screening strategies: no screening, standard chest X-ray, AI-assisted chest X-ray, low-dose CT, and AI-assisted low-dose CT. The goal was to identify which strategy offered the best value for money across different age groups and smoking profiles in the Korean population.
A combined decision tree and Markov simulation model. The study constructed a mixed model using TreeAge Pro 2022, integrating a decision tree for initial screening strategy selection with a Markov model for lifetime disease progression tracking. Four scenarios were simulated: the current Korean guideline target (ages 54-74), and three expanded populations (ages 40-80, 50-80, and 60-80) aligned with US Preventive Services Task Force recommendations.
Each scenario modeled a hypothetical cohort of 10,000 individuals stratified by smoking status: non-smokers, light smokers (less than 30 pack-years), and heavy smokers (30 or more pack-years). Hazard ratios for lung cancer incidence were applied based on smoking status relative to non-smokers, with heavy smokers carrying a hazard ratio of 29.90.
Key diagnostic performance inputs were drawn from published literature. CXR alone had a sensitivity of 0.470 and specificity of 0.780. AI-assisted CXR increased sensitivity to 0.696 and specificity to 0.940, substantially improving both detection and precision. LDCT achieved the highest sensitivity at 0.810, while LDCT plus AI reached 0.820 -- a modest improvement over LDCT alone.
Quality-adjusted life years were used as the effectiveness measure, with utility weights for each health state (healthy, local disease, regional disease, distant disease, disease-free, and death) drawn from published Korean studies. All costs were expressed in 2022 USD using an exchange rate of 1265 KRW per USD, with a willingness-to-pay threshold of $32,409.9 per QALY, reflecting one times South Korea's GDP per capita.
AI-assisted chest X-ray was cost-effective across all scenarios. The incremental cost-effectiveness ratio for AI-assisted CXR versus CXR alone ranged from $8,679 to $10,030 per QALY gained across the four age-group scenarios, all well below the Korean willingness-to-pay threshold of $32,410 per QALY. This makes AI-assisted CXR the only strategy consistently identified as cost-effective.
Standard CXR alone was less favorable, with ICERs of $10,335 to $19,420 per QALY versus no screening, and the incremental gain over no screening was less consistent. When comparing CXR to no screening, CXR achieved ICERs that varied widely by age group, performing better in older populations where lung cancer incidence is higher.
LDCT-based strategies performed poorly from a cost-effectiveness standpoint. Adding AI to LDCT produced ICERs of $599,724 to $671,780 per QALY versus AI-assisted CXR, far exceeding the willingness-to-pay threshold. The marginal benefit of AI for LDCT was negligible because LDCT already has high baseline sensitivity, leaving little room for improvement that could justify the additional cost.
Average cost-effectiveness ratios for AI-assisted CXR versus no screening ranged from $9,950 to $15,601 per QALY across scenarios, all well within acceptable thresholds. The older age group scenario (ages 60-80) yielded the most favorable ACER of $9,950 per QALY, confirming that targeting higher-risk, older populations maximizes cost-effectiveness.
Age at screening start was the most sensitive parameter. Deterministic sensitivity analysis revealed that the age at which screening begins was the single most influential parameter affecting the ICER. Older screening populations consistently showed more favorable ICERs, reflecting higher underlying lung cancer incidence and therefore greater absolute benefit from early detection.
The initial stage distribution of detected lung cancers (regional and distant) was the second most sensitive parameter. When cancers are detected at later stages, the cost-effectiveness of screening diminishes because treatment costs are higher and survival benefits smaller.
Probabilistic sensitivity analysis using Monte Carlo simulation with 10,000 iterations showed that AI-assisted CXR had a 91.8% probability of being the most cost-effective strategy in the primary Korean guideline scenario (ages 54-74). In the broader age-range scenarios, this probability ranged from 76.2% to 83.9%, confirming robust cost-effectiveness even under parameter uncertainty.
The cost-effectiveness acceptability curve showed that AI-assisted CXR remained the optimal strategy with greater than 50% probability even when the willingness-to-pay threshold was reduced to $140,000, demonstrating strong economic robustness across a wide range of threshold assumptions.
The diagnostic gap explains the economic difference. CXR alone has a sensitivity of only 47%, roughly half that of LDCT at 81%. AI integration raises CXR sensitivity to 69.6%, substantially narrowing this gap. Combined with CXR's much lower cost (approximately $7.30 per scan versus $118.60 for LDCT), this moderate diagnostic improvement generates favorable cost-effectiveness that LDCT cannot match even with AI enhancement.
For LDCT, AI adds only marginal sensitivity improvement (from 0.810 to 0.820), providing insufficient incremental benefit to justify the already high baseline cost. The cost of LDCT scanning substantially influences the ICER, and a threshold analysis found that LDCT screening would become cost-effective only if the scan price were reduced to approximately $128.
LDCT's lower specificity (0.690 versus 0.940 for AI-assisted CXR) also contributes to higher downstream costs from false-positive follow-up investigations. High false-positive rates, overdiagnosis, and radiation exposure concerns are recognized limitations of LDCT that CXR with AI partially avoids through superior specificity.
The AI model in this study was modeled as operating independently for CXR (based on a standalone deep learning algorithm study), while LDCT AI was modeled as assistive to radiologists. This distinction reflects different implementation models: standalone AI for CXR could address radiologist workforce shortages, particularly in settings where radiologist availability is limited.
AI-assisted CXR offers a scalable and affordable national screening option. The findings support integrating AI-based CAD into chest radiography for lung cancer screening as a cost-effective, accessible, and scalable alternative to LDCT-only programs. This approach could extend screening benefits to the general population beyond the current high-risk eligibility criteria.
Key limitations include the model's broader target population compared to the current Korean national program (which only targets heavy smokers aged 54-74), limiting direct comparability. The smoking cessation benefit was also not incorporated -- since mortality rates roughly double when patients fail to quit after early detection, inclusion of cessation effects would likely further improve the cost-effectiveness of any screening strategy.
The specificity parameter for AI-assisted CXR was derived from a single validation study and may be somewhat overestimated compared to broader multicenter evidence. Robust external validation of AI models is critical before national implementation, as the economic value of AI screening is highly sensitive to both sensitivity (which drives early detection benefit) and specificity (which determines false-positive downstream costs).
From a public health perspective, AI-assisted CXR is particularly attractive for countries with limited resources or radiologist shortages where LDCT implementation is not feasible. Digital chest radiography is widely available, low-cost, and minimally invasive, making it an actionable platform for expanding population-level lung cancer screening when enhanced by AI detection algorithms.