A Growing Clinical Challenge. Nearly 15 million Americans are now eligible for annual lung cancer screening with low-dose CT imaging following expanded 2021 guidelines. As CT use grows, incidentally discovered pulmonary nodules -- spots on the lung that may or may not be cancerous -- are becoming increasingly common, creating a substantial diagnostic burden.
The Cost of Getting It Wrong. Currently 10-15% of surgical lung resections are performed for benign (non-cancerous) disease. These unnecessary surgeries expose patients to operative risk without therapeutic benefit. Conversely, misclassifying a cancerous nodule as low-risk delays diagnosis and treatment, worsening outcomes for potentially curable early-stage lung cancers.
Current Classification Framework. The American College of Chest Physicians guidelines categorize pulmonary nodules as low-probability (below 5% malignancy risk), intermediate-probability (5-65%), or high-probability (above 65%). Accurate classification is essential to guide appropriate workup and avoid both over-treatment of benign nodules and under-treatment of malignant ones.
AI as a Potential Solution. The Lung Cancer Prediction (LCP) Score is an FDA-cleared AI-based radiomic platform that uses quantitative CT imaging features to estimate malignancy probability. A 12-reader study showed that AI assistance improved accurate nodule classification compared to clinician reading alone. However, the cost-effectiveness of this AI tool had not been quantified.
Framework and Perspective. Researchers constructed a decision analysis model comparing AI-assisted nodule risk stratification versus clinician assessment alone. The model was built from a payer perspective with a lifetime horizon, meaning it captures all healthcare costs and survival outcomes over a patient's lifetime rather than just short-term effects.
The Base Case Patient. The model's base case was a 60-year-old medically operative candidate with a 1.1 cm incidentally discovered pulmonary nodule. A malignancy prevalence of 65% was assumed, reflecting the higher-risk population typically seen in dedicated pulmonary nodule and thoracic surgery clinics.
Model Structure. The decision tree modeled each possible pathway following initial risk classification: surveillance, PET-CT evaluation, minimally invasive surgical lobectomy or wedge resection, or combinations thereof. Risk stratification probabilities for both clinician-alone and AI-assisted approaches were derived from a published 12-reader validation study of the LCP Score.
Primary Outcome Measure. The primary outcome was cost per life-year gained (LYG), with a standard willingness-to-pay (WTP) threshold of $100,000 per life-year used to define cost-effectiveness. The incremental cost-effectiveness ratio (ICER) comparing AI-assisted to clinician-alone strategies was the key comparative metric.
Cost Parameters. All costs were derived from Centers for Medicare and Medicaid Services (CMS) reimbursement rates and converted to 2021 US dollars. The AI tool cost was modeled at $650 per use (the CMS reimbursement rate). Surgical costs included $8,821 for wedge resection and $12,944 for lobectomy. Downstream cancer treatment costs were substantial, ranging from $192,182 for stage II-IIIA adjuvant therapy to over $420,000 for stage IV chemotherapy.
Cancer Treatment Assumptions. Anticancer medication costs were based on NCCN guidelines for adenocarcinoma without targetable mutations. Stage I cancers required no adjuvant therapy. Stage II-IIIA patients received cisplatin, pemetrexed, and pembrolizumab. Unresectable stage III patients received carboplatin, paclitaxel, and durvalumab. Stage IV patients received pembrolizumab-based regimens costing up to $420,088.
Survival Outcomes by Stage. Life years were calculated using established actuarial methods. Benign disease patients had a life expectancy of 16.44 years from the time of diagnosis. Stage I lung cancer patients had 16.32 life years, while stage II patients had 8.42, stage III resectable 5.04, and stage IV 3.34 life years -- reflecting the strong survival benefit of early cancer detection.
Comprehensive Sensitivity Testing. Researchers conducted one-way sensitivity analyses varying each parameter by at least plus or minus 20%. Probabilistic sensitivity analysis used 400 Monte Carlo simulation iterations with prespecified parameter distributions. A two-way sensitivity analysis examined FDG-PET specificity and use frequency, accounting for geographic variation in endemic fungal lung disease that reduces PET specificity.
Base Case Results. In the primary analysis with 65% malignancy prevalence, AI-assisted risk stratification resulted in 0.03 additional life years compared to clinician-alone evaluation. The incremental cost was only $114. This yielded an ICER of $4,485 per life-year gained -- far below the standard $100,000 willingness-to-pay threshold, confirming strong cost-effectiveness.
The Critical Malignancy Threshold. The most influential factor in the model was malignancy prevalence. As pre-test probability of malignancy decreased, the ICER rose. AI support remained cost-effective at the $100,000/LYG threshold down to a malignancy prevalence of 5%. Below 5% prevalence, AI support exceeded the willingness-to-pay threshold and was no longer cost-effective.
Clinician Accuracy Matters. When clinician accuracy was varied while holding AI accuracy constant, AI support became less cost-effective as clinician performance improved. AI support exceeded the $100,000/LYG threshold when clinicians correctly identified malignant nodules as high-risk more than 65% of the time. In the baseline reader study, clinicians achieved this only 47% of the time, indicating substantial room for AI to add value.
Probabilistic Analysis. The probabilistic sensitivity analysis confirmed the findings: AI assistance was cost-effective in over 50% of model iterations when the willingness-to-pay threshold exceeded $4,058. At the standard $100,000/LYG threshold, AI assistance was cost-effective in 71% of iterations, indicating robust cost-effectiveness across a wide range of parameter uncertainty.
Matching AI Use to Clinical Setting. Reported malignancy rates range from 3% for all incidentally discovered nodules to 60-80% in dedicated pulmonary nodule and thoracic surgery clinics. Since AI is cost-effective at malignancy prevalence of 5% or higher, it is clearly appropriate for specialist settings but requires more careful assessment in general screening populations where prevalence may approach the 5% threshold.
Lower Clinician Accuracy Benefits More from AI. The cost-effectiveness of AI support increases as the gap between clinician and AI accuracy widens. In settings with limited access to subspecialty-trained pulmonologists and thoracic surgeons, AI is likely to accelerate appropriate referrals for patients with likely malignant pathology. In high-volume specialist centers with very accurate clinicians, the marginal benefit of AI is smaller.
Geographic Considerations. In regions with endemic fungal lung disease (such as histoplasmosis in the US Midwest), FDG-PET scans have reduced specificity, leading to more false-positive results. In these settings, AI-assisted risk stratification becomes increasingly cost-effective because more accurate upfront classification reduces the need for PET scanning and its associated misclassification.
Unaccounted Downstream Costs. The model did not include costs of treating disease recurrence after primary therapy. Since delayed diagnosis leads to later-stage disease with much higher treatment costs, incorporating these downstream costs would likely make AI support appear even more cost-effective than the current analysis shows.
Guideline-Based Model Assumptions. The model assumes guideline-based care, which does not fully reflect real-world practice. Significant variation exists in how clinicians manage pulmonary nodules, and practice often deviates from guidelines. The model was intentionally idealized to assess the ceiling of AI benefit under optimal conditions.
Accuracy Estimates From a Single Study. Clinician and AI accuracy parameters were derived from a single 12-reader study analyzing 300 cancer-enriched nodules. This cohort was not a consecutive clinical cohort, meaning real-world performance of both clinicians and AI could differ in varying settings. Prospective trials in consecutive clinical cohorts are needed.
The Behavior Change Problem. The cost-effectiveness of AI-supported risk stratification entirely depends on whether AI recommendations actually change clinical management. If clinicians receive AI scores but do not modify their decisions, the improved accuracy provides no survival benefit. A pragmatic clinical trial is needed to measure how AI tools actually influence clinician behavior and patient outcomes.
Institutional Costs Not Captured. The analysis reflects the payer perspective and does not include institutional costs such as AI platform licensing fees or implementation expenses. These real-world adoption costs could affect the calculus for healthcare systems evaluating whether to deploy AI nodule evaluation tools.
A Clear Cost-Effectiveness Finding. This study provides the first formal cost-effectiveness analysis of AI-assisted pulmonary nodule risk stratification, finding it cost-effective at an ICER of $4,485/LYG in the base case -- well below standard healthcare value thresholds. This finding supports broader adoption of AI nodule evaluation tools in appropriate clinical settings.
The Stakes of Accurate Evaluation. With 10-15% of lung resections currently performed on benign tissue and the survival difference between stage I and stage IV lung cancer representing over 13 years of life expectancy, the imperative for accurate nodule risk classification could not be greater. AI tools that reduce both unnecessary surgeries and delayed diagnoses serve both patient safety and population health goals.
The Promise of Automation. As indeterminate pulmonary nodules continue to be identified at increasing rates through expanded screening programs, the clinical burden of manual risk stratification will grow. AI tools used in an automated or semi-automated workflow could reduce physician burden while maintaining or improving classification accuracy -- a scalable solution to a growing problem.
Next Steps. Clinical trials should test AI performance across diverse clinical contexts varying nodule size, shape, location, cancer prevalence, and fungal disease rates. As AI becomes more deeply integrated into clinical practice, ongoing evaluation of real-world performance and cost-effectiveness is essential to ensure the technology continues to benefit patients.