Artificial intelligence models may outperform traditional regression models in predicting lung cancer risk, according to a new systematic review and meta-analysis of 140 studies.
The findings suggested that artificial intelligence (AI) tools, especially those incorporating imaging data, may offer more precise risk stratification for screening programs.
Investigators analyzed 185 traditional regression models and 64 AI models developed to assess future lung cancer risk in individuals without a current or past diagnosis. Among these, 16 AI models and 65 traditional models had been externally validated. Model performance was measured using the area under the receiver operating characteristic curve (AUC).
AI models demonstrated a pooled AUC of 0.82 across external validations compared with 0.73 for traditional models. Models using low-dose computed tomography (LDCT) scans achieved a pooled AUC of 0.85, indicating improved accuracy when imaging data were included.
Prediction accuracy varied by time frame. For 1-year lung cancer incidence, AI models achieved the highest pooled AUC of 0.91. For 5- to 6-year predictions, AI models had an AUC of 0.79, while traditional models reached 0.75.
AI model performance was also consistent across population types. In the general population, the pooled AUC was 0.84. Among ever-smokers, it was 0.81. Traditional regression models had a pooled AUC of 0.73 for both groups.
Externally validated AI models relied on a range of input data. Half included imaging such as LDCT or chest x-rays. Others were based on epidemiologic risk factors like age and smoking history or clinical biomarkers.
Despite their improved performance, both AI and traditional models were associated with high risk of bias (ROB), particularly in participant selection and analysis methods. Among AI models, 83% were rated as high ROB compared with 66% of traditional models.
AI models were generally more recent, with many developed after 2020. Most were validated in only one or two data sets, often from the U.S. or United Kingdom. In contrast, traditional models like PLCOm2012 have been validated in more than 25 independent data sets.
Direct comparisons within the same data sets further supported AI’s performance advantage. In one study involving ever-smokers, an AI model achieved an AUC of 0.91, whereas a traditional regression model reached just 0.59.
The results suggested AI may help improve lung cancer screening by identifying high-risk individuals more accurately, particularly when imaging data are available. However, the investigators noted that broader validation across diverse populations is needed before AI models can be widely implemented in clinical settings.
The investigators highlighted the need for standardized development practices and improved data transparency to reduce bias. Future research should focus on prospective validations, especially in underrepresented populations, to determine the practical utility of AI-based risk prediction models in real-world clinical workflows.
The authors declared no conflicts of interest.
Source: JACR
