A recent study published in Cancers conducted a meta-analysis to evaluate artificial intelligence (AI) potential in early lung cancer detection.
Study: AI-Driven Models for Diagnosing and Predicting Outcomes in Lung Cancer: A Systematic Review and Meta-Analysis. Image Credit: metamorworks/Shutterstock.com
Background
Lung cancer is a global health issue with high mortality rates due to late diagnosis. Current techniques for early detection include computed tomography (CT) scans; however, benign lesions and radiologists experience impact performance. Innovative strategies are required to improve prognosis and survival rates.
AI models could improve these approaches by enhancing precision and efficiency, reducing false positive and negative cases, and providing complementary techniques to existing ones.
AI-assisted diagnostic systems in healthcare, particularly for lung cancer, may enhance diagnostic accuracy, stability, and work efficiency.
About the study
In the present meta-analysis, researchers assessed the efficacy of AI models in the early identification of pulmonary cancer, highlighting their potential for improved diagnostic accuracy and analyzing their strengths, limitations, and comparative advantages over traditional methods.
The team searched the PubMed, Science Direct, Embase, and Google Scholar databases to retrieve relevant records published in English through October 2023.
Two researchers independently screened records using predetermined criteria to select high-quality studies and resolved discrepancies by consensus or consulting a third researcher.
The team included original research articles of studies assessing AI performance for detecting lung cancer at an early stage and reporting the findings as performance metrics such as specificity, sensitivity, and accuracy.
They excluded studies with inadequate information on the performance of AI models, commentaries, conference abstracts lacking primary information, and reviews.
The team extracted data on the study setting, design, AI model used, data source, performance metrics, validation method, and outcomes.
They followed the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines for analysis and evaluated study quality using the Quality Assessment of Diagnostic Accuracy Studies (QUADAS-2) tool.
The bias risk domains assessed in the studies concerned patient selection, index test, reference standard, flow, and timing, reported as low, unclear, or high risk.
The researchers assessed heterogeneity in the studies using the I2 statistic and chi-square tests. They performed random-effects modeling for the meta-analysis and analyzed forest plots of pooled diagnostic metrics for AI models.
Results
Initially, the team identified 1,024 records, assessed 116 for eligibility, and excluded 326 duplicates and 28 studies published in non-English languages.
Only 39 records met the eligibility criteria, showcasing diverse AI model applications for lung cancer detection and highlighting varying strengths among studies.
The study reveals artificial intelligence potential for diagnosing lung cancer in the initial stage, with a pooled sensitivity of 0.87 and a specificity of 0.87, indicating high accuracy in identifying true positives and negatives.
However, the team observed heterogeneity in the studies due to differences in study populations, data sources, and model specifications. The team found low bias risks in patient selection, index tests, and reference standards but higher bias risks in flow and timing.
The studies showed that AI models, particularly recurrent neural networks (RNN) and convolutional neural networks (CNN), can improve lung cancer prediction accuracy, decreasing false positives and lowering the impact of missing data.
Other AI models used included DL, DBN, machine learning (ML), logistic regressions (LL), random forest classifiers (RF), naïve Bayesian systems (NBS), Bayesian networks (BN), and decision trees.
A study from China used three-dimensional deep learning models on computed tomography scans, achieving sensitivity, specificity, and overall diagnostic accuracy of 75%, 82%, and 89%, respectively.
Another study in China used the Support Vector Machine-Least Absolute Linkage and Selection Operator (SVM-LASSO) on lung image database consortium (LIDC)-image database resource initiative (IDRI) data and attained 85% accuracy, 12% higher than Lung-Reporting and Data System (RADS).
Another study conducted in China used a three-dimensional customized mixed link network (CMixNet) for data obtained from the LIDC-IDRI and lung nodule analysis (LUNA-16) datasets, achieving sensitivity and specificity of 94% and 91%, respectively, demonstrating better results than existing techniques used to detect lung cancer.
The findings highlighted the trade-offs between specificity and sensitivity and the strengths of AI-based techniques.
While a few studies highlighted AI’s potential to overcome particular challenges, others emphasized the reliability and efficiency of artificial intelligence in pulmonary cancer screening, benefiting healthcare professionals and patients.
Conclusions
The study findings showed that AI models effectively detect early lung cancer, identify positives and negatives, and improve prognosis.
However, heterogeneity in studies underscores the need for standardized protocols. Future research should focus on refining AI models, considering challenges, and collaborating with researchers, clinicians, and policymakers to establish guidelines and standards for AI systems in pulmonary cancer screening.
Addressing these challenges will advance AI technologies, ultimately facilitating early pulmonary cancer diagnosis and prompt management.