Association of a high-speed, knowledge-based AI tool at the primary care level with zero-miss triage and detection of incidental pulmonary nodules.

270

Background: In mainland China, the high accessibility and remarkably low out-of-pocket cost of chest computed tomography (CT) scans (approximately $26 USD) have led to widespread screening and a subsequent surge in the detection of incidental pulmonary nodules (IPNs). However, a severe shortage of thoracic specialists and inadequate specialized training among primary care physicians have created a critical bottleneck. This systemic gap disproportionately impacts older adults, who often lack the capacity to independently navigate complex healthcare networks to seek specialized care, making them highly vulnerable to critical delays in the treatment of high-risk nodules. To address this, we developed a novel artificial intelligence (AI) tool designed to assist primary care providers in identifying high-risk patients for prompt referral, bypassing the need for heavy DICOM image data transfers. Methods: We developed an ultra-fast, text-based AI diagnostic tool utilizing a knowledge-base architecture rather than traditional data-heavy deep learning models. This knowledge-driven approach significantly reduced the dependency on massive datasets, allowing the model to be robustly constructed and trained using a highly curated cohort of only 500 expertly annotated CT reports. The algorithm extracts and interprets radiographic terminology directly from the text to stratify IPN risk. To confirm the tool's clinical reliability and generalizability, it was subsequently tested and validated on a large, independent real-world cohort of over 3,000 cases. Results: By leveraging a knowledge-base design and focusing exclusively on report text, the AI tool demonstrated exceptional computational efficiency, achieving a processing speed of 10 reports per second. In the large-scale validation cohort (n > 3,000), the tool achieved an impressive overall accuracy of 96.3% in risk stratification. Most crucially, the model yielded a false-negative rate of 0%. This zero-miss rate ensures that absolutely no high-risk patients are overlooked or misclassified as low-risk, guaranteeing maximum patient safety in grassroots medical settings. Conclusions: By employing a knowledge-base architecture, this ultra-fast, report-based AI tool achieves high accuracy and a zero-miss rate with minimal training data. By completely eliminating false negatives, it empowers primary care physicians to make reliable referral decisions. This tool is exceptionally meaningful for older adults with cancer; by providing expert-level triage directly at the primary care level, it overcomes their diminished capacity to independently seek out specialists, effectively preventing life-threatening delays in high-risk cases while protecting low-risk older patients from the hazards of overtreatment and unnecessary invasive procedures.