Performance of an adult-trained AI tool for intracranial hemorrhage detection on head CT in children aged 6-17 years.
Authors
Affiliations (5)
Affiliations (5)
- Department of Radiology, Yale University, New Haven, 330 Cedar Street, TE 2-212, 06520, USA. [email protected].
- Department of Radiology, Texas Children's Hospital, Houston, USA.
- Department of Radiology, Baylor College of Medicine, Houston, USA.
- Department of Radiology, Yale University, New Haven, 330 Cedar Street, TE 2-212, 06520, USA.
- AIDOC Medical, New York, USA.
Abstract
Most commercially available artificial intelligence (AI) tools in radiology are trained and approved for adult use, creating an access gap for pediatric patients. Intracranial hemorrhage (ICH) detection is a common adult AI application without pediatric FDA clearance. To evaluate the performance of an FDA-cleared, adult-trained AI tool for ICH detection on non-contrast head CT (NCHCT) in pediatric patients aged 6-17 years. This retrospective, multi-institution study analyzed consecutive pediatric NCHCTs performed between January 2017 and November 2022 across 21 sites. Inclusion criteria were patient age 6-17 years and adequate imaging quality. Radiology reports were classified as ICH-positive or ICH-negative using a validated natural language processing (NLP) tool. The AI tool analyzed DICOM images independently. Discordant AI-NLP cases underwent blinded adjudication by three radiologists to establish ground truth. Performance metrics includingsensitivity, specificity, accuracy, positive predictive value (PPV), and negative predictive value (NPV) were calculated with Wilson 95% confidence intervals (CIs). The cohort included 1,996 NCHCTs (768 females, 1,223 males, 5 unknown). ICH prevalence was 8.6% (172/1,996). Compared with ground truth, AI achieved 94.2% sensitivity (162/172, 95% CI, 89.6-97.2%), 94.7% specificity (1,727/1,824, 95% CI, 93.6-95.7%), 94.6% accuracy (1,889/1,996, 95% CI, 93.6-95.6%), 62.5% PPV (162/259, 95% CI, 57.8-67.0%), and 99.4% NPV (1,727/1,737, 95% CI, 99.0-99.7%). AI correctly identified ICH in cases missed by radiologists, but false positives were common, most often due to streak artifact (21.6%) and misclassified anatomy (18.6%). Interrater agreement for ground truth adjudication was substantial (κ=0.683). An adult-trained AI tool demonstrated high sensitivity, specificity, and accuracy for ICH detection in pediatric patients aged 6-17 years, comparable to its adult performance. Selective adaptation of adult-trained AI tools could expand access to AI-assisted triage for certain pediatric populations, potentially reducing delays in critical imaging interpretation. However, prospective validation is required before clinical deployment.