Purpose or Case Report: To evaluate and quantify the use of uncertainty language in pediatric radiology report impressions, focusing on variation of uncertainty by modality, patient demographics, and care setting.
Methods & Materials: We analyzed 310,789 pediatric radiology reports from a tertiary care center over a 12-month period (10/2023-09/2024). Reports were assessed based on the first impression point, including information on modality, care setting, patient sex, and age. GPT-4o (OpenAI, USA) was used to generate an expanded list of uncertainty terms (n=54) and categorized these terms into five probability levels: Very high probability (>90%), High probability (>75% and <90%), Intermediate probability (>25% and <75%), Low probability (<25% and >5%), and Very low probability (<5%) as outlined in Shinagare et al.’s Diagnostic Certainty Scale. Each term was assigned the median probability value of its range for statistical analysis. We characterized the average level of uncertainty across different modalities, patient demographics, and clinical settings, using a one-sample t-test to determine statistical significance (set at P <0.001).
Results: The top five uncertainty phrases identified were, “likely,” “compatible with,” “consistent with,” “may be,” and “may represent.” The average certainty probability across all reports was 97.0%. Reports for patients aged 0-1 (96.7%, P<0.001), 1-5 (96.8%, P<0.001), outside imaging consults (94.0%, P<0.001), MR (94.0%, P<0.001), CT (95.8%, P<0.001), ultrasound (96.7%, P<0.001), and emergency cases (96.6%, P<0.001) exhibited significantly higher uncertainty compared to the overall average. No significant difference in uncertainty was observed between the average and the male (96.9%, P=0.22) and female (97.0%, P=0.20) cohorts.
Conclusions: Cross sectional modalities such as MRI and CT, low resolution exams such as ultrasound, as well as emergency and outside imaging consults, showed notably higher uncertainty, potentially reflecting the complexities of these cases and/or the limitations to the interpretations related to modality and clinical context. Patient sex did not appear to influence the use of uncertainty language, which serves as a control for the analysis. These findings highlight the importance of standardizing uncertainty communication across different clinical contexts, which could improve diagnostic clarity and aid clinical decision-making. Future studies should explore the clinical impact of these uncertainty variations on patient outcomes and decision-making processes.