Clinical Evaluation of An Artificial Intelligence Algorithm For the Flagging of Incidental Pulmonary Emboli on Routine Contrast-Enhanced Chest CT

Kiran Batra, MD

Summary

Purpose

 To evaluate an incidental pulmonary embolism ( iPE ) Artificial Intelligence (AI) algorithm in patients referred for routine, contrast-enhanced CT (CECT)of the chest at two healthcare systems served by the same radiologists. We examined the performance of the algorithm, impact of different patient populations and conditions associated with misclassification.


Methods and Materials

The iPE algorithm was prospectively applied to Contrast-enhanced Chest CT (CECCT )exams between September 2019 and February 2020 at a safety net hospital (SNH) and academic referral hospital (ARH). Clinical reads were performed by radiologists serving both institutions. A consensus overread was performed on algorithm and/or clinical report positive exams. In addition, exams were reviewed to assess for etiologies which could contribute to misclassification.


Results

The iPE algorithm analyzed 7,466 CECT exams (mean age: 59±15, 3,811 women). The algorithm had a sensitivity of 83.3%, specificity of 99.8%, and negative predictive value of 99.7%. Significant differences in age, gender, percentage of outpatients, percentage of urgent and stat cases, and prevalence of associated cancer diagnosis between the two institutions did not alter the positive rates for detection of iPE (SNH: 1.26% vs ARH: 1.77%, P=NS). Several factors contributed to misclassification, including anatomy altered by prior surgery and rim enhancement of metastatic lesions.


Conclusions

The AI-algorithm had moderate sensitivity but a high specificity and negative predictive value in detecting iPE. The detection rate was similar at 2 hospitals with different patient populations and referral patterns. We also identified several sources of misclassification which could serve as targets to potentially improve the performance of the algorithm


Clinical Relevance/Application

 Incidental pulmonary embolism (iPE) is a finding on routine CT exams ordered for other reasons and is associated with increased mortality and morbidity. An Artificial Intelligence (AI) algorithm that flags iPE could be beneficial to radiologists. However, the variability in contrast timing, underlying pathology, and patient referral pose significant challenges to an algorithm for iPE.

Popular Posts
Subscribe to our Blog
Follow us on social
AI Consultation Request
Get your complimentary consultation with an AI expert