Theoretical Clinical Utility of Advanced Practice Provider Use of an Artificial Intelligence Radiomics-Based Tool for Pulmonary Nodule Evaluation and Management

Abstract

Abstract RATIONALE: Advanced practice providers (APPs) are integral members of many pulmonary nodule (PN) clinics. We aimed to determine how a commercially available artificial intelligence (AI) radiomics-based tool that generates a Lung Cancer Prediction (LCP) score from 1 to 10 influences APP PN evaluation and management. METHODS: In this retrospective multi-reader multi-case study, 6 APPs independently evaluated 300 anonymized, axial-view CT scans each containing an indeterminate PN (1,800 evaluated PNs; 50% malignancy prevalence). Without any additional clinical information, for each case APPs were asked to provide an estimate of malignancy risk (0%-100%) and a management recommendation (no follow-up, short- or long-term CT surveillance, immediate CT or PET imaging, non-surgical biopsy, or surgical resection) before and after receiving the AI-generated LCP score. We compared malignancy risk estimates and management recommendations with and without use of the AI tool using descriptive statistics, area under the curve (AUC), and reclassification table analysis. RESULTS: Compared to benign PN cases (N=900), malignant PN cases (N=900) were associated with a higher LCP score (8.5 vs 4.6; P<0.001) generated by the AI tool. With use of the AI tool, mean APP malignancy risk estimates increased for malignant PNs (68.9% vs 55.7%; P<0.001) and decreased for benign PNs (21.6% vs 23.7%; P<0.001), corresponding to an increase in average AUC from 0.79 to 0.88 (P<0.001). A higher proportion of malignant PNs were classified as high (>65%) risk (63.4% vs 46.2%; P<0.001) and recommended for an invasive biopsy or surgical resection (71.6% vs 54.9%; P<0.001) with use of the AI tool (Table). Notably, among PN cases in which use of the AI tool was associated with a change in management recommendation from non-invasive imaging to an invasive procedure, 82% (164 of 201) represented malignancy, compared to 33% (14 of 43) among cases in which the opposite management recommendation change occurred. In contrast, while benign PNs were more often classified as low (<5%) risk (39.0% vs 34.3%; P<0.001) with use of the AI tool, there was no significant difference in the proportion recommended for an invasive procedure (18.1% vs 17.2%; P=0.49). CONCLUSIONS: Use of a commercially available AI radiomics-based tool for PN evaluation was associated with increased APP diagnostic accuracy and invasive diagnostic procedure recommendation for malignant PNs. These results suggest that use of this AI tool has theoretical clinical utility, and that future prospective, randomized controlled clinical trials should be conducted to assess its use in routine clinical practice.

Ähnliche Arbeiten