congastro2023-40″>https://doi.org/10.46613/>congastro2023-40
This work is licensed under CC BY 4.0
Robles C
1
, Baquerizo J
1
, Puga M
1
, Cunto D
1
, Egas M
1
, Arevalo M
1
, Mendez J
1
, Alcivar J
1
, Del Valle R
1
, Alvarado H
1
, Tabacelia D
1
, Carvajal J
1
, Pitanga H
1
1Instituto Ecuatoriano de Enfermedades Digestivas (IECED), Guayaquil, Ecuador
BACKGROUND: Two digital single-operator cholangioscopy (DSOC)-based artificial intelligence (AI) models have been worldwide proposed to identify neoplasia. Both have similar or even higher diagnostic accuracy (DxA) than expert endoscopists or target biopsies. Direct visualization of biliary strictures through DSOC-guided pCLE has also demonstrated high DxA. There are no head-to-head studies comparing both resources.
AIM: To compare DxA for identifying neoplasia in indeterminate biliary lesions using a novel DSOC-AI model against DSOC-guided pCLE.
METHODS: Historic cohort. Adults underwent a video-recorded DSOC-guided pCLE (06/2014 to 11/2021). The absence of videos, biopsies, or twelve-month follow-ups were excluded. For DSOC direct visualization, Carlos Robles-Medranda classification constituted criteria for neoplasia. For DSOC-guided pCLE, Miami and Paris classifications constituted criteria for neoplasia or inflammatory disorders, respectively. DSOC-AI model analysed DSOC videos offline (AIWorks-Cholangioscopy system). The DSOC-AI model considered neovasculature as a neoplasia criterion. The gold standard for neoplasia was clinical evolution, imaging, or surgical specimen during twelve-month follow-up. DxA were calculated in terms of sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), observed agreement, and area under the curve (AUC).
RESULTS: Ninety patients were selected, median age was 66.4 ± 13.7 years, 56.7% female. Tumour suspicion was the most common indication (55.6%). The DSOC-AI model reached a 97.7% sensitivity, 75% specificity, 98.8% PPV, 60% NPV and 96.7% observed agreement. The DSOC-guided pCLE reached a 94.2% sensitivity, 100% specificity, 100% PPV, 44.4% NPV and 94.4% observed agreement. AUC for DSOC-AI model was 0.79, DSOC direct visualization 0.74 (P=.763), DSOC-guided pCLE 0.72 (P=.634) and DSOC and pCLE-guided biopsy 0.83 (P=.809).
CONCLUSION: The DSOC-AI model demonstrated an offline DxA similar to DSOC-guided pCLE. It is advisable to design a larger multicentric head-to-head trial with a proportional sample among neoplastic and non-neoplastic cases for the best estimation of obtained results.