Initial Experience with an Artificial Intelligence System for Pulmonary Embolism Response Team Activation and Coordination in a Large Urban Health System

Sunday, March 5, 2023

3:45 PM – 3:54 PM

Location: Phoenix Convention Center, 226ABC

Presenting Author(s)

Timothy Carlon, MD, MBA (he/him/his)

Resident
Icahn School of Medicine At Mount Sinai

Disclosure(s): No financial relationships to disclose

Author/Co-author(s)

Daryl Goldman, MD

Resident
Icahn School of Medicine at Mount Sinai
Ricki A. Gottlieb, MD

Resident
Icahn School of Medicine at Mount Sinai
JW

Jennifer M. Watchmaker, MD, PhD

Resident Physician
Mount Sinai Hospital
DM

David Mendelson, MD

Professor of Radiology
Icahn School of Medicine at Mount Sinai
Robert A. Lookstein, MD

Executive Vice Chair; Diagnostic, Molecular, and Interventional Radiology
Icahn School of Medicine at Mount Sinai

Disclosure(s):

Timothy Carlon, MD, MBA: No financial relationships to disclose

Daryl Goldman, MD: No financial relationships to disclose

Ricki A. Gottlieb, MD: No financial relationships to disclose

Robert A. Lookstein, MD: No relevant disclosure to display

Purpose: To evaluate the performance of a novel artificial intelligence (AI) system for pulmonary embolism response team (PERT) activation and coordination under real-world conditions in a large urban health system.

Materials and Methods:

All CT pulmonary angiograms between January and August 2022 were triaged using an AI-based automated PERT activation algorithm (AIDoc, Tel-Aviv, Israel). AI PERT activation was based on the location of clot (central vs. peripheral) and CT evidence of right ventricular strain (RV to LV ratio greater than 1.0). A retrospective review was conducted of all institutional PERT activations over the same period. PERT decisions including patient transfers within the health system and use of advanced interventions were recorded. Performance characteristics of the AI algorithm for identifying true PERT activations, patients requiring transfer, and patients requiring intervention were calculated.

Results: Pulmonary embolism was identified on 954 CT pulmonary angiograms, and the AI algorithm sent an automated PERT notification on 155 (16.2%) of these. The institutional PERT was formally activated for 121 patients, of whom 38 underwent at least one intervention beyond anticoagulation alone (3 systemic thrombolysis, 4 surgical thrombectomy, 6 catheter-directed thrombolysis, 19 percutaneous thrombectomy, 15 inferior vena cava filter, 2 ECMO). Thirty-two patients who initially presented to satellite facilities were transferred to a hospital with both endovascular and open cardiothoracic surgery capabilities. The sensitivity and specificity of the AI algorithm for identifying true PERT activations were 62.8% and 90.5% respectively. Among patients requiring transfer for possible intervention, sensitivity was 84.4% and specificity 88.1%. Among patients who underwent any intervention, sensitivity was 78.9% and specificity 86.4%. When catheter-directed intervention is considered separately, sensitivity and specificity were 95.8% and 85.8% respectively.

Conclusion:

Our system’s initial experience demonstrates high sensitivity and specificity of the AI algorithm in identifying candidates for endovascular pulmonary embolism intervention despite poor correlation with PERT activations overall. Ongoing research will further assess the clinical benefits and optimal uses for this technology.