Assessing the Feasibility of an Automated Suggestion System for Communicating Critical Findings from Chest Radiology Reports to Referring Physicians

Abstract

Time-sensitive communication of critical imaging findings like pneumothorax or pulmonary embolism to referring physicians is important for patient safety. However, radiology findings are recorded in free-text format, relying on verbal communication that is not always successful. Natural language processing can provide automated suggestions to radiologists that new critical findings be added to a followup list. We present a pilot assessment of the feasibility of an automated critical finding suggestion system for radiology reporting by assessing suggestions made by the pyConTextNLP algorithm. Our evaluation focused on the false alarm rate to determine feasibility of deployment without increasing alert fatigue. pyConTextNLP identified 77 critical findings from 1,370 chest exams. Review of the suggested findings demonstrated a 7.8% false alarm rate. We discuss the errors, which would be challenging to address, and compare pyConTextNLP's false alarm rate to false alarm rates of similar systems from the literature.

Automatic detection of actionable findings and communication mentions in radiology reports using natural language processing

“…We are not the first to automatically detect actionable findings in radiology reports. Previous studies have reported different approaches, often with excellent results [2][3][4][5][6][7][8][9][10][11]. However, almost all of these studies focused on the detection of a limited set of findings, often related to a specific modality or organ system.…”

“…Monitoring the detection and communication of actionable findings in reports delivered in routine clinical care is important for quality control and auditing purposes, but requires laborious manual review that may not be feasible at scale. A considerable number of studies have used natural language processing to automate the detection of actionable findings in radiology reports [2][3][4][5][6][7][8][9][10][11]. However, almost all of these studies focused on the detection of a limited set of findings (often related to a specific modality or organ system) or even a single finding.…”

Applications of natural language processing tools in the surgical journey

“…When applied to radiology reports, NLP algorithms similarly achieve accuracy >90% through a variety of different applications. These include detection of intraabdominal fluid suggestive of surgical site infection from CT reports, detection of bone metastasis from bone scintigraphy reports and characterisation of other surgical pathologies including periprosthetic fractures (56)(57)(58). Despite this, accuracy of NLP is highly variable, with one study identifying poor accuracy in identifying critical features on thyroid ultrasound including echogenicity (27%) and margins (58.9%) (59).…”

A Data-driven Framework for Learning and Visualizing Characteristics of Thrombotic Event Phenotypes from Clinical Texts

“…4 The pyConText knowledge base was further developed to detect deep vein thrombosis as well as stroke and its risk factors. [5][6][7][8][9][10] Moreover, using the web-based query-building tool called Data Discovery and Query Builder (DDQB), Tien et al have shown how such expressions and rule-logic for negation can be leveraged to classify thrombotic events 30-days following hip and knee surgeries. 11 Their rule-based approach achieved high results: recall (97%) and specificity (99%) for deep vein thrombosis, recall (97%) and specificity (100%) for pulmonary embolism, and recall (100%) and specificity (99%) for myocardial infarction.…”