Author of 1 Presentation
SS 5.10 - Prediction of splenomegaly in >100,000 structured oncologic radiology reports using natural language processing
Abstract
Purpose
To develop and assess the accuracy of a natural language processing (NLP) model to identify splenomegaly from structured CT radiology reports at a tertiary cancer center.
Material and methods
In an IRB-approved, retrospective study, all CT chest/abdomen/pelvis reports (July 2009 to April 2019) adhering to departmental structured template were included. The SPLEEN subsection was extracted and those with default ‘unremarkable’ text were excluded from training. For patients with colorectal cancers (CRC), hepatobiliary cancers (HB), leukemia, Hodgkin’s lymphoma (HL) and non-HL (NHL), 1920 of 105,042 reports were annotated as positive or negative/uncertain for splenomegaly. Model training was performed on 1536 and model accuracy was tested on 384 reports. The prediction model was then applied to the remaining reports to calculate frequencies of splenomegaly.
Results
In the annotated reports, splenomegaly was present in 42.2%. After training, the splenomegaly classifier achieved 94% overall accuracy, 94.6% precision (positive predictive value) and 94% recall (sensitivity). When the model was applied to all unannotated reports, the predicted frequency of splenomegaly for CRC patients was 8.7% (5275/60462), HB 17.7% (2210/12506), leukemia 31.5% (1684/5340), HL 6.1% (390/6386) and NHL 9.2% (1866/20348).
Conclusion
NLP can predict splenomegaly from structured radiology reports after training from a limited sample of annotated text. At our institution, the frequency of splenomegaly in CRC patients was similar to HL and NHL patients, and lower than both patients with HB cancers and leukemia. Validation with splenic volumetry is ongoing.
Video-on-demand
Author of 1 Presentation
SS 5.10 - Prediction of splenomegaly in >100,000 structured oncologic radiology reports using natural language processing (ID 418)
Abstract
Purpose
To develop and assess the accuracy of a natural language processing (NLP) model to identify splenomegaly from structured CT radiology reports at a tertiary cancer center.
Material and methods
In an IRB-approved, retrospective study, all CT chest/abdomen/pelvis reports (July 2009 to April 2019) adhering to departmental structured template were included. The SPLEEN subsection was extracted and those with default ‘unremarkable’ text were excluded from training. For patients with colorectal cancers (CRC), hepatobiliary cancers (HB), leukemia, Hodgkin’s lymphoma (HL) and non-HL (NHL), 1920 of 105,042 reports were annotated as positive or negative/uncertain for splenomegaly. Model training was performed on 1536 and model accuracy was tested on 384 reports. The prediction model was then applied to the remaining reports to calculate frequencies of splenomegaly.
Results
In the annotated reports, splenomegaly was present in 42.2%. After training, the splenomegaly classifier achieved 94% overall accuracy, 94.6% precision (positive predictive value) and 94% recall (sensitivity). When the model was applied to all unannotated reports, the predicted frequency of splenomegaly for CRC patients was 8.7% (5275/60462), HB 17.7% (2210/12506), leukemia 31.5% (1684/5340), HL 6.1% (390/6386) and NHL 9.2% (1866/20348).
Conclusion
NLP can predict splenomegaly from structured radiology reports after training from a limited sample of annotated text. At our institution, the frequency of splenomegaly in CRC patients was similar to HL and NHL patients, and lower than both patients with HB cancers and leukemia. Validation with splenic volumetry is ongoing.