Core Concepts
Annotation inconsistencies in unstructured death investigation notes can lead to misattributed suicide causes, hindering effective suicide prevention strategies. This study proposes an empirical NLP approach to detect these inconsistencies, identify problematic instances, and demonstrate the effectiveness of correcting them.
Abstract
This study aimed to uncover annotation inconsistencies in unstructured death investigation notes from the National Violent Death Reporting System (NVDRS) and their impact on suicide cause attribution.
Key highlights:
Demonstrated the existence of data annotation inconsistencies across states, leading to performance disparities in suicide crisis prediction systems.
Introduced a method to identify problematic instances responsible for these inconsistencies through a cross-validation-like paradigm.
Found that for Ohio, 14.8% of the annotations for Family Relationship Crisis, 13.9% for Physical Health Crisis, and 1.5% for Mental Health Crisis were potential mistakes; for Colorado, 7.7%, 4.9%, and 2.0% respectively.
Showed that removing the problematic instances improved model performance and generalizability, with an average increase of 2.2% in the average micro F1 scores on the test set of other states.
Manually rectified 159 potential mistakes in Ohio's Family Relationship Crisis annotations, finding 89 to be actual mis-labelings, which led to a 4.2% increase in the average micro F1 score on the test set of other states and a 3.5% increase on Ohio's test set.
Analyzed the risk of bias in the data annotations, observing differences in the Odds Ratios for demographic subgroups (age, race, sex) before and after removing the identified mistakes.
The findings highlight the importance of addressing annotation inconsistencies in unstructured death investigation notes to improve the accuracy and reliability of suicide cause attribution, ultimately supporting more effective suicide prevention strategies.
Stats
The National Violent Death Reporting System (NVDRS) dataset contains 267,804 recorded suicide death incidents from 2003 to 2020 across all 50 U.S. states, Puerto Rico, and the District of Columbia.
Quotes
"Recent studies suggested the annotation inconsistencies within the NVDRS and the potential impact on erroneous suicide-cause attributions."
"Our results showed that incorporating the target state's data into training the suicide-crisis classifier brought an increase of 5.4% to the F1 score on the target state's test set and a decrease of 1.1% on other states' test set."