Sawyer Stribley
Research Mentor: Peter Larson
Mentor Department: Injury Prevention Center/School of Public Health/SEAS, Public Health
Author(s): Sawyer Stribley, Jaemin Jeon, Elora Shakoor, Peter Larson
Session: Session 6 (3:00 PM – 3:50 PM)
Presentation Type: Poster 46
Abstract
(1) Statement of Purpose MDHHS records information for all decedents located in Michigan upon death, including ICD-10 codes and, in the case of injury related deaths, cause of death descriptions (CDDs). In cases with multiple contributing conditions, up to 18 ICD-10 codes may be recorded, and the underlying mechanism or sequence of events leading to death is often ambiguous when relying on codes alone. This ambiguity can lead to misclassification of injury mechanisms and biased mortality estimates in injury research. The purpose of this study is to evaluate whether semantic modeling of CDDs can be used to assess the consistency of ICD-10 injury classifications and to identify dominant mechanisms and event sequences underlying injury-related deaths. (2) Methods and Approach We obtained publicly available, de-identified death records for all Michigan decedents ranging from 2006-2025. Free-text cause-of-death descriptions were cleaned and standardized to reduce linguistic variability while preserving semantic content. We vectorized the remaining sentences to conduct analyses using both keyword-based and embedding-based approaches. In addition to hierarchical clustering, we compared semantic clustering from pre-trained and untrained models. Model outputs were evaluated in relation to ICD-10–based injury classifications to assess concordance and areas of divergence. (3) Results and Conclusions The dataset includes over 1.3 million death records, of which a substantial subset reflects injury-related causes such as falls, motor vehicle incidents, occupational injuries, homicide, suicide, and substance-related deaths. Preliminary analyses indicate that semantic clustering of cause-of-death descriptions produces coherent groupings that align with major injury mechanisms while also revealing heterogeneity within ICD-10 categories. Both pre-trained and untrained “from scratch†models are able to organize death descriptions by semantic meaning and sentence structure. Models can assess chronology, and determine order of events leading to death. (4) Innovation and Significance to the Field Vital records are a cornerstone of injury surveillance, yet reliance on ICD-10 codes alone may obscure injury mechanisms and pathways, particularly in complex or multi-factor deaths. This study demonstrates a scalable, reproducible strategy for incorporating free-text cause-of-death narratives into injury classification and validation efforts. Integrating semantic analysis with traditional coding systems has the potential to improve injury surveillance accuracy and strengthen downstream epidemiologic research.
