Number
816
Name
A Machine Learning Tool for Assessing Student Clinical Reasoning in Predicting Cause of Death from Multi-System Anatomical Cadaveric Observations
Date & Time
Sunday, June 7, 2026, 5:30 PM - 7:00 PM
Location Name
Oglethorpe Ballroom
Speakers
Authors
Adrian Bozocea, B.S., Medical College of Georgia, Augusta University, Augusta, Georgia
Amogh Gadekar, B.S., Medical College of Georgia, Augusta University, Augusta, Georgia
Shivum Lal, B.S., Medical College of Georgia, Augusta University, Augusta, Georgia
Morganne Manuel, Ph.D., Medical College of Georgia, Augusta University, Augusta, Georgia
Olivia Wireman, Ph.D., Medical College of Georgia, Augusta University, Augusta, Georgia
Shannon Barwick, Ph.D., Medical College of Georgia, Augusta University, Augusta, Georgia
Mindy Johnson, Ph.D., Medical College of Georgia, Augusta University, Augusta, Georgia
Presentation Topic(s)
Technology and Innovation
Description
PURPOSE
Cadaveric dissection remains central to medical education, offering
opportunities to explore anatomical factors relevant to patient outcomes. At
the Medical College of Georgia, students conduct postmortem investigations
that generate datasets encompassing anatomical variations, gross pathology,
and histopathology. However, these datasets remain difficult to use in
determining disease patterns and contextualizing donor demographics. This study
presents a reproducible supervised machine learning framework for
heterogeneous cadaveric data to compare model predictions against student
clinical reasoning, support future mixed-methods investigations, and
establish metrics for anatomy education quality improvement.
METHODS
This retrospective study utilized 337 anatomical observations across five
domains (demographics, musculoskeletal, cardiopulmonary-hematologic,
gastrointestinal-genitourinary, brain/head) from 49 donors in two cohorts.
The 2020-2021 cohort (n=24) served as training data and 2021-2022 (n=25) as
independent testing data. A three-stage feature engineering pipeline reduced
337 observations (69.2% categorical, 16.4% integer, 14.3% free-text) to 49-53
clinically interpretable predictors using quality filtering, statistical
feature selection, and composite feature creation. Models were trained to
predict five cause-of-death (COD) categories (cardiovascular, respiratory,
neoplasm, neurologic, multisystem) with class imbalance correction and
evaluated against student predictions.
RESULTS
Eleven algorithms were evaluated in baseline and tuned configurations, with
ensemble methods applied to top models. Random Forest achieved highest
performance (F1-score: 0.70) and 87% accuracy on independent tests versus 68%
student accuracy. Models performed best for cardiovascular and neurologic COD
and struggled with respiratory cases. Feature-importance analysis identified
heart enlargement, age-inappropriate thoracic and brain pathology,
lung/pleural abnormalities, and organ weights as key discriminators.
CONCLUSION
This study establishes the first reproducible supervised machine learning
framework for heterogeneous cadaveric datasets that outperformed student
reasoning. These findings highlight its potential as an educational reference
tool and a foundation for future anatomy curriculum and research.
Cadaveric dissection remains central to medical education, offering
opportunities to explore anatomical factors relevant to patient outcomes. At
the Medical College of Georgia, students conduct postmortem investigations
that generate datasets encompassing anatomical variations, gross pathology,
and histopathology. However, these datasets remain difficult to use in
determining disease patterns and contextualizing donor demographics. This study
presents a reproducible supervised machine learning framework for
heterogeneous cadaveric data to compare model predictions against student
clinical reasoning, support future mixed-methods investigations, and
establish metrics for anatomy education quality improvement.
METHODS
This retrospective study utilized 337 anatomical observations across five
domains (demographics, musculoskeletal, cardiopulmonary-hematologic,
gastrointestinal-genitourinary, brain/head) from 49 donors in two cohorts.
The 2020-2021 cohort (n=24) served as training data and 2021-2022 (n=25) as
independent testing data. A three-stage feature engineering pipeline reduced
337 observations (69.2% categorical, 16.4% integer, 14.3% free-text) to 49-53
clinically interpretable predictors using quality filtering, statistical
feature selection, and composite feature creation. Models were trained to
predict five cause-of-death (COD) categories (cardiovascular, respiratory,
neoplasm, neurologic, multisystem) with class imbalance correction and
evaluated against student predictions.
RESULTS
Eleven algorithms were evaluated in baseline and tuned configurations, with
ensemble methods applied to top models. Random Forest achieved highest
performance (F1-score: 0.70) and 87% accuracy on independent tests versus 68%
student accuracy. Models performed best for cardiovascular and neurologic COD
and struggled with respiratory cases. Feature-importance analysis identified
heart enlargement, age-inappropriate thoracic and brain pathology,
lung/pleural abnormalities, and organ weights as key discriminators.
CONCLUSION
This study establishes the first reproducible supervised machine learning
framework for heterogeneous cadaveric datasets that outperformed student
reasoning. These findings highlight its potential as an educational reference
tool and a foundation for future anatomy curriculum and research.
Presentation Tag(s)
Student Presentation