Predicting Causes of Death from Structured Electronic Health Records Using Machine Learning

Tuesday, August 27, 2024

2:00 PM – 2:15 PM CEST

Location: Conv Hall I D

Presenting Author(s)

Mohammed Al-Garad-, PhD

Research Assistant Professor
Vanderbilt University Medical Center, United States

Slides

Co-Author(s)

RR

Ruth M. Reeves, PhD

Health Services Resesearch Scientist
Vanderbilt University Medical Center
Nashville, United States
Rishi J. Desai, PhD

Associate Professor
Division of Pharmacoepidemiology and Pharmacoeconomics, Department of Medicine, Brigham and Women’s Hospital and Harvard Medical School, Boston, Massachusetts
Boston, United States
ML

Michele LeNoue-Newton

Vanderbilt University Medical Center, United States
DP

Daniel Park

Vanderbilt University Medical Center, United States
Shirley V. Wang, PhD (she/her/hers)

Associate Professor
Division of Pharmacoepidemiology and Pharmacoeconomics, Department of Medicine, Brigham and Women's Hospital, Harvard Medical School
Boston, United States
JM

Judith C. Maro, PhD

Associate Professor
Department of Population Medicine, Harvard Medical School and Harvard Pilgrim Health Care Institute
Boston, United States
Candace C. Fuller, MPH, PhD

Research Scientist
Department of Population Medicine, Harvard Pilgrim Health Care Institute
Boston, United States
Kueiyu Joshua Lin, MD, ScD

Assistant Professor
Division of Pharmacoepidemiology and Pharmacoeconomics, Department of Medicine, Brigham and Women's Hospital, Harvard Medical School
Boston, United States
José J. Hernández-Muñoz, RPh, MPH, MSc, PhD

Pharmacoepidemiologist
Office of Surveillance and Epidemiology, Center for Drug Evaluation and Research, U.S. Food and Drug Administration, Silver Spring, USA
The Woodlands, United States
Aida Kuzucan, PharmD, PhD

Epidemiologist
US Food and Drug Administration
Baltimore, United States
XW

Xi Wang

Office of Surveillance and Epidemiology, Center for Drug Evaluation and Research, US Food and Drug Administration, Silver Spring, MD, United States
HP

Haritha Pillai

Division of Pharmacoepidemiology and Pharmacoeconomics, Department of Medicine, Brigham and Women’s Hospital and Harvard Medical School, Boston, Massachusetts, USA, United States
KN

Kerry Ngan

Division of Pharmacoepidemiology and Pharmacoeconomics, Department of Medicine, Brigham and Women’s Hospital, Harvard Medical School, Boston, MA, United States
JW

Jill Whitaker, RN-BC, MSN

Nurse Informatics Data Analyst
Vanderbilt University Medical Center, United States
JD

Jessica Deere

Vanderbilt University Medical Center, United States
MM

Michael F. McLemore, RN

Nurse Informatics Data Analyst
Vanderbilt University Medical Center, United States
DW

Dax Westerman, MS

Principal Application Developer
Vanderbilt University Medical Center
Nashville, United States
Michael E. Matheny, MD, MS, MPH

Professor of Biomedical Informatics
Vanderbilt University Medical Center
Nashville, United States

Background: Rapidly identifying causes of death (CoD) is important for the surveillance of medical products but is limited by the delayed availability of cause of death information for patients. Structured electronic health record (EHR) data represents a rich source of patient medical history and care delivery information but has significant variability.

Objectives: This study evaluates the capacity of machine learning (ML) using structured EHR data to predict CoD.

Methods: As part of work with the Sentinel Innovation Center development network, we enrolled a cohort of 13,796 VUMC patients with a last encounter, a recorded death, and matching death records in the National Death Index (NDI) from 2019-2021. We collected structured data for 6 months prior to the last encounter including diagnosis codes, procedures, laboratory result codes, outpatient and inpatient medications, vital signs and demographics. A total of 2,048 features were available for ML prediction. The underlying condition for the CoD recorded by the NDI was re-coded into one of 52 standardized Rankable CoD per the National Center for Health Statistics. Data were split 80/20 into train/test sets, stratified by CoD labels, and four ML models were trained on this data: XGBoost, Random Forest, KNN, and SVM. Weighted AUC and weighted F-measure were employed as evaluation metrics.

Results: Within the cohort, the top 5 CoD were malignant neoplasms (30.1%), heart disease (15.9%), COVID-19 (7.6%), accidents (7.6%), and cerebrovascular disease (4.4%). The XGBoost algorithm outperformed other ML models in predicting the top 15 causes of death out of 52 categories. XGBoost achieved an average weighted AUC of 0.86 and a weighted F-measure of 0.74, while Random Forest, KNN, and SVM achieved average weighted AUCs of 0.79, 0.65, and 0.73, and weighted F-measures of 0.72, 0.51, and 0.59, respectively. Using XGBoost, "heart diseases" and "nephritis, nephrotic syndrome, and nephrosis" had the highest AUC scores at 0.99 and 0.96, respectively. Meanwhile, "cerebrovascular disease" and "essential hypertension and hypertensive renal disease" had AUC scores of 0.67.

Conclusions: This work demonstrates that CoD can be predicted with good discrimination using only high dimensional features extracted from structured EHR data. XGBoost was the top performer because it could handle sparse data and nonlinear relationships. Some CoDs, such as cerebrovascular disease, may be harder to predict because of close correlation with other diseases. Our future project aims to create a model that combines structured EHR data and unstructured clinical notes for a more effective mortality predictor and integrate these models into probabilistic CoD estimation.