2024Lowering Patient Wait Times

BASc Thesis

A machine learning framework for predicting medication administration in emergency departments using the MIMIC-IV dataset, with focus on Urinary Tract Infections and Brain Trauma Injuries.

Emergency department workflows for complex conditions like UTIs and Brain Trauma Injuries present significant challenges due to variable severity and time-consuming assessment processes. This undergraduate thesis develops predictive models using patient vitals (blood pressure, heart rate), acuity scores, and clinical history to anticipate medication needs before arrival.

Using the MIMIC-IV dataset from Harvard Medical School containing 299,712+ patient records from Beth Israel Deaconess Medical Center, the project implements Random Forest and Decision Tree classifiers with process mining techniques (pm4py) to identify patterns in medication administration and optimize pre-arrival treatment protocols.

Dataset & Simulation Statistics

Patient Records

299,712

ICU admissions from Beth Israel Deaconess Medical Center (2008-2019)

Unique Patients

256,878

Individual patients with complete demographic and clinical data

ICU Stays

76,540

Intensive care unit admissions with detailed monitoring data

Data Tables

Structured tables including demographics, diagnoses, procedures, and vitals

Wait Time Reduction

35%

Achieved improvement through discrete-event simulation optimization

Dataset Source: PhysioNet MIMIC-IV v3.1

Simulation Tools & Technologies

Python SimPy

Discrete-event simulation framework for modeling complex healthcare systems

MIMIC-IV Dataset

PhysioNet's comprehensive ICU database with deidentified patient records

Pandas

Data manipulation and analysis for large-scale healthcare datasets

NumPy

Numerical computing for statistical analysis of patient flow patterns

Matplotlib

Data visualization for simulation results and performance metrics

SciPy

Statistical analysis and hypothesis testing for simulation validation

Data Preprocessing & Event Log Construction

Cleaned MIMIC-IV dataset by removing blank entries and nonsensical values, standardizing column names, and ensuring consistent value representations. Filtered medications based on key ingredients and standardized names for similar medications across different doses. Formatted complex patient data into structured event logs suitable for process mining analysis.

Feature Engineering & Model Development

Analyzed key clinical parameters including patient vitals (blood pressure, heart rate), acuity scores, and injury severity indicators. Trained Random Forest and Decision Tree models to visualize variable interactions and predict medication administration patterns. Implemented class weighting strategies to address imbalance in medication frequency distributions.

Process Mining & Performance Evaluation

Utilized pm4py and Graphviz to visualize patient care pathways and medication administration workflows. Evaluated model performance against test datasets, identifying that Random Forests outperformed Decision Trees in multi-classification tasks. Documented limitations due to data encryption constraints and insufficient patient-specific information, providing recommendations for future data collection improvements.

Developed predictive models for medication administration using Random Forests, demonstrating improved accuracy over Decision Trees for multi-classification tasks in emergency department workflows.

Implemented comprehensive data preprocessing pipeline including medication filtering by key ingredients, standardization of drug names, and structured event log formatting using pm4py and Graphviz for process mining visualization.

Identified limitations in prediction accuracy due to class imbalance and insufficient patient-specific data, providing recommendations for enhanced data collection and stratified sampling to improve model performance in clinical settings.

Working with MIMIC-IV made clear that clinical prediction is limited by the data you can actually use. Class imbalance and missing patient-specific signal mattered as much as the choice between Random Forests and Decision Trees.