DETECT 1-Year Pilot Evaluation

Public health informatics project focused on evaluating the 1-year performance of the DETECT elder mistreatment screening tool embedded in EMS software. Built fuzzy-matched subject IDs, linked data across EMS and APS systems, and conducted multi-level fidelity and reporting analyses using R. Fully reproducible pipeline with annotated narrative code and manuscript draft.

Project Summary

This pilot project evaluated DETECT, a screening questionnaire embedded in MedStar’s EMS electronic care record (ePCR) software to assist EMS medics in identifying potential elder mistreatment (EM) in community dwelling older adults. The study assessed data from February 1, 2017 to February 1, 2018, using EMS encounter records and Texas Adult Protective Services (APS) to explore fidelity, reporting patterns, and linkage outcomes.

• Tool: DETECT EM screening questionnaire
  
• Institution: University of Texas Health Sciences Center, School of Public Health (Dallas Campus)
• PI: Dr. M. Brad Cannell
• Partners: MedStar Mobile Healthcare & Texas APS
  
• Objective: Evaluate DETECT’s field fidelity and reporting patterns
  
• Data Timeframe: Feb 2017 - Feb 2018
  
• Project Involvement: March 2022 - Sep 2023
  
• Status: Analysis completed; manuscript drafted (unpublished)

Tech Stack & Project Constraints

Despite the scale, complexity, and sensitivity of the DETECT dataset (~30K EMS records and ~18K APS records), all work was completed on a single local workstation with no access to server-grade infrastructure or high-performance computing. This required efficient memory management, modular workflow design, and rigorous reproducibility — even under constraints.

Core Tools & Libraries (R)

• tidyverse (incl. lubridate, tidyr, stringr, dplyr, forcats) — data wrangling, transformation, parsing
  
• fastLink — probabilistic record linkage & fuzzy matching
  
• codebookr — automated codebook generation
  
• mice — missingness pattern diagnostics
  
• questionr, readxl, vctrs, data.table, ggplot2 — analysis, input handling, and visualization
  
• here — reproducible project file paths

Collaboration & Cloud Tools

• GitHub — version-controlled, team-based code development
  
• OneDrive — secure cloud storage of PHI-compliant data sets
  
• Quarto (.qmd) — narrative coding and reproducible reporting
  
• Microsoft Word, Canva, Visio — manuscript drafting and visual presentation (per PI request)

By combining open-source tooling, privacy-conscious cloud platforms, and a fully local analysis pipeline, I built a scalable and reproducible workflow capable of handling sensitive public health data with minimal system overhead.

My Contributions

Data Wrangling

• Standardized date formats, validated birthdates
  
• Parsed and cleaned address and name strings
  
• Transformed race/ethnicity into meaningful categoricals
  
• Flagged contextual cues in free-text comment fields

Fuzzy Matching

• Performed subject-level fuzzy matching within EMS data using fastLink
• Performed subject-level fuzzy matching between EMS and APS data using fastLink
  
• Linked EMS and APS datasets via cross-matching identifiers and temporal criteria
  
• Created robust subject IDs with minimal manual review burden

Analysis Pipeline

• Developed consort flow tables
  
• Analyzed item-level DETECT response patterns
  
• Assessed fidelity between screening and reporting intent/outcomes
  
• Profiled demographic characteristics from linked records

Reproducible Reporting

• Authored comprehensive QMD files with narrative-style annotations
  
• Generated full codebooks using codebookr for all derived datasets

Manuscript Preparation

• Wrote full-length scholarly manuscript summarizing findings and methods for PI

Code Highlights

Data Wrangling

File	Description
medstar_compliance_01_cleaning.qmd	Cleaned race/ethnicity, address strings, ZIP codes (36,304 → 28,228 cleaned rows)
medstar_epcr_cleaning_02_fastLink.qmd	Benchmarked variable selection for optimal fuzzy matching
medstar_epcr_cleaning_03_unique_ids.qmd	Validated subject IDs (28,228 rows, 16,565 unique subjects)
aps_cleaning_01_initial_clean.qmd	Cleaned APS records (18,152 entries, 11,178 unique subjects)

Fuzzy Matching & Linkage

File	Description
merge_aps_medstar_01_fastLink.qmd	Matched subjects across EMS and APS
merge_aps_medstar_02_group_ids.qmd	Created 2,126 valid cross-set group IDs
merge_aps_medstar_03_refining_observations.qmd	Temporally matched EMS and APS events based on intake/investigation timelines

Analysis

File	Focus	Description
analysis_medstar_aps_01_consort_tables.qmd	Consort Tables	Tracks subject flow between EMS and APS records
analysis_medstar_aps_02_detect_response_patterns.qmd	Screening Responses	Explores DETECT item-level response patterns
analysis_medstar_aps_03_fidelity_agreement.qmd	Fidelity & Agreement	Assesses alignment between screening, reporting, and APS follow-up
analysis_medstar_aps_04_demographics_analyses.qmd	Demographics	Summarizes age, gender, and other key traits

Reflections

This project blended public health informatics with careful statistical wrangling and real-world ambiguity. From parsing messy free-text entries to developing a highly scalable fuzzy matching framework, my work ensured linkage quality while surfacing meaningful insights into DETECT’s field performance. Though the manuscript was never published, the structure, code, and process remain directly transferable to future interdisciplinary screening studies.