KB and ML-NLP Performance on Kentucky EMS Opioid Overdose Data (2018-2022)

Peter Rock's dataset compares knowledge-based and machine learning methods for identifying suspected opioid overdoses in Emergency Medical Services data. It contains performance metrics from an oversampled dataset of 2,327 weighted EMS encounters from Kentucky between 2018 and 2022. The data is stored in an XLS file and was last updated in April 2026.

Use Cases

• Benchmarking NLP model performance against traditional knowledge-based definitions for overdose identification based on the described evaluation metrics.
• Training or validating ML-NLP models for clinical text classification using EMS patient care narratives mentioned in the description.
• Analyzing the trade-offs between sensitivity and specificity in overdose surveillance based on structured versus unstructured data approaches discussed in the results.

Strengths

• Dataset is based on 2,327 weighted EMS encounters, providing a substantive sample for analysis.
• Ground truth labels were determined by manual review from EMS experts, suggesting a reliable benchmark.
• Performance metrics (sensitivity, specificity, F1-score) for five KB definitions and ML-NLP models are included, allowing for direct comparison.

Limitations

• Row count is unknown, which may limit suitability assessment for large-scale modeling.
• Column-level documentation is absent; field semantics must be inferred after download.
• The 5.5 KB file size indicates a very small dataset, likely containing only summary results rather than raw case data.

Provenance

Source

Kentucky State EMS data (2018–2022)

Collection Method

Secondary data from an oversampled dataset, with ground truth labels determined by EMS expert manual review.

Time Range

2018 to 2022

Freshness

Last updated 2026-04-28 17:25:27; freshness should be verified.

Geography

Kentucky, USA