Telehealth Notes for Diabetes Patients with Physical Activity Extraction Benchmark

Data from 943 patients collected over 12 years in the DiabMemory system, supplemented by 100 synthetic notes, were analyzed for physical activity information extraction. The dataset was created by Fabian Wiesmüller and last updated in May 2026. It includes pseudonymized free-text notes from a diabetes telehealth platform.

Use Cases

• Benchmarking rule-based NLP algorithms against LLMs based on extraction performance metrics described in the study
• Augmenting structured patient activity records using information extracted from unstructured clinical notes
• Evaluating the trade-offs between computational efficiency and accuracy for local NLP approaches
• Studying methods for preserving patient privacy in clinical NLP tasks using pseudonymization

Strengths

• Includes data from 943 patients spanning a 12-year period
• Benchmarks local methods (regex and Mistral LLM) against GPT-4.1, providing comparative performance metrics
• Patient privacy was preserved via pseudonymization and local deployment, avoiding third-party cloud services

Limitations

• Row count is unknown, which may limit suitability assessment
• The study notes limitations from a rather small test set and the use of synthetic data
• Column-level documentation is absent; field semantics must be inferred after download

Provenance

Source

DiabMemory system

Collection Method

Data collected from a telehealth platform, with notes pseudonymized and supplemented by synthetic data.

Time Range

Data collected over 12 years

Freshness

Last updated 2026-05-07 05:42:33; freshness should be verified