Diabetes Telehealth Activity Notes Augmented via NLP

943 diabetes patients' structured telehealth data was augmented with physical activity information extracted from their free-text notes over a 12-year period. Fabian Wiesmüller published this research in 2026, benchmarking local rule-based and Mistral LLM methods against GPT-4.1. The dataset includes 100 synthetically generated notes used for benchmarking the extraction algorithms.

Use Cases

• Benchmarking local NLP models against cloud-based LLMs based on the described synthetic note evaluation.
• Developing rule-based algorithms for clinical text information extraction based on the described regex methods.
• Augmenting structured patient records with unstructured data insights based on the described integration task.
• Studying privacy-preserving data processing in healthcare based on the described pseudonymization approach.

Strengths

• Data spans 12 years of collection from 943 patients, providing longitudinal context.
• Includes 100 synthetic notes specifically created for benchmarking model performance.
• Local methods achieved a minimum F1-score of 0.84 for the primary extraction task.
• Rule-based data augmentation achieved an F1-score of 0.73, surpassing a local LLM.

Limitations

• The test set is described as 'rather small', which may limit generalizability of findings.
• Benchmarking relied partly on synthetic data, which may not fully represent real-world notes.
• Column-level documentation is absent; field semantics must be inferred after download.

Provenance

Source

Data collected from the DiabMemory telehealth system.

Collection Method

Structured patient data was augmented using NLP extraction from free-text notes.

Time Range

Data collected over 12 years.

Freshness

Last updated 2026-05-07 05:42:33