Clinical Data for Diabetes Risk Prediction with NLP-Derived Features

A dataset of 1,879 individuals used to investigate integrating NLP with clinical data for type 2 diabetes risk prediction. The data includes structured clinical variables and unstructured textual entries processed with a BERT-based NLP pipeline. The study, authored by Yaoyan Lu and last updated in 2026, validated the integrated model on a post-2020 cohort of 939 individuals.

Use Cases

• Train a diabetes risk classifier based on structured clinical variables like BMI, HbA1c, and blood pressure.
• Develop NLP models to extract latent risk factors from unstructured clinical notes, such as symptom descriptions and lifestyle notes.
• Build hybrid predictive models that integrate NLP-derived features with traditional clinical data.
• Benchmark machine learning algorithms (logistic regression, random forest, XGBoost, neural networks) on a multimodal clinical dataset.
• Conduct temporal validation studies on medical prediction models using the post-2020 cohort data.

Strengths

• Dataset size of 1,879 individuals provides a basis for model development.
• Includes a temporal validation cohort of 939 individuals from a post-2020 period.
• Performance metrics for the integrated model are reported, including an AUC-ROC of 0.92 and accuracy of 88.2%.
• Model robustness was assessed via bootstrap confidence intervals, sensitivity analysis, and optimism adjustment.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count for the primary dataset is known, but the exact number of rows per individual or feature set is unknown.
• The dataset is very small at 137.4 KB, indicating limited scope and likely a summary or processed feature set rather than raw data.

Provenance

Source

Yaoyan Lu via figshare.

Collection Method

Analyzed from a public dataset; structured variables were normalized and unstructured text was processed via a BERT-based NLP pipeline.

Time Range

Primary data collection period unspecified; includes a temporal validation cohort from a post-2020 period.

Freshness

Last updated 2026-05-28 06:18:36; freshness should be verified.