The added value of text from Dutch general practitioner notes in predictive modeling

Area of Science:

• Health Informatics
• Clinical Prediction Modeling
• Natural Language Processing

Background:

• General practitioner (GP) electronic health records (EHRs) contain valuable unstructured clinical notes.
• Developing accurate prognostic prediction models is crucial for improving patient care in primary care settings.

Purpose of the Study:

• To evaluate the utility of Dutch unstructured EHR data, alongside structured data, for building prognostic prediction models.
• To assess the performance of models using text data, structured data, or a combination of both.

Main Methods:

• Trained and validated 84 prediction models for 4 clinical problems using diverse text representations and algorithms.
• Utilized observational GP EHR data, with internal and external validation across different EHR systems.

Main Results:

• Models using only text data performed comparably to or better than structured data alone in two prediction tasks.
• Combining structured and text data consistently outperformed models using either data type alone in these tasks.
• Minimal performance variation observed across different text representations and algorithms.

Conclusions:

• Unstructured EHR data can yield well-performing clinical prediction models independently.
• Integrating unstructured data with structured data offers significant added value for predictive accuracy.
• This research underscores the importance of clinical NLP in non-English languages and cross-EHR system validation.