Linking Symptom Inventories Using Semantic Textual Similarity.

Area of Science:

• Neuroscience
• Medical Informatics

Background:

• Traumatic brain injury (TBI) assessment relies on numerous symptom inventories.
• Inconsistency across these inventories hinders comparable diagnostics and outcome prediction.
• Lack of standardization complicates data aggregation and analysis in TBI research.

Purpose of the Study:

• To develop and evaluate a novel approach using semantic textual similarity (STS) to harmonize TBI symptom data.
• To link symptom descriptions and scores across disparate TBI assessment tools.
• To improve the comparability and utility of TBI assessment data.

Main Methods:

• Utilized four pretrained deep learning models to perform STS on symptom description pairs.
• Ranked symptom similarities based on conceptual likeness to bridge incongruent inventories.
• Applied the STS approach to predict symptom severity across four inventories for 6,607 participants from 16 international data sources.

Main Results:

• The STS approach achieved 74.8% accuracy across five prediction tasks.
• This method outperformed other tested models in screening symptom description pairs.
• Correlation and factor analysis confirmed that the properties of the original scales were largely preserved post-conversion.

Conclusions:

• Semantic textual similarity offers a powerful method for harmonizing TBI assessment data.
• This approach can aid expert decision-making and improve the reliability of TBI diagnostics.
• Incorporating contextual semantic information enhances TBI outcome prediction and research comparability.