Using natural language processing to facilitate the harmonisation of mental health questionnaires: a validation study using real-world data

Background: Pooling data from different sources will advance mental health research by providing larger sample sizes and allowing cross-study comparisons; however, the heterogeneity in how variables are measured across studies poses a challenge to this process.

Methods: This study explored the potential of natural language processing (NLP) to harmonise different mental health questionnaires by matching individual questions based on their semantic content. Using the Sentence-BERT model, we calculated the semantic similarity (cosine index) between 741 pairs of questions from 5 questionnaires. Drawing on data from a representative UK sample of adults (N=2,058), we calculated a Spearman rank correlation for each of the same pairs of items, and then estimated the correlation between the cosine values and Spearman coefficients. We also used network analysis to explore the model’s ability to uncover structures within the data and metadata.

Results: We found a moderate overall correlation (r = .48, p <.001) between the two indices. In a holdout sample, the cosine scores predicted the real-world correlations with a small degree of error (MAE= 0.05, MedAE = 0.04 , RMSE = 0.064) suggesting the utility of NLP in identifying similar items for cross-study data pooling. Our NLP model could detect more complex patterns in our data, but it needed manual rules to decide which edges to include in the network.

Conclusions: This research shows that it is possible to quantify the semantic similarity between pairs of questionnaire items from their meta-data, and these similarity indices correlate with how participants would answer the same two items. This highlights the potential of NLP to facilitate cross-study data pooling in mental health research. Nevertheless, researchers are cautioned to verify the psychometric equivalence of matched items.