Large language models for epidemiological research via automated machine learning: a case study and method comparison from the British National Child Development Study (Preprint)

BACKGROUND Large language models have had a huge impact on natural language processing (NLP) in recent years. However, their application in epidemiological research is still limited to analysis of electronic health records and social media data. OBJECTIVE To demonstrate the potential beyond these domains, we aimed to develop prediction models based on texts collected in an epidemiological cohort and compared their performance to classical regression methods. METHODS We used data from the British National Child Development Study, where 10,567 11-year-old children wrote essays about how they imagined themselves as 25-year-olds. Fifteen percent of the dataset was set aside as a test set for performance evaluation. Pre-trained language models were fine-tuned using AutoTrain (by Hugging Face) to predict current reading comprehension score (0-35) and future body mass index (BMI) and physical activity (active vs. inactive) at the age of 33. We then compared their predictive performance (accuracy or discrimination) with linear and logistic regression models including demographic and lifestyle factors of the parents and the children between birth and age 11 as predictors. RESULTS NLP clearly outperformed linear regression when predicting reading comprehension score (RMSE=3.89 [95% CI: 3.74, 4.05] for NLP vs. 4.14 [3.98, 4.30] and 5.41 [5.23, 5.58] for regression models with and without general ability score as predictor). Predictive performance for physical activity was similarly poor for the two methods (AUC ROC=0.55 [0.52, 0.60] for both), but slightly better than random assignment, while linear regression clearly outperformed the NLP approach when predicting BMI (RMSE=4.38 [4.02, 4.74] for NLP vs. 3.85 [3.54, 4.16] for regression). The NLP approach did not perform better than simply assigning the mean BMI from the training set as predictors. CONCLUSIONS Our study demonstrated the potential of using large language models to utilize text collected in epidemiological studies. The performance of the approach appeared to depend on how directly the topic of the text was related to outcome. Open-ended questions specifically designed to capture certain health concepts and lived experiences in combination with NLP methods should receive more attention in future epidemiological studies.