NLP-Based Pain Prediction Using Machine Learning and Boosted Models: A Comparative Analysis of TF-IDF and BoW Representations with Headache Data

Abstract

Pain assessment traditionally relies on subjective self-reports, which can introduce variability due to differences in individual pain tolerance and perception. The aim of this study is to predict a personalized pain description encompassing pain severity, sensory impact, and threshold using Natural Language Processing (NLP) techniques and machine learning (ML) models trained on patient-reported pain descriptions and demographic data. Specifically, Bag-of-Words (BoW) and Term Frequency-Inverse Document Frequency (TF-IDF) representations were employed to transform textual pain narratives into numerical features for multi-output regression modeling. The dataset consists of 153 patients from different regions of Türkiye, with pain descriptions and demographic information used as predictors. A comprehensive ML framework, including boosting models, tree-based models, linear models was evaluated using R2, MSE, MAE, and MAPE metrics. Results indicate that boosting-based models, particularly Gradient Boosting (R2?=?0.989, MSE?=?0.056) and XGBoost (R2?=?0.982, MSE?=?0.090), achieved the highest predictive accuracy when TF-IDF features were used. While the inclusion of self-reported headache severity improved model performance, high accuracy was still attainable without it, suggesting that linguistic features alone can provide robust pain sensitivity predictions. TF-IDF proved superior to BoW in boosting models, while BoW performed better in linear regression models, underscoring the importance of feature representation. This study highlights the potential of NLP-driven ML models for automated pain assessment, reducing reliance on subjective self-reports and contribute to the growing field of AI-driven healthcare, offering a scalable and objective approach to pain prediction that can inform personalized pain management strategies. © 2025 Elsevier B.V., All rights reserved.