Natural Language Processing for Phenotyping: A Feasibility Study in Predicting ASA Physical Status from Preoperative Clinical Narratives (Motivated by the Study on ASA Classification Prediction from Preoperative Notes by Chung et al.)

This report explores the use of natural language processing to automate clinical phenotyping by predicting the American Society of Anesthesiologists physical status classification based on preoperative evaluation notes. Motivated by a recent study that demonstrated the feasibility of text-based severity prediction, we assess how unstructured clinical narratives can serve as rich, standalone inputs for risk stratification. Using a large, standardized dataset of preoperative notes, we trained and evaluated multiple machine learning models, including random forests, linear classifiers, word embedding models, and transformer-based deep learning architectures. Performance was benchmarked using standard classification metrics, and interpretability was examined through Shapley value analysis of influential clinical terms. The results showed that models using the full narrative outperformed those using isolated sections, with the highest accuracy achieved by domain-adapted deep learning models. Most misclassifications occurred between adjacent severity categories, and reviewer analyses indicated the models often made clinically plausible predictions. Our findings demonstrate that natural language processing can effectively extract phenotypic signals from clinical text, reducing reliance on structured data and manual review. We highlight essential dataset characteristics, model evaluation strategies, and future directions for improving accuracy, generalizability, and clinical adoption of automated text-based phenotyping methods.