Textionnaire: An NLP-Based Questionnaire Analysis Method for Complex and Ambiguous Task Decision Support

Questionnaires are one of the most popular means of gathering information in many domains. Traditional health surveys, particularly those focusing on specific populations of patients or certain conditions tend to have sample sizes that are too small to apply advanced artificial intelligence methods. In this study, we introduce a method that enables researchers to apply deep learning algorithms on small questionnaire datasets. We introduce a method to represent the information of fixed response questions of questionnaires in text format analogous to a word presentation for a number stored in expanded form. Then we use a pre-trained, domain-relevant deep learning model for questionnaire classification. We applied this method to two databases: i) a hospital-maintained database of mental health questionnaires linked with hospital administrative resource uti-lization data, and ii) a similarly structured publicly available dataset, the National Survey on Drug Usage and Health. In all experiments, the deep pre-trained model had a better Area Under Receiver Operating Characteristic Curve (AUROC) than a 3-layer artificial neural network (ANN). The proposed method enhanced classification performance by 22.95 % on Mental Health private data and 16.679% on small sub-sample of NSDUH data compared to a baseline ANN machine learning algorithm. Also, our technique improved the performance by up to 15.38% when tested for generalizability of classification.