Developing a Machine-learning Algorithm for Classifying Parents’ Responses to Web-based Early Language Screening

Objectives: This study aims to address the issue of decreased reliability caused by inaccurate responses in web-based early language screening tests for infants and toddlers by proposing a method for detecting the sincerity of parental responses using a machinelearning algorithm. Methods: The language development screening test used in this study was developed so that parents could respond directly online, and a total of 453 parents participated in the screening test. Additionally, 226 randomly generated responses were created using a random number generator. According to the criteria set by the researchers, the response types were categorized into sufficient effort responses, insufficient effort responses, and randomly generated responses. Various machine learning algorithms, including XGBoost, LightGBM, and MLP, were used to train the response data, and the accuracy of the model was assessed using K-fold cross-validation. Results: Based on the established criteria, the responses were classified as 220 sufficient effort responses, 233 insufficient effort responses, and 226 randomly generated responses. The ensemble voting demonstrated a high accuracy of approximately 89.7% in detecting insufficient effort responses and randomly generated responses, effectively classifying parents’ insufficient effort responses. Conclusion: The machine-learning-based approach proposed in this study can be effectively utilized not only in web-based language screening for infants and toddlers but also in detecting insufficient effort responses in self-reported surveys. This approach is expected to enhance the reliability of research data and facilitate more accurate assessments.