Random Forest-Based Prediction of Self-Reported Headache Complaint Indicators Among College Students Using Daily Activity and IoT Sensor Data

Abstract

Headache complaints among college students may be associated with daily activity patterns and environmental conditions. This study aimed to model self-reported headache complaint indicators using daily activity questionnaire data and Internet of Things environmental data without positioning the output as a clinical diagnosis. Environmental data were recorded using BME280, BH1750, and MQ-135 sensors, while daily activity data were collected using a self-report questionnaire. Sensor readings were aggregated by date and integrated with questionnaire responses to form 305 records from 59 respondents. Random Forest was optimized using Randomized Search CV and evaluated against Decision Tree and K-Nearest Neighbors under three feature scenarios Internet of Things features, daily activity features, and combined features. SMOTE was applied only to the training data, and model differences were assessed using the McNemar test and Wilcoxon signed-rank test. Random Forest achieved the highest overall performance in the daily activity questionnaire scenario, with 81.52% accuracy, 84.96% F1-score, and 85.32% mean cross-validation F1-score. In the combined scenario, Random Forest obtained 77.17% accuracy and 81.08% F1-score. Statistical testing showed significant differences only in selected McNemar comparisons, while Wilcoxon tests on cross-validation F1-scores were not significant across all comparisons. Daily activity data were more informative than date-level environmental sensor data in this dataset. The findings should be interpreted as exploratory numerical performance results rather than evidence of clinical causality or universal model superiority.

GS Cited Analysis