Machine learning approached a 14-item shortened version of the Positive And Negative Sleep Appraisal Measure (PANSAM-14).

backgroundThis study aimed to develop a shortened version of the Positive And Negative Sleep Appraisal Measure (PANSAM) that accurately predicts the total score across its four subscales using a machine learning-based approach to streamline the subscales of PANSAM using eXtreme Gradient Boosting (XGBoost).

methodsWe collected data from 1,000 participants in South Korea through an online survey, measuring the PANSAM scores for each subscale. To identify the most representative items within each subscale, we used eXtreme Gradient Boosting (XGBoost), which can assess the predictive strength of each item based on the R² score. Additionally, we assigned optimal weights using the Symbolic Regression-Based Clinical Score Generator (SymScore) to ensure a refined and interpretable scoring system.

resultsWe developed the SymScore-based PANSAM-14, selecting 14 representative items across the four subscales: Subscale 1 (Items 12, 15, 16, and 24), Subscale 2 (Items 10, 18, and 21), Subscale 3 (Items 3, 11, and 19), and Subscale 4 (Items 1, 13, 25, and 29). These selected items demonstrated high accuracy in predicting subscale scores (R² = 0.94, 0.92, 0.94, and 0.94). We then developed a simple and interpretable scoring table using SymScore, achieving performance in predicting the total score comparable to the XGBoost-based version (R² = 0.93, 0.93, 0.94, and 0.95) while offering a practical and interpretable alternative.

conclusionThe SymScore-based PANSAM-14 exhibits high predictive accuracy for the total subscale scores. It can be used as a useful, reliable, and valid tool for assessing individuals' dysfunctional beliefs about sleep.