Short-form video (SFV) has become a globally popular form of entertainment in recent years, appearing on major social media platforms. However, current research indicate that short video addiction can lead to numerous negative effects on both physical and psychological health, such as decreased attention span and reduced motivation to learn. Additionally, Short-form Video Addiction (SFVA) has been linked to other issues such as a lack of psychological support in real life, family or academic pressure, and social anxiety. Currently, the detection of SFVA typically occurs only after users experience negative consequences. Therefore, we aim to construct a short video addiction dataset based on social network behavior and design an early detection framework for SFVA. Previous mental health detection research on online social media has mostly focused on detecting depression and suicidal tendency. In this study, we propose the first early detection framework for SFVA EarlySD. We first introduce large language models (LLMs) to address the common issues of sparsity and missing data in graph datasets. Meanwhile, we categorize social network behavior data into different modalities and design a heterogeneous social network structure as the primary basis for detecting SFVA. We conduct a series of quantitative analysis on short video addicts using our self-constructed dataset, and perform extensive experiments to validate the effectiveness of our method EarlySD, using social data and heterogeneous social graphs in the detection of short video addiction.