Critical transitions are abrupt, highly destructive, and exceptionally difficult-to-predict systemic shifts that frequently lead to massive casualties and extensive property damage. Complex systems—ranging from ecological and biological to financial and climate systems—exhibit tipping points at which they can undergo drastic changes. Predicting these systemic tipping points remains a complex and pressing challenge for the contemporary scientific community. Generic early warning signals (EWS) derived from dynamical systems theory exhibit inconsistent performance when applied to real-world noisy data. Recent studies have demonstrated that deep learning classifiers trained on synthetic data can enhance predictive performance. However, to the best of our knowledge, neither approach leverages historical, system-specific data. To address this, we propose a machine learning framework based on surrogate data, wherein classifiers are trained on empirical data from historical transitions. Concurrently, we introduce spatiotemporal diffusion and relaxation time as predictive indicators, both of which serve as effective precursors to critical transitions. Compared to traditional metrics such as variance and autocorrelation, the relaxation time indicator consistently exhibits an increasing trend preceding both oscillatory and non-oscillatory bifurcations, thereby demonstrating broader applicability. Ultimately, our objective is to identify early warning signals prior to critical transitions, providing a scientific basis and robust reference for policymakers and managers in formulating disaster mitigation strategies and making informed decisions.