Abstract: Abstract: The Haibei Tibetan Autonomous Prefecture, located in the central Qilian Mountains, is widely underlain by permafrost. With ongoing climate warming, thaw-related geohazards occur more frequently, posing serious risks to lives and property; [Objective] susceptibility assessment is therefore essential to support sustainable development and early-warning efforts. [Methods] Taking Haibei Prefecture as the study area, we first screened 16 candidate predictors, analyzed their correlation and importance, and then applied dimensionality-reduction techniques to optimize the indicator set. Three machine-learning models were subsequently trained and evaluated. [Conclusion] The main findings are as follows: (1) The Kruskal–Wallis test indicates that distance to roads, multi-year mean air temperature, thawing index, elevation, and snow-cover days are the primary controls on thaw-related geohazards in Haibei. (2) Unlike conventional geological hazards, rainfall is generally not the decisive hazard-formative condition for thaw-related geohazards; therefore, susceptibility assessment should focus on indicators that characterize the thermal state of frozen ground and the intensity of engineering disturbances. (3) Dimensionality reduction effectively removed redundancy and high inter-correlation among variables while preserving the dominant information content of the original data; the resulting fused factors showed markedly higher importance, thereby improving the indicator set. (4) All three machine-learning models achieved better predictive performance after indicator optimization via dimensionality reduction, with the logistic-regression model performing best among the approaches.