Landslides pose significant hazards in the mountainous region of Sikkim, India, necessitating accurate susceptibility mapping to mitigate risks. This study applies four machine learning models: Boosted Tree (BT), Gradient Boosting Machine (GBM), K-Nearest Neighbour (KNN), and Multilayer Perceptron (MLP) to develop a detailed landslide susceptibility map. Feature selection was performed using correlation analysis, the Boruta model, and multicollinearity tests, which identified 13 key landslide conditioning factors based on 1456 landslide inventory points. The GBM model demonstrated the highest predictive performance with an AUC of 0.99, followed by BT (AUC: 0.965), MLP (AUC: 0.940), and KNN (AUC: 0.895) in the testing dataset. The confusion matrix validation confirmed that GBM outperformed other models, achieving the highest F1 score (0.894) and accuracy (89.4%), followed by BT with an F1 score of 0.874 and accuracy of 87.8%. KNN and MLP displayed lower performance, with KNN showing an F1 score of 0.724 and accuracy of 72.6%, and MLP significantly underperforming with an F1 score of 0.096 and accuracy of 48.6%. Statistical significance testing using the Wilcoxon Signed-Rank Test revealed significant differences between BT and MLP (p = 0.018), while other model pairs exhibited no statistically significant performance differences. Additionally, the variable importance analysis highlighted Diurnal Temperature Range (DTR) as the most critical factor influencing landslide occurrence (43.99%), followed by elevation (21.59%). These findings provide valuable insights for policymakers and government authorities, enabling them to take necessary measures for effective landslide management in the vulnerable areas of Sikkim, confirming the efficacy of machine learning models for geohazard assessments.