Acute Lymphocytic Leukaemia (ALL) is identified as one of the leading causes of childhood death by the World Health Organisation. Although survival analysis of childhood ALL plays a major role in understanding patients’ outcomes, it is less popular among medical practitioners. Especially in ALL data, most patients survive for a long time, and hence the data is complicated with censored observations. In this context, traditional survival models are less efficient, and machine learning models have been identified to be better. This study aimed to compare the performance of both traditional survival models and machine learning methods specifically designed for censored data in predicting survival and identifying risk factors. Data from 1,567 ALL patients treated at the Maharagama National Cancer Institute between 2016 and 2023 were analysed, with a censoring rate of 83%. Prior to modelling, data were weighted with the Inverse Probability of Censoring Weighting (IPCW) technique to handle imbalance, representing a novel application in Sri Lankan medical studies where this method has not yet been reported. Six survival models, including Cox proportional hazard, Cox with elastic-net regularisation, Cox with XG-Boost, Random Survival Forest (RSF) and Survival Support Vector Machine (SVM), were evaluated using the concordance index and time-dependent AUC. RSF with IPCW showed superior performance, closely estimating overall survival at 83.37%, where the Kaplan-Meier estimate of 83.49%. Initial WBC count, risk stratification, and age at diagnosis were identified as significant risk factors through feature importance analysis. Furthermore, among the two main subtypes of Acute Lymphoblastic Leukaemia, B-ALL patients showed a higher survival rate (83.4%) than T-ALL patients (78.5%).