Survival prediction for brain tumour patients is one of the crucial tasks in the medical field. Accurate survival prediction provides valuable information on the severity and seriousness of the patient and helps in treatment planning by assessing the results based on tumour characteristics and patient-specific factors. In this study, the 3D autoencoder model is used to capture both low-level features such as pixel-wise intensity variations and textural patterns, and high-level features such as overall tumour shape, spread, and relationships with surrounding brain structures by using a multi-level encoding-decoding strategy. The study uses BraTS 2020 dataset. The proposed 3D autoencoder model achieved a Dice similarity coefficient of 89% for tumour segmentation. For survival prediction, multiple feature sets were explored independently and in combination. Initially, radiomic statistical features such as skewness, kurtosis, and tumour region volumes were individually analyzed using machine learning algorithms. The separate c-index score for only radiomic features was 0.61 while the c-index score for only tumour volumes features was 0.62. Subsequently, a combined feature set was constructed by integrating 512 latent features extracted from the encoder portion of the segmentation model along with tumour region volumes and radiomic statistics, resulting in a total of 640 features. Although 640 features were obtained, only 20 features were chosen using random forest feature selection. This comprehensive feature fusion approach demonstrated improved prediction performance, achieving a concordance index (C index) of 0.73 and a mean average error (MAE) of 212 using a random forest regressor.