Early prediction of student academic performance can help higher education institutions identify students at risk of dropping out and provide timely support. This dissertation proposes an evolutionary higher-order extreme learning machine (EHO-ELM) algorithm for predicting student academic success and dropout. The proposed approach extends the traditional extreme learning machine (ELM) method by incorporating higher-order multi-cube units (MCUs) into a single-layer neural network (SLNN). The performance of ELM can be affected by the random initialization of hidden-layer weights and thresholds. In addition, ELM typically adopts low-order neurons, which limits the SLNN’s ability to capture more complex interactions amongst input variables. MCUs allow the model to capture nonlinear relationships while avoiding the rapid increase in parameters that can occur with conventional higher-order neurons. EHO-ELM uses a modified genetic algorithm (GA) with self-adaptive parameters to circumvent the suboptimal generalization performance associated with hidden-layer initialization and to identify appropriate network structures. The evolutionary process generates, trains, and evaluates a population of MCU-based SLNNs and optimizes their hidden-layer weights, thresholds, and sub-cube configurations. The output weights are calculated analytically using the Moore–Penrose pseudoinverse, retaining the computational efficiency and structural simplicity that characterize the ELM algorithm. The proposed system is evaluated using a dataset containing a range of student-related variables, including demographic characteristics, higher-education pathways, family and financial circumstances, early academic performance, and macroeconomic conditions. Its performance is compared with 14 alternative machine-learning methods (ELM, online sequential ELM, MCU-ELM, a decision tree, and 10 backpropagation-based neural-network algorithms). The experimental results show that EHO-ELM achieves higher classification accuracy than the other methods in the student-outcome prediction task. The statistical significance of these differences is further examined using the Wilcoxon signed-rank test. Overall, the findings suggest that combining evolutionary optimization with higher-order neural architectures can improve the performance of ELM-based models in educational data mining applications.