Predicting the academic standing of a student at the graduation time can be very useful, for example, in helping institutions select among candidates, or in helping potentially weak students in overcoming educational challenges. Most studies use individual course grades to represent college performance, with a recent trend towards using grade point average (GPA) per semester. It is unknown however which of these representations can yield the best predictive power, due to the lack of a comparative study. To answer this question, a case study is conducted that generates two sets of classification models, using respectively individual course grades and GPAs. Comprehensive sets of experiments are conducted, spanning different student data, using several well-known machine learning algorithms, and trying various prediction window sizes. Results show that using course grades yields better accuracy if the prediction is done before the third term, whereas using GPAs achieves better accuracy otherwise. Most importantly, variance analysis on the experiment results reveals interesting insights easily generalizable: individual course grades with short prediction window induces noise, and using GPAs with long prediction window causes over-simplification. The demonstrated analytical approach can be applied to any dataset to determine when to use which college performance representation for enhanced prediction.

The increasing availability of high temporal resolution neuroimaging data has increased the efforts to understand the dynamics of neural functions. Until recently, there are few studies on generative models supporting classification and prediction of neural systems compared to the description of the architecture. However, the requirement of collapsing data spatially and temporally in the state-of-the art methods to analyze functional magnetic resonance imaging (fMRI), electroencephalogram (EEG) and magnetoencephalography (MEG) data cause loss of important information. In this study, we addressed this issue using a topological data analysis (TDA) method, called Mapper, which visualizes evolving patterns of brain activity as a mathematical graph. Accordingly, we analyzed preprocessed MEG data of 83 subjects from Human Connectome Project (HCP) collected during working memory n-back task. We examined variation in the dynamics of the brain states with the Mapper graphs, and to determine how this variation relates to measures such as response time and performance. The application of the Mapper method to MEG data detected a novel neuroimaging marker that explained the performance of the participants along with the ground truth of response time. In addition, TDA enabled us to distinguish two task-positive brain activations during 0-back and 2-back tasks, which is hard to detect with the other pipelines that require collapsing the data in the spatial and temporal domain. Further, the Mapper graphs of the individuals also revealed one large group in the middle of the stimulus detecting the high engagement in the brain with fine temporal resolution, which could contribute to increase spatiotemporal resolution by merging different imaging modalities. Hence, our work provides another evidence to the effectiveness of the TDA methods for extracting subtle dynamic properties of high temporal resolution MEG data without the temporal and spatial collapse.