Patients living with HIV are eligible for treatment when their CD4 count is less than 350 cells/mm 3 . The patients receive antiretroviral treatment (ART) which they need to take every day for the rest of their life. To maintain treatment effcacy, it is necessary to avoid the event of treatment failure. In order to assist physicians monitoring HIV patients this paper propose temporal data mining to predict treatment outcome by providing visual representation of prediction results. Temporal abstraction is used to classify time series data into discrete categories, each represented typically with a symbol. Artificial neural networks are used in this study where the problem of unbalanced data size occurs during the learning process. Two under-sampling techniques are proposed. With the nearest samples to cluster center technique, accuracy is achieved at levels higher than 85% and the discovered patterns correspond with real world diagnosis where viral load is the primary feature to predict treatment outcome.