Biomedical big data are usually high dimensional and collected in the form of a continuous influx of new features. Online Feature Selection (OFS) is a promising way to manage and analyze such data, as OFS circumvents the huge computation cost brought by simultaneously considering all the features, and can also dynamically maintain a distribution-fitting feature subset on the fly. However, almost all the OFS solutions are based on a naive premise that all features are of the same type, overlooking the fact that real biomedical data set usually consists of heterogeneous numerical and categorical features. This paper therefore proposes a new approach to Online Heterogeneous Feature Selection (OHFS), which dynamically maintains a feature subset that maximizes the number of neighborhood sets where all the objects within each neighborhood set are of the same class. To appropriately partition the objects into neighborhood sets, a density-guided relation is proposed, which adaptively forms non-overlapping neighborhood sets by detecting spatially compact objects. A unified density measure is also presented to avoid information loss in processing heterogeneous features. It turns out that the proposed approach features parameter- free, interpretability, and efficiency. It is capable of maintaining a concise feature subset while receiving any type of feature. Extensive experimental evaluations demonstrate its superiority.