Identifying changes in the parameters of a dynamical system can be vital in many diagnostic and sensing applications. Sensitivity vector fields (SVFs) are one way of identifying such parametric variations by quantifying their effects on the morphology of a dynamical system's attractor. In many cases, SVFs are a more effective means of identification than commonly employed modal methods. Previously, it has only been possible to construct SVFs for a given dynamical system when a full set of state variables is available. This severely restricts SVF applicability because it may be cost prohibitive, or even impossible, to measure the entire state in high-dimensional systems. Thus, the focus of this paper is constructing SVFs with only partial knowledge of the state by using time-delay coordinate embeddings. Local models are employed in which the embedded states of a neighborhood are weighted in a way referred to as embedded point cloud averaging. Application of the presented methodology to both simulated and experimental time series demonstrates its utility and reliability.