We propose a data-driven method for geometric feature detection in 3D scattered point clouds and its integration into an adaptive spline-approximation framework. For each point, we compute geometric quantities derived from the covariance matrix of its local neighborhood; these features are then used as inputs to a multi-layer perceptron applied independently to every point. This results in an effective technique for identifying points near edge features in the input model. The detected features guide an adaptive reconstruction scheme based on Locally Refined B-splines (LR B-splines), exploiting their anisotropic refinement capabilities while avoiding the cost of traditional error-estimation-driven adaptive loops. Numerical experiments demonstrate that the proposed approach improves the robustness of the feature detection and the quality of the reconstruction compared with other local spline refinement strategies.