Sensor-integration enables the creation of smart structures that can feel and provide insights into the loads they experience. To accurately reconstruct loads and deformations from strain measurements, so-called shape sensing and load sensing, solving inverse problems using computationally efficient techniques is necessary. Physics-informed neural networks (PINNs) have recently gained popularity for ad-dressing these challenges. However, many studies have not examined the effects of varying load case data and sensor positions in real structures compared to those used during training. This paper presents a sensitivity study of a PINN designed for solving inverse problems (iPINN) related to deformation reconstruction from strain measurements in shape sensing. The study analyses the impact of three different factors on the prediction accuracy on the example of a rectangular aluminum tube undergoing bending loads. First, the impact of different training datasets is analyzed for this part for three cases of load scenarios: one with only bending loads, one with bending and overlaid torsional loads, and one with a mix of both. The data is derived from a finite element (FE) simulation with different load amplitudes. As second factor, the influence of different sensor quantities and positions is analyzed on the example of configurations with 5, 10, 15, and 50 strain rosettes. Last, the influence of deviations from sensor positions used during training is investigated. The results show that there is still a dependency on training datasets for prediction accuracy, especially to transfer predictions from torsional to pure bending loads. Besides, a high influence on sensor positions and measurements was found.