Accurate prediction of cadmium (Cd) ecotoxicity to and accumulation in soil biota is important in soil health. However, very limited information on Cd ecotoxicity on naturally contaminated soils. Herein, we investigated soil Cd ecotoxicity using Folsomia candida, a standard single-species test animal, in 28 naturally Cd-contaminated soils, and the back-propagation neural network (BPNN) model was used to predict Cd ecotoxicity to and accumulation in F. candida. Soil total Cd and pH were the primary soil properties affecting Cd toxicity. However, soil pH was the main factor when the total Cd concentration was ˂ 3 mg kg− 1. Interestingly, correlation analysis and the K-spiked test confirmed nutrient potassium (K) was essential for Cd accumulation, highlighting the significance of studying K in Cd accumulation. The BPNN model showed greater prediction accuracy of collembolan survival rate (R2 = 0.797), reproduction inhibitory rate (R2 = 0.827), body Cd concentration (R2 = 0.961), and Cd bioaccumulation factor (R2 = 0.964) than multiple linear regression models. Then the developed BPNN model was used to predict Cd ecological risks in 57 soils in southern China. Compared to multiple linear regression models, the BPNN models can better identify high-risk regions. This study highlights the potential of BPNN as a novel and rapid tool for the evaluation and monitoring of Cd ecotoxicity in naturally contaminated soils.