Abstract: The study of \alpha-decay half-lives is of great significance for revealing nuclear structure and quantum tunneling mechanisms. It also provides essential theoretical support for the discovery of new nuclides, the prediction of superheavy elements, and various applications in nuclear technology. In this work, two feedforward fully connected neural network models are employed to predict \alpha-decay half-lives. Model ML1 takes the proton number Z, neutron number N, even-odd effect term \delta, and \alpha-decay energy Q_\alpha as inputs. Model ML2 extends ML1 by incorporating the minimum angular momentum term l(l+1). Both models share identical network architectures and hyperparameter settings, allowing a systematic comparison of the impact of angular momentum on model performance. The study shows that both ML1 and ML2 exhibit good predictive performance for \alpha-decay half-lives, with root-mean-square deviations (RMS) of 0.46 and 0.44 logarithmic units relative to the experimental values, respectively. These results are significantly better than those obtained with the finite-range droplet model (FRDM) combined with the Viola-Seaborg (VS) formula. In addition, both models are capable of successfully reproducing the sudden changes in \alpha-decay half-lives for nuclei located near shell closures, such as Z=82 and N=126. Due to the limited number of nonzero angular momentum \alpha-decay samples in the experimental data, although ML2 incorporates the angular momentum term, the RMS difference compared to ML1 is small. This indicates that in neural networks with sufficient parameters, the angular momentum information may be overshadowed by the optimization of other parameters, resulting in limited improvement in prediction accuracy.