Recently, deep learning based methods have shown potential as alternative approaches for lung time difference electrical impedance tomography (tdEIT) reconstruction other than traditional regularized least square methods, that have inherent severe ill-posedness and low spatial resolution posing challenges for further interpretation. However, the validation of deep learning reconstruction quality is mainly focused on simulated data rather than in vivo human chest data, and on image quality rather than clinical indicator accuracy. In this study, a variational autoencoder is trained on high-resolution human chest simulations, and inference results on an EIT dataset collected from 22 healthy subjects performing various breathing paradigms are benchmarked with simultaneous spirometry measurements. The deep learning reconstructed global conductivity is significantly correlated with measured volume-time curves with correlation > 0.9. EIT lung function indicators from the reconstruction are also highly correlated with standard spirometry indicators with correlation > 0.75.Clinical Relevance— Our deep learning reconstruction method of lung tdEIT can predict lung volume and spirometry indicators while generating high-resolution EIT images, revealing potential of being a competitive approach in clinical settings.