Smart cities, smart grids, and in general Smart human-centric EcoSystems (SES) emerge from the co-existence of many systems that operate according to independently owned specifications, and evolve over time. SES may fail despite the correct behavior of the systems that comprise the SES. Fully testing SES on testbed to prevent failures in production is impossible, and scenarios that may lead to catastrophic SES failures are unavoidable. In this paper we frame the core issues of SES failures, and propose Smart human-centric Ecosystem Monitoring, SEM, an approach that predicts SES failures to enable corrective actions for mitigating the catastrophic effects of failures. SEM identifies failure-prone scenarios from the reconstruction error of SES indicators, that is, metric values that SEM collects from SES at constant frequency. SEM computes the reconstruction error with a suitably trained denoising autoencoder combined with a Transformer architecture. The results of experimenting with a peer-to-peer ride-sharing ecosystem operating in San Francisco confirm that SEM can effectively predict SES failures early enough to activate preventing actions, and indicate the generalizability of SEM with continual learning.