Reliable near-time forecast of COVID-19 hospital admissions can help enable effective resource management which is vital in reducing pressure from healthcare services. The use of mobile network data has come to attention in response to COVID-19 pandemic leveraged on their ability in capturing people social behaviour. Crucially, we show that there are latent features in irreversibly anonymised and aggregated mobile network data that carry useful information in relation to the spread of SARS-CoV-2 virus. We describe development of the forecast models using such features for near-time prediction of COVID-19 hospital admissions. In a case study, we verified the approach for two hospitals in Sweden, Sahlgrenska University Hospital and Södra Älvsborgs hospital, working closely with the experts engaged in the hospital resource planning. Importantly, the results of the forecast models were used in year 2021 by logisticians at the hospitals as one of the main inputs for their decisions regarding resource management.