Sleep　staging　plays　a　significant　role　in　diagnosing　sleep-related　diseases　and　assessing　sleep　quality.　Recent　years　have　witnessed　a　remarkable　advancement　in　deep　learning　for　automatic　sleep　staging.　However,　the　classification　accuracy　for　certain　stages　remains　unsatisfactory　due　to　imbalanced　sleep　data.　Moreover,　the　sleep　stages　are　easily　confused　with　each　other,　further　leading　to　low　classification　accuracy.　To　address　the　issues,　we　propose　a　multi-task-based　feature-reconstruction　sleep　staging　method,　comprising　a　feature　extraction　module　and　a　multi-task　module.　In　the　multi-task　module,　specifically,　the　main　task　branch　encodes　the　features　extracted　by　the　feature　extraction　module　for　sleep　staging,　while　the　auxiliary　task　branch　randomly　masks　the　features　before　performing　the　same　encoding.　Subsequently,　based　on　the　classification　results　of　the　main　task,　a　dimensional　reconstruction　is　employed　to　reconstruct　the　original　features　from　the　encoded　features　for　confused　sleep　stages.　The　overall　loss,　formed　by　combining　the　classification　loss　and　the　reconstruction　loss,　is　utilized　to　constrain　the　model.　We　evaluate　the　performance　of　our　proposed　model　using　three　public　datasets,　and　the　results　demonstrate　the　effectiveness　of　the　method　in　solving　the　problem　of　low　accuracy　in　confused　sleep　stages　classification.