The foetal sheep brain develops organised sleep states from 115-120 d gestational age (dGA, term 150 dGA) alternating between REM and NREM sleep. We aimed to investigate whether maturation of REM or NREM sleep generating structures leads to the development of distinct sleep states. The electrocorticogram (ECoG) was recorded from five unanaesthetised chronically instrumented foetal sheep in utero and was analysed every 5th day between 115-130 dGA by two different non-linear methods. We calculated a non-linear prediction error which quantifies the causality of the ECoG and applied bispectral analysis which quantifies non-linear interrelations of single frequency components within the ECoG signal. The prediction error during REM sleep was significantly higher than during NREM sleep at each investigated age (P<0.0001) coincidental with poor organisation of the rhythmic pattern in the ECoG during REM sleep. At 115 dGA, organised sleep states defined behaviourally were not developed yet. The prediction error, however, showed already different states of electrocortical activity that were not detectable using power spectral analysis. The prediction error of the premature NREM sleep ECoG decreased significantly during emergence of organised sleep states between 115 and 120 dGA and continued to decrease after the emergence of distinct sleep states (P<0.05). The prediction error of the premature REM sleep ECoG did not change until 120 dGA and began to increase at 125 dGA (P<0.05). Using bispectral analysis, we showed couplings between delta waves (1.5-4 Hz) and frequencies in the range of spindle waves (4-8 and 8-12 Hz) during NREM sleep that became closer during development. The results show that maturation of ECoG synchronisation mediating structures is important for the development of organised sleep states. The further divergence of the prediction error of NREM and REM sleep after development of organised sleep states reveals continuous functional development. Thus, complementary application of non-linear ECoG analysis to power spectral analysis provide new insights in the collective behaviour of the neuronal network during the emergence of sleep states.