Epileptic　seizure　prediction　has　been　widely　investigated　to　provide　timely　warning　for　patients　with　epilepsy　in　clinical　practice.　However,　current　methods　have　limited　prediction　capabilities　as　it　remains　a　grand　challenge　to　properly　handle　the　intensive　intra-individual　and　inter-individual　variabilities　along　with　the　information　fusion　of　EEG　features　from　various　domains.　To　address　the　issues,　this　study　proposes　a　multi-domain　EEG　feature　fusion　network　with　supervised　contrastive　learning　(namely　MDFFNet-SCL)　for　epileptic　seizure　prediction.　MDFFNet-SCL　first　employs　five　types　of　data　transformations　to　generate　the　positive　samples　of　anchor　EEG　samples.　Multi-domain　EEG　features　are　then　computed　to　reveal　the　multi-level　information　related　to　epileptic　seizure,　which　is　fed　into　EEG　encoders　with　multi-scale　convolution　network　to　independently　extract　the　deep　EEG　representations.　In　consideration　of　the　inter-domain　and　intra-domain　correlations,　deep　canonical　correlation　analysis　(DCCA)　is　exploited　to　merge　the　deep　EEG　representations　in　multiple　domains　and　form　the　shared　representations.　Finally,　MDFFNet-SCL　applies　cross-entropy　classification　loss　combined　with　CCA　and　SCL　losses　for　model　training,　where　CCA　and　SCL　losses　are　utilized　to　increase　the　correlation　between　multi-domain　features　and　alleviate　the　intra-individual　and　inter-individual　variabilities.　Experimental　results　on　two　public　EEG　datasets　(CHB-MIT,　Kaggle)　indicate　that　(1)　multi-domain　feature　fusion　and　supervised　contrastive　learning　schemes　have　demonstrated　superiority　in　capturing　discriminative　EEG　features　between　the　preictal　and　interictal　states,　(2)　MDFFNet-SCL　significantly　outperforms　the　state-of-the-art　counterparts　with　accuracy　and　AUC　up　to　98.70%,99.89%　for　CHB-MIT　EEG　dataset　and　96.95%,99.27%　for　Kaggle　EEG　dataset.　The　solution　can　be　extended　to　design　a　more　general　&　robust　framework　for　cross-subject　or　cross-dataset　seizure　prediction.　©　2025　Elsevier　Ltd