Preeclampsia is a complex pregnancy disorder that poses significant health risks to both mother and fetus. Despite its clinical importance, the underlying molecular mechanisms remain poorly understood. In this study, we developed an integrative deep learning and bioinformatics approach to identify potential biomarkers for preeclampsia. Three microarray datasets related to preeclampsia were initially analyzed to select a preliminary gene subset based on $P$-values. Feature selection was then performed in two consecutive rounds: first, the Fisher score method was applied to extract significant genes, followed by the minimum Redundancy Maximum Relevance method to refine the subset further. These selected gene subsets were trained using our proposed Attention-based Convolutional Neural Network (AttCNN), which achieved the highest classification accuracy compared with other models. From the experiments, a set of 58 common genes was identified between differentially expressed genes and the final optimized subset. Here, Gene Ontology and KEGG pathway enrichment analyses highlighted key biological processes and pathways associated with preeclampsia. Subsequently, a protein-protein interaction network was constructed, identifying 10 hub genes: TSC22D1, IRF3, MME, SRSF10, SOD1, HK2, ERO1L, SH3BP5, UBC, and ZFAND5. Further analysis of gene regulatory networks, including transcription factor-gene, gene-microRNA, and drug-gene interactions, revealed that seven hub genes (HK2, SRSF10, SOD1, ERO1L, IRF3, MME, and SH3BP5) were strongly associated with preeclampsia. Molecular docking analysis showed that HK2, SH3BP5, and SOD1 exhibited significant binding affinities with two preeclampsia drugs. These findings suggest that the identified hub genes hold promise as biomarkers for early prognosis, diagnosis, and potential therapeutic targets for preeclampsia.