Neuroscience　studies　have　demonstrated　that　inherent　interactions　in　human　brain　are　associated　with　various　brain　diseases.　The　functional　connectivity　(FC)　analysis　in　the　resting-state　functional　magnetic　resonance　imaging　(rs-fMRI)　has　attracted　increasing　attention.　Recently,　the　application　of　deep　learning　algorithms　has　shown　great　potential　in　exploring　the　diagnosis　biomarker　for　brain　disease.　However,　the　high　dimensionality　and　small　sample　size　properties　of　the　FC　patterns　usually　restrict　the　training　of　deep　learning　models　and　limit　the　model　generalization　performance.　To　solve　this　problem,　this　paper　proposed　a　novel　prediction　model　with　a　two-stage　design　to　extract　and　classify　FC　features　for　the　brain　diseases　diagnosis.　Our　model　employed　a　flexible　configuration　that　contained　a　deep　learning-based　FC　feature　extraction　stage　followed　by　a　separate　classification　stage.　In　the　first　stage,　stacked　sparse　denoising　autoencoders　were　employed　to　extract　and　learn　the　abstract　feature　representations　of　the　initial　FC　patterns　with　both　unsupervised　and　supervised　learning　steps.　In　the　second　stage,　the　learnt　features　were　used　to　train　a　separate　simple　classifier　for　brain　disease　classification.　We　conducted　experiments　on　a　large-scale　fMRI　dataset　to　construct　FC　patterns　and　distinguish　autism　patients　from　healthy　controls.　Results　showed　that　our　two-stage　model　outperformed　the　competing　classification　methods　and　showed　robust　generality　for　different　brain　nodes　parcellation.　Our　study　indicated　that　this　flexible　two-stage　model　design　can　further　improve　the　classification　performance　and　provide　a　promising　approach　for　the　computer-aided　diagnosis　of　brain　disorders.　©　2021　IEEE.