Functional magnetic resonance imaging (fMRI) is a powerful tool for diagnosing neurological disorders. However, accurately distinguishing disease-related features from confounding covariates (e.g., age, gender, site) and individual variability remains a challenge. To tackle this problem, we propose a novel graph disentanglement learning (GDL) framework that decomposes the latent features from fMRI images into 3 components: disease-related features, covariate-related features, and individual variations. The covariate-related features are learned by aligning 2 subject similarity matrices between the features and the true covariates. The disease-related features are guided by a classification loss. We validate our method on 3 fMRI datasets: ADHD-200, schizophrenia (SCZ), and Presbycusis. The method outperforms existing approaches by an average of 0.5%, 1.7%, and 2.1% in accuracy on the 3 datasets respectively. Ablation studies confirm that our model is robust to hyperparameter selection. The disease-associated regions identified by our model align with established clinical findings. These results suggest that GDL is a promising tool for fMRI-based disease diagnosis and biomarker discovery. The code is publicly available at https://github.com/perpetualmachine/GDL_MICCAI .