Recently, Electronic Health Records (EHR) have become valuable for enhancing diagnosis prediction. Despite the effectiveness of existing deep learning based methods, one unified embedding fails to capture multiple disease causes of a patient. Even though naive adoption of multi-head attention could produce multiple cause vectors, a strong correlation between these cause representations might mislead the model to learning statistical spurious dependencies between cause vectors and diagnosis predictions. Hence, in this work, we propose a novel Multi-Cause Learning framework for Diagnosis Prediction, named MulDiag. Our Multi-Cause Network extracts multiple cause representations for a patient. We introduce HSIC (Hilbert-Schmidt Independence Criterion) to measure the dependencies among each pair of cause representations. Further, sample re-weighting techniques are utilized to conduct cause decorrelation. Experimental results on a publicly available dataset demonstrate the effectiveness of our method.