Identifying important biomarkers for complex diseases is crucial for developing novel prevention, diagnosis, and treatment strategies. This paper proposes PermFIT, a permutation-based feature importance test, to estimate and test feature importance within complex machine learning models (deep neural networks, random forests, support vector machines). PermFIT is computationally efficient and doesn't require model refitting. Extensive numerical studies demonstrate PermFIT's validity and its ability to improve prediction accuracy. Applications to TCGA kidney tumor data and HITChip atlas data showcase its practical utility in identifying key biomarkers and enhancing model performance.

Xinlei Mi, Baiming Zou, Fei Zou, Jianhua Hu