INTRODUCTION: Machine learning classification algorithms have shown human-level performance, but noise in the dataset (e.g. due to measurement errors, bad data processing, random error) can reduce accuracy of the predictions. OBJECTIVES: To investigate how: 1) different degrees of artificial noise and 2) selected variable transformations affect performance of five classification algorithms. METHODS: A synthetic dataset in preclinical Alzheimer’s Disease setting was used. It contained data of 2336 cognitively normal (at baseline) subjects, six validated baseline predictors, and a binary outcome indicating development of symptomatic AD in the future. Five classification algorithms - Generalised Linear model with LASSO penalty (GLM), Random Forest (RF), Support Vector Machine (SVM), Tree based and Linear based Gradient Boosting (GBT, GBL) – were trained on (1) reference and (2) transformed (trigonometric, exponential/logarithmic, multiplicative transformations on predictors) datasets and tested on (1) auxiliary noisy validation datasets of various degree noise and on the (2) transformed validation dataset, respectively. Eighty-to-twenty 10-fold cross-validation approach was used. RESULTS: Accuracy of prediction (F1-Score) of algorithms trained and validated on noise-free datasets was: GBL=0.97, GBT=0.96, RF=0.91, GLM=0.80; SVM=0.80. Percentage change in F1-Score when validated on dataset with the highest level of noise was: GLM=-2.2%, SVM=-0.2%, RF=-41.5%, GBT=-75.8%, GBL=-78.0%. Percentage change in accuracy when validated on dataset with transformations was: RF=6.59%, GBL=0.31%, GLM=0.1%, GBT=0.1%, SVM=-0.1%. CONCLUSIONS: (1) All algorithms can extract data patterns and maintain their predictive performance under extreme transformations. (2) Boosting and tree-based algorithms drastically outperform SVM and LASSO algorithms in noise-free data (3) SVM and LASSO face no decrease in performance, while Tree-based algorithms can be rendered unusable by heavy noise. (4) Ability to identify the source and degree of noise in the validation dataset is necessary to select appropriate classification algorithm. (5) Data-specific tuning is critical to assure high performance even under heavy noise.