Sample size calculation for training ensemble machine learning models on health data.

Abstract

Health research studies often suffer from small sample sizes, and training machine learning (ML) models requires large datasets. There is a dearth of literature on determining the adequate sample size for using ML models. We developed an empirically derived sample size calculator for ensemble ML models: random forests and two gradient-boosted decision trees (light gradient boosting machine [LGBM] and extreme gradient boosting [XGBoost]). This predicts the sample size required to achieve a pre-defined level of prognostic performance with a certain probability. Prognostic performance is defined as the sample area under the ROC curve (ROC-AUC) relative to the optimal model trained on the full (population) dataset. Our calculator's accuracy was compared to three common heuristics and a statistical approach to sample size calculation. For example, the median relative error sample size prediction was 25% to achieve 85% of the optimal performance with 90% certainty for LGBM. Our model has significantly better accuracy than other methods for tree-based ensemble ML models.