The influence of sample size, modelling algorithms, and hyperparameters on bootstrap-based prediction stability in clinical prediction model development: An empirical study using the HOPE dataset

Abstract Background The stability of clinical prediction models has emerged as an important factor in model development. Models producing unstable or unreliable predictions can ultimately lead to improper clinical decisions and harm patients when applied in clinical practice. Previous studies have highlighted the role of sample size in prediction stability, with larger samples delivering increased stability. However, the role of other modelling decisions remains unclear. This study considered three modifiable factors influencing the stability of predictions: sample size, modelling algorithm (i.e., parametric regression and machine learning algorithms), and hyperparameter tuning. Methods We used the previously published HOPE dataset to develop a prediction model for anxiety at palliative care admission, incorporating 34 candidate predictors. To examine the impact of sample size, we generated seven datasets of varying sizes through random sampling. A total of 21 modelling scenarios were evaluated by combining these datasets with three modelling approaches: LASSO logistic regression, random forest, and XGBoost. Models were tuned using cross-validation and assessed for discrimination (AUC) and calibration. Prediction stability was evaluated using a bootstrap-based framework, including the mean absolute prediction error (MAPE) and related instability plots. Results In this clinical dataset, larger sample sizes were associated with lower average MAPE and a higher proportion of predictions with MAPE ≤ 5% for LASSO, with limited additional changes beyond the reference size ( n = 3,875). LASSO also showed lower AUC optimism and calibration slopes closer to 1 across increasing sample sizes. Random forest and XGBoost showed higher discrimination than LASSO in some scenarios, but had higher average MAPE, a lower proportion of predictions with MAPE ≤5%, and calibration slopes further from 1 across sample sizes. Conservative hyperparameter settings, such as shallower trees and slower learning rates, were associated with lower MAPE but slightly lower discrimination. Conclusions Our results provide empirical insights into bootstrap-based prediction stability in clinical prediction model development. Model stability appeared to depend not only on sample size but also on modelling approach and hyperparameter tuning. For machine learning models, considering stability alongside marginal gains in accuracy may be important. Further research should explore alternative validation strategies and optimal tuning practices to guide robust clinical prediction model development.