Skip to main content

Table 1 Main strengths and weaknesses of popular predictive models.

From: A comparative analysis of predictive models of morbidity in intensive care unit after cardiac surgery – Part I: model planning

Model Strengths Weaknesses
BL Easy to construct (quick learning and low computational overhead); low sensitivity to missing data; recursive updating. Low performance with clearly non-normal data or manifestly non homoscedastic distributions; poor calibration.
BQ Easy to construct (quick learning and low computational overhead); low sensitivity to missing data; recursive updating. Low performance with clearly non-normal data; poor calibration.
k NN Very intuitive; no statistical assumption about the data; good classification if number of samples is large enough. Critical choice of neighbourhood size and metric; large storage requirements and time consuming for large databases.
LR Parsimony (few model parameters); interpretability of the parameters in terms of odds. Outliers can affect results significantly; certain assumptions about predictors; difficult updating.
ISS Very simple use in clinical practice; strong intuitive appeal; widespread use in heart surgery. Worse performance than more complex models; difficult customization and updating.
ANN No statistical assumption about data; ability to estimate non-linear relationships between input data and outputs. Long training process; experience needed to determine network topology; poor interpretability; difficult updating.
  1. BL, Bayes linear; BQ, Bayes quadratic; kNN, k-nearest neighbour; LR, logistic regression; ISS, integer scoring systems; ANN, artificial neural network.