Scoring Algorithms for a Computer-Based Cognitive Screening Tool: An Illustrative Example of Overfitting Machine Learning Approaches and the Impact on Estimates of Classification Accuracy

Computerized cognitive screening tools, such as the self-administered Computerized Assessment of Memory Cognitive Impairment (CAMCI), require little training and ensure standardized administration and could be an ideal test for primary care settings. We conducted a secondary analysis of a data set including 887 older adults (M age = 72.7 years, SD = 7.1 years; 32.1% male; M years education = 13.4, SD = 2.7 years) with CAMCI scores and independent diagnoses of mild cognitive impairment (MCI). A study by the CAMCI developers used a portion of this data set with a machine learning decision tree model and suggested that the CAMCI had high classification accuracy for MCI (sensitivity = 0.86, specificity = 0.94). We found similar support for accuracy (sensitivity = 0.94, specificity = 0.94) by overfitting a decision tree model, but we found evidence of lower accuracy in a cross-validation sample (sensitivity = 0.62, specificity = 0.66). A logistic regression model, however, discriminated modestly in both training (sensitivity = 0.72, specificity = 0.80) and cross-validation data sets (sensitivity = 0.69, specificity = 0.74). Evidence for strong accuracy when overfitting a decision tree model and substantially reduced accuracy in cross-validation samples was replicated across 500 bootstrapped samples. In contrast, the evidence for accuracy of the logistic regression model was similar in the training and cross-validation samples. The logistic regression model produced accuracy estimates consistent with other published CAMCI studies, suggesting evidence for classification accuracy of the CAMCI for MCI is likely modest. This case study illustrates the general need for cross-validation and careful evaluation of the generalizability of machine learning models.