Predicting unplanned medical visits among patients with diabetes: translation from machine learning to clinical implementation

Table 4 Training performance of classifiers with optimized parameters, presented as confusion matrices and balanced accuracy ± standard deviation across five-fold cross-validation runs

Classifier	Predicted: no unplanned visits	Predicted: ≥ 1 unplanned visit
Linear discriminant analysis
Actual: No Unplanned Visits	50.7% ± 1.1%	49.3% ± 1.1%
Actual: ≥ 1 Unplanned Visit	24.7% ± 0.7%	75.3% ± 0.7%
Average	63.0 ± 0.2%
Quadratic discriminant analysis
Actual: No Unplanned Visits	83.0% ± 0.2%	17.1% ± 0.2%
Actual: ≥ 1 Unplanned Visit	56.2% ± 0.2%	43.8% ± 0.2%
Average	63.4% ± 0.1%
Linear SVM
Actual: No Unplanned Visits	71.3% ± 0.8%	28.7% ± 0.8%
Actual: ≥ 1 Unplanned Visit	39.6% ± 0.7%	60.4% ± 0.7%
Average	65.8% ± 0.1%
Radial SVM
Actual: No Unplanned Visits	67.0% ± 1.1%	33.0% ± 1.1%
Actual: ≥ 1 Unplanned Visit	21.4% ± 0.4%	78.6% ± 0.4%
Average	72.8% ± 0.1%
Single hidden layer NN
Actual: No Unplanned Visits	50.8% ± 28.7%	49.2% ± 28.7%
Actual: ≥ 1 Unplanned Visit	31.5% ± 20.2%	68.5% ± 20.2%
Average	59.7% ± 7.9%
Triple hidden layer DNN
Actual: No Unplanned Visits	65.4% ± 15.0%	34.6% ± 15.0%
Actual: ≥ 1 Unplanned Visit	36.5% ± 14.3%	63.5% ± 14.3%
Average	64.4% ± 0.8%
XG boost
Actual: No Unplanned Visits	38.8% ± 35.3%	61.2% ± 35.3%
Actual: ≥ 1 Unplanned Visit	12.9% ± 11.2%	87.2% ± 11.2%
Average	63.0% ± 11.9%
Logistic regression
Actual: No Unplanned Visits	60.4% ± 0.2%		39.6% ± 0.2%
Actual: ≥ 1 Unplanned Visit	29.8% ± 0.2%		70.2% ± 0.2%
Average	65.3% ± 0.2%

Basic cross-validation was run for classifiers without hypermarameters (linear and quadratic discriminant analysis, logistic regression) and nested cross-validation for classifiers with hyperparameters (linear and radial SVM, single-layer NN and triple-layer DNN) to optimize hyperparameters
Cross-validation matrices show the training performance with respect to the actual class (rows) against the predicted class (columns), with ± standard deviation across cross-validation runs. DNN deep nets. NN neural nets. SVM support vector machines. XG boost extreme gradient boosting

ISSN: 1472-6947