Table 2.

Weighted F1 scores for classification fields and mean accuracy for token extractor fields on full training data sample (n = 2066)

Data elements	Logistic regression	AdaBoost classifier	Random forest	SVM	CNN	LSTM	Majority class accuracy
Gleason grade—primary	0.978	0.971	0.941	0.932	0.981	0.628	0.709
Gleason grade—secondary	0.958	0.943	0.913	0.912	0.968	0.576	0.467
Gleason grade—tertiary	0.923	0.930	0.844	0.886	0.930	0.741	0.901
Tumor histology	0.989	0.995	0.995	0.993	0.995	0.994	0.991
Cribriform pattern	0.963	0.981	0.963	0.968	0.987	0.966	0.997
Treatment effect	0.981	0.979	0.981	0.981	0.981	0.973	0.985
Tumor margin status	0.941	0.953	0.888	0.918	0.950	0.630	0.799
Benign margin status	0.977	0.975	0.972	0.981	0.978	0.967	0.997
Perineural invasion	0.944	0.978	0.938	0.929	0.972	0.613	0.771
Seminal vesicle invasion	0.943	0.974	0.940	0.965	0.976	0.784	0.904
Extraprostatic extension	0.954	0.953	0.882	0.939	0.961	0.778	0.712
Lymph node status	0.983	0.952	0.983	0.973	0.986	0.824	0.570
Mean weighted F1 across classification models	0.961	0.965	0.937	0.948	0.972	0.790	0.817
T stage	0.951	0.954	0.948	–	–	–	–
N stage	0.954	0.954	0.948	–	–	–	–
M stage	0.972	0.969	0.969	–	–	–	–
Estimate tumor volume	0.605	0.765	0.873	–	–	–	–
Prostate weight	0.846	0.855	0.914	–	–	–	–
Mean accuracy for token extractor models	0.866	0.899	0.930	–	–	–	–

CNN, convolutional neural network; LSTM, long short-term memory neural network; SVM, support vector machine.