Joeky Senders

135 Comparing NLP methods TABLE 2. Model performance according to the area under the receiver operating curve and accuracy, compared to logistic regression as benchmark. Model AUC (95%-CI) p* Accuracy (95%-CI) p* 1D-convolutional neural network 0.93 (0.90–0.95) 0.02 85 (81–88) <0.001 LASSO regression 0.92 (0.89–0.94) 0.02 83 (80–87) <0.001 LSTM 0.91 (0.88–0.94) 0.12 87 (84–90) <0.001 GRU 0.91 (0.88–0.93) 0.18 86 (82–89) <0.001 Logistic regression 0.88 (0.85–0.92) - 64 (60–68) - Multilayer perceptron 0.87 (0.84–0.90) 0.36 80 (76–83) <0.001 Abbreviations: 1D=one dimensional; AUC=area under the receiver operating curve; CI=confidence interval; LASSO=least absolute shrinkage and selection operator; LSTM=long-short term memory; GRU=gated recurrent unit. *Corrected for multiple testing using the Benjamini-Hochberg procedure. FIGURE 2. Calibration plot for all natural language processing models. Abbreviations: Conv1D: one-dimensional convolutional neural network; GRU: gated recurrent unit; LSTM=long-short term memory; MLP=multilayer perceptron.