Joeky Senders

149 Deep Learning and NLP learning curves Results A total of 7000 pathology reports from 5242 patients were retrieved. Among these patients, 2316 (44.2%) were diagnosed with glioma, 1412 (26.9%) with meningioma, and 1514 with cerebral metastasis (28.9%). Baseline characteristics for the training, validation, and test set are provided in Table 2. A statistically significant difference (p=0.038) was found in the mean age across the cohorts. This difference was deemed to be of little clinical significance (57.0 years in the training set versus 56.4 years in the validation set) and likely the result of the large cohort sizes. TABLE 2. Cohort characteristics Patient Characteristics Description Training set (n = 2621) Validation set (n = 1310) Test set (n = 1311) p No. % No. % No. % Patient age <50 728 27.8 423 32.3 394 30.1 0.060 >70 519 19.8 236 18.0 251 19.1 50-70 1374 52.4 651 49.7 666 50.8 Mean ± SD 57.0 ± 14.4 56.4 ± 14.2 56.7 ± 14.7 0.038 Sex Female 1474 56.2 720 55.0 738 56.3 0.716 Male 1147 43.8 590 45.0 573 43.7 Reports per patient Median [IQR] 1 [1 - 1] 1 [1 – 1] 1 [1 – 1] 0.683 Histopathological diagnosis Glioma 1142 43.6 574 43.8 600 45.8 0.697 Meningioma 712 27.2 362 27.6 338 25.8 Metastasis 767 29.3 374 28.5 373 28.5 Abbreviations: IQR=interquartile range; No.=sample size; p=p-value; SD=standard deviation The neural network architecture developed in the current study, ClinicalTextMiner, demonstrated a steeper learning curve than regression-based models (Figure 1, Table 3) and other competing deep learning models (Figure 2). Regression-based algorithms required 200-400 and 800-1500 training examples to reach the AUC performance thresholds of 0.95 and 0.98, respectively. ClinicalTextMiner reached these thresholds with 100 and 200 examples, respectively, corresponding to a learning capacity that is two to eight times more efficient. The best performing CNN architecture reached the AUC performance threshold of 0.95 after training with at least 400 training examples and did not reach the performance threshold of 0.98. Furthermore, its