Joeky Senders

136 Chapter 7 TABLE 3. Model performance according to the calibration slope and intercept. A calibration intercept of 0 with a calibration slope of 1 is considered as perfect calibration. Model Calibration Slope (95%-CI) Intercept (95%-CI) 1D-convolutional neural network 0.90 (0.81 – 1.00) 0.03 (-0.04 – 0.09) LASSO regression 1.06 (0.95 – 1.17) -0.06 (-0.14 – 0.01) LSTM 0.86 (0.78 – 0.95) 0.05 (-0.02 – 0.11) GRU 0.92 (0.83 – 1.02) -0.02 (-0.09 – 0.05) Logistic regression 4.57 (3.94 – 5.20) -2.19 (-2.57– -1.80) Multilayer perceptron 1.14 (0.98 – 1.29) -0.01 (-0.10 – 0.08) 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. Discussion NLP constitutes a subfield of artificial intelligence that focuses on enabling computers to understand and process human languages. 1 Machine learning is another branch of artificial intelligence that focuses on enabling computer algorithms to learn from experience. 10 At their intersection, NLP harnessed with machine learning algorithms can learn how to process language by training on a vast number of labeled examples. 11 Among various NLP methods, the bag-of-words approach combined with a LASSO regression model demonstrated the best overall performance in extracting an equally- distributed, binary outcome from free-text clinical reports. NLP has already been explored for the analysis of radiology reports of brain tumor patients, as well as other cancer types. Cheng et al. used NLP to analyze free-text radiology reports for tumor status classification. 12 Their NLPmodel had 80.6% sensitivity and 91.6% specificity in determining whether tumors had progressed, regressed, or remained stable. NLP for the analysis of radiology reports has also been explored in the context of other cancer types including hepatocellular carcinomas, 13–16 breast cancer, 17–20 lung cancer, 21–23 and other abdominal or pelvic tumors. 11,24–27 All studies that provided sufficient insight into their modeling approach utilized a bag-of-words approach. To our knowledge, the current study presents the first sequence-based NLP approach for the analysis of free-text radiology reports in oncology patients, as well as the first head-to-head comparison of sequence-based and bag-of-words models for medical text analysis.