Comparing deep learning and concept extraction based methods for patient phenotyping from clinical narratives.

Natural language processing with deep learning for medical adverse event detection from free-text medical narratives: A case study of detecting total hip replacement dislocation

To develop and evaluate an internally validated natural language processing (NLP) model to determine guideline adherence of antibiotic decisions in the emergency department (ED) for skin and soft tissue infections (SSTIs). This cross-sectional pilot study developed and applied a random forest (RF)/NLP model to classify clinical narratives of patients with skin infections as requiring either methicillin-resistant Staphylococcus aureus (MRSA) or non-MRSA antibiotic coverage. Conducted at a quaternary care children's hospital with an annual volume of 50,000 ED visits, our study included patients aged 1 to 18 years presenting for SSTIs to the ED from July 1, 2018, through June 30, 2023. Main outcomes included the NLP model's sensitivity, specificity, and receiver operator characteristic (ROC) curve compared to gold standard manual physician review of the medical records. A total of 1588 patients were part of the training data set, with an additional 423 patients utilized for validation. The RF model achieved an area under the curve (AUC) of 0.99, 97.0% sensitivity, and 94.9% specificity. In the validation data set, the model achieved 96.6% sensitivity and 90.1% specificity. Performance remained strong despite absent/missing patient history in certain narratives. The NLP model demonstrated that automated analysis of clinical narratives for determining guideline adherence is feasible. Despite missing data in narratives, the model's high performance suggests potential for broader application. Given the rise in antibiotic-resistant infections and the role of judicious antibiotic use, developing automated systems through NLP can significantly contribute to health care delivery and patient safety. The methodology of this study provides a feasible, sustainable path for similar applications in the emergency setting.

447 Background: Mortality from bladder cancer (BC) increases exponentially once it invades the muscle. At the population level, accurate delineation of these patients is challenging. Methods: To develop and validate a natural language processing (NLP) model for automatically identifying muscle-invasive BC (MIBC) patients, aiding in population-based BC research. All patients with a CPT code for transurethral resection of bladder tumor (TURBT) (N = 76,060) were selected from the Department of Veterans Affairs (VA) Corporate Data Warehouse database. A sample of 600 patients (with 2,337 full-text notes) who had TURBT and confirmed pathology results were selected for NLP model development (500 patients) and validation (100 patients). Muscle-invasion (yes/no), unknown, or no cancer, were confirmed by detailed chart review of pathology notes. The NLP performance was assessed by calculating the sensitivity, positive predictive value (PPV), and overall accuracy at the individual note and patient levels. Results: In the validation cohort, the NLP model had overall accuracy of 88% and 92% at the note and patient levels. Specifically, PPV and specificity for predicting muscle-invasion on note level were 83% and 70%, respectively. The model classified non-muscle invasive BC (NMIBC) with 98% sensitivity at both the note and patient levels. Although the sensitivity for MIBC was 70% for note-level determination, the sensitivity was 86% when evaluated at the patient level. When applying the model to 71,200 patients VA-wide, the model classified 13,642 (19%) as having MIBC and 47,595 (66%) as NMIBC. The NLP model was able to identify invasion status for 96% TURBT patients at the population level. Inherent limitations include relatively small training set given the size of the VA population. Conclusions: This NLP model for identifying muscle-invasion at the population level had high accuracy. The NLP model may be a practical and accurate tool for efficiently identifying BC invasion status and may potentially aid in population-based BC research in the VA.

The highly structured nature of medical reports makes them feasible for automated large-scale patient identification. This study aimed to develop a natural language processing (NLP) model to retrospectively retrieve patients with presence and history of carotid stenosis (CS) using their ultrasound reports. Ultrasound reports from our institution between January 2016 and December 2017 were selected. To process the texts, we developed a parser to divide the raw text into fields. For baseline method, we used bag-of-n-grams and term frequency inverse document frequency as the features and used linear classifiers. Logistic regression was performed as the baseline model. Convolution and recurrent neural networks (CNN; RNN) with attention mechanism were applied to the dataset to improve the classification accuracy. We had 1220 ultrasound reports for training and 307 for testing, totaling to 1527 reports. For predicting history of CS, both CNN and RNN-attention models had a significantly higher specificity than logistic regression. In addition, RNN-attention also had a significantly higher F1 score and accuracy. For predicting presence of carotid stenosis, all models achieved above 93% accuracy. RNN-attention achieved a 95.4% accuracy, although the difference with logistic regression was not statistically significant. RNN-attention had a statistically significant higher specificity than logistic regression. We developed linear, CNN, and RNN models to predict history and presence of CS from ultrasound reports. We have demonstrated NLP to be an efficient, accurate approach for large-scale retrospective patient identification, with applications in long-term follow-up of patients and clinical research studies. • Natural language processing models using both linear classifiers and neural networks can achieve a good performance, with an overall accuracy above 90% in predicting history and presence of carotid stenosis. • Convolution and recurrent neural networks, especially with additional features including field awareness and attention mechanism, have superior performance than traditional linear classifiers. • NLP is shown to be an efficient approach for large-scale retrospective patient identification, with applications in long-term follow-up of patients and further clinical research studies.

Deep learning (also known as deep structured learning, hierarchical learning, or deep machine learning) is a branch of machine learning based on a set of algorithms that attempt to model high-level abstractions in data by using a deep graph with multiple processing layers, composed of multiple linear and nonlinear transformations. Deep learning has been characterized as a class of machine learning algorithms with the following characteristics [257]: • They use a cascade of many layers of nonlinear processing units for feature extraction and transformation. Each successive layer uses the output from the previous layer as input. The algorithms may be supervised or unsupervised and applications include pattern analysis (unsupervised) and classification (supervised). • They are based on the (unsupervised) learning of multiple levels of features or representations of the data. Higher-level features are derived from lower-level features to form a hierarchical representation. • They are part of the broader machine learning field of learning representations of data. • They learn multiple levels of representations that correspond to different levels of abstraction; the levels form a hierarchy of concepts. These definitions have in common: multiple layers of nonlinear processing units and the supervised or unsupervised learning of feature representations in each layer, with the layers forming a hierarchy from low-level to high-level features. Various deep learning architectures such as deep neural networks, convolutional deep neural networks, deep belief networks (DBN), and recurrent neural networks have been applied to fields like computer vision, automatic speech recognition, natural language processing, audio recognition, and bioinformatics where they have been shown to produce state-of-the-art results on various tasks. In this chapter we start in Section 7.1 with an introduction, giving a brief history of this field, the relevant literature, and its applications. Then we study some basic concepts of deep learning such as convolutional neural networks, recurrent neural networks, backpropagation algorithm, restricted Boltzmann machines, and deep learning networks in Section 7.2. Then we illustrate three examples for Apache Spark implementation for mobile big data (MBD), user moving pattern extraction, and combination with nonparametric Bayesian learning, respectively, in Sections 7.3 through 7.5. Finally, we have summary in Section 7.6.

Understanding treatment patterns and effectiveness for patients with metastatic prostate cancer (mPCa) is dependent on accurate assessment of metastatic status. The objective was to develop a natural language processing (NLP) model for identifying patients with mPCa and evaluate the model's performance against chart-reviewed data and an International Classification of Diseases (ICD) 9/10 code-based method. In total, 139,057 radiology reports on 6,211 unique patients from the Department of Veterans Affairs were used. The gold standard was metastases by detailed chart review of radiology reports. NLP performance was assessed by sensitivity, specificity, positive predictive value, negative predictive value, and date of metastases detection. Receiver operating characteristic curves was used to assess model performance. When compared with chart review, the NLP model had high sensitivity and specificity (85% and 96%, respectively). The NLP model was able to predict patient-level metastasis status with a sensitivity of 91% and specificity of 81%, whereas sensitivity and specificity using ICD9/10 billing codes were 73% and 86%, respectively. For the NLP model, date of metastases detection was exactly concordant and within < 1 week in 55% and 58% of patients, compared with 8% and 17%, respectively, using the ICD9/10 billing codes method. The area under the curve for the NLP model was 0.911. A limitation is the NLP model was developed on the basis of a subset of patients with mPCa and may not be generalizable to all patients with mPCa. This population-level NLP model for identifying patients with mPCa was more accurate than using ICD9/10 billing codes when compared with chart-reviewed data. Upon further validation, this model may allow for efficient population-level identification of patients with mPCa.

Background Diabetic Retinopathy (DR) remains a leading cause of preventable blindness, with rising prevalence particularly in low-resource settings. Traditional DR screening relies on specialist ophthalmologic assessment, posing barriers to early detection. Recent advances in deep learning have enabled the development of automated tools that may enhance the scalability and accessibility of DR screening. Objectives To compare the performance of two convolutional neural network (CNN) architectures-EfficientNet-B5 and ResNet-50-for automated DR detection using publicly available retinal fundus image datasets. Methods We utilized a transfer learning approach on two large datasets: EyePACS 2015 (n = 35,126) for training/validation and APTOS 2019 (n = 3,662) for external testing. Images were preprocessed with resizing, cropping, and contrast enhancement. The EfficientNet-B5 and ResNet-50 models were independently trained and evaluated using key metrics: accuracy, sensitivity, specificity, area under the ROC curve (AUC), and Quadratic Weighted Kappa (QWK), with classification across DR severity levels. Results EfficientNet-B5 achieved a QWK of 92%, sensitivity of 98%, specificity of 93%, and AUC of 96% on the test dataset. ResNet-50 demonstrated similarly high performance with an accuracy of 96%, QWK of 86%, sensitivity of 94%, specificity of 97%, and AUC of 96%. Both models showed robust generalization across severity grades, indicating strong discriminative capabilities. Conclusions The results highlighted that our models are viable architectures for AI-driven DR screening, with EfficientNet-B5 showing slightly higher agreement scores and ResNet-50 offering stronger specificity. These models hold promise for deployment in DR screening programs, especially in resource-limited settings where ophthalmologic expertise is scarce. Future work should focus on clinical validation and integration into public health systems to improve early DR detection and reduce vision-related morbidity. Key messages • Integrating artificial intelligence (AI) into diabetic retinopathy screening can enhance early detection and support equitable access to eye care, particularly in resource-limited settings. • AI-based screening tools offer a scalable and effective approach to strengthen public health strategies aimed at preventing vision loss from diabetic complications.

In the field of data security, biometric security is a significant emerging concern. The multimodal biometrics system with enhanced accuracy and detection rate for smart environments is still a significant challenge. The fusion of an electrocardiogram (ECG) signal with a fingerprint is an effective multimodal recognition system. In this work, unimodal and multimodal biometric systems using Convolutional Neural Network (CNN) are conducted and compared with traditional methods using different levels of fusion of fingerprint and ECG signal. This study is concerned with the evaluation of the effectiveness of proposed parallel and sequential multimodal biometric systems with various feature extraction and classification methods. Additionally, the performance of unimodal biometrics of ECG and fingerprint utilizing deep learning and traditional classification technique is examined. The suggested biometric systems were evaluated utilizing ECG (MIT-BIH) and fingerprint (FVC2004) databases. Additional tests are conducted to examine the suggested models with:1) virtual dataset without augmentation (ODB) and 2) virtual dataset with augmentation (VDB). The findings show that the optimum performance of the parallel multimodal achieved 0.96 Area Under the ROC Curve (AUC) and sequential multimodal achieved 0.99 AUC, in comparison to unimodal biometrics which achieved 0.87 and 0.99 AUCs, for the fingerprint and ECG biometrics, respectively. The overall performance of the proposed multimodal biometrics outperformed unimodal biometrics using CNN. Moreover, the performance of the suggested CNN model for ECG signal and sequential multimodal system based on neural network outperformed other systems. Lastly, the performance of the proposed systems is compared with previously existing works.

Medical educators pay increasing attention to the potential utility of narrative data for assessment, but lack of efficient and standardized ways of interpreting the data have limited its use. Natural language processing (NLP) algorithms could provide new insights for using narrative data within assessment especially identifying at-risk or low-performing students. Assessment data were reviewed from 16 cohorts of medical students from the University of Cincinnati College of Medicine (graduating classes of 2006-2022). A T-score Average (TSA) was calculated for each student based on clerkship assessment data. Narrative data from core clerkship evaluations in responses to the prompt "any opportunities for improvement" were utilized for analysis. The narrative data and calculated TSA were then used to train and test 4 NLP models with the goal of utilizing NLP to identify at-risk students as defined by the bottom 10% average TSA score. Based on typical NLP methods, the developed NLP models all performed adequately in identifying student performance with an overall accuracy of 0.8 across all 4 models. However, none of the NLP models was able to identify students within the bottom 10% of performance. During this process, we uncovered the presence of "copy/paste" behavior, a previously undocumented phenomenon within narrative data where preceptors duplicated comments from 1 student to the next. Training NLP models including "copy/paste" comments improved NLP ability to identify students within the bottom and top 10% of performance. NLP models were unable to accurately identify at-risk students. Model accuracy increased with the inclusion of "copy/paste" comments, indicating there may be some discriminatory functionality within aspects of narrative data beyond keywords such as lexical diversity and word quantity. Future work will explore how these other aspects of narrative data associate with performance and utilize more advanced large language models of analysis.Based on typical natural language processing (NLP) methods, the tested NLP models performed adequately in identifying student performance with an overall accuracy of 0.8 across all models. However, none of the NLP models was able to identify students within the bottom 10% of performance, revealing widespread "copy/paste" behavior in evaluations that may limit narrative data's discriminatory utility.

To compare the diagnostic efficiencies of deep learning single-modal and multi-modal for the classification of benign and malignant breast mass lesions. We retrospectively collected data from 203 patients (207 lesions, 101 benign and 106 malignant) with breast tumors who underwent breast magnetic resonance imaging (MRI) before surgery or biopsy between January 2014 and October 2020. Mass segmentation was performed based on the three dimensions-region of interest (3D-ROI) minimum bounding cube at the edge of the lesion. We established single-modal models based on a convolutional neural network (CNN) including T2WI and non-fs T1WI, the dynamic contrast-enhanced (DCE-MRI) first phase was pre-contrast T1WI (d1), and Phases 2, 4, and 6 were post-contrast T1WI (d2, d4, d6); and Multi-modal fusion models with a Sobel operator (four_mods:T2WI, non-fs-T1WI, d1, d2). Training set (n=145), validation set (n=22), and test set (n=40). Five-fold cross validation was performed. Accuracy, sensitivity, specificity, negative predictive value, positive predictive value, and area under the ROC curve (AUC) were used as evaluation indicators. Delong's test compared the diagnostic performance of the multi-modal and single-modal models. All models showed good performance, and the AUC values were all greater than 0.750. Among the single-modal models, T2WI, non-fs-T1WI, d1, and d2 had specificities of 77.1%, 77.2%, 80.2%, and 78.2%, respectively. d2 had the highest accuracy of 78.5% and showed the best diagnostic performance with an AUC of 0.827. The multi-modal model with the Sobel operator performed better than single-modal models, with an AUC of 0.887, sensitivity of 79.8%, specificity of 86.1%, and positive prediction value of 85.6%. Delong's test showed that the diagnostic performance of the multi-modal fusion models was higher than that of the six single-modal models (T2WI, non-fs-T1WI, d1, d2, d4, d6); the difference was statistically significant (p = 0.043, 0.017, 0.006, 0.017, 0.020, 0.004, all were greater than 0.05). Multi-modal fusion deep learning models with a Sobel operator had excellent diagnostic value in the classification of breast masses, and further increase the efficiency of diagnosis.

Following recent advancements in deep learning and artificial intelligence, spoken language identification applications are playing an increasingly significant role in our day-to-day lives, especially in the domain of multi-lingual speech recognition. In this article, we propose a spoken language identification system that depends on the sequence of feature vectors. The proposed system uses a hybrid Convolutional Recurrent Neural Network (CRNN), which combines a Convolutional Neural Network (CNN) with a Recurrent Neural Network (RNN) network, for spoken language identification on seven languages, including Arabic, chosen from subsets of the Mozilla Common Voice (MCV) corpus. The proposed system exploits the advantages of both CNN and RNN architectures to construct the CRNN architecture. At the feature extraction stage, it compares the Gammatone Cepstral Coefficient (GTCC) feature and Mel Frequency Cepstral Coefficient (MFCC) feature, as well as a combination of both. Finally, the speech signals were represented as frames and used as the input for the CRNN architecture. After conducting experiments, the results of the proposed system indicate higher performance with combined GTCC and MFCC features compared to GTCC or MFCC features used individually. The average accuracy of the proposed system was 92.81% in the best experiment for spoken language identification. Furthermore, the system can learn language-specific patterns in various filter size representations of speech files.

Recent advancements in deep learning techniques tailored for autonomous driving in Indian road conditions are crucial for revolutionizing transportation systems. Indian roads present unique challenges, including unpredictable traffic patterns, diverse road infrastructures, and challenging weather conditions. Deep learning is pivotal in addressing these challenges, focusing on perception, decision-making, and control. Various architectures, including Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and Deep Reinforcement Learning (DRL), are analyzed for their efficacy in autonomous driving. Sensor fusion techniques enhance perception capabilities, such as lidar, radar, and camera data integration. Additionally, advancements in semantic segmentation and object detection algorithms improve scene understanding and obstacle recognition. Collaborative efforts between academia, industry, and government agencies are essential for accelerating the deployment of autonomous driving technology in Indian road conditions. In conclusion, leveraging cutting-edge deep learning methodologies enables safe and efficient navigation of autonomous vehicles, ushering in a transformative future in transportation systems.

IntroductionThe electrocardiogram (ECG) is a valuable tool for the diagnosis of myocardial ischemia as it presents distinctive ischemic patterns. Deep learning methods such as convolutional neural networks (CNN) are employed to extract data-derived features and to recognize natural patterns. Hence, CNN enable an unbiased view on well-known clinical phenomenon, e.g., myocardial ischemia. This study tested a novel, hypothesis-generating approach using pre-trained CNN to determine the optimal ischemic parameter as obtained from the highly susceptible intracoronary ECG (icECG).MethodThis was a retrospective observational study in 228 patients with chronic coronary syndrome. Each patient had participated in clinical trials with icECG recording and ST-segment shift measurement at the beginning (i.e., non-ischemic) and the end (i.e., ischemic) of a one-minute proximal coronary artery balloon occlusion establishing the reference. Using these data (893 icECGs in total), two pre-trained, open-access CNN (GoogLeNet/ResNet101) were trained to recognize ischemia. The best performing CNN during training were compared with the icECG ST-segment shift for diagnostic accuracy in the detection of artificially induced myocardial ischemia.ResultsUsing coronary patency or occlusion as reference for absent or present myocardial ischemia, receiver-operating-characteristics (ROC)-analysis of manually obtained icECG ST-segment shift (mV) showed an area under the ROC-curve (AUC) of 0.903±0.043 (p<0.0001, sensitivity 80%, specificity 92% at a cut-off of 0.279mV). The best performing CNN showed an AUC of 0.924 (sensitivity 93%, specificity 92%). DeLong-Test of the ROC-curves showed no significant difference between the AUCs. The underlying morphology responsible for the network prediction differed between the trained networks but was focused on the ST-segment and the T-wave for myocardial ischemia detection.ConclusionsWhen tested in an experimental setting with artificially induced coronary artery occlusion, quantitative icECG ST-segment shift and CNN using pathophysiologic prediction criteria detect myocardial ischemia with similarly high accuracy.

Like any other science, research in natural language processing (NLP) depends on the ability to draw correct conclusions from experiments. A key tool for this is statistical significance testing: We use it to judge whether a result provides meaningful, generalizable findings or should be taken with a pinch of salt. When comparing new methods against others, performance metrics often differ by only small amounts, so researchers turn to significance tests to show that improved models are genuinely better. Unfortunately, this reasoning often fails because we choose inappropriate significance tests or carry them out incorrectly, making their outcomes meaningless. Or, the test we use may fail to indicate a significant result when a more appropriate test would find one. NLP researchers must avoid these pitfalls to ensure that their evaluations are sound and ultimately avoid wasting time and money through incorrect conclusions.This book guides NLP researchers through the whole process of significance testing, making it easy to select the right kind of test by matching canonical NLP tasks to specific significance testing procedures. As well as being a handbook for researchers, the book provides theoretical background on significance testing, includes new methods that solve problems with significance tests in the world of deep learning and multidataset benchmarks, and describes the open research problems of significance testing for NLP.The book focuses on the task of comparing one algorithm with another. At the core of this is the p-value, the probability that a difference at least as extreme as the one we observed could occur by chance. If the p-value falls below a predetermined threshold, the result is declared significant. Leaving aside the fundamental limitation of turning the validity of results into a binary question with an arbitrary threshold, to be a valid statistical significance test, the p-value must be computed in the right way. The book describes the two crucial properties of an appropriate significance test: The test must be both valid and powerful. Validity refers to the avoidance of type 1 errors, in which the result is incorrectly declared significant. Common mistakes that lead to type 1 errors include deploying tests that make incorrect assumptions, such as independence between data points. The power of a test refers to its ability to detect a significant result and therefore to avoid type 2 errors. Here, knowledge of the data and experiment must be used to choose a test that makes the correct assumptions. There is a trade-off between validity and power, but for the most common NLP tasks (language modeling, sequence labeling, translation, etc.), there are clear choices of tests that provide a good balance.Beginning with a detailed background on significance testing, the book then shows the reader how to carry out tests for specific NLP tasks. There is a mix of styles, with the first four chapters providing reference material that will be extremely useful to both new and experienced researchers. Here, it is easy to find the material related to a given NLP task. The next two chapters discuss more recent research into the application of significance tests to deep neural networks and for testing across multiple datasets. Alongside open research questions, these later chapters provide clear guidelines on how to apply the proposed methods. It is this mix of background material and reference guidelines that I believe makes this book so compelling and nicely self-contained.The introduction in Chapter 1 motivates the need for a comprehensive textbook and outlines challenges that the later chapters address more deeply. The theoretical background material begins in Chapter 2, which introduces core concepts, including hypothesis testing, type 1 and type 2 errors, validity and power, and p-values. The reader does not need to have any prior knowledge of statistical significance tests to follow this part. However, experienced readers could still benefit from reading this chapter, as concepts such as p-values are widely misunderstood and misused (Amrhein, Greenland, and McShane 2019).The significance tests themselves are introduced in Chapter 3, categorized into parametric and nonparametric tests. The chapter begins with the intuitively simple paired z-test, then builds up to more commonly-applied techniques, showing the connections and assumptions that each test makes. Step-by-step algorithms help the reader to implement each test. Although the chapter does cite uses of tests in NLP research, the main purpose is to present the theory behind each test and point out their differences.Chapter 4 provides perhaps the most handy part of the book for reference: a correspondence between common NLP tasks and statistical tests. Each task is discussed in terms of the evaluation metrics used, then a decision tree is introduced to guide the reader toward a choice between a parametric test, bootstrap or randomization test, or sampling-free nonparametric test. Section 4.3 then links each NLP evaluation measure to a specific significance test, presenting a large table that helps readers identify which test they need for a specific task. Particular considerations for each task are also pointed out to provide more detail about the appropriate options. The final part of this chapter describes the issue of p-hacking, in which dataset sizes are increased until a significance threshold is reached without consideration for biases in the data (discussed, for example, in Hofmann [2015]). The chapter proposes a simple solution to ensure robust significance testing with large datasets.Where Chapter 4 presents well-established methods, Chapter 5 introduces the current research question of how best to apply statistical significance testing to deep learning. Non-convex loss functions, stochastic optimization, random initialization, and a multitude of hyperparameters limit the conclusions we can draw from a single test run of a deep neural network (DNN). This chapter, which is based on the authors’ ACL paper (Dror, Shlomov, and Reichart 2019), explains how the comparison process can be overhauled to provide more meaningful evaluations. Beginning by explaining the difficulties of evaluating DNNs, the chapter then introduces criteria for a comparison framework, then discusses the limitations of current methods. Reimers and Gurevych (2018) have previously tackled this problem, but their approach has limited power and does not provide a confidence score. Other works, such as Clark et al. (2011), compare DNNs using a collection of statistics, such as the mean or standard deviation of performance across runs. This book shows how such an approach violates the assumptions of the significance tests. The authors propose almost stochastic dominance as the basis for a better alternative. The chapter explains how to use the proposed method, evaluates it in an empirical case study, and finally analyzes the errors made by each testing approach.Large NLP models are often tested across a range of datasets, which presents another problem for standard significance testing. Chapter 6 discusses the challenges of assessing two questions: (1) On how many datasets does algorithm A outperform algorithm B? (2) On which datasets does A outperform B? Applying standard significance tests individually to each dataset and counting the number of significant results is likely to overestimate the total number of significant results, as this chapter explains. The authors then present a new framework for replicability analysis, based on partial conjunction testing, and discuss two variants (Bonferroni and Fisher) for when the datasets are independent or dependent. They introduce a method based on Benjamini and Heller (2008) to count the number of datasets where one method outperforms another, then show how to use the Holm procedure (Holm 1979) to identify which datasets these are. Chapter 6 provides a lot of detailed background on the proposed replicability analysis framework, and the later sections again link the process to specific NLP case studies, and step-by-step summaries help the reader to apply the methodology. Extensive empirical results illustrate the very substantial differences in outcomes between the proposed approach and standard procedures.The final two chapters present open problems and conclude, showing that the topic has many interesting research questions of its own, such as problems when performing cross-validation, and the limited statistical power of replicability analysis.Overall, I highly recommend this book to a wide range of NLP researchers, from new students to seasoned experts who wish to ensure that they compare methods effectively. The book is excellent as both an introduction to the topic of significance testing and as a reference to use when evaluating your results. For anyone with further interest in the topic, it also points the way to future work. If one could level any criticism at this book at all, it is that it does not deeply discuss the basic flaws of significance testing or what the alternatives might be. For now, though, significance testing is an integral part of NLP research and this book provides a great resource for researchers who wish to perform it correctly and painlessly.

In this paper, we present a scene classification method based on vision transformers. These types of networks, which are now the standard models in natural language processing (NLP) do not rely on convolution block as in convolutional neural networks (CNNs). Alternatively, they are based on a mechanism known as multi-head self-attention (MSA), which captures the contextual relations between image pixels regardless of their spatial distance. At the first step, the images under analysis are split into patches, then converted to sequence by flattening and embedding. The embedding position is encoded and added to the sequence to preserve the order of the patches. Then, the resulting sequence is fed to several MSA layers for generating the final representation. To increase the classification performance, we employed several data augmentation strategies to expand the size and the diversity of the training data. Additionally, we show experimentally that we can compress the network by pruning half of its layers while keeping the competing performance. We further investigate the performance of the data-efficient image transformers (DeiT), a version of the model that is trained by knowledge distillation with less amount of data. Experimental results on two remote sensing datasets show that vision transformers can outperform state-of-the-art methods based on CNNs.