Search

Skin cancer has a significant impact on the lives of many individuals annually and is recognized as the most prevalent type of cancer. In the United States, an estimated annual incidence of approximately 3.5 million people receiving a diagnosis of skin cancer underscores its widespread prevalence. Furthermore, the prognosis for individuals afflicted with advancing stages of skin cancer experiences a substantial decline in survival rates. This paper is dedicated to aiding healthcare experts in distinguishing between benign and malignant skin cancer cases by employing a range of machine learning and deep learning techniques and different feature extractors and feature selectors to enhance the evaluation metrics. In this paper, different transfer learning models are employed as feature extractors, and to enhance the evaluation metrics, a feature selection layer is designed, which includes diverse techniques such as Univariate, Mutual Information, ANOVA, PCA, XGB, Lasso, Random Forest, and Variance. Among transfer models, DenseNet-201 was selected as the primary feature extractor to identify features from data. Subsequently, the Lasso method was applied for feature selection, utilizing diverse machine learning approaches such as MLP, XGB, RF, and NB. To optimize accuracy and precision, ensemble methods were employed to identify and enhance the best-performing models. The study provides accuracy and sensitivity rates of 87.72% and 92.15%, respectively. [ABSTRACT FROM AUTHOR]

Due to the recent advances in the Internet and communication technologies, network systems and data have evolved rapidly. The emergence of new attacks jeopardizes network security and make it really challenging to detect intrusions. Multiple network attacks by an intruder are unavoidable. Our research targets the critical issue of class imbalance in intrusion detection, a reflection of the real-world scenario where legitimate network activities significantly out number malicious ones. This imbalance can adversely affect the learning process of predictive models, often resulting in high false-negative rates, a major concern in Intrusion Detection Systems (IDS). By focusing on datasets with this imbalance, we aim to develop and refine advanced algorithms and techniques, such as anomaly detection, cost-sensitive learning, and oversampling methods, to effectively handle such disparities. The primary goal is to create models that are highly sensitive to intrusions while minimizing false alarms, an essential aspect of effective IDS. This approach is not only practical for real-world applications but also enhances the theoretical understanding of managing class imbalance in machine learning. Our research, by addressing these significant challenges, is positioned to make substantial contributions to cybersecurity, providing valuable insights and applicable solutions in the fight against digital threats and ensuring robustness and relevance in IDS development. An intrusion detection system (IDS) checks network traffic for security, availability, and being non-shared. Despite the efforts of many researchers, contemporary IDSs still need to further improve detection accuracy, reduce false alarms, and detect new intrusions. The mean convolutional layer (MCL), feature-weighted attention (FWA) learning, a bidirectional long short-term memory (BILSTM) network, and the random forest algorithm are all parts of our unique hybrid model called MCL-FWA-BILSTM. The CNN-MCL layer for feature extraction receives data after preprocessing. After convolution, pooling, and flattening phases, feature vectors are obtained. The BI-LSTM and self-attention feature weights are used in the suggested method to mitigate the effects of class imbalance. The attention layer and the BI-LSTM features are concatenated to create mapped features before feeding them to the random forest algorithm for classification. Our methodology and model performance were validated using NSL-KDD and UNSW-NB-15, two widely available IDS datasets. The suggested model's accuracies on binary and multi-class classification tasks using the NSL-KDD dataset are 99.67% and 99.88%, respectively. The model's binary and multi-class classification accuracies on the UNSW-NB15 dataset are 99.56% and 99.45%, respectively. Further, we compared the suggested approach with other previous machine learning and deep learning models and found it to outperform them in detection rate, FPR, and F-score. For both binary and multiclass classifications, the proposed method reduces false positives while increasing the number of true positives. The model proficiently identifies diverse network intrusions on computer networks and accomplishes its intended purpose. The suggested model will be helpful in a variety of network security research fields and applications. [ABSTRACT FROM AUTHOR]

Objective: To establish whether or not a natural language processing technique could identify two common inpatient neurosurgical comorbidities using only text reports of inpatient head imaging. Materials and methods: A training and testing dataset of reports of 979 CT or MRI scans of the brain for patients admitted to the neurosurgery service of a single hospital in June 2021 or to the Emergency Department between July 1–8, 2021, was identified. A variety of machine learning and deep learning algorithms utilizing natural language processing were trained on the training set (84% of the total cohort) and tested on the remaining images. A subset comparison cohort (n = 76) was then assessed to compare output of the best algorithm against real-life inpatient documentation. Results: For "brain compression", a random forest classifier outperformed other candidate algorithms with an accuracy of 0.81 and area under the curve of 0.90 in the testing dataset. For "brain edema", a random forest classifier again outperformed other candidate algorithms with an accuracy of 0.92 and AUC of 0.94 in the testing dataset. In the provider comparison dataset, for "brain compression," the random forest algorithm demonstrated better accuracy (0.76 vs 0.70) and sensitivity (0.73 vs 0.43) than provider documentation. For "brain edema," the algorithm again demonstrated better accuracy (0.92 vs 0.84) and AUC (0.45 vs 0.09) than provider documentation. Discussion: A natural language processing-based machine learning algorithm can reliably and reproducibly identify selected common neurosurgical comorbidities from radiology reports. Conclusion: This result may justify the use of machine learning-based decision support to augment provider documentation. [ABSTRACT FROM AUTHOR]

In recent years, artificial intelligence (AI) has shown promising applications in various scientific domains, including biochemical analysis research. However, the effectiveness of AI in modeling small-scale, imbalanced datasets remains an open question in such fields. This study explores the capabilities of eight basic AI algorithms, including ridge regression, logistic regression, random forest regression, and others, in modeling a small, imbalanced clinical dataset (total n = 387, class 0 = 27, class 1 = 360) related to the records of the biochemical blood tests from the patients with multiple wasp stings (MWS). Through rigorous evaluation using k-fold cross-validation and comprehensive scoring, we found that none of the models could effectively model the data. Even after fine-tuning the hyperparameters of the best-performing models, the results remained below acceptable thresholds. The study highlights the challenges of applying AI to small-scale datasets with imbalanced groups in biochemical or clinical research and emphasizes the need for novel algorithms tailored to small-scale data. The findings also call for further exploration into techniques such as transfer learning and data augmentation, and they underline the importance of understanding the minimum dataset scale required for effective AI modeling in biochemical contexts. [ABSTRACT FROM AUTHOR]

Computer-aided diagnosis techniques based on deep learning in skin cancer classification have disadvantages such as unbalanced datasets, redundant information in the extracted features and ignored interactions of partial features among different convolutional layers. In order to overcome these disadvantages, we propose a skin cancer classification model named EFFNet, which is based on feature fusion and random forests. Firstly, the model preprocesses the HAM10000 dataset to make each category of training set images balanced by image enhancement technology. Then, the pre-training weights of the EfficientNetV2 model on the ImageNet dataset are fine-tuned on the HAM10000 skin cancer dataset. After that, an improved hierarchical bilinear pooling is introduced to capture the interactions of some features between the layers and enhance the expressive ability of features. Finally, the fused features are passed into the random forests for classification prediction. The experimental results show that the accuracy, recall, precision and F1-score of the model reach 94.96%, 93.74%, 93.16% and 93.24% respectively. Compared with other models, the accuracy rate is improved to some extent and the highest accuracy rate can be increased by about 10%. [ABSTRACT FROM AUTHOR]

To analyse large numbers of texts, social science researchers are increasingly confronting the challenge of text classification. When manual labeling is not possible and researchers have to find automatized ways to classify texts, computer science provides a useful toolbox of machine-learning methods whose performance remains understudied in the social sciences. In this article, we compare the performance of the most widely used text classifiers by applying them to a typical research scenario in social science research: a relatively small labeled dataset with infrequent occurrence of categories of interest, which is a part of a large unlabeled dataset. As an example case, we look at Twitter communication regarding climate change, a topic of increasing scholarly interest in interdisciplinary social science research. Using a novel dataset including 5,750 tweets from various international organizations regarding the highly ambiguous concept of climate change, we evaluate the performance of methods in automatically classifying tweets based on whether they are about climate change or not. In this context, we highlight two main findings. First, supervised machine-learning methods perform better than state-of-the-art lexicons, in particular as class balance increases. Second, traditional machine-learning methods, such as logistic regression and random forest, perform similarly to sophisticated deep-learning methods, whilst requiring much less training time and computational resources. The results have important implications for the analysis of short texts in social science research. [ABSTRACT FROM AUTHOR]

Background context: Academic meetings serve as an opportunity to present and discuss novel ideas. Previous studies have identified factors predictive of publication without generating predictive models. Machine learning (ML) presents a novel tool capable of generating these models. As such, the objective of this study was to use ML models to predict subsequent publication of abstracts presented at a major surgical conference. Study design/setting: Database study. Methods: All abstracts from the North American Spine Society (NASS) annual general meetings (AGM) from 2013–2015 were reviewed. The following information was extracted: number of authors, institution, location, conference category, subject category, study type, data collection methodology, human subject research, and FDA approval. Abstracts were then searched on the PubMed, Google Scholar, and Scopus databases for publication. ML models were trained to predict whether the abstract would be published or not. Quality of models was determined by using the area under the receiver operator curve (AUC). The top ten most important factors were extracted from the most successful model during testing. Results: A total of 1119 abstracts were presented, with 553 (49%) abstracts published. During training, the model with the highest AUC and accuracy metrics was the partial least squares (AUC of 0.77±0.05, accuracy of 75.5%±4.7%). During testing, the model with the highest AUC and accuracy was the random forest (AUC of 0.69, accuracy of 67%). The top ten features for the random forest model were (descending order): number of authors, year, conference category, subject category, human subjects research, continent, and data collection methodology. Conclusions: This was the first study attempting to use ML to predict the publication of complete articles after abstract presentation at a major academic conference. Future studies should incorporate deep learning frameworks, cognitive/results-based variables and aim to apply this methodology to larger conferences across other fields of medicine to improve the quality of works presented. [ABSTRACT FROM AUTHOR]

Functional and motility-related gastrointestinal (GI) disorders affect nearly 40% percent of the population. Disturbances of GI myoelectric activity have been proposed to play a significant role in these disorders. A significant barrier to usage of these signals in diagnosis and treatment is the lack of consistent relationships between GI myoelectric features and function. A potential cause of this issue is the use of arbitrary classification criteria, such as percentage of power in tachygastric and bradygastric frequency bands. Here we applied automatic feature extraction using a deep neural network architecture on GI myoelectric signals from free-moving ferrets. For each animal, we recorded during baseline control and feeding conditions lasting for 1 h. Data were trained on a 1-dimensional residual convolutional network, followed by a fully connected layer, with a decision based on a sigmoidal output. For this 2-class problem, accuracy was 90%, sensitivity (feeding detection) was 90%, and specificity (baseline detection) was 89%. By comparison, approaches using hand-crafted features (e.g., SVM, random forest, and logistic regression) produced an accuracy from 54% to 82%, sensitivity from 46% to 84% and specificity from 66% to 80%. These results suggest that automatic feature extraction and deep neural networks could be useful to assess GI function for comparing baseline to an active functional GI state, such as feeding. In future testing, the current approach could be applied to determine normal and disease-related GI myoelectric patterns to diagnosis and assess patients with GI disease. [ABSTRACT FROM AUTHOR]

Chromatin interaction analysis by paired-end tag sequencing (ChIA-PET) can capture genome-wide chromatin interactions mediated by a specific DNA-associated protein. The ChIA-PET experiments have been applied to explore the key roles of different protein factors in chromatin folding and transcription regulation. However, compared with widely available Hi-C and ChIP-seq data, there are not many ChIA-PET datasets available in the literature. A computational method for accurately predicting ChIA-PET interactions from Hi-C and ChIP-seq data is needed that can save the efforts of performing wet-lab experiments. Here we present DeepChIA-PET, a supervised deep learning approach that can accurately predict ChIA-PET interactions by learning the latent relationships between ChIA-PET and two widely used data types: Hi-C and ChIP-seq. We trained our deep models with CTCF-mediated ChIA-PET of GM12878 as ground truth, and the deep network contains 40 dilated residual convolutional blocks. We first showed that DeepChIA-PET with only Hi-C as input significantly outperforms Peakachu, another computational method for predicting ChIA-PET from Hi-C but using random forests. We next proved that adding ChIP-seq as one extra input does improve the classification performance of DeepChIA-PET, but Hi-C plays a more prominent role in DeepChIA-PET than ChIP-seq. Our evaluation results indicate that our learned models can accurately predict not only CTCF-mediated ChIA-ET in GM12878 and HeLa but also non-CTCF ChIA-PET interactions, including RNA polymerase II (RNAPII) ChIA-PET of GM12878, RAD21 ChIA-PET of GM12878, and RAD21 ChIA-PET of K562. In total, DeepChIA-PET is an accurate tool for predicting the ChIA-PET interactions mediated by various chromatin-associated proteins from different cell types. Author summary: Various techniques have been widely used to model and investigate three-dimensional (3D) genomes, such as Hi-C, ChIA-PET, and HiChIP. Unlike Hi-C, ChIA-PET can capture genome-wide chromosomal contacts mediated by a predefined DNA-associated protein. Compared with an adequate resource of Hi-C data captured from different cell lines for various species, only a limited number of ChIA-PET data sets are publicly available. Here we developed a deep learning method using Hi-C and ChIP-seq data to predict ChIA-PET directly. Our evaluation results demonstrate that our learned models can successfully predict ChIA-PETs mediated by a different DNA-associated protein for a separate cell line. Our method is an effective alternative to obtaining ChIA-PET contacts from a wet lab. [ABSTRACT FROM AUTHOR]

Detection of Premature Ventricular Contractions (PVC) is of crucial importance in the cardiology field, not only to improve the health system but also to reduce the workload of experts who analyze electrocardiograms (ECG) manually. PVC is a non-harmful common occurrence represented by extra heartbeats, whose diagnosis is not always easily identifiable, especially when done by long-term manual ECG analysis. In some cases, it may lead to disastrous consequences when associated with other pathologies. This work introduces an approach to identify PVCs using machine learning techniques without feature extraction and cross-validation techniques. In particular, a group of six classifiers has been used: Decision Tree, Random Forest, Long-Short Term Memory (LSTM), Bidirectional LSTM, ResNet-18, MobileNetv2, and ShuffleNet. Two types of experiments have been performed on data extracted from the MIT-BIH Arrhythmia database: (i) the original dataset and (ii) the balanced dataset. MobileNetv2 came in first in both experiments with high performance and promising results for PVCs' final diagnosis. The final results showed 99.90% of accuracy in the first experiment and 99.00% in the second one, despite no feature detection techniques were used. The approach we used, which was focused on classification without using feature extraction and cross-validation techniques, allowed us to provide excellent performance and obtain better results. Finally, this research defines as first step toward understanding the explanations for deep learning models' incorrect classifications. [ABSTRACT FROM AUTHOR]

Annual monitoring of the spatial distribution of cultivated land is important for maintaining the ecological environment, achieving a status quo of land resource management, and guaranteeing agricultural production. With the gradual development of remote sensing technology, it has become a common practice to obtain cultivated land boundary information on a large scale with the help of satellite Earth observation images. Traditional land use classification methods are affected by multiple types of land cover, which leads to a decrease in the accuracy of cultivated land mapping. In contrast, although the current advanced methods (such as deep learning) can obtain more accurate cultivated land mapping results than traditional methods, such methods often require the use of a massive amount of training samples, large computing power, and highly complex model tuning processes, increasing the cost of mapping and requiring the involvement of more professionals. This has hindered the promotion of related methods in mapping institutions. This paper proposes a method based on time series vector features (MTVF), which uses vector thinking to establish the features. The advantage of this method is that the introduction of vector features enlarges the differences between the different land cover types, which overcomes the loss of mapping accuracy caused by the influences of the spectra of different ground objects and ensures the calculation efficiency. Moreover, the MTVF uses a traditional method (random forest) as the classification core, which makes the MTVF less demanding than advanced methods in terms of the number of training samples. Sentinel-2 satellite images were used to carry out cultivated land mapping for 2020 in northern Henan Province, China. The results show that the MTVF has the potential to accurately identify cultivated land. Furthermore, the overall accuracy, producer accuracy, and user accuracy of the overall study area and four sub-study areas were all greater than 90%. In addition, the cultivated land mapping accuracy of the MTVF is significantly better than that of the maximum likelihood, support vector machine, and artificial neural network methods. [ABSTRACT FROM AUTHOR]

Scientific research is shedding light on the interaction of the gut microbiome with the human host and on its role in human health. Existing machine learning methods have shown great potential in discriminating healthy from diseased microbiome states. Most of them leverage shotgun metagenomic sequencing to extract gut microbial species-relative abundances or strain-level markers. Each of these gut microbial profiling modalities showed diagnostic potential when tested separately; however, no existing approach combines them in a single predictive framework. Here, we propose the Multimodal Variational Information Bottleneck (MVIB), a novel deep learning model capable of learning a joint representation of multiple heterogeneous data modalities. MVIB achieves competitive classification performance while being faster than existing methods. Additionally, MVIB offers interpretable results. Our model adopts an information theoretic interpretation of deep neural networks and computes a joint stochastic encoding of different input data modalities. We use MVIB to predict whether human hosts are affected by a certain disease by jointly analysing gut microbial species-relative abundances and strain-level markers. MVIB is evaluated on human gut metagenomic samples from 11 publicly available disease cohorts covering 6 different diseases. We achieve high performance (0.80 < ROC AUC < 0.95) on 5 cohorts and at least medium performance on the remaining ones. We adopt a saliency technique to interpret the output of MVIB and identify the most relevant microbial species and strain-level markers to the model's predictions. We also perform cross-study generalisation experiments, where we train and test MVIB on different cohorts of the same disease, and overall we achieve comparable results to the baseline approach, i.e. the Random Forest. Further, we evaluate our model by adding metabolomic data derived from mass spectrometry as a third input modality. Our method is scalable with respect to input data modalities and has an average training time of < 1.4 seconds. The source code and the datasets used in this work are publicly available. Author summary: The gut microbiome can be an indicator of various diseases due to its interaction with the human system. Our main objective is to improve on the current state of the art in microbiome classification for diagnostic purposes. A rich body of literature evidences the clinical value of microbiome predictive models. Here, we propose the Multimodal Variational Information Bottleneck (MVIB), a novel deep learning model for microbiome-based disease prediction. MVIB learns a joint stochastic encoding of different input data modalities to predict the output class. We use MVIB to predict whether human hosts are affected by a certain disease by jointly analysing gut microbial species-relative abundance and strain-level marker profiles. Both of these gut microbial features showed diagnostic potential when tested separately in previous studies; however, no research has combined them in a single predictive tool. We evaluate MVIB on various human gut metagenomic samples from 11 publicly available disease cohorts. MVIB achieves competitive performance compared to state-of-the-art methods. Additionally, we evaluate our model by adding metabolomic data as a third input modality and we show that MVIB is scalable with respect to input feature modalities. Further, we adopt a saliency technique to interpret the output of MVIB and identify the most relevant microbial species and strain-level markers to our model predictions. [ABSTRACT FROM AUTHOR]

Background: Cardiac dysrhythmias (CD) affect millions of Americans in the United States (US), and are associated with considerable morbidity and mortality. New strategies to combat this growing problem are urgently needed. Objectives: Predicting CD using electronic health record (EHR) data would allow for earlier diagnosis and treatment of the condition, thus improving overall cardiovascular outcomes. The Guideline Advantage (TGA) is an American Heart Association ambulatory quality clinical data registry of EHR data representing 70 clinics distributed throughout the US, and has been used to monitor outpatient prevention and disease management outcome measures across populations and for longitudinal research on the impact of preventative care. Methods: For this study, we represented all time-series cardiovascular health (CVH) measures and the corresponding data collection time points for each patient by numerical embedding vectors. We then employed a deep learning technique–long-short term memory (LSTM) model–to predict CD from the vector of time-series CVH measures by 5-fold cross validation and compared the performance of this model to the results of deep neural networks, logistic regression, random forest, and Naïve Bayes models. Results: We demonstrated that the LSTM model outperformed other traditional machine learning models and achieved the best prediction performance as measured by the average area under the receiver operator curve (AUROC): 0.76 for LSTM, 0.71 for deep neural networks, 0.66 for logistic regression, 0.67 for random forest, and 0.59 for Naïve Bayes. The most influential feature from the LSTM model were blood pressure. Conclusions: These findings may be used to prevent CD in the outpatient setting by encouraging appropriate surveillance and management of CVH. [ABSTRACT FROM AUTHOR]

Advances in remote sensing and machine learning enable increasingly accurate, inexpensive, and timely estimation of poverty and malnutrition indicators to guide development and humanitarian agencies' programming. However, state of the art models often rely on proprietary data and/or deep or transfer learning methods whose underlying mechanics may be challenging to interpret. We demonstrate how interpretable random forest models can produce estimates of a set of (potentially correlated) malnutrition and poverty prevalence measures using free, open access, regularly updated, georeferenced data. We demonstrate two use cases: contemporaneous prediction, which might be used for poverty mapping, geographic targeting, or monitoring and evaluation tasks, and a sequential nowcasting task that can inform early warning systems. Applied to data from 11 low and lower-middle income countries, we find predictive accuracy broadly comparable for both tasks to prior studies that use proprietary data and/or deep or transfer learning methods. [ABSTRACT FROM AUTHOR]

Gadolinium-enhancing lesions reflect active disease and are critical for in-patient monitoring in multiple sclerosis (MS). In this work, we have developed the first fully automated method to segment and count the gadolinium-enhancing lesions from routine clinical MRI of MS patients. The proposed method first segments the potential lesions using 2D-UNet from multi-channel scans (T1 post-contrast, T1 pre-contrast, FLAIR, T2, and proton-density) and classifies the lesions using a random forest classifier. The algorithm was trained and validated on 600 MRIs with manual segmentation. We compared the effect of loss functions (Dice, cross entropy, and bootstrapping cross entropy) and number of input contrasts. We compared the lesion counts with those by radiologists using 2,846 images. Dice, lesion-wise sensitivity, and false discovery rate with full 5 contrasts were 0.698, 0.844, and 0.307, which improved to 0.767, 0.969, and 0.00 in large lesions (>100 voxels). The model using bootstrapping loss function provided a statistically significant increase of 7.1% in sensitivity and of 2.3% in Dice compared with the model using cross entropy loss. T1 post/pre-contrast and FLAIR were the most important contrasts. For large lesions, the 2D-UNet model trained using T1 pre-contrast, FLAIR, T2, PD had a lesion-wise sensitivity of 0.688 and false discovery rate 0.083, even without T1 post-contrast. For counting lesions in 2846 routine MRI images, the model with 2D-UNet and random forest, which was trained with bootstrapping cross entropy, achieved accuracy of 87.7% using T1 pre-contrast, T1 post-contrast, and FLAIR when lesion counts were categorized as 0, 1, and 2 or more. The model performs well in routine non-standardized MRI datasets, allows large-scale analysis of clinical datasets, and may have clinical applications. [ABSTRACT FROM AUTHOR]