Macro-averaged Precision Research Articles

Quantitatively assessing the impact of the quality of SNOMED CT subtype hierarchy on cohort queries

Abstract Objective SNOMED CT provides a standardized terminology for clinical concepts, allowing cohort queries over heterogeneous clinical data including Electronic Health Records (EHRs). While it is intuitive that missing and inaccurate subtype (or is-a) relations in SNOMED CT reduce the recall and precision of cohort queries, the extent of these impacts has not been formally assessed. This study fills this gap by developing quantitative metrics to measure these impacts and performing statistical analysis on their significance. Material and Methods We used the Optum de-identified COVID-19 Electronic Health Record dataset. We defined micro-averaged and macro-averaged recall and precision metrics to assess the impact of missing and inaccurate is-a relations on cohort queries. Both practical and simulated analyses were performed. Practical analyses involved 407 missing and 48 inaccurate is-a relations confirmed by domain experts, with statistical testing using Wilcoxon signed-rank tests. Simulated analyses used two random sets of 400 is-a relations to simulate missing and inaccurate is-a relations. Results Wilcoxon signed-rank tests from both practical and simulated analyses (P-values < .001) showed that missing is-a relations significantly reduced the micro- and macro-averaged recall, and inaccurate is-a relations significantly reduced the micro- and macro-averaged precision. Discussion The introduced impact metrics can assist SNOMED CT maintainers in prioritizing critical hierarchical defects for quality enhancement. These metrics are generally applicable for assessing the quality impact of a terminology’s subtype hierarchy on its cohort query applications. Conclusion Our results indicate a significant impact of missing and inaccurate is-a relations in SNOMED CT on the recall and precision of cohort queries. Our work highlights the importance of high-quality terminology hierarchy for cohort queries over EHR data and provides valuable insights for prioritizing quality improvements of SNOMED CT's hierarchy.

A hybrid transformer and attention based recurrent neural network for robust and interpretable sentiment analysis of tweets

Sentiment analysis is a pivotal tool in understanding public opinion, consumer behavior, and social trends, underpinning applications ranging from market research to political analysis. However, existing sentiment analysis models frequently encounter challenges related to linguistic diversity, model generalizability, explainability, and limited availability of labeled datasets. To address these shortcomings, we propose the Transformer and Attention-based Bidirectional LSTM for Sentiment Analysis (TRABSA) model, a novel hybrid sentiment analysis framework that integrates transformer-based architecture, attention mechanism, and recurrent neural networks like BiLSTM. The TRABSA model leverages the powerful RoBERTa-based transformer model for initial feature extraction, capturing complex linguistic nuances from a vast corpus of tweets. This is followed by an attention mechanism that highlights the most informative parts of the text, enhancing the model’s focus on critical sentiment-bearing elements. Finally, the BiLSTM networks process these refined features, capturing temporal dependencies and improving the overall sentiment classification into positive, neutral, and negative classes. Leveraging the latest RoBERTa-based transformer model trained on a vast corpus of 124M tweets, our research bridges existing gaps in sentiment analysis benchmarks, ensuring state-of-the-art accuracy and relevance. Furthermore, we contribute to data diversity by augmenting existing datasets with 411,885 tweets from 32 English-speaking countries and 7,500 tweets from various US states. This study also compares six word-embedding techniques, identifying the most robust preprocessing and embedding methodologies crucial for accurate sentiment analysis and model performance. We meticulously label tweets into positive, neutral, and negative classes using three distinct lexicon-based approaches and select the best one, ensuring optimal sentiment analysis outcomes and model efficacy. Here, we demonstrate that the TRABSA model outperforms the current seven traditional machine learning models, four stacking models, and four hybrid deep learning models, yielding notable gain in accuracy (94%) and effectiveness with a macro average precision of 94%, recall of 93%, and F1-score of 94%. Our further evaluation involves two extended and four external datasets, demonstrating the model’s consistent superiority, robustness, and generalizability across diverse contexts and datasets. Finally, by conducting a thorough study with SHAP and LIME explainable visualization approaches, we offer insights into the interpretability of the TRABSA model, improving comprehension and confidence in the model’s predictions. Our study results make it easier to analyze how citizens respond to resources and events during pandemics since they are integrated into a decision-support system. Applications of this system provide essential assistance for efficient pandemic management, such as resource planning, crowd control, policy formation, vaccination tactics, and quick reaction programs.

Detecting cardiac states with wearable photoplethysmograms and implications for out-of-hospital cardiac arrest detection.

Out-of-hospital cardiac arrest (OHCA) is a global health problem affecting approximately 4.4million individuals yearly. OHCA has a poor survival rate, specifically when unwitnessed (accounting for up to 75% of cases). Rapid recognition can significantly improve OHCA survival, and consumer wearables with continuous cardiopulmonary monitoring capabilities hold potential to "witness" cardiac arrest and activate emergency services. In this study, we used an arterial occlusion model to simulate cardiac arrest and investigated the ability of infrared photoplethysmogram (PPG) sensors, often utilized in consumer wearable devices, to differentiate normal cardiac pulsation, pulseless cardiac (i.e., resembling a cardiac arrest), and non-physiologic (i.e., off-body) states. Across the classification models trained and evaluated on three anatomical locations, higher classification performances were observed on the finger (macro average F1-score of 0.964 on the fingertip and 0.954 on the finger base) compared to the wrist (macro average F1-score of 0.837). The wrist-based classification model, which was trained and evaluated using all PPG measurements, including both high- and low-quality recordings, achieved a macro average precision and recall of 0.922 and 0.800, respectively. This wrist-based model, which represents the most common form factor in consumer wearables, could only capture about 43.8% of pulseless events. However, models trained and tested exclusively on high-quality recordings achieved higher classification outcomes (macro average F1-score of 0.975 on the fingertip, 0.973 on the finger base, and 0.934 on the wrist). The fingertip model had the highest performance to differentiate arterial occlusion pulselessness from normal cardiac pulsation and off-body measurements with macro average precision and recall of 0.978 and 0.972, respectively. This model was able to identify 93.7% of pulseless states (i.e., resembling a cardiac arrest event), with a 0.4% false positive rate. All classification models relied on a combination of time-, power spectral density (PSD)-, and frequency-domain features to differentiate normal cardiac pulsation, pulseless cardiac, and off-body PPG recordings. However, our best model represented an idealized detection condition, relying on ensuring high-quality PPG data for training and evaluation of machine learning algorithms. While 90.7% of our PPG recordings from the fingertip were considered of high quality, only 53.2% of the measurements from the wrist passed the quality criteria. Our findings have implications for adapting consumer wearables to provide OHCA detection, involving advancements in hardware and software to ensure high-quality measurements in real-world settings, as well as development of wearables with form factors that enable high-quality PPG data acquisition more consistently. Given these improvements, we demonstrate that OHCA detection can feasibly be made available to anyone using PPG-based consumer wearables.

Monitoring water quality parameters of freshwater aquaculture ponds using UAV-based multispectral images

Monitoring water quality is crucial for water exchange, precise feeding, and quality control of water products in freshwater aquaculture. In light of the issue of spatial heterogeneity in freshwater aquaculture pond waters and the constraints of conventional sensor detection techniques and traditional machine learning models. In this study, UAV multispectral images were combined with four machine learning algorithms (Ridge, XGBoost, CatBoost, RF) and the Stacking model to model the estimation of Chlorophyll a (Chl-a) and Turbidity and map their spatial distribution. The findings indicate that, in contrast to machine learning models, the Stacking model of water quality parameter performs better with higher accuracy. Meanwhile,for Chl-a and Turbidity the optimal sub-model combination in the Stacking model varies, with the most effective estimation model for Chl-a concentration identified as RF-XGB-Ridge (R2 = 0.84, RMSE=1.882 µg/L, MAE=3.433 µg/L and Slope = 0.791). As to Turbidity, the RF-CAB-Ridge model demonstrates superior performance, with macro-averaged precision (macro-p) of 93.3 %, macro-averaged recall (macro-R) of 88.8 %, macro-averaged F1-score (macro-F1) of 0.895, and Kappa coefficient of 0.813. Furthermore, the results of the joint analyses, which included measured samples and management measures at the test site, demonstrated that the spatial distribution maps of Chl-a and Turbidity were in alignment with the current status of water quality at the test site. This consistency was observed across both temporal and spatial scales. The results demonstrate that the integration of UAV multispectral images with the Stacking model can enhance the precision of water quality parameter models, facilitates the examination of the spatial and temporal distribution of water quality parameters and the underlying influencing factors, and advances the capability for dynamic monitoring of water quality parameters in freshwater aquaculture regions. Concurrently, it offers fundamental theoretical and methodological assistance for the precise regulation of water quality in freshwater aquaculture ponds and the formulation of optimal production management strategies.

Text Analytics on YouTube Comments for Food Products

YouTube is a popular social media platform in the contemporary digital landscape. The primary focus of this study is to explore the underlying sentiment in user comments about food-related videos on YouTube, specifically within two pivotal food categories: plant-based and hedonic product. We labeled comments using sentiment lexicons such as TextBlob, VADER, and Google’s Sentiment Analysis (GSA) engine. Comment sentiment was classified using advanced Machine-Learning (ML) algorithms, namely Support Vector Machines (SVM), Multinomial Naive Bayes, Random Forest, Logistic Regression, and XGBoost. The evaluation of these models encompassed key macro average metrics, including accuracy, precision, recall, and F1 score. The results from GSA showed a high accuracy level, with SVM achieving 93% accuracy in the plant-based dataset and 96% in the hedonic dataset. In addition to sentiment analysis, we delved into user interactions within the two datasets, measuring crucial metrics, such as views, likes, comments, and engagement rate. The findings illuminate significantly higher levels of views, likes, and comments in the hedonic food dataset, but the plant-based dataset maintains a superior overall engagement rate.

Deep transfer learning-based bird species classification using mel spectrogram images.

The classification of bird species is of significant importance in the field of ornithology, as it plays an important role in assessing and monitoring environmental dynamics, including habitat modifications, migratory behaviors, levels of pollution, and disease occurrences. Traditional methods of bird classification, such as visual identification, were time-intensive and required a high level of expertise. However, audio-based bird species classification is a promising approach that can be used to automate bird species identification. This study aims to establish an audio-based bird species classification system for 264 Eastern African bird species employing modified deep transfer learning. In particular, the pre-trained EfficientNet technique was utilized for the investigation. The study adapts the fine-tune model to learn the pertinent patterns from mel spectrogram images specific to this bird species classification task. The fine-tuned EfficientNet model combined with a type of Recurrent Neural Networks (RNNs) namely Gated Recurrent Unit (GRU) and Long short-term memory (LSTM). RNNs are employed to capture the temporal dependencies in audio signals, thereby enhancing bird species classification accuracy. The dataset utilized in this work contains nearly 17,000 bird sound recordings across a diverse range of species. The experiment was conducted with several combinations of EfficientNet and RNNs, and EfficientNet-B7 with GRU surpasses other experimental models with an accuracy of 84.03% and a macro-average precision score of 0.8342.

Computer Vision Method for Automatic Detection of Microstructure Defects of Concrete.

The search for structural and microstructural defects using simple human vision is associated with significant errors in determining voids, large pores, and violations of the integrity and compactness of particle packing in the micro- and macrostructure of concrete. Computer vision methods, in particular convolutional neural networks, have proven to be reliable tools for the automatic detection of defects during visual inspection of building structures. The study's objective is to create and compare computer vision algorithms that use convolutional neural networks to identify and analyze damaged sections in concrete samples from different structures. Networks of the following architectures were selected for operation: U-Net, LinkNet, and PSPNet. The analyzed images are photos of concrete samples obtained by laboratory tests to assess the quality in terms of the defection of the integrity and compactness of the structure. During the implementation process, changes in quality metrics such as macro-averaged precision, recall, and F1-score, as well as IoU (Jaccard coefficient) and accuracy, were monitored. The best metrics were demonstrated by the U-Net model, supplemented by the cellular automaton algorithm: precision = 0.91, recall = 0.90, F1 = 0.91, IoU = 0.84, and accuracy = 0.90. The developed segmentation algorithms are universal and show a high quality in highlighting areas of interest under any shooting conditions and different volumes of defective zones, regardless of their localization. The automatization of the process of calculating the damage area and a recommendation in the "critical/uncritical" format can be used to assess the condition of concrete of various types of structures, adjust the formulation, and change the technological parameters of production.

Hybrid knowledge and data driven approach for prioritizing sewer sediment cleaning

The efficient assessment of sewer sediment condition is important for municipalities to formulate prioritization strategies for cleaning initiatives. However, manual assessment methods are plagued by inherent subjective and inaccuracy. To address these deficiencies, this paper introduces a hybrid approach integrating both knowledge-based principles and data-driven techniques for Sewer Sediment Cleaning Priority Assessment (SSCPA). The proposed approach exhibits a notable level of assessment accuracy, achieving macro-average precision, recall, and F1-score metrics of 87.9%, 88.0%, and 88.0%, respectively. These findings underscore the efficacy of SSCPA as a valuable tool for evaluating sewer sediment conditions, thereby enhancing the decision-making process for cleaning prioritization efforts. Future research should incorporate the probability of failure as a pivotal factor and explore the temporal dynamics of sewer sediment for more comprehensive insight.

Network intrusion detection system by applying ensemble model for smart home

The exponential advancements in recent technologies for surveillance become an important part of life. Though the internet of things (IoT) has gained more attention to develop smart infrastructure, it also provides a large attack surface for intruders. Therefore, it requires identifying the attacks as soon as possible to provide a secure environment. In this work, the network intrusion detection system, by applying the ensemble model (NIDSE) for Smart Homes is designed to identify the attacks in the smart home devices. The problem of classifying attacks is considered a classification predictive modeling using eXtreme gradient boosting (XGBoosting). It is an ensemble approach where the models are added sequentially to correct the errors until no further improvements or high performance can be made. The performance of the NIDSE is tested on the IoT network intrusion (IoT-NI) dataset. It has various types of network attacks, including host discovery, synchronized sequence number (SYN), acknowledgment (ACK), and hypertext transfer protocol (HTTP) flooding. Results from the cross-validation approach show that the XGBoosting classifier classifies the nine attacks with micro average precision of 94% and macro average precision of 85%.

Pengembangan Aplikasi Analisis Sentimen Penerapan PPKM Level 3

This study aims to analyze tweets about the imposition of Community Activity Restrictions (PPKM) level 3 to find the number of positive, negative, and neutral responses by building webbased software to conduct sentiment analysis on tweets data with the keywords “PPKM level 3". This research consists of two increments. The first increment is modeling and sentiment analysis and the second increment is making a website for visualizing the results. The results in the first increment are a model to perform sentiment analysis with an accuracy of 76.59% for training and for testing, with an accuracy of f1-score 75% with the test results on the macro average precision 82%, recall 73%, f1-score 75% with support of 20 and weighted average precision 80%, recall 75%, f1-score 75% with support of 20 and dataset analysis results with 30.82% positive sentiments, 17.80% neutral sentiments, and 51.36% negative sentiments. The result in the second increment is a website for visualization.

Traffic Fingerprints for Homogeneous IoT Traffic Based on Packet Payload Transition Patterns

Traffic fingerprint was considered an effective security protection mechanism in IoT scenarios because it can be used to automatically identify accessed devices. However, the results of replication experiments show that the classic traffic fingerprints based on simple network traffic attribute features have a significantly lower ability to identify accessed devices in real 5G IoT scenarios compared to what was stated in traditional IoT scenarios. The growing homogenization of IoT traffic caused by the application of 5G is believed to be the reason for the poor ability of traditional traffic fingerprints to identify 5G IoT terminals. Studying an enhanced traffic fingerprint is necessary to accommodate the homogeneous Internet of Things traffic. In addition, during the reproducing experiments, we noticed that the solution of overlap is a key factor that restricts the recognition ability of one-vs-all multi-classifiers, and the efficiency of existing methods still has some room for optimization. Based on targeted improvements to these two issues, we proposed an enhanced IoT terminal traffic fingerprint based on packet payload transition patterns to improve the device recognition ability in homogeneous IoT traffic. Additionally, we designed an improved solution for overlap based on density centers to expedite decision making. According to the experimental results, when compared with the existing traffic fingerprint, the proposed traffic fingerprint in this study demonstrated a Macro-Average Precision of close to 90% for network traffic from real 5G IoT terminals. The proposed overlap solution based on the density centers reduced the decision-making time from hundreds of seconds to tens of seconds while ensuring decision-making accuracy.

Blood Cell Revolution: Unveiling 11 Distinct Types with ‘Naturalize’ Augmentation

Artificial intelligence (AI) has emerged as a cutting-edge tool, simultaneously accelerating, securing, and enhancing the diagnosis and treatment of patients. An exemplification of this capability is evident in the analysis of peripheral blood smears (PBS). In university medical centers, hematologists routinely examine hundreds of PBS slides daily to validate or correct outcomes produced by advanced hematology analyzers assessing samples from potentially problematic patients. This process may logically lead to erroneous PBC readings, posing risks to patient health. AI functions as a transformative tool, significantly improving the accuracy and precision of readings and diagnoses. This study reshapes the parameters of blood cell classification, harnessing the capabilities of AI and broadening the scope from 5 to 11 specific blood cell categories with the challenging 11-class PBC dataset. This transformation facilitates a more profound exploration of blood cell diversity, surpassing prior constraints in medical image analysis. Our approach combines state-of-the-art deep learning techniques, including pre-trained ConvNets, ViTb16 models, and custom CNN architectures. We employ transfer learning, fine-tuning, and ensemble strategies, such as CBAM and Averaging ensembles, to achieve unprecedented accuracy and interpretability. Our fully fine-tuned EfficientNetV2 B0 model sets a new standard, with a macro-average precision, recall, and F1-score of 91%, 90%, and 90%, respectively, and an average accuracy of 93%. This breakthrough underscores the transformative potential of 11-class blood cell classification for more precise medical diagnoses. Moreover, our groundbreaking “Naturalize” augmentation technique produces remarkable results. The 2K-PBC dataset generated with “Naturalize” boasts a macro-average precision, recall, and F1-score of 97%, along with an average accuracy of 96% when leveraging the fully fine-tuned EfficientNetV2 B0 model. This innovation not only elevates classification performance but also addresses data scarcity and bias in medical deep learning. Our research marks a paradigm shift in blood cell classification, enabling more nuanced and insightful medical analyses. The “Naturalize” technique’s impact extends beyond blood cell classification, emphasizing the vital role of diverse and comprehensive datasets in advancing healthcare applications through deep learning.

Machine Learning-Assisted Discovery of Novel Anti-HIV Drug Candidates: An Analysis Using Molecular Datasets

Human Immunodeficiency Virus (HIV) serves as a crisis of global public health, necessitating new anti-HIV agents due to the virus's rapid mutation and subsequent drug resistance. While current Combination Antiretroviral Therapy (CART) has helped control the infection and mortality rates, traditional drug development approaches are costly and inefficient. This study aims to address this issue by applying machine learning algorithms for lead compound discovery using the extensive, quality-assured DTP Antiviral Screen Databases. Three molecular datasets—Extended-Connectivity Fingerprints (ECFP), Simplified Molecular-Input Line-Entry System (SMILES), and 2D molecular IMAGES—were processed using Principal Component Analysis (PCA), train-test splitting, and dataset balancing. Six machine learning algorithms were employed, including linear and nonlinear models, optimized through 5-fold cross-validation. The Area Under the Receiver Operating Characteristic (AUROC) curve was utilized to evaluate the models' performance, as well as the macro averaged precision, averaged recall, averaged F1 score, and balanced accuracy metrics. The ensemble models were constructed from the top-performing individual models. The best individual model, a SVM model trained on the ECFP dataset, achieved performance metrics of 0.78 on macro-averaged precision; 0.68 on macro-averaged recall; 0.72 on macro-averaged F1 score; 0.71 on balanced accuracy; and 0.75 on AUROC when evaluated on the testing data. The best ensemble model, fused with SVM, kNN, and logistic regression trained on the ECFP dataset, achieved performance metrics of 0.82 on macro-averaged precision; 0.67 on macro-averaged recall; 0.72 on macro-averaged F1 score; and 0.70 on balanced accuracy when evaluated on the testing data. The models were then applied to a Pubmed-extracted drug dataset, identifying several promising anti-HIV drug candidates, fulfilling the study's objective to improve the efficiency and success rate of new anti-HIV drug screening and discovery. In summary, this research demonstrates the transformative potential of machine learning in accelerating and optimizing the drug discovery process for HIV treatment.

Research on the Identification of a Winter Olympic Multi-Intent Chinese Problem Based on Multi-Model Fusion

Aiming at addressing the inability of traditional web technologies to effectively respond to Winter-Olympics-related user questions containing multiple intentions, this paper explores a multi-model fusion-based multi-intention recognition model BCNBLMATT to solve this problem. The model is proposed to address the characteristics of complex semantics, strong contextual relevance, and a large number of informative features of the Chinese problem text related to the Winter Olympics, as well as the limitations of the traditional word vector model, such as insufficient expression in the textual representation and the relative concern mechanism of feature expression. The BCNBLMATT model first obtains a comprehensive feature vector representation of the problem text through BERT. Then, a multi-scale text convolutional neural network model and a BiLSTM-Multi-heads attention model (a joint model combining a bidirectional long- and short-term attention network with a multi-head attention mechanism) are used to capture local features at more scales and contextually critical information features at more levels. Finally, the two obtained kinds of features are concatenated and fused to obtain richer and more comprehensive information about the problem text features, which improves the model’s performance in the multi-attention recognition task. Comparative experiments on the Winter Olympics Chinese question dataset and the MixATIS question dataset show that the BCNBLMATT model significantly improves the three metrics of macro-averaged precision, macro-averaged recall, and macro-averaged F1 value and exhibits better generalization. This study provides an effective solution to the multi-intent recognition task for Winter Olympic problems, overcomes the limitations of traditional models, and provides new ideas for improving the performance of multi-intent recognition.

A Resilience Approach for Diagnosing and Predicting HBV-Related Diseases Based on Blood Tests

Chronic hepatitis B virus (HBV) infection, which threatens global public health, is a major contributor to liver-related morbidity and mortality. Examinations for liver diseases related to chronic HBV infection—including laboratory tests, ultrasounds, computed tomography (CT), and liver biopsies—may take up medical resources, particularly since they overlap in most instances. Thus, there is an urgent need to establish an economical and effective diagnosis method in order to streamline the medical process for HBV-related diseases. Using complex network models constructed based on clinical blood tests, we provide such a method by defining the novel measure of functional resilience to assess patients’ liver conditions. By combining network models and dynamics, we discovered the pivotal items and their corresponding thresholds, which can guide further research on preventing disease deterioration in critical states of these diseases. The macro-averaged precision of our method, functional resilience, is 84.74%, whereas the macro-averaged precision of physicians’ experience without assistance from imaging or biopsy is 55.63%. From an economic perspective, our approach could save the equivalent of at least 30 USD per visit for most Chinese patients and at least 400 USD per visit for most US patients, compared with general diagnostic methods. Globally, this will add to savings of at least 10.5 billion USD annually. Our method can comprehensively evaluate the condition of patients’ livers and help avert the waste of medical resources during the diagnosis of liver disease by reducing excessive imaging exams.

DATA MINING IMPLEMENTATION FOR DETECTION OF ANOMALIES IN NETWORK TRAFFIC PACKETS USING OUTLIER DETECTION APPROACH

The large number of data packet records of network traffic can be used to evaluate the quality of a network as well as to analyze the occurrence of anomalies in the network, both related to network security and network performance. Based on the data obtained, the occurrence of anomalies in computer networks can not be detected specifically on which traffic packets. Meanwhile, to monitor network traffic packets manually will require a lot of time and resources, making it difficult to detect potential anomaly events more specifically. This study analyzes network packet traffic data to see records that include anomalies with an outlier detection approach, using the Isolation Forest algorithm to detect outliers on network traffic packet data, with the result that minority data are of the outliers type of 1,643 records (4.86%), while inliers are 32,098 records (95.13%). Then check and filter the expert attributes that contain expert information. The outlier detection results were classified using 5 algorithms as comparison, namely Random Forest Classifier, Support Vector Machine, Decision Tree Classifier, K-Nearest Neighbor, and Bernoulli Naive Bayes. The Random Forest algorithm has the highest score for accuracy, macro average precision, and macro average f1-score, namely 0.9962067330488383; 0.78; and 0.82. The classification model can be used to classify samples with labels "inliers", "outliers", "Error", and "warning outliers". There are labels that have scores for precision, recall, and f1-scrore that are not too high, namely the labels “error” (0.50; 1.00; and 0.67) and “warning outlier” (0.64; 0 .70; 0.67). The resulting classification model is used for prototype development that facilitates the process of investigating potential network traffic packet anomalies more specifically.

Modeling driving styles of online ride-hailing drivers with model identifiability and interpretability

The online ride-hailing industry has grown rapidly and has greatly improved travel efficiency worldwide. The driving styles of the drivers of ride-hailing vehicles largely determine the safety and comfort of passengers. Therefore, identifying and analyzing such driving styles are important tasks in order to maintain travel safety and efficiency. However, few studies have focused specifically on online ride-hailing drivers’ driving styles, and those few usually suffer from a lack of representativeness of influential factors as well as from failure to balance the identification accuracy and interpretability of the model applied. To address these issues, we propose an interpretable machine learning method to model the driving styles of online ride-hailing drivers. The proposed method considers both model identifiability and interpretability. First, we conducted naturalistic driving tests in Nanjing, China to collect driver data, vehicle kinematic data, and video data. Then, we extracted driver features, environmental features, and online ride-hailing driving features as the influential factors. Next, we used the CatBoost algorithm to identify three types of driving styles: aggressive, normal, and cautious. Finally, we used the Shapley additive explanation (SHAP) algorithm (1) to explore the effects of the influential factors on the driving styles from four perspectives (feature importance, total effect, main effect, and interaction effect) and (2) to analyze the differences in driving styles between online ride-hailing drivers and ordinary drivers. In addition, to verify the effectiveness of CatBoost, we compared it to three typical machine learning algorithms: extreme gradient boosting (XGBoost), artificial neural network, and support vector machine. The results show that CatBoost significantly outperformed the other three algorithms with macro-average precision, recall, and F1-score values of 0.818, 0.881, and 0.842, respectively. Also, SHAP was able to explain the complex nonlinear relationship between the driving styles and the influential factors. Overall, the distance, driving event, driver age, driving task, and duration features are of relatively high importance. This study provides an effective and innovative method to identify and analyze the driving styles of online ride-hailing drivers. Importantly, this method can help the online ride-hailing industry to monitor, analyze, and improve drivers’ driving behavior and thus improve travel safety and efficiency.

Establishment and Evaluation of Intelligent Diagnostic Model for Ophthalmic Ultrasound Images Based on Deep Learning.

The goal of the work described here was to construct a deep learning-based intelligent diagnostic model for ophthalmic ultrasound images to provide auxiliary analysis for the intelligent clinical diagnosis of posterior ocular segment diseases. The InceptionV3-Xception fusion model was established by using two pre-trained network models-InceptionV3 and Xception-in series to achieve multilevel feature extraction and fusion, and a classifier more suitable for the multiclassification recognition task of ophthalmic ultrasound images was designed to classify 3402 ophthalmic ultrasound images. The accuracy, macro-average precision, macro-average sensitivity, macro-average F1 value, subject working feature curves and area under the curve were used as model evaluation metrics, and the credibility of the model was assessed by testing the decision basis of the model using a gradient-weighted class activation mapping method. The accuracy, precision, sensitivity and area under the subject working feature curve of the InceptionV3-Xception fusion model on the test set reached 0.9673, 0.9521, 0.9528 and 0.9988, respectively. The model decision basis was consistent with the clinical diagnosis basis of the ophthalmologist, which proves that the model has good reliability. The deep learning-based ophthalmic ultrasound image intelligent diagnosis model can accurately screen and identify five posterior ocular segment diseases, which is beneficial to the intelligent development of ophthalmic clinical diagnosis.

Persistent and occasional: Searching for the variable population of the ZTF/4MOST sky using ZTF Data Release 11

Aims. We present a variability-, color-, and morphology-based classifier designed to identify multiple classes of transients and persistently variable and non-variable sources from the Zwicky Transient Facility (ZTF) Data Release 11 (DR11) light curves of extended and point sources. The main motivation to develop this model was to identify active galactic nuclei (AGN) at different redshift ranges to be observed by the 4MOST Chilean AGN/Galaxy Evolution Survey (ChANGES). That being said, it also serves as a more general time-domain astronomy study. Methods. The model uses nine colors computed from CatWISE and Pan-STARRS1 (PS1), a morphology score from PS1, and 61 single-band variability features computed from the ZTF DR11 g and r light curves. We trained two versions of the model, one for each ZTF band, since ZTF DR11 treats the light curves observed in a particular combination of field, filter, and charge-coupled device (CCD) quadrant independently. We used a hierarchical local classifier per parent node approach-where each node is composed of a balanced random forest model. We adopted a taxonomy with 17 classes: non-variable stars, non-variable galaxies, three transients (SNIa, SN-other, and CV/Nova), five classes of stochastic variables (lowz-AGN, midz-AGN, highz-AGN, Blazar, and YSO), and seven classes of periodic variables (LPV, EA, EB/EW, DSCT, RRL, CEP, and Periodic-other). Results. The macro-averaged precision, recall, and F1-score are 0.61, 0.75, and 0.62 for the g-band model, and 0.60, 0.74, and 0.61, for the r-band model. When grouping the four AGN classes (lowz-AGN, midz-AGN, highz-AGN, and Blazar) into one single class, its precision-recall, and F1-score are 1.00, 0.95, and 0.97, respectively, for both the g and r bands. This demonstrates the good performance of the model in classifying AGN candidates. We applied the model to all the sources in the ZTF/4MOST overlapping sky (−28 ≤ Dec ≤ 8.5), avoiding ZTF fields that cover the Galactic bulge (|gal_b| ≤ 9 and gal_l ≤ 50). This area includes 86 576 577 light curves in the g band and 140 409 824 in the r band with 20 or more observations and with an average magnitude in the corresponding band lower than 20.5. Only 0.73% of the g-band light curves and 2.62% of the r-band light curves were classified as stochastic, periodic, or transient with high probability (Pinit ≥ 0.9). Even though the metrics obtained for the two models are similar, we find that, in general, more reliable results are obtained when using the g-band model. With it, we identified 384 242 AGN candidates (including low-, mid-, and high-redshift AGN and Blazars), 287 156 of which have Pinit ≥ 0.9.

Comparing Machine Learning Techniques for Predictions of Motorway Segment Crash Risk Level

Motorways are typically the safest road environment in terms of injury crashes per million vehicle kilometres; however, given the high severity of crashes occurring therein, there is still space for road safety improvements. The objective of this study is to compare the classification performance of five machine learning techniques for predictions of crash risk levels of motorway segments. To that end, data on crash risk levels, driving behaviour metrics, and road geometry characteristics of 668 motorway segments were exploited. The utilized dataset was divided into training and test subsets, with a proportion of 75% and 25%, respectively. The training subset was used to train the models, whereas the test subset was used for the evaluation of their performance. The response variable of the models was the crash risk level of the considered motorway segments, while the predictors were various road design characteristics and naturalistic driving behaviour metrics. The techniques considered were Logistic Regression, Decision Tree, Random Forest, Support Vector Machine, and K-Nearest Neighbours. Among the five techniques, the Random Forest model achieved the best classification performance (overall accuracy: 89.3%, macro-averaged precision: 89.0%, macro-averaged recall: 88.4%, macro-averaged F1 score: 88.6%). Moreover, the Shapley additive explanations were calculated in order to assist with the interpretation of the model’s outcomes. The findings of this study are particularly useful as the Random Forest model could be used as a highly promising proactive road safety tool for identifying potentially hazardous motorway segments.