Best-performing Model Research Articles

Android malware defense through a hybrid multi-modal approach

The rapid proliferation of Android apps has given rise to a dark side, where increasingly sophisticated malware poses a formidable challenge for detection. To combat this evolving threat, we present an explainable hybrid multi-modal framework. This framework leverages the power of deep learning, with a novel model fusion technique, to illuminate the hidden characteristics of malicious apps. Our approach combines models (leveraging late fusion approach) trained on attributes derived from static and dynamic analysis, hence utilizing the unique strengths of each model. We thoroughly analyze individual feature categories, feature ensembles, and model fusion using traditional machine learning classifiers and deep neural networks across diverse datasets. Our hybrid fused model outperforms others, achieving an F1-score of 99.97% on CICMaldroid2020. We use SHAP (SHapley Additive exPlanations) and t-SNE (t-distributed Stochastic Neighbor Embedding) to further analyze and interpret the best-performing model. We highlight the efficacy of our architectural design through an ablation study, revealing that our approach consistently achieves over 99% detection accuracy across multiple deep learning models. This paves the way groundwork for substantial advancements in security and risk mitigation within interconnected Android OS environments.

P-TIMA: a framework of T witter threat intelligence mining and analysis based on a prompt-learning NER model

Abstract Open-source information platforms such as Twitter continuously provide the latest threat intelligence, including new vulnerabilities and in-the-wild exploitations of advanced persistent threat (APT) groups. Automated extraction of threat intelligence from Twitter has become crucial for defenders to access up-to-date threat knowledge. However, existing studies mainly rely on supervised learning methods to extract threat intelligence knowledge, such as entities, which require a large amount of annotated data. This paper presents Threat Intelligence Mining and Analysis based on Prompt Learning (P-TIMA), a framework specifically crafted for extracting and analyzing threat intelligence from Twitter. P-TIMA employs our innovative few-shot entity recognition method, SecEntPrompt (SEP), built on prompt learning, to extract vulnerability intelligence from Twitter. Additionally, P-TIMA analyzes and profiles the overarching vulnerability intelligence obtained from Twitter, along with in-the-wild exploitation intelligence of APT groups. The SEP improves the average entity recognition F1 score by 3.62-4.40 compared with the best-performing comparison model and outperforms the method based on the large language model on recognition performance and inference time. To validate our framework, we apply P-TIMA to extract vulnerability-related threat intelligence from real Twitter data. Through case studies, we then analyze trends in vulnerability threats and the exploitation capabilities of APT groups. In conclusion, our framework provides a more efficient and accurate method for extracting threat intelligence from Twitter, enabling defenders to stay up-to-date with the latest threat trends and helping them improve their defense strategies against cyber attacks.

Enhancing Sex Estimation Accuracy with Cranial Angle Measurements and Machine Learning.

The development of current sexing methods largely depends on the use of adequate sources of data and adjustable classification techniques. Most sex estimation methods have been based on linear measurements, while the angles have been largely ignored, potentially leading to the loss of valuable information for sex discrimination. This study aims to evaluate the usefulness of cranial angles for sex estimation and to differentiate the most dimorphic ones by training machine learning algorithms. Computed tomography images of 154 males and 180 females were used to derive data of 36 cranial angles. The classification models were created by support vector machines, naïve Bayes, logistic regression, and the rule-induction algorithm CN2. A series of cranial angle subsets was arranged by an attribute selection scheme. The algorithms achieved the highest accuracy on subsets of cranial angles, most of which correspond to well-known features for sex discrimination. Angles characterizing the lower forehead and upper midface were included in the best-performing models of all algorithms. The accuracy results showed the considerable classification potential of the cranial angles. The study demonstrates the value of the cranial angles as sex indicators and the possibility to enhance the sex estimation accuracy by using them.

Developing large language models to detect adverse drug events in posts on X

ABSTRACT Adverse drug events (ADEs) are one of the major causes of hospital admissions and are associated with increased morbidity and mortality. Post-marketing ADE identification is one of the most important phases of drug safety surveillance. Traditionally, data sources for post-marketing surveillance mainly come from spontaneous reporting system such as the Food and Drug Administration Adverse Event Reporting System (FAERS). Social media data such as posts on X (formerly Twitter) contain rich patient and medication information and could potentially accelerate drug surveillance research. However, ADE information in social media data is usually locked in the text, making it difficult to be employed by traditional statistical approaches. In recent years, large language models (LLMs) have shown promise in many natural language processing tasks. In this study, we developed several LLMs to perform ADE classification on X data. We fine-tuned various LLMs including BERT-base, Bio_ClinicalBERT, RoBERTa, and RoBERTa-large. We also experimented ChatGPT few-shot prompting and ChatGPT fine-tuned on the whole training data. We then evaluated the model performance based on sensitivity, specificity, negative predictive value, positive predictive value, accuracy, F1-measure, and area under the ROC curve. Our results showed that RoBERTa-large achieved the best F1-measure (0.8) among all models followed by ChatGPT fine-tuned model with F1-measure of 0.75. Our feature importance analysis based on 1200 random samples and RoBERTa-Large showed the most important features are as follows: “withdrawals”/”withdrawal”, “dry”, “dealing”, “mouth”, and “paralysis”. The good model performance and clinically relevant features show the potential of LLMs in augmenting ADE detection for post-marketing drug safety surveillance.

Applying machine learning approaches for predicting obesity risk using US health administrative claims database

IntroductionBody mass index (BMI) is inadequately recorded in US administrative claims databases. We aimed to validate the sensitivity and positive predictive value (PPV) of BMI-related diagnosis codes using an electronic...

Interpretable prediction of remanence in sintered NdFeB through machine learning strategy

The effects of composition on magnetic properties of sintered NdFeB are generally revealed by trial-and-error method, which is actually time-consuming and resource-intensive. It has been demonstrated that data-driven machine learning (ML) strategy has a great potential in accelerating processes of material research. In this study, accurate prediction of remanence in sintered NdFeB magnets has been achieved by a high-quality and few-shot dataset. We used a set of physical and chemical properties of element as inputs to build ML models based on element classification knowledge. Specifically, according to the position of elements, we divided them into three categories and then calculated the average properties of three positions separately. The strong generalization ability of best-performing ML model was further verified in nineteen unknown samples with only 0.724 % error. In addition, the symbolic regressor (SR) and SHapley Additive exPlanation (SHAP) algorithm were employed to explain the model results from physical perspective. In short, a high-efficiency and feasible ML strategy was developed for predicting remanence of sintered NdFeB magnets. We believe that this work will show great prospects in the composition design and performance screening of sintered NdFeB magnets.

Task decomposition strategy based on machine learning for boosting performance and identifying mechanisms in heterogeneous activation of peracetic acid process

Heterogeneous activation of peracetic acid (PAA) process is a promising method for removing organic pollutants from water. Nevertheless, this process is constrained by several complex factors, such as the selection of catalysts, optimization of reaction conditions, and identification of mechanism. In this study, a task decomposition strategy was adopted by combining a catalyst and reaction condition optimization machine learning (CRCO-ML) model and a mechanism identification machine learning (MI-ML) model to address these issues. The Categorical Boosting (CatBoost) model was identified as the best-performing model for the dataset (1024 sets and 7122 data points) in this study, achieving an R2 of 0.92 and an RMSE of 1.28. Catalyst composition, PAA dosage, and catalyst dosage were identified as the three most important features through SHAP analysis in the CRCO-ML model. The HCO3– is considered the most influential water matrix affecting the k value. The errors between all reverse experiment results and the predictions of the CRCO-ML and MI-ML models were <10 % and 15 %, respectively. This interdisciplinary work provides novel insights into the design and application of the heterogeneous activation of PAA process, significantly contributing to the rapid development of this technology.

Utilizing machine learning and hemagglutinin sequences to identify likely hosts of influenza H3Nx viruses

Influenza is a disease that represents both a public health and agricultural risk with pandemic potential. Among the subtypes of influenza A virus, H3 influenza virus can infect many avian and mammalian species and is therefore a virus of interest to human and veterinary public health. The primary goal of this study was to train and validate classifiers for the identification of the most likely host species using the hemagglutinin gene segment of H3 viruses. A five-step process was implemented, which included training four machine learning classifiers, testing the classifiers on the validation dataset, and further exploration of the best-performing model on three additional datasets. The gradient boosting machine classifier showed the highest host-classification accuracy with a 98.0 % (95 % CI [97.01, 98.73]) correct classification rate on an independent validation dataset. The classifications were further analyzed using the predicted probability score which highlighted sequences of particular interest. These sequences were both correctly and incorrectly classified sequences that showed considerable predicted probability for multiple hosts. This showed the potential of using these classifiers for rapid sequence classification and highlighting sequences of interest. Additionally, the classifiers were tested on a separate swine dataset composed of H3N2 sequences from 1998 to 2003 from the United States of America, and a separate canine dataset composed of canine H3N2 sequences of avian origin. These two datasets were utilized to look at the applications of predicted probability and host convergence over time. Lastly, the classifiers were used on an independent dataset of environmental sequences to explore the host identification of environmental sequences. The results of these classifiers show the potential for machine learning to be used as a host identification technique for viruses of unknown origin on a species-specific level.

Development of DRIP - drought representation index for CMIP climate model performance, application to Southeast Brazil

With the escalating impacts of drought events driven by climate change, reducing the uncertainty of drought projections becomes critical for enhancing risk management and adaptation strategies. This study aimed to develop an index for assessing the performance of CMIP6 Global Climate Models in simulating meteorological drought scenarios across regional hydrological systems, intended to provide more reliable information for management purposes. Named the ‘Drought Representation Index for CMIP Climate Model Performance’ (DRIP), this index evaluates CMIP models' performance to represent drought severity, duration, and return period. DRIP was used to select CMIP models and create an ensemble of the best-performing models (E-DRIP) to improve the reliability of drought projections. E-DRIP was then compared with a general ensemble of available CMIP6 models (E-CMIP). We applied this method in Southeast Brazil, a region known for its climate uncertainties and low predictability; specifically, it was implemented within the Paraíba do Sul River Basin, a nationally strategic watershed in a highly populated and industrialized area, which has recently faced unprecedented drought-related water crises. Results showed that DRIP effectively assessed the individual performance of CMIP models, which exhibited considerable variability, and identified the top-performing models for a multi-model ensemble. Additionally, the E-DRIP ensemble significantly reduced uncertainties in drought projections, achieving an average reduction of 63 % in the study area compared to E-CMIP. Furthermore, the proposed method enables evaluations across any standardized drought index scale, reference period, or threshold, and can be readily adapted to other hydrological systems.

Application of multimodal learning in daylight provision and view quality assessment of residential building layouts

Optimizing the layout of residential buildings based on daylight performance and view quality is crucial to visual comfort and well-being of building occupants. Machine Learning (ML) methods offer valuable support for performance-based decision-making process at the early-stage building design. In this study, a novel workflow is introduced to integrate ML models into the architectural design process. With the designer’s input floor layout designs, the presented multimodal ML model predicts daylight provision and view quality, which are then translated into practical visual representations by a post-processing step. This approach allows input designs to be evaluated by the ML model, leading to enhanced design decisions while preserving the designer’s autonomy. Results for the best-performing model, implementing ResNet50 and a fully connected network, led to a Mean Square Error (MSE) of 0.0440 and 0.0478, and an R2 score of 0.7411 and 0.7815 for the daylight and view metrics, respectively. The results of the daylight and view predictive models are further interpreted according to different apartment categories and at various resolutions. These results indicate that the method could be viable for predicting daylight provision and view quality in early design tools, providing designers with faster feedback that supports informed decision-making during design iterations. Ultimately, the challenges of the study and further improvements are discussed.

Performance Analysis of YOLO and Detectron2 Models for Detecting Corn and Soybean Pests Employing Customized Dataset

One of the most challenging aspects of agricultural pest control is accurate detection of insects in crops. Inadequate control measures for insect pests can seriously impact the production of corn and soybean plantations. In recent years, artificial intelligence (AI) algorithms have been extensively used for detecting insect pests in the field. In this line of research, this paper introduces a method to detect four key insect species that are predominant in Brazilian agriculture. Our model relies on computer vision techniques, including You Only Look Once (YOLO) and Detectron2, and adapts them to lightweight formats—TensorFlow Lite (TFLite) and Open Neural Network Exchange (ONNX)—for resource-constrained devices. Our method leverages two datasets: a comprehensive one and a smaller sample for comparison purposes. With this setup, the authors aimed at using these two datasets to evaluate the performance of the computer vision models and subsequently convert the best-performing models into TFLite and ONNX formats, facilitating their deployment on edge devices. The results are promising. Even in the worst-case scenario, where the ONNX model with the reduced dataset was compared to the YOLOv9-gelan model with the full dataset, the precision reached 87.3%, and the accuracy achieved was 95.0%.

Study on risk factors of impaired fasting glucose and development of a prediction model based on Extreme Gradient Boosting algorithm.

The aim of this study was to develop and validate a machine learning-based model to predict the development of impaired fasting glucose (IFG) in middle-aged and older elderly people over a 5-year period using data from a cohort study. This study was a retrospective cohort study. The study population was 1855 participants who underwent consecutive physical examinations at the First Affiliated Hospital of Soochow University between 2018 and 2022.The dataset included medical history, physical examination, and biochemical index test results. The cohort was randomly divided into a training dataset and a validation dataset in a ratio of 8:2. The machine learning algorithms used in this study include Extreme Gradient Boosting (XGBoost), Support Vector Machines (SVM), Naive Bayes, Decision Trees (DT), and traditional Logistic Regression (LR). Feature selection, parameter optimization, and model construction were performed in the training set, while the validation set was used to evaluate the predictive performance of the models. The performance of these models is evaluated by an area under the receiver operating characteristic (ROC) curves (AUC), calibration curves and decision curve analysis (DCA). To interpret the best-performing model, the Shapley Additive exPlanation (SHAP) Plots was used in this study. The training/validation dataset consists of 1,855 individuals from the First Affiliated Hospital of Soochow University, yielded significant variables following selection by the Boruta algorithm and logistic multivariate regression analysis. These significant variables included systolic blood pressure (SBP), fatty liver, waist circumference (WC) and serum creatinine (Scr). The XGBoost model outperformed the other models, demonstrating an AUC of 0.7391 in the validation set. The XGBoost model was composed of SBP, fatty liver, WC and Scr may assist doctors with the early identification of IFG in middle-aged and elderly people.

Mitochondrial-related genes as prognostic and metastatic markers in breast cancer: insights from comprehensive analysis and clinical models.

Breast cancer (BC) constitutes a significant peril to global women's health. Contemporary research progressively suggests that mitochondrial dysfunction plays a pivotal role in both the inception and advancement of BC. However, investigations delving into the correlation between mitochondrial-related genes (MRGs) and the prognosis and metastasis of BC are still infrequent. Utilizing data from the TCGA database, we employed the "limma" R package for differential expression analysis. Subsequently, both univariate and multivariate Cox regression analyses were executed, alongside LASSO Cox regression analysis, to pinpoint prognostic MRGs and to further develop the prognostic model. External validation (GSE88770 merged GSE425680) and internal validation were further conducted. Our investigation delved into a broad spectrum of analyses that included functional enrichment, metabolic and immune characteristics, immunotherapy response prediction, intratumor heterogeneity (ITH), mutation, tumor mutational burden (TMB), microsatellite instability (MSI), cellular stemness, single-cell, and drug sensitivity analysis. We validated the protein and mRNA expressions of prognostic MRGs in tissues and cell lines through immunohistochemistry and qRT-PCR. Moreover, leveraging the GSE102484 dataset, we conducted differential gene expression analysis to identify MRGs related to metastasis, subsequently developing metastasis models via 10 distinct machine-learning algorithms and then selecting the best-performing model. The division between training and validation cohorts was set at 70% and 30%, respectively. A prognostic model was constructed by 9 prognostic MRGs, which were DCTPP1, FEZ1, KMO, NME3, CCR7, ISOC2, STAR, COMTD1, and ESR2. Patients within the high-risk group experienced more adverse outcomes than their counterparts in the low-risk group. The ROC curves and constructed nomogram showed that the model exhibited an excellent ability to predict overall survival (OS) for patients and the risk score was identified as an independent prognostic factor. The functional enrichment analysis showed a strong correlation between metabolic progression and MRGs. Additional research revealed that the discrepancies in outcomes between the two risk categories may be attributed to a variety of metabolic and immune characteristics, as well as differences in intratumor heterogeneity (ITH), tumor mutational burden (TMB), and cancer stemness indices. ITH, TIDE, and IPS analyses suggested that patients possessing a low-risk score may exhibit enhanced responsiveness to immunotherapy. Additionally, distant metastasis models were established by PDK4, NRF1, DCAF8, CHPT1, MARS2 and NAMPT. Among these, the XGBoost model showed the best predicting ability. In conclusion, MRGs significantly influence the prognosis and metastasis of BC. The development of dual clinical prediction models offers crucial insights for tailored and precise therapeutic strategies, and paves the way for exploring new avenues in understanding the pathogenesis of BC.

Model Predictive Control of a Stand-Alone Hybrid Battery-Hydrogen Energy System: A Case Study of the PHOEBUS Energy System

Model predictive control is a promising approach to robustly control complex energy systems, such as hybrid battery-hydrogen energy storage systems that enable seasonal storage of renewable energies. However, deriving a mathematical model of the energy system suitable for model predictive control is difficult due to the unique characteristics of each energy system component. This work introduces mixed integer linear programming models to describe the nonlinear multidimensional operational behavior of components using piecewise linear functions. Furthermore, this paper develops a new approach for deriving a strategy for seasonal storage of renewable energies using cost factors in the objective function of the optimization problem while considering degradation effects. An experimentally validated simulation model of the PHOEBUS Energy System is utilized to compare the performance of two model predictive controllers with a hysteresis band controller such as utilized for the real-world system. Furthermore, the sensitivity of the model predictive controller to the prediction horizon length and the temporal resolution is investigated. The prediction horizon was found to have the highest impact on the performance of the model predictive controller. The best-performing model predictive controller with a 14-day prediction horizon and perfect foresight increased the total energy stored at the end of the year by 18.9% while decreasing the degradation of the electrolyzer and the fuel cell.

Abstract B019: A machine learning approach to predicting smoking cessation outcomes among Spanish-speaking smokers who completed a culturally targeted intervention

Abstract Introduction: Among the Hispanic/Latinx population in the US, four of five leading causes of death are attributed to smoking. Few evidence-based interventions have been developed to improve smoking cessation for this population. Our team has demonstrated the efficacy of a culturally targeted, extended self-help intervention for Spanish-speaking smokers, estimating a smoking abstinence rate of 33% at 24 months, compared to 24% for usual care. Further efforts are needed to enhance this efficacious, low-cost, scalable intervention approach. Machine learning (ML) is one approach to increase understanding of the predictors of treatment outcomes and inform strategies to improve the intervention. Toward that goal, this secondary data analysis utilized the decision tree (DT) model to predict self-reported 7-day point prevalence abstinence at 18-month follow-up for participants receiving our intervention. Method: Data from participants who reported smoking status at an 18-month follow-up (N=332 of 714 enrolled, 36% abstinent) were randomly split (80:20) into training and test datasets based on smoking status. We entered demographics, psychosocial, and smoking variables collected at baseline as predictors in the model. In addition to handling missing values and normalizing numeric features, we employed the Synthetic Minority Over-sampling Technique (SMOTE) combined with Tomek links (SMOTE-Tomek) to improve inter-class separability. Recursive Feature Elimination (RFE) using DT was implemented to identify the most relevant predictors. The cross-validation (CV) pipeline incorporated preprocessing, feature selection, handling class-imbalance, over-sampling, and model training with a DT as the interpretable classifier. Hyperparameters were tuned using stratified K-fold (K=5) CV to optimize parameters using Grid search. F1 score was selected as the primary metric as it accounts for each class performance by combining the precision and recall metrics. Finally, the performance of the classifier was evaluated in 20% of the unseen dataset. Results: The DT classifier for the CV achieved the F1 score of 0.73 [95% CI 0.66, 0.77], revealing a reasonably good performance in identifying smokers at 18-months. RFE selected familism, age, affect, and confidence as the most relevant predictors of smoking status. Individuals who had lower familism scores (&amp;lt;138), combined with higher age (&amp;gt;50), were more likely to be smokers at 18 months. Individuals who had higher familism scores (&amp;gt;138) combined with higher negative affect (&amp;gt;42) were more likely to be smokers. Conclusion: This study provides the first step toward personalized care for smoking cessation among the Hispanic/Latinx population and demonstrates the potential of DT classifier to predict cessation outcomes among individuals who received and completed a culturally and linguistically targeted smoking cessation intervention. Our future analysis will compare this model with conventional statistical modeling approaches and other ML algorithms to identify the best-performing parsimonious model. Citation Format: Ranjita Poudel, K. Ruwani M. Fernando, Matthew B. Schabath, Steven K. Sutton, Thomas H. Brandon, Issam El Naqa, Vani N. Simmons. A machine learning approach to predicting smoking cessation outcomes among Spanish-speaking smokers who completed a culturally targeted intervention [abstract]. In: Proceedings of the 17th AACR Conference on the Science of Cancer Health Disparities in Racial/Ethnic Minorities and the Medically Underserved; 2024 Sep 21-24; Los Angeles, CA. Philadelphia (PA): AACR; Cancer Epidemiol Biomarkers Prev 2024;33(9 Suppl):Abstract nr B019.

Mortality Risk Assessment Using Deep Learning-Based Frequency Analysis of EEG and EOG in Sleep

Abstract Study Objectives To assess whether the frequency content of electroencephalography (EEG) and electrooculography (EOG) during nocturnal polysomnography (PSG) can predict all-cause mortality. Methods Power spectra from PSGs of 8,716 participants, included from the MrOS Sleep Study and the Sleep Heart Health Study (SHHS), were analyzed in deep learning-based survival models. The best-performing model was further examined using SHapley Additive Explanation (SHAP) for data-driven sleep-stage specific definitions of power bands, which were evaluated in predicting mortality using Cox Proportional Hazards models. Results Survival analyses, adjusted for known covariates, identified multiple EEG frequency bands across all sleep stages predicting all-cause mortality. For EEG, we found an all-cause mortality hazard ratio (HR) of 0.90 (CI95% 0.85–0.96) for 12–15 Hz in N2, 0.86 (CI95% 0.82–0.91) for 0.75–1.5 Hz in N3, and 0.87 (CI95% 0.83–0.92) for 14.75–33.5 Hz in REM sleep. For EOG, we found several low-frequency effects including an all-cause mortality HR of 1.19 (CI95% 1.11–1.28) for 0.25 Hz in N3, 1.11 (CI95% 1.03–1.21) for 0.75 Hz in N1, and 1.11 (CI95% 1.03–1.20) for 1.25–1.75 Hz in Wake. The gain in the concordance index (C-index) for all-cause mortality is minimal, with only a 0.24% increase: The best single mortality predictor was EEG N3 (0-0.5 Hz) with C-index of 77.78% compared to 77.54% for confounders alone. Conclusion Spectral power features, possibly reflecting abnormal sleep microstructure, are associated with mortality risk. These findings add to a growing literature suggesting that sleep contains incipient predictors of health and mortality.

Capabilities of GPT-4 in ophthalmology: an analysis of model entropy and progress towards human-level medical question answering

BackgroundEvidence on the performance of Generative Pre-trained Transformer 4 (GPT-4), a large language model (LLM), in the ophthalmology question-answering domain is needed.MethodsWe tested GPT-4 on two 260-question multiple choice question...

A Complex Background SAR Ship Target Detection Method Based on Fusion Tensor and Cross-Domain Adversarial Learning

Synthetic Aperture Radar (SAR) ship target detection has been extensively researched. However, most methods use the same dataset division for both training and validation. In practical applications, it is often necessary to quickly adapt to new loads, new modes, and new data to detect targets effectively. This presents a cross-domain detection problem that requires further study. This paper proposes a method for detecting SAR ships in complex backgrounds using fusion tensor and cross-domain adversarial learning. The method is designed to address the cross-domain detection problem of SAR ships with large differences between the training and test sets. Specifically, it can be used for the cross-domain detection task from the fully polarised medium-resolution ship dataset (source domain) to the high-resolution single-polarised dataset (target domain). This method proposes a channel fusion module (CFM) based on the YOLOV5s model. The CFM utilises the correlation between polarised channel images during training to enrich the feature information of single-polarised images extracted by the model during inference. This article proposes a module called the cross-domain adversarial learning module (CALM) to reduce overfitting and achieve adaptation between domains. Additionally, this paper introduces the anti-interference head (AIH) which decouples the detection head to reduce the conflict of classification and localisation problems. This improves the anti-interference and generalisation ability in complex backgrounds. This paper conducts cross-domain experiments using the constructed medium-resolution SAR full polarisation dataset (SFPD) as the source domain and the high-resolution single-polarised ship detection dataset (HRSID) as the target domain. Compared to the best-performing YOLOV8s model among typical mainstream models, this model improves precision by 4.9%, recall by 3.3%, AP by 2.4%, and F1 by 3.9%. This verifies the effectiveness of the method and provides a useful reference for improving cross-domain learning and model generalisation capability in the field of target detection.

Personalized Prediction of Long-Term Renal Function Prognosis Following Nephrectomy Using Interpretable Machine Learning Algorithms: Case-Control Study.

Acute kidney injury (AKI) is a common adverse outcome following nephrectomy. The progression from AKI to acute kidney disease (AKD) and subsequently to chronic kidney disease (CKD) remains a concern; yet, the predictive mechanisms for these transitions are not fully understood. Interpretable machine learning (ML) models offer insights into how clinical features influence long-term renal function outcomes after nephrectomy, providing a more precise framework for identifying patients at risk and supporting improved clinical decision-making processes. This study aimed to (1) evaluate postnephrectomy rates of AKI, AKD, and CKD, analyzing long-term renal outcomes along different trajectories; (2) interpret AKD and CKD models using Shapley Additive Explanations values and Local Interpretable Model-Agnostic Explanations algorithm; and (3) develop a web-based tool for estimating AKD or CKD risk after nephrectomy. We conducted a retrospective cohort study involving patients who underwent nephrectomy between July 2012 and June 2019. Patient data were randomly split into training, validation, and test sets, maintaining a ratio of 76.5:8.5:15. Eight ML algorithms were used to construct predictive models for postoperative AKD and CKD. The performance of the best-performing models was assessed using various metrics. We used various Shapley Additive Explanations plots and Local Interpretable Model-Agnostic Explanations bar plots to interpret the model and generated directed acyclic graphs to explore the potential causal relationships between features. Additionally, we developed a web-based prediction tool using the top 10 features for AKD prediction and the top 5 features for CKD prediction. The study cohort comprised 1559 patients. Incidence rates for AKI, AKD, and CKD were 21.7% (n=330), 15.3% (n=238), and 10.6% (n=165), respectively. Among the evaluated ML models, the Light Gradient-Boosting Machine (LightGBM) model demonstrated superior performance, with an area under the receiver operating characteristic curve of 0.97 for AKD prediction and 0.96 for CKD prediction. Performance metrics and plots highlighted the model's competence in discrimination, calibration, and clinical applicability. Operative duration, hemoglobin, blood loss, urine protein, and hematocrit were identified as the top 5 features associated with predicted AKD. Baseline estimated glomerular filtration rate, pathology, trajectories of renal function, age, and total bilirubin were the top 5 features associated with predicted CKD. Additionally, we developed a web application using the LightGBM model to estimate AKD and CKD risks. An interpretable ML model effectively elucidated its decision-making process in identifying patients at risk of AKD and CKD following nephrectomy by enumerating critical features. The web-based calculator, found on the LightGBM model, can assist in formulating more personalized and evidence-based clinical strategies.

Pressure distribution based 2D in-bed keypoint prediction under interfered scenes

In-bed pose estimation holds significant potential in various domains, including healthcare, sleep studies, and smart homes. Pressure-sensitive bed sheets have emerged as a promising solution for addressing this task considering the advantages of convenience, comfort, and privacy protection. However, existing studies primarily rely on ideal datasets that do not consider the presence of common daily objects such as pillows and quilts referred to as interference, which can significantly impact the pressure distribution. As a result, there is still a gap between the models trained with ideal data and the real-life application. Besides the end-to-end training approach, one potential solution is to recognize the interference and fuse the interference information to the model during training. In this study, we created a well-annotated dataset, consisting of eight in-bed scenes and four common types of interference: pillows, quilts, a laptop, and a package. To facilitate the analysis, the pixels in the pressure image were categorized into five classes based on the relative position between the interference and the human. We then evaluated the performance of five neural network models for pixel-level interference recognition. The best-performing model achieved an accuracy of 80.0% in recognizing the five categories. Subsequently, we validated the utility of interference recognition in improving pose estimation accuracy. The ideal model initially shows an average joint position error of up to 30.59 cm and a Percentage of Correct Keypoints (PCK) of 0.332 on data from scenes with interferences. After retraining on data including interference, the error is reduced to 13.54 cm and the PCK increases to 0.747. By integrating interference recognition information, either by excluding the parts of the interference or using the recognition results as input, the error can be further minimized to 12.44 cm and the PCK can be maximized up to 0.777. Our findings represent an initial step towards the practical deployment of pressure-sensitive bed sheets in everyday life.