Inference Model Research Articles

Elucidating Development Trajectories of Brain Functional Abnormalities in Major Depressive Disorder Utilizing a Data-Driven Disease Progression Model.

Concerns have arisen regarding the heterogeneity of patients with major depressive disorder (MDD), particularly when the varying disease progression trajectories among individuals are overlooked. Recognizing these distinct trajectories is crucial for personalized assessments and accurate disease progression predictions in MDD, posing a significant challenge in clinical practice. We utilized a data-driven subtype and stage inference (SuStaIn) model to infer trajectories based on cross-sectional amplitude of low-frequency fluctuations (ALFF) derived from resting-state functional magnetic resonance imaging data of 833 patients with MDD and 834 healthy controls. Based on distinct trajectories, two subtypes of MDD were identified: Subtype 1 showed declining ALFF from paracentral lobule (PCL) to thalamus to medial orbitofrontal cortex (OFCmed), with higher core depression scores and gray matter atrophy, whereas Subtype 2 had an opposing trajectory, with initial OFCmed ALFF decrease gradually extending to PCL. Our findings contribute to a better understanding of MDD heterogeneity and facilitate precise disease progression predictions.

Automatic generation of monitoring report based on large language model and knowledge graph inference

Automatic generation of monitoring report based on large language model and knowledge graph inference

Effects of the triglyceride-glucose index on non-alcoholic fatty liver disease: Causal evidence from longitudinal cohort studies.

Effects of the triglyceride-glucose index on non-alcoholic fatty liver disease: Causal evidence from longitudinal cohort studies.

Light absorption of black and brown carbon in eastern China.

Light absorption of black and brown carbon in eastern China.

Multiword target-independent transformer-based model for financial sentiment analysis in colloquial Cantonese

Tokenization process decomposes a multi-word-span instrument name into several tokens and the transformer attention mechanism handles each token individually, thus hindering the treatment of the related tokens as a single entity. The existence of multiple instruments in a single message further exaggerates the complications and results in low predictive performance. This study proposed the use of sequentially tagged target-independent sentinel tokens to encapsulate multiword instrument aspects for natural language inference model fine-tuning. The encapsulation not only facilitated the attention mechanism to handle an instrument name as a single entity but also enabled the model to handle unseen instruments effectively. Our empirical analysis was based on 5,178 manually annotated instrument–sentiment pairs originated from finance discussion board messages that addressed sentiments of one to four instruments in a single post. The proposed approach consistently outperformed the direct bidirectional encoder representations from transformers (BERT) based approach in terms of recall, precision, and F1-score when handling financial commentaries written in colloquial Cantonese. This study demonstrated the potential benefits of target-independent sentinel token encapsulation for natural language inference. The underlying logic of multiword target-independent encapsulation was expected to hold for other languages, including Chinese, Japanese, and Thai.

Masked graph autoencoder-based multi-agent dynamic relational inference model for trajectory prediction

Masked graph autoencoder-based multi-agent dynamic relational inference model for trajectory prediction

Urban building-level positioning using data-driven algorithms enhanced by spatial variations in sensor features

ABSTRACT As individuals spend most of their time indoors, determining whether a mobile device is located indoors or outdoors – and identifying the specific building it is in – is essential for enabling building-level location-based services and fine-grained human activity analysis. However, existing indoor positioning techniques often rely on dedicated infrastructure or dense signal fingerprinting, limiting their scalability across diverse urban environments. To address this, we propose a lightweight, data-driven framework for building-level mobile device location recognition that integrates indoor/outdoor (I/O) classification and building matching using limited sensor data. A random forest model is trained on a structured, scene-diverse sample library to classify I/O status based on multi-sensor features. For devices identified as indoors, building identification is performed using a Bayesian inference model that incorporates prior knowledge derived from anonymous crowdsourced data, leveraging spatial heterogeneity in sensor feature distributions across candidate buildings. Experiments conducted in three Chinese cities demonstrated that I/O classification achieved over 90% accuracy, and building matching based on crowdsourced data achieved at least 70% precision using only satellite or Wi-Fi features. Our approach requires no infrastructure deployment or extensive labeled data, offering a scalable and practical solution for building-level location inference across large and heterogeneous regions.

Manipulation Test for Multidimensional RDD

ABSTRACTThe causal inference model for the regression discontinuity design (RDD) relies on assumptions that imply the continuity of the density of the assignment (running) variable. The test for this implication is commonly referred to as the manipulation test and is regularly reported in applied research to strengthen the design's validity. The multidimensional RDD (MRDD) extends the RDD to contexts where treatment assignment depends on several running variables. This paper introduces a manipulation test for the MRDD. First, it develops a theoretical model for causal inference with the MRDD, which is used to derive a testable implication on the conditional marginal densities of the running variables. Then, it constructs the test for the implication based on a quadratic form of a vector of statistics separately computed for each marginal density. Finally, the proposed test is compared with alternative procedures commonly employed in applied research.

Cloud-Based Melanoma Detection Using Deep Learning

Abstract: Early detection of melanoma significantly improves patient outcomes, and integrating modern AI with cloud computing offers a powerful solution to this challenge. This paper presents the design and development of a cloud-based melanoma detection system leveraging deep learning models for accurate and real-time diagnosis. The system utilizes convolutional neural networks (CNNs) trained on a large dataset of dermoscopic and clinical images to classify skin lesions. It is deployed on a scalable cloud infrastructure, enabling global access, remote image upload, and rapid inference. A modular APIdriven architecture ensures seamless communication between the frontend, model inference engine, and cloud storage. Emphasis is placed on data security, user authentication, and compliance with medical data privacy standards. Experimental evaluation demonstrates high accuracy and efficiency, supporting its application in telemedicine, dermatology clinics, and remote healthcare settings.

Secure AI Infrastructure: Building Trustworthy AI Systems in Distributed Environments

As enterprises increasingly deploy artificial intelligence to drive customer experiences, business intelligence, and automation, ensuring the security of AI infrastructure has become paramount. Distributed AI systems must not only be scalable and performant they must also be trustworthy, protecting sensitive data and model integrity across dynamic, cloud-native environments. This article explores critical components of secure AI infrastructure, highlighting strategies and technologies for building resilient systems that withstand sophisticated threats. From securing data pipelines with encryption and access controls to protecting model training environments and inference endpoints, a comprehensive defense-in-depth approach addresses the unique security challenges of AI systems. Privacy-preserving techniques like federated learning and differential privacy enable organizations to balance utility with data protection requirements. Proper governance frameworks incorporating model inventories, version control, and ethical considerations establish the foundation for responsible AI deployment. Through practical implementation examples, including a case study from the financial services sector, this article demonstrates how organizations can create AI systems that protect against emerging threats while maintaining operational effectiveness across diverse computing environments.

Scaling machine learning and operations research models for omni-channel retail in the cloud: A framework for real-time decision optimization

This article examines how cloud-native architectures enable retailers to scale machine learning and operations research models across omni-channel environments. It explores the transformation from monolithic on-premise systems to flexible cloud platforms, highlighting how distributed computing frameworks address the computational demands of retail-scale ML model training and inference. The discussion covers architectural patterns for real-time data processing, distributed training techniques, auto-scaling inference architectures, and parallelization strategies for complex optimization problems. The integration of predictive ML insights with prescriptive OR optimization is presented as a critical capability, with various integration patterns examined including sequential, feedback loop, stochastic, and joint learning approaches. Data pipelines connecting predictive and prescriptive models are explored alongside event-driven architectures for cross-channel decision workflows and API design patterns for unified retail intelligence systems. Implementation challenges and technical debt considerations complete the analysis, focusing on both architectural principles and organizational factors that influence successful adoption of cloud-scaled retail analytics

Dental disease detection using deep learning with X-ray

Dental afflictions, in another way investigated and treated in a convenient way, can influence weighty energy issues and a prejudiced quality of history. Conventional plans of disease are very weak on expert reading of dental radiographs or dispassionate figures, that understand expected time-consuming and dependent on something human instability. This paper presents a survey of a mechanical arrangement for dental affliction detection established the use of deep education procedures improved accompanying Generative Adversarial Networks (GANs). Convolutional Neural Network (CNN) is employed for exact categorization of various dental environments like sunken or decayed areas, periodontitis, and gingivitis from intraoral concepts and dental radiographs. To surmount the disadvantage of limited and unstable dossier, GANs are employed to create artificial finest different dental countenances, filling out the preparation dataset and leading to model inference. Additionally, bureaucracy takes advantage of explicable AI plans to visualize and stress the distressed domains, that aid dental experts during dispassionate administrative. Experimental effects story that the bestowed foundation provides extreme accuracy, veracity, and recall, beat normal methods. This research focal points the potential of joining GANs and deep knowledge to offer powerful, ascendable, and evident-time dental affliction disease.

Bayesian inference for nonlinear mixed-effects location scale and interval-censoring cure-survival models: An application to pregnancy miscarriage.

Motivated by a pregnancy miscarriage study, we propose a Bayesian joint model for longitudinal and time-to-event outcomes that takes into account different complexities of the problem. In particular, the longitudinal process is modeled by means of a nonlinear specification with subject-specific error variance. In addition, the exact time of fetal death is unknown, and a subgroup of women is not susceptible to miscarriage. Hence, we model the survival process via a mixture cure model for interval-censored data. Finally, both processes are linked through the subject-specific longitudinal mean and variance. A simulation study is conducted in order to validate our joint model. In the real application, we use individual weighted and Cox-Snell residuals to assess the goodness-of-fit of our proposal versus a joint model that shares only the subject-specific longitudinal mean (standard approach). In addition, the leave-one-out cross-validation criterion is applied to compare the predictive ability of both models.

An Intelligent Intrusion Detection System in IoV Using Machine Learning and Deep Learning Models

ABSTRACTCreating an internet of vehicles (IoV) intelligent intrusion detection system (IDS) that can identify and stop a wide range of new and evolving cyberattacks is a challenging task. IoV generates vast amounts of data from sensors and communication protocols, necessitating an IDS capable of handling various data formats and extracting useful information. Real‐time attack identification and mitigation require efficient model inference with minimal latency. Cyberattackers employ sophisticated techniques, necessitating an IDS capable of learning and adapting to new threats. By oversampling the minority class with SMOTE, SMOTEBoost increases the training data available for learning its features, leading to better detection of minority class instances. The boosting component focuses on training weak learners on misclassified instances, further improving specificity toward the minority class and reducing false positives. The paper describes a way to choose important features for finding intrusions in the IoV that includes nonnegative latent factor dimensionality‐minimizing intraclass compactness (NLF‐DMIC). For classification, this paper proposes a four‐model decision tree (DT), a random forest (RF), an enhanced LSTM network with a conventional long‐short‐term memory. Tested on the CICIDS‐2018 and Car‐Hacking datasets, the suggested approach showed the best intrusion detection performance. The ILSTM outperformed various ML and DL approaches, achieving 98.68% accuracy with SMOTEBOOST on the CICIDS2018 dataset and 98.87% accuracy with SMOTEBOOST on the Car‐Hacking dataset. An intelligent IDS for IoV using machine learning (ML) and deep learning (DL) models enhances recognition accuracy, reduces false alarms, and adapts to dynamic threats, ensuring robust, real‐time security for connected vehicles.

SiamAHG: adaptive hierarchical graph attention for lightweight siamese tracking

SiamAHG: adaptive hierarchical graph attention for lightweight siamese tracking

Bayesian network structure learning by opposition-based learning

As a classical basic model for causal inference, Bayesian networks are of vital importance both in artificial intelligence with uncertainty and interpretability. The significant status of Bayesian networks in these research orientations depends on its topological structure, namely directed acyclic graphs. Bayesian network structure learning is a well-known NP-hard problem, and its computation accuracy is still worth being further studied. In this paper, we propose a new Bayesian network structure learning algorithm, OP-PSO-DE, which combines Particle Swarm Optimization(PSO) and Differential Evolution to search for the optimal structure. Since the computation complexity of BN structure learning increases exponentially with the number of nodes, the proposed algorithm incorporates opposition-based learning to narrow the search space of heuristic algorithms, which can effectively accelerate the searching process. Experimental results show that the proposed algorithm achieves better performances than other state-of-the-art structure learning algorithms when the sample size is 500. The source code of the paper can be found at this link: https://github.com/sunbaodan-hrbeu/paper_code.

Closing the gap in the clinical adoption of computational pathology: a standardized, open-source framework to integrate deep-learning models into the laboratory information system

BackgroundDigital pathology (DP) has revolutionized cancer diagnostics and enabled the development of deep-learning (DL) models aimed at supporting pathologists in their daily work and improving patient care. However, the clinical adoption of such models remains challenging. Here, we describe a proof-of-concept framework that, leveraging Health Level 7 (HL7) standard and open-source DP resources, allows a seamless integration of both publicly available and custom developed DL models in the clinical workflow.MethodsDevelopment and testing of the framework were carried out in a fully digitized Italian pathology department. A Python-based server-client architecture was implemented to interconnect through HL7 messaging the anatomic pathology laboratory information system (AP-LIS) with an external artificial intelligence-based decision support system (AI-DSS) containing 16 pre-trained DL models. Open-source toolboxes for DL model deployment were used to run DL model inference, and QuPath was used to provide an intuitive visualization of model predictions as colored heatmaps.ResultsA default deployment mode runs continuously in the background as each new slide is digitized, choosing the correct DL model(s) on the basis of the tissue type and staining. In addition, pathologists can initiate the analysis on-demand by selecting a specific DL model from the virtual slide tray. In both cases, the AP-LIS transmits an HL7 message to the AI-DSS, which processes the message, runs DL model inference, and creates the appropriate visualization style for the employed classification model. The AI-DSS transmits model inference results to the AP-LIS, where pathologists can visualize the output in QuPath and/or directly as slide description in the virtual slide tray.ConclusionsTaken together, the developed integration framework through the use of the HL7 standard and freely available DP resources offers a standardized, portable, and open-source solution that lays the groundwork for the future widespread adoption of DL models in pathology diagnostics.

Profiling and determining household’s factors influencing crop diversity and animal species inclusion in home gardening, Southeast Nigeria

Home gardening is a game-changer for ensuring food security, benefiting both farming and non-farming households alike. Households that engage in gardening are not just growing crops; they are fostering diverse ecosystems by cultivating multiple plant species and raising animals. However, the key household characteristics that drive this diversity remain underexplored in existing research. This study takes a deep dive into the factors influencing gardening diversity, shedding light on the socio-demographic elements that shape household decisions on food production. By analysing data from 698 home garden households through a self-administered questionnaire, we employed inferential statistical models, including Ordered Probit and Negative Binomial Regression, to unravel the drivers of crop and animal diversity in home gardening. The findings indicate a majority of households fall into the medium (52.4%) and high (41.1%) categories of crop diversity. Key determinants of crop diversity include household location, size, garden land availability, production season, and the participation of family labour. Meanwhile, factors such as home-ownership and gardening experience play a crucial role in the decision to rear animals. These insights carry significant implications. To integrate both crop cultivation and animal rearing successfully, households must possess specific characteristics beyond simply growing food crops. The research underscores that while all gardening households engage in crop production, not all extend their efforts to animal husbandry. Crucially, socio-economic conditions shape household decisions on what and how to produce. By understanding and addressing these conditions, policymakers, agricultural experts, and communities can enhance home gardening diversity, particularly in the Global South boosting both nutrition and resilience against food insecurity. Investing in home gardening isn’t just about growing food; it’s about cultivating a more sustainable, diverse, and food-secure future.

Exploring treatment effects and fluid resuscitation strategies in septic shock: a deep learning-based causal inference approach

Septic shock exhibits diverse etiologies and patient characteristics, necessitating tailored fluid management. We aimed to compare resuscitation strategies using normal saline, Ringer’s lactate, and albumin, and to determine which patient factors are associated with improved outcomes. We analyzed septic shock patients from the MIMIC-IV database, categorizing them by the fluid administered: normal saline, Ringer’s lactate, albumin, or their combinations. A deep learning-based causal inference model estimated treatment effects on in-hospital mortality and kidney outcomes (defined as a doubling of creatinine or the initiation of kidney replacement therapy). Multivariable logistic regression was then applied to the individual treatment effects to identify patient characteristics linked to better outcomes for Ringer’s lactate and additional albumin infusion compared to normal saline alone. Among 13,527 patients, 17.8% experienced in-hospital mortality and 16.2% developed kidney injury. Ringer’s lactate reduced mortality by 2.33% and kidney injury by 1.41% compared to normal saline. Adding albumin to normal saline further reduced mortality by 1.20% and kidney outcomes by 0.71%. The combination of Ringer’s lactate and albumin provided the greatest benefit (mortality: −3.07%, kidney injury: −3.00%). Patients with high SOFA scores, low albumin, or high lactate levels benefited more from normal saline, whereas those with low eGFR or on vasopressors were less likely to benefit from albumin. Ringer’s lactate, particularly when combined with albumin, is superior to normal saline in reducing mortality and kidney injury in septic shock patients, underscoring the need for personalized fluid management based on patient-specific factors.

Effective Multitasking Segmentation Using Vision Transformer Network for Malformed Images

Background: Effectively and efficiently segmenting human body areas in malformed images is of tremendous practical importance. Recent techniques typically utilize transformers to extract discriminative functions. However, the worldwide interest in these transformers, regular consequences due to the lack of image feature details, and high computational costs result in suboptimal segmentation accuracy and low inference speeds. Objective: To obtain better human segmentation by enhancing model inference speed, this paper applies the Human Spatial Prior Module (HSPM) and the Dynamic Token Pruning Module (DTPM) to deal with real-time three-dimensional environments. Materials and Methods: In this paper, we have applied the Human Spatial Prior Module (HSPM) and the Dynamic Token Pruning Module (DTPM), designed to seize human-unique features in malformed images by dynamically extracting complex information. Meanwhile, the DTPM enhances the inference pace by pruning the point tokens at each Vision Transformer layer (VT). Unlike conventional cropping techniques, the pruned tokens are retained through function maps and selectively reactivated in subsequent layers to enhance model performance. Results: To show the effectiveness of our Vision Transformer for Malformed Images (VT-MI), we expand the dataset with an input image size of 512 x 512 and conduct behavior experiments on this prolonged dataset as well as on the COCO, MPII Human Pose, and a few real-time captured image datasets. Our technique achieves a mIoU growth of 86.5 and an FPS improvement of 7.9 on the COCO dataset with 93.21% accuracy, together with mIoU (person) of 86.11 with dynamic token pruning and reactivation on the entire dataset, while additionally decreasing the model size by using approximately 120 GFLOPs. Conclusion: This work contributes deeper insights into vision transformer networks and provides a clear direction towards both image classification and dense prediction tasks. Moreover, extensive experiments help in showcasing comparative analysis that benchmarks the frameworks contributing to image analysis advancements.