Artificial Intelligence-based Approach Research Articles

An artificial intelligence-based approach for identifying rare disease patients using retrospective electronic health records applied for Pompe disease.

We retrospectively screened 350,116 electronic health records (EHRs) to identify suspected patients for Pompe disease. Using these suspected patients, we then describe their phenotypical characteristics and estimate the prevalence in the respective population covered by the EHRs. We applied Symptoma's Artificial Intelligence-based approach for identifying rare disease patients to retrospective anonymized EHRs provided by the "University Hospital Salzburg" clinic group. Within 1 month, the AI screened 350,116 EHRs reaching back 15 years from five hospitals, and 104 patients were flagged as probable for Pompe disease. Flagged patients were manually reviewed and assessed by generalist and specialist physicians for their likelihood for Pompe disease, from which the performance of the algorithms was evaluated. Of the 104 patients flagged by the algorithms, generalist physicians found five "diagnosed," 10 "suspected," and seven patients with "reduced suspicion." After feedback from Pompe disease specialist physicians, 19 patients remained clinically plausible for Pompe disease, resulting in a specificity of 18.27% for the AI. Estimating from the remaining plausible patients, the prevalence of Pompe disease for the greater Salzburg region [incl. Bavaria (Germany), Styria (Austria), and Upper Austria (Austria)] was one in every 18,427 people. Phenotypes for patient cohorts with an approximated onset of symptoms above or below 1 year of age were established, which correspond to infantile-onset Pompe disease (IOPD) and late-onset Pompe disease (LOPD), respectively. Our study shows the feasibility of Symptoma's AI-based approach for identifying rare disease patients using retrospective EHRs. Via the algorithm's screening of an entire EHR population, a physician had only to manually review 5.47 patients on average to find one suspected candidate. This efficiency is crucial as Pompe disease, while rare, is a progressively debilitating but treatable neuromuscular disease. As such, we demonstrated both the efficiency of the approach and the potential of a scalable solution to the systematic identification of rare disease patients. Thus, similar implementation of this methodology should be encouraged to improve care for all rare disease patients.

Computed Tomography Assessment of Coronary Atherosclerosis: From Threshold-Based Evaluation to Histologically Validated Plaque Quantification.

Arterial plaque rupture and thrombosis is the primary cause of major cardiovascular and neurovascular events. The identification of atherosclerosis, especially high-risk plaques, is therefore crucial to identify high-risk patients and to implement preventive therapies. Computed tomography angiography has the ability to visualize and characterize vascular plaques. The standard methods for plaque evaluation rely on the assessment of plaque burden, stenosis severity, the presence of positive remodeling, napkin ring sign, and spotty calcification, as well as Hounsfield Unit (HU)-based thresholding for plaque quantification; the latter with multiple shortcomings. Semiautomated threshold-based segmentation techniques with predefined HU ranges identify and quantify limited plaque characteristics, such as low attenuation, non-calcified, and calcified plaque components. Contrary to HU-based thresholds, histologically validated plaque characterization, and quantification, an emerging Artificial intelligence-based approach has the ability to differentiate specific tissue types based on a biological correlate, such as lipid-rich necrotic core and intraplaque hemorrhage that determine plaque vulnerability. In this article, we review the relevance of plaque characterization and quantification and discuss the benefits and limitations of the currently available plaque assessment and classification techniques.

141P Unveiling the likely pharmacological mechanisms of brigatinib on brain metastasis in ALK+ patients with non-small cell lung cancer: A systems biology and artificial intelligence-based approach

141P Unveiling the likely pharmacological mechanisms of brigatinib on brain metastasis in ALK+ patients with non-small cell lung cancer: A systems biology and artificial intelligence-based approach

Development of a deep stacked ensemble with process based volatile memory forensics for platform independent malware detection and classification

Malware has become more complicated in its purpose and abilities over time, demanding continuous progress in detection and defense technologies. Malware designers use anti-analysis obfuscation techniques, including packing and encryption, to evade detection and hinder the analysis process. Current malware detection methods have shortcomings; thus, an alternative dynamic platform-independent scheme is proposed to extract harmful hardware impressions. This scheme includes extracting and converting a file from process memory dumps into an image. A combined structural and statistical image textural analysis is performed by designing a hybrid local and global feature descriptor. The hybrid feature descriptor helps to improve the data training ability of the proposed deep-stacked ensemble model by reducing input dimensions. A deep-stacked ensemble model is developed by combining prediction outputs from weak learners (CNNs) and feeding them into a meta-learner (MLP) as learning input. An explainable artificial intelligence-based approach is employed to interpret and validate the final results of the proposed scheme. Evaluations are conducted using three datasets: the publicly available Dumpware10 dataset, which contains 3686 samples from 10 different malware families; the publicly available CIC-MalMem-2022 dataset, which includes 2,916 samples from 15 different obfuscated malware families; and a real-world dataset, which contains 2375 samples of both malware and benign android apps. Experimental outcomes show that the proposed scheme achieved 99.1 % accuracy in analyzing Windows malware memory dumps, 94.3 % accuracy in analyzing Android malware memory dumps, and 99.8 % accuracy in analyzing Windows obfuscated malware memory dumps. The final results indicate that our vision-based system provides an excellent defense against malicious programs.

PRD-MADDPG: An efficient learning-based algorithm for orbital pursuit-evasion game with impulsive maneuvers

This paper comprehensively investigates the problem of impulsive orbital pursuit-evasion games (OPEGs) by using an artificial intelligence-based approach. First, the mathematical model for the impulsive OPEGs in which the pursuer and evader both perform their orbital maneuvers by imposing the impulsive velocity increments is constructed. Second, the problem of impulsive OPEGs is transformed into a bilateral optimization problem with a minimum–maximum optimization index in terms of terminal time and multiple constraints such as maneuverability, total fuel consumption, and mission time, etc. To determine the optimal impulsive maneuvers for both sides, a PRD-MADDPG (Predict-Reward-Detect Multi-Agent Deep Deterministic Policy Gradient) algorithm in the frame of multi-agent reinforcement learning is designed. This novel algorithm uses the basic MADDPG to achieve the strategies training and learning, and applies the supplemental PRD to predict the change of game state during the interval between two adjacent impulsive maneuvers and incorporate these information into the algorithm training in the form of predicted reward. Finally, some pursuit-evasion missions near the Geosynchronous Earth Orbit are numerically analyzed to verify the validness and effectiveness of the algorithm. The results prove that the PRD-MADDPG algorithm is very efficient to find applicable strategies even considering rather complex constraints. It is also shown that the learning-based strategies can be effectively applied in the extended scenarios which are not seen in the training process.

The Design of Convolutional Neural Networks Model for Classification of Ear Diseases on Android Mobile Devices

An otorhinolaryngologist (ORL) or general practitioner diagnoses ear disease based on ear image information. However, general practitioners refer patients to ORL for chronic ear disease because the image of ear disease has high complexity, variety, and little difference between diseases. An artificial intelligence-based approach is needed to make it easier for doctors to diagnose ear diseases based on ear image information, such as the Convolutional Neural Network (CNN). This paper describes how CNN was designed to generate CNN models used to classify ear diseases. The model was developed using an ear image dataset from the practice of an ORL at the University of Mataram Teaching Hospital. This work aims to find the best CNN model for classifying ear diseases applicable to android mobile devices. Furthermore, the best CNN model is deployed for an Android-based application integrated with the Endoscope Ear Cleaning Tool Kit for registering patient ear images. The experimental results show 83% accuracy, 86% precision, 86% recall, and 4ms inference time. The application produces a System Usability Scale of 76.88% for testing, which shows it is easy to use. This achievement shows that the model can be developed and integrated into an ENT expert system. In the future, the ENT expert system can be operated by workers in community health centres/clinics to assist leading health them in diagnosing ENT diseases early.

Bioinformatics investigation on blood-based gene expressions of Alzheimer's disease revealed ORAI2 gene biomarker susceptibility: An explainable artificial intelligence-based approach

The progressive, chronic nature of Alzheimer's disease (AD), a form of dementia, defaces the adulthood of elderly individuals. The pathogenesis of the condition is primarily unascertained, turning the treatment efficacy more arduous. Therefore, understanding the genetic etiology of AD is essential to identifying targeted therapeutics. This study aimed to use machine-learning techniques of expressed genes in patients with AD to identify potential biomarkers that can be used for future therapy. The dataset is accessed from the Gene Expression Omnibus (GEO) database (Accession Number: GSE36980). The subgroups (AD blood samples from frontal, hippocampal, and temporal regions) are individually investigated against non-AD models. Prioritized gene cluster analyses are conducted with the STRING database. The candidate gene biomarkers were trained with various supervised machine-learning (ML) classification algorithms. The interpretation of the model prediction is perpetrated with explainable artificial intelligence (AI) techniques. This experiment revealed 34, 60, and 28 genes as target biomarkers of AD mapped from the frontal, hippocampal, and temporal regions. It is identified ORAI2 as a shared biomarker in all three areas strongly associated with AD's progression. The pathway analysis showed that STIM1 and TRPC3 are strongly associated with ORAI2. We found three hub genes, TPI1, STIM1, and TRPC3, in the network of the ORAI2 gene that might be involved in the molecular pathogenesis of AD. Naive Bayes classified the samples of different groups by fivefold cross-validation with 100% accuracy. AI and ML are promising tools in identifying disease-associated genes that will advance the field of targeted therapeutics against genetic diseases.

An Artificial Intelligence-Based Approach to Assess and Classify University Students' Attendance in Live Sessions

An Artificial Intelligence-Based Approach to Assess and Classify University Students' Attendance in Live Sessions

Tuning OER Electrocatalysts toward LOM Pathway through the Lens of Multi-Descriptor Feature Selection by Artificial Intelligence-Based Approach

From the thermodynamic and kinetic viewpoints, the oxygen evolution reaction (OER) is central to the production of hydrogen through electrocatalytic water splitting process. As a result, extensive research is carried out with an aim to develop (preferably low cost) electrocatalysts for efficient OER reaction. Two main mechanisms, the lattice oxygen mediated mechanism (LOM) and the adsorbate evolution mechanism (AEM), are proposed for the OER process, but the former seems to be favored as the latter is surrounded by a number of discrepancies. The LOM is based on the oxidation and reduction chemistry of lattice oxygen anions. Both from fundamental and application perspectives, identification and selection of crucial electronic and structural features to rationally tune the LOM for OER process and determine the rate of the OER reaction are essential but remain a challenge. As yet, this has largely been attempted by trial and error (synthesis and performance evaluation) approaches and, hence, has been tied up by tediousness and inefficiency. Considering the availability of large chemical space and huge probabilities in fine-tuning of electronic and structural attributes, the use of artificial intelligence (AI) can be efficacious. To predict promising material features that enable rational design of OER electrocatalysts, electrocatalytic performance maps have been developed based on reactivity multiple descriptors. The reactivity multiple descriptors are numerous and are classified into several families. Therefore, modern materials informatics technique is urgently required to enable large-scale data mining to rapidly screen and select the best physical features from multiple descriptors for a rational design of LOM-induced OER electrocatalysts. In this context, feature selection by artificial intelligence-based approach can be a solution to such challenge. Yet to the best of our knowledge, no attempt has been made before to tune electrocatalysts for the OER through LOM reaction mechanism following feature selection by artificial intelligence-based approach.

Forecasting green roofs’ potential in improving building thermal performance and mitigating urban heat island in the Mediterranean area: An artificial intelligence-based approach

Green roofs are widely used in hot or cold climates mainly because they are capable to improve the energy efficiency of buildings and, when implemented at a large scale, reducing air pollution and the urban heat island effect (UHI) in urban contexts.Artificial Neural Network (ANN) black-box algorithms are a valid alternative to studying complex systems. However, the literature highlights - quite surprisingly – none of the available research refers to coupling ANNs and green roofs in the Mediterranean area, where green roofs are instead considered one of the most suitable technologies to reduce the high cooling demand.Therefore, the objective of this research work is to create and validate an ANN for the prediction of the monthly green roof’s internal and external surface temperatures and the monthly internal air temperature, starting from different green roof parameters and climatic variables. Specifically, the ANN was created with reference to a Mediterranean climate considering an existing green roof on a building of the University of Palermo characterized by a cooling demand predominance; 180 green roof configurations, obtained by varying the characteristic parameters of vegetation (plant height, leaf area index and leaf reflectivity) and the substrate thickness and thermophysical properties (lightweight and heavyweight), were dynamically simulated on an hourly basis to build the training dataset. In addition, other 72 green roof configurations were simulated to generate the dataset for the validation purpose of the ANN accuracy. The optimal ANN-related architecture consists of 90 neurons with one hidden layer and guarantees very high accuracy predictions.The outcomes of this research represent a useful tool to determine the thermal response of green roofs and their impact on the building energy demand and indoor thermal comfort and UHI mitigation.

Artificial intelligence-based approach for cluster identification in a CFB riser

Cluster evolution dominates the gas–solid interaction and heat and mass transfer in risers, but cluster identification is still a challenge. Based on the unsupervised artificial intelligence approach, a two-step K-means cluster identification method is proposed to distinguish the clusters in a riser. With the consideration of the instantaneous particle motion and position information, it is able to capture the phenomenon of particles attaching and detaching a cluster. The estimated solid fraction threshold is located in the region where the slope of the solid fraction is the smallest, and approximately makes the numbers of particles detaching and attaching the clusters equal. The solid fraction threshold is about 0.204. It is found that the particles penetrating a cluster are in a more violent turbulent movement. Larger gas velocity and solid flux result in stronger interactions between clusters. The proposed method provides a new way to identify the cluster.

Automated brain tumour segmentation from multi-modality magnetic resonance imaging data based on new particle swarm optimisation segmentation method.

Segmentation of brain tumours is a complex problem in medical image processing and analysis. It is a time-consuming and error-prone task. Therefore, computer-aided detection systems need to be developed to decrease physicians' workload and improve the accuracy of segmentation. This paper proposes a level set method constrained by an intuitive artificial intelligence-based approach to perform brain tumour segmentation. By studying 3D brain tumour images, a new segmentation technique based on the Modified Particle Swarm Optimisation (MPSO), Darwin Particle Swarm Optimisation (DPSO), and Fractional Order Darwinian Particle Swarm Optimisation (FODPSO) algorithms were developed. The introduced technique was verified according to the MICCAI RASTS 2013 database for high-grade glioma patients. The three algorithms were evaluated using different performance measures: accuracy, sensitivity, specificity, and Dice similarity coefficient to prove the performance and robustness of our 3D segmentation technique. The result is that the MPSO algorithm consistently outperforms the DPSO and FO DPSO.

PHSPP02 Presentation Time: 2:40 PM: Automated Channel Identification for Efficient Brachytherapy Treatment Planning

<h3>Purpose</h3> For interstitial brachytherapy, channel reconstruction during the process of brachytherapy treatment planning can be an arduous and time-consuming task, which becomes more difficult and error-prone when 20 to 30 different catheters with curvatures are involved and overlap in 3D in the patient anatomy. The goal of this project is to develop an artificial intelligence-based approach for automated channel identification using patient CT scans, reducing both the time and potential for human error in brachytherapy treatment planning. <h3>Materials and Methods</h3> Three sets of CT scans were used as well as a baseline model for training, which were acquired with a set of binary-coded "dummy wires" markers <i>in situ</i>. Each of the 12 available dummy wires was designed with a unique pattern of radio-opaque markers and/or spacing strands which was mapped to a channel identification array. An identification algorithm was developed which involved first manually specifying the 3D DICOM coordinates of the centroids of each of the marker strands within the dummy wires. A nearest neighbor search was performed using a k-d tree method for each centroid acting as a potential starting point within a channel. Using the knowledge from the baseline model of each of the 12 potential dummy wire markers, the system predicted the most probable trajectory through the centroids. Then, to assign the proper channel number, the identified pattern of the applicable neighboring points was correlated with the channel identification array defined previously. This process was repeated until the entire configuration of channel identities was determined. Finally, a process of elimination was used to output the correct channel labels and locations. <h3>Results</h3> In this feasibility study, the developed algorithm was able to suitably provide channel identities and return the locations to the user when the proper centroid locations were input into the model. The model ran within seconds and required no training data other than the layout and spacing of the strands within the coded dummy-wire marker system. The final output was able to be graphically represented as labeled reconstructions of the catheters in a 3D space. <h3>Conclusions</h3> The outcomes of this preliminary work suggested that there is a future for automated aid in the channel identification process based on coded dummy-wire markers for more efficient treatment planning in brachytherapy. Using only the centroids of the markers, the channels could be accurately reconstructed. Future work will involve creating a 3D CT scan-based image detection algorithm to automate the process of defining the centroid locations without any <i>a priori</i> knowledge.

An Artificial Intelligence-Based Approach to Social Data-Aware Optimization for Enterprise Management

Enterprise management has always been a hot issue in society. In today’s society, enterprises are no longer individuals cut off from society and no longer have profit as their sole purpose, but need to exist and develop in combination with social information. With the increasing competition among enterprises, the original enterprise management methods can no longer meet the needs of sustainable development of enterprises, nor can they effectively utilize social data. To better understand the daily emotions of corporate workers, we use a multimodal emotion recognition method in this paper. Multimodal emotion recognition refers to the recognition of human emotional states through different modal information such as speech, visual, and text related to human emotional expressions, which has important research significance in the fields of human-computer interaction, artificial intelligence, and emotional computing and has received much attention from researchers. Given the great success of deep learning methods developed in recent years for various tasks, various deep neural networks are now used to learn high-level representations of emotional features for multimodal emotion recognition. The analysis of employee sentiment is complemented by traditional management methods that make full use of social data. In this paper, based on the study of a single enterprise management model, the proposed model contains five substructure modules, starting with feature inputs, extracting features at different levels through three convolutional modules and outputting recognition results through a softmax classifier. The focus is on how to utilize social data, while combining deep learning with traditional enterprise management methods to fill the research gap in this area in academia.

A Novel Artificial Intelligence-Based Approach for Quantitative Assessment of Angiogenesis in the Ex Ovo CAM Model.

Simple SummaryAngiogenesis is crucial in tissue regeneration and a relevant factor in tumor growth. Consequently, there is a demand for methods to assess angiogenesis. The chorioallantoic membrane (CAM) assay is a widely used in vivo model for the study of angiogenesis. However, there is no established gold standard to evaluate the vascularisation of the CAM, which limits the comparability of the different studies. In this manuscript, we present a novel approach to address this topic.Angiogenesis is a highly regulated process. It promotes tissue regeneration and contributes to tumor growth. Existing therapeutic concepts interfere with different steps of angiogenesis. The quantification of the vasculature is of crucial importance for research on angiogenetic effects. The chorioallantoic membrane (CAM) assay is widely used in the study of angiogenesis. Ex ovo cultured chick embryos develop an easily accessible, highly vascularised membrane on the surface. Tumor xenografts can be incubated on this membrane enabling studies on cancer angiogenesis and other major hallmarks. However, there is no commonly accepted gold standard for the quantification of the vasculature of the CAM. We compared four widely used measurement techniques to identify the most appropriate one for the quantification of the vascular network of the CAM. The comparison of the different quantification methods suggested that the CAM assay application on the IKOSA platform is the most suitable image analysis application for the vasculature of the CAM. The new CAM application on the IKOSA platform turned out to be a reliable and feasible tool for practical use in angiogenesis research. This novel image analysis software enables a deeper exploration of various aspects of angiogenesis and might support future research on new anti-angiogenic strategies for cancer treatment.

Using artificial intelligence to avoid human error in identifying embryos: a retrospective cohort study.

To determine whether convolutional neural networks (CNN) can be used to accurately ascertain the patient identity (ID) of cleavage and blastocyst stage embryos based on image data alone. A CNN model was trained and validated over three replicates on a retrospective cohort of 4889 time-lapse embryo images. The algorithm processed embryo images for each patient and produced a unique identification key that was associated with the patient ID at a timepoint on day 3 (~ 65hours post-insemination (hpi)) and day 5 (~ 105 hpi) forming our data library. When the algorithm evaluated embryos at a later timepoint on day 3 (~ 70 hpi) and day 5 (~ 110 hpi), it generates another key that was matched with the patient's unique key available in the library. This approach was tested using 400 patient embryo cohorts on day 3 and day 5 and number of correct embryo identifications with the CNN algorithm was measured. CNN technology matched the patient identification within random pools of 8 patient embryo cohorts on day 3 with 100% accuracy (n = 400 patients; 3 replicates). For day 5 embryo cohorts, the accuracy within random pools of 8 patients was 100% (n = 400 patients; 3 replicates). This study describes an artificial intelligence-based approach for embryo identification. This technology offers a robust witnessing step based on unique morphological features of each embryo. This technology can be integrated with existing imaging systems and laboratory protocols to improve specimen tracking.

Scour depth prediction at bridge piers using metaheuristics-optimized stacking system

This paper proposes a novel artificial intelligence-based approach, called the metaheuristics-optimized stacking system (MOSS), to assist civil engineers in estimating scour depth at bridge piers. MOSS integrates the advantages of hybrid and stacking ensemble schemes by combining a metaheuristic optimization algorithm with efficient machine learning (ML) models. The metaheuristic algorithm simultaneously finds the optimal values of all hyperparameters of constituent ML models in the stacking ensemble to generate the most effective system. The developed MOSS was established by integrating a forensic-based investigation (FBI) algorithm, least squares support vector regression (LSSVR), and radial basis function neural network (RBFNN). The efficiency of the MOSS is verified comprehensively with reference to three case studies of both laboratory data and field data that cover various levels of complexity and types of pier foundations. The performance of MOSS is compared to that of other single ML models, conventional voting ensemble, hybrid models, empirical methods, and mathematical approach. The analytical results of cross-validation reveal that MOSS was the most reliable approach, achieving the best values of all performance evaluation metrics. MOSS exhibited mean absolute percentage errors of 7.127%, 29.195% and 13.131% in the prediction of scour depth using a laboratory dataset, a field dataset, and a complex pier foundations dataset, respectively, and these values are at least 36%, 19% and 41% lower than those obtained using other approaches. The automated predictive analytics revealed the robustness, efficiency, and stability of MOSS. The novelty and contributions of this study include (1) providing a general stacking framework and re-defining the establishment of an ensemble ML model, and (2) developing a highly promising tool that greatly outperforms currently available methods to help civil engineers estimate the scour depth around bridge piers.

Accurate thermal prediction model for building-integrated photovoltaics systems using guided artificial intelligence algorithms

Progress in development of building-integrated photovoltaic systems is still hindered by the complexity of the physics and materials properties of the photovoltaic (PV) modules and its effect on the thermal behavior of the building. This affects not only the energy generation, as its active function and linked to economic feasibility, but also the thermal insulation of the building as part of the structure’s skin. Traditional modeling methods currently presents limitations, including the fact that they do not account for material thermal inertia and that the proposed semi-empirical coefficients do not define all types of technologies, mounting configuration, or climatic conditions. This article presents an artificial intelligence-based approach for predicting the temperature of a poly-crystalline silicon PV module based on local outdoor weather conditions (ambient temperature, solar irradiation, relative outdoor humidity and wind speed) and indoor comfort parameters (indoor temperature and indoor relative humidity) as inputs. A combination of two algorithms (Grammatical Evolution and Differential Evolution) guides to the creation of a customized expression based on the Sandia model. Different data-sets for a fully integrated PV system were tested to demonstrate its performance on three different types of days: sunny, cloudy and diffuse, showing relative errors of less than 4% in all cases and including night time. In comparison to Sandia model, this method reduces the error by up to 11% in conditions of variability of sky over short time intervals (cloudy days).

Artificial Intelligence-Based Approach for Misogyny and Sarcasm Detection from Arabic Texts.

Social media networking is a prominent topic in real life, particularly at the current moment. The impact of comments has been investigated in several studies. Twitter, Facebook, and Instagram are just a few of the social media networks that are used to broadcast different news worldwide. In this paper, a comprehensive AI-based study is presented to automatically detect the Arabic text misogyny and sarcasm in binary and multiclass scenarios. The key of the proposed AI approach is to distinguish various topics of misogyny and sarcasm from Arabic tweets in social media networks. A comprehensive study is achieved for detecting both misogyny and sarcasm via adopting seven state-of-the-art NLP classifiers: ARABERT, PAC, LRC, RFC, LSVC, DTC, and KNNC. To fine tune, validate, and evaluate all of these techniques, two Arabic tweets datasets (i.e., misogyny and Abu Farah datasets) are used. For the experimental study, two scenarios are proposed for each case study (misogyny or sarcasm): binary and multiclass problems. For misogyny detection, the best accuracy is achieved using the AraBERT classifier with 91.0% for binary classification scenario and 89.0% for the multiclass scenario. For sarcasm detection, the best accuracy is achieved using the AraBERT as well with 88% for binary classification scenario and 77.0% for the multiclass scenario. The proposed method appears to be effective in detecting misogyny and sarcasm in social media platforms with suggesting AraBERT as a superior state-of-the-art deep learning classifier.

Determining Carina and Clavicular Distance-Dependent Positioning of Endotracheal Tube in Critically Ill Patients: An Artificial Intelligence-Based Approach.

Early and accurate prediction of endotracheal tube (ETT) location is pivotal for critically ill patients. Automatic and timely detection of faulty ETT locations from chest X-ray images may avert patients' morbidity and mortality. Therefore, we designed convolutional neural network (CNN)-based algorithms to evaluate ETT position appropriateness relative to four detected key points, including tracheal tube end, carina, and left/right clavicular heads on chest radiographs. We estimated distances from the tube end to tracheal carina and the midpoint of clavicular heads. A DenseNet121 encoder transformed images into embedding features, and a CNN-based decoder generated the probability distributions. Based on four sets of tube-to-carina distance-dependent parameters (i.e., (i) 30-70 mm, (ii) 30-60 mm, (iii) 20-60 mm, and (iv) 20-55 mm), corresponding models were generated, and their accuracy was evaluated through the predicted L1 distance to ground-truth coordinates. Based on tube-to-carina and tube-to-clavicle distances, the highest sensitivity, and specificity of 92.85% and 84.62% respectively, were revealed for 20-55 mm. This implies that tube-to-carina distance between 20 and 55 mm is optimal for an AI-based key point appropriateness detection system and is empirically comparable to physicians' consensus.