ML-based Approach Research Articles

Comparative performance and machine learning-based optimization of TSA configurations for NH3 removal from NH3 cracking gas

To meet the H2 standard used in fuel cells, removing ammonia (NH3) from NH3 cracking gas to less than 0.1 ppm still faces technical challenges. In this study, novel configurations of a three-bed temperature swing adsorption (TSA) process using zeolite 4A and machine learning (ML)-based optimization were developed for effective NH3 removal. Three TSA configurations, consisting of a cooling gas, three-bed TSA, heat exchangers, an expander, and/or a compressor, were designed and evaluated using the TSA model with zeolite 4A pellets after the model validation with a reference. In a techno-economic comparison, the configuration using H2 pressure swing adsorption (PSA) tail gas as the cooling gas and NH3 TSA off-gas from the cooling step as the heating gas for energy recovery (TSA-TGER) performed the best. Dynamic behavior and sensitivity analyses were conducted to elucidate the characteristics of the TSA-TGER configuration. Using five main operating variables selected from the Pearson correlation method, the developed artificial neural network model could precisely predict the results with a reduction in computational cost of 1800 times compared with the process simulation. Finally, at the optimum condition found from the ML-based optimization, the TSA-TGER configuration consumed 2174.8 MJ/tNH3 and 162.33 $/tNH3 to produce an H2 mixture with less than 0.1 ppm NH3, indicating that the NH3 removal cost contributed to only approximately 0.98% of the referred H2 production cost (3580 $/tH2) via the NH3-to-H2 process. These results provide guidelines for designing an effective NH3 removal configuration from NH3 cracking gases. The proposed ML-based optimization approach can also be applied to other purification processes.

Machine learning-based prediction of in-hospital death for patients with takotsubo syndrome: The InterTAK-ML model.

Takotsubo syndrome (TTS) is associated with a substantial rate of adverse events. We sought to design a machine learning (ML)-based model to predict the risk of in-hospital death and to perform a clustering of TTS patients to identify different risk profiles. A ridge logistic regression-based ML model for predicting in-hospital death was developed on 3482 TTS patients from the International Takotsubo (InterTAK) Registry, randomly split in a train and an internal validation cohort (75% and 25% of the sample size, respectively) and evaluated in an external validation cohort (1037 patients). Thirty-one clinically relevant variables were included in the prediction model. Model performance represented the primary endpoint and was assessed according to area under the curve (AUC), sensitivity and specificity. As secondary endpoint, a K-medoids clustering algorithm was designed to stratify patients into phenotypic groups based on the 10 most relevant features emerging from the main model. The overall incidence of in-hospital death was 5.2%. The InterTAK-ML model showed an AUC of 0.89 (0.85-0.92), a sensitivity of 0.85 (0.78-0.95) and a specificity of 0.76 (0.74-0.79) in the internal validation cohort and an AUC of 0.82 (0.73-0.91), a sensitivity of 0.74 (0.61-0.87) and a specificity of 0.79 (0.77-0.81) in the external cohort for in-hospital death prediction. By exploiting the 10 variables showing the highest feature importance, TTS patients were clustered into six groups associated with different risks of in-hospital death (28.8% vs. 15.5% vs. 5.4% vs. 1.0.8% vs. 0.5%) which were consistent also in the external cohort. A ML-based approach for the identification of TTS patients at risk of adverse short-term prognosis is feasible and effective. The InterTAK-ML model showed unprecedented discriminative capability for the prediction of in-hospital death.

Combining electromyographic and electrical impedance data sets through machine learning: A study in D2-mdx and wild-type mice.

Needle impedance-electromyography (iEMG) assesses the active and passive electrical properties of muscles concurrently by using a novel needle with six electrodes, two for EMG and four for electrical impedance myography (EIM). Here, we assessed an approach for combining multifrequency EMG and EIM data via machine learning (ML) to discriminate D2-mdx muscular dystrophy and wild-type (WT) mouse skeletal muscle. iEMG data were obtained from quadriceps of D2-mdx mice, a muscular dystrophy model, and WT animals. EIM data were collected with the animals under deep anesthesia and EMG data collected under light anesthesia, allowing for limited spontaneous movement. Fourier transformation was performed on the EMG data to provide power spectra that were sampled across the frequency range using three different approaches. Random forest-based, nested ML was applied to the EIM and EMG data sets separately and then together to assess healthy versus disease category classification using a nested cross-validation procedure. Data from 20 D2-mdx and 20 WT limbs were analyzed. EIM data fared better than EMG data in differentiating healthy from disease mice with 93.1% versus 75.6% accuracy, respectively. Combining EIM and EMG data sets yielded similar performance as EIM data alone with 92.2% accuracy. We have demonstrated an ML-based approach for combining EIM and EMG data obtained with an iEMG needle. While EIM-EMG in combination fared no better than EIM alone with this data set, the approach used here demonstrates a novel method of combining the two techniques to characterize the full electrical properties of skeletal muscle.

Quantifying the Efficacy of Magnetic Nanoparticles for MRI and Hyperthermia Applications via Machine Learning Methods.

Magnetic nanoparticles are a prospective class of materials for use in biomedicine as agents for magnetic resonance imagining (MRI) and hyperthermia treatment. However, synthesis of nanoparticles with high efficacy is resource-intensive experimental work. In turn, the use of machine learning (ML) methods is becoming useful in materials design and serves as a great approach to designing nanomagnets for biomedicine. In this work, for the first time, an ML-based approach is developed for the prediction of main parameters of material efficacy, i.e., specific absorption rate (SAR) for hyperthermia and r1 /r2 relaxivities in MRI, with parameters of nanoparticles as well as experimental conditions as descriptors. For that, a unique database with more than 980 magnetic nanoparticles collected from scientific articles is assembled. Using this data, several tree-based ensemble models are trained to predict SAR, r1 and r2 relaxivity. After hyperparameter optimization, models reach performances of R2 = 0.86, R2 = 0.78, and R2 = 0.75, respectively. Testing the models on samples unseen during the training shows no performance drops. Finally, DiMag, an open access resource created to guide synthesis of novel nanosized magnets for MRI and hyperthermia treatment with machine learning and boost development of new biomedical agents, is developed.

Spatial attention-based residual network for human burn identification and classification

Diagnosing burns in humans has become critical, as early identification can save lives. The manual process of burn diagnosis is time-consuming and complex, even for experienced doctors. Machine learning (ML) and deep convolutional neural network (CNN) models have emerged as the standard for medical image diagnosis. The ML-based approach typically requires handcrafted features for training, which may result in suboptimal performance. Conversely, DL-based methods automatically extract features, but designing a robust model is challenging. Additionally, shallow DL methods lack long-range feature dependency, decreasing efficiency in various applications. We implemented several deep CNN models, ResNeXt, VGG16, and AlexNet, for human burn diagnosis. The results obtained from these models were found to be less reliable since shallow deep CNN models need improved attention modules to preserve the feature dependencies. Therefore, in the proposed study, the feature map is divided into several categories, and the channel dependencies between any two channel mappings within a given class are highlighted. A spatial attention map is built by considering the links between features and their locations. Our attention-based model BuRnGANeXt50 kernel and convolutional layers are also optimized for human burn diagnosis. The earlier study classified the burn based on depth of graft and non-graft. We first classified the burn based on the degree. Subsequently, it is classified into graft and non-graft. Furthermore, the proposed model performance is evaluated on Burns_BIP_US_database. The sensitivity of the BuRnGANeXt50 is 97.22% and 99.14%, respectively, for classifying burns based on degree and depth. This model may be used for quick screening of burn patients and can be executed in the cloud or on a local machine. The code of the proposed method can be accessed at https://github.com/dhirujis02/Journal.git for the sake of reproducibility.

Machine Learning Analysis of Microtensile Bond Strength of Dental Adhesives.

Dental adhesives provide retention to composite fillings in dental restorations. Microtensile bond strength (µTBS) test is the most used laboratory test to evaluate bonding performance of dental adhesives. The traditional approach for developing dental adhesives involves repetitive laboratory measurements, which consumes enormous time and resources. Machine learning (ML) is a promising tool for accelerating this process. This study aimed to develop ML models to predict the µTBS of dental adhesives using their chemical features and to identify important contributing factors for µTBS. Specifically, the chemical composition and µTBS information of 81 dental adhesives were collected from the manufacturers and the literature. The average µTBS value of each adhesive was labeled as either 0 (if <36 MPa) or 1 (if ≥36 MPa) to denote the low and high µTBS classes. The initial 9-feature data set comprised pH, HEMA, BisGMA, UDMA, MDP, PENTA, filler, fluoride, and organic solvent (OS) as input features. Nine ML algorithms, including logistic regression, k-nearest neighbor, support vector machine, decision trees and tree-based ensembles, and multilayer perceptron, were implemented for model development. Feature importance analysis identified MDP, pH, OS, and HEMA as the top 4 contributing features, which were used to construct a 4-feature data set. Grid search with stratified 10-fold cross-validation (CV) was employed for hyperparameter tunning and model performance evaluation using 2 metrics, the area under the receiver operating characteristic curve (AUC) and accuracy. The 4-feature data set generated slightly better performance than the 9-feature data set, with the highest AUC score of 0.90 and accuracy of 0.81 based on stratified CV. In conclusion, ML is an effective tool for predicting dental adhesives with low and high µTBS values and for identifying important chemical features contributing to the µTBS. The ML-based data-driven approach has great potential to accelerate the discovery of new dental adhesives and other dental materials.

Towards a Universal Privacy Model for Electronic Health Record Systems: An Ontology and Machine Learning Approach

This paper proposed a novel privacy model for Electronic Health Records (EHR) systems utilizing a conceptual privacy ontology and Machine Learning (ML) methodologies. It underscores the challenges currently faced by EHR systems such as balancing privacy and accessibility, user-friendliness, and legal compliance. To address these challenges, the study developed a universal privacy model designed to efficiently manage and share patients’ personal and sensitive data across different platforms, such as MHR and NHS systems. The research employed various BERT techniques to differentiate between legitimate and illegitimate privacy policies. Among them, Distil BERT emerged as the most accurate, demonstrating the potential of our ML-based approach to effectively identify inadequate privacy policies. This paper outlines future research directions, emphasizing the need for comprehensive evaluations, testing in real-world case studies, the investigation of adaptive frameworks, ethical implications, and fostering stakeholder collaboration. This research offers a pioneering approach towards enhancing healthcare information privacy, providing an innovative foundation for future work in this field.

Detecting Smart Contract Vulnerabilities with Combined Binary and Multiclass Classification

The development of Distributed Ledger Technology (DLT) is pushing toward automating decentralized data exchange processes. One of the key components of this evolutionary step is facilitating smart contracts that, in turn, come with several additional vulnerabilities. Despite the existing tools for analyzing smart contracts, keeping these systems running and preserving performance while maintaining a decent level of security in a constantly increasing number of contracts becomes challenging. Machine Learning (ML) methods could be utilized for analyzing and detecting vulnerabilities in DLTs. This work proposes a new ML-based two-phase approach for the detection and classification of vulnerabilities in smart contracts. Firstly, the system’s operation is set up to filter the valid contracts. Secondly, it focuses on detecting a vulnerability type, if any. In contrast to existing approaches in this field of research, our algorithm is more focused on vulnerable contracts, which allows to save time and computing resources in the production environment. According to the results, it is possible to detect vulnerability types with an accuracy of 0.9921, F1 score of 0.9902, precision of 0.9883, and recall of 0.9921 within reasonable execution time, which could be suitable for integrating existing DLTs.

An ML-Based Approach to Reconstruct Heart Rate from PPG in Presence of Motion Artifacts.

The heart rate (HR) is a widely used clinical variable that provides important information on a physical user's state. One of the most commonly used methods for ambulatory HR monitoring is photoplethysmography (PPG). The PPG signal retrieved from wearable devices positioned on the user's wrist can be corrupted when the user is performing tasks involving the motion of the arms, wrist, and fingers. In these cases, the obtained HR is altered as well. This problem increases when trying to monitor people with autism spectrum disorder (ASD), who are very reluctant to use foreign bodies, notably hindering the adequate attachment of the device to the user. This work presents a machine learning approach to reconstruct the user's HR signal using an own monitoring wristband especially developed for people with ASD. An experiment is carried out, with users performing different daily life activities in order to build a dataset with the measured signals from the monitoring wristband. From these data, an algorithm is applied to obtain a reliable HR value when these people are performing skill improvement activities where intensive wrist movement may corrupt the PPG.

An ML-Based Approach for HCF Life Prediction of Additively Manufactured AlSi10Mg Considering the Effects of Powder Size and Fatigue Damage

As a popular technique, additive manufacturing (AM) has garnered extensive utilization in various engineering domains. Given that numerous AM metal components are exposed to fatigue loads, it is of significant importance to investigate the life prediction methodology. This study aims to investigate the high-cycle fatigue (HCF) behavior of AM AlSi10Mg, taking into account the influences of powder size and fatigue damage, and a novel ML-based approach for life prediction is presented. First, the damage-coupled constitutive model and fatigue damage model are derived, and the Particle Swarm Optimization method is employed for the material parameters’ calibration of M AlSi10Mg. Second, the numerical implementation of theoretical models is carried out via the development of a user-defined material subroutine. The predicted fatigue lives of AM AlSi10Mg with varying powder sizes fall within the triple error band, which verifies the numerical method and the calibrated material parameters. After that, the machine learning approach for HCF life prediction is presented, and the Random Forest (RF) and K-Nearest Neighbor (KNN) models are employed to predict the fatigue lives of AM AlSi10Mg. The RF model achieves a smaller MSE and a larger R2 value compared to the KNN model, signifying its superior performance in predicting the overall behavior of AM AlSi10Mg. Under the same maximum stress, a decrease in the stress ratio from 0.5 to −1 leads to a reduction in fatigue life for both powder sizes. As the powder size decreases, the rate of damage evolution accelerates, leading to shorter fatigue life.

A Modified DSR Protocol Using Deep Reinforced Learning for MANETS

Wireless Networks like Mobile Adhoc Networks (MANETs) have been under extensive research over the past few years with congestion control as one of the most important features to ensure efficient and fair sharing of network resources among users. Machine learning (ML) has achieved a success rate in addressing large-scale and complex problems and researchers have begun to shift their attention from the rule-based method to an ML-based approach to handle the complex needs of future networks where conventional rule-based approaches tend to become inefficient and ineffective. To handle the problems in congestion control, precise notification should be generated as the backpressure transmission process. Although backpressure-based routing algorithms give optimal throughput, they typically have poor delay performance under moderate loads. This may be because packets are being sent over longer routes unnecessarily. Furthermore, the existing backpressure-based optimisation algorithms require every node to compute differential backlogs for every destination queue with the corresponding destination queue at every adjacent node. The proposed algorithm proves to be a cross-layer protocol for wireless MANET that generalises the channel access management and routing process which includes traffic management, connection maintenance and distributed scheduling for concurrent transmission. The joint congestion control method with scheduling algorithm has enhanced active radio communication network by interchanging scheduling schema with adaptation modelling and also the optimum congestion dominance and flow control model is being designed using deep reinforced learning.

Integrating experiments, finite element analysis, and interpretable machine learning to evaluate the auxetic response of 3D printed re-entrant metamaterials

Metamaterials have received extensive attention in fundamental and applied research over the past two decades due to their unique mechanical behavior. This paper presents an interpretable machine learning (ML) approach for efficient response prediction of three-dimensional (3D)-printed metamaterials. However, developing such an ML-based model requires a large consistent, representative, balanced, and complete dataset. To this extent, an experimentally validated finite element analysis (FEA) approach is implemented to generate 8096 non-self-intersecting re-entrant honeycomb structures by varying the mesoscale geometrical features to obtain the corresponding Poisson's ratios. This dataset is leveraged to develop a feed-forward multilayer perceptron-based predictive model. The developed ML model shows excellent predictive efficacy on the unseen test dataset. Shapely additive explanation (SHAP) is then used for model interpretation. SHAP results show that the slant cell length is the dominant input feature dictating the model output whereas cell angle and vertical cell length show mixed trends signifying that other input features influence their effect on the model output. Moreover, cell thickness does not significantly influence the model output when compared to other input features. Overall, the integrated numerical simulation-experiment-interpretable ML-based predictive approach presented here can be leveraged to design and develop metamaterials for a wide range of engineering applications.

Analysis of Conditioning Factors in Cuenca, Ecuador, for Landslide Susceptibility Maps Generation Employing Machine Learning Methods

Landslides are events that cause great impact in different parts of the world. Their destructive capacity generates loss of life and considerable economic damage. In this research, several Machine Learning (ML) methods were explored to select the most important conditioning factors, in order to evaluate the susceptibility to rotational landslides in a sector surrounding the city of Cuenca (Ecuador) and with them to elaborate landslide susceptibility maps (LSM) by means of ML. The methods implemented to analyze the importance of the conditioning factors checked for multicollinearity (correlation analysis and VIF), and, with an ML-based approach called feature selection, the most important factors were determined based on Classification and Regression Trees (CART), Feature Selection with Random Forests (FS RF), and Boruta and Recursive Feature Elimination (RFE) algorithms. LSMs were implemented with Random Forests (RF) and eXtreme Gradient Boosting (XGBoost) methods considering a landslide inventory updated to 2019 and 15 available conditioning factors (topographic (10), land cover (3), hydrological (1), and geological (1)), from which, based on the results of the aforementioned analyses, the six most important were chosen. The LSM were elaborated considering all available factors and the six most important ones, with the previously mentioned ML methods, and were compared with the result generated by an Artificial Neural Network with resilient backpropagation (ANN rprop-) with six conditioning factors. The results obtained were validated by means of AUC-ROC value and showed a good predictive capacity for all cases, highlighting those obtained with XGBoost, which, in addition to a high AUC value (&gt;0.84), obtained a good degree of coincidence of landslides at high and very high susceptibility levels (&gt;72%). Despite the findings of this research, it is necessary to study in depth the methods applied for the development of future research that will contribute to developing a preventive approach in the study area.

A holistic and proactive approach to forecasting cyber threats

Traditionally, cyber-attack detection relies on reactive, assistive techniques, where pattern-matching algorithms help human experts to scan system logs and network traffic for known virus or malware signatures. Recent research has introduced effective Machine Learning (ML) models for cyber-attack detection, promising to automate the task of detecting, tracking and blocking malware and intruders. Much less effort has been devoted to cyber-attack prediction, especially beyond the short-term time scale of hours and days. Approaches that can forecast attacks likely to happen in the longer term are desirable, as this gives defenders more time to develop and share defensive actions and tools. Today, long-term predictions of attack waves are mostly based on the subjective perceptiveness of experienced human experts, which can be impaired by the scarcity of cyber-security expertise. This paper introduces a novel ML-based approach that leverages unstructured big data and logs to forecast the trend of cyber-attacks at a large scale, years in advance. To this end, we put forward a framework that utilises a monthly dataset of major cyber incidents in 36 countries over the past 11 years, with new features extracted from three major categories of big data sources, namely the scientific research literature, news, blogs, and tweets. Our framework not only identifies future attack trends in an automated fashion, but also generates a threat cycle that drills down into five key phases that constitute the life cycle of all 42 known cyber threats.

Detecting Remote Access Network Attacks Using Supervised Machine Learning Methods

Remote access technologies encrypt data to enforce policies and ensure protection. Attackers leverage such techniques to launch carefully crafted evasion attacks introducing malware and other unwanted traffic to the internal network. Traditional security controls such as anti-virus software, firewall, and intrusion detection systems (IDS) decrypt network traffic and employ signature and heuristic-based approaches for malware inspection. In the past, machine learning (ML) approaches have been proposed for specific malware detection and traffic type characterization. However, decryption introduces computational overheads and dilutes the privacy goal of encryption. The ML approaches employ limited features and are not objectively developed for remote access security. This paper presents a novel ML-based approach to encrypted remote access attack detection using a weighted random forest (W-RF) algorithm. Key features are determined using feature importance scores. Class weighing is used to address the imbalanced data distribution problem common in remote access network traffic where attacks comprise only a small proportion of network traffic. Results obtained during the evaluation of the approach on benign virtual private network (VPN) and attack network traffic datasets that comprise verified normal hosts and common attacks in real-world network traffic are presented. With recall and precision of 100%, the approach demonstrates effective performance. The results for k-fold cross-validation and receiver operating characteristic (ROC) mean area under the curve (AUC) demonstrate that the approach effectively detects attacks in encrypted remote access network traffic, successfully averting attackers and network intrusions.

On Some Novel Similarity-Based Functions Used in the ML-Based q-RASAR Approach for Efficient Quantitative Predictions of Selected Toxicity End Points.

The novel quantitative read-across structure-activity relationship (q-RASAR) approach uses read-across-derived similarity functions in the quantitative structure-activity relationship (QSAR) modeling framework in a unique way for supervised model generation. The aim of this study is to explore how this workflow enhances the external (test set) prediction quality of conventional QSAR models by the incorporation of some novel similarity-based functions as additional descriptors using the same level of chemical information. To establish this, five different toxicity data sets, for which QSAR models were reported previously, have been considered in the q-RASAR modeling exercise, which uses chemical similarity-derived measures. The identical sets of chemical features along with the same compositions of training and test sets as reported previously were used in the present analysis for ease of comparison. The RASAR descriptors were calculated based on a chosen similarity measure with the default setting of relevant hyperparameter(s) and were then clubbed with the original structural and physicochemical descriptors, and the number of selected features was further optimized by employing a grid search technique applied on the respective training sets. These features were then used to develop multiple linear regression (MLR) q-RASAR models that show enhanced predictivity as compared to the QSAR models developed previously. Moreover, various other ML algorithms like support vector machine (SVM), linear SVM, random forest, partial least squares, and ridge regression were also employed using the same feature combinations as used in the MLR models to compare the prediction qualities. The q-RASAR models for five different data sets possess at least one of the RASAR descriptors, RA function, gm, and average similarity, suggesting that these are important determinants of similarities that contribute to the development of predictive q-RASAR models, as also evident from the SHAP analysis of the models.

Amplified radio-over-fiber system linearization using recurrent neural networks

Ubiquitous communication is an emergent feature of the future sixth generation of mobile communication (6G) networks. One of the main challenges of this upcoming mobile network is providing broadband communication wherever connectivity is necessary, including rural areas. Up to now, all previous generations of mobile networks have not satisfactorily accomplished this task, and the high cost of deploying and maintaining complex communication infrastructure in remote areas is one of the main reasons. The centralized radio access network might play an important role in overcoming it, since it enables all baseband centralization in a central office (CO), simplifying network deployment and reducing operating/maintenance costs. Additionally, in this scenario, the radio over fiber (RoF) system might be used for transporting analog signals from the CO to a simplified remote base station, composed only of an optical detector and radio-frequency front-end. However, the Mach–Zehnder modulator and power amplifier, commonly employed in RoF systems, introduce undesired nonlinear effects that can severely degrade overall system performance and prohibitively increase the out-of-band emission. We investigate the use of machine learning (ML) algorithms applied to the linearization of an electrical-amplified RoF system. Particularly, a memory recurrent neural network (RNN) linearization is proposed and compared with a memoryless multilayer perceptron linearization. The root mean square error vector magnitude, normalized mean square error, and adjacent channel leakage ratio metrics have been calculated to evaluate the performance of our ML-based approach. Numerical results demonstrate promising linearization performance when the RNN memory depth is equal to or higher than the amplified-RoF system memory depth.

Quantitative Prediction of Inorganic Nanomaterial Cellular Toxicity via Machine Learning.

Organic chemistry has seen colossal progress due to machine learning (ML). However, the translation of artificial intelligence (AI) into materials science is challenging, where biological behavior prediction becomes even more complicated. Nanotoxicity is a critical parameter that describes their interaction with the living organisms screened in every bio-related research. To prevent excessive experiments, such properties have to be pre-evaluated. Several existing ML models partially fulfill the gap by predicting whether a nanomaterial is toxic or not. Yet, this binary categorization neglects the concentration dependencies crucial for experimental scientists. Here, an ML-based approach is proposed to the quantitative prediction of inorganic nanomaterial cytotoxicity achieving the precision expressed by 10-fold cross-validation (CV) Q2 = 0.86 with the root mean squared error (RMSE) of 12.2% obtained by the correlation-based feature selection and grid search-based model hyperparameters optimization. To provide further model flexibility, quantitative atom property-based nanomaterial descriptors are introduced allowing the model to extrapolate on unseen samples. Feature importance is calculated to find an interpretable model with optimal decision-making. These findings allow experimental scientists to perform primary in silico candidate screening and minimize the number of excessive, labor-intensive experiments enabling the rapid development of nanomaterials for medicinal purposes.

Abstract TP102: Automated Assessment Of Ischemic Core On Noncontrast Computed Tomography: A Comparative Analysis With CT Perfusion

Introduction: Application of machine learning (ML) algorithms has shown promising results in estimating ischemic core volumes using routine non-contrast CT (NCCT). We aimed to assess the performance of the e-Stroke Suite software (Brainomix, Oxford, United Kingdom) in assessing ischemic core volumes on NCCT compared to CTP in patients with acute ischemic stroke. Methods: In this retrospective study, consecutive patients with anterior circulation large vessel occlusions who underwent pretreatment NCCT and CTP and posttreatment MRI were included. Ischemic core volumes were automatically calculated on NCCTs using e-Stroke Suite (Brainomix) which uses a combination of traditional 3D graphics and ML classification techniques to identify ischemic core voxels. Estimated ischemic core volumes were also automatically calculated from CTP using Olea Sphere (Olea Medical SAS, SP23) using a combination of rCBF&lt;25% and differential Time-to-peak (dTTP)&gt;5 sec. Estimated core volumes were compared against the final infarct volume on posttreatment MRI in patients who achieved successful reperfusion (mTICI ≥2b). Results: 83 patients [52 female; age (mean ± SD): 73.1 ±15.3] were included. The estimated ischemic core volumes (mean ± SD) were 18.9 ± 13.5 mL on NCCT and 17.5 ± 16.5 mL on CTP, not significantly different (p=0.54) and demonstrated significant correlation (r=0.51, p&lt;0.001) ( Figure 1 ). Among patients with successful recanalization (n=49), there was no significant difference in estimated ischemic core volume between NCCT vs. CTP (p=0.80) and NCCT vs. MRI (p=0.38). There was significant correlation between estimated ischemic core volume on NCCT vs. CTP (r=0.75, p&lt;0.001) and vs. final MRI infarct volume (r=0.75, p&lt;0.001). Conclusions: Results show estimated ischemic core volumes obtained automatically by ML-based approach (Brainomix) on NCCT correlates well with ischemic core volumes on acute CTP and with post-treatment MR infarct volume.

MACHINE LEARNING-BASED PREDICTION OF GEBOES SCORE AND HISTOLOGIC IMPROVEMENT AND REMISSION THRESHOLDS IN ULCERATIVE COLITIS

Abstract BACKGROUND Histology is emerging as a key therapeutic endpoint for ulcerative colitis driven by associations between histologic response and long-term outcomes. However, existing scoring systems are subjective and consequently have variable inter- and intra-reader variability. Geboes scoring is a well-established system for ulcerative colitis histologic assessment that has previously been used to define thresholds for histo-endoscopic mucosal improvement (Geboes Score ≤3.1, together with Mayo score 0 or 1) and histologic remission (Geboes Score &amp;lt;2). Here we report the first machine learning (ML)-based prediction of the Geboes Score, and Geboes Score-derived thresholds of histologic improvement and remission, directly from whole slide images (WSI) of hematoxylin and eosin (H&amp;E)-stained mucosal biopsies. METHODS 3,148 WSI were scored by three expert gastrointestinal pathologists and the median consensus score was used to determine the Geboes score for each slide as ground truth. ML models were trained on median consensus scores to predict the Geboes score and subscores for each slide. Model performance vs. pathologist median consensus score was measured using accuracy and the F1 score, which accounts for both false positive and false negative errors. RESULTS The ML-based model performance, measured against median consensus scores of three pathologists, showed strong performance at predicting overall Geboes Score, with a quadratic kappa of 0.89. The model was also able to predict both histologic improvement and histologic remission with high accuracy. For predicting histological improvement as defined by a Geboes Score of ≤3.1, the model showed accuracy of 0.92 and F1 score of 0.92 (Figure 1). For predicting histological remission as defined by a Geboes Score of &amp;lt; 2, the model showed accuracy of 0.91 and F1 score of 0.89 (Figure 2). CONCLUSIONS We report a ML-based approach for predicting Geboes score and Geboes score-based key thresholds of histologic improvement and histologic remission. Model predictions show high accuracy compared to median consensus pathologist scores. This approach may enable standardized, reproducible and accurate prediction of these clinically relevant thresholds to better measure histologic disease activity and treatment response in clinical trials.