Multi-class Model Research Articles

MKELM based multi-classification model for foreign accent identification

The automatic identification of foreign accents can play a crucial role in various speech systems, including speaker identification, e-learning, telephone banking, and more. Additionally, it can greatly enhance the robustness of Automatic Speech Recognition (ASR) systems. Non-native accents in speech signals are characterized by distinct pronunciations, prosody, and voice characteristics of the speaker. However, automatically identifying foreign accents poses significant challenges, particularly in the context of multi-class modeling. Multi-classification models face difficulties in achieving high performance and dealing with computational challenges when confronted with multi-dimensional and unbalanced datasets, such as those with more than two accents. Furthermore, the choice of features remains a bottleneck problem for Foreign Accent Identification (FAID), further hindering performance in these tasks. Consequently, the accuracy of current systems is typically low. To address these challenges, this paper proposes a framework based on the Multi-Kernel Extreme Learning Machine (MKELM) model for the multi-classification of FAID. The MKELM model utilizes a novel weighted scheme to classify various non-native English accents, including Arabic, Chinese, Korean, French, and Spanish. The model first combines Mel-frequency cepstral coefficients (MFCCs) and prosodic features as input, trains pairwise binary classifiers independently, and subsequently employs a weighting scheme to distinguish between classes and identify accents. Through experiments, the proposed model achieves an accuracy rate of 84.72% using a paired weighting scheme. In contrast, the accuracy rate drops to 66.5% when employing the traditional non-weighted multi-classification scheme. A comparison with other models demonstrates the significant advantages of the proposed model in FAID multi-class classification, showcasing improved accuracy, reduced computational complexity (requiring fewer computations, faster learning rates, and shorter training time), and enhanced stability compared to state-of-the-art classification methods.

Facial recognition models for identifying genetic syndromes associated with pulmonary stenosis in children.

Williams-Beuren syndrome, Noonan syndrome, and Alagille syndrome are common types of genetic syndromes (GSs) characterized by distinct facial features, pulmonary stenosis, and delayed growth. In clinical practice, differentiating these three GSs remains a challenge. Facial gestalts serve as a diagnostic tool for recognizing Williams-Beuren syndrome, Noonan syndrome, and Alagille syndrome. Pretrained foundation models (PFMs) can be considered the foundation for small-scale tasks. By pretraining with a foundation model, we propose facial recognition models for identifying these syndromes. A total of 3297 (n = 1666) facial photos were obtained from children diagnosed with Williams-Beuren syndrome (n = 174), Noonan syndrome (n = 235), and Alagille syndrome (n = 51), and from children without GSs (n = 1206). The photos were randomly divided into five subsets, with each syndrome and non-GS equally and randomly distributed in each subset. The proportion of the training set and the test set was 4:1. The ResNet-100 architecture was employed as the backbone model. By pretraining with a foundation model, we constructed two face recognition models: one utilizing the ArcFace loss function, and the other employing the CosFace loss function. Additionally, we developed two models using the same architecture and loss function but without pretraining. The accuracy, precision, recall, and F1 score of each model were evaluated. Finally, we compared the performance of the facial recognition models to that of five pediatricians. Among the four models, ResNet-100 with a PFM and CosFace loss function achieved the best accuracy (84.8%). Of the same loss function, the performance of the PFMs significantly improved (from 78.5% to 84.5% for the ArcFace loss function, and from 79.8% to 84.8% for the CosFace loss function). With and without the PFM, the performance of the CosFace loss function models was similar to that of the ArcFace loss function models (79.8% vs 78.5% without PFM; 84.8% vs 84.5% with PFM). Among the five pediatricians, the highest accuracy (0.700) was achieved by the senior-most pediatrician with genetics training. The accuracy and F1 scores of the pediatricians were generally lower than those of the models. A facial recognition-based model has the potential to improve the identification of three common GSs with pulmonary stenosis. PFMs might be valuable for building screening models for facial recognition. Key messages What is already known on this topic: Early identification of genetic syndromes (GSs) is crucial for the management and prognosis of children with pulmonary stenosis (PS). Facial phenotyping with convolutional neural networks (CNNs) often requires large-scale training data, limiting its usefulness for GSs. What this study adds: We successfully built multi-classification models based on face recognition using a CNN to accurately identify three common PS-associated GSs. ResNet-100 with a pretrained foundation model (PFM) and CosFace loss function achieved the best accuracy (84.8%). Pretrained with the foundation model, the performance of the models significantly improved, although the impact of the type of loss function appeared to be minimal. How this study might affect research, practice, or policy: A facial recognition-based model has the potential to improve the identification of GSs in children with PS. The PFM might be valuable for building identification models for facial detection.

Predicting multiple taste sensations with a multiobjective machine learning method

Taste perception plays a pivotal role in guiding nutrient intake and aiding in the avoidance of potentially harmful substances through five basic tastes - sweet, bitter, umami, salty, and sour. Taste perception originates from molecular interactions in the oral cavity between taste receptors and chemical tastants. Hence, the recognition of taste receptors and the subsequent perception of taste heavily rely on the physicochemical properties of food ingredients. In recent years, several advances have been made towards the development of machine learning-based algorithms to classify chemical compounds’ tastes using their molecular structures. Despite the great efforts, there remains significant room for improvement in developing multi-class models to predict the entire spectrum of basic tastes. Here, we present a multi-class predictor aimed at distinguishing bitter, sweet, and umami, from other taste sensations. The development of a multi-class taste predictor paves the way for a comprehensive understanding of the chemical attributes associated with each fundamental taste. It also opens the potential for integration into the evolving realm of multi-sensory perception, which encompasses visual, tactile, and olfactory sensations to holistically characterize flavour perception. This concept holds promise for introducing innovative methodologies in the rational design of foods, including pre-determining specific tastes and engineering complementary diets to augment traditional pharmacological treatments.

Machine-learning approach for estimating passenger car equivalent factors using crowdsourced data

Passenger car equivalent (PCE) factors are used by the Highway Capacity Manual (HCM) to convert truck volumes to equivalent passenger car volumes and are typically calculated using multi-class volumes collected from traffic sensors. However, this requires costly sensor installations that provide limited spatial coverage. Therefore, this study proposes a novel approach to estimate PCE volumes using crowdsourced and open data. A multi-class volume estimation model (TS-SAE-XGB) is proposed to estimate passenger car and truck volumes, and single unit truck ratios. These parameters are input to a PCE interpolation algorithm which estimates PCE values using HCM methods. A spatial leave-one-out cross validation was conducted to compare the proposed model against five other machine learning models when estimating PCE values. The TS-SAE-XGB model estimated PCE and heavy vehicle factors with a MAPE of 6.22% and 3.03%, respectively, providing transportation professionals a practical method of estimating freeway PCE values where sensors are unavailable.

Evaluation of adulteration in soy-based beverages by water addition using chemometrics applied to ATR-FTIR spectroscopy

The adulteration of soy-based beverages (SBBs) by adding water to increase profitability is a fraudulent practice that requires urgent solutions to ensure product integrity and consumer trust. Therefore, the use of infrared spectroscopy (ATR-FTIR) associated with chemometrics methods can be a quick and advantageous alternative to this problem. In this study, the one-class and multiclass methods applied to ATR-FTIR data to classify a set of 80 SBBs samples were used. The unequal dispersed classes (UNEQ), soft independent modeling of class analogy (SIMCA), data driven SIMCA (DD-SIMCA), and one-class random forest (OC-RF) methods were used for one-class modeling. Models were constructed using the non-adulterated samples as target class (TA) and the adulterated samples as non-target class (NT). The k-nearest neighbors (k-NN), partial least squares discriminant analysis (PLS-DA), dual class random forest (DC-RF), and dual class random forest with Monte Carlo sampling (DC-RF-MC) methods were used for multiclass modeling. For k-NN and PLS-DA, samples were organized into four classes (non-adulterated samples, adulterated with 5% v∙v−1, 10% v∙v−1, and 20% v∙v−1 of water). DC-RF models used the same class settings as one-class models. DD-SIMCA, PLS-DA, and DC-RF-MC showed accuracy of 100%. The results show the feasibility of ATR-FTIR and chemometrics models to identify adulterations by adding water.

Using Machine Learning Multiclass Classification Technique to Detect IoT Attacks in Real Time.

This paper presents a real-time intrusion detection system (IDS) aimed at detecting the Internet of Things (IoT) attacks using multiclass classification models within the PySpark architecture. The research objective is to enhance detection accuracy while reducing the prediction time. Various machine learning algorithms are employed using the OneVsRest (OVR) technique. The proposed method utilizes the IoT-23 dataset, which consists of network traffic from smart home IoT devices, for model development. Data preprocessing techniques, such as data cleaning, transformation, scaling, and the synthetic minority oversampling technique (SMOTE), are applied to prepare the dataset. Additionally, feature selection methods are employed to identify the most relevant features for classification. The performance of the classifiers is evaluated using metrics such as accuracy, precision, recall, and F1 score. The results indicate that among the evaluated algorithms, extreme gradient boosting achieves a high accuracy of 98.89%, while random forest demonstrates the most efficient training and prediction times, with a prediction time of only 0.0311 s. The proposed method demonstrates high accuracy in real-time intrusion detection of IoT attacks, outperforming existing approaches.

A supervised multiclass framework for mineral classification of Iberian beads.

Research on personal adornments depends on the reliable characterisation of materials to trace provenance and model complex social networks. However, many analytical techniques require the transfer of materials from the museum to the laboratory, involving high insurance costs and limiting the number of items that can be analysed, making the process of empirical data collection a complicated, expensive and time-consuming routine. In this study, we compiled the largest geochemical dataset of Iberian personal adornments (n = 1243 samples) by coupling X-ray fluorescence compositional data with their respective X-ray diffraction mineral labels. This allowed us to develop a machine learning-based framework for the prediction of bead-forming minerals by training and benchmarking 13 of the most widely used supervised algorithms. As a proof of concept, we developed a multiclass model and evaluated its performance on two assemblages from different Portuguese sites with current mineralogical characterisation: Cova das Lapas (n = 15 samples) and Gruta da Marmota (n = 10 samples). Our results showed that decisión-tres based classifiers outperformed other classification logics given the discriminative importance of some chemical elements in determining the mineral phase, which fits particularly well with the decision-making process of this type of model. The comparison of results between the different validation sets and the proof-of-concept has highlighted the risk of using synthetic data to handle imbalance and the main limitation of the framework: its restrictive class system. We conclude that the presented approach can successfully assist in the mineral classification workflow when specific analyses are not available, saving time and allowing a transparent and straightforward assessment of model predictions. Furthermore, we propose a workflow for the interpretation of predictions using the model outputs as compound responses enabling an uncertainty reduction approach currently used by our team. The Python-based framework is packaged in a public repository and includes all the necessary resources for its reusability without the need for any installation.

SMART: Supervised multi-class image retargeting generative model based on a long-range sampling strategy

Content-aware image retargeting (CAIR) techniques are crucial in multimedia processing for displaying images on various devices while preserving visually salient contents with desirable visual effects. There are discrete and continuous algorithms. For the former, the artefacts happen when the foreground proportion is larger than the retargeting ratio; for the latter, the salient regions are prone to be squeezed. In this paper, we reformulate the retargeting process into sampling the salient signal and reconstruction under aesthetic supervision, the supervised multi-class image retargeting reconstruction (SMART) framework. The target images can be represented into complementary parts, the masked and unmasked ones, according to the saliency influences in the encoder phrase. The long-range sampling algorithm is proposed to calculate similarities through an 8-connected planar path while considering spatial distance and feature correlation. The sampled embeddings in latent space reconstruct the retargeted images under supervised signals for aesthetic quality. The semantic loss Lsem from the pretrained CLIP model can maintain consistency for both content and semantics. The supervised loss, Lir, is introduced to ensure the retargeted qualities are close to the preferred labels. Then, we release a new retargeting dataset comprising seven image classes (animal, building, car, flower, indoor, landscape and people) with supervised labels collected from designers for further aesthetic retargeting study. The ablation studies are conducted to confirm the effectiveness of the new dataset, and comparative experiments with state-of-the-art baselines demonstrate the advantages of the proposed method.

Application of multi-label convolutional neural networks on enhancement of reliability of classification, a case study on bearing fault diagnostics

In contrast to the mutli-class classification where one and only one class would be assigned to each instance, multi-label classification allows for more flexible classification and possibility of having multiple labels for instance. This property enables us to model the dependencies between the classes such as effect of one label on the probability of occurrence of the other label. Considering these connections between the classes can help us to enhance the reliability of classifier and handle the contradictory prediction. Much like how we can assign multiple classes to an instance, we can also have an instance that remains unlabeled. This class is typically referred to as the "neutral" label, allowing us to bluntly state that we lack sufficient information to label the instance with adequate certainty. Leveraging these properties, a 1D convolutional neural network (CNN) has been proposed to classify common faults occurring in the bearings. Network has the capability to assign a Confidence level between 0 to 1 to each class independently. Later, a threshold has been set on each class to identify the instance labels if any. Subsequently, instances with contradictory predicted label have been flagged as undetermined and removed from the predicted result. thus, the remaining classified instances can be labeled with sufficient confident. In the end, the multi-label CNN model performance has been compared with equivalent multi-class CNN model. Technique has been evaluated using MAFAULDA open dataset which is composed of 1951 multivariate time-series acquired by a triaxial accelerometer, a microphone, and a tachometer using SpectraQuest’s Machinery Fault Simulator (MFS) experimentally simulating different bearing states including normal operation as well as inner and outer race bearing faults. A data reduction scheme has been performed on the dataset to make the diagnostic more challenging and evaluate the performance of CNN in more realistic condition with insufficient data.

Performance evaluation of multiclass classification models for ToN-IoT network device datasets

Internet of things (IoT) technology has empowered tangible objects to establish internet connections, facilitating data exchange with computational capabilities. With significant potential across sectors like healthcare, environmental monitoring, and industrial control, IoT represents a promising technological advancement. This study explores datasets from ToN-IoT’s IoT devices, focusing on multi-class classification, including normal and attack classes, with an additional aim of identifying potential attack sub-classes. Datasets comprise various IoT devices, such as refrigerators, garage doors, global positioning systems (GPS) sensors, motion lights, modbus devices, thermostats, and weather sensors. Comparative analysis is conducted between two prominent multiclass classification models, extreme gradient boosting (XGBoost) and light gradient boosting machine (LightGBM), utilizing accuracy and computational time metrics as evaluation criteria. Research findings highlight that the LightGBM model achieves superior accuracy at 78%, surpassing XGBoost 74.31%. However, XGBoost demonstrates an advantage with a shorter computational time of 1.23 seconds, compared to LightGBM 6.79 seconds. This study not only provides insights into multiclass classification model selection but also underscores the crucial consideration of the trade-off between accuracy and computational efficiency in decision-making. Research contributes to advancing our understanding of IoT security through effective classification methodologies. The findings offer valuable information for researchers and practitioners, emphasizing the nuanced decisions needed when selecting models based on specific priorities like accuracy and computational efficiency.

Using AutoML and generative AI to predict the type of wildfire propagation in Canadian conifer forests

Accurate prediction of wildfire behaviour is critical for safe and effective fire management and suppression. This study focused on developing and evaluating machine learning (ML) models based on a dataset collected of outdoor experimental fires in Canadian conifer forests. Binary classification models (surface fire vs. crown fire) and multi-class fire type models (surface, passive crown, active crown) were created using ensemble tree methods (Random Forest, XGBoost) and automated ML (AutoGluon, TabPFN). Generative adversarial networks (GAN) were used to generate synthetic data to overcome imbalances in the type of fire distribution. The results show automated ML methods applied to the binary problem to perform best of all tested methods, with an overall accuracy of 91%. TabPFN had the highest accuracy for the multi-class problem, with the use of GAN improving model fit. Results show overfitting issues with some of the ML models, highlighting the need of independent evaluation when ML models are developed. The TabPFN models have potential for application in supporting fire management, namely in fuel management and identifying situations with high fire spread and intensity potential that can impact firefighter safety.

Automatic classification and segmentation of multiclass jaw lesions in cone-beam CT using deep learning.

To develop and validate a modified deep learning (DL) model based on nnU-Net for classifying and segmenting five-class jaw lesions using cone-beam CT (CBCT). A total of 368 CBCT scans (37 168 slices) were used to train a multi-class segmentation model. The data underwent manual annotation by two oral and maxillofacial surgeons (OMSs) to serve as ground truth. Sensitivity, specificity, precision, F1-score, and accuracy were used to evaluate the classification ability of the model and doctors, with or without artificial intelligence assistance. The dice similarity coefficient (DSC), average symmetric surface distance (ASSD), and segmentation time were used to evaluate the segmentation effect of the model. The model achieved the dual task of classifying and segmenting jaw lesions in CBCT. For classification, the sensitivity, specificity, precision, and accuracy of the model were 0.871, 0.974, 0.874, and 0.891, respectively, surpassing oral and maxillofacial radiologists (OMFRs) and OMSs, approaching the specialist. With the model's assistance, the classification performance of OMFRs and OMSs improved, particularly for odontogenic keratocyst (OKC) and ameloblastoma (AM), with F1-score improvements ranging from 6.2% to 12.7%. For segmentation, the DSC was 87.2% and the ASSD was 1.359 mm. The model's average segmentation time was 40 ± 9.9 s, contrasting with 25 ± 7.2 min for OMSs. The proposed DL model accurately and efficiently classified and segmented five classes of jaw lesions using CBCT. In addition, it could assist doctors in improving classification accuracy and segmentation efficiency, particularly in distinguishing confusing lesions (eg, AM and OKC).

Malware Detection for Internet of Things Using One-Class Classification.

The increasing usage of interconnected devices within the Internet of Things (IoT) and Industrial IoT (IIoT) has significantly enhanced efficiency and utility in both personal and industrial settings but also heightened cybersecurity vulnerabilities, particularly through IoT malware. This paper explores the use of one-class classification, a method of unsupervised learning, which is especially suitable for unlabeled data, dynamic environments, and malware detection, which is a form of anomaly detection. We introduce the TF-IDF method for transforming nominal features into numerical formats that avoid information loss and manage dimensionality effectively, which is crucial for enhancing pattern recognition when combined with n-grams. Furthermore, we compare the performance of multi-class vs. one-class classification models, including Isolation Forest and deep autoencoder, that are trained with both benign and malicious NetFlow samples vs. trained exclusively on benign NetFlow samples. We achieve 100% recall with precision rates above 80% and 90% across various test datasets using one-class classification. These models show the adaptability of unsupervised learning, especially one-class classification, to the evolving malware threats in the IoT domain, offering insights into enhancing IoT security frameworks and suggesting directions for future research in this critical area.

Dual-stage deep learning for sangac optical fiber sensing multi-event detection and localization

This paper introduces a dual-stage deep learning structure for multi-event detection and localization (DDMDL) for a loop-based Sagnac interferometer optical fiber sensing system to overcome the difficulties of detecting and localizing multiple events. The DDMDL approach comprises of a combination of (i) a detection stage made of a multi-class classification model that quantifies the events, and (ii) a localization stage made of separate deep-learning regression models tasked with precisely localizing the events. By addressing the multi-event localization as a combined multi-class classification and multi-regression problem, our approach enables more accurate prediction of event locations. We evaluate our DDMDL model in a broad range of test scenarios across three signal-to-noise ratio (SNR) levels: low, medium, and high, all outside the model’s training dataset. The model had high detection (i.e., classification) accuracy in simulation, with rates of 90%–99% for single-event scenarios, 94%–98% for two-event scenarios, and 82%–99% for three-event scenarios at the three SNR levels. The precision of localization (i.e., regression) was evaluated by Mean Absolute Error (MAE). The results showed that single-event scenarios could be localized within a range of 35–50 m, two-event scenarios within a range of 82–106 m, and three-event scenarios within a range of 131–151 m. These findings were consistent across different SNR levels, ranging from low to high, over a 50 km sensing fiber length. Experimental validations confirmed the DDMDL model’s practical applicability, with detection accuracy of 80% in single-event scenarios and localization accuracy with MAEs ranging between 32 and 65 m. For two-event scenarios, the model achieved an 87% success rate, with MAEs ranging between 122 m and 204 m, emphasizing the model’s potential effectiveness in different applications.

Class conditioned text generation with style attention mechanism for embracing diversity

In the field of artificial intelligence and natural language processing (NLP), natural language generation (NLG) has significantly advanced. Its primary aim is to automatically generate text in a manner resembling human language. Traditional text generation has mainly focused on binary style transfers, limiting the scope to simple transformations between positive and negative tones or between modern and ancient styles. However, accommodating style diversity in real scenarios presents greater complexity and demand. Existing methods usually fail to capture the richness of diverse styles, hindering their utility in practical applications. To address these limitations, we propose a multi-class conditioned text generation model. We overcome previous constraints by utilizing a transformer-based decoder equipped with adversarial networks and style-attention mechanisms to model various styles in multi-class text. According to our experimental results, the proposed model achieved better performance compared to the alternatives on multi-class text generation tasks in terms of diversity while it preserves fluency. We expect that our study will help researchers not only train their models but also build simulated multi-class text datasets for further research.

Comprehensive comparisons of gradient-based multi-label adversarial attacks

Adversarial examples which mislead deep neural networks by adding well-crafted perturbations have become a major threat to classification models. Gradient-based white-box attack algorithms have been widely used to generate adversarial examples. However, most of them are designed for multi-class models, and only a few gradient-based adversarial attack algorithms specifically designed for multi-label classification models. Due to the correlation between multiple labels, the performance of these gradient-based algorithms in generating adversarial examples for multi-label classification is worthy of analyzing and evaluating comprehensively. In this paper, we first transplant five typical gradient-based adversarial attack algorithms in the multi-class environment to the multi-label environment. Secondly, we comprehensively compared the performance of these five attack algorithms and the other four existing multi-label adversarial attack algorithms by experiments on six different attack types, and evaluated the transferability of adversarial examples generated by all algorithms under two attack types. Experimental results show that, among different attack types, the majority of multi-step attack algorithms have higher attack success rates compared to one-step attack algorithms. Additionally, these gradient-based algorithms face greater difficulty in augmenting labels than in hiding them. For transfer experimental results, the adversarial examples generated by all attack algorithms exhibit weaker transferability when attacking other different models.

Broiler health monitoring technology based on sound features and random forest

The existing broiler health monitoring technology has problems such as low automation, unstable monitoring results, and low practical value, making it difficult to provide timely and reliable broiler health monitoring results. The broiler sound signal can provide feedback on their health. A widely validated and correct experience is to analyze the frequency of coughs in a segment of broiler sound signal to determine the health of the broiler group. Based on this, in this paper, the authors proposed a new broiler health monitoring technology based on sound detection. The broiler health monitoring problem is cleverly transformed into a multi-classification problem, which can be solved by identifying the sound types in broiler sound signals. Specifically, the audio signal collection system was designed to complete signal collection and preliminary signal filtering. Wiener filtering was used for deep signal filtering. The 60-dimensional sound features with good performance from three aspects, time-frequency domain, Mel-Frequency Cepstral Coefficients, and sparse representation were extracted, and a preliminary data set was created. Min-max normalization was used to align the numerical distribution of the data set, and a high-quality data set was created. Multi-classification models based on different classification algorithms and neural networks were trained, and the best-performing Random Forest was obtained, thus parameter optimization was carried out, and the optimal multi-classification model was obtained, achieving a classification accuracy of 91.14%. The visualization platform was built to process the classification results of the multi-classification model, completing majority voting processing and cough rate calculation, thereby achieving broiler health monitoring. In addition, the definitions of cough rate and prediction accuracy were newly proposed. A large number of experiments have verified the feasibility of the broiler health monitoring technology proposed in this paper, with an average prediction accuracy of 98.97% achieved.

Accurate prediction of drug combination risk levels based on relational graph convolutional network and multi-head attention

BackgroundAccurately identifying the risk level of drug combinations is of great significance in investigating the mechanisms of combination medication and adverse reactions. Most existing methods can only predict whether there is an interaction between two drugs, but cannot directly determine their accurate risk level.MethodsIn this study, we propose a multi-class drug combination risk prediction model named AERGCN-DDI, utilizing a relational graph convolutional network with a multi-head attention mechanism. Drug-drug interaction events with varying risk levels are modeled as a heterogeneous information graph. Attribute features of drug nodes and links are learned based on compound chemical structure information. Finally, the AERGCN-DDI model is proposed to predict drug combination risk level based on heterogenous graph neural network and multi-head attention modules.ResultsTo evaluate the effectiveness of the proposed method, five-fold cross-validation and ablation study were conducted. Furthermore, we compared its predictive performance with baseline models and other state-of-the-art methods on two benchmark datasets. Empirical studies demonstrated the superior performances of AERGCN-DDI.ConclusionsAERGCN-DDI emerges as a valuable tool for predicting the risk levels of drug combinations, thereby aiding in clinical medication decision-making, mitigating severe drug side effects, and enhancing patient clinical prognosis.

Development and validation of an automated machine learning model for the multi-class classification of diabetic retinopathy, central retinal vein occlusion and branch retinal vein occlusion based on color fundus photographs

IntroductionAutomated machine learning (AutoML) is a novel artificial intelligence (AI) strategy that enables clinicians without coding experience to develop their own AI models. This study assessed the discriminative performance of AutoML in differentiating diabetic retinopathy (DR), central retinal vein occlusion (CRVO) and branch retinal vein occlusion (BRVO) from normal fundi using color fundus photographs (CFPs). MethodsWe carried out AutoML model design using CFPs retrieved from a publicly available CFP data set (3200 labelled images). The retrieved CFPs were reviewed for quality and then uploaded to the Google Cloud Vertex AI platform for AutoML training and testing. We trained a multi-class classification model to differentiate DR, CRVO, BRVO from normal fundi using 875 CFPs and externally validated the model using 210 CFPs obtained from another dataset. Performance metrics, including area under receiver operator curve (AUROC) and sensitivity were reported. We then compared the AutoML model to state-of-the-art deep learning (DL)-based DR and RVO multi-class models identified through a literature search. ResultsOur AutoML model showed high discriminative performance in the multi-class classification of DR, CRVO and BRVO based on CFPs, with an AUROC, precision and recall reaching 0.995, 95.4% and 95.4% respectively at the 0.5 confidence threshold. The per-label sensitivity and specificity, respectively, were normal fundi (97.5%, 100%), DR (100%, 93.88%), CRVO (66.67%, 100%) and BRVO (71.43%, 98.73%). Our AutoML model generally showed similar performance to the state-of-the-art DL classifiers. ConclusionOur AutoML model can detect DR, CRVO, and BRVO in CFPs with good diagnostic accuracy and is a potentially useful screening tool.

Dynamic Exception Points for Fair Liver Allocation

There are disparities in access to livers based on transplant patients’ height, which disproportionately affects Hispanics, Asians, and women (across all ethnicities), because short patients can receive transplants from a smaller pool of available deceased donors for medical reasons. Reduced likelihood of transplantation leads to higher mortality rates and longer waiting times. We analyze fairness within the current U.S. liver allocation system where patients receive priority dynamically, based on their model for end-stage liver disease (MELD) scores, which reflect the severity of liver disease. We propose a simple adjustment, providing additional (exception) points based on height and MELD score, that can be easily implemented in practice, which materially reduces the disparity without sacrificing overall efficiency. We model the liver allocation system as a multiclass fluid model of overloaded queues with heterogeneous servers. We impose explicit equity constraints for all static patient classes, that is, height. We characterize the optimal solution under the objective of minimizing pretransplant mortality. The discretized version of the optimal policy is numerically solved using estimates from clinical data and a detailed simulation study demonstrates its effectiveness. The optimal policy, called the equity adjusted mortality risk policy, advocates ranking patients based on their short-term mortality risk adjusted for equity among height classes. Interpretation of the shadow prices of equity constraints in the optimal control problem as MELD exception points is novel in the transplant context since they can be seamlessly mapped into the existing system. Our simulations show that for women, the disparity can be almost completely eliminated. Hispanics and Asians greatly benefit from receiving these MELD exception points also. Our work provides a remedy to reduce the disparities in access to liver transplantation within the MELD-based allocation. Our approach can help the on-going analysis of the continuous distribution model for livers because it also considers aspects of candidate biology, notably height and body surface area. Funding: M. Akan was supported by the National Science Foundation [Grant CMMI-1334194] and the Carnegie Mellon University (CMU) [Onetto Fellowship in Operations Management]. Supplemental Material: The online appendix is available at https://doi.org/10.1287/serv.2023.0092 .