Feature Selection Research Articles

Grad-CAM-Based Feature Selection and Dementia Classification Algorithm Using Voice Data

Grad-CAM-Based Feature Selection and Dementia Classification Algorithm Using Voice Data

Enhanced Phishing URLs Detection using Feature Selection and Machine Learning Approaches

Enhanced Phishing URLs Detection using Feature Selection and Machine Learning Approaches

A neighborhood classifier based on adaptive radius selection and attribute reduction

Due to the capacity to process numerical data and generate highly interpretable results, neighborhood rough set theory has been widely used for data classification. However, there are two drawbacks that limit its performance. One is that the existing neighborhood classifiers suffer from insufficient information extraction due to the lack of a training process with feature selection. The other is that the neighborhood radius requires artificial parameters to adapt to different data distributions. To overcome the above shortcomings, a neighborhood classifier based on adaptive radius selection and attribute reduction (NRS-ASAR) is proposed. First, the K-nearest neighbor algorithm and the majority voting principle are introduced to eliminate noisy samples. Next, the training correlation radius is defined based on attribute reduction in which a relationship between the neighborhood radius and the features is established. Then, to overcome the subjectivity of the parameter in neighborhood classifier, an adaptive neighborhood radius called nearest neighborhood radius is defined. Finally, experimental results demonstrate that NRS-ASAR has better classification performance.

Derivation and Validation of Prediction of Preterm preeclampsia using machine learning algorithms.

To develop machine learning models for predicting preterm preeclampsia using the information available before 23 weeks' gestation. This was a secondary analysis of the Nulliparous Pregnancy Outcomes Study: Monitoring Mothers-to-Be cohort. We considered 131 features available before 23 weeks including maternal demographics, obstetrics and family history, social determinants of health, physical activity, nutrition, and early second-trimester ultrasound. Our primary outcome was preterm preeclampsia before 37 weeks. The dataset was randomly split into a training set (70%) and a validation set (30%). Machine learning models using glmnet, multilayer perceptron, random forest, XGBoost, and LightGBM models were developed. Using the machine learning approach that achieved the best area under the curves (AUC), we developed the final model. Further feature selection was conducted from the top 25 important features based on SHapley Additive exPlanations values. The performance of the final model was assessed using the validation dataset. Of 9,467 individuals, 219 (2.3%) had preterm preeclampsia. The AUC of the XGBoost model was the highest (AUC 0.749 [95% confidence interval 0.736-0.762]) compared with other models. Therefore, XGBoost was used to develop models using fewer variables. The XGBoost model with the 8 features (in order of importance: mean uterine artery pulsatility index in the early second trimester, chronic hypertension, pregestational diabetes, uterine artery notch, systolic and diastolic blood pressure in the first trimester, body mass index, and maternal age) was chosen as the final model as it had an AUC of 0.741 (95% confidence interval 0.730-0.752) which was not inferior to the original model (P=0.58). The final model in the validation dataset had an AUC of 0.779 (95% confidence interval 0.722-0.831). An online application of the final model was developed (https://kawakita.shinyapps.io/Preterm_preeclampsia/). Machine learning algorithms using information available before 23 weeks can accurately predict preterm preeclampsia before 37 weeks.

Automated derivation of diagnostic criteria for lung cancer using natural language processing on electronic health records: a pilot study

BackgroundThe digitisation of healthcare records has generated vast amounts of unstructured data, presenting opportunities for improvements in disease diagnosis when clinical coding falls short, such as in the recording of patient symptoms. This study presents an approach using natural language processing to extract clinical concepts from free-text which are used to automatically form diagnostic criteria for lung cancer from unstructured secondary-care data.MethodsPatients aged 40 and above who underwent a chest x-ray (CXR) between 2016 and 2022 were included. ICD-10 and unstructured data were pulled from their electronic health records (EHRs) over the preceding 12 months to the CXR. The unstructured data were processed using named entity recognition to extract symptoms, which were mapped to SNOMED-CT codes. Subsumption of features up the SNOMED-CT hierarchy was used to mitigate against sparse features and a frequency-based criteria, combined with univariate logarithmic probabilities, was applied to select candidate features to take forward to the model development phase. A genetic algorithm was employed to identify the most discriminating features to form the diagnostic criteria.Results75002 patients were included, with 1012 lung cancer diagnoses made within 12 months of the CXR. The best-performing model achieved an AUROC of 0.72. Results showed that an existing ‘disorder of the lung’, such as pneumonia, and a ‘cough’ increased the probability of a lung cancer diagnosis. ‘Anomalies of great vessel’, ‘disorder of the retroperitoneal compartment’ and ‘context-dependent findings’, such as pain, statistically reduced the risk of lung cancer, making other diagnoses more likely. The performance of the developed model was compared to the existing cancer risk scores, demonstrating superior performance.ConclusionsThe proposed methods demonstrated success in leveraging unstructured secondary-care data to derive diagnostic criteria for lung cancer, outperforming existing risk tools. These advancements show potential for enhancing patient care and results. However, it is essential to tackle specific limitations by integrating primary care data to ensure a more thorough and unbiased development of diagnostic criteria. Moreover, the study highlights the importance of contextualising SNOMED-CT concepts into meaningful terminology that resonates with clinicians, facilitating a clearer and more tangible understanding of the criteria applied.

Comparison of different machine learning methods in river streamflow estimation using isovel contours and hydraulic variables

ABSTRACT This study evaluates several machine learning (ML) models for estimating streamflow in the Osage River, Missouri, USA, and the Severn River, UK, using hydraulic input variables. These input variables are the flow section area (A), wetted perimeter (Pw ), water surface width (W), and velocity parameter (U) derived from isovel contours. Before using these models, the influential variables are selected using the sequential feature selection (SFS) method to reduce model complexity while maintaining performance. The two most important input variables identified were A and U, reducing the number of inputs from four to two. The results showed that the river’s flow conditions affect the accuracy of ML models used for estimating streamflow. Linear models such as MLR and ANFIS perform better in steady flow conditions (Osage River). In contrast, decision tree-based and non-linear models such as SVR with a radial basis kernel function (SVR_RBF) are better for unsteady flow conditions (Severn River). The finding suggests simpler models outperform complex deep-learning approaches (such as LSTM) for estimating streamflow by hydraulic variables. By selecting appropriate and efficient ML models, hydrometer stations can significantly improve the accuracy of streamflow estimation, leading to more informed decision-making in water management.

An improved conditional relevance and weighted redundancy feature selection method for gene expression data

An improved conditional relevance and weighted redundancy feature selection method for gene expression data

Advancements in evaporation prediction: introducing the Gated Recurrent Unit–Multi-Kernel Extreme Learning Machine (MKELM)–Gaussian Process Regression (GPR) model

Predicting evaporation is an essential topic in water resources management. It is critical to plan irrigation schedules, optimize hydropower production, and accurately calculate the overall water balance. Thus, researchers have developed many prediction models for predicting evaporation. Despite the development of these models, there are still unresolved challenges. These challenges include selecting the most important input parameters, handling nonstationary data, extracting critical information from data, and quantifying the uncertainty of predicted values. Thus, the main aim of this study is to address these challenges by developing a new prediction model. The new prediction model, named Gated Recurrent Unit–Multi-Kernel Extreme Learning Machine (MKELM)–Gaussian Process Regression (GPR), was used to predict one-month ahead evaporation in the Kashafrood basin, Iran. This model was executed in multiple stages. First, a feature selection algorithm was used to determine the most critical input parameters. A data processing technique was then employed to decompose nonstationary data into stationary intrinsic mode functions (IMFs). The GRU model then processed these components to extract their essential information. In the following step, the extracted information was inserted into the MKELM model to predict evaporation. Finally, the GPR model quantified the uncertainty of predicted values. Our research also introduces a new optimizer called the Salp Swarm Optimization Algorithm–Sine Cosine Optimization Algorithm. This algorithm was used to tune the model parameters. This algorithm's performance and the prediction models’ accuracy were evaluated using several error indices. According to the study results, the GRU–MKELM–GPR model performed better than other models in predicting monthly evaporation. It improved the training and testing mean absolute error values of the other models by 21%-43% and 8.2–33%, respectively. Moreover, the new model improved the R2 (R-squared or coefficient of determination) values of other models by 5–12%. Generally, the main findings of this paper included the superior performance of the new model in predicting evaporation data and the superior performance of a new optimizer in adjusting model parameters. These findings highlighted the effectiveness of the suggested model in addressing the challenges associated with evaporation prediction.

Interrelated dynamic biased feature selection and classification model using enhanced gorilla troops optimizer for intrusion detection

An Intrusion Detection System (IDS) is a valuable tool for network security since it can identify attacks, intrusions, and other types of illegal access. Excessive and irrelevant data slows down the classification process and eventually weakens the system's capacity to make informed decisions when IDS is monitoring a huge volume of network traffic. Innovative approaches are utilized to create large amounts of data and a lot of network traffic in order to test its effectiveness. A vital step in machine learning is feature selection. By determining which features are most essential for describing the dataset and its initial attributes, feature selection seeks to improve intrusion prediction performance while simultaneously delving deeper into the stored data. Making a feature selection is similar to fixing an optimization problem without a clear definition when users don't know where to begin. Finding the fewest characteristics required to describe the dataset, the original features, and to conduct classification is the primary purpose of feature selection, which also aims to improve prediction performance and acquire a deeper understanding of the stored data. As a result, researchers have been focusing on feature-selection issues recently, especially in light of the massive growth in available databases. Metaheuristic algorithms using a learning model have been the subject of studies to optimize feature selection difficulties. This research uses Enhanced Gorilla Troops Optimizer (EGTO), for enhancing the feature selection process and then performing classification. This research presents a Interrelated Dynamic Biased Feature Selection Model using Enhanced Gorilla Troops Optimizer (IDBFS-EGTO) for generation of feature vector set for intrusion detection. Despite its apparent success in handling a wide range of practical problems, it risks getting mired in local optima and premature convergence when faced with more difficult optimization challenges that is overcome with EGTO. The EGTO approach, which uses a collection of operators to strike a more steady equilibrium between exploitation and exploration. The proposed model generates relevant feature subset for machine learning model for accurate detection and classification of intrusions in the network. The proposed model achieved 98.4 % accuracy in intrusion detection and 98.6 % accuracy in EGTO optimization classification. The proposed model is improved by 3.8 % in feature weight allocation accuracy and 1.2 % in detection accuracy levels. The proposed model is compared with the traditional models and the results represent that the proposed model performance is high.

Feature selectivity and invariance in marsupial primary visual cortex.

A fundamental question in sensory neuroscience revolves around how neurons represent complex visual stimuli. In mammalian primary visual cortex (V1), neurons decode intricate visual features to identify objects, with most being selective for edge orientation, but with half of those also developing invariance to edge position within their receptive fields. Position invariance allows cells to continue to code an edge even when it moves around. Combining feature selectivity and invariance is integral to successful object recognition. Considering the marsupial-eutherian divergence 160million years ago, we explored whether feature selectivity and invariance was similar in marsupials and eutherians. We recovered the spatial filters and non-linear processing characteristics of the receptive fields of neurons in wallaby V1 and compared them with previous results from cat cortex. We stimulated the neurons in V1 with white Gaussian noise and analysed responses using the non-linear input model. Wallabies exhibit the same high percentage of orientation selective neurons as cats. However, in wallabies we observed a notably higher prevalence of neurons with three or more filters compared to cats. We show that having three or more filters substantially increases phase invariance in the V1s of both species, but that wallaby V1 accentuates this feature, suggesting that the species condenses more processing into the earliest cortical stage. These findings suggest that evolution has led to more than one solution to the problem of creating complex visual processing strategies. KEY POINTS: Previous studies have shown that the primary visual cortex (V1) in mammals is essential for processing complex visual stimuli, with neurons displaying selectivity for edge orientation and position. This research explores whether the visual processing mechanisms in marsupials, such as wallabies, are similar to those in eutherian mammals (e.g. cats). The study found that wallabies have a higher prevalence of neurons with multiple spatial filters in V1, indicating more complex visual processing. Using a non-linear input model, we demonstrated that neurons with three or more filters increase phase invariance. These findings suggest that marsupials and eutherian mammals have evolved similar strategies for visual processing, but marsupials have condensed more capacity to build phase invariance into the first step in the cortical pathway.

CT-based delta-radiomics for predicting pathological response to neoadjuvant immunochemotherapy in esophageal squamous cell carcinoma: a multicenter study

BackgroundThe study aimed to investigate the predictive value of delta-radiomics derived from computed tomography (CT) for pathological complete response (pCR) to neoadjuvant immunochemotherapy (NICT) among patients with esophageal squamous cell carcinoma (ESCC), helping clinicians determine whether to modify the neoadjuvant treatment strategy, proceed to surgery, or forgo surgery altogether.MethodsA total of 140 ESCC patients from two institutions (Database 1 = 93; Database 2 = 47) who underwent NICT and surgery were retrospectively included in the study. The training set consisted of patients from Database 1, while the testing set included patients from Database 2. All patients underwent contrast-enhanced CT scans before the start of the treatment and prior to the operation. The delta-radiomics features were calculated as the relative net change of radiomics features between the two-time points. Feature selection was conducted using Pearson correlation analysis, intraclass correlation coefficients, and the fivefold cross-validation with least absolute shrinkage and selection analysis. Four models were established, comprising a clinical model, a pre-treatment radiomics model, a delta-radiomics model, and a mixed model. Area under the curve (AUC) and decision curve analysis were used to assess the performance and the clinical value of the models.ResultsLess than half of the tumors (40/140, 28.6%) showed pCR following NICT. The delta-radiomics model displayed AUC of 0.827 and 0.790 in the training and testing set for predicting pCR, which was superior to the clinical model based on age and clinical tumor node metastasis (cTNM) stage (0.758 and 0.615) and the pre-treatment radiomics model (0.787 and 0.621). Furthermore, the delta-radiomics model demonstrated a more excellent AUC value in the testing set than the mixed model (0.847 and 0.719), which integrated clinical and delta-radiomics features.ConclusionsThe delta-radiomics model exhibited better diagnostic performance in preoperative prediction of pCR for NICT in ESCC patients compared to the clinical, pre-treatment radiomics, and mixed models.

NetQDA: Local Network-Guided High-Dimensioanl Quadratic Discriminant Analysis

Quadratic Discriminant Analysis (QDA) is a well-known and flexible classification method that considers differences between groups based on both mean and covariance structures. However, the connection structures of high-dimensional predictors are usually not explicitly incorporated into modeling. In this work, we propose a local network-guided QDA method that integrates the local connection structures of high-dimensional predictors. In the context of gene expression research, our method can identify genes that show differential expression levels as well as gene networks that exhibit different connection patterns between various biological state groups, thereby enhancing our understanding of underlying biological mechanisms. Extensive simulations and real data applications demonstrate its superior performance in both feature selection and outcome classification compared to commonly used discriminant analysis methods.

Development of a spectral repository for the identification of western Himalayan medicinal plants using machine learning techniques

The identification of medicinal plant species is a crucial task for assessing the status of our bioresources. Conventional methods primarily rely on taxonomy and laboratory-based instruments, which are time-consuming and require the requisite expertise. Thus, there is an escalating demand for efficient techniques that can quickly identify these precious species. The advent of Hyperspectral Remote Sensing (HRS) with artificial intelligence has significantly increased the scope of HRS techniques by offering rapid and precise plant identification. This study utilised non-imaging HRS handheld sensors to build a spectral repository for 10 important medicinal plant species from diverse locations across Indian Himalayan states, representing varying altitudinal and ecological conditions. The spectral repository encompasses 1237 distinct spectral signatures obtained from the leaves and canopies of the targeted plant species. Subsequently, an identification model has been developed using Random Forest (RF) with several feature selection methods, and it has been revealed that the RF model, coupled with wrapper-based feature selection, is an effective combination for classifying the targeted plant species. The calibration and test datasets accounted for accuracies of 87.87% and 91.39%, respectively, with corresponding kappa coefficients of 0.85 and 0.89. Furthermore, the developed RF model was applied to ‘PRISMA’ satellite data to identify Saussurea costus crops in farmers' croplands, achieving a classification accuracy of 81.31% and a kappa coefficient of 0.76. Therefore, the study highlights the potential of integrating RF, in-situ HRS, and satellite HRS for the non-destructive, precise, and accurate identification of medicinal plants that can significantly contribute to biodiversity conservation and sustainable resource management.

Application of machine learning in breast cancer survival prediction using a multimethod approach

Breast cancer is one of the most prevalent cancers with an increasing trend in both incidence and mortality rates in Iran. Survival analysis is a pivotal measure in setting appropriate care plans. To the best of our knowledge, this study is pioneering in Iran, introducing a multi-method approach using a Deep Neural Network (DNN) and 11 conventional machine learning (ML) methods to predict the 5 year survival of women with breast cancer. Supplying data from two centers comprising a total of 2644 records and incorporating external validation further distinguishes the study. Thirty-four features were selected based on a literature review and common variables in both datasets. Feature selection was also performed using a p value criterion (< 0.05) and a survey involving oncologists. A total of 108 models were trained. According to external validation, the DNN model trained with the Shiraz dataset, considering all features, exhibited the highest accuracy (85.56%). While the DNN model showed superior accuracy in external validation, it did not consistently achieve the highest performance across all evaluation metrics. Notably, models trained with the Shiraz dataset outperformed those trained with the Tehran dataset, possibly due to the lower number of missing values in the Shiraz dataset.

Integrating omics data and machine learning techniques for precision detection of oral squamous cell carcinoma: evaluating single biomarkers

IntroductionEarly detection of oral squamous cell carcinoma (OSCC) is critical for improving clinical outcomes. Precision diagnostics integrating metabolomics and machine learning offer promising non-invasive solutions for identifying tumor-derived biomarkers.MethodsWe analyzed a multicenter public dataset comprising 61 OSCC patients and 61 healthy controls. Plasma metabolomics data were processed to extract 29 numerical and 47 ratio features. The Extra Trees (ET) algorithm was applied for feature selection, and the TabPFN model was used for classification and prediction.ResultsThe model achieved an area under the curve (AUC) of 93% and an overall accuracy of 76.6% when using top-ranked individual biomarkers. Key metabolic features significantly differentiated OSCC patients from healthy controls, providing a detailed metabolic fingerprint of the disease.DiscussionOur findings demonstrate the utility of integrating omics data with advanced machine learning techniques to develop accurate, non-invasive diagnostic tools for OSCC. The study highlights actionable metabolic signatures that have potential applications in personalized therapeutics and early intervention strategies.

Enhancing Intrusion Detection Systems with Dimensionality Reduction and Multi-Stacking Ensemble Techniques

The deployment of intrusion detection systems (IDSs) is essential for protecting network resources and infrastructure against malicious threats. Despite the wide use of various machine learning methods in IDSs, such systems often struggle to achieve optimal performance. The key challenges include the curse of dimensionality, which significantly impacts IDS efficacy, and the limited effectiveness of singular learning classifiers in handling complex, imbalanced, and multi-categorical traffic datasets. To overcome these limitations, this paper presents an innovative approach that integrates dimensionality reduction and stacking ensemble techniques. We employ the LogitBoost algorithm with XGBRegressor for feature selection, complemented by a Residual Network (ResNet) deep learning model for feature extraction. Furthermore, we introduce multi-stacking ensemble (MSE), a novel ensemble method, to enhance attack prediction capabilities. The evaluation on benchmark datasets such as CICIDS2017 and UNSW-NB15 demonstrates that our IDS surpasses current models across various performance metrics.

Unveiling Lung Diseases in CT Scan Images With a Hybrid Bio‐Inspired Mutated Spider‐Monkey and Crow Search Algorithm

ABSTRACTBio‐inspired computer‐aided diagnosis (CAD) has garnered significant attention in recent years due to the inherent advantages of bio‐inspired evolutionary algorithms (EAs) in handling small datasets with elevated precision and reduced computational complexity. Traditional CAD models face limitations as they can only be developed post‐outbreak, relying on datasets that become available during such events such as the COVID‐19 pandemic. The scarcity of data for emerging diseases poses a substantial challenge to achieving elevated precision with conventional deep‐learning algorithms. Furthermore, even when datasets are available, employing deep learning for class‐based classification is arduous, necessitating model retraining, in this paper, we propose a novel hybrid algorithm that leverages the strengths of the crow search algorithm (CSA) and the spider monkey optimization (SMO) algorithm to create an optimised spider monkey crow search (OSM‐CS) algorithm. We developed a CAD tool that maps each input CT image to a high‐dimensional vector by extracting four categories of features: high contrast, polynomial decomposition, textural, and pixel statistics. The proposed OSM‐CS algorithm is employed as a feature selection method. Our experimental results demonstrate the effectiveness of the OSM‐CS algorithm, achieving an impressive accuracy of 98.2% when coupled with an AdaBoost classifier for multi‐class classification and 99.93% for binary classification. This performance surpasses that of state‐of‐the‐art (SOTA) deep learning models and recently published hybrid algorithms, underscoring the potential of the OSM‐CS algorithm as a powerful tool in the realm of CAD.

DeepAIPs-Pred: Predicting Anti-Inflammatory Peptides Using Local Evolutionary Transformation Images and Structural Embedding-Based Optimal Descriptors with Self-Normalized BiTCNs.

Inflammation is a biological response to harmful stimuli, playing a crucial role in facilitating tissue repair by eradicating pathogenic microorganisms. However, when inflammation becomes chronic, it leads to numerous serious disorders, particularly in autoimmune diseases. Anti-inflammatory peptides (AIPs) have emerged as promising therapeutic agents due to their high specificity, potency, and low toxicity. However, identifying AIPs using traditional in vivo methods is time-consuming and expensive. Recent advancements in computational-based intelligent models for peptides have offered a cost-effective alternative for identifying various inflammatory diseases, owing to their selectivity toward targeted cells with low side effects. In this paper, we propose a novel computational model, namely, DeepAIPs-Pred, for the accurate prediction of AIP sequences. The training samples are represented using LBP-PSSM- and LBP-SMR-based evolutionary image transformation methods. Additionally, to capture contextual semantic features, we employed attention-based ProtBERT-BFD embedding and QLC for structural features. Furthermore, differential evolution (DE)-based weighted feature integration is utilized to produce a multiview feature vector. The SMOTE-Tomek Links are introduced to address the class imbalance problem, and a two-layer feature selection technique is proposed to reduce and select the optimal features. Finally, the novel self-normalized bidirectional temporal convolutional networks (SnBiTCN) are trained using optimal features, achieving a significant predictive accuracy of 94.92% and an AUC of 0.97. The generalization of our proposed model is validated using two independent datasets, demonstrating higher performance with the improvement of ∼2 and ∼10% of accuracies than the existing state-of-the-art model using Ind-I and Ind-II, respectively. The efficacy and reliability of DeepAIPs-Pred highlight its potential as a valuable and promising tool for drug development and research academia.

Enhanced Occupational Safety in Agricultural Machinery Factories: Artificial Intelligence-Driven Helmet Detection Using Transfer Learning and Majority Voting

The objective of this study was to develop an artificial intelligence (AI)-driven model for the detection of helmet usage among workers in tractor and agricultural machinery factories with the aim of enhancing occupational safety. A transfer learning approach was employed, utilizing nine pre-trained neural networks for the extraction of deep features. The following neural networks were employed: MobileNetV2, ResNet50, DarkNet53, AlexNet, ShuffleNet, DenseNet201, InceptionV3, Inception-ResNetV2, and GoogLeNet. Subsequently, the extracted features were subjected to iterative neighborhood component analysis (INCA) for feature selection, after which they were classified using the k-nearest neighbor (kNN) algorithm. The classification outputs of all networks were combined through iterative majority voting (IMV) to achieve optimal results. To evaluate the model, an image dataset comprising 662 images of individuals wearing helmets and 722 images of individuals without helmets sourced from the internet was constructed. The proposed model achieved an accuracy of 90.39%, with DenseNet201 producing the most accurate results. This AI-driven helmet detection model demonstrates significant potential in improving occupational safety by assisting safety officers, especially in confined environments, reducing human error, and enhancing efficiency.

A Multimodal Machine Learning Model in Pneumonia Patients Hospital Length of Stay Prediction

Hospital overcrowding, driven by both structural management challenges and widespread medical emergencies, has prompted extensive research into machine learning (ML) solutions for predicting patient length of stay (LOS) to optimize bed allocation. While many existing models simplify the LOS prediction problem to a classification task, predicting broad ranges of hospital days, an exact day-based regression model is often crucial for precise planning. Additionally, available data are typically limited and heterogeneous, often collected from a small patient cohort. To address these challenges, we present a novel multimodal ML framework that combines imaging and clinical data to enhance LOS prediction accuracy. Specifically, our approach uses the following: (i) feature extraction from chest CT scans via a convolutional neural network (CNN), (ii) their integration with clinically relevant tabular data from patient exams, refined through a feature selection system to retain only significant predictors. As a case study, we applied this framework to pneumonia patient data collected during the COVID-19 pandemic at two hospitals in Naples, Italy—one specializing in infectious diseases and the other general-purpose. Under our experimental setup, the proposed system achieved an average prediction error of only three days, demonstrating its potential to improve patient flow management in critical care environments.