Stack Model Research Articles

Explainable machine learning methods to predict postpartum depression risk

Postpartum depression (PPD) is a type of depression that mothers have following childbirth due to hormonal changes, psychological transition to parenting, and exhaustion. This depression strikes either during/or in the first year following childbirth. It is also a frequently disregarded medical condition that must be treated right away as it might have major repercussions. Machine learning (ML) and artificial intelligence (AI) are tools that healthcare professionals can utilize to anticipate this condition more rapidly and correctly. Consequently, we have demonstrated how to use explainable artificial intelligence (XAI) methods and heterogeneous classifiers to predict postpartum depression in mothers who have recently given birth. The K-Nearest Neighbor (KNN) model and the customized stack model outperformed all other classifiers. KNN model obtained 97% accuracy, 98% recall, and 95% precision and the stack model obtained 97% accuracy, 100% recall, and 94% precision, respectively. A set of frameworks and resources known as explainable artificial intelligence (XAI) facilitates the comprehension and interpretation of predictions made by machine learning algorithms. Four distinct XAI techniques: ELI5, Shapley Additive Values (SHAP), Local Interpretable Model-agnostic Explanations (LIME), and Anchor – have been used to interpret the model predictions. Explainability, interpretability, accountability, and transparency are crucial parameters of XAI, ensuring that machine learning models provide understandable and trustworthy results to users and stakeholders. The goal of this interdisciplinary research is to develop an automated diagnosis framework with tools that can transform therapy for postpartum depression leading to suicide attempts and empower medical professionals to offer mothers individualized, high-quality care.

Monitoring soil salinity in coastal wetlands with Sentinel-2 MSI data: Combining fractional-order derivatives and stacked machine learning models

Monitoring soil salinity is essential for understanding the behavior of coastal wetland ecosystems and implementing effective management strategies. Despite the advantages of the Multi-Spectral Instrument (MSI) data for large-scale, high-frequency soil salinity monitoring, challenges remain in data preprocessing and model construction. We combined fractional-order derivative (FOD) technology with stacked machine learning models to monitor and map soil salinity using Sentinel-2 MSI data. The base models included Elastic Net Regression, Support Vector Regression, Artificial Neural Network, Extreme Gradient Boosting, and Random Forest, with Non-Negative Least Squares as the meta-learner. The results showed that low-order FOD enhanced image gradients and maintained a high peak signal-to-noise ratio, thereby improving the correlation with soil salinity. Notably, the 0.25-order FOD showed the best performance, increasing the correlation coefficient with soil salinity by up to 13 %. The stacked machine learning models effectively combined the strengths of different base models, enhancing prediction accuracy by more than 8 % compared to single models. Furthermore, combining stacked models with FOD further improved prediction accuracy, with an increase in R² of up to 9 %. The combination of 0.25-order FOD and the stacked machine learning model achieved the best performance (R² = 0.82, RMSE = 10.19 ppt, RPD = 2.38, RPIQ = 4.69). This approach provides a reference for rapid and effective large-scale digital mapping of soil salinity in coastal wetlands.

Prediction of Multi-Pharmacokinetics Property in Multi-Species: Bayesian Neural Network Stacking Model with Uncertainty.

Pharmacokinetic (PK) properties of a drug are vital attributes influencing its therapeutic effectiveness, playing an important role in the drug development process. Focusing on the difficult task of predicting PK parameters, we compiled an extensive data set comprising parameters across multiple species. Building upon this groundwork, we introduced the PKStack ensemble model to predict PK parameters across diverse species. PKStack integrates a variety of base models and includes uncertainty in its predictions. We also manually collected PK data from animals as an external test set. We predicted a total of 45 tasks for nine PK parameters in five species, and in general, the prediction accuracy was better for intravenous injections, including parameters such as human Vd (R2 = 0.72, RMSE = 0.31), human CL (R2 = 0.52, RMSE = 0.32), and others. In addition to predictive accuracy, we also considered the interpretability of the results and the definition of the model's application domain. Based on the findings, our model has great potential for practical applications in drug discovery.

Comparative analysis of climate-induced habitat shift of economically significant species with diverse ecological preferences in the Northwest Pacific

The Northwest Pacific Ocean is the most productive fishing ground in the Pacific Ocean, with a continuous rise in water temperature since 1990. We developed stacked species distribution models (SSDMs) to estimate the impacts of climate change on the distribution dynamics of economically significant species under three climate change scenarios for the periods 2040-2060 and 2080-2100. Overall, water temperature is the most important factor in shaping the distribution patterns of species, followed by water depth. The predictive results indicate that all the species show a northward migration in the future, and the migration distance varies greatly among species. Most pelagic species will expand their habitats under climate change, implying their stronger adaptability than benthic species. Tropical fishes are more adaptable to climate change than species in other climate zones. Though limitations existed, our study provided baseline information for designing a climate-adaptive, dynamic fishery management strategy for maintaining sustainable fisheries.

A Real-Time End-to-End Framework with a Stacked Model Using Ultrasound Video for Cardiac Septal Defect Decision-Making

Echocardiography is the gold standard for the comprehensive diagnosis of cardiac septal defects (CSDs). Currently, echocardiography diagnosis is primarily based on expert observation, which is laborious and time-consuming. With digitization, deep learning (DL) can be used to improve the efficiency of the diagnosis. This study presents a real-time end-to-end framework tailored for pediatric ultrasound video analysis for CSD decision-making. The framework employs an advanced real-time architecture based on You Only Look Once (Yolo) techniques for CSD decision-making with high accuracy. Leveraging the state of the art with the Yolov8l (large) architecture, the proposed model achieves a robust performance in real-time processes. It can be observed that the experiment yielded a mean average precision (mAP) exceeding 89%, indicating the framework’s effectiveness in accurately diagnosing CSDs from ultrasound (US) videos. The Yolov8l model exhibits precise performance in the real-time testing of pediatric patients from Mohammad Hoesin General Hospital in Palembang, Indonesia. Based on the results of the proposed model using 222 US videos, it exhibits 95.86% accuracy, 96.82% sensitivity, and 98.74% specificity. During real-time testing in the hospital, the model exhibits a 97.17% accuracy, 95.80% sensitivity, and 98.15% specificity; only 3 out of the 53 US videos in the real-time process were diagnosed incorrectly. This comprehensive approach holds promise for enhancing clinical decision-making and improving patient outcomes in pediatric cardiology.

Landslide Displacement Prediction Stacking Deep Learning Algorithms: A Case Study of Shengjibao Landslide in the Three Gorges Reservoir Area of China

Computational models enable accurate, timely prediction of landslides based on the monitoring data on-site as the development of artificial intelligence technology. The most existing prediction methods focus on finding a single prediction algorithm with excellent performance or an integrated and efficient hyperparameter optimization algorithm with a highly accurate regression prediction algorithm. In order to break through the limitation of generalization of prediction models, this paper proposes an ensemble model that combines deep learning algorithms, with a stacking framework optimized with the sliding window method. Multiple deep learning algorithms are set as the first layer of the stacking framework, which is optimized with the sliding window method to avoid confusion in the time order of datasets based on time series analysis. The Shengjibao landslide in the Three Gorges Reservoir is used as a case study. First, the cumulative displacement is decomposed into a trend and a periodic term using a moving average method. A single-factor and a multi-factor superposition model based on multiple deep learning algorithms are used to predict the trend and periodic term of the displacement, respectively. Finally, the predicted values of the trend and periodic terms are added to obtain the total predicted landslide displacement. For monitoring point ZK2-3, the values of RMSE and MAPE of the total displacement prediction with the stacking model are 15.93 mm and 0.54%, and the values of RMSE and MAPE of the best-performing individual deep learning model are 20.00 mm and 0.64%. The results show that the stacking model outperforms other models by combining the advantages of each individual deep learning algorithm. This study provides a framework for integrating landslide displacement prediction models. It can serve as a reference for the geological disaster prediction and the establishment of an early warning system in the Three Gorges Reservoir Area.

A Stacked Multimodality Model Based on Functional MRI Features and Deep Learning Radiomics for Predicting the Early Response to Radiotherapy in Nasopharyngeal Carcinoma

BackgroundThis study aimed to construct and assess a comprehensive model that integrates MRI-derived deep learning radiomics, functional imaging (fMRI), and clinical indicators to predict early efficacy of radiotherapy in nasopharyngeal carcinoma (NPC). MethodsThis retrospective study recruited NPC patients with radiotherapy from two Chinese hospitals between October 2018 and July 2022, divided into a training set (hospital I, 194 cases), an internal validation set (hospital I, 82 cases), and an external validation set (hospital II, 40 cases). We extracted 3404 radiomic features and 2048 deep learning characteristics from MRI modalities - T1WI, CE-T1WI, T2WI, and T2WI/FS. Additionally, both the ADC and its maximum (ADCmax) from DWI imaging, along with TBF and its maximum (TBFmax) from ASL imaging were derived. We used four classifiers (LR, XGBoost, SVM and KNN) and stacked algorithm as model construction methods. The receiver operating characteristic (ROC) curve (AUC) and decision curve analysis (DCA) was used to assess models. ResultsThe manual radiomics model leveraging XGBoost and the deep learning model based on KNN (the AUCs in the training set: 0.909, 0.823, respectively) showed better predictive efficacy than other machine learning algorithms. The stacked model that integrated MRI-based deep learning radiomics, fMRI, and hematological indicators, has a large improvement of AUC in the training set [0.984 (95%CI: 0.972 - 0.996)], the internal validation set [0.936 (95%CI: 0.885 - 0.987)], and the external validation set [0.959 (95%CI: 0.901-1)]. ConclusionsOur research has developed a clinical-radiomics integrated model based on MRI which can predict early radiotherapy response in NPC and provide guidance for personalized treatment.

Enhanced Phishing Detection: An Ensemble Stacking Model with DT-RFECV and SMOTE

Enhanced Phishing Detection: An Ensemble Stacking Model with DT-RFECV and SMOTE

Parameter Estimation of Proton Exchange Membrane Fuel Cells Using Chaotic Newton-Raphson-Based Optimizer

This paper presents a novel improved metaheuristic algorithm, Chaotic Newton-Raphson-Based Optimizer (CNRBO), for the Proton Exchange Membrane Fuel Cells (PEMFCs) parameter extraction problem. In this study, ten distinct chaotic maps are integrated within the Newton-Raphson-based optimizer (NRBO) to improve its performance. The chaotic maps are employed to define the probability of the trap avoidance operator (TAO). The proposed CNRBO algorithm is validated using standard benchmark optimization problems. Results proved its advantages over the standard NRBO algorithm in terms of accuracy, robustness, and convergence speed. For PEMFC parameters extraction using the proposed CNRBO algorithm, the objective function to be minimized is the sum of squared errors (SSE) between estimated and measured stack voltages across various data points. To validate the effectiveness and reliability of the proposed CNRBO, its performance is compared with recent algorithms on eight different PEMFC stack models including NedStackPS6, BCS 500W, Horizon 500W, 250W stack, Avista SR-12 500W, Temasek 1KW, Ballard Mark V 5KW and Horizon H-1000XP stack. Statistical measures are employed to assess the superiority and robustness of the proposed CNRBO. Statistical analyses demonstrate that the proposed CNRBO algorithm outperforms existing algorithms in terms of accuracy, search capability, and convergence speed, solidifying its position as a powerful tool for PEMFC parameter estimation.

Stack-AVP: A Stacked Ensemble Predictor Based on Multi-view Information for Fast and Accurate Discovery of Antiviral Peptides

AVPs, or antiviral peptides, are short chains of amino acids capable of inhibiting viral replication, preventing viral entry, or disrupting viral membranes. They represent a promising area of research for developing new antiviral therapies due to their potential to target a broad spectrum of viruses, incorporating those resistant to traditional antiviral drugs. However, traditional experimental methods for identifying AVPs are often costly and labour-intensive. Thus far, multiple computational methods have been introduced for the in silico identification of AVPs, but these methods still have certain shortcomings. In this study, we propose a novel stacked ensemble learning framework, termed Stack-AVP, for fast and accurate AVP identification. In Stack-AVP, we investigated heterogeneous prediction models, which were trained with 12 commonly used machine learning algorithms coupled with a wide range of multiple feature encoding schemes. Subsequently, these prediction models were adopted to generate multi-view features providing class information and probability information. Finally, we applied our feature selection method to determine the best feature subset for the construction of the final stacked model. Comparative assessments on the independent test dataset revealed that Stack-AVP surpassed the performance of current state-of-the-art methods, with an accuracy of 0.930, MCC of 0.860, and AUC of 0.975. Furthermore, it was found that our multi-view features exhibited a crucial mechanism to improve the prediction performance of AVPs. To facilitate experimental scientists in performing high-throughput identification of AVPs, the prediction sever Stack-AVP is publicly accessible at https://pmlabqsar.pythonanywhere.com/Stack-AVP.

Hyperspectral Inversion of Soil Cu Content in Agricultural Land Based on Continuous Wavelet Transform and Stacking Ensemble Learning

Heavy metal pollution in agricultural land poses significant threats to both the ecological environment and human health. Therefore, the rapid and accurate prediction of heavy metal content in agricultural soil is crucial for environmental protection and soil remediation. Acknowledging the limitations of traditional single linear or nonlinear machine learning models in terms of prediction accuracy, this study developed an ensemble learning model that integrates multiple linear or nonlinear learning models with a random forest (RF) model to improve both the prediction accuracy and reliability. In this study, we selected a typical copper (Cu) polluted area in the Pearl River Delta of Guangdong Province as the research site and collected Cu content data and indoor soil reflectance spectral data from 269 surface soil samples. First, the soil spectral data were preprocessed using Savitzky–Golay (SG) smoothing, multiplicative scattering correction (MSC), and continuous wavelet transform (CWT) to reduce noise interference. Next, principal components analysis (PCA) was employed to reduce the dimensionality of the preprocessed spectral data, eliminating redundant features and lowering the computational complexity. Finally, based on the dimensionality-reduced data and Cu content, we established a stacked ensemble learning model, where the base models included SVR, PLSR, BPNN, and XGBoost, with RF serving as the meta-model to estimate the soil heavy metal content. To evaluate the performance of the stacking model, we compared its prediction accuracy with that of individual models. The results indicate that, compared to the traditional machine learning models, the prediction accuracy of the stacking model was superior (R2 = 0.77; RMSE = 7.65 mg/kg; RPD = 2.29). This suggests that the integrated algorithm demonstrates a greater robustness and generalization capability. This study presents a method to improve soil heavy metal content estimation using hyperspectral technology, ensuring a robust model that supports policymakers in making informed decisions about land use, agriculture, and environmental protection.

The short-term wind power prediction based on a multi-layer stacked model of BO-CNN-BiGRU-SA

Wind power generation is influenced by various meteorological factors, exhibiting significant volatility and unpredictability. This variability presents considerable challenges for accurate wind power forecasting. In this study, we propose an innovative method for short-term wind power prediction that integrates a Bayesian-optimized Convolutional Neural Network (CNN), Bidirectional Gated Recurrent Units (BiGRU), and a Self-Attention Mechanism (SA) within a multi-layer architecture. Initially, we preprocess features using Pearson correlation analysis and input them into the CNN to investigate complex nonlinear spatial relationships among multiple feature variables and the current load. Subsequently, the BiGRU captures long-term dependencies from both forward and backward time sequences. Finally, we implement the Self-Attention Mechanism to weigh the features and generate the predicted wind power. We optimize the model's numerous hyperparameters utilizing a Bayesian algorithm. Through comparative ablation experiments with varying time segment lengths on wind farm datasets from four regions, our method significantly outperforms 11 models, including Long Short-Term Memory (LSTM), and surpasses several state-of-the-art (SOTA) prediction models, such as iTransformer, PatchTST, Non-stationary Transformers, TSMixer, and DLinear. The highest coefficient of determination (R²) achieved was 0.981, with the Symmetric Mean Absolute Percentage Error (SMAPE), Root Mean Square Error (RMSE), and Mean Absolute Error (MAE) decreasing by 11.22% to 62.04% compared to other models. The results demonstrate the predictive accuracy and generalization performance of our proposed model.

A Stacked Neural Network Model for Damage Localization.

Traditional vibration-based damage detection methods often involve human intervention in decision-making, therefore being time-consuming and error-prone. In this study, we propose using Artificial Neural Networks (ANNs) to detect patterns in the structural response and create accurate predictions. The features extracted from the response signal are the Relative Frequency Shifts (RFSs) of the first eight weak-axis bending vibration modes, and the predictions refer to the damage location. To increase the accuracy of the predictions, we propose a novel stacked neural network approach, capable of detecting damage locations with high accuracy. The dataset used for training involves, as input data, the RFSs calculated with an original method for numerous damage locations and severities. The following models were used as building blocks for our stacked approach: Multilayer Perceptron, Recurrent Neural Network, Long Short-term Memory, and Gated Recurrent Units. The entire beam was thus split into segments and each network was trained in this stacked model on one beam segment. All results obtained with the models are also compared to a standard neural network trained on the entire beam. The results obtained show that the model that performs the best contains 14 stacked two-layer feedforward networks.

The cascade integration model based on machine learning to predict gestational diabetes

Abstract Machine learning has significant superiority in disease prediction and auxiliary clinical decisions due to its powerful data analysis and exploration capabilities. Making accurate predictions about gestational diabetes mellitus (GDM) with clinical data is not always an easy task. In previous work, GDM was predicted using machine learning models and integrated learning models such as Stacking. This work improves on previous work by developing a new integrated learning model building approach for better exploiting the benefits of machine learning models. Initially, the clinical data set is normalized. Then, according to the principle of removing the redundant features of each machine learning model, the first nine high-importance features of the five single models are filtered respectively. Finally, the GDM Cascade integration prediction model is constructed and compared with the Blending model and Stacking model, it is obvious that the proposed model construction method has superior performance and the AUC value reaches 0.9536.

Smart grid stability prediction using Adaptive Aquila Optimizer and ensemble stacked BiLSTM

BackgroundSmart grids, characterized by their ability to integrate renewable energy sources and manage the dynamic balance between supply and demand, require sophisticated prediction models to maintain stability. Traditional machine learning (ML) models often fall short in predicting the highly variable nature of smart grid operations. MethodsThis study introduces an ensemble stacked bidirectional Long Short-Term Memory model enhanced by a proposed Adaptive Aquila Optimizer (AAO). The AAO uses a Sigmoid Factor to balance exploration and exploitation, adapting the transition from broad searches to focused ones based on iteration progress. It is utilized for feature selection by identifying and excluding irrelevant and redundant features and methodically evaluates seven key hyperparameters to fine-tune the model's performance. Additionally, a weighted voting mechanism is employed to aggregate predictions in the ensemble model. ResultsMultiple rounds of empirical experiments using different sets of optimizers and configurations, supported by the visualization capabilities of TensorBoard, demonstrate significant improvements in the performance of the AAO-BiLSTM model. The results show profound potential with accuracy, precision, recall, and F1-score rates of 99.55%, surpassing both traditional ML algorithms and state-of-the-art approaches.

Prompt Recovery for Large Language Models

Understanding and recovering prompts in large language models (LLMs) is vital for addressing concerns related to privacy, copyright, and beyond. However, there is a lack of extensive research in this area. To fill this gap, we implemented model stacking techniques, such as utilizing mean prompts and embedding models, tailored for specific datasets. While these individual models were designed for particular datasets, our combined stacking model demonstrated improved accuracy in prompt recovery across diverse datasets. Although there was a slight decline in performance on the initial dataset, our comprehensive evaluation across multiple LLMs and prompt benchmarks indicates that our stacking model exceeds the performance of existing methods. Notably, this approach uses a single LLM without depending on external resources, making it an efficient and accessible solution for prompt recovery.

PredPVP: A Stacking Model for Predicting Phage Virion Proteins Based on Feature Selection Methods

Background: Phage therapy has a broad application prospect as a novel therapeutic method, and Phage Virion Proteins (PVP) can recognize the host and bind to surface receptors, which is of great significance for the development of antimicrobial drugs for the treatment of infectious diseases caused by bacteria. In recent years, several PVP predictors based on machine learning have been developed, which usually use a single feature to train the learner. In contrast, higher dimensional feature representations tend to contain more potential sequence information. Methods: In this work, we construct a stacking model PredPVP for PVP prediction by combining multiple features and using feature selection methods. Specifically, the sequence is first encoded using seven features. For this high-dimensional feature representation, three feature selection methods wereutilized to remove redundant features, then integrated with eight machine learning algorithms. Finally, probability features and class features (PCFs) generated by 24 base models were put into logistic regression (LR) to train the model. Results: The results of the independent test set indicate that PredPVP has higher performance compared to other existing predictors, with an AUC of 93.4%. Conclusion: We expect PredPVP to be used as a tool for large-scale PVP recognition, providing a new way for the development of novel antimicrobials and accelerating its application in actual treatment. The datasets and source codes used in this study are available at https://github.com/caoqian23/PredPVP.

A hybrid transformer and attention based recurrent neural network for robust and interpretable sentiment analysis of tweets

Sentiment analysis is a pivotal tool in understanding public opinion, consumer behavior, and social trends, underpinning applications ranging from market research to political analysis. However, existing sentiment analysis models frequently encounter challenges related to linguistic diversity, model generalizability, explainability, and limited availability of labeled datasets. To address these shortcomings, we propose the Transformer and Attention-based Bidirectional LSTM for Sentiment Analysis (TRABSA) model, a novel hybrid sentiment analysis framework that integrates transformer-based architecture, attention mechanism, and recurrent neural networks like BiLSTM. The TRABSA model leverages the powerful RoBERTa-based transformer model for initial feature extraction, capturing complex linguistic nuances from a vast corpus of tweets. This is followed by an attention mechanism that highlights the most informative parts of the text, enhancing the model’s focus on critical sentiment-bearing elements. Finally, the BiLSTM networks process these refined features, capturing temporal dependencies and improving the overall sentiment classification into positive, neutral, and negative classes. Leveraging the latest RoBERTa-based transformer model trained on a vast corpus of 124M tweets, our research bridges existing gaps in sentiment analysis benchmarks, ensuring state-of-the-art accuracy and relevance. Furthermore, we contribute to data diversity by augmenting existing datasets with 411,885 tweets from 32 English-speaking countries and 7,500 tweets from various US states. This study also compares six word-embedding techniques, identifying the most robust preprocessing and embedding methodologies crucial for accurate sentiment analysis and model performance. We meticulously label tweets into positive, neutral, and negative classes using three distinct lexicon-based approaches and select the best one, ensuring optimal sentiment analysis outcomes and model efficacy. Here, we demonstrate that the TRABSA model outperforms the current seven traditional machine learning models, four stacking models, and four hybrid deep learning models, yielding notable gain in accuracy (94%) and effectiveness with a macro average precision of 94%, recall of 93%, and F1-score of 94%. Our further evaluation involves two extended and four external datasets, demonstrating the model’s consistent superiority, robustness, and generalizability across diverse contexts and datasets. Finally, by conducting a thorough study with SHAP and LIME explainable visualization approaches, we offer insights into the interpretability of the TRABSA model, improving comprehension and confidence in the model’s predictions. Our study results make it easier to analyze how citizens respond to resources and events during pandemics since they are integrated into a decision-support system. Applications of this system provide essential assistance for efficient pandemic management, such as resource planning, crowd control, policy formation, vaccination tactics, and quick reaction programs.

Mobile Application and Machine Learning-Driven Scheme for Intelligent Diabetes Progression Analysis and Management Using Multiple Risk Factors

Diabetes mellitus is a chronic disease that affects over 500 million people worldwide, necessitating personalized health management programs for effective long-term control. Among the various biomarkers, glycated hemoglobin (HbA1c) is a crucial indicator for monitoring long-term blood glucose levels and assessing diabetes progression. This study introduces an innovative approach to diabetes management by integrating a mobile application and machine learning. We designed and implemented an intelligent application capable of collecting comprehensive data from diabetic patients, creating a novel diabetes dataset named DiabMini with 127 features of 88 instances, including medical information, personal information, and detailed nutrient intake and lifestyle. Leveraging the DiabMini, we focused the analysis on HbA1c dynamics due to their clinical significance in tracking diabetes progression. We developed a stacking model combining eXtreme Gradient Boosting (XGBoost), Support Vector Classifier (SVC), Extra Trees (ET), and K-Nearest Neighbors (KNN) to explore the impact of various influencing factors on HbA1c dynamics, which achieved a classification accuracy of 94.23%. Additionally, we applied SHapley Additive exPlanations (SHAP) to visualize the contributions of risk factors to HbA1c dynamics, thus clarifying the differential impacts of these factors on diabetes progression. In conclusion, this study demonstrates the potential of integrating mobile health applications with machine learning to enhance personalized diabetes management.

Auto encoder-based defense mechanism against popular adversarial attacks in deep learning.

Convolutional Neural Network (CNN)-based models are prone to adversarial attacks, which present a significant hurdle to their reliability and robustness. The vulnerability of CNN-based models may be exploited by attackers to launch cyber-attacks. An attacker typically adds small, carefully crafted perturbations to original medical images. When a CNN-based model receives the perturbed medical image as input, it misclassifies the image, even though the added perturbation is often imperceptible to the human eye. The emergence of such attacks has raised security concerns regarding the implementation of deep learning-based medical image classification systems within clinical environments. To address this issue, a reliable defense mechanism is required to detect adversarial attacks on medical images. This study will focus on the robust detection of pneumonia in chest X-ray images through CNN-based models. Various adversarial attacks and defense strategies will be evaluated and analyzed in the context of CNN-based pneumonia detection. From earlier studies, it has been observed that a single defense mechanism is usually not effective against more than one type of adversarial attack. Therefore, this study will propose a defense mechanism that is effective against multiple attack types. A reliable defense framework for pneumonia detection models will ensure secure clinical deployment, facilitating radiologists and doctors in their diagnosis and treatment planning. It can also save time and money by automating routine tasks. The proposed defense mechanism includes a convolutional autoencoder to denoise perturbed Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD) adversarial images, two state- of-the-art attacks carried out at five magnitudes, i.e., ε (epsilon) values. Two pre-trained models, VGG19 and VGG16, and our hybrid model of MobileNetV2 and DenseNet169, called Stack Model, have been used to compare their results. This study shows that the proposed defense mechanism outperforms state-of-the-art studies. The PGD attack using the VGG16 model shows a better attack success rate by reducing overall accuracy by up to 67%. The autoencoder improves accuracy by up to 16% against PGD attacks in both the VGG16 and VGG19 models.