Evaluation of Machine Learning Methods Developed for Prediction and Diagnosis of Pneumonia: A Systematic Review

ABSTRACT

Background and Aims

With the increasing prevalence of pneumonia, machine learning (ML) models have been increasingly utilized to diagnose, predict, and treat pneumonia due to their ability to manage complex datasets. This systematic review evaluates the performance and quality of ML models developed for pneumonia prediction, diagnosis, and treatment, following the statistical reporting guidelines of Assel et al. (2018).

Methods

On 15 January 2024, a systematic review was conducted in PubMed, Scopus, Web of Science, and Google Scholar using the PRISMA checklist. Articles developing or validating ML models for pneumonia were included. Performance metrics, including accuracy, sensitivity, specificity, and area under the curve (AUC), were extracted with confidence intervals where available.

Results

Of 11,545 screened articles, 42 studies evaluating 125 ML models were included. For pneumonia diagnosis, DenseNet achieved the highest accuracy of 94% (95% CI: 92%–96%), while Random Forest and XGBoost were the most effective for prediction, with AUCs of 0.96 (95% CI: 0.94–0.98) and 0.97 (95% CI: 0.95–0.99), respectively. Neural networks (n = 15) showed a peak accuracy of 98.94% (95% CI: 97.5%–99.5%), and ResNet demonstrated superior performance with an accuracy of 99.63% (95% CI: 98.8%–99.9%). Potential biases in datasets and limited generalizability were noted.

Conclusion

ML algorithms significantly improve pneumonia diagnosis and prediction, optimizing clinical decision‐making. However, data set biases and generalizability challenges highlight the need for standardized reporting and robust validation.

1. Introduction

An epidemic is the happening of a more significant number of injuries, diseases, or other medical problems than expected in a particular place among many or a particular group of people in a given period. One of the most dangerous epidemics in the past is pneumonia, which accounts for 14% of all deaths among children under 5 years of age [1]. Pneumonia is categorized as infectious and noninfectious based on different pathogens, where infectious pneumonia is classified into bacteria, viruses, mycoplasma, chlamydial pneumonia, and others, while noninfectious pneumonia is immune‐related pneumonia, aspiration pneumonia caused by physical and chemical factors, and radiation pneumonia are classified [2]. Pneumonia as community‐acquired pneumonia (CAP), HAP (Nosocomial Pneumonia), and ventilator associated pneumonia (VAP) are classified based on different infections, among which CAP forms a major part. Exposure to air pollution is a risk factor for severe pneumonia [3]. Overcrowding, contamination, and unsanitary environments lead to pneumonia in underdeveloped and developing countries with few medical resources [3, 4]. Pneumonia mortality is strongly age‐related, and the prevalence of pneumonia grows dramatically with age [5]. Malnutrition significantly contributes to stroke‐associated pneumonia (SAP) by weakening the immune system, impairing infection resistance, and slowing healing. Stroke patients often struggle with swallowing, aspiration, and decreased physical activity, leading to poor nutritional intake and subsequent malnutrition, which compromises the body's defenses and fosters pneumonia development. Research indicates that gut microbiota—the intestinal microorganisms—affect immune response and pneumonia susceptibility. Dysbiosis, an imbalance in the gut microbiome, is associated with increased inflammation and higher respiratory infection risks, including pneumonia, particularly in stroke patients. The gut‐brain‐lung axis, linking gut and respiratory health, suggests that a healthy gut microbiota may help prevent SAP by enhancing immune function and reducing inflammation [6, 7, 8].

Pneumonia has been a serious human health problem over the years because children and adults die every year, many researchers have performed strongly and are still trying to find the best way to diagnose pneumonia [6]. Diagnosing pneumonia and PTE presents a certain complexity due to the similarity of their symptoms, cough, shortness of breath, and chest pain [7, 8]. Also, the signs and symptoms given are often subjective and nonspecific. For this reason, the criteria or auxiliary diagnoses specified in the guidelines [9, 10] are updated every year. However, wearing a mask, living a healthy lifestyle, acceptable hygiene, and avoiding smoking are some safeguards that can be taken to prevent pneumonia [11]. Varied approaches are used to detect pneumonia, including chest radiography, blood culture, sputum culture, fluid sample, bronchoscopy, and pulse oximetry [12]. However, diagnostic radiological techniques for lung diseases include chest X‐ray imaging, computed tomography (CT), and magnetic resonance imaging (MRI), among which chest X‐ray imaging is useful and reasonable. Because it is more functional and portable in the hospital, and exposure to radioactivity has a lower dose for patients. However, there is a severe shortage of trained radiologists. X‐ray imaging is also preferred over CT imaging because CT imaging takes seriously longer than X‐ray imaging, and high‐quality CT scanners may not be obtainable in many underdeveloped areas. On the other hand, X‐ray is the most typical and general diagnostic imaging technique that plays an important role in clinical care and epidemiological studies [13, 14].

One of the hopeful techniques for predicting and diagnosis of diseases recently is artificial intelligence. Computer‐aided diagnosis using artificial intelligence‐based solutions is increasingly widespread these days [15, 16, 17, 18]. Artificial intelligence (AI) is considered an important aspect for predicting pneumonia in patients, and also AI appears promising in the diagnosis of pneumonia in chest X‐rays (CXR). Therefore, MDA‐PSP uses the integrated important signs and CXR of patients to predict pneumonia status using a trained model and organize them using deep learning [11, 19]. ML is very common in health care informatics in disease diagnosis and prognosis [20, 21, 22, 23]. However, the current methods for identifying pneumonia have low accuracy and their use may delay antimicrobial treatment. However, ML can be integrated with electronic medical record (EHR) systems to identify information and help prompt clinical decision‐making [24]. Automatic classification of disease in X‐ray images is demanding due to the lack of access to enormous amounts of annotated data and efficient ML algorithms to learn their specific attributes. Multiple data streams and methods can be used to enhance disease prediction accuracy. Textual data from X‐ray diagnostic images were combined with annotated image data for chest disease classifier training [25]. Unlike the usual strategy of direct disease classification, deep learning can be used to create residual maps for abnormal diseases along with normal images [26]. Of course, these deep learning algorithms need a lot of data to assemble an accurate model. Unfortunately, access to such a large volume of labeled data is another important problem for ML approaches in the medical field [27, 28, 29]. Electronic clinical decision support and mobile health tools as ways to coordinate patient care with treatment and management procedures aligned with guidelines have appeared [30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40]. Another concern with this disease is that sometimes the attributes that describe the existence of this disease are often combined with other diseases, and therefore radiologists challenge the diagnosis of this disease. To handle this problem, the need for artificial intelligence (AI) is felt to detect the response to pneumonia treatment in affected patients [11, 41, 42]. Therefore, the purpose of this systematic review was to investigate and estimate ML algorithms used to diagnose, prevent, and treat pneumonia.

2. Methods

2.1. Literature Search and Screening

A systematic review was conducted on 15 January 2024, searching PubMed, Scopus, Web of Science, and Google Scholar using terms “Machine Learning” and “Pneumonia” and their MeSH terms, following the PRISMA checklist. No time restrictions were applied.

2.2. Eligibility Criteria

Included studies were English‐language articles developing or validating ML models for pneumonia diagnosis, prediction, or treatment. Excluded were non‐journal articles (e.g., conference abstracts, reviews, letters), articles with unavailable full texts, or those lacking relevance to the study objectives. Two researchers independently reviewed articles, with disputes resolved by a third author.

2.3. Data Extraction and Synthesis

Data were extracted on publication year, first author, country, study objective, data set size, ML algorithms, and performance metrics (accuracy, sensitivity, specificity, AUC, F1 score). Accuracy was defined as the proportion of correct predictions, sensitivity as the true positive rate, specificity as the true negative rate, AUC as the model's discriminative ability, and F1 score as the harmonic mean of precision and recall. Following Assel et al. (2018), confidence intervals (95% CI) were extracted where available to quantify uncertainty. Cross‐validation techniques (e.g., k‐fold), data set splitting (training, validation, test sets), and methods to address overfitting (e.g., regularization) or class imbalance (e.g., SMOTE) were noted. Two researchers independently extracted data, with discrepancies resolved through discussion. In addition to performance metrics, information on ethical considerations, such as ethical approval and informed consent, was extracted from included studies where reported.

3. Results

3.1. Study Selection

Figure 1 shows the approach of finding and preferring research based on the PRISMA chart. A total of 11,545 related articles were selected for review and 3780 duplicate studies were removed. The publications were then screened using their titles and abstracts. By completing the review and removing the articles that had nothing to do with the research objective, 250 full‐text articles were evaluated, and finally, 42 articles remained for study. Follow the results of the search strategy in Table 1.

Figure 1.

Screening of studies based on the PRISMA chart.

Table 1.

Outcomes of search strategies for every database.

#	PubMed database search approach	Results
1	“pneumonia” [MeSH Terms]	36,149
2	“pneumonia” [Title/Abstract] OR “Pneumocystis” [Title/Abstract]	16,368
3	“Machine learning” [MeSH Terms]	12,015
4	(((((“machine learning”[Title/Abstract]) OR (“deep learning”[Title/Abstract])) OR (“supervised machine learning”[Title/Abstract])) OR (“support vector machine”[Title/Abstract])) OR (“Unsupervised Machine Learning”[Title/Abstract]))	17,152
5	1 OR 2	45,104
6	3 OR 4	18,034
7	5 AND 6	4471
#	Scopus database search approach	Results
1	TITLE‐ABS‐KEY (“pneumonia”) OR (“Pneumocystis”)	524,940
2	TITLE‐ABS‐KEY (“machine learning”) OR TITLE‐ABS‐KEY (“deep learning”) OR TITLE‐ABS‐KEY (“supervised machine learning”) OR TITLE‐ABS‐KEY (“support vector machine”) OR TITLE‐ABS‐KEY (“Unsupervised Machine Learning”)	1,037,665
3	1 AND 2	4576
#	Google Scholar database search approach	Results
1	“machine learning” OR “deep learning” OR “supervised machine learning” OR “unsupervised machine learning” OR “support vector machine”) AND (“pneumonia” OR “Pneumocystis”
774
#	Web of Science database search approach	Results
1	TS= (“pneumonia”) OR TS= (“ Pneumocystis”)	173,820
2	TS= (“machine learning”) OR TS= (“deep learning”) OR TS= (“supervised machine learning”) OR TS= (“unsupervised machine learning”) OR TS= (“support vector machine”)	452,752
3	1 AND 2	1724

3.2. Characteristics of the Study

Among the scanned articles, 10 articles in China [43, 44, 45, 46, 47, 48, 49, 50, 51, 52], 9 articles in India [53, 54, 55, 56, 57, 58, 59, 60], 3 articles in America [61, 62, 63], three articles in Egypt [64, 65, 66], 3 articles in Turkey [67, 68, 69], 3 articles in Saudi Arabia [70, 71, 72], 2 articles in Pakistan [73, 74] and other studies in Japan [75], Germany [76], Great Britain [77], Algeria [78], Georgia [79], Romania [80], Korea [81], Bangladesh [9] has been done. ML algorithms used in most studies include convolutional neural network (CNN) [49, 50, 51, 52, 58, 59, 66, 79, 80, 81], support vector machine (SVM) [9, 45, 47, 48, 50, 55, 73, 78], KNN [9, 47, 52, 74, 78], random forest (RF) [9, 47, 48, 55, 62, 70, 78], logistic regression (LR) [47, 48, 55, 61], DT decision tree) [54, 77, 82], XGBoost [47, 48, 61, 62, 67, 78].

3.3. The Effect of Machine Learning Algorithms on Pneumonia Diagnosis

According to Table 2, half of the included studies analyzed the influence of using ML algorithms on the diagnosis of pneumonia diseases. In the study of Sourab et al. [9], a convolutional neural network (CNN) architecture with 22 layers was created for this proposed method, and three separate ML techniques were used to extract and categorize the learned features (CNN model was used. Support Vector Machine, Random Forest Classifier, and K‐Nearest Neighbor). After various classifications of RF, KNN, and SVM classifiers employing the training characteristics of the CNN model, the accuracy obtained was 99.52%, 96.55%, and 97.32% respectively. This proposed CNN‐RF hybrid method is comparative to other conventional methods with an accuracy of 99.52% and an AUC score of 98.7%. In the study of Liu et al. [43], who proposed the multi‐branch fusion learning (MBFL) method for pneumonia detection from chest X‐ray images, several experiments were conducted, and the average classification accuracy was 95.61. The highest single‐class accuracy for COVID‐19 is obtained with an accuracy of 99.10%. Using the MBFL model, better results were obtained for rapid pneumonia screening. In the study of Longjiang et al. [44], a pediatric clinical set was constructed using Rsnet‐50 to differentiate viral from bacterial pneumonia in the first test set of the best‐performing classifier and the sensitivity was 79 and the specificity was 88.9. In the article by Mabrouk et al. [64], which shows a computer‐aided classification of pneumonia, that uses the mobile learning symbol NetV2, DensNet 169, and Vision Transformer, the proposed EL approach from another. The existing advanced techniques performed better and achieved 93.91 accuracy and 93.88 F‐one score in the test phase. In the study of Manickam et al. [45] with the aim of pre‐processing the input image of the chest to identify the presence of pneumonia using segmentation based on unit architecture and classification of pneumonia as bacteria or virus using data models. The net image is done. In Masud et al.'s study [70], a recent approach was presented to determine the presence of pneumonia and identify its type through chest radiography analysis. A three‐class classification was based on the attributes including various information from the samples, and then the final classification was done using the random forest classifier. Finally, the proposed method was tested on a widely employed chest radiograph data set to estimate its performance. As a result, the proposed model can categorize the data set sample with a classification accuracy of 86.30 and an F‐one score of 86.03. In Muhammad et al.'s study [73], classification algorithms such as LR, SVM, ANN, NB, and KNN were evaluated, with transfer learning using InceptionV3 achieving the highest accuracy at 97.19%. Muralidhar et al. [55] highlighted the use of ML and DL for intelligent pneumonia diagnosis predictions, employing chest X‐rays and Mendelian images for training. They used various DL models (CNN, Xception, Inception, ResNetV2, VGG16, ResNet50) and ML algorithms (SVM, RF, DT, AB), finding that the hybrid and SVM models achieved the highest classification accuracy at 97.70%. In Postalcioglu et al.'s study [67], a software tool evaluated X‐ray images for enhancement techniques, with CatBoost yielding the best results in speed (0.7 s) and accuracy (83%). In Rixe et al.'s study [61], techniques like wood embedding, XGBoost, SVM, naive Bayes, and LRlite GBM were evaluated, with naive Bayes showing the highest sensitivity at 93.5% and XGBoost achieving the highest F1 score of 72.4%. For pneumonia diagnosis, naive Bayes again recorded the highest sensitivity at 80.1%, while SVM had the best F1 score at 53%. Harika et al. [82] utilized a CNN model, achieving accuracy rates between 77.35% and 87.64% in training and 81.46% to 90.62% in testing. Hedhoud et al. [78] proposed a CNN and XGBoost‐based model for classifying viral and bacterial pneumonia images, achieving 89% accuracy and 85% sensitivity. In the study by Jiang et al. [46], enhancement pre‐processing was used to address the low volume of a poorly balanced data set, demonstrating that data augmentation significantly improves model accuracy in convolutional neural networks. Kanwal's study [74] tested five machine learning classifiers—Fine tree, Linear Discriminant, weighted KNN, wide neural network, and ensemble bagged tree—using eight PPG signal features, with weighted KNN accurately predicting 9 out of 10 subjects and improving clinical decision‐making. Kareem et al. [77] explored the use of CNN, KNN, ResNet, CheXNet, and ANN for pneumonia diagnosis and conducted a survey on how hospitals and medical institutions could collaborate to train ML models with their datasets for more effective and accurate disease diagnosis. In the study by Kaya et al. [68], a CNN framework was proposed for accurate pneumonia diagnosis, achieving 98.94% accuracy and a 99.12% F1 score on the general test data set. Kumar et al. [56] employed a CNN to automatically differentiate between pneumonia and normal cases, attaining 91% validity for pneumonia classification. Larsen et al. [83] developed a model with 90% sensitivity, 77% specificity, 82.5% accuracy, and an F1 score of 0.86, aiding diagnosis for those with physician access. Han et al. [63] demonstrated their ResNet18 model's superiority on the RSNA pneumonia diagnosis data set, achieving 85.4% accuracy, 0.84 F1 score, and 0.87 AUC, outperforming several advanced models. In Elshnnawy et al'‘s study [65], four models were developed using various deep learning methods, including ResNet152V2 and MobileNetV2. Their framework achieved impressive metrics: 99.22% accuracy, 99.43% precision, 99.44% F1 score, 99.44% recall, and 99.77% AUC, with ResNet152V2 outperforming the others. MobileNetV2, CNN, and LSTM‐CNN models also exceeded 91% in various metrics. Celik et al. [69] utilized Multilayer Perceptron (MLP) and k‐Nearest Neighbors (k‐NN) algorithms for pneumonia detection, finding that MLP outperformed k‐NN across all evaluation criteria, achieving 95.673% accuracy, 95.706% F1 score, and 99.006% AUC. In the study by Chouhan et al. [57], five different models were analyzed, leading to the proposal of an ensemble model that surpasses the individual models with an accuracy of 96.4% and recall of 99.62% in diagnosing pneumonia. Anai et al. [75] utilized two deep learning models; the first achieved a precision‐recall AUC of 0.929 with 50.0% sensitivity and 92.4% specificity for fatal pneumonia classification. Using an external validation set of 100 chest X‐ray images, the model recorded 68% sensitivity, 86% specificity, 77% accuracy, and an F1 score of 74.7%. The second model showed sensitivity, specificity, and accuracy of 39.6%, 92.8%, and 82.7%, respectively, with external validation at 38.0% and 92.0%. Almaslukh et al. [71] proposed a lightweight deep learning technique using pre‐trained DenseNet‐121 for feature extraction and a deep neural network (DNN) with random search fine‐tuning. DenseNet‐121 was chosen for its effective representation of lung characteristics. This method achieved an accuracy of 98.90%, improving by 0.47% over the previous approach on the same data set. In the study by Ali et al. [66], an AI model diagnosed pneumonia from chest X‐ray images with 92% sensitivity and specificity, achieving an accuracy of 92.75%. Singh et al. [58] developed a QCSA network using spatial and channel attention mechanisms, attaining 94.53% accuracy and 0.89 AUC on Kaggle's X‐ray data set. Stephen et al. [81] employed data augmentation to improve a CNN model, reporting a training accuracy of 95.31% and a validation accuracy of 93.73%. In Sushant et al.‘s study [80], the research models used are CNN, Dens Net, VGG16, ResNet50, and Inception Net. All 5 models are trained with 30 cycles and among all models, Dense Net shows an accuracy of 94.00%, which is higher than the other 4 models. In the study of Szepesi et al. [79], the results obtained by the network in the Kaggle competition are categorized first with the following criteria: 97.2% accuracy, 97.3% recall, 97.4% precision, and AUC = 0.982. In Vetrithangam et al.‘s study [59], the principal objective of designing and deploying this modified ResNet152v2 model to predict pneumonia from chest X‐rays is to achieve high accuracy while minimizing computational complexity and reducing calculation time. This model had a better performance compared to the existing approaches and an accuracy of 99.77%, sensitivity of 99.86%, specificity of 95.4%, and precision of 99.86% were obtained. In Wang et al.'s study [49], they entered the X‐ray image with texture features into the modified VGG16 model, C‐VGG, and then added the fusion method to C‐VGG. For fusion, they added a feature to obtain FC‐VGG. The model achieved an accuracy of 92.19% in recognizing children's pneumonia images, an average accuracy of 93.44%, an average recall of 92.19%, and an average F1 coefficient of 92.81%. In Wang et al.'s study [50], they present a deep regression framework for automatic pneumonia screening that uses multichannel images and multivariate information to simulate the Pneumonia clinical screening process. In this study, the RCNN model with accuracy, sensitivity, specificity, and F1 score of the baseline are 0.923, 0.911, 0.936, and 0.917 respectively. In Yang et al.'s study [51], the results show that LeNet5 and AlexNet models achieved better pneumonia diagnosis for small data sets, while MobileNet models and ResNet18 were more appropriate for pneumonia detection for big datasets. In the study of Yaseliani et al. [52], the best classifier model was presented using a weight of 0.4 for the SVM‐RBF classifier and a weight of 0.1 for the LR classifier, with an accuracy of 98.55%, a precision of 98.72%, 99.30% recall and 99.01% F1 score. This model had the most pleasing performance in terms of all performance measures compared to the hybrid model with FC layers and all ML classifiers. In Zeban et al.'s study [60], the deep learning method was implemented as a wavelet scattering network as a lung pneumonia classification model. The proposed system specifies that the wavelet scattering network has classified chest X‐ray images with 98% accuracy.

3.4. The Effect of Machine Learning Algorithms in Predicting Pneumonia

Out of 42 studies, six examined machine learning (ML) algorithms for disease prediction (Table 3), focusing on high‐risk patient identification [53, 54], predicting 30‐day readmission risks for pneumonia [62], and forecasting pneumonia after liver transplantation [48]. Swetha KR's study utilized big data techniques and various ML models (SVM, RF, LR, DT, CNN), with CNN demonstrating significant accuracy [53]. Meena K's research categorized the population based on virus effects into high‐risk, low‐risk, and moderate groups, achieving 98%–99% accuracy, with ResNext achieving the highest accuracy among CNN models [54]. Yu and colleagues applied a rule‐based model alongside other ML models to predict 30‐day rehospitalization risk, finding XGBoost and the rule‐based model most effective, with XGBoost providing the best predictions [62]. Effah CY's team tested eight ML models based on biomarkers and laboratory parameters, identifying RF and XGBoost as the best predictors in both original and SMOTE datasets [47]. Chen CJ et al. developed a model for predicting postoperative pneumonia in liver transplant patients, achieving the best performance with 14 common variables in XGBoost [48]. Alharbi AH et al. introduced an enhanced BoxENet model, which excelled in speed and accuracy in both binary and multi‐classification tasks [72].

3.5. The Effect of Machine Learning Algorithms in the Treatment of Pneumonia

Among 42 studies, in one case, input studies were used in the field of pneumonia cure. In this study, König R. et al. created a decision tree based on etiological and clinical parameters. It was found that the patients who were treated according to the recommended treatment had a lower mortality rate (27%) compared to the observed standard of care [76].

4. Discussion

4.1. The Principal Findings

Pneumonia is an acute respiratory infection that is one of the most common diseases in the world with a high mortality rate in the last few decades. Due to the complexity of lung diseases, the diagnosis of pneumonia in chest X‐ray is highly dependent on the eye of an experienced radiologist. Consequently, there is great potential for AI algorithms to further assist and improve detection. ‎AI has been able to play a significant role in the diagnosis, prediction, and cure of pneumonia and is a practical guide for physicians. As a result, ML systems that combine and interpret an enormous amount of complex and enormous data have been able to play a meaningful role in the diagnosis, prediction, and treatment of pneumonia. Accordingly, this systematic review aimed to gather the most acceptable available research information in the field of diagnosis, prognosis, and treatment of pneumonia. The evidence from this review shows that in all studies, ML techniques have a practical and positive approach to facilitating disease diagnosis prediction and treatment. According to the findings of our study, out of a total of 42 studies, in 35 of the input studies, ML algorithms were used to diagnose pneumonia and all studies show the high capability of ML algorithms. were diagnosed with pneumonia [9, 43, 44, 45, 46, 49, 50, 51, 52, 55, 56, 57, 58, 59, 60, 61, 63, 64, 65, 66, 67, 68, 69, 70, 71, 73, 74, 75, 77, 78, 79, 80, 81, 82, 83]. Thus, even in K. Larsen's study, it was stated that a deep learning model, a modified convolutional neural network, could receive input in the form of X‐rays and make a diagnosis with the same accuracy as a physician compared to a pre‐diagnosed image. The data was gathered based on accuracy measurement as well as more specific criteria such as specificity and sensitivity, and the results showed that the tested algorithm was able to accurately identify uncommon lung findings on average in 82.5% of cases. The model separately reached a maximum specificity of 98.5% and a maximum sensitivity of 90%, and the highest simultaneous values of these two criteria were 90% sensitivity and 78.5% specificity [83]. Among the ML algorithms used in the included studies, CNN models [9, 45, 46, 49, 50, 51, 52, 55, 56, 57, 58, 59, 61, 63, 64, 65, 66, 68, 73, 77, 78, 79, 80, 81, 83], SVM [9, 45, 49, 55, 73, 78], KNN [9, 60, 62, 77] are the most used. Nevertheless, among all the used algorithms, the ResNet152v2 algorithm with 99.77% accuracy [59] and the combined CNN‐RF algorithm with 99.52% accuracy [9] were more effective for pneumonia diagnosis and differentiation. On the other hand, the SVM algorithm (RBF) had the least effect in diagnosing pneumonia with an accuracy of 51.13% [53]. In the study of Sourab SY et al., they found that the combination of RF, KNN, and SVM machine learning algorithms with CNN has a more pleasing performance than using a CNN algorithm alone to optimize pneumonia diagnosis. Meanwhile, the CNN‐RF hybrid model showed higher accuracy [9]. One of the challenges of conventional data analysis techniques is the rapid increase of multidimensional clinical data and the difficulty of integrating clinical information. Since the clinical information of different fields is heterogeneous in terms of structure and meaning, the combination and integration of ML algorithms seems to be more effective for better diagnosis of diseases and differentiation between their types. In this context, the reliability and structure of national electronic health records can directly affect the performance of ML algorithms, especially in low‐resource healthcare settings [84].

On the other hand, in 6 cases of input studies, ML algorithms were used regarding the capability to predict pneumonia disease [47, 48, 53, 54, 62, 72], which include cases related to identifying high‐risk patients [53, 54], predicting the risk of 30‐day readmission in patients with pneumonia [62] and predicting pneumonia after liver transplantation [48]. In this case, Meena K and colleagues provided a solution for physicians to predict the effect of the virus as high risk, low risk, and moderate among the testing population via various deep learning techniques such as convolutional neural networks. CNN), artificial neural network (ANN), and recurrent neural networks using short‐term memory cells that examined 3000 CT images of approved pneumonia patients, and 98%‐99% accuracy was achieved [54]. In the study of Huang Y et al., a rule‐based model and other machine learning (ML) models, including decision tree, random forest, LASSO, and XGBoost, found that the performance of machine learning models for predicting readmission is different in pneumonia patients, and among these, XGboost performed better than the rule‐based model based on AUROC [62]. Also, in the study of Chen CJ et al., they developed a model for predicting postoperative pneumonia in patients after orthotopic liver transplantation (OLT) using ML. Among the six ML models including logistic regression (LR), support vector machine (SVM), random forest (RF), adaptive boosting (AdaBoost), extreme gradient boosting (XGBoost) and gradient boosting machine (GBM), XGBoost model with AUC 0.734 (sensitivity: 52.6%; specificity: 77.5%) in the test set and with 14 common variables (INR, HCT, PLT, ALB, ALT, FIB, WBC, PT, serum Na + , TBIL, anesthesia time, preoperative length of stay, total fluid transfer and operative time) had the best performance [48]. Even in the study of Effah CY et al., several experimentations were performed using eight ML models to predict pneumonia based on biomarkers, laboratory parameters, and physical characteristics. The results of which showed that biomarkers such as C‐reactive protein and procalcitonin have the most discriminating power and the set of machine learning models such as RF with (accuracy = 92.0%, precision = 91.3%, recall = 96.0%, f1‐Score = 93.6%) and XGBoost has the highest performance accuracy with (accuracy = 90.8%, precision = 92.6%, recall = 92.3%, f1‐score = 92.4%) In the original data set, it obtained AUCs of 0.96 and 0.97, and in the SMOTE data set, RF and XGBoost obtained the highest prediction results with f1 scores of 92.0% and 91.2%, respectively. Also, an AUC of 0.97 was obtained for both RF and XGBoost models [47]. Similar approaches have also been explored in predictive rehabilitation research systems, where structured data collection enables personalized patient stratification and risk estimation [85, 86, 87]. It seems that using XGBoost [47, 48, 62] and CNN [53, 54] can be very useful in predicting pneumonia. In addition, according to Alharbi AH et al.'s study, the presented enhanced BoxENet model achieves more acceptable than other models in binary and multi‐classification models, and the improved BoxENet compared to other models in training and also has a higher speed in classification [72]. In addition, among 42 studies, ML algorithms were used in the field of treating pneumonia in one of the included studies. In the study by König R et al., they developed a decision tree based on etiological and clinical parameters, which are already available to support a personalized decision for or against macrolides for the best clinical outcome for each patient. This prospective, multinational CAPNETZ study aimed to enhance the primary outcome of 180‐day survival. As a result of a simple decision tree of patient attributes, including chronic cardiovascular and chronic respiratory diseases, as well as the number of leukocytes in respiratory secretions at the time of registration, it was found that patients without cardiovascular disease or patients with concomitant respiratory diseases and high leukocyte counts in secretions Inhalation benefit from macrolide cure. ‎Patients who were treated according to the recommended cure had a lower mortality (27%) compared to the observed standard of care. Consequently, the classification of macrolide treatment in patients following a simple treatment law may lead to a significant reduction in mortality in CAP [76]. It seems that the role of ML algorithms in enhancing disease diagnosis was more than disease prediction and treatment. Meanwhile, the use of combined methods can optimize differential diagnoses and promote the decision‐making process of health care providers. In addition, the identification of the factors of disease transmission is very important for the implementation of control and prevention measures, because the identification of these factors can bring predictions that help to make informed decisions in the treatment of pneumonia, public awareness campaigns, and health professional education. While numerous machine learning options exist, ChatGPT, a widely recognized and freely accessible ML model, can significantly contribute to pneumonia diagnosis by supporting healthcare professionals in their interpretation of symptoms, patient history, and diagnostic results. This model has the capacity to examine extensive datasets, recognize early indicators, and aid ML algorithms in data preprocessing and analysis. By seamlessly integrating with clinical processes, ChatGPT enhances decision‐making and ensures that diagnostic models are trained using high‐quality, pertinent data [88, 89, 90, 91].

In addition to the applications of machine learning in the diagnosis and prediction of pneumonia, recent studies have demonstrated that deep learning algorithms, such as temporal convolutional networks (TCNs), exhibit remarkable performance in analyzing multimodal physiological signals (e.g., EEG, EOG, and EMG) for emotion recognition [87]. This study highlights the capability of TCNs in modeling complex temporal data, which could potentially be applied to analyzing clinical signals related to pneumonia, such as respiratory or heart rate data. This underscores the potential of deep learning algorithms for integrating multimodal data in disease diagnosis and may serve as a new research direction for enhancing pneumonia prediction models. Respiratory‐related signals, such as electrocardiogram (ECG) and photoplethysmography (PPG), have shown promise in enhancing the accuracy of machine learning models for pneumonia diagnosis, particularly when integrated with wearable sensors [92, 93].

4.2. Strengths and Limitations

This study has several strengths. First, this study has been one of the few articles that have investigated machine learning methods in pneumonia. Second, in our search, we not only looked at pneumonia but also about COVID‐19, since the number of final articles reached 292 articles and the volume of tables increased, we excluded COVID‐19 from our study. Third. We carefully reviewed all the studies that used different types of ML algorithms in pneumonia based on all indicators so that the results would be helpful for future researchers. Fourth. The country filter was not considered for searching articles. Therefore, studies from all LMICs and high‐income countries were suitable for this review. Among the limitations, since English is the international and common language of the world, our review exclusively included English language studies. As a result, this evaluation did not consider related articles written in other languages of the world, and we may have missed an article in this review.

5. Conclusion

Pneumonia disease, although it is a serious and acute disease, with the use of artificial intelligence, can be identified earlier to help the patient and his family. This technology also acts as an effective associate for physicians and other healthcare providers in diagnosing pneumonia and can assist health system managers in decreasing cure costs.

Author Contributions

Azam Kheirdoust wrote the protocol, helped with the search strategy, and wrote the article's first draft. Fatemeh Barzanouni took part in authoring the article's first draft, revising it, and managing the search strategy. Alireza Rasoulian and Fatemeh Barzanouni took part in the data extraction and article quality evaluation processes. Kosar Ghaddaripouri and Aynaz Esmailzadeh took involved in the management of the study selection, data extraction, and article quality evaluation. Mohammad Reza Mazaheri Habibi wrote the study protocol and helped with study design. The final draft of the manuscript has been reviewed and approved by all authors. [Mohammad Reza Mazaheri Habibi] is the sole responsible party for the accuracy and integrity of the data analysis and has full access to all study data.

Conflicts of Interest

The authors declare no conflicts of interest.

Transparency Statement

The lead author Mohammad Reza Mazaheri Habibi affirms that this manuscript is an honest, accurate, and transparent account of the study being reported; that no important aspects of the study have been omitted; and that any discrepancies from the study as planned (and, if relevant, registered) have been explained.

Supporting information

supmat.

Kheirdoust A., Barzanouni F., Rasoulian A., et al., “Evaluation of Machine Learning Methods Developed for Prediction and Diagnosis of Pneumonia: A Systematic Review,” Health Science Reports 8 (2025): 1‐10, 10.1002/hsr2.71446.

Data Availability Statement

Data and materials are available from the corresponding author upon reasonable request.