1 Introduction

The airway diseases have been called the worldwide source of sickness, as they damage the air tubes that carry oxygen and other gases into and out of the lungs. These diseases generally induce a constriction or obstruction of the airways. People with airway illnesses often remark that they are "trying to breathe out through a straw." Airway disease is a fundamental cause of mortality and disability as it touches at least 65 million people and kills 3 million individuals yearly. Hence it is ranked as the 3rd most prevalent reason for death around the globe overall [1]. As shown in Fig. 1, chronic respiratory disorders fatalities rose from 3.32 million in 1990 to 3.91 million in 2017. Annually, the age-standardized mortality rate for chronic respiratory diseases fell by 2.41% (2.28% to 2.55%), but in 195 nations, the annual fatality rates attributable to chronic respiratory illnesses have shown a great variation. The highest mortality was seen in regions with a low Socio-demographic index [2]. Particulate matter pollution, such as pneumoconiosis and asthma, has been the leading cause of COPD fatalities in areas with a low Socio-demographic index [3]. In addition, one of the studies also states that 40,000 people have been suffering from cystic fibrosis in the US [4], 10 million tuberculosis cases have been estimated, and 1.4 million are reported TB deaths in 2019 [5]. In contrast, lung cancer has been termed the leading cause of cancer for approx 350 deaths each day in the US [6].

Fig. 1
figure 1

Global impact of airway diseases [2]

Full size image

After considering these numbers, several physicians, clinicians, and medical experts from all over the world have tried their best possible way to predict, detect and diagnose these dreadful diseases, for which they used conventional and AI-based approaches that are described in the upcoming subsections.

1.1 Conventional Methods to Diagnose Airway Diseases

Small airways account for between 10 and 25% of total airway resistance in healthy lungs, with their contribution to total airway resistance increasing significantly in multiple airway disorders. It is because the internal diameter of the tiny airways is less than 2 mm, and there is no cartilage [5]. Hence in Table 1, a few traditional ways such as spirometry, body plethysmography, impulse oscillometry, and washout tests used for diagnosing airway disorders in patients are shown, along with their brief description and drawbacks.

Table 1 Traditional approaches to diagnose airway diseases
Full size table

1.2 AI Techniques to Predict Airway Diseases

Artificial intelligence has shown tremendous growth in the health sector in today's era via its various applications, as shown in Fig. 2. After having a brief idea about the drawbacks of traditional approaches to diagnosing airway diseases, artificial intelligence has demonstrated its efficiency and excellent performance in automatic image categorization using multiple machine and deep learning algorithms to detect numerous airway illnesses [10]. Researchers from various disciplines are steadily amassing evidence to support the use of AI in diagnosing airway sickness using models that can learn and make choices utilizing massive input data [11]. The fundamental rationale for using learning models is that these approaches learn by constructing a more abstract representation of input (unlike classical machine learning) in which the model collects information automatically and produces more accurate results [12].

Fig. 2
figure 2

Role of AI in healthcare

Full size image

Researchers from many fields are progressively accumulating data to support the usage of AI to diagnose airway illness. Researchers have applied various algorithms of the machine and deep learning, as discussed in Sect. 5, to conduct an extensive assessment to detect multiple airway diseases. A corresponding spike has been seen in the medical applications of AI, particularly in pulmonology. Applications of AI to the global problem of airway illnesses can meet the most priority-based requirements highlighted for accurately detecting, diagnosing, and providing the best treatments [13]. A deep learning system generated results nearly equal to a group of thoracic radiologists identifying fibrotic lung disease in research [14]. Another study [15] found that a neural network developed by Google scientists was just as good as radiologists in detecting cancerous lung nodules. A similar model [16] could identify and predict acute respiratory illness episodes and death in smokers along with chronic obstructive pulmonary disease (COPD). Large amounts of well-structured data are required to create and validate AI algorithms, and the algorithms must operate with data of varying quality [17]. Clinicians must grasp how AI works in the context of multiple disorders such as asthma and chronic obstructive pulmonary disease. As a result, it would be interesting to witness the pros of AI (Artificial Intelligence) developed for doctors and patients due to its usage in medical practice in the future [18]. Other AI tools have also been used tremendously in medical science, such as robotics, natural language processing, expert systems, etc. In robotics, more than 2 lakh robots have been installed annually and are used to perform prostate surgery, head-neck surgery, update patient records, etc. Likewise, natural language processing (NLP) helps to analyze, understand and classify the unstructured format of clinical documentation and the interaction of patients with doctors via bots [19].

Thusly, to fully understand the role of techniques using artificial intelligence concepts to predict and diagnose airway diseases, a systematic literature review (SLR) has been conducted based on the AI learning models to predict such illnesses. The study is structured into various sections. Initially, a brief description of airway diseases and their global impact has been briefed in Sect. 1. This section also includes the traditional approaches as well as the role of AI in diagnosing airway diseases. Section 2 presents the qualities of the paper used in the study, while Sect. 3 presents the framework to predict airway diseases using AI-based techniques. Section 4 provides a comparison of various methods based on their dataset, applied algorithms with their results in terms of different metrics such as precision, accuracy, F1-score, etc., and the limitations of multiple airway diseases such as Mesothelioma, cystic fibrosis, emphysema, pulmonary edema, pneumoconiosis, lung cancer, pulmonary embolism, tuberculosis, asthma, and covid-19. The discussion segment, which answers the research questions from Sect. 2, is handled in Sect. 5. Finally, Sect. 6 concludes the study that assists researchers in determining the optimal technique for detecting disorders and the future scope.

2 Materials and Methods

Full-text archives in the English language of six different publication databases have been searched between 2010 and 2022 such as (ScienceDirect (https://www.sciencedirect.com), Google Scholar (https://scholar.google.co.in), Scopus (https://www.scopus.com), Web of Science (http://isiwebofknowledge.com), PubMed (http://www.ncbi.nlm.nih.gov/pubmed), EMBASE (https://www.embase.com). The paper has been searched using the keywords "airway diseases", "cystic fibrosis"," pneumoconiosis"," pulmonary embolism", "pulmonary edema", "Mesothelioma", "lung cancer ", "tuberculosis", "asthma ", "covid-19"," emphysema", "machine learning" "artificial intelligence," "deep learning," as well as combinations of these keywords. In addition to this, the articles have also been chosen based on inclusion and exclusion parameters (Table 2) which are based on various metrics such as time duration, scholarly articles from which the paper can be accessed, research to highlight the problem, a comparator for analyzing and comparing the researchers' work, methodologies to demonstrate the strategies that had been used in their articles, and finally research design to analyze the results.

Table 2 Inclusion standards and exclusion standards
Full size table

In this systematic literature review, PRISMA (Preferred reporting items for systematic Reviews and Meta-Analyses) guidelines have been applied in which four phases have been used to select the research papers (Fig. 3), i.e., Identification- in which the identification of records is carried out by accessing various repositories, Screening- in which the papers are selected transparently by assaying the decisions that are made at different stages of the systematic review, Eligibility- in which all full-length articles are evaluated [20] and finally Included- in which the final selected articles are included to write the review paper. PRISMA is chosen as it helps improve the reporting and transparency of systematic reviews and meta-analyses. It is also useful for the readers to understand how the authors have filtered out the selected papers by using keywords, year of publication, language, etc., to frame any article.

Fig. 3
figure 3

PRISMA flow chart

Full size image

As far as this paper is concerned, for a better understanding of the state of research on machine and deep learning in airway disease detection, peer-reviewed papers stated that these algorithms had played a vital role in predicting such disorders. In addition to this, a few research questions have also been framed that were investigated in the study:

RQ 1: Year wise analysis of predicting multiple airway diseases using AI based techniques.

RQ 2: How doctors are being helped by deep and machine learning techniques in detecting the airway diseases?

RQ 3: Which ML and DL techniques are broadly applied to predict airway diseases?

RQ 4: Name the characteristics that manipulate the quality of prediction models based on deep and machine learning?

3 Framework to Predict Multiple Airway Diseases

In this section, various phases to predict and classify airway diseases have been mentioned and diagrammatically shown in Fig. 4.

Fig. 4
figure 4

Predicting airway diseases using multiple learning models

Full size image

Dataset: The foremost step is to gather the images from various repositories or datasets so that the system can classify them by learning them. It is essential to feed the system with many images for better classification. The data for predicting airway diseases like pulmonary embolism, emphysema, tuberculosis, pulmonary edema, pneumoconiosis, cystic fibrosis, asthma, covid 19, and lung cancer has been collected from multiple data sources X-ray, CTscan (Computed Tomography), histopathology image, etc. Table 3 shows the detailed description of the dataset for all the respective airway diseases.

Table 3 Dataset of multiple airway diseases
Full size table

Pre-processing: After collecting the images from various repositories, it is our prime duty to pre-process them before training our system with them. It is because the images could be blurred, noisy or their features are not visible, and training the system with such images can affect the accuracy performance and generate the wrong output, which would be risky for people's health. There are various techniques by which the image can be pre-processed, such as Contrast Limited Adaptive Histogram Equalization (CLAHE) for increasing the image contrast, geometric transformations, image filtering, etc. [69]. Data pre-processing also includes data augmentation to increase the dataset size without collecting the new data and reduce overfitting. Data normalization is the process of organizing the data in a structured way. Multiple images are put into a common statistical distribution in terms of size and pixel values so that there can be a change in the range of pixel intensity values [70]. Resizing images is a critical pre-processing step since neural networks receive inputs of the same size. All images need to be resized to a fixed size before sending them as an input to the convolution neural network. After this, feature extraction could be conducted on the training images so that those features will feed the learning models to identify or predict the class of any new image. This process produces images that have been altered or modified and will be utilized in the training phase [71].

Learning models: Modern systems are deemed artificially intelligent when they use the machine and deep learning methods, which allows the computer (the machine) to understand tasks from an ever-changing dataset. Thanks to recent breakthroughs in learning algorithms and processing speed machines, deep learning models have become possible for many prediction problems [72]. Hence in this phase, machine learning algorithms, deep learning algorithms, transfer learning, and ensemble learning models can be selected based on various factors such as the size of the dataset, complexity of the data, etc. CNN (convolution neural network), which is excellent for finding image patterns, is a perfect algorithm for classifying things (and many other tasks involving images). CNN's, like neural networks in the real brain, are made up of neurons with trainable biases and weights that receive various inputs [73]. The information's weighted sum is then calculated. After that, the weighted total is input into an activation function, which results. As far as transfer learning is concerned, VGG16 (Visual Geometry Group), VGG19, MobileNetV2, ResNet50 (Residual Neural Network), and other pre-trained models are frequently trained on massive datasets, a standard benchmark in the computer vision area. These models can be used directly to predict new tasks or as part of a model's training process. Moreover, these techniques, such as transfer and ensemble techniques, are also used to minimize training time, enhance the accuracy of the classification and prevent modeling errors in the system [74].

Classification: In this phase, the trained model will determine which class a picture belongs to, such as emphysema, cystic fibrosis, pneumoconiosis, pulmonary embolism, pulmonary edema, asthma, mesothelioma, tuberculosis, covid-19, and lung cancer. If the image does not correspond to either of these disorders, the model should suggest that it is a normal lung image. To verify the system's performance, the model that correctly classifies the image should be evaluated using specific evaluative metrics such as accuracy, loss, precision, recall, F1 score, and so on (Table 4) [75,76,77,78,79].

Table 4 Evaluation metrics to test system performance
Full size table

4 Background

The work done by researchers to forecast the numerous airway diseases has been discussed in three subsections, i.e., pulmonary edema, pulmonary embolism, and covid 19; Mesothelioma and lung cancer; asthma, tuberculosis, and cystic fibrosis; emphysema and pneumoconiosis. To undertake comparative analysis, the goal or objective of their study has been initially briefed. Further, the datasets, procedures, outcomes, findings, and limits are given in the tabular form, i.e., from Table 5, 6, 7, 8, followed by the overall analysis.

Table 5 Comparative analysis of the techniques used to predict pulmonary edema, pulmonary embolism, covid-19
Full size table
Table 6 Comparative analysis of the techniques used to predict Mesothelioma and lung cancer
Full size table
Table 7 Comparative analysis of the techniques used to predict asthma, tuberculosis, cystic fibrosis
Full size table
Table 8 Comparative analysis of the techniques used to predict emphysema, pneumoconiosis
Full size table

4.1 Role of AI to Predict Pulmonary Edema, Pulmonary Embolism, Covid-19

Medical pictures are one of the first diagnoses since they can disclose specific undetected pathologic changes. Still, the absence of publicly available datasets and benchmark studies makes it impossible to compare and define the best detection systems. To solve this, Islam et al. [79] employed several datasets to analyze the effectiveness of models on various disorders, including pulmonary edema. To find the anomalies in chest X-rays, the investigator's employed trained classifiers. According to Liao et al. [80], one major obstacle in analyzing medical pictures is limited ground truth labels. As a result, they designed and tested a semi-supervised learning system for estimating pulmonary edema to aid therapeutic choices in congestive heart failure. To tackle the problem, they created a Bayesian model (Eq. 1) that learns the probabilistic feature representations from the complete picture collection and uses them to forecast the degree of edema

$$P\left(c|x\right)= \frac{P\left(x|c\right)P(c)}{P(x)}$$
(1)

where \(P\left(c|x\right)\) is the posterior probability, \(P\left(x|c\right)\) is the likelihood, \(P(c)\) is the class prior probability, and \(P(x)\) is the predictor prior probability,

The authors maximized the log probability of the data to construct the probabilistic feature representation (Eq. 2) that has been learned from all images in predicting pulmonary edema.

$$\mathrm{log}p\left(x,y; \theta \right)= \sum_{i=1}^{{N}_{i}}\mathrm{log}p({x}_{i},{y}_{i}; \theta )+ \sum_{i={N}_{i}+1}^{N}\mathrm{log}p({x}_{i}; \theta )$$
(2)

Here \(\theta\) is the parameters, N is the total number of images, y is the edema severity label, x represent the single image.

According to Hong et al. [81], patients with pulmonary edema have many consequences for the rest of their lives. Physicians must deal with stress, concern, and inconvenience when treating such patients because they may suffocate if they are not treated. As a result, the author advocated a comprehensive investigation on the application of auditory classification algorithms for the automated identification of excessive lung water that engorges the alveolar beds to address these issues. The authors developed a unique approach using recursive feature eliminations, logistic regression (Eq. 3), and principal component analysis (Eqs. 4–6) to validate the learned data with supplemented samples from local hospitals.

$${Y}_{i}={\beta }_{0}+{\beta }_{1}{X}_{i}+ {\varepsilon }_{i}$$
(3)

Here \({Y}_{i}\) is dependent variable, \({\beta }_{0}\) is population y-intercept, \({\beta }_{1}\) is population slope coefficient, \({X}_{i}\) is an independent variable, and \({\varepsilon }_{i}\) is random term error. Given a dataset of N centered observations in a d-dimensional space

$$X=\left\{{x}_{1} , {x}_{2} , {x}_{3} , \dots \dots \dots ., {x}_{N}\right\},\sum_{k=1}^{N}{x}_{k}=0, {x}_{k}\in {R}^{d}$$
(4)

PCA diagonalizes the covariance matrix where C is a covariance matrix,

$$C= \frac{1}{N} \sum_{K=1}^{N}{X}_{K}{X}_{K}^{T}$$
(5)

To solve the Eq. (5) using eigen values

$$\lambda v=Cv =>\lambda \left({X}_{k}.v\right)=\left({X}_{k} .Cv\right), where k=\mathrm{1,2} ,\dots .,N$$
(6)

Lin et al. [82] investigated the diagnosis and progression of illnesses to integrate bio-fluid-based infrared spectroscopy into the clinical area. The authors looked at using Fourier transform infrared micro spectroscopy to detect abrupt cardiac death. Assessing the degree of pulmonary edema is complex, and detecting it in chest radiographs allows doctors to make prompt treatment decisions for patients. Using the large-scale clinical MIMIC-CXR (Chest X-ray Scan) database, Kumar et al. [83] tested different supervised and semi-supervised deep learning approaches to detect the degree of edema from radiological pictures. Furthermore, the authors evaluated three ways to alleviate class imbalance during implementation: weighted cross entropy loss (Eqs. 7–8), class aware sampling, and random minority oversampling. Kumar et al. [84] reported a texture analysis of chest X-ray to detect pulmonary edema in chest X-rays automatically. To put it another way, the authors were able to tell the difference between a chest X-ray with indications of pulmonary edema and a normal chest X-ray.

$${l}_{i}= -\sum_{c \in C}{y}_{i}, c \mathrm{log}({p}_{i}, c)$$
(7)
$${ l}^{w}=\sum_{i}{m}_{i}{l}_{i}$$
(8)

where \({l}_{i}\) is the ith window's component of the multiclass cross-entropy loss, \({l}^{w}\) is the weighted cross-entropy function, p is the probability that the window i belongs to class c as predicted by the given model, \({m}_{i}\) is the weight applied to the \({i}^{th}\) component. According to Hayat et al. [85], quick screening of pulmonary edema patients is required such that radiologists can make a prediction as soon as feasible. However, depending on specialists' knowledge of reasoning impedes the diagnostic process. As a result, the author created a deep learning-based architectural model to detect the presence of acute pulmonary edema in chest X-ray pictures. Brestel et al. [86] sought to deliver expert-level information to every chest X-ray image right away. To achieve expert-level automatic interpretation of regular chest X-rays, they used a machine learning approach and discussed the results using a robust approach of clinical validation.

Kiourt et al. [87] used transfer learning methodologies to adopt and analyze some of the most common convolutional neural network designs to get decent model accuracy for detecting pulmonary embolism in CT scans. Deep convolutional neural network (DCNN) exhibited good results in identifying critical abnormalities in CT images, including intracranial haemorrhage, acute brain ischemia, and essential abdomen findings, according to Weikert et al. [88]. An end-to-end, fully convolutional network (i.e., U-Net) was created by Liu et al. [89] to segment clots and determine clot volume in CTPA (CT pulmonary angiogram). This study aimed to evaluate U-ability Net's to identify clots in terms of efficiency and accuracy and compute the APE clot load (acute pulmonary embolism). Rucco et al. [90] suggested CAD (computer-aided drugs) for the mathematical theory of Hyper-networks and Q-analysis, allowing for a dimensional patient dataset description. The findings were utilized for feature selection in the artificial neural network training stage. The authors were able to identify and diagnose pulmonary embolism in that manner.

Since we know that the number of Covid-19 cases has risen dramatically in the last year, it's become even more critical to track and identify healthy and infected persons quickly and precisely. Many existing detection approaches are ineffective in detecting viral patterns. As a result, Chen [91] developed an excellent classification approach for detecting COVID-19 viral sequences. Ismael et al. [92] employed an AI-based technique that effectively monitors various lung illnesses. For deep feature extraction, the authors operated pre-trained CNN models with SVM (support vector machine) classifiers with different kernel functions such as linear, quadratic, cubic, and Gaussian, used for covid-19 classification.

Similarly, according to Punn et al. [93], artificial intelligence experts have concentrated their professional knowledge on constructing mathematical models for assessing the epidemic condition utilizing state-wide shared data. As a result, the authors advocated a few AI models to analyze everyday exponential behaviour and the projection of future corona virus reach-ability across countries using real-time data. The data science community has developed many machine learning (ML) models to improve Covid-19's diagnostic capabilities. Most are based on computed tomography (CT) images or chest X-rays. Cabitza et al. [94] used various classifiers by including machine learning techniques for blood-test results, which are generally available in clinical practices. Panwar et al. [95] focused on observational lockdown analysis and said that artificial intelligence technologies were necessary to defeat such a solution. With this in mind, the scientists presented a CNN-based algorithm, emphasizing that lockdown isn't the only way to combat the covid19 epidemic. Elaziz et al. [96] demonstrated a technique for accurately classifying covid-19 chest X-ray pictures. The characteristics of orthogonal moment features and feature selection approaches determine the classification strategy. The authors created a novel feature selection approach based on several assessment techniques to improve the behavior of Mamta ray foraging optimization. Several aspects may be used to identify viral infections based on imaging patterns. Wang et al. [97] hypothesized that CNN might aid in identifying distinctive characteristics that would be difficult to detect using visual recognition alone.

4.2 Role of AI to Predict Mesothelioma, Lung Cancer

Alam et al. [101] aimed to look for clinical, radiological, and histological variables in malignant Mesothelioma. The authors suggested a novel framework for identifying prognostic indicators utilizing non-invasive and cost-effective methods based on various techniques. According to the authors, their suggested framework would aid medical professionals and healthcare experts in detecting malignant Mesothelioma early and treating it more effectively by including crucial prognostic markers. In their study, Latif et al. [102] looked at the risk factors for malignant Mesothelioma. The scientists employed a dataset that included healthy people and Mesothelioma patients, but only Mesothelioma patients were chosen for symptom identification. According to the authors, these findings will aid in managing MM-related co-morbidities such as cardiovascular disease, cancer-related mental distress, diabetes, anemia, and hypothyroidism. Choudhry et al. [103] employed artificial intelligence-based algorithms to offer the best system for Malignant Pleural Mesothelioma early identification and prognosis (MPM). According to the authors, decision tree models, random forest, have a risk of overfitting; hence, they created a model that can diagnose with or without pricey biopsy data to overcome the flaws outlined above.

Similarly, Alam et al. [104] concentrated on investigating MM risk variables. The scientists included ill and healthy people in their study, resulting in a larger dataset. The dataset has a class imbalance problem, with the number of malignant Mesothelioma patients being much lower than the number of healthy people. Furthermore, the numerical attributes were categorized as nominal attributes, and association rules were created in the dataset. To detect malignant Mesothelioma, Gupta et al. [105] compared numerous machine learning algorithms with different feature sets to address the class imbalance problem. To achieve this, the authors used three sampling techniques: resampling, synthetic minority oversampling technique (SMOTE), and adaptive synthetic sampling (ADASYN). They also used other dimension reduction approaches, such as the ordinary least square method (OLS), principal component analysis (PCA), and random forest feature selection (RFFS), as well as genetic algorithms, to determine the exact collection of features. MesoNet was developed by Courtiol et al. [106]. It successfully predicted survival rates of Mesothelioma patients using whole-slide digitized images with no need for pathologist-provided locally labeled regions. The researchers also confirmed that the method was more suitable for estimating the patient's survival rate than conventional pathological approaches.

Several researchers have proposed several CNN strategies for lung cancer detection. These models, however, could not give the expected detection accuracy. Sori et al. [107] utilized a retraining technique, i.e., multi-phase CNN. The initial training was identical to the usual one. It was based on fine-tuning, which just used the selected parts of the model to describe and learn distinct morphology of lung nodule characteristics for contextual information, as shown in Fig. 5. Much work has gone into developing computer-assisted diagnosis and detection methods to increase the diagnostic quality for lung cancer detection categorization. As a result, Asuntha et al. [108] attempted to identify malignant lung nodules in the input picture and categorize lung cancer according to severity. The authors employed several optimal feature extraction approaches such as local binary pattern (LBP), scale-invariant feature transform (SIFT), and others to extract textural, geometric, and intensity features. Machine learning models improve the model's performance by learning from previous experiences. These models also seek to identify practical factors and their relationships. As a result, Xie et al. [109] discovered clinical metabolic markers that demonstrated significant differences between lung tumor patients and healthy persons. According to the authors, biomarkers were also employed to distinguish between histological subtypes and illness degrees.

Fig. 5
figure 5

Architecture based on multi-pathway CNN to detect lung cancer [107]

Full size image

Because they had a restricted lung cancer image dataset that did not work out using typical transfer learning and data augmentation techniques, Ausawalaithong et al. [110] implemented the transfer learning model twice to improve its performance. The first occurred when the model was transferred from the domain of public image to the chest X-ray. The model was then applied to lung cancer for the second time. According to the authors, multi-transfer learning solved the sample's limited size and produced greater results on the test than classic transfer learning. Shakeel et al. [111] introduced an intelligent machine learning approach to improvise the lung detection process. The CT scan-based lung pictures were continually evaluated using a multilayer brightness-preserving method that reduced image noise and improved lung image quality. Due to the segmentation process's relevance, the scientists adopted a multilayer augmented deep neural network technique to extract the cancer-affected area. Pradhan et al. [112] researched to predict various illnesses to make a judgment on lung cancer prediction. The authors also provided a comprehensive analysis of several machine learning algorithms to assess their competence and performance in predicting lung cancer and, as a result, detecting lung cancer with IoT integration. In a retrospective study of the USA National Lung Screening Trial (NLST), Kadir et al. [113] indicated the influence of lung imaging reporting and data system (lung RADS). Although lung RADS has been shown to lower the total number of benign nodules during screening, its categorization task has proven complicated. As a result, the authors addressed the issue by recommending that radiologists and pulmonary medicine specialists use computer-assisted technology as a tool. As a result, the authors analyzed their study's progress in developing and validating the lung cancer predictive model and nodule categorization.

4.3 Role of AI to Predict Asthma, Tuberculosis, Cystic Fibrosis

Awal et al. [114] applied multiple learning models to investigate the parameters that characterize asthma diagnosis and prediction. BOMLA (Bayesian Optimisation-based Machine Learning Framework for Asthma), a new machine learning technique, had been developed for identifying asthma. Khasha et al. [115] set out to identify asthma control levels, and they only looked at research that used supervised approaches like classification models in data mining. Their primary goal was to improve the effectiveness of classification algorithms by investigating the impact of daily clinical data's time-series/time-sequences dynamics on asthma control level detection in patients. Zhang et al. [116] identified severe asthma exacerbations based on freely accessible daily monitoring data. Compared to previously published models, the authors hypothesized that a predictive algorithm created utilizing machine learning techniques and an extensive training dataset of daily monitoring data would yield greater accuracy for identifying asthma exacerbations, as shown in Fig. 6.

Fig. 6
figure 6

Ways to predict asthma exacerbation patients

Full size image

Several studies have employed Mahalanobis–Taguchi System (MTS) for intelligent illness detection with reasonable accuracy, according to Zhan et al. [117]. Their research aimed to see if MTS could be used to diagnose asthma based on regular blood measurements from healthy people and people living with asthma.

Tuberculosis can be cured by diagnosing it at its early stages, and to make it happen, the main requirement is to use the diagnostic technologies properly. Hence, Dasanayaka et al. [118] presented a model that can detect TB using deep generative adversal network. The chest X-ray images were chosen based on the subjective and objective quality assessment metrics. The objective quality assessment metrics were selected as peak signal-to-noise ratio (PSNR) (calculated by Eqs. 9–10), and the radiologists performed a subjective evaluation.

$$PSNR=10{log}_{10 } \left(\frac{{MAX}^{2}}{MSE}\right)$$
(9)

And to calculate (mean square error) MSE, we have an Eq. 10 where I and K are the observed and predicted values, m &n are the data points respectively

$$MSE= \frac{1}{mn} \sum_{i=0}^{m-1}\sum_{j=0}^{n-1}[I \left(i,j\right)- k\left(i,j\right)]$$
(10)

According to Panicker et al. [119], traditional microscopic sputum smear screening for TB (tuberculosis) diagnosis is time-consuming and error-prone. To make the diagnostic procedure more accessible, the scientists employed an image binarization approach and a modified convolution neural network to detect TB by identifying the pixels in the picture corresponding to bacilli. Gao et al. [120] concentrated on using cutting-edge deep learning approaches to analyze CT pulmonary images. The authors identified five categories of severity for tuberculosis to track therapy effectiveness. The latter was the subject of Hwa et al. [121]. They detected tuberculosis using contrast-enhanced canny edge detected (CEED-Canny) X-ray images which generated edge detection in lung X-ray pictures. Li et al. [122] developed a new model for generating quantitative computed tomography images in a clinic for diagnosing pulmonary TB. Based on CT data, the scientists used a fine-tuned 3D CNN model to classify pulmonary TB lesion areas. Their mission was to digitally assess the spatial position of each lesion, the confidence of each infection, the presence of calcification, lesion type classifications, overall infection likelihood, and practical volume of the left and right lungs. According to Lai et al. [123], machine learning technologies identify and diagnose various disorders. The effectiveness of artificial neural networks and support vector machine in predicting the development of prostate, ovarian, breast, and liver cancer were outstanding. As a result, the researchers utilized machine learning-based three algorithms to predict anti-tuberculosis drug-induced hepatotoxicity. According to Barros et al. [124], determining the severity of illnesses is necessary to improve patient quality of life, efficiently manage health resources, and so on. As a result, they focused on evaluating nine machine learning models, including KNN (calculated by Eq. 11), Nave Bayes, and decision trees, on improving TB prognosis to predict the chance of mortality using patient geographical, medical, and laboratory data. To summarise, the authors used feature selection techniques to identify the most relevant fields, used randomized search techniques to choose the optimal hyper-parameters of the machine learning model, and proposed an ensemble learning model to achieve better results.

$$d\left(p,q\right)=d\left(q,p\right)=\sqrt{{({q}_{1}-{p}_{1})}^{2}+{({q}_{2}+{p}_{2})}^{2}+ \dots .. + {({q}_{n}+ {p}_{n})}^{2}}=\sqrt{\sum_{i=1}^{n}{({q}_{i}-{p}_{i})}^{2}}$$
(11)

Here the two points in Euclidean n-space are p and q. According to Das et al. [125], a chest X-ray is a potential signal for detecting TB. However, the shortage of competent radiologists in limited resource areas exacerbates the problem. As a result, the authors wanted to employ an end-to-end deep learning TB screening tool based on chest X-ray images so that their system could adjust to changed data over time.

The cystic fibrosis transmembrane conductance regulator is widely distributed in the colon and plays a crucial function in controlling gut secretion viscosity and pH. As a result, Malagelada et al. [126] evaluated gut function using imaging methods often used to detect anatomical abnormalities of the digestive system. The authors also predicted that combining internal (endoluminal) and exterior (MRI) imaging gives a new viewpoint on the relationship between cystic fibrosis anatomical and functional findings. Zucker et al. [127] investigated if a deep CNN model might aid automated Brasfield rating of radiographic images of a chest for cystic fibrosis sufferer that would be analogous to that of a radiologist. Marques et al. [128] developed a texture classification-based technique to detect cystic fibrosis anomalies. The researchers used convolutional neural networks, as shown in Fig. 7, in two ways, in which the first one detected aberrant tissues, and the next determined the sort of structural abnormalities. The authors also presented a network that solely used patch-wise annotations to compute pixel-wise heatmaps of present irregularities.

Fig. 7
figure 7

Architecture of cascade approach for the detection of cystic fibrosis [128]

Full size image

Dio et al. [129] stated that physicians were given tools to recognize airway disorders in an automated and quick manner. Their goal was to produce a proof-of-concept and the first favorable results that might lead to rapid, reliable, and automatic detection of such disorders, similar to how the Sweat Chloride Test was used to diagnose Cystic Fibrosis (CF). Likewise, Zhou et al. [130] proposed a novel technique for cystic fibrosis diagnosis based on combining desorption electrospray ionization mass spectrometry with gradient boosted decision trees (calculated by Eqs. 12, 13) that divided the set of data into sub-sets, where S, X is a discrete random variable and P is a probability.

$$E\left(S\right)= \sum_{i=1}^{c}-{\mathrm{p}}_{\mathrm{i}}{log}_{2}{p}_{i}$$
(12)

and entropy using the frequency table of two attributes is given as

$$E\left(S,X\right)=\sum_{c\in X}P\left(c\right)E(c)$$
(13)

to analyze perspiration samples.

4.4 Role of AI to Predict Emphysema, Pneumoconiosis

Emphysema is a condition that causes difficulties in breathing and needs to be detected early with computed tomography scans and primary function testing. On the other hand, the challenges involved with specific diagnostic methods have prompted additional computer-assisted treatments to flourish. As a result, Mondal et al. [131] used deep learning networks to conduct automated pulmonary emphysema diagnosis, resulting in increased detection accuracy. Bortsova et al. [132] investigated a weakly labeled strategy comparable to multiple instance learning. Using this method, the authors measured emphysema by leveraging %age labels and applying existing information on the nature of these labels. The authors developed an approach that used a custom loss to learn the intervals and a label proportion problem-specific architecture (LPP). Deep learning has made significant progress in various complex tasks of processing images. In light of this, Humphries et al. [133] used the Fleischner technique to classify emphysema using chest CT image processing. The authors' primary goal was to see if emphysema patterns at the participant level may predict disability and death when identified using a deep learning algorithm. Srivastava et al. [134] employed neural networks to help medical experts diagnose Chronic Obstructive Pulmonary Disease by providing a complete and systematic assessment of clinical pulmonary audio data. The authors used ten splits of K-fold Cross-Validation (defined by Eq. 14) to escalate the presentation of the existing deep neural networks for N observations.

$$let k=\left\{1\dots N\right\}\to \left\{1\dots L\right\}$$
(14)

which is an indexing function that indicates the division to which randomization assigns observation ‘i’. The fitted function that is computed with the kth part of the removed data is denoted by \({\widehat{\mathrm{f}}}^{-\mathrm{k}\left(\mathrm{i}\right)} \left(\mathrm{x}\right)\). The estimate prediction error using cross validation (CV) is shown by Eq. (15)

$$CV\left(f\right)=\frac{1}{N} \sum_{i=1}^{N}L ({y}_{i}, {\widehat{f}}^{-k\left(i\right)} \left({x}_{i }\right))$$
(15)

According to Nyboe et al. [135], automatic evaluation of emphysema presence might produce highly steady forecasts at a far cheaper cost, so it could be a viable alternative for an expert to do the assessments. As a result, they proposed a multiple instance learning (MIL) technique for emphysema detection. They also looked at whether emphysema at the scan level is sufficient to train the system to predict emphysema occurrence at the area level. Improving the detection of extreme inactivity (EI) in COPD patients can reduce morbidity and death. Except for patients with evident EI, detecting such conduct in a real-life session is impossible. As a result, the authors such as Aguilaniu et al. [136] presented a machine learning technique to test for excessive inactivity. They created a prediction system that could accurately identify EI patients who would benefit the most from therapies like pulmonary rehabilitation. Gonzalez et al. [137] introduced a biomarker estimated approach based on deep learning techniques that relied on a regression network. In this system, the input was supplied to algorithm photos having the structure where the biomarker was generated, and the output was shown directly as the biomarker value. The two indicators used to demonstrate the suggested regression architectures were emphysema and bone mineral density (BMD), which are medically necessary. Pena et al. [138] described a technique for autonomously estimating emphysema areas in patients with chronic obstructive pulmonary disease (COPD) using High-Resolution Computed Tomography (HRCT) pictures that don’t require manually labeled records for training. The authors wanted to use HRCT images without local annotations to detect emphysema sites in COPD patients.

Wang, X et al. [139] wanted to see if deep learning could be used to consider pneumoconiosis on digital chest radiographs as well as contrast its performance to that of qualified radiologists. The scientists used a conventional deep convolutional neural network to analyze chest X-ray pictures and verify them with various parameters. Furthermore, the scientists requested two trained radiologists to assess the testing dataset separately and compare their results to the computerized system. Hao et al. [65] presented an electronic health record-based pneumoconiosis radiography dataset. According to the authors, recent research has focused on machine learning for computer-aided detection. These have attained remarkable precision, with the artificial neural network (ANN) doing exceptionally well.

Nevertheless, wide use in clinical practice has been challenging due to unbalanced samples and a lack of readability. Hence to address such issues, the authors initially created a pneumoconiosis-based radiograph dataset, which included both un-favorable and favorable representatives. Secondly, deep convolutional diagnostic methodologies were examined in identifying pneumoconiosis, and balanced training was used to improve recall. Pneumoconiosis diagnosis is based mainly on chest radiographic images, according to Wang et al. [140], and there is substantial disagreement across clinicians. Zhang et al. [141] set out to create an artificial intelligence (AI)-based model that would aid physicians in pneumoconiosis diagnosis as well as grading using chest radiographic images. The chest radiograph system was created with the help of a training cohort and validated with the help of an independent assessment cohort. Their groundbreaking research evaluated the possibility and effectiveness of AI-assisted radiography diagnosis and screening in the field of occupational lung disease. To surmount the issues of insufficient annotation pneumoconiosis data and increase the accuracy of pneumoconiosis diagnostics, Zheng et al. [142] presented two transfer learning models. They also demonstrated various pre-processing techniques for improving the quality and accuracy of X-rays, such as segmentation of lungs and amplification of data. Many computer-aided investigations on pneumoconiosis classification algorithms have been offered, according to Zhang et al. [143], but most of them were based on lung pictures. As a result, the authors provided a technique for diagnosing pneumoconiosis using wrist pulse signals gathered from non-pneumoconiosis individuals and pneumoconiosis patients. Machine learning approaches were utilized to process and assay the pulses of non-pneumoconiosis persons and the pneumoconiosis patient. Other than specialist radiologists, Wang et al. [144] noted a lack of sequential, automatic, and primary procedures for identifying and analyzing the evolution of pneumoconiosis in every coal miner. Consequently, the authors presented the most recent research findings from a study to address the challenges described by creating Computer-Aided Diagnosis (CAD) tools to identify pneumoconiosis using chest X-rays automatically.

4.5 Overall analysis

The best techniques have been filtered out based on their respective accuracy in Table 9 after observing the performance of the models that have been used for the detection and diagnosis of multiple airway diseases such as pulmonary edema, pulmonary embolism, cystic fibrosis, pneumoconiosis, lung cancer, asthma, covid-19, Mesothelioma, tuberculosis, and emphysema.

Table 9 Overall comparison of models
Full size table

5 Discussion

5.1 RQ 1: Year Wise Analysis of Predicting Multiple Airway Diseases Using AI Techniques

As demonstrated in Fig. 8, one hundred fifty-five (155) papers that range over 12 years, i.e. from (2010 to 2022)* have been studied for identifying different types of airway diseases using AI techniques which include pulmonary edema, cystic fibrosis, emphysema, Mesothelioma, pneumoconiosis, pulmonary embolism, lung cancer, covid19, tuberculosis, and asthma.

Fig. 8
figure 8

Distribution of papers

Full size image

One hundred papers have been taken from recent years i.e. 2019 to 2022, and the remaining fifty-five papers have been taken from 2010 to 2018. This research intends to create a broader sense of the various types of AI and AI-derived techniques, such as machine and deep learning techniques for detecting and diagnosing multiple airway diseases.

*- (As per the sources, the papers from the past ten years can be considered for SLR, but in this paper, we had to include some other information, such as a dataset of any particular disease, etc. for which we studied the papers atmost from the past 12 years.)

5.2 RQ 2: How Doctors are Being Helped by Deep and Machine Learning Techniques in Detecting the Airway Diseases?

The goal of an AI system is to organically learn a function and continue to improve without being actively taught. Machine learning and deep learning are artificial intelligence technologies that allow systems to detect patterns and connections between data and desired outputs [145]. These techniques connect multiple chunks of information discovered from the facts without requiring explicit human description [146]. At its most basic level, a machine and deep learning e-based strategy lead to a better diagnosis by analyzing a more excellent range of data than a physician.

We know that medical treatment is substantially affected by the increasing amount of healthcare data which makes it very difficult for pulmonologists to handle and analyze it for treating their patients manually. Hence, to improvise such a patient-doctor relationship, clinicians use AI techniques to predict, classify, and diagnose diseases efficiently. The stats have been shown in Fig. 9, which clearly defines the accuracy of AI techniques in predicting and analyzing the diseases as compared to the traditional methods used by pulmonologists.

Fig. 9
figure 9

Analysis between Pulmonologists and AI techniques in predicting the diseases

Full size image

One advantage of employing AI to interpret diagnostic exams such as imaging is the ability to evaluate tests done in geographically isolated or underserved places [147]. This may result in a more early and accurate diagnosis, as well as a referral to expert treatment at an earlier stage of the disease, potentially influencing the prognosis Radiographs from these centers, on the other hand, can be remotely submitted and analyzed by a single central system using AI [148]. In Fig. 10, it has been shown that artificial intelligence techniques have successfully proven to be better than pulmonologists in predicting airway disorders.

Fig. 10
figure 10

Airway Diseases Prediction by AI and Pulmonologists

Full size image

5.3 RQ 3: Which ML and DL Techniques are Broadly Applied to Predict Airway Diseases?

Patients who require early diagnosis and treatment can benefit from AI-driven sickness detection models that help medical companies produce improved diagnostic tools. As a result, the literature mentions the techniques for diagnosing airway disorders such as cystic fibrosis, pneumoconiosis, pulmonary edema, pulmonary embolism, covid-19, emphysema, tuberculosis, lung cancer, asthma, and mesothelioma. Linear regression, support vector machine (SVM), random forest (RF), decision tree (DT), nave Bayes (NB), fuzzy particle, logistic regression, and ensemble learning are the most commonly used machine learning models in the literature. Convolutional Neural Networks (CNN) are the most often utilized deep learning models for illness diagnosis. Furthermore, artificial neural networks (ANN) and transfer learning techniques such as VGGNet, ResNet, and others have been extensively used as shown in Fig. 11.

Fig. 11
figure 11

Machine and deep learning based prediction models

Full size image

5.4 RQ 4: Name the Characteristics that Manipulate the Quality of Prediction Models Based on Deep and Machine Learning?

AI algorithms in the medical industry are essential, notably for identifying a disease from the medical database. Many firms employ these approaches to predict illnesses early and better medical diagnostics. Irrespective of their continual progress, there are specific problems that remain. On the one hand, machine and deep learning algorithms can handle any complicated situation, but on the other hand, they also demand more research effort for practical implementations [149]. Hence, certain limitations occur while utilizing machine and deep learning models to diagnose airway illnesses gathered from research gaps and described below.

Data Paucity One of the most pervasive issues in artificial intelligence is a shortage of high-quality data. Every business will encounter this difficulty throughout the AI deployment process. Thus, in the future, open-source datasets of multiple airway diseases should be considered, though they occasionally lack quality but represent promising solutions for organizations [150]. In addition, synthetic data should be created to increase data security and privacy; data augmentation should be considered to enhance the size of the dataset without accumulating additional data; and finally, transfer learning techniques should be used when we have enough training data. Modeling errors: Two main restricting errors are overfitting and underfitting. This is called overfitting, when a model learns the information and noise in the training dataset to the point that it degrades its performance on a new dataset. In underfitting, models cannot train or generalize new datasets, thus impairing the system's performance [151]. Therefore, algorithms like pre trained models should be incorporated that work on system modeling errors in the future. Perpetual improvement of models: At times, optimizing a model's performance might be difficult. This is because optimizing a model's performance makes it more accurate in predicting and the most dependable and acceptable in artificial intelligence. Developing an AI-based model is not difficult for engineers while verifying its performance is critical to obtaining accurate and trustworthy results [152]. Hence to improve it, in the future, the appropriate amount of data should be utilized, the proper algorithms should be used, and models should be verified and evaluated appropriately. Transfer learning can also enhance the performance of AI-based models [153]. Class imbalance: An unbalanced classification issue is when the distribution of instances across recognized classes is uneven or biased [154]. Imbalanced classifications provide a problem to predict the class because most machine learning methods for classification are built on the premise of an equal number of instances for each class. So, in the future, Re-sampling, K fold cross-validation, or pre-trained models can be used to deal with class imbalance issues [155].

6 Conclusion and Future Work

Between 2010 and 2022, 155 studies were chosen from six digital libraries that can be accessed online, and four questions were investigated after going through them. Researchers analyzed a variety of technical developments that might be used to enhance AI-based models in the field of pulmonology. This article discusses the influence of machine learning and deep learning techniques on analyzing the information related to airway disorders. Furthermore, the paper details several academics' rigorous research efforts to demonstrate how machine learning and deep learning models help to detect or categorize various airway illnesses. The literature's systematic review is presented in a tabular format, with each column indicating the dataset, methodology, and outcomes used by the researchers and their limitations. After accumulating the constraints encountered by researchers for the prediction of airway problems, an attempt has been made to determine the usage of the latest models that may be incorporated in the future to enhance the system's performance. The lasso method, CNN, Decision Tree, GAP Net, and other algorithms were examined by certain researchers. However, their findings revealed that the models could not distinguish between diseases or detect anomalies in their data. Aside from that, several models only worked with limited data sets, failed to pre-process data, and refused to provide localization information for the final image. The model's overall performance is hampered due to these limitations, which must be addressed. Several approaches used in research publications, such as random forest and logistic regression, have the lowest prediction accuracy because of modeling flaws such as overfitting and underfitting. Many researchers have also struggled to categorize data using DFD-Net, Fuzzy Particle effectively, Swarm Optimization to detect lung cancer, CNN to detect pneumoconiosis, Dense Net, Inception V3, ResNet, and DNN to diagnose pulmonary embolism, and Canny Edge to detect tuberculosis. It's also worth noting that researchers have only focused on one or two airway disorders for prediction, which limits users' capacity to distinguish between other types of airway difficulties.

Although artificial technology has innumerable benefits, its flaws and limitations may restrict its applications, notably in the healthcare business. But there is also scope for improvement in this field to overcome the issues such as the best selection of models which can be further improved or innovated to provide better output, and optimization techniques should be added to the network to provide the optimal results. The most important task while using the learning models is correctly classifying the data. Hence, such a methodology that connects the data's features to improve the classification model and generates correct output should be included. Besides this, deep learning techniques should be improved to be applied to a dataset where they can detect various airway diseases using less processing time. In addition, the conventional learning methods should also be considered in the future as they can be used to improve the airway disease detection model.