Key Words:
Artificial intelligence; Allergic rhinitis; Diagnosis; Deep learning; Machine learning; Ensemble learning Introduction
Allergic rhinitis (AR) is a common chronic inflammation of the upper respiratory tract. It has been considered as a type of stubborn disease that seriously affects people’s daily lives. The prevalence of this disease is showing a high trend globally. About 500 million people worldwide suffer from AR, with the highest prevalence in developed regions such as Western Europe, Northern Europe and North America, generally 12-30%[1]. An AR epidemiological survey of Chinese adults showed that it rose from 11.1% in 2005 to 17.6% in 2011[2,3]. It is a type I allergic disease mediated by IgE with multiple cytokines involved. The pathogenesis of AR is related to many factors, and the specific pathogenesis is not yet clear. Various cells, proteins and cytokines produced by the patient’s body may participate in or promote the occurrence and development of AR.
The typical symptoms of AR are paroxysmal sneezing, watery nasal discharge, itchy nose and stuffy nose, which may be accompanied by eye symptoms including itchy eyes, tearing, redness and burning sensation, etc. The main signs of AR are bilateral swelling of the nasal mucosa, edema of the lower turbinate, and a lot of watery discharge in the nasal cavity. The main signs of AR are bilateral nasal mucosa pale and edema, inferior turbinate edema, and a large amount of watery discharge in the nasal cavity[4]. The allergic signs of the eye are mainly hyperemia and edema, and AR patients accompany with asthma, eczema and dermatitis also have other signs of lungs and skin. In addition to symptoms and signs, the diagnosis of this disease also depends on the detection of allergens, including in vivo tests (skin prick test SPT) and in vitro tests (blood tIgE and sIgE tests), and nasal provocation test[5]. In addition, nasal secretion smears and sIgE in nasal lavage fluid are also helpful for clinical diagnosis[6]. Endoscopy or computed tomography (CT) can observe changes in signs such as hypertrophy of the turbinate, swelling of the mucosa, and help to diagnosis of diseases such as sinusitis and nasal polyps[7].
The diagnosis of AR is mainly based on symptoms and signs, as well as laboratory tests, but due to the limitations of outpatient conditions in China, some tests are not routinely operated, such as nasal provocation test, nasal secretion smear, etc[8]. Although the nasal provocation test is the gold standard for the diagnosis of AR, it has risks and is not clinically used as a routine method. Based on medical history,it can be divided into intermittent AR: symptom onset <4 d/week, or <4 consecutive weeks and persistent AR: symptom onset ≥ 4 d/week, and ≥ 4 consecutive weeks. And according to the severity of the symptoms, it also can be divided into mild AR: mild symptoms, no significant impact on quality of life (including sleep, daily life, work and study) and Moderate-severe AR: severe symptoms, affecting quality of lifesignificantly (including sleep , daily life, work and study).
Although there are more feasible diagnostic criteria, in clinical practice, experienced doctors are still required to make an accurate diagnosis based on medical history, examination, living habits, etc. However, due to individual differences and limitations of inspection methods, inconsistencies in diagnosis may still occur. Artificial intelligence(AI) is a cutting-edge and cross-disciplinary disciplinethat develops theories, methods, technologies, and application systems for simulating,extending, and expanding human intelligence[9].AI has been widely used in various industries in recent years, and has developed powerful mathematical models algorithm such as decision trees, naive Bayes and artificial neural networks (ANN), which are used in intelligent control, pattern recognition, prediction and other fields. In recent years, ensemble learning can organically combine multiple prediction results obtained by multiple single learning models to obtain more accurate, stable and strong final results. And ensemble learning models such as Boosting, Bagging and Random Forest(RF) have been proposed one after another and applied to various types of data sets. This study hopes to explore the application of AI ensemble learning in AR clinical diagnosis through the deep learning of ensemble learning models in big data, data analysis of more than 2,000 clinical cases in outpatient service in combination with the typical characteristics of Chinese AR.
Materials and Method
1.Sample source
clinical samples of nasal inflammation came from Tongji Hospital and Shanghai Anting Hospital, and the data collection time was 2019.4.1-2020.3.31. A total of 2231 case data were collected. The collected cases were patients with a preliminary diagnosis of suspected AR. Among them, 1335 were male (59.84%) and the average age was (35.39±19.71) years; 896 were female (40.16%) and the average age was (37.69±17.94) years old. All patients’ Clinical history were obtained, including time, name, age, gender, course of disease, four symptoms: sneezing, runny nose, itchy nose, stuffy nose, two eye symptoms. The physical signs include nasal polyps and nasal secretions. Blood tests include blood routine examination, total IgE, allergen SIgE, and CT imaging tests.
This study mainly collected cases of AR and included 6 types of diseases with similar symptoms: Rhinosinusitis (RS), Chronic rhinitis(RS), upper respiratory tract infection (URI), nasal septum deviation (NSD), adenoid hypertrophy (AH) and others (OTH contains nasal tumors, etc.) and collected clinical data such as medical history, clinical symptoms, allergen detection and imaging.
The diagnosis of AR combined with medical history and clinical symptoms can be divided into four types: mild intermittent, mild persistent, moderate - severe intermittent and moderate-severe persistent. The clinical symptom score was calculated using the total nasal and ocular symptom scores (TNSS and TOSS), which were scored from four aspects: stuffy nose, runny nose, itchy nose, and sneezing.Finally, it is divided into four gradesas 0: no symptoms; 1 : mild; 2 : moderate; 3 : severe[10].
2.Experimental setup and algorithm structure design
The data records a total of 66 features including 16 symptoms and signs including eye symptoms, nasal cavity examination, and runny nose. The presence or absence of symptoms and signs are represented by 1 or 0respectively. The classification method based on association rules used in the framework is compared with other classification methods, the former is the decision tree induction method (C4.5) and the latter is the probability classification method [11].
The classification of AR symptoms is a special multi-marker learning problem, that is, a patient may be combined with other diseases at the same time. And at the same time, some labels are mutex. For example, a patient without AR should not be diagnosed with intermittent mild classification, or a patient cannot have both intermittent and persistent AR. To solve such multi-label classification problems, problem conversion method and algorithm adaptation method are usually used.
Both transformation ideas were used in this study. Convert traditional multi-label classification into multiple binary classification problems with equal number of labels, and then use various basic machine learning algorithms to train each model to build an ensemble classification model based on multi-label classification,as shown in Figure 1. Table 1 shows the different classification methods used for various rhinitis samples and types in the comprehensive classification model.
One-Hot-Encoding is used in the analysis to encode all cases. One-hot encoding is also known as one-bit effective encoding.This method is to encode N states with N-bit 0-1 features. Each state has its own 0-1 feature bit, and at any time, only one valid. One-Hot coding can handle non-continuous numerical features, and to some extent, it also expands the features. For example, case A has clinical symptoms such as ophthalmia, turbinate hypertrophy, and clear secretions. The value of case A under these symptoms is 1, and there is no tearing, pale mucosa, or mucosal congestion. The value under these symptoms is 0. Finally, the doctor diagnosed the patient as AR and nasal septum deviation,so these corresponding values are 1, and the values of other symptoms that have not appeared are 0. The specific data form of the case is shown in Table 2.All case data were processed as a symptom-diagnosis input vector for the symptom classification model
3.Unbalanced data processing
For multi-category classification, class imbalance methods include SMOTE, ADASYN, All-KNN and other methods[12]. For the multi-label classification of rhinitis, the included patients with AR accounted for 95.1% of the total patients, and a few patients with similar rhinitis symptoms had a label lower than 10% of the total sample number. This makes it difficult to achieve a balanced distribution of all categories of data. If oversampling SMOTE is applied to a small number of labels, the number of AR labels will be increased, the imbalance of the overall rhinitis symptoms data will be exacerbated, and the overall classification accuracy will be reduced[13].Analysis of the actual clinical data collected shows that if the training set and the test set are divided into minority labels, the minority samples in the test set will be reduced,which will lead to the increase of the influence of the single classification result on the comprehensive classification and affect the balance ofprediction markers and sample size. To this end, ADASYN algorithm is adopted in this study to deal with unbalanced rhinitis sample datato ensure a balanced strategy for AR and its labels,effectively improve the classification accuracy of most similar rhinitis diseases AR and its labels,and also increasing the classification accuracy of a few otherunbalanced rhinitis cases[14]. The unbalanced split of rhinitis sample data is shown in Figure 2.
4.Ensemble analysis of clinical data
To evaluate the prediction results of AR samples, select the confusion matrix comprehensive indicators: true positive (TP), false negative (FN), false positive (FP) and true negative(TN), and use precision, sensitivity, specificity, G-Mean=sqrt(Sensitivity×Specificity), F1-Score, area under ROC curve AUC and other parameters together as predictive evaluation indicators.
This study proposed a heterogeneous ensemble rhinitis classifier model (Adaptive Random Forest-Out Of Bag-Easy Ensemble, ARF-OOBEE), which can identify a variety of disease, such as sinusitis (RS) (binary variable), The severity or persistence (ordered variable) of AR (AR), etc. This model effectively avoids the interference between multi-label type classification and multi-class symptom classification by converting heterogeneous multi-output classification problems into multi-label classification problems and 2 multi-class classification problems, and two or more indexing or typing labels for the same patient at the same time. Multiple models are used to train the classifier for the same batch of data, and the final ensemble classifier is obtained by using the ensemble learning algorithm. At the same time, 6 common machine learning classification algorithms were selected for comparative experiments, including Naive Bayes (NB)[15], Support Vector Machine (SVM)[16,17], Logistic Regression (LR)[18], Multilayer Perceptron (MLP)[19], Deep Forest (GCForest)[20], eXtreme Gradient boosting (XGBoost)[21].