Asthma is a heterogeneous disease and usually coincident with chronic airway inflammation. Asthma can be diagnosed based upon the patient’s history of respiratory symptoms such as wheezing, chest tightness, shortness of breath, and cough which may vary among the patient population in terms of intensity, time, and decreased expiratory airflow.

Allergic asthma is the easiest phenotype of asthma to diagnose with symptoms usually becoming apparent in childhood [1]. Globally, the prevalence of asthma in adults is estimated to be 4.3% [1]; in 2012, an estimated 300 million adults suffered from asthma, and by 2025, this figure will increase to 400 million. Moreover, the annual mortality rate of asthma is estimated to be about 250,000 [2]. In Iran, the prevalence of asthma is around 5.5% [3]. Every year, according to the World Health Organization (WHO), 15 million disability-adjusted life-years are lost because of the disease [4]. Those who develop asthma often have allergic conditions or a family history of allergic conditions such as eczema, food allergies, drug allergies, or allergic rhinitis. These patients normally respond well to inhaled corticosteroid treatments [2]. Like mast cells, different immunologic cells such as eosinophils, lymphocytes, and neutrophils have a role in the process of airway inflammation [5]. Genetic factors are among the risk factors of asthma, in addition to environmental factors such as exposure to allergens which may also exacerbate asthma symptoms and access to health care services [2].

There is a need for health care providers and governments to work collectively to improve control of asthma symptoms [2]. In traditional medicine, Crocus sativus L, which is more commonly known as saffron, has been used as a treatment for heart disease, depression, stress, and sleep disorders [5]. C sativus L possesses antioxidant [6] and anti-inflammatory properties [7]. Its active components—safranal and crocin—have demonstrated beneficial anti-inflammatory and antioxidant effects [5].

Several studies [6,8,9] have reported the effects of C sativus L on asthmatic patients. Zilaee et al [6] studied, in a randomized clinical trial, the effects of C sativus L supplements on clinical symptoms, Asthma Severity Score, blood pressure, and lipid profiles of patients with mild or moderate persistent allergic asthma. Although these studies showed aggregate effects, it was not possible to predict rare and serious effects for individuals.

Recommender systems, also known as recommendation engines, are used in online personalized predictive models and have been increasingly implemented in many areas of application to extract useful information from data; however, most of the available approaches that rely upon traditional statistical outcomes are unable to extract crucial knowledge [10,11] such as severity reduction estimates for patients with asthma [6]. Therefore, recommender systems can overlook significant effects of C sativus L supplements in patients with allergic asthma.

Recently, researchers have begun to estimate the effectiveness of clinical medicine using machine learning methods [12]. Machine learning techniques can be used to approximate the treatment effects of medicines [13,14]. An important application of machine learning in medicine is the development of automated risk-prediction algorithms to guide clinical care [15]. These algorithms can be used to integrate and interpret complex biomedical and health care data in scenarios where traditional statistical methods may not work [7].

To address the current limitations, a clinical prediction system based upon machine learning algorithms was developed to estimate the level of effect of C sativus L supplements in patients with allergic asthma. To classify clinical improvement in patients, we developed a model that determines the potential effect of C sativus L supplements on individual patients with asthma by extracting the factors with the greatest effect from clinical features, hematologic features, anti-inflammatory features, and Asthma Severity Score.

The best fitness value that was achieved for each output factor is shown in Table 2. Model performance is shown in Figures 4-7. The results were achieved with 100 epochs for training process.

As shown in Figure 4-7, the model’s training performance reached the acceptable fitness value with the MSE lower than 5% for each factor. Furthermore, Figure 8 shows the training performance of the overall model which reached a fitness level of 3.3322 × 10^–9. This result illustrates an accuracy greater than 99% for training. Table 3 shows the most effective factors that were found by the genetic algorithm. The selected parameters were used as input in the genetic algorithm–modified neural network to predict the effect on each clinical, hematologic, or anti-inflammatory parameter. FEV₁/FVC ratio, waking up due to asthma symptoms, hs-CRP, eosinophil, basophil, weight, and smoking history were selected most frequently by the genetic algorithm process to estimate the level of alleviation after C sativus L supplement therapy. Table 4 shows the optimized architecture of the model for the test set (hidden layers, neurons, training and fitting functions), and estimated rate of risk after supplement therapy for each factor are presented. According to Table 4, the predictions for waking up due to asthma symptoms and basophil with the MSE of 0.004 and 0.045 showed the highest and lowest accuracy, respectively, among the factors in the testing phase. A pattern recognition network was selected by the genetic algorithm for most experiments (patternnet in MATLAB). Furthermore, conjugate gradient backpropagation with Polak-Ribière updates (traincgp in MATLAB) was frequently selected for training in the genetic algorithm optimization process.

According to Table 4, the best accuracy was for waking due to asthma symptoms; the model classified the patients in terms of level of alleviation with an accuracy greater than 99% (MSE=0.0004) in testing. For this prediction model, 6 factors (FEV₁/FVC ratio, frequency of waking due to asthma symptoms, hs-CRP, eosinophil, basophil, weight, smoking condition) out of 18 input parameters were identified and used to classify the patients in five groups from low to high level of effect.

In this study, a clinical system to predict the effectiveness of C sativus L supplements in patients with allergic asthma was designed. The findings indicated that the model sufficiently estimated the level of alleviation in patients with asthma and the level of variation of asthma clinical, anti-inflammatory, and hematological factors.

Research [25] conducted on healthy individuals indicated that doses up to 400 mg per day of C sativus L were safe. In addition, a study [26] noted that a daily dosage of 100 mg of C sativus L in patients with metabolic syndrome reduced anti-HSP70 levels. Anti-HSP has been demonstrated to be a risk factor for asthma and has been correlated with asthma severity [8,26]. Another study [6] showed the beneficial effects of 100 mg C sativus L per day on clinical and immunologic factors as well as on symptom severity in asthma.

In a study [17], it was confirmed that it is possible to create a predictive model with machine learning algorithms that can outperform experts. Predictions can be in the form of patient classification, disease diagnosis, drug composition, and reactions to drugs. Prediction of treatment effects using machine learning have been used to develop clinical drug prediction systems [12,27]. To the best of our knowledge, our clinical prediction system is the first system that uses input features selection and classifies effectiveness of C sativus L supplement therapy.

The input features have an important role in the efficient implementation of prediction problems [28]. It is possible to use data mining methods to decrease the data dimensions, choose optimal features, and achieve better system precision [29]. In a study [15], partial least square regression was used to identify 9 out of 48 prognostic factors that were correlated to persistent asthma. Moreover, multilayer and probabilistic neural networks topologies were studied to find the best prediction accuracy [15]. In another study [29], data analysis was performed to select 13 effective factors to use in an ANN asthma diagnostic model [29]. We used a genetic algorithm feature selection model to find pertinent features. Genetic algorithms were recently developed to compare different feature selection methods and may be useful for feature selection when the problem has an exponential search space. The many advantages of genetic algorithms for feature selection are highlighted in the literature [30,31]. In contrast to that of other asthma diagnostic studies, the feature selection in our system is aimed to optimize the classification of asthma patients with supplement therapy into different groups in terms of the level of effect of C sativus L supplement.

ANN-based models have recently been used as a robust technique to classify patients with asthma, chronic obstructive pulmonary disease, or normal lung function based on measurement of lung condition and symptoms [7,32,33].

Analyses of the mean accuracies of the genetic algorithm–modified neural network predictor with feature selection included selecting effective risk factors and architecture components using an allergic asthma dataset for which the best MSE of 0.0004 was obtained. It was also clear that feature selection improved the accuracy of prediction for all asthma risk factors as well as the accuracy of the classification of patients in terms of level of alleviation.

Our study was able to show the importance prioritizing factors to predict the variation level of variation of factors in the patients with allergic asthma. Frequency of salbutamol inhaler use, frequency of waking due to asthma symptoms, and weight were the input features that were selected most often by the genetic algorithm to predict the level of variation in allergic asthma risk factors for the patients using C sativus L supplement therapy. Therefore, the initial value of those factors have an important role in predicting the level of alleviation of disease severity.

Several studies have listed the significant and nonsignificant effects of C sativus L supplements on the demographic, anthropometric, and clinical characteristics of asthma [6,8,9]. Hosseini et al [9] showed that C sativus L significantly affected anti-HSP and hs-CRP factor serum levels . Moreover, they showed that it increased the pulmonary volumes FEV₁, FVC, FEV₁/FVC, and FEF_25%-75% [9]. They also showed that there was no significant effect on the eosinophil count [18]. The genetic algorithm–modified neural network system in this study accurately predicted the level of C sativus L effects for anti-HSP (96.5%), hs-CRP (98.9%), FEV₁ (98.1%), FVC (97.5%), FEV₁/FVC (97%), FEF_25%-75% (96.7%), eosinophil (97.5%), and basophil (95.5%). Moreover, in general the patients were classified into 5 groups from low to high level of alleviation which reached accuracies of 99.9% in both training and testing experiments. FEV₁/FVC ratio, frequency of waking up due to asthma symptoms, hs-CRP, eosinophil, basophil, weight, and smoking history were selected as effective factors to estimate the level of alleviation of asthma in patients after the C sativus L supplement therapy.

The results also confirmed that it is possible to rely upon the prediction process described in this paper for the early prediction of level of variation of asthma factors. This study is the first to evaluate the classification accuracy of C sativus L supplement effect on the patients with asthma through feature selection. Our genetic algorithm–modified neural network can predict the effects of using C sativus L supplement on patients with asthma. This study indicated the importance of prioritizing each factor in predicting the effect of supplement therapy on allergic asthma.

By assessing risk, this method can be viewed as an important innovation to ensure that asthma is controlled and that serious complications are avoided. This study contributes to helping doctors to identify early on which factors will improve during treatment.

It would be interesting to evaluate the genetic algorithm–modified neural network on additional groups of patients with allergic asthma or other types of asthma. In future work, development of intelligent models including heuristic algorithms for feature selection with other machine learning techniques is recommended. Clinical prediction systems can also be applied to predict or to simulate the effectiveness of other medicines on other conditions, especially where different initial patient conditions may alter the course of treatment.