Method for predicting severe acute exacerbation of bronchiectasis using artificial intelligence

Abstract

The present invention relates to a method for the prognosis of acute exacerbation in bronchiectasis using an artificial intelligence model. The method processes clinical data such as BSI and / or FACED scores and demographic information as input, handles missing values, and standardizes the data. The standardized data undergo four-fold cross-validation and are analyzed using a multilayer perceptron (MLP) model to acquire outputs through which the acute exacerbation of bronchiectasis can be prognosed. Through the process of analyzing and training with the input data, weights are assigned to individual clinical indicators, thereby improving the accuracy of the output, that is, the prediction results.

Description

A method for predicting severe acute exacerbations of bronchiectasis using artificial intelligence.

[0001] The present invention relates to a method for predicting severe acute exacerbation of bronchiectasis using artificial intelligence, and more specifically, to a method for more sensitively and accurately predicting a severe acute exacerbation state requiring hospitalization by analyzing the clinical characteristics of a patient, including the bronchiectasis severity index (BSI), using artificial intelligence (AI).

[0002]

[0003] Meanwhile, the present invention was made possible through research supported by the University Technology Management Promotion Project of the Ministry of Science and ICT (Project Unique Number: 2710006818, Project Number: RS-2024-00451411) and research supported by the Group Research Support Project of the Ministry of Science and ICT (Project Unique Number: 2710046053, Project Number: NR049557).

[0004] Bronchiectasis is a chronic lung disease characterized by permanent dilation of the airways, leading to chronic respiratory symptoms and recurrent acute exacerbations. Bronchiectasis is a highly prevalent chronic respiratory disease worldwide, and its prevalence has recently been increasing. Furthermore, patients with bronchiectasis are known to have a lower quality of life and higher medical utilization compared to the general population, resulting in a significant burden of illness.

[0005] Patients with bronchiectasis typically complain of recurrent cough and purulent sputum. Hemorrhage from the inflamed airway mucosa can occur in 50-70% of cases, resulting in blood-tinged sputum. In severe cases, bleeding from thickened bronchial arteries can lead to massive hemoptysis. Systemic symptoms such as fatigue, weight loss, and muscle pain may also accompany these symptoms. Patients with bronchiectasis often experience acute exacerbations, triggered by certain infections. These exacerbations are characterized by increased sputum production, increased purulence, and frequent bloody sputum production. Systemic symptoms, including fever, may also be prominent. Rarely, clubbing may occur, and chronic hypoxia can lead to dyspnea and cyanosis. These frequent exacerbations can lead to a decreased quality of life, decreased lung function, increased medical care utilization, and even increased mortality.

[0006] Factors that predict the severity and prognosis of bronchiectasis include the bronchiectasis severity index (BSI) and the FACED (Forced Expiratory Volume in one second, FEV1 % predicted, Age, Chronic colonization, Extension, and Dyspnea) indices, and both indices are known to predict the prognosis of bronchiectasis well. However, the FACED index is suitable for predicting the long-term risk of death over 15 years, and the BSI is known to better predict the risk of acute exacerbation, hospitalization, and death of bronchiectasis than the FACED index (Journal of the Korean Society of Internal Medicine, Vol. 95, No. 3, pp. 141-150, 2020).

[0007] Acute exacerbations of bronchiectasis can be diagnosed by examining sputum production, respiratory rate, respiratory rate, and auscultation (e.g., crackles and abnormal breath sounds). BSI can also be used to predict the severity of acute exacerbations, but these are insufficient for assessing the severity and prognosis. In particular, severe exacerbations (requiring hospitalization or emergency room visits), the most dangerous of these, are clinically crucial but difficult to predict.

[0008] In addition, existing methods such as BSI and FACED predict the prognosis of patients by summing the numerical results obtained through simple conditional statements for each factor. However, since these methods are simple summations of each item, they do not consider complex interactions between individual items.

[0009] In addition, there is a prior literature regarding a system for predicting the occurrence of acute exacerbations in patients with chronic obstructive pulmonary disease (Chinese Patent Publication No. 107451390) and a prior literature regarding an application for predicting acute exacerbations of chronic respiratory diseases (Korean Patent Publication No. 10-2019-0115330) that uses medical history data, clinical feature data, and lifestyle patterns to predict acute exacerbations of chronic respiratory diseases. However, these prior literatures do not describe predictions using a specific artificial intelligence model, and it is difficult to know about bronchiectasis.

[0010] Accordingly, the inventors of this study aimed to predict prognosis considering the overall patient condition by learning the complex interactions between clinical factors through an AI-based prognostic model. By analyzing clinical characteristics, including the BSI score, using AI to more sensitively and accurately predict severe acute exacerbations, and applying this to clinical practice, we expect to be able to tailor treatment plans to each patient's individual characteristics, thereby reducing hospitalization rates and the burden of disease.

[0011] The purpose of the present invention is to provide a method for predicting acute exacerbation of bronchiectasis using an artificial intelligence model.

[0012] To achieve the above purpose,

[0013] The present invention comprises a data receiving step in which a processor of a computer device receives data including clinical indicators from a server connected to the computer device;

[0014] A step in which the processor extracts clinical characteristic data related to acute exacerbation of bronchiectasis from the received data; and

[0015] A method for predicting the prognosis of acute exacerbation of bronchiectasis using an artificial intelligence model is provided, which includes a step of predicting acute exacerbation of bronchiectasis through the result value obtained by inputting the extracted clinical feature data into an artificial intelligence prediction algorithm.

[0016] In addition, the present invention provides a device for predicting the prognosis of acute exacerbation of bronchiectasis for performing the above prediction method.

[0017] The present invention develops an artificial intelligence-based prognostic prediction model that learns the complex interactions between various clinical factors and enables prognostic prediction considering the overall condition of the patient. This method complements the shortcomings of existing methods by enabling feature extraction that considers the correlation between each factor, thereby achieving higher prediction accuracy. Experimental results show that the prediction accuracy based on BSI is 96.7% and that based on FACED is 73.0%, while the method of the present invention shows a significantly higher prediction accuracy of 98.3%.

[0018] Figure 1 is a missing value distribution graph organized by analyzing missing values ​​in the data.

[0019] Figure 2 is a schematic diagram for 4-fold cross-validation.

[0020] Figure 3a shows the results of a 4-fold cross-validation analysis using the MLP model using BSI as input.

[0021] Figure 3b shows the results of a 4-fold cross-validation analysis using the MLP model using all clinical data features, including BSI, as input values.

[0022] Figure 4 is a ROC curve graph when analyzed with XGBoost, LR, or MLP.

[0023] Figure 5 illustrates the structure of an artificial intelligence-based bronchiectasis prediction model.

[0024] Figure 6 is a graph that summarizes the SHAP (Shapley Additive exPlanations) values ​​that confirm the contribution of each clinical feature.

[0025] Figure 7 is a graph that organizes the average SHAP values ​​to confirm the relative contribution of each clinical feature.

[0026] Hereinafter, the present invention will be described in detail.

[0027] The present invention comprises a data receiving step in which a processor of a computer device receives data including clinical indicators from a server connected to the computer device;

[0028] A step in which the processor extracts clinical characteristic data related to acute exacerbation of bronchiectasis from the received data; and

[0029] A method for predicting the prognosis of acute exacerbation of bronchiectasis using an artificial intelligence model is provided, which includes a step of predicting acute exacerbation of bronchiectasis through the output value obtained by inputting the above-mentioned extracted clinical feature data into an artificial intelligence prediction algorithm.

[0030] The above acute exacerbation condition refers to a case where three or more of the major symptoms, such as cough, increased sputum volume / viscosity, purulent sputum, dyspnea, decreased exercise capacity, fatigue, discomfort, and hemoptysis, worsen and last for more than 48 hours. If the acute exacerbation condition occurs, pneumonia, etc. may occur, which may lead to death, so a change in treatment method or outpatient treatment and hospitalization may be required. In the present invention, it is desirable to predict a severe acute exacerbation requiring hospitalization.

[0031] The above clinical data may include demographic data, smoking history, underlying disease, spirometry data, related treatment history, laboratory data, bronchiectasis severity index (BSI), and FACED (Forced Expiratory Volume in one second, FEV1 % predicted, Age, Chronic colonization, Extension, and Dyspnea), and preferably includes BSI and / or FACED.

[0032] The above prediction method includes a step of preprocessing data, which involves identifying missing values. Missing values ​​refer to values ​​not indicated in the data. The method may additionally include a step of replacing missing values ​​of continuous features with the mean, and replacing missing values ​​of categorical features with the mode.

[0033] Data preprocessed in this manner may additionally include a data standardization step using the following calculation formula.

[0034]

[0035] Xnew represents the standardized data value, x represents the original data, μ represents the mean, and σ represents the standard deviation.

[0036] The above artificial intelligence model may be an extreme gradient boosting (XGBoosting), logistic regression (LR), or multilayer perceptron (MLP) model as a prediction algorithm, and is preferably a multilayer perceptron model.

[0037] Using artificial intelligence, input data is processed based on a multilayer perceptron (MLP), the input data is a vectorized value of clinical factors, and the output data can provide the prognosis of acute exacerbation of bronchiectasis as a value between 0 (mild) and 1 (severe).

[0038] The structure of the artificial intelligence model used in this method consists of a total of three MLP blocks and one dense layer. Each MLP block consists of a dense layer for feature extraction, batch normalization, a rectified linear unit (ReLU) activation function, and a dropout layer to prevent overfitting. The feature information obtained through the above process is output as a value between 0 and 1 through the sigmoid activation function. Finally, if the output value is greater than the threshold value set by the user, the prognosis of acute exacerbation can be determined as severe (1), otherwise as mild (0).

[0039] Weights can be given to features with high contributions among input data, and the more features with high contributions are included, the higher the accuracy of predicting the prognosis of acute exacerbation.

[0040] A method for predicting the prognosis of acute exacerbation of bronchiectasis using an artificial intelligence model according to one embodiment of the present invention may also be implemented in the form of a computer-readable recording medium recording a program for execution on a computer. The computer-readable recording medium may be any available medium that can be accessed by a computer, and may include both volatile and nonvolatile media, removable and non-removable media. In addition, the computer-readable recording medium may include a computer storage medium. The computer storage medium may include both volatile and nonvolatile, removable and non-removable media implemented by any method or technology for storing information such as computer-readable instructions, data structures, program modules, or other data.

[0041] The exemplary modules, steps, or combinations thereof described in this specification may be implemented in electronic hardware (e.g., digital designs designed through coding), software (e.g., various forms of applications including program instructions), or a combination thereof. Whether hardware and / or software is implemented may vary depending on design constraints imposed on the user terminal.

[0042]

[0043] In addition, the present invention provides a device for predicting the prognosis of acute exacerbation of bronchiectasis, including a program for performing the above prediction method.

[0044] The above device may be a device having a program installed therein for executing the above prediction method and may include a device capable of outputting a result value.

[0045] The device may include an artificial neural network and process weights of the artificial neural network. For example, the computing device may quantize the weights of the artificial neural network and encode at least a portion of the quantized weights. The device may be, but is not limited to, a smartphone, tablet PC, PC, smart TV, mobile phone, media player, server, microserver, or other mobile or non-mobile computing device equipped with an AI program for processing the weights of the artificial neural network and including a voice recognition function.

[0046]

[0047] In a specific embodiment of the present invention, data including the clinical characteristics of 492 patients with bronchiectasis were obtained, and a four-fold cross-validation was performed using the clinical characteristics as input values ​​to predict the prognosis of acute exacerbation of bronchiectasis (Fig. 2). At this time, the models for analysis were XGBoosting, LR, or MLP, and the results from each model were compared. As a result, it was confirmed that the prediction of acute exacerbation prognosis was the most accurate when MLP was used (Table 7). In addition, the prediction results were compared by using different types of clinical characteristics as input values, and it was confirmed that the accuracy of the prediction results increased when BSI and FACED were included in the input values. To confirm the clinical characteristics with such high contributions, the Shapley Additive exPlanations (SHAP) technique was used to quantify the influence of factors and visualize them. As a result, it was confirmed that the contributions were high in the order of BMI, increased sputum concentration, history of severe exacerbation, tuberculosis, and pneumonia (Figs. 6 and 7). The present invention shows that by analyzing the clinical characteristics of patients with bronchiectasis using an artificial intelligence model, the prognosis of acute exacerbation of patients with bronchiectasis can be determined with high accuracy.

[0048]

[0049] Hereinafter, the present invention will be described in more detail through specific examples.

[0050] However, the following examples are only illustrative of the present invention, and the content of the present invention is not limited by the following examples.

[0051]

[0052] <Example 1> Clinical diagnosis and analysis of patients (baseline characteristics analysis)

[0053] We conducted a study of 492 patients enrolled in the Korean Multicenter Bronchiectasis Audit and Research Collaboration (KMBARC) with 1-year follow-up data. KMBARC is a prospective, non-interventional, observational cohort study of noncystic bronchiectasis in Korea. Patients aged 18 years or older with bronchiectasis in one or more lobes as detected by chest CT were included in the study. Patients with cystic bronchiectasis, traction bronchiectasis due to interstitial lung disease (ILD), or pregnancy were excluded.

[0054] We analyzed the clinical characteristics of 492 patients with bronchiectasis to establish a baseline. Baseline data included age, sex, body mass index (body mass index (BMI) (weight in kilograms divided by height in meters squared), smoking history, comorbidities, medication history, and laboratory results. Symptoms were assessed using the modified Medical Research Council (mMRC) questionnaire and the Bronchial Ejaculation Health Questionnaire (BHQ) for quality of life. Respiratory function was assessed at patient enrollment, and spirometry before and after bronchodilator administration was performed according to the American Thoracic Society / European Respiratory Society criteria. Absolute values ​​of FEV1 and forced vital capacity (FVC) were recorded, and the percentage of predicted values ​​for FEV1 and FVC were calculated using reference equations derived from a representative Korean sample.

[0055] To evaluate the sputum volume (ml / day) and color of patients with bronchiectasis, the sputum volume was estimated to be 50 ml if it could be contained in a small glass, and 200 ml if it could be contained in a cup. The sputum color was evaluated using a sputum color chart. Sputum color was classified as 1 (mucopurulent), 2 (mucopurulent sputum), 3 (purulent), or 4 (severely purulent). Microbiological analysis of sputum specimens was performed according to standard methods, and semi-qualitative bacterial culture was performed. If the Murray and Washington criteria were met, Gram stain was performed prior to sputum culture. Nontuberculous mycobacterial (NTM) lung disease was diagnosed using microbiological analysis using the microbiological criteria provided by the American Thoracic Society and the Infectious Diseases Society of America. The radiological severity of bronchiectasis was analyzed using the Modified Reiff index, and the clinical status and severity of bronchiectasis were evaluated using the FACED and BSI indices.

[0056] The median age of the patients was 65 years (IQR, 56–74 years), 233 (47.4%) were male, and 312 (63.4%) were never-smokers. The most common comorbidity was chronic obstructive pulmonary disease (COPD) (43.9%), followed by tuberculosis (35.6%) and asthma (25.4%).

[0057] Parameters for lung function were predicted forced vital capacity (FVC) 72% (56-88), predicted FEV1% 63% (44-82), and FEV1 / FVC ratio 65% (51-79). The amount of sputum was 22 ml (0-59), and mucoid was the most common sputum type at 63.8%, followed by mucopurulent sputum at 29.6%. Pseudomonas aeruginosa, which can cause infection, was found in 22% of patients with bronchiectasis. The median modified Medical Research Council (mMRC), BHQ index, and modified Reiff index were 1 (0-2), 64.2 (53.6-74.6), and 6 (2-10), respectively. The median BSI and FACED scores were 2 (0-4) and 6 (3-9), respectively. Mucolytics were the most commonly used treatment (70%), followed by LAMA+LABA (50.5%) and ICS+LABA (31.8%). Ten patients (0-3) and 0 (0-1) experienced acute exacerbations and severe exacerbations, respectively. Detailed information on other clinical parameters is presented in Table 1.

[0058] N = 492Age, years64.5 (55.5-73.5)Sex, males233 (47.4)BMI, kg / m 223.1 (19.5-26.7)Smoking historyNever smoker312 (63.4)Ex-smoker160 (32.5)Current smoker19 (3.9)ComorbiditiesChronic obstructive pulmonary disease216 (43.9)Asthma125 (25.4)History of pulmonary tuberculosis174 (35.6)Diabetes mellitus55 (11.2)Stroke9 (2.3)Pertussis44 (9.0)NTM-PD48 (9.8)Rheumatoid arthritis32 (6.5)Sinusitis42 (8.5)Gastroesophageal reflux disease71 (14.5)SpirometryFVC, L2.6 (1.8-3.4)FVC, %72 (56-88)FEV1, L1.7 (1.1-2.3)FEV1, %63 (44-82)FEV1 / FVC, %65 (51-79)Sputum concentrationMucoid293 (63.8)Mucopurulent136 (29.6)Purulent26 (5.7)Severe purulent4 (0.9)Sputum volume22 (0-59)mMRC1 (0-2)BHQ score64.2 (53.6-74.6)Modified Reiff score6 (2-10)MicrobiologyP. aeroginosa108 (22)BSI score2 (0-4)FACED score6 (3-9)Laboratory findingsWBC count, / L7300 (4800-9800)Neutrophil count, / μL4378 (2274-6482)Eosinophil count, / μL196 (26-366)Hemoglobin13 (12-14)Platelet count243 (164-321)hs-CRP, mg / dL1.5 (0-5.4)Respiratory treatment398 (81.9)Long-term oxygen therapy17 (3.5%)Theophylline41 (10.1)ICS16 (5.1)ICS + LABA105 (31.8)LABA + LAMA191 (50.5)LAMA68 (20.9)Oral corticosteroid10 (3.3)Mucolytics285 (70.0)Macrolide36 (7.7)Number of any exacerbations in previous year1 (0-3)Number of exacerbations in previous year0 (0-1).

[0059] Statistical analysis was performed to determine the correlation between each data feature and acute exacerbation. Continuous features were tested using the t-test, while categorical features were tested using the chi-square test or Fisher's exact test. The results are summarized in Tables 2 and 3.

[0060] FeatureExcluding missing valuesMissing valueTotalSevere AE_0Severe AE_1P-valueAge492064.5 ± 964.7 ± 9.163.5 ± 80.36BMI490223.1 ± 3.623.2 ±3.522 ± 4.10.018Sputum volume4613122 ± 36.917.4 ± 25.556.2 ± 73.8< 0.001WBC2472457.2 ± 2.47.1 ± 2.48.1 ± 2.20.041HB24724513.2 ± 1.413.3 ± 1.312.6 ± 1.30.011PLT246246243.3 ± 78.9237.7 ± 74.9290.2 ± 95.9< 0.001CRP1803121.5 ± 3.91.5 ± 4.11.2 ± 1.90.606esr10039227.7 ± 25.324.5 ± 20.663.4 ± 440.042FEV1FVC492064.9 ± 13.765.6 ± 13.660.1 ± 13.50.005FEV1_base49201.7 ± 0.61.7 ± 0.61.3 ± 0.6< 0.001FEV1_per_base492062.5 ± 19.464.2 ± 18.748.8 ± 19.4< 0.001FVC_base49202.6 ± 0.82.6 ± 0.82.1 ± 0.8< 0.001FVC_per_base492071.9 ± 16.173.2 ± 15.361.4 ± 18< 0.001dlco_base18231014.1 ± 514.2 ± 513.2 ± 50.404Dlco_perc_base18231073.2 ± 21.473.8 ± 21.467.6 ± 20.60.235mMRC479131.2 ± 0.81.1 ± 0.81.8 ± 1.1< 0.001Reiff48846.4 ± 4.16.1 ± 48.8 ± 4.6< 0.001BSI48846.1 ± 3.45.2 ± 2.312.8 ± 3.1< 0.001FACED49202 ± 1.61.9 ± 1.63.1 ± 1.4< 0.001입원횟수49200.2 ± 0.60.2 ± 0.40.9 ± 1.1< 0.001WBC_base3121807.3 ± 2.57.2 ± 2.48.4 ± 2.60.005Neu_per_base32316958.8 ± 11.658.4 ± 10.862.6 ± 160.139Lym_per_base32316926.9 ± 10.230.5 ± 9.925.1 ± 11.30.003Eos_per_base3231692.8 ± 2.32.8 ± 2.32.7 ± 2.30.751Ney_base3081844378 ± 2104.34253.4 ± 1984.45416.5 ± 2742.30.024Lym_base3081842084.9 ± 794.22096.3 ± 793.71989.6 ± 804.20.467Eos_base308184196 ± 170.1194.9 ± 168.9204.7 ± 181.70.755.

[0061] FeatureExcluding missing valuesMissing valueTotalSevere AE_0Severe AE_1P-valueSex4920259(52.6%)225(51.6%)34(60.7%)0.253Smoking491119(3.9%)17(3.9%)2(3.6%)0.9760(32.6%)141(32.4%)19(33.9%)312(63.5%)277(63.7%)35(62.5%)Sputum_concentration45933293(63.8%)273(67.2%)20(37.7%)<0.001136(29.6%)114(28.1%)22(41.5%)26(5.7%)17(4.2%)9(17%)4(0.9%)2(0.5%)2(3.8%)COPD4920216(43.9%)184(42.4%)32(57.1%)0.048asthma4920125(25.4%)104(23.9%)21(37.5%)0.041tb4893174(35.6%)148(34.2%)26(46.4%)0.098stroke3921009(2.3%)9(2.6%)0(0%)0.607CVD4920154(31.3%)138(31.7%)16(28.6%)0.753Malig491143(8.8%)39(9%)4(7.1%)0.805DM491155(11.2%)48(11%)7(12.7%)0.878pneumonia4893209(42.7%)168(38.8%)41(73.2%)<0.001pertussis489344(9%)41(9.5%)3(5.4%)0.445NTM489348(9.8%)41(9.5%)7(12.5%)0.632RA489332(6.5%)29(6.7%)3(5.4%)1sinusitis492042(8.5%)39(8.9%)3(5.4%)0.456gerd489371(14.5%)61(14.1%)10(17.9%)0.581pseudomanas4920108(22%)81(18.6%)27(48.2%)<0.001Itot486617(3.5%)6(1.4%)11(20%)<0.001Theophylline4058741(10.1%)30(8.5%)11(22%)0.006ICS31617616(5.1%)10(3.5%)6(17.6%)0.004ICS.LABA330162105(31.8%)90(30.4%)15(44.1%)0.152LAMA32616668(20.9%)58(19.7%)10(31.2) %)0.196LAMA.LABA378114191(50.5%)161(48.3%)30(66.7%)0.032LABA3071857(2.3%)7(2.5%)0(0%)1OCS30019210(3.3%)7(2 .6%)3(10.3%)0.061mucolytic40785285(70%)250(69.6%)35(72.9%)0.766physiotherapy4801216(3.3%)10(2.4%)6(10.9%)0 .006Macrolide4682436(7.7%)28(6.7%)8(15.4%)0.047Cystic4920258(52.4%)221(50.7%)37(66.1%)0.043Baseline(severe presence or absence of AE)4920105(21.3%)75(17.2%)30(53.6%)<0.001.

[0062] Before analyzing the collected data, the presence of missing values ​​was checked and removed as a preprocessing step. The distribution of missing values ​​is shown in Figure 1.

[0063]

[0064] <Example 2> Quantitative evaluation

[0065] To establish a model to predict acute exacerbation or severe acute exacerbation of bronchiectasis, various artificial intelligence models were used to verify their predictive ability.

[0066] Specifically, extreme gradient boosting (XGBoost), logistic regression (LR), or multilayer perceptron (MLP) were used to design acute exacerbation prediction models. Each model has its own characteristics and provides different advantages in capturing patterns and making predictions. The hyperparameters for each model used for exacerbation prediction are summarized in Table 4. In addition, each model was implemented using the Scikit-learn 1.1.2 and TensorFlow 2.4.1 libraries in Python 3.8 and the CUDA 11.0.3 toolkit on a desktop computer with an Intel Core i7-12700, @ 2.10 GHz central processing unit (CPU), and an NVIDIA GeForce RTX3090 Ti graphics processing unit (GPU).

[0067] ModelHyperparametersvalues rate1.0e-3batch size8loss functionweighted binary cross entropy

[0068] To evaluate the performance of each model, 20% of the dataset was used as a test dataset, and the effectiveness of the MLP was evaluated using 4-fold cross-validation to effectively control model overfitting. When performing four validation iterations, the data, excluding the test data (test, 20%), was divided into training (60%) and validation (validation, 20%) for each iteration (Fig. 2).

[0069] Additionally, four quantitative model performance metrics were adopted: sensitivity, specificity, F1 score, and area under the curve (AUROC). The ROC was calculated by plotting (1-specificity) on the x-axis and sensitivity on the y-axis, and each evaluation metric was calculated using the formulas below.

[0070] Sensitivity = TP / (TP+FN)

[0071] Specificity = TN / (TN+FP)

[0072] F1-score = 2X(precision X sensitivity) / (precision + sensitivity)

[0073] Precision = TP / (TP+FP)

[0074] The above TP, FP, TN, and FN represent the number of true positives, false positives, true negatives, and false negatives, respectively, and all indicators have values ​​between 0 and 1, with 1 being the optimal value.

[0075]

[0076] The results of the 4-fold cross-validation of the MLP model showed that the MLP model using BSI as input showed an average sensitivity of 0.90, a specificity of 0.93, an F1 score of 0.74, and an AUROC of 0.97 (Table 5 and Fig. 3a). In addition, the MLP model using all features as input showed an average sensitivity of 0.96, a specificity of 0.95, an F1 score of 0.80, and an AUROC of 0.98 (Table 6 and Fig. 3b).

[0077] SensitivitySpecificityF1-scoreAUROC1st Fold0.90910.93180.74070.96752nd Fold0.90910.93180.74070.96753rd Fold0.90910.93180.74070.96754th Fold0.90910.93180.74070.9675Average0.90910.93180.74070.9675

[0078] Abbreviations: MLP, multilayer perceptron; BSI, Bronchiectasis Severity Index; AUROC, area under the receiver operating characteristic.

[0079] SensitivitySpecificityF1-scoreAUROC1st Fold1.00000.95450.84620.98762nd Fold0.90910.92050.71430.98043rd Fold0.90910.95450.80000.97424th Fold1.00000.95450.84620.9897Average0.9545 ± 0.050.9460 ± 0.010.8016 ± 0.050.9830 ± 0.01

[0080] Abbreviations: MLP multilayer perceptron; AUROC, area under the receiver operating characteristic.

[0081]

[0082] <Example 3> Comparison of prediction performance

[0083] To identify a model that more accurately predicts acute exacerbations of bronchiectasis, we compared the performance of clinical indicators and models. FACED and BSI were used as clinical indicators, and the exacerbation prediction rates using FACED and BSI were significantly higher in sensitivity (0.73 vs. 0.91), specificity (0.63 vs. 0.93), F1-score (0.31 vs. 0.74), and area under the curve (AUROC) (0.73 vs. 0.97), indicating that the BSI score was more effective than the FACED score in predicting exacerbations (Table 7).

[0084] In addition, we wanted to analyze using three analysis models (XGBoost, LR, MLP) and compare the prediction accuracy according to the analysis model. XGBoost is an abbreviation for "extreme gradient boosting" and is a popular learning algorithm belonging to the gradient boosting family of machine learning techniques. It is a method to compensate for the errors of the previous learners by adding new weak learners one by one to minimize the loss of the existing tree formed by combining weak learners, and through this process, it is a method to train the model quickly and efficiently by optimizing.

[0085] Logistic regression (LR) is a supervised machine learning algorithm developed for classification problems. It is based on the concept of fitting a sigmoid function to data to estimate the probability of an instance belonging to a specific class. The sigmoid function maps values ​​in the range of 0 to 1, making it suitable for probability estimation. In logistic regression analysis, the coefficient associated with each predictor variable represents the log probability of the target class for each unit change in that predictor variable, assuming all other variables are held constant. Therefore, each coefficient can provide insight into the direction and magnitude of each predictor's influence on the outcome.

[0086] Multilayer perceptrons (MLPs) are widely used for solving various machine learning tasks due to their outstanding performance. They consist of multiple layers of interconnected neurons, configured in a feedforward fashion, and employ weighted sums and activation functions to capture the complex, nonlinear relationships between input features and output predictions. Through backpropagation and gradient descent, MLPs are trained to minimize a loss function and adjust weights and biases. This makes them a powerful tool for learning complex patterns and making accurate predictions in diverse applications, such as image recognition, natural language processing, and time series analysis.

[0087] XGBoost and LR differ in that the user must directly input the features of the input factors, whereas MLP automatically derives and analyzes the features when the input factors are input as is. In addition, XGBoost and LR perform linear analysis, whereas MLP performs nonlinear analysis.

[0088] As a result of analysis using three models, it was confirmed that the sensitivity, specificity, F1-score, and AUROC value were better when using the MLP model than when using XGBoost and LR (Table 7), and as shown in Figure 4, when confirmed through the ROC curve graph, it can be confirmed that the AUC value when using the MLP model is the largest, which can be seen through this that the accuracy of the MLP model is the highest.

[0089] ClassifierSensitivitySpecificityF1-scoreAUROCClinical indicatorFACED0.72730.62500.30770.7304BSI0.90910.93180.74070.9675ModelXGBoost 0.90900.92330.72600.9613LR0.90910.92330.72220.9700MLP0.95450.94600.80160.9830

[0090] <Example 4> Model learning results according to input values

[0091] The input values ​​are vectorized values ​​of various clinical factors such as demographic data such as the patient's gender, age, and BMI, smoking history, underlying disease, spirometry data, related treatment history, and laboratory data, and the output data is a predicted value of the prognosis of bronchiectasis between 0 (mild) and 1 (severe). The structure of the artificial intelligence model used in this method consists of a total of three MLP blocks, one dense layer, batch normalization, ReLU activation function, and a dropout layer to prevent overfitting (Fig. 5).

[0092] <4-1> Learning Outcome 1

[0093] ICS, ICS / LABA, Reiff, BSI, and FACED were entered as input values, and data preprocessing was performed to replace missing values ​​with the mean or mode, and analysis was performed using an MLP model. The following calculation formula was used to standardize the data.

[0094]

[0095] The results of the predictive analysis are shown in Table 8.

[0096] SensitivitySpecificityF1-scoreAUROC1st Fold0.90910.98860.90910.96702nd Fold0.90910.87500.62500.92423rd Fold1.00000.97730.91670.99594th Fold0.81820.79550.47370.8864Average0.90910.90910.73110.9478

[0097] <4-2> Learning Outcome 2

[0098] ICS, Reiff, BSI, and FACED were entered as input values, and data preprocessing was performed to replace missing values ​​with the mean or mode, and analysis was performed using an MLP model. Table 9 shows the results of the predictive analysis.

[0099] SensitivitySpecificityF1-scoreAUROC1st Fold0.90910.97730.86960.96382nd Fold0.90910.90910.68970.99173rd Fold0.90911.00000.95240.97524th Fold0.90910.89770.66670.9866Average0.90910.94600.79460.9793

[0100] <4-3> Learning Outcome 3

[0101] ICS, cystic, BSI, and FACED were entered as input values, and data preprocessing was performed to replace missing values ​​with the mean or mode, and analysis was performed using an MLP model. Table 10 shows the results of the predictive analysis.

[0102] SensitivitySpecificityF1-scoreAUROC1st Fold0.90910.92050.71430.93292nd Fold0.90911.00000.95240.91743rd Fold0.90911.00000.95240.92254th Fold0.90910.84090.57140.9845Average0.90910.94030.79760.9393

[0103] <4-4> Learning Outcome 4

[0104] In Example 4-3, ICS, cystic, and BSI were input as input values, excluding FACED. Data preprocessing was performed to replace missing values ​​with the mean or mode, and analysis was performed using an MLP model. Table 11 shows the results of the prediction analysis.

[0105] SensitivitySpecificityF1-scoreAUROC1st Fold0.90910.84090.57140.98552nd Fold0.90910.88640.64520.98973rd Fold0.90910.97730.86960.95144th Fold0.90910.98860.90910.9762Average0.90910.92330.74880.9757

[0106] <4-5> Learning Outcome 5

[0107] In Example 4-3, ICS, cystic, and FACED were input as input values, excluding BSI. Data preprocessing was performed to replace missing values ​​with the mean or mode, and analysis was performed using an MLP model. Table 12 shows the results of the predictive analysis.

[0108] SensitivitySpecificityF1-scoreAUROC1st Fold0.81200.80680.48650.86572nd Fold0.72730.67050.33330.86673rd Fold0.72730.73860.38100.86164th Fold0.81820.90910.64290.8874Average0.77270.78130.46100.8704

[0109] <4-6> Learning Outcome 6

[0110] In Example 4-3, ICS was entered as an input value, excluding BSI and FACED, and data preprocessing was performed to replace missing values ​​with the mean or mode, and analysis was performed using an MLP model. Table 13 shows the results of the predictive analysis.

[0111] SensitivitySpecificityF1-scoreAUROC1st Fold0.72730.72730.37210.86362nd Fold0.81820.75000.42860.81923rd Fold0.81820.92050.66670.85744th Fold1.00000.89770.62070.8481Average0.79550.82390.52200.8471

[0112] Excluding the clinical indicators FACED and BSI, the sensitivity, specificity, F1-score, and AUROC were all significantly lower when predicting malignant deterioration, suggesting that the prediction results are less reliable.

[0113]

[0114] <Example 5> Analysis of factors affecting prediction

[0115] To determine how much each feature contributed to the prediction of acute exacerbations, we visualized the Shapley Additive exPlanations (SHAP) values. A summary plot showing SHAP values, considering all numerical data values, revealed that the most influential features in the MLP were BSI, increased sputum concentration, history of severe exacerbations, tuberculosis, and pneumonia (Figure 6). The relative importance ranking of the features is shown in Figure 7.

[0116] In conclusion, it can be seen that acute exacerbation prognosis can be predicted with high accuracy by inputting clinical characteristics including BSI and analyzing them using a multilayer perceptron (MLP) model.

Claims

1. A method for predicting the prognosis of acute exacerbation of bronchiectasis using an artificial intelligence model, A data receiving step in which a processor of a computer device receives data including clinical indicators from a server connected to the device; A step in which the processor extracts clinical features related to acute exacerbation of bronchiectasis from the received data; and A step of predicting acute exacerbation of bronchiectasis by inputting the extracted clinical features into the prediction algorithm of an artificial intelligence model and using the resulting values; A method for predicting the prognosis of acute exacerbation of bronchiectasis using an artificial intelligence model, including:

2. In paragraph 1, A method for predicting the prognosis of acute exacerbation of bronchiectasis using an artificial intelligence model, wherein the acute exacerbation is defined as the worsening of three or more symptoms selected from the group consisting of cough, increased sputum volume / viscosity, purulent sputum, dyspnea, decreased exercise capacity, fatigue, discomfort, and hemoptysis, which are major symptoms of bronchiectasis, and lasts for more than 48 hours.

3. In paragraph 2, A method for predicting the prognosis of acute exacerbation of bronchiectasis using an artificial intelligence model, wherein the above acute exacerbation is a severe acute exacerbation requiring hospitalization.

4. In paragraph 1, A method for predicting the prognosis of acute exacerbation of bronchiectasis using an artificial intelligence model, wherein the above clinical indicator is a BSI or FACED score.

5. In paragraph 4, A method for predicting the prognosis of acute exacerbation of bronchiectasis using an artificial intelligence model, wherein the above clinical indicators are BSI and FACED scores.

6. In paragraph 1, A method for predicting the prognosis of acute exacerbation of bronchiectasis using an artificial intelligence model, wherein the above clinical indicators are demographic information, smoking history, underlying disease, spirometry information, related treatment history, laboratory test results, BSI, and FACED.

7. In paragraph 1, The above artificial intelligence model is a method for predicting the prognosis of acute exacerbation of bronchiectasis using an artificial intelligence model, wherein the artificial intelligence model uses a multilayer perceptron (MLP) model as a prediction algorithm.

8. In paragraph 1, A method for predicting the prognosis of acute exacerbation of bronchiectasis using an artificial intelligence model, wherein the above prediction method additionally includes a step of training and learning using extracted clinical features.

9. In paragraph 8, A method for predicting the prognosis of acute exacerbation of bronchiectasis using an artificial intelligence model, wherein the above training and learning steps are performed through 4-fold cross-validation.

10. In paragraph 1, A method for predicting the prognosis of acute exacerbation of bronchiectasis using an artificial intelligence model, wherein the above prediction method additionally includes a data standardization step using the following mathematical formula.

11. In paragraph 1, A method for predicting the prognosis of acute exacerbation of bronchiectasis using an artificial intelligence model, wherein the above result value is derived as a value between 0 and 1, where 0 represents normal and 1 represents acute exacerbation.

12. A device for predicting the prognosis of acute exacerbation of bronchiectasis by performing the method of paragraph 1.

Images