Validation of a modified oxygen nebulized inhalation method in airway surface anesthesia by comparative analysis via scalable broad learning Authors: Yanping Sun, Tao Chen, Kai Shi, Shenglin Mo, Yufei Deng, Lihua Liao, Huazhou Chen Published: 2025 ABSTRACT AbstractIn clinical practice for the diagnosis of pulmonary tuberculosis (PTB), bronchoscopy is typically performed under airway surface anaesthesia. The effectiveness of this anaesthesia is closely associated with the smoothness of bronchoscopy diagnosis, as well as the incidence and severity of related adverse events. To enhance the efficacy of airway surface anaesthesia, the modified oxygen nebulized inhalation (MONI) method is developed. Derived from the traditional oxygen nebulized inhalation (ONI) procedure, this modified method improves the refined selection of a nebulizer and precise control of the oxygen flowing speed.To validate the advantages of the MONI method, this study compared it with the traditional ONI method through data experiments. Patients undergoing bronchoscopic operation were divided into two groups: one group received MONI for anaesthesia, and the other received ONI. Six key clinical items were recorded during the procedure. A comparative analysis was then conducted on the grouped data using machine learning models. A parameter-scalable broad learning system (BLS) architecture is proposed for feature extraction from raw data, with the optimal analytical model determined by minimizing the loss function value. Both the number of virtual input nodes and the number of neurons in the hidden layer are set as tunable parameters to optimize the model. Scoring data for the target clinical items were input into the system, transformed for BLS network training, and then used to generate predictions. Comparative analysis of the BLS output predictions showed that the data recorded from the MONI group performed better than that from the ONI group.Furthermore, the optimal models were validated to be significant for prediction and could explain how the network output data correlates with each of the six clinical items. Thus, we conclude that the proposed MONI method can practically enhance the effect of airway surface anaesthesia, which will facilitate the diagnosis of pulmonary tuberculosis (PTB). The scalable BLS model is prospective to provide advanced artificial intelligence support to detection procedures, thereby contributing to the effective prevention of infectious diseases.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-025-28908-4.Keywords: Bronchoscopy, Airway surface anaesthesia, Modified oxygen nebulized inhalation, Data analysis, Broad learning system, Parametric scalable trainingSubject terms: Clinical trial design, Data mining, Machine learning Introduction Pulmonary tuberculosis (PTB) is a chronic infectious disease transmitted by the respiratory tract, which seriously affects the health of people in China1. It is regarded as one of the key diseases to combat. The definitive diagnosis of PTB is based on the finding of mycobacterium tuberculosis in respiratory specimens2. In clinical practice, sputum or alveolar lavage collected from the patients is used as the pathogenetic specimen, but the sputum samples are small so that it may not adequately reflect the whole physiology of a lung, thus alveolar lavage is frequently used in respiratory medicine. However, the detection on alveolar lavage usually has an over higher positive rate than actual, which leads to low accuracy in pathological judgement3. To address this issue, for the patients with PTB, the use of online bronchoscopy is advocated as the prior choice for a rapid test, in combination with the judgement by alveolar lavage sampling, such that the accuracy of pathogenetic diagnosis can be probably improved. Bronchoscopy allows in-field observation of the tracheal course and tracheal mucosa, for further assessment of the patient’s lungs. It serves as one of the key primary means for diagnosing tuberculosis and assessing its condition4. Nevertheless, bronchoscopy is a kind of mini-invasive test, during which patients may experience adverse effects such as haemoptysis, cardiac arrhythmia, airway spasm and hypoxia5. Then, bronchoscopy needs accompanied by surface anaesthesia of the airway in advance. Several studies have shown that the effectiveness of airway anaesthesia is directly related to the smoothness of the bronchoscopy procedure, the occurrence and magnitude of the associated adverse effects6–8. Therefore, it is very important to do a good job for airway surface anaesthesia before bronchoscopy detection and therapy. Lidocaine is currently the main medicament used for airway surface anaesthesia. The common method to deliver Lidocaine into the airway for anaesthesia involves nebulizing it into the clinical oxygen vapor, then it will be injected into the airway along with the oxygen-driven nebulized inhalation (ONI). With current clinical instruction9,10, the specific procedures of ONI are as follows: The patient is available at a sitting or semi-recumbent position. An oxygen-driven nebulizer is used, filled with 10 ml oxygen, ready to nebulize a 2% Lidocaine solution. The oxygen injection flow speed is controlled at 6–8 L/min, to blow the liquid Lidocaine as the nebulized status. Clinicians (doctors or nurses) guide the patient to inhale the nebulized gas in the way of breathing deeply in through the mouth and slowly out through the nose, for a duration of 15–20 min. This allows the Lidocaine to be well nebulized and pushed into the airways and lungs. However, in the long-term clinical practice, the ONI method is found having some inevitable drawbacks. For the first thing, there is no exact quantitative value, but only a fuzzy range to control the oxygen injection flow speed. Therefore, medical staff can only adjust the oxygen flow by personal empirical knowledge. For the second, patients undergoing ONI are prone to have some adverse reactions, such as nausea, vomiting, chest tightness, shortness of breath, etc., which will lead to a decrease in the patient’s ability to comply with nebulization. For the third, the Lidocaine-contained mist tends to adhere to the patient’s face, conjunctiva of the eyes, mouth and nose, leading to a reduction in the amount of medicament being inhaled and the potential for the medication to cause adverse to the conjunctival tissue. And, for the most important thing, it is hardly to fix the effective medicament dose for patients as every patient is specifically different. Actually, the dose of nebulized Lidocaine particles adhering to the lungs is the important factor to influence the anaesthesia effect. Thus, the uncertainty of medicament dose is the direct cause of ineffective or less effective anaesthesia in ONI. To improve the anaesthetic effect, a key issue to address is how to increase the effective deposition dose of nebulized lidocaine in the lungs11. Many research works revealed that the main influencing factors include: the size of nebulized particles, the way that patient co-operate in inhalation, characteristic spraying matters of the nebulizers12–14. In clinical application, the particle size is related to the flowing speed of the oxygen-driven nebulizer. If the dose of nebulized Lidocaine inhaled per unit of time is too large, the patient absorbs too much amount, so that the patient feels uncomfortable. Excessively strong medicament inhalation will irritate the patient’s throat, causing symptoms including coughing, chest tightness, and dizziness15. A patient-adopted nebulizer with suitable characteristics is mostly decisive by the attending doctor or the operating nurse, which is originated from the standardized diagnostic and treatment process, or from the basic and professional knowledge of the doctor/nurse16. Concerning on these factors, a modified method of ONI is proposed in this paper. We aimed to improve the oxygen nebulized inhalation method by quantifying the particle size of the nebulized medicament and adjusting the oxygen flow rate to achieve appropriate release of the medicament mist. The primary goals of these modifications were to enhance the efficacy of airway surface anaesthesia and reduce the incidence of adverse reactions in patients. In clinical practice, we train the doctors and nurses to learn about the technical improvement, and promote the basic knowledge to patients and their accompanying family members to raise their pervasive awareness, in order to lift the comfort of patients when receiving PTB examination and treatment. To validate the improved efficacy of airway surface anaesthesia achieved by the modified oxygen nebulized inhalation (ONI) method (hereafter denoted as MONI), this study monitored patients’ physiological responses during the progression of bronchoscopic diagnosis. As a result, a group of clinical data from the patients undergoing the MONI diagnosis are targeted. Meanwhile, another group of data from the patients undergoing traditional ONI test are also collected for comparative analysis. Analytical method is investigated for data mining to validate the improvement of MONI over the ONI method. Machine learning models has been applied17. Broad learning system (BLS, i.e. a broad learning network architecture) is popular in machine learning due to its high flexibility, simple network construction and fast computation18. It has been widely used in various fields such as image denoising19, mechanic control20, eco-agricultural environment21 and etc. During BLS training process, the model tends to be more easily improved by horizontally extending the input neurons than by stacking the hidden layers. Therefore, BLS has an advantage in reducing the model dimensionality when dealing with a batch of complex data. However, the training of BLS still relies on data-driven training. Increasingly research evidences show that the BLS model automatically generates extensively large number of horizontal extended neurons during its training process22. The excessive redundant neural nodes may lead to the decrease of the training speed and the reduction of the model’s generalization ability23. Therefore, parameter tuning is on vital demand for the control of BLS model optimization. In this study, a scalable BLS network architecture is proposed for data analysis to validate the effect of airway surface anaesthesia by using MONI method. For model enhancement, the structural parameters are designed to be tunable for scalable optimization. A key characteristic of scalable broad learning is to generate a suitable number of extended number of neurons, which are combined with the original input nodes to form the final virtual input layer. The parametric-scalable network model is applied to validate the improvement of MONI method, through the combined training of feature extraction, network activation and transformed prediction. Data experiment is launched for the validation of the proposed methodologies as the scientific support to the detection and prevention of infectious diseases, such as PTB. The modified method of oxygen nebulized inhalation Ethics statement The study protocol was approved by the Medical Ethics Committee of Liuzhou People’s Hospital, with the ethical approval number: KY2023-018-01. The requirement for informed consent was waived by the Medical Ethics Committee of Liuzhou People’s Hospital. This study was conducted in accordance with the Declaration of Helsinki as well as the relevant guidelines and regulations in China. The privacy of all subjects was protected. Only investigators had the permission to review personal medical records, and data anonymity was applied during data management. Operation procedure On the basis of the traditional ONI method, the factors affecting the effective dose of inhaled medicament are taken more into account, then the MONI method is proposed. The main contributory feature of MONI lies in the parameter-quantitative selection of a nebulizer and the oriented quantitative control of the oxygen flowing speed. (I) The selection of a nebulizer. Awareness of the parameters of a nebulizer is vital before to choose. The mouthpiece-type nebulizer is a popular model for portable medical application. There are two key parameters that directly affect the degree of the comfortable feeling of a patient include the total volume capacity of the nebulizer and the aerosol diameter of the inhalation hole. Accumulating experiences show that the mouthpiece-type nebulizer is regulated in specification of 10 ml in volume capacity with the inhalation hole in diameter of 1–5 μm to well release the nebulized medicament particles. For the seldom special case when the patient is unable to use a mouthpiece-type nebulizer, a mask with good seal can be chosen as an alternative to reduce the medicament spread to face and eyes. (II) The control of oxygen flowing speed. The nebulizer is filled with 10 ml 2% Lidocaine solution. Oxygen gas is blown into the nebulizer at a controlled speed, for fully mix with the Lidocaine, and then enter the patient’s airway. The size of mist Lidocaine particles formed by oxygen-driven nebulization is decisively related to the pressure and speed of the oxygen airflow. The faster the airflow, the smaller the medicament particles and the larger the amount of released nebulized medication. Several studies have shown that surface nebulized anaesthesia is most effective when the optimal nebulized medicament particles deposited in the lung are 1–5 μm in diameter24,25; The oxygen flowing speed of 6 L/min benefits to maintain the diameter of the nebulized particles at 1–5 μm, and it does not decrease the nebulization effect26. Therefore, the oxygen flowing speed of 6 L/min is selected as the optimum driven nebulization speed, and, in actual practice, the speed was gradually tuned up to 6 L/min in a changing manner from slow to appropriate, to spray the anaesthetic medicament into the expected optimal nebulized aerosol morphology. The appropriate control of the oxygen flowing speed is prospectively good to raise the deposition rate of the nebulized particles in the airway, thereby to increase the effective dose of the nebulized medicament, thus to improve the effect of airway surface anaesthesia. The performing steps of MONI is similar to the traditional ONI, but the equipment parameters and the oxygen flow are controlled through the procedures. Before operation, the attending doctor or the prime care nurse explains to the target patient the purpose of the oxygen-driven nebulized inhalation. If the patient is nervous about the treatment, or emotionally unstable, the doctor/nurse will talk to the patient, to relieve his/her tension. This is something like about the psychological preparation. At the beginning, the patient is instructed to clean his/her mouth, and to stay in sitting or semi-recumbent position. Then the patient is taught to inhale the nebulized Lidocaine aerosol obeying the MONI method in the following steps. During nebulization, the patient needs to put the mouthpiece-type nebulizer attached on his/her mouth, and to inhale gently so that the nebulizer fits snugly around the mouth and remains airtight without leakage. The nebulizer should not be tilted or shaken during the inhalation, so as to less produce air bubbles, which reduce the inhalation efficiency of the medication. Firstly, the patient inhales the aerosol, and contracts the abdomen to enable the chest expanded maximally. In this way, the medicament particles easily enter both lungs through the main airway and the left and right bronchial tubes along with the inhalation movement. Next, the patient needs to hold the breath for 2–3 s, to allow the inhaled aerosolized particles to be deposited in all branches of the airways27. Then, the patient can exhale the breath slowly through his/her nose, with mouth closed, to gradually relax the abdomen, so that the medication can be naturally passed to the distal bronchioles, resulting in an increase of the effective inhaled dose of the medicament. Overall, the slow mouth-inhalation and nose-exhalation movements should be kept in a continuous cycle at the frequency of 8–10 times per minute. The whole nebulization process should maintain for 15–20 min, to drive the glottis open well, making it easier for the bronchoscope to pass smoothly through the vocal fold. If the patient is not feeling well, there can be a short break for 0.5–1 min in between. The evaluation items and their scorings The effectiveness of oxygen-driven nebulized anaesthesia directly affects the smoothness of subsequent bronchoscopic interventions for surgery or examination. A good anaesthetic result can shorten the duration of the surgery/examination, improve the comfortableness of patient while undergoing PTB screening and treatment, and also reduce the incidence of intraoperative and postoperative adverse responses from the patient. The bronchoscopic examination will begin in about 15 min after the nebulization has been completed. The effectiveness of nebulized anaesthesia is assessed by collecting a series of clinical observation items that record the patient’s physiological response when the bronchoscope enters the patient’s airway. These items include: the open status of glottis (OSG), the smoothness of bronchoscope passing through the vocal fold (SBP), the counts of patient’s coughing (CPC), and the record of other responses from the patient during bronchoscopy (ROR). Additionally, the patient’s intra-operative blood oxygen saturation (BOS) and heart rate counts (HRC) are also recorded. The open status of glottis (OSG): Glottis is the gate for the medical tubes interventional to go to the trachea, the bronchi and the lung. A bronchoscope must go through the gate to work. The open status of glottis is important for bronchoscopic operation28. Nebulized anaesthesia is a manner to drive the glottis to a relax and good open status. But different anaesthesia effects will lead to different glottis open status. The operator reads the visualized image and makes the record according to the standardization. The representative visual cases of OSG are regarded as the standardized examples (See Fig. 1). The good or fairly good OSG support to perform the bronchoscopic examination, while the poor (including very poor) situations denies the operation. Visualization of the open status of the glottis. The smoothness of bronchoscope passing through the vocal fold (SBP): It is a kind of operational feedback tactile sense to the operator. The doctor pushes the bronchoscope into the airway during the operation. His hand will get the counterpart touching sense from the tube whether it moves smoothly or in struggle, then he makes the record. The counts of patient’s coughing (CPC)29: It is a counted number to record the times of patient coughing during bronchoscopy. The way of standardization is to range the numbers in qualitative means: no-cough (absolutely no cough), seldom-cough (less than 5 times) and much-cough (more than 5 times). The record of other responses from the patient during bronchoscopy (ROR)30: It is classified in 5 different cases. (Case I) No response; (Case II) Mild reaction (such as face expression of a little pain); (Case III) Moderate reaction (such as face expression of middle pain and slight headshaking); (Case IV) Verbal protest and heavy headshaking, and (Case V) Physical protesting behaviors, like prevention by hands and legs. The blood oxygen saturation (BOS): It is the percentage of oxygenated haemoglobin to total haemoglobin in the blood. It is an important quantitative index to diagnosis for hypoxaemia. Medical criterion tells that a person is evaluated by the BOS value in three segmented ranges31: The range of over 95% ensures the person has saturated oxygen in blood; the range of [90%, 95%) means that the person’s blood is under saturation, but has no obvious abnormal behavioral symptom; and the range of less than 90% tells that the person is lack of BOS and suffering from hypoxaemia, probably leading to several symptoms of oxygen deprivation such as dizziness, weakness, shortness of breath. The heart rate counts (HRC): It counts the heart beating times per minute. The normal HRC is 60–100/min32. Bronchoscopy intervention causes tension and fear in patients, leading to an unstable heart rate and, in severe cases, causing cardiopulmonary disorders or inducing some cardiovascular diseases. Thus, during the bronchoscopic examination, the HRC is maintained on normal monitoring. If the patient’s heart beats faster than normal, it is endorsed that 20% higher is acceptable. If the heart beats slower, a decrease of 10 is on the tolerated boundary. Instantly, the detection should be suspended when the HRC is over exceeded out of the acceptable line. For data records, except for the normal, the HRC ranges of [50, 60) and (100, 120] are regarded as acceptably normal, while the others are abnormal. In practice, the records of OSG and SBP are from the operator (i.e. the attending doctor). The CPC and ROR are the routine records of the patient’s responses by nurses. The BOS and HRC are derived from medical specialized instruments33. Assuming that these 6 items are available for digital scorings, the effectiveness and efficiency of the traditional ONI and the MONI methods are firstly scored, and then compared for airway surface anaesthesia. These data will be considered in joint-use for comprehensive evaluation on the effect of airway surface anaesthesia. For data analysis, these recorded evaluation items are rescaled by digital scorings for different cases, to support the enforcement of machine learning methodologies and modeling strategies. A within-five-star scoring criterion is designed. The surgery-regarded worst case is scored 1, the best case is scored 5, and the stepwise cases in between were scored by equal intervals. The items and their corresponding rescaled scoring values are listed as in Table 1; Specially for the OSG scoring, we deliberately separated it into two scoring levels, where the “Fairly good” and “Good” levels were both scored 5 because they support bronchoscopic examination, while the “Poor” and “Very poor” levels were scored 1 as they deny the operation. Table 1Different cases of the evaluation items and their corresponding rescaled scorings.ItemsData recognition for endorsed levelsOSG caseFairly good & Good(Support)Poor & Very poor(Deny)OSG scoring51SBP caseSmoothStruggledSBP scoring51CPC caseNo-coughSeldom-coughMuch-coughCPC scoring531ROR caseCase ICase IICase IIICase IVCase VROR scoring54321BOS caseSaturation()Under saturation()Lack of BOS()BOS scoring531HRC caseNormal()Acceptably normal()Abnormal()HRC scoring531 Different cases of the evaluation items and their corresponding rescaled scorings. Fairly good & Good (Support) Poor & Very poor (Deny) Saturation () Under saturation () Lack of BOS () Normal () Acceptably normal () Abnormal () Data acquisition This is a kind of retrospective study. The experimental data is from the clinical observations of patients during the diagnosis. The data was collected in the time duration from April 2023 to March 2024, in the Bronchoscopy Clinical Examination department of Liuzhou People’s Hospital (a Level-3 A hospital) in Guangxi Province, China. Based on their medical history, the patients were selected who were suffering from respiratory infectious diseases like PTB. Practically, a total of 200 participants were selected. These patients are aged from 18 to 60. To launch the comparative experiment, 100 of these patients are randomly selected and grouped for MONI operation, the remaining 100 patients grouped for ONI. As abovementioned, the practical operation steps are similar for MONI and ONI; the main difference is that the MONI method is proposed to strictly quantitative control on the oxygen flow and the nebulizer’s inherent parameter. The participants were prohibited from eating for 4–6 h before the operation and from drinking water for the last 2 h. Then a doctor or a nurse guided and assisted the patients went through the oxygen-driven nebulized anaesthesia by following the MONI or ONI steps, respectively. After nebulization, the primary care nurse delivered the patient to the operation room. On the way, the nurse informed the patient the key points of the operation, the possible occasional physical reaction and the correct way to cope with them. The nurse is also responsible to relieve the patient’s tension and encourage them to be active in co-operation with the doctor when under treatment. In the operation room, the patient should lie flat on the examination platform in the cervical supine position. The nurse gently covers the patient’s eyes with a mask to protect them from the irrigation fluid that would be used subsequently. Successively, the nurse made a nasal catheter oxygen inhalation to the patient. A moderate oxygen flow was controlled by the scoring when for MONI, but adjusted to the patient’s physical condition when for ONI. Simultaneously, the patient was equipped with the continuous BOS monitoring and HRC monitoring were performed. The instrumental setup for bronchoscopy operation is showed in Fig. 2. It is composed of the CV-290 endoscopic system and a BF-260 electronic bronchoscope (produced by Olympus inc., JP). The bronchoscope needs to be cleaned and strictly sterilized before intervention into the human body, following the instructions of the “Regulation for Cleaning and Disinfection Technique of Flexible Endoscope” (a China’s Health Industry Standard, No. WS 507–2016). Commonly, the bronchoscope tube goes into the patient’s tracheological airway system through the nose. Before entering the nose, the bronchoscope is fully lubricated at its front end to reduce friction to the airway surface. The instrumental setup in the operation scenery. (a) The instrumental setup, (b) The CV-290 endoscopic system, (c) The BF-260 bronchoscope. In the path, the tube goes through the glottis into the main airway, with selection into the left or right brunches, and until the next branches. Combined with the CT-scanned markers, the bronchoscope rapidly finds the concerning locations and identifies the possible lesions, where the tube will lavage the area, make micro-graphic observation and take high-resolution photographs. Also, the tube will suck the airway secretions and take the specimen of lavage fluid out of the body for in-vitro medical pathology analysis. The whole process of bronchoscopic intervention in the body generally takes less than 10 min. The tube will withdraw from the body along its original route, with simultaneous monitoring on the patient’s physiological conditions evaluated by the scoring items. The bronchoscopic operation is in charge by the attending doctor, and assisted by a surgery nurse. The primary care nurse usually sits in to monitor the patient’s reaction, and recorded the operation data. When the operation was finished, the patient was required to rest in the observation room for half an hour. The medical staff observed the patient for any possible adverse reactions, such as cough, nausea, vomiting, chest tightness, shortness of breath or airway spasm, and made these records to be added as the post operative data. In summary, each bronchoscopic operation performed on a patient generates a set of data, which serves as a recorded sample. The data includes multiple variables including OSG, SBP, CPC, ROR, BOS and HRC. The MONI group data and the ONI group data are collected respectively. But it is not easy to identify which method is better for nebulized anaesthesia by reading the raw data. Methodology should be investigated to mine the information of data features, for effective comparative analysis of the anaesthesiatic effect by MONI and ONI. Methodologies for data analysis The effect of airway surface anaesthesia is strongly related to the recorded operative data as well as to the patients’ physiological data. Such data exhibits distinct multi-level characteristics and diversity. Given the comprehensiveness and complexity of evaluating anaesthesia effectiveness, advanced machine learning (or statistical) methods are required for data analysis. This study is to build up an architectural machine learning model based on a broad learning system (BLS). The model is established with its parameters tunable for optimization, in order to enhance the evaluation of airway surface anaesthesia effectiveness by automatic parameter adjustment for intelligent analysis. The BLS-based model is trained based on the collected data of the six scoring items of OSG, SBP, CPC, ROR, BOS, HRC. For model optimization, the sessions of feature extraction, network activation and transformed prediction of the BLS network are combined for adaptively tuning their structural parameters. By using the scalable BLS model, the comparative evaluation between the ONI and MONI methods can be prospectively developed into a data-driven smart analytical manner. The broad learning network architecture The broad learning system (BLS) is an extended architecture of a fully connected network with random weight (FCNRW), which is to generate an additional layer to broaden the input data34; herein, the additional layer is called extended layer. The structure of a BLS network is showed in Fig. 3. The recorded patients’ data are imported to the BLS network through the input layer. A typical FCNRW is to delivered the input data directly to the hidden layer by the transformed mapping of a neural perceptron. In counterpart, the BLS is especially functional to deliver the input data to form the extended layer, so that the input data can be meticulously explained. The input layer and the extended layer are combined into a renewal input layer, which is called a virtual input layer, as to extensively accumulate the input data information. Next, the data contained in the virtual input layer are flattened, dropped out and joint for a series of perceptron computations, being connected to a hidden layer for feature selection, and then the informative variables are transformed for a Softmax activation to yield the final output. The structure of a broad learning network. In algorithmic details, supposing that the input data is , where represents the number of the sampling patients, and points to the number of nodes in the input layer for receiving the data of variables for each record. As is introduced in Sect. 2.2, there are 6 scoring items on concerning for the patients, which means that the data has six different variables, thus is constantly valued as 6. The extended layer is produced for mapping to generate a series of extended input nodes, where the number of extended nodes (denoted as ) is tunable for necessary model optimization. The mapping from the input layer to the extended node is formulated as, where is a activation function for extended data mappings; and are the adaptive tuning weight and the bias, respectively; their matrix forms are denoted as and ; and is the generated data of the i-th extended node. For convenient denotation, the data of all the extended input nodes are aggregately marked as a matrix , namely, . According to the BLS principle, the input data and the extended input data are combined for fusion computation by following the FCNRW structural flow. Therefore, the accumulated data in the total virtual input layer is mapped to the hidden layer, and the data features are extracted in the way of neural perceptron calculation, which is expressed in the equation, where is the activation function for hidden activation; and represent the adaptive tunable weights, respectively applied for and ; the combination form indicates that the two weights are for joint training; is the bias; and represents the output data of the j-th hidden neural node. Similarly, the all-in-one output of the hidden layer is denoted as the matrix . Finally, the calculated features selected out from the hidden layer are transformed to the Softmax unit for quantitative regression or qualitative clustering computation. As the BLS network is functional to identify the effective features of the data, then the features are regarded as informative. Thus, a simple linear model applied in Softmax probably suits for model prediction. Suppose that the ultimate output of BLS is , which can be linearly predicted as where counts for the regression coefficients, and is the bias; is the activation function for prediction. For the demand to evaluate the effect of nebulized anaesthesia to the airway surface, the BLS architectural model need to be trained with the recorded data. The model training target is to optimize the weights for observing the maximum value of . The reason is that Y is originated from the digitalized scorings of the 6 recorded evaluation items, while the scoring standard use a higher score to mark a better case. Therefore, the training of the model can be transformed to solve the following optimization problem, where represents the trained optimal values of all the weights applied in the BLS. When the optimizational BLS model is identified by observing the optimal weights, all of the samples (i.e. the targeted patients) are immediately predicted. From the abovementioned modeling deduction progress, it can be easily derived that the BLS model for data-driven scoring to the effect of airway surface nebulized anaesthesia. Individually, for the -th patient (for ), the input data only applies to the targeted sample . The prediction performs as the following equations, where the network weights applied the previously trained optimal set . Subsequently, the model prediction on can be used to evaluate the effect of nebulized anaesthesia, and further for the comparison between the MONI method and the traditional ONI. The activation functions The proposed broad learning system tries to train the connection weights between cross-layer neural nodes for optimization. The tuning progress is originally driven by the information extracted from the input data (i.e. the medical staff’s records of the sample patients). Notably, the activation functions are influencing the data for delivery and for the output. As abovementioned, the function is used to activate the perceptrons for the mapping to generate the extended data; for the mapping to transfer the data to hidden neurons; and for the mapping to compute the output data. By considering the input of each layer and the expected output information transformed through the layers, the ELU function, the Sigmoid function and the Leaky ReLU function (See Fig. 4) are applied respectively for , and to train and test the scalable BLS architecture. The activation functions used for layered neural perceptron in BLS. (a) The ELU functionfor the mapping X→Z, (b) The Sigmoid functionfor the mapping X, Z→H, (c) The Leaky ReLU functionfor the mapping H→Y. Indicators for the scalable BLS modeling Using the above-proposed methodologies, the MONI method and the traditional ONI method are evaluated by inputting the scoring data of the 6 recorded clinical items (i.e. the OSG, SBP, CPC, ROR, BOS and HRC) in to the BLS network architecture, and reading the output of the network. As is known from the algorithmic procedure, the final output data is endorsed as comprehensively containing the feature information extracted from the 6 items. It can be used to evaluate the grouped data in the effect of airway surface anaesthesia. To validate whether the resultant data of is suitable to evaluate the performance of the BLS architecture, a loss function is proposed by drawing on the concept of Spearman’s rank correlation coefficient35. The function is constructed in similarity to the minus term of the Spearman’s equation, namely, . where represents each of the items in {OSG, SBP, CPC, ROR, BOS, HRC} in turn; represents the rank difference between and after ascendingly sorting by Spearman’s rank; is used to calculate the second norm of a vector whose elements record the -th scoring item of all targeted samples; is the number of concerning evaluation items; in this study, we have ; the operator means to have the accumulated summation of all these 6 items. As a higher Spearman’s correlation coefficient indicates there is better explainability in ; accordingly, the loss value, which is equivalent to the minus term of Spearman’s equation, is more preferable to optimize the model when it is calculated smaller. Moreover, a significance test is vital for the comparative evaluation of the anaesthesia effect between MONI and ONI. The test is numerically based on the output data of the parameter-tunable BLS model. In statistic concept, Friedman’s test (F-test) is an appreciable evaluation method to validate the significant difference for comparison of two dimension-consist groups of data36. It is meticulously suitable to finish the comparison task in this study. Practically, the variance value is calculated for the MONI group and for the ONI group, respectively, and then the F value (the standard statistical indicator for Friedman’s test) is deduced. Theoretically, the boundary F value can be identified from the Friedman’s test table. The significance test for the comparison of MONI and ONI is alternatively performed by directly comparing the calculated sample F value () to the table-determined boundary F value (). For conclusion, means that there is no significant difference between the MONI group data and the ONI group data; otherwise, indicates that there is a significant difference between the two sets of data. Model design and parameter tuning The sample data from clinical records are available for analysis by using the proposed scalable BLS architectural method. For convenient comparison, the MONI group data and the ONI group data are respectively input to the model for network perceptron training. The data is in size of 100 samples for each group applied for BLS model training. The BLS produced 6 neuron nodes in the input layer to receive the rescaled scorings on OSG, SBP, CPC, ROR, BOS and HRC. In modeling, the input data was firstly used to generate a series of extended nodes as to form the extended layer. This generating process applied the based perceptron transforming way, working with the ELU function for activation. The number of extended nodes were planned for tuning as changeable from 3 consecutively to 20 for model optimization. Each extended node was transformed from the 6 input nodes with adaptively weight auto-adjustment. Successively, the extended layer and the original input layer were combined as the broadened virtual input layer, the number of virtual input nodes () are designed accordingly as tuning from 9 to 26. This is one of the structural parameters which are decisive to the data-driven optimization of the BLS model. Next, the data in the virtual input layer were further delivered to the hidden layer for feature extraction. In the duration, the hidden layer was designed as composed with a tunable number of neuron nodes (). The architecture was fully tested for the structural variation of changing from 1 to 8. This is the other important parameter for scalable training of the BLS model. The informative feature at each hidden node was transformed in the perceptron mode from a series of data stored at the tunable number of virtual input neural nodes. Sigmoid function was applied for the activation of this transformation. Finally, the feature data in the hidden layer were moved to the Softmax unit for calculation of the network output value () according to Eq. (3). The activation in the Softmax applied the Leaky ReLU function. Successively, the loss value is obtained by Eq. (9), to identify the information of the targeted 6 clinical items that are endorsed in . The BLS model was designed as scalable with its two vital structural parameters as tunable. With network linkage, there generated plenty of perceptron sub-models with thousands of linking weights being adaptively adjusted to their optimally applicable values. The weights were tuned during the modeling training process, and optimized obeying the computation rule defined by Eq. (4). As the results, Fig. 5 shows the visualization of the BLS model optimization by joint grid search of the tuning structural parameters of and . In general comparative viewpoint, input of the MONI group data is able to observe lower loss values than the input of ONI group, which indicates that the MONI data performed better than the ONI data. Visualization of model optimization by joint tuning of the BLS structural parameters for the comparison between MONI and ONI data. Results and discussions After parameter optimization for the scalable BLS model, it is obtained that the most optimized BLS model for MONI data was found with and , there deducing the minimal loss value of 2.253; while the best model for ONI data obtained the minimal loss value of 2.677, acquired from the structural parameters of and . These two optimized BLS structures were identified. Moreover, the output values of the models were calculated by adaptively adjusting their network linking weights. Figure 6 shows the blue line of the output values for the 100 MONI-group samples and the red line for the 100 ONI-group samples. In the figure, the gaps between the two lines were filled with light blue color (where the blue line is higher than the red line) or with yellow color (where the red line is higher). It can be learned from Fig. 6 that, for individuals (i.e. the patient individuals), there are most samples that have higher output values from the MONI data than from the ONI data, although there are still a few exceptions that are lower. By visualization, the resultant blue line representing MONI data is generally higher in average than the red line representing the ONI data, and the light blue area is obviously larger that the yellow fill-in. We conclude that the prediction values from the optimized BLS network indicate that MONI data is better than ONI data in the clinical viewpoint, by noting that the BLS output is the comprehensive information transformed from the original scoring data of the 6 important clinical items of OSG, SBP, CPC, ROR, BOS and HRC. The predicted output values of the 100 individual samples from the optimized BLS models, respectively for the MONI data and for the ONI data. Furthermore, F-test was applied to validate the significant difference between the MONI and ONI data groups in this comparison issue. The sample F values () were calculated based on the variances of the output data observed from the optimized BLS models, respectively identified for MONI and ONI. The calculated equals to 2.237. And it is easy to find that the boundary F value () is 2.198. The result of ensures that the difference between the MONI and ONI data is significant, thus to prove that the conclusion is solid for the fact that the clinical observation from MONI data is better than that from the ONI data. To validate the efficiency of the proposed BLS network, we have performed the ANOVA analysis using the SPSS Statistics software (v.26), which is traditional statistical method consistent with standard clinical research practices. ANOVA is to validate the significance of the difference between the two methods regarding each scoring item. The validation results are presented in Table 2. We can learn from the table that, by ANOVA analysis on each item, the differences of MONI and ONI method are significant for the scorings of SBP, CPC, BOS and ROR; for these four items, the MONI method was validated obviously better than ONI. However, the difference between the two methods regarding the OSG and HRC items are not significant. In comparison to ANOVA, the BLS optimization approach is able to conduct a comprehensive modeling distinguishable result by fusion of the 6 clinical items of OSG, SBP, CPC, ROR, BOS and HRC. The proposed method is able to establish factor-combination analytical models better than the traditional statistical method in accordance with each item individually. Table 2The ANOVA analytical results by using SPSS software.Model testEstimated marginal scoringsF valueSignificanceMONI groupONI groupOSG2.1830.1124.6804.560SBP9.6070.0014.5603.800CPC9.4020.0013.6503.080ROR8.2360.0013.8503.160BOS5.9910.0054.7674.535HRC1.7640.1864.9404.973 The ANOVA analytical results by using SPSS software. By feature extraction of the parameter-scalable BLS architecture, the information of the targeted 6 clinical scoring items were transformed through the network and ultimately stored in the output data . To reveal the explainability of to each of the items, we calculated the Pearson’s correlation coefficient between the values deduced from the most optimized BLS models and each of the 6 items of OSG, SBP, CPC, ROR, BOS and HRC, respectively. Figure 7 shows the correlation results for the optimized model trained by the MONI data (sub-figure (a)) and for the model by ONI data (sub-figure (b)). It can be seen from Fig. 7 that the BLS information-transformed output data is highly correlated to the clinical items of SBP, CPC, ROR and BOS, while lowly correlated to the items of OSG and HRC. These results indicates that the proposed scalable BLS architecture adaptively suits to fast extract the feature information from the clinical recorded data, especially for the SBP, CPC, ROR and BOS items. Correlation coefficient between the output value of the optimal BLS models and the targeted scoring items. (a) for the MONI data, (b) for the ONI data. The fact that the network output value fails to explain the OSG and HRC data is consistent to the results from traditional ANOVA analysis by SPSS software. These correlation results yield the efficiency and reliability of the proposed BLS network optimizational method. Actually, the reason is further explainable. When undergoing clinical surgery, the patient may have psychological tension. This tension is caused by the operation which is taken as the abnormal invasion to human body. It directly affects the instability of heart rate. Similarly, due to this psychological tension, the patient does not naturally relax his/her glottis, thus indirectly leading to that the glottis does not open well in good respondance to professional medical operation. In contrast, the other four items can be well explained by statistical network modelling because they are not highly related to the physiological response of the patient undergoing surgery. Typically, the SBP, CPC and ROR items are highly correlated with surgeon skill, whereas the BOS item is mostly related to the effectiveness of airway surface anaesthesia. The correlation analysis revealed another aspect to compare the proposed MONI method and the traditional ONI method. MONI and ONI are both successful to be explicated by high correlation coefficients (> 0.8) for the detection of SBP, CPC, ROR and BOS items, and there is little difference by comparison. However, the correlation difference on OSG and HRC are significant although the values are not high enough. The differences validate that the MONI method, compared to the ONI method, is able to reduce the influence on surgery monitoring from the patients’ physiological response, and improve the information validity of the clinical data by data mining technologies. These model analytical results highlight that AI methodologies are effective in supporting the precise monitoring of the operation procedures during the process. The design of scalable tuning for BLS architectural parameters represents the first attempt to introduce AI-assisted clinical bronchoscopic diagnosis, providing a new approach to developing bronchoscopy-related anaesthesia operations for precise clinical evaluation. The characteristic of scalable broad learning model is to generate an appropriate number of extended neurons to deliver the extracted features as informative signals for model improvement. However, the AI-assisted scalable BLS model has limitations. Firstly, the extracted features are difficult to explain in relation to the pathological basis; the results are data-driven outputs. Secondly, the BLS structural parameters are fully automatically adjusted to the data, based on the assumption that the clinical observed items can be digitally scored. On this basis, the scalable training of BLS with auto-adjusted parameters makes it convenient to launch real-time data analysis in clinical operations. With modeling support from the proposed methodologies, data experiments can be conducted to facilitate the validation of detection and prevention strategies for PTB-like infectious diseases during clinical practice. By quantitative comparison on scoring items between the MONI and ONI methods, we validated the modified method as an acceptable improved choice for airway surface anaesthesia. The approach holds prospect for the emergence and development of intelligent control in clinical anaesthesia operations. Conclusions In this study, comparative analysis was carried out for MONI and ONI methods in support for airway surface anaesthesia by data computational experiment. Data was collected from 100 clinical cases in the MONI group and 100 in the ONI group, respectively. The 6 clinical items of OSG, SBP, CPC, ROR, BOS and HRC were recorded for each patient. The scoring data of these items were delivered to a scalable BLS network architecture for data analysis. By BLS optimization, the output prediction data , which comprehensively interprets the information of the 6 concerning items, is used to evaluate the airway surface anaesthetic effects for the comparison between the MONI and ONI methods. Results show that the MONI data with precise controls performed better than the ONI data. Furthermore, the optimal BLS models were proved significant by the statistical F-test theory. Successively, the correlation coefficients were calculated between the BLS output data and the 6 clinical scored items, to confirm the explainability of the output data. The resultant correlation coefficients for SBP, CPC, ROR and BOS items are high enough to validate the feasibility of the network model and the scalable training strategy. The significant difference in correlations between MONI and ONI for OSG and HRC implies that the proposed MONI method can practically enhance the anaesthetic effect on the airway surface, as revealed by data modeling using the BLS-based network architecture. The experimental results imply that studying adaptive model training methods may yield an improved approach for detecting infectious diseases. Nevertheless, BLS is a data-driven modelling architecture that assumes the clinical observations can be digitally scored. The output prediction results indicate that high scores reflect a good anaesthetic effect for the MONI method, which validates its improvement for the airway surface anaesthetic effect. In reality, the extracted BLS features are difficult to link to pathological explanations. Therefore, more work is needed to find an explainable statistical model that adheres to the pathological basis, and more metrics are needed to improve the scoring tactics. With the current development of AI, we have conducted the first trial to quantitatively analysis clinical observation data. The proposed scalable BLS modeling technologies have been validated beneficial for the clinical examination of PTB. In the future, we anticipate more AI + clinical/medical technologies will contribute to the prevention and diagnosis of PTB and other infectious diseases. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Acknowledgements This work was supported by National Natural Science Foundation of China (62365008), Scientific Research Projects of Guangxi Health Commission (Z-B20221296, Z-B20221309). Author contributions Yanping Sun: Conceptualization, Formal analysis, Funding acquisition, Investigation, Project administration, Resources, Validation, Writing - Original Draft, Writing - Review & Editing. Tao Chen: Data Curation, Methodology, Software, Visualization, Writing - Original Draft, Writing - Review & Editing. Kai Shi; Data Curation, Formal analysis, Methodology, Software, Visualization, Writing - Original Draft. Shenglin Mo: Formal analysis, Funding acquisition, Investigation, Validation. Yufei Deng: Data Curation, Resources, Visualization. Lihua Liao: Investigation, Validation. Huazhou Chen: Conceptualization, Funding acquisition, Methodology, Project administration, Supervision, Writing - Review & Editing. All authors read and approved the final manuscript. Data availability The datasets used during the current study available from the corresponding author on reasonable request. Code availability The proposed methods and algorithms were implemented via MATLAB programming, and the main codes supporting the findings are available as supplementary materials. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Yanping Sun and Tao Chen have contributed equally to this work. References Associated Data Supplementary Materials Supplementary Material 1 Data Availability Statement The datasets used during the current study available from the corresponding author on reasonable request. The proposed methods and algorithms were implemented via MATLAB programming, and the main codes supporting the findings are available as supplementary materials.