Infant asthma three-branch classification auxiliary judgment method and medium

Abstract

The present application relates to a kind of infantile asthma three branch classification auxiliary judgment method and medium, it is related to data processing technical field.The method includes: the tidal breathing data and demographic data of infant are collected and preprocessed, the tidal breathing data and demographic data after preprocessing are extracted and grouped log likelihood ratio, obtain the log likelihood ratio corresponding to each group;Each log likelihood ratio is fused by hierarchical bayes, and the fusion result is obtained, and the fusion result is based on three branch classification auxiliary judgment of asthma risk.Compared with prior art, infantile asthma can be more objectively and accurately based on tidal breathing data auxiliary judgment.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically to a method and medium for assisting in the three-branch classification of infant asthma. Background Technology

[0002] Asthma is the most common chronic respiratory disease in childhood, mainly characterized by chest tightness, wheezing, cough, and difficulty breathing. In 2021, there were approximately 300 million asthma patients worldwide, with a higher prevalence in children than adults. In 2019, the number of children diagnosed with asthma globally reached an unprecedented nearly 22 million. Therefore, early diagnosis and intervention in infants and young children are crucial for improving long-term prognosis. However, accurate identification of asthma in infants and young children aged 3 years and under remains a significant clinical challenge in pediatric respiratory medicine. Currently, the clinical diagnosis of asthma in infants and young children heavily relies on the experience of the attending physician, the interpretation of the child's symptoms (such as wheezing) described by the guardian, and a presumptive diagnosis after ruling out other causes; the value of auxiliary tests is limited. Clinically, exhaled nitric oxide (FeNO) testing and tidal breathing pulmonary function testing are mainly used for infants and young children aged 3 years and under. While FeNO testing, as a biomarker of airway inflammation, has some value in the diagnosis of asthma in adults and older children, its applicability is limited for infants and young children aged 5 years and under, and it is not yet widely adopted in some hospitals. Tidal breathing pulmonary function testing, as a passive and non-invasive method of respiratory function assessment, has long been used to evaluate the lung function of infants and young children, as well as uncooperative subjects. Studies have demonstrated that several measurements based on tidal breathing (such as time to peak emission ratio (tPTEF / tE), volume to peak emission ratio (vPTEF / vE), TEF25 / VE, TEF50, and TEF75) can reflect airway obstruction and small airway function to some extent, and are statistically associated with respiratory function trajectories or asthma-related outcomes in subsequent age groups (e.g., tPTEF / tE is associated with later airway obstruction or asthma risk). These studies provide physiological and epidemiological evidence for using tidal breathing parameters for early screening.

[0003] Currently, commonly used clinical tidal breathing pulmonary function testing equipment rapidly provides quantitative results by calculating single or a few traditional indicators (respiratory rate, tidal volume, PEF, tPTEF / tE, TEF25 / VE, etc.), featuring simplicity, rapid calculation, and ease of clinical implementation. Standardized studies have been conducted on related equipment and measurement methods, making them suitable for passive measurements without subject cooperation. Various instruments and analysis software for tidal breathing measurements are available for clinical application (including the engineering implementation of data acquisition, filtering, and effective respiratory segment identification processes). These tools support routine measurements and preliminary analysis, facilitating widespread adoption.

[0004] Although the existing tidal breathing test has some application value, the technology still has significant shortcomings, which limits its widespread application and diagnostic performance in the diagnosis and treatment of infant asthma. Summary of the Invention

[0005] The technical problem to be solved by this application is to provide a method and medium for assisting in the three-branch classification of infant asthma, which has the characteristics of being able to more objectively and accurately assist in the diagnosis of infant asthma based on tidal breathing data.

[0006] Firstly, a method for assisting in the three-branch classification of infantile asthma is provided, including:

[0007] Tidal breathing data and demographic data of infants and young children were collected and preprocessed. The collected tidal breathing data included tidal breathing test data and tidal breathing loops, and the collected demographic data included gender, age, height and weight.

[0008] Heterogeneous features were extracted and grouped log-likelihood ratios were obtained from the preprocessed tidal breathing data and demographic data to obtain the log-likelihood ratios for each group. The extracted heterogeneous features included traditional tidal breathing test data features, higher-order derived coupling features, curve morphology features, and demographic data features. The higher-order derived coupling features refer to the ratio-type respiratory dynamics index formed by cross-domain combination and nonlinear transformation of multiple basic dimensions in the tidal breathing signal. The curve morphology features refer to the structural information of the curve's shape, symmetry, angle changes, local slope, and overall configuration extracted by geometric fitting and modeling the complete two-dimensional flow-volume loop during tidal breathing.

[0009] Hierarchical Bayesian fusion is performed on each log-likelihood ratio to obtain the fusion result;

[0010] Asthma risk is assessed using a three-branch classification based on the fusion results; the three-branch classification results include high asthma risk, low asthma risk, and an indeterminate range.

[0011] In a second aspect, one embodiment provides a computer-readable storage medium storing a program that can be loaded and executed by a processor as a method for assisting in the three-branch classification of infant asthma.

[0012] The beneficial effects of this invention are:

[0013] Because heterogeneous feature extraction and grouped log-likelihood ratio acquisition are performed on preprocessed tidal breathing data and demographic data, obtaining the log-likelihood ratio for each group allows for the application of different probability models suitable for different feature types. Employing distribution assumptions that better reflect the actual data yields more accurate and stable parameter estimates. Furthermore, hierarchical Bayesian fusion of the log-likelihood ratios enables unified inference across different statistical models and feature types, giving the final output probability a clear Bayesian physical meaning, which can be directly used as input for subsequent three-branch decision mechanisms. Because the three-branch classification structure... The results include high asthma risk, low asthma risk, and uncertainty intervals, enabling a three-branch decision-making mechanism that better aligns with medical risk control. Because the extracted heterogeneous features include higher-order derived coupling features, they can simultaneously reflect the relative intensity, synchronicity, and proportional relationships between respiratory physiological processes, exhibiting higher sensitivity in the early stages of subtle pathological changes, thus revealing potential structural information that cannot be expressed by a single original feature. Furthermore, because the extracted heterogeneous features include curve morphology features, they can focus on characterizing the geometric morphology of expiratory phase lesions, thereby compensating for the shortcomings of traditional test parameters in representing geometric features and improving the completeness of the tidal breathing feature system. Therefore, the proposed solution can more objectively and accurately assist in the diagnosis of infantile asthma based on tidal breathing data. Attached Figure Description

[0014] Figure 1 This is a schematic flowchart of an embodiment of the method for assisting in the three-branch classification of infant asthma according to this application;

[0015] Figure 2 This is a schematic diagram of a two-dimensional flow-volume loop curve for tidal breathing according to an embodiment of this application;

[0016] Figure 3 This is a schematic flowchart of a method for heterogeneous feature extraction and grouped log-likelihood ratio acquisition according to an embodiment of this application;

[0017] Figure 4 This application Figure 1 A flowchart illustrating one embodiment of step S30 in the illustrated embodiment. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to specific embodiments and accompanying drawings. Similar elements in different embodiments are referred to by associated similar element reference numerals. In the following embodiments, many details are described to facilitate a better understanding of this application. However, those skilled in the art will readily recognize that some features may be omitted in different situations, or may be replaced by other elements, materials, or methods. In some cases, certain operations related to this application are not shown or described in the specification. This is to avoid obscuring the core parts of this application with excessive description. For those skilled in the art, detailed description of these related operations is not necessary; they can fully understand the related operations based on the description in the specification and general technical knowledge in the art.

[0019] Furthermore, the features, operations, or characteristics described in the specification can be combined in any suitable manner to form various embodiments. At the same time, the steps or actions in the method description can be rearranged or adjusted in a manner obvious to those skilled in the art. Therefore, the various orders in the specification and drawings are only for the clear description of a particular embodiment and do not imply a necessary order, unless otherwise stated that a particular order must be followed.

[0020] The serial numbers assigned to components in this article, such as "first" and "second", are used only to distinguish the objects being described and have no sequential or technical meaning.

[0021] To facilitate the explanation of the inventive concept of this application, the tidal breath test technology will be briefly described below.

[0022] The applicant found in their research that, on the one hand, traditional indicators, mostly based on peak values, time ratios, or single-point ratios, are insufficient to comprehensively describe the morphological details of respiratory waveforms (such as curve asymmetry, local depressions / protrusions, slope changes, etc.). Therefore, when pathological changes are only reflected in the waveform microstructure or dynamic patterns, these indicators are often insufficiently sensitive, leading to missed diagnoses or blurred discrimination boundaries. On the other hand, although there have been attempts in recent years to apply machine learning methods to respiratory signal analysis and disease identification (including feature engineering-based classifiers, dimensionality reduction methods such as PPCA, and some deep learning / embedded methods), most are single models or end-to-end black-box methods, lacking a systematic framework for modeling and fusing different categories of information in a probabilistically compatible manner. Many existing studies or tools only provide numerical indicators or threshold judgments, lacking a systematic approach that incorporates clinical risks (such as the different costs of missed diagnoses and misdiagnoses) into the decision-making process; there are also few standardized implementations of "uncertainty" as a clear output (i.e., three-way decision), making it difficult to handle gray area samples and the allocation of examination resources in clinical practice.

[0023] In view of the above problems, this application provides a method and medium for auxiliary judgment of the three-branch classification of infant asthma. First, tidal breathing data and demographic data of infants are collected and preprocessed. Since heterogeneous features are extracted and grouped log-likelihood ratios are obtained from the preprocessed tidal breathing data and demographic data, the log-likelihood ratios corresponding to each group are obtained. This allows for the application of different probability models suitable for different feature types, and the use of distribution assumptions that are more closely aligned with the actual data can produce more accurate and stable parameter estimates. Furthermore, since hierarchical Bayesian fusion is performed on the various log-likelihood ratios to obtain the fusion result, unified inference between different statistical models and different feature types can be achieved, making the final output probability clearly defined. The Bayesian physical meaning can be directly used as input for the subsequent three-branch decision-making mechanism. Since the three-branch classification results include high asthma risk, low asthma risk, and an uncertain interval, a three-branch decision-making mechanism more consistent with medical risk control can be obtained. Because the extracted heterogeneous features include higher-order derived coupling features, they can simultaneously reflect the relative intensity, synchronicity, and proportional relationships between respiratory physiological processes, exhibiting higher sensitivity in the early stages of subtle pathological changes, thus revealing potential structural information that cannot be expressed by a single original feature. Furthermore, because the extracted heterogeneous features include curve morphology features, they can focus on characterizing the geometric morphology of expiratory phase lesions, thereby compensating for the insufficient representation of geometric features by traditional test parameters and improving the completeness of the tidal breathing feature system. Therefore, the proposed solution can more objectively and accurately assist in the diagnosis of infantile asthma based on tidal breathing data.

[0024] The present application solution will be described in detail below with reference to specific embodiments.

[0025] This application provides an auxiliary method for the three-branch classification of asthma in infants and young children. Please refer to... Figure 1 ,include:

[0026] Step S10: Collect and preprocess tidal breathing data and demographic data of infants and young children.

[0027] In this embodiment of the application, the collected tidal breathing data includes tidal breathing test data and tidal breathing loop (flow-volume curve), and the collected demographic data includes gender, age, height, and weight.

[0028] Those skilled in the art will understand that, since the infant asthma three-branch classification auxiliary judgment method of this application is based on a trained heterogeneous feature fusion model based on a hierarchical Bayesian framework, the training data prepared before model training includes infants with and without asthma, and the collected tidal breathing data also includes tidal breathing test data and tidal breathing loops. Specifically, the subjects collected from infants with asthma include infants with three types of asthma (infection-induced asthma, non-infection-induced asthma, and cough-variant asthma).

[0029] In the preprocessing, the preprocessing of demographic data includes encoding, and the preprocessing of tidal breathing data includes filtering out noise and outliers.

[0030] Step S20: Extract heterogeneous features and obtain grouped log-likelihood ratios from the preprocessed tidal breathing data and demographic data to obtain the corresponding log-likelihood ratios for each group.

[0031] In this embodiment, the extracted heterogeneous features include traditional tidal breathing test data features, higher-order derived coupling features, curve morphology features, and demographic data features. Among them, the higher-order derived coupling features and curve morphology features are obtained based on the tidal breathing loop.

[0032] The characteristics of conventional tidal breathing test data consist of 15 values ​​directly output by the conventional tidal breathing test, including: respiratory rate, tidal volume / kg, inspiratory time, expiratory time, ratio of inspiratory time to expiratory time (IST / EXE), time to PTEF (TPTEF), tPTEF%tE(mean), tPTEF%tE(SD), tPTEF%tE(highest), VPEF%VE(mean), VPEF%VE(SD), VPEF%VE(highest), mean inspiratory flow, mean exspiratory flow, and peak exspiratory flow.

[0033] Those skilled in the art will understand that the above 15 types of traditional tidal breathing test data characteristics are only a preferred embodiment of this application, and those skilled in the art can add or subtract them based on actual needs.

[0034] Higher-order derived coupling features refer to ratio-based respiratory dynamics indices formed by cross-domain combination and nonlinear transformation of multiple basic dimensions in tidal breathing signals.

[0035] The higher-order derived coupling features proposed in this application differ from the single-point or single-dimensional parameters used in traditional clinical practice. These coupling features can simultaneously reflect the relative intensity, synchronicity, and proportional relationships between respiratory physiological processes, thereby revealing potential structural information that cannot be expressed by a single original feature. Since the core pathological mechanism of asthma is usually manifested by the interaction of multiple factors such as increased airway resistance, decreased small airway compliance, and restricted dynamic flow rate during the expiratory phase, higher-order derived coupling features can capture these coupling effects in tidal breathing with mild symptoms, and therefore have higher sensitivity in the early stages of subtle pathological changes.

[0036] In this embodiment, the higher-order derived coupling features include: the ratio of the area enclosed by the flow-volume loop to its perimeter, the ratio of the peak expiratory flow rate of the flow-volume loop to its tidal volume, the ratio of the expiratory area to the inspiratory area, the ratio of the area of ​​the flow-volume loop in the 50%–100% volume range after expiration to the entire expiratory area, the ratio of the peak expiratory flow rate to the peak inspiratory flow rate, the ratio of the instantaneous flow rate at 25% expiratory volume to the total expiratory volume, the ratio of the instantaneous flow rate at 50% expiratory volume to the instantaneous flow rate at 50% inspiratory volume, the ratio of the instantaneous flow rate at 75% expiratory volume to the instantaneous flow rate at 75% inspiratory volume, the ratio of the instantaneous flow rate at 25% expiratory volume to the highest expiratory flow rate, the ratio of the instantaneous flow rate at 25% expiratory volume to the instantaneous flow rate at 75% expiratory volume, and the ratio of the curvature of the curve to the right of the highest point of expiratory flow rate on the flow-volume loop curve to the average curvature of the post-expiratory phase segment.

[0037] Those skilled in the art will understand that the above-described 11 types of higher-order derived coupling features are merely a preferred embodiment of this application, and those skilled in the art can add or subtract features based on actual needs. The following describes these 11 types of higher-order derived coupling features.

[0038] Regarding the ratio of the area enclosed by the flow-volume ring to its circumference: In asthmatic patients with expiratory limitation, the F–V (flow-volume) ring often exhibits localized depressions or an early decrease in flow, leading to a reduction in area while the circumference decreases disproportionately, thus decreasing the area-to-circumference ratio. Therefore, the ratio of the area enclosed by the flow-volume ring to its circumference can be used to reflect the "fatness" and regularity of the curve. A healthy tidal ring is usually regularly shaped, with its area and circumference in a certain proportion; airway obstruction or uncoordinated breathing can cause the ring to become irregular, exhibiting depressions or increased boundary complexity, thereby altering the area-to-circumference ratio.

[0039] Regarding the ratio of peak expiratory flow rate to tidal volume in the flow-volume loop: In asthmatic patients, airway obstruction restricts expiratory flow, leading to a significant decrease in PTEF (Total Physical Expiratory Flow Rate), while tidal volume (VT) remains relatively stable due to compensatory deepening or prolongation of expiratory time. This results in a significant decrease in the expiratory height-to-width ratio (peak expiratory flow rate to tidal volume). Therefore, the ratio of peak expiratory flow rate to tidal volume in the flow-volume loop can be used to reflect the maximum exhaust efficiency per unit tidal volume. Under normal circumstances, the expiratory height-to-width ratio remains within a relatively narrow range.

[0040] Regarding the ratio of expiratory area to inspiratory area: Asthma often presents with expiratory limitation, a decreased expiratory area ratio (expiratory / inspiratory < normal), or a relatively increased inspiratory area ratio. Therefore, the ratio of expiratory area to inspiratory area can be used to reflect the balance of the two-phase respiratory volume and flow distribution. In normal individuals, the inspiratory and expiratory areas are nearly balanced; airway obstruction can lead to a relative decrease in expiratory area or an increase in inspiratory compensation.

[0041] The ratio of the area of ​​the flow-volume (F–V) loop within the 50%–100% post-expiratory volume range to the total expiratory area: Asthmatic patients typically exhibit a concave curve at end-expiratory time, with an earlier peak expiratory time. This ratio reflects the characteristics of changes in the post-expiratory phase. A reduction in airflow at end-expiratory time significantly alters this ratio. Therefore, the ratio of the flow-volume loop area within the 50%–100% post-expiratory volume range to the total expiratory area is a sensitive indicator of small airway obstruction.

[0042] Regarding the ratio of peak expiratory flow rate to peak inspiratory flow rate (PTEF / PTIF): When the airway is obstructed, PTE decreases, leading to a decrease in the PTE / PTIF value. Therefore, the ratio of peak expiratory flow rate to peak inspiratory flow rate can be used to directly measure the symmetry of inspiratory and expiratory flow rates and airway patency. A decrease in peak expiratory flow rate can indicate expiratory limitation.

[0043] Regarding the ratio of instantaneous flow rate at 25% expiratory volume to total expiratory volume (TEF25 / VE): airway obstruction can reduce expiratory flow rate early, leading to a decrease in TEF25 / VE, and is sensitive to mild obstruction. Therefore, the ratio of instantaneous flow rate at 25% expiratory volume to total expiratory volume can be used to reflect the relationship between flow rate and volume in the early to low flow rate phase of expiration, serving as an indicator of local flow-volume efficiency.

[0044] The ratio of instantaneous flow rate at 50% expiratory volume to instantaneous flow rate at 50% inspiratory volume (TEF50 / TIF50): Mid-expiratory limitation is common in asthma, and the TEF50 / TIF50 ratio is usually decreased, serving as an indicator of mid-expiratory obstruction. Therefore, the ratio of instantaneous flow rate at 50% expiratory volume to instantaneous flow rate at 50% inspiratory volume can be used to measure the symmetry of inspiratory / expiratory flow rates at the midpoint of breathing, and after reducing inter-individual tidal volume differences, it better reflects airway dynamic differences.

[0045] The ratio of instantaneous flow rate at 75% expiratory volume to instantaneous flow rate at 75% inspiratory volume (TEF75 / TIF75): TEF75 / TIF75 decreases significantly in cases of substantial airway obstruction or small airway involvement. Its sensitivity for mild lesions is lower than TEF25 / VE, but it is indicative of severe involvement. Therefore, the ratio of instantaneous flow rate at 75% expiratory volume to instantaneous flow rate at 75% inspiratory volume can reflect the flow rate comparison between end-expiratory (near residual volume) and end-expiratory phases. Significant restriction at end-expiratory phase can indicate severe small airway obstruction or dynamic compression.

[0046] For the ratio of instantaneous flow rate at 25% expiratory volume to peak expiratory flow rate (TEF25 / PEF): Airway obstruction significantly restricts exhalation early on, resulting in a lower TEF25 / PEF ratio, reflecting an increased rate of flow rate decline. Therefore, the ratio of instantaneous flow rate at 25% expiratory volume to peak expiratory flow rate can be used to relatively measure the proportion of peak flow rate in early exhalation, describing whether the flow rate declines rapidly in early exhalation.

[0047] The ratio of instantaneous flow rate at 25% expiratory volume to that at 75% expiratory volume (TEF25 / TEF75): In asthma, small airway collapse or obstruction leads to a more significant decrease in flow rate during late expiration, increasing the TEF25 / TEF75 ratio. Therefore, the ratio of instantaneous flow rate at 25% expiratory volume to that at 75% expiratory volume can be used to measure the flow rate change between the two phases of expiration; under normal circumstances, the flow rates in the two phases are approximately equal.

[0048] The ratio of the curvature of the curve to the right of the point of highest expiratory flow on the flow-volume loop to the average curvature of the post-expiratory phase: This ratio can amplify the local abnormality of a sharp indentation in the post-expiratory phase, improving the specificity for asthma identification. The decrease in flow rate in the later expiratory phase causes the post-expiratory phase to tilt significantly towards the horizontal axis, even showing indentation. Therefore, the ratio of the curvature of the curve to the right of the point of highest expiratory flow on the flow-volume loop to the average curvature of the post-expiratory phase can be used to reflect the local unevenness of the post-expiratory phase.

[0049] Based on the aforementioned higher-order derived coupling features, it is evident that these ratio features possess inherent normalization properties, effectively eliminating the interference of individual differences in age, body size, and lung capacity on absolute quantitative indicators, thereby enhancing cross-individual and cross-device stability and generalization ability. Their construction process is independent of traditional clinical parameter systems, thus introducing complementary and non-redundant information dimensions into the model. Through cross-dimensional construction, proportional expression, and enhanced structural sensitivity, higher-order derived coupling features enable this application to capture more discriminative airflow dynamics patterns.

[0050] Curve morphology features refer to the structural information of the curve's shape, symmetry, angle changes, local slope, and overall configuration extracted by geometric fitting and modeling the complete two-dimensional flow-volume loop during tidal breathing.

[0051] In one embodiment, considering that asthma lesions mostly occur during the expiratory phase, to enhance the analysis of expiratory features, this embodiment fits the tidal breathing loop into an expiratory pentagon composed of five key feature points. This construction aims to focus on characterizing the geometric morphology of the lesions during the expiratory phase, thereby compensating for the inadequacy of traditional test parameters in representing geometric features and improving the completeness of the tidal breathing feature system.

[0052] Specifically, methods for obtaining curve shape features may include:

[0053] Step S100: Perform geometric fitting modeling of the complete two-dimensional flow-volume loop during tidal breathing, and extract the shape of the curve; please refer to... Figure 2 Take the origin A of the curve as the origin of the coordinate system, and take the maximum value of the vertical coordinate of the breathing stage curve as the highest point C of the curve.

[0054] Step S200: Using the highest point as the fulcrum, the curve is divided into the left and right sides of the highest breathing point. Two-dimensional straight line fitting is performed on the curves of the left and right sides respectively to obtain the optimal point B on the left and the optimal point D on the right. Connect the origin A to the optimal point B on the left, the optimal point B on the left to the highest point C, the highest point C to the optimal point D on the right, the optimal point D on the right to the projection E of point D on the x-axis, and point E to point A, thus obtaining the pentagonal feature of the tidal breathing curve.

[0055] In one embodiment, a two-dimensional straight line fitting is performed on the curve on the left to obtain the optimal point B on the left, including:

[0056] Step S211: Using the left side of the highest point as the index range, traverse each point on the curve as a candidate point.

[0057] In one specific embodiment, the x-coordinates of the exhalation phase curves are first arranged in ascending order. Then, spline interpolation is used to generate a high-density interpolated curve. The curve coordinates can be defined as (xnew, ynew), and the index of the highest point is highest_idx. Each point on the curve is traversed within the index range [5, highest_idx−5) to the left of the highest point as a candidate point. In one embodiment, the points on the traversed curves avoid both ends to reduce boundary effects.

[0058] Step S212: For any candidate point among the candidate points, calculate the slope of the straight line from the origin to the candidate point to obtain the first interval straight line equation, calculate the slope of the straight line from the candidate point to the highest point to obtain the second interval straight line equation, and thus obtain the first interval straight line equation and the second interval straight line equation for each candidate point.

[0059] Step S213: For any candidate point among the candidate points, predict the data points on the curve between the origin and the candidate point point by point based on the first interval straight line equation to obtain the first predicted value of each data point. Predict the data points on the curve between the highest point and the candidate point point by point based on the second interval straight line equation to obtain the second predicted value of each data point. Calculate the first root mean square error based on all the first predicted values ​​and the second root mean square error based on all the second predicted values. Calculate the weighted linear combination error based on the first root mean square error and the second root mean square error to obtain the weighted linear combination error of each candidate point.

[0060] In one embodiment, the weighted linear combination error calculated based on the first root mean square error and the second root mean square error can be expressed as:

[0061] ,

[0062] in, This represents the weighted linear combination error of any candidate point. and These represent the first root mean square error and the second root mean square error, respectively. and Indicates weight, The weights can be adjusted according to the curve fit.

[0063] Step S214: Select the candidate point with the smallest weighted linear combination error as the optimal point B on the left.

[0064] In one embodiment, performing a two-dimensional straight line fitting on the curve on the right to obtain the optimal point D on the right may include: extracting data points on the curve on the right starting from the highest point C of the curve, fitting a linear function using the least squares method to obtain the slope and intercept representing the trend at the end of respiration, and obtaining a fitted straight line; inputting the rightmost x-coordinate of the curve to the fitted linear function to obtain the extrapolated position of the fitted straight line on the rightmost side of the curve, and taking this extrapolated position as the optimal point D on the right.

[0065] By connecting the five points A, B, C, D, and E obtained above in sequence, we can obtain the pentagonal characteristic of the tidal breathing curve.

[0066] Step S300: Based on the above pentagonal characteristics, calculate the distance to BC, the angle of ∠BCD, the angle of ∠CAE, the angle of ∠ABD, the angle of ∠ACD, the slope of line CD, the root mean square error of line CD and the curve to the right of point C, the mean absolute error of line CD and the curve to the right of point C, the area of ​​the pentagon, and the ratio of the area to the perimeter of the pentagon.

[0067] Those skilled in the art will understand that the above-described 10 types of curve morphological features are merely a preferred embodiment of this application, and may be added to or subtracted from them based on actual needs. The following describes these 10 types of curve morphological features.

[0068] Based on the aforementioned pentagonal characteristics, the BC distance represents the Euclidean distance between point B (the optimal segmentation point on the left) and point C (the peak of expiration), representing the "slope length" of the later stage of the rising phase of the curve. In asthmatic children, due to increased airway resistance, the flow velocity in the rising phase of expiration increases rapidly, leading to an earlier peak and a shorter BC distance. In healthy infants and young children, the expiratory rise is gentler, resembling an approximate arc, with the peak appearing later, resulting in a longer BC distance. Therefore, the BC distance can sensitively capture patterns of "changes in the peak rise slope" or "gentle pre-peak flattening," helping to identify pathological signals of expiratory initiation or early restriction.

[0069] The angle ∠BCD represents the angle formed by points B (the best-fit point on the left), C (the peak point), and D (the rightmost fitted point), reflecting the sharpness of the expiratory curve at its peak (peak sharpness). In normal breathing, the expiratory curve is approximately elliptical, with expiratory flow rising and falling slowly, forming near-symmetry. In asthmatic patients, expiratory flow rises rapidly and falls slowly, the angle decreases, forming a sharp peak. Therefore, the angle ∠BCD can measure the sharpness of the expiratory peak after airflow limitation, effectively identifying the characteristic morphology of the expiratory segment.

[0070] The angle ∠CAE represents the angle formed by points A (origin), C (peak point), and E (right-end projection point), reflecting the inclination of the overall expiratory curve relative to the horizontal axis. A smaller angle indicates a faster return to zero flow during expiration (full expiration) and a more symmetrical expiratory curve, indicative of normal breathing. A larger angle indicates an earlier peak expiratory volume and slower residual volume expiration, commonly seen in asthma patients. Therefore, the angle ∠CAE sensitively reflects expiratory delay and residual volume retention, serving as a global indicator of decreased ventilation efficiency.

[0071] The angle ∠ABD is formed by points A (origin), B (best-fit point on the left), and D (rightmost fitted point). It reflects the degree of geometric deflection between the ascending and descending directions of exhalation and is highly sensitive to increased airway resistance, premature closure of small airways, and a sharp decrease in flow velocity at the end of the exhalation. Therefore, the angle ∠ABD can reflect the overall opening shape and asymmetry of the exhalation process.

[0072] The angle ∠ACD represents the angle formed by points A (origin), C (peak point), and D (rightmost fitted point). It measures the deviation of the downward direction of the curve after the expiratory peak from the overall starting direction of exhalation. It can sensitively capture the rapid decay phenomenon after the peak caused by pathological mechanisms such as small airway obstruction, dynamic airway collapse, and unstable flow rate in the late expiratory phase. The angle decreases significantly during rapid decay in the later stages of exhalation; therefore, the angle ∠ACD is more sensitive to changes in the morphology of the later part of the curve and can serve as a supplementary angle.

[0073] The slope of the straight line CD represents the slope of the line from the peak point C to the extrapolated point D on the right, characterizing the rate of decrease in flow rate at the end of expiration. A high slope indicates that the flow rate changes rapidly with volume (rapid decrease in the post-expiratory phase), while a low slope or close to 0 indicates that the flow rate in the later stages tends to level off or decrease slowly. Asthmatic patients experience a more rapid decrease in flow rate at the end of expiration (the absolute value of the slope is larger), and the more severe the obstruction, the larger the slope. Healthy individuals have a generally symmetrical elliptical expiratory curve with a smaller slope. Therefore, the slope of the straight line CD can be used as a direct kinetic quantity to quantify the dynamic characteristics of the post-expiratory phase. It is a direct geometric indicator for assessing airway resistance / compliance abnormalities and has good discriminative power in differentiating asthma grades (mild / moderate / severe).

[0074] The root mean square error (RMSE) of the curve to the right of point C on line CD represents the root mean square error of the fit between the curve to the right of point C and line CD. It reflects the degree of deviation of the curve fluctuation in the right-side descending segment from the linear fit. It is very sensitive to instantaneous large fluctuations in the curve (such as artifacts from coughing or swallowing). Asthmatic patients often exhibit irregular fluctuations and expiratory instability in their curves, leading to larger errors. Therefore, the RMSE of the curve to the right of point C on line CD can characterize the instability of airflow at the end of expiration; a larger RMSE indicates poor airway compliance and severe airway restriction.

[0075] The mean absolute error (MAE) of the straight line CD and the curve to the right of point C is similar to the root mean square (RMS) error, but it is less sensitive to abnormal fluctuations and reflects the overall trend deviation. It reflects the average deviation level of the curve as a whole and is not sensitive to large error points, thus making it more robust. Therefore, the MAE can serve as a robust supplement to the RMS error, improving the algorithm's noise resistance and robustness to abnormal patterns. The combined use of RMS and MAE can simultaneously reflect the magnitude of the deviation and outlier stability.

[0076] The area of ​​the pentagon is calculated by taking the vertices of the polygon in the order of (A–B–C–D–E). It represents the overall comprehensive value of the pentagon fitted during the exhalation phase and is less affected by small-scale curve fluctuations. A decrease in area or a change in shape suggests a decline in overall expiratory capacity or curve distortion.

[0077] In the area-to-perimeter ratio of a pentagon, the perimeter is the sum of the lengths of all sides. The area / perimeter ratio measures compactness. A more elongated or irregular exhalation curve will affect this ratio. It is sensitive to changes in overall shape and can be used as a supplementary indicator.

[0078] Based on the aforementioned curve morphology characteristics, it is evident that, compared to single-point features such as peak flow rate and time ratio relied upon in traditional clinical practice, geometric features directly affect the continuous morphology of the original respiratory curve. They can capture local abnormalities and structural deformations caused by pathological factors such as increased small airway resistance, dynamic airway collapse, and end-expiratory restriction. Especially in the low-flow, low-cooperation testing context of infant tidal breathing, pathological changes often do not directly manifest as a decrease in absolute peak value, but rather as increased curvature, enhanced irregularity, or morphological shifts near the peak in the later expiratory phase. Therefore, compared to traditional indicators, geometric features offer higher pathological sensitivity and earlier identification capabilities.

[0079] Furthermore, morphological features offer the advantage of strong comparability across devices and individuals, independent of respiratory volume or body size, and can achieve highly stable structural comparisons after scale-space fitting and registration. In the embodiments of this application, by constructing multiple geometric metrics, a highly interpretable, low-dimensional, and physically meaningful feature family is formed, providing structured, non-redundant, and clinically interpretable high-value input for subsequent geometric feature modeling and three-branch decision-making, thereby improving the accuracy of triadic classification.

[0080] In real-world clinical settings, infants and young children exhibit significant physiological heterogeneity in tidal breathing signals. This heterogeneity is not solely determined by respiratory mechanics but is also systematically influenced by demographic factors such as age, height, weight, and sex. Lung structure and airway compliance change rapidly with age, and the same disease state may present with different tidal breathing patterns at different age levels. Height and weight directly affect thoracic biomechanics and ventilation, resulting in inherent shifts in peak expiratory volume, tidal volume, and time-proportion characteristics across individuals of different body types. Sexual differences also lead to baseline differences in lung capacity, flow patterns, and airway physiology. Without explicit modeling of these factors, classification models will inevitably misinterpret normal physiological differences as pathological changes, thereby reducing sensitivity and specificity. Furthermore, demographic variables are independent covariates with low noise and high reliability, and can be used as cross-individual calibration factors to improve the comparability of characteristics. Therefore, the demographic data in this embodiment include sex, age, height, and weight.

[0081] The applicant discovered in their research that the four types of features—traditional tidal breathing test data, higher-order derived coupling features, curve morphology features, and demographic data—exhibit significant differences in numerical range, statistical distribution, scale, and physiological generation mechanisms. If a single model is used to jointly model all features, fitting must be performed under uniform data transformation and distribution assumptions, making it difficult to simultaneously consider the statistical characteristics of different feature categories. This could lead to increased model bias, decreased fitting quality for key features, and weakened overall interpretability. Furthermore, due to the limited sample size of infant respiratory data, it is difficult to collect an effective dataset sufficient to complete a complex model. Using a single model would require extensive data training to prevent overfitting. Therefore, directly modeling all features using a single model is unlikely to achieve stable and reliable classification performance under practical conditions.

[0082] The applicant also discovered in their research that, in order to fuse multiple grouped features into a single output, using direct averaging / weighted averaging of sub-model probabilities can mix uncalibrated probabilities together, generating erroneous evidence, because the probability scales and calibration states of different grouped features are typically different. Large probability estimation biases will directly propagate to the final result; using a fusion model (such as XGBoost or a neural network model) reduces interpretability and is prone to overfitting. Therefore, to ensure model robustness, lightweight design, and interpretability, the specific embodiments of this application use the log-likelihood ratio (LLR) as the fusion criterion.

[0083] Therefore, in this embodiment of the application, the log-likelihood ratio of each group is obtained by grouping. This allows different probability models to be used for different feature types, and the distribution assumptions that are closer to the actual data can produce more accurate and stable parameter estimates (lower model bias and variance).

[0084] Those skilled in the art will understand that when constructing a probability model, a group modeling approach is also used to obtain the sub-models corresponding to each group for obtaining the log-likelihood ratio.

[0085] In one embodiment, please refer to Figure 3 Methods for extracting heterogeneous features and obtaining grouped log-likelihood ratios from preprocessed tidal breathing data and demographic data may include:

[0086] Step S1000: Perform Gaussian discriminant analysis on the traditional tidal breathing test data to calculate the log-likelihood ratio and obtain the first log-likelihood ratio.

[0087] Traditional tidal breathing features exhibit stable, unimodal, and approximately Gaussian statistical structures, with linearly modelable correlations. Gaussian discriminant analysis (GDA) can simultaneously model the mean difference and covariance structure of features, possessing stronger discriminative power than Naive Bayes and higher interpretability and better small-sample robustness than more complex machine learning models. Furthermore, its discriminant function can directly generate the log-likelihood ratio (LLR), naturally compatible with the applied hierarchical Bayesian fusion mechanism. Therefore, GDA is the most suitable modeling method for these traditional parameters.

[0088] Specifically, step S1000 may include: performing distribution analysis and numerical preprocessing on the traditional tidal breathing test data features (15 continuous features), including: performing Box-Cox transformation on skewed or boundary features to improve their approximate normality; uniformly applying Z-score standardization to all continuous features to ensure balanced weights for each dimension in covariance matrix estimation; and suppressing the influence of extreme points on parameter estimation by threshold pruning based on median absolute deviation (MAD) for outliers. The preprocessed feature set is considered as a continuous vector (15 dimensions).

[0089] During model training, if the number of asthma samples in the training set is... The number of healthy samples is The total number of samples n is Then the class prior estimate can be expressed as:

[0090] ,

[0091] ,

[0092] in, Represents the class prior estimate of asthma samples. This represents the class prior estimate of a healthy sample.

[0093] Assuming that the feature vector of each class follows a Gaussian distribution, estimate the multivariate Gaussian distribution parameters, i.e., the mean vector, for samples from the asthma group and the healthy group, respectively. Covariance Matrix (in After obtaining the two parameters, the probability of k can be obtained. Since the covariance structures of the two groups of samples are similar, a linear discriminant model (LDA) is constructed using a shared covariance matrix.

[0094] For each sample x, calculate the log-likelihood ratio (LLR) as the diagnostic output for that feature group. Then:

[0095] ,

[0096] in, Represents the log-likelihood ratio for any sample x; Let x represent the probability that any sample x has asthma. Let represent the probability that any sample x is healthy; m and b are both intermediate variables. , , This represents the covariance matrix shared by the two classes. for The inverse matrix; T denotes the transpose.

[0097] In one embodiment, to ensure probabilistic comparability of the model when fusing different feature groups at different levels, the GDA output is further subjected to Plattscaling to obtain a first log-likelihood ratio, thereby optimizing the matching degree between the output probability and the true diagnostic result.

[0098] Step S2000: The log-likelihood ratio of the higher-order derived coupling features is calculated using a class-conditional model that combines a class-conditional model with a multivariate Student-t distribution, to obtain the second log-likelihood ratio.

[0099] To address the numerical characteristics of higher-order derived coupling feature groups, this application embodiment combines a class-conditional multivariate Student-t model with a multivariate Student-t distribution to uniformly model the class-conditional distribution of higher-order derived coupling features when constructing and solving the sub-model corresponding to the second log-likelihood ratio. This class-conditional multivariate Student-t model can simultaneously handle the mixed value ranges, potential strong correlations, and heavy-tailed or outlier fluctuations of higher-order derived coupling features under limited sample conditions, thus achieving the best balance between statistical expressiveness and robustness. Compared with traditional methods such as Naive Bayes, univariate Laplace models, or Gaussian models, multivariate Student-t can characterize the joint structure between features through the covariance matrix, avoiding evidence duplication or weight bias caused by the independence assumption. Simultaneously, by shrinking the degree-of-freedom parameter and covariance, parameter discriminability and estimation stability can be maintained even with only a few hundred cases per class. Therefore, this model is better suited than other common statistical models to capture the overall pattern differences in proportional features between asthmatic and healthy infants and can significantly improve the reliability and discriminative performance of subsequent LLR calculations and hierarchical Bayesian fusion.

[0100] However, the applicant found in the research that, due to the different theoretical domains of higher-order derived coupling features, it is difficult to guarantee the availability of subsequent multivariate Student-t models.

[0101] Therefore, in this embodiment of the application, before obtaining the second log-likelihood ratio based on the categorical conditional multivariate Student-t model, the mathematical properties of each feature in the higher-order derived coupling features are first subjected to a monotonically invertible transformation, so that all features are mapped to the real number domain (-∞, +∞), and then input into the categorical conditional multivariate Student-t model.

[0102] In one embodiment, a method for performing a monotonically invertible transformation on the mathematical properties of individual features in a higher-order derived coupling feature may include:

[0103] For features located in (0, 1), the logit transformation is used; for features in other ranges, the natural logarithm transformation is used, which can be expressed as:

[0104] ,

[0105] in, Let j represent any higher-order derived coupling feature. Let represent the monotonically invertible transformation result of any higher-order derived coupling feature j. This refers to the logit function. This represents a very small constant that avoids being zero, and can be taken as 10⁻⁶.

[0106] Z-score standardization is performed on each result after the monotonically invertible transformation. The transformations of these two processes are completely reversible and do not destroy the discriminative information of the original features. The statistical distribution of the processed features across each category (asthma / healthy) is processed using a multivariate Student-t distribution to capture heavy-tailed behavior, inter-feature correlations, and multidimensional structural differences, obtaining the class conditional density. The second log-likelihood ratio is then calculated based on the obtained class conditional density, which can be expressed as:

[0107] ,

[0108] in, Let represent the log-likelihood ratio for any sample z, i.e., the second log-likelihood ratio; Let z represent the probability that any sample z has asthma. This represents the probability that any sample z is healthy.

[0109] In step S3000, the curve morphology features are used as input to the Class-conditional Principal Component Analysis (PPCA) model to calculate the log-likelihood ratio and obtain the third log-likelihood ratio.

[0110] As one embodiment of this application, a Class-conditional Probability-Based Analysis (PPCA) model is introduced to probabilistically discriminate highly correlated geometric features between the asthma group and the healthy group. This method effectively handles the strong correlations and potential low-dimensional structures commonly found in tidal breathing curves while preserving key structural information of the geometric features. Compared to traditional Gaussian models, PPCA explicitly introduces latent variables to perform low-rank modeling of features, avoiding singular or ill-conditioned covariance matrices due to high collinearity, thus significantly improving the stability of density estimation and log-likelihood ratio calculation. Unlike general PCA, PPCA outputs explicit class-conditional probability densities within a probabilistic framework, making it more suitable for subsequent LLR fusion and three-branch decision systems, and appropriate for clinical data with limited sample sizes.

[0111] In one embodiment of this application, the 10 types of curve morphology features need to be preprocessed before being input into the conditional probability principal component analysis (PPCA) model. Since the root mean square error of the curve to the right of line CD and point C, and the mean absolute error of the curve to the right of line CD and point C (representing the statistical characteristics of curve fitting error) are two types of error features, they typically exhibit significant skewness, heavy-tailed characteristics, and strong anomaly sensitivity. If they are directly standardized with geometric variables such as distance, angle, and slope, it will cause problems such as dimensional imbalance, covariance matrix estimation bias, and improper dominance of principal component directions by error features. Ultimately, this will cause the latent subspace of the PPCA model to deviate from the true geometric structure, resulting in a weight imbalance in the obtained log-likelihood ratio (LLR). Therefore, before inputting these two types of features into the PPCA model, logarithmic transformation and robust standardization are required to ensure that the error features participate in the conditional density estimation in a stable, symmetrical, and controllable manner within the probability model. Simultaneously, the other 8 types of features out of the 10 types are used to form a sample feature vector, and z-score standardization is performed on each dimension to eliminate the influence of dimensions. By concatenating the preprocessed dimensions in order, we can obtain a transformed input vector with a dimension of 10. This vector is then input into a principal component analysis model with a conditional probability class to obtain the third log-likelihood ratio.

[0112] Step S4000: Use demographic data features as input to the Bayesian logistic regression model, calculate the log-likelihood ratio, and obtain the fourth log-likelihood ratio.

[0113] Because demographic data features (age, height, weight, gender) have low dimensionality, heterogeneous distribution, and their relationship with disease risk is usually non-linear or weakly correlated, directly using traditional linear discriminant analysis or Gaussian discriminant analysis will be limited by distribution assumptions and will be difficult to accurately characterize the classification boundary; while Naive Bayes relies on the feature independence assumption and is not suitable for variables with inherent correlations (such as height and weight).

[0114] Therefore, in one embodiment of this application, a Bayesian logistic regression (BLR) model is introduced in the modeling section of demographic data features to capture the probabilistic influence of individual basic characteristics on the risk of infant asthma. BLR can naturally handle binary and continuous variables within the discriminative framework of logistic regression, and by applying a Gaussian prior to the parameters, it quantifies and automatically shrinks the uncertainty of coefficients in small sample situations, effectively suppressing overfitting and improving the stability of the probability output. Compared with non-probabilistic models such as support vector machines and decision trees, BLR can directly output the posterior probability of the class, seamlessly integrating with the log-likelihood ratio fusion and three-branch decision mechanism of this invention. Therefore, choosing BLR as the classification model for demographic features has comprehensive advantages such as theoretical feasibility, numerical stability, and high compatibility with the overall architecture.

[0115] By using grouped modeling, each sub-model estimates a smaller set of parameters, avoiding direct estimation of high-dimensional nonparametric densities. This makes the model more robust and computationally efficient with limited samples. Simultaneously, each group of models corresponds to a clear physiological or morphological meaning (e.g., "geometric group → abnormal curve morphology"), making the output more interpretable and easier for doctors to understand and verify. If new features (such as biomarkers or other characteristics) are added in the future, only sub-modules need to be added and interfaces provided; no overall retraining is required, facilitating long-term maintenance and clinical iteration.

[0116] The log-likelihood ratio (LLR) is essentially a logarithmic measure of "category evidence," a statistic used in Bayesian decision-making. It directly characterizes which category is more likely to generate observational data and how strong that evidence is. While the outputs of different sub-models may differ in form, they can all be converted to LLR, thus providing a unified interface for evidence across heterogeneous models. The product of the sub-likelihoods calculated from multiple sub-models is the joint likelihood; taking the logarithm of this yields the sum of the sub-LLRs. This additive form is highly suitable for linear weighting, interpretation, and three-branch decision-making. Furthermore, the sign and magnitude of the LLR intuitively indicate the strength of the evidence supporting or opposing asthma, allowing for a group-by-group demonstration of each sub-model's contribution to the final judgment, facilitating physician review and tracing.

[0117] Step S30: Perform hierarchical Bayesian fusion on each log-likelihood ratio to obtain the fusion result.

[0118] Multi-source feature fusion typically assumes that each feature group is independent and uses a Naive Bayes framework to directly fuse the log-likelihood ratios (LLRs) of each group. However, the applicant discovered in its research that in the embodiments of this application, both higher-order coupled features and traditional tidal breathing data features use flow rate and volume ratio values, resulting in a mathematical correlation between the two sets of features, violating the conditional independence assumption of Naive Bayes. If the direct fusion method is still used, the information contained in the higher-order coupled features will be counted repeatedly, leading to model overfitting, decreased generalization ability, and artificially inflated diagnostic confidence after fusion, affecting the reliability of clinical applications.

[0119] In view of this, this application proposes a hierarchical Bayesian fusion method to obtain the fusion result. In one embodiment, please refer to... Figure 4 ,include:

[0120] Step S301: Use a linear Bayesian discriminant structure to perform hierarchical fusion of each log-likelihood ratio to obtain the log-posterior ratio.

[0121] In step S301, a linear Bayesian discriminant structure is used for hierarchical fusion in the fusion layer, mapping each log-likelihood ratio of the input to the final log-posterior ratio, which can be expressed as:

[0122] ,

[0123] in, Represents the log-posterior ratio vector. This represents the log-likelihood ratio vector formed by the individual log-likelihood ratios; represents the fusion weight vector, which represents the evidence contribution of each log-likelihood ratio calculation model; T represents the transpose; q represents the fusion bias, which is used to adjust the global prior ratio.

[0124] In one embodiment, to avoid overfitting of the weights with a limited number of samples, the present invention sets an isotropic Gaussian prior for the fusion weights.

[0125] The introduction of weight vectors is equivalent to introducing prior confidence coefficients into the fusion layer, performing Bayesian redistribution of diagnostic confidence from features from different sources. This ensures that the contribution of each sub-model is automatically adjusted according to its actual discriminative ability, without relying on manual experience. This mechanism allows the model to maintain a stable posterior probability output even when sample information is incomplete.

[0126] Step S302: Based on the sigmoid function, the log-posterior ratio is processed to obtain the posterior probability.

[0127] Step S303: Use the posterior probability as the fusion result.

[0128] Through the above processing, unified reasoning between different statistical models and different feature types is achieved, so that the probability of the final output has a clear Bayesian physical meaning and can be directly used as the input of the subsequent three-branch decision mechanism.

[0129] Step S40: Based on the fusion results, perform a three-branch classification to assist in determining asthma risk. The three-branch classification results include high asthma risk, low asthma risk, and an indeterminate range.

[0130] After obtaining the fusion results, to achieve a three-way decision-making mechanism that conforms to medical risk control, a threshold adaptive learning method based on the medical loss matrix is ​​further introduced. This method automatically determines the optimal first decision threshold α and second decision threshold β by minimizing the expected medical loss on the validation dataset. The first decision threshold is used to determine whether the asthma risk is high-risk, and the second decision threshold is used to determine whether the asthma risk is low-risk. The first decision threshold is greater than the second decision threshold, thereby achieving an optimal balance between the missed diagnosis rate, the misdiagnosis rate, and the further examination rate.

[0131] The three-branch classification of asthma risk is used to assist in the assessment of asthma risk based on the fusion results. This includes: determining the relationship between the fusion results and preset assessment thresholds, which include a first assessment threshold and a second assessment threshold, with the first assessment threshold being greater than the second assessment threshold; if the fusion results are greater than or equal to the first assessment threshold, the asthma risk is determined to be high; if the fusion results are less than or equal to the second assessment threshold, the asthma risk is determined to be low; if the fusion results are less than the first assessment threshold but greater than the second assessment threshold, the asthma risk is determined to be in the uncertain region, requiring further examination and assessment.

[0132] To obtain the aforementioned first and second decision thresholds, correspondingly, during model training, medical loss values ​​for different diagnostic results are set based on clinical risk preferences, including: the first loss value CFN for a person who is actually ill but is judged to be healthy (missed diagnosis), the second loss value CFP for a person who is actually healthy but is judged to be ill (misdiagnosis), and the third loss value CFT for a person who is judged to be "needing further examination" (uncertain decision).

[0133] The above-mentioned loss values ​​can be determined based on the experience of medical experts, economic costs, or clinical safety requirements, satisfying the condition CFN > CFP > CFT.

[0134] To make the diagnostic strategy optimal in a medical sense, embodiments of this application automatically learn α and β by minimizing the expected loss of the validation set samples.

[0135] Based on the three-branch classification auxiliary judgment method for infant asthma provided in the above embodiments, since it extracts heterogeneous features and obtains grouped log-likelihood ratios from preprocessed tidal breathing data and demographic data, it obtains the log-likelihood ratios corresponding to each group. This allows for the application of different probability models suitable for different feature types, and the use of distribution assumptions that are more closely aligned with the actual data can produce more accurate and stable parameter estimates. Furthermore, since it performs hierarchical Bayesian fusion on the various log-likelihood ratios to obtain the fusion result, it enables unified inference between different statistical models and different feature types, giving the final output probability a clear Bayesian physical meaning, which can be directly used for subsequent three-branch decision-making. The input to the mechanism is as follows: Because the three-branch classification results include high asthma risk, low asthma risk, and an uncertain interval, a three-branch decision-making mechanism that better aligns with medical risk control can be obtained. The extracted heterogeneous features include higher-order derived coupling features, enabling the simultaneous reflection of the relative intensity, synchronicity, and proportional relationships between respiratory physiological processes, exhibiting higher sensitivity in the early stages of subtle pathological changes, thus revealing potential structural information that cannot be expressed by a single original feature. Furthermore, the extracted heterogeneous features include curve morphology features, allowing for a focused characterization of the lesion geometry during the expiratory phase, thereby compensating for the insufficient representation of geometric features by traditional test parameters and enhancing the completeness of the tidal breathing feature system. Therefore, the proposed solution can more objectively and accurately assist in the diagnosis of infantile asthma based on tidal breathing data.

[0136] One embodiment of this application provides a computer-readable storage medium storing a program, the stored program including methods that can be loaded by a processor and processed in any of the above embodiments.

[0137] Those skilled in the art will understand that all or part of the functions of the various methods in the above embodiments can be implemented by hardware or by computer programs. When all or part of the functions in the above embodiments are implemented by computer programs, the program can be stored in a computer-readable storage medium, which may include: read-only memory, random access memory, disk, optical disk, hard disk, etc., and the program is executed by a computer to achieve the above functions. For example, the program can be stored in the memory of a device, and when the program in the memory is executed by the processor, all or part of the above functions can be achieved. In addition, when all or part of the functions in the above embodiments are implemented by computer programs, the program can also be stored in a server, another computer, disk, optical disk, flash drive, or external hard drive, etc., and can be downloaded or copied to the memory of a local device, or the system of the local device can be updated. When the program in the memory is executed by the processor, all or part of the functions in the above embodiments can be achieved.

[0138] The above examples illustrate the present invention only to aid in understanding it and are not intended to limit the scope of the invention. Those skilled in the art can make various simple deductions, modifications, or substitutions based on the principles of this invention.

Claims

1. A method for auxiliary diagnosis of the three-branch classification of infantile asthma, characterized in that, include: Data collection and preprocessing of tidal breathing and demographic data of infants and young children; The collected tidal breathing data includes tidal breathing test data and tidal breathing loops, and the collected demographic data includes gender, age, height, and weight. Heterogeneous features were extracted and grouped log-likelihood ratios were obtained from the preprocessed tidal breathing data and demographic data to obtain the log-likelihood ratios for each group. The extracted heterogeneous features included traditional tidal breathing test data features, higher-order derived coupling features, curve morphology features, and demographic data features. The higher-order derived coupling features refer to the ratio-type respiratory dynamics index formed by cross-domain combination and nonlinear transformation of multiple basic dimensions in the tidal breathing signal. The curve morphology features refer to the structural information of the curve's shape, symmetry, angle changes, local slope, and overall configuration extracted by geometric fitting and modeling the complete two-dimensional flow-volume loop during tidal breathing. Hierarchical Bayesian fusion is performed on each log-likelihood ratio to obtain the fusion result; Asthma risk is assessed using a three-branch classification based on the fusion results; the three-branch classification results include high asthma risk, low asthma risk, and an indeterminate range.

2. The method for auxiliary judgment of the three-branch classification of infant asthma as described in claim 1, characterized in that, The higher-order derived coupling features include: the ratio of the area enclosed by the flow-volume loop to its perimeter; the ratio of the peak expiratory flow rate of the flow-volume loop to its tidal volume; the ratio of the expiratory area to the inspiratory area; the ratio of the area of ​​the flow-volume loop in the 50%–100% volume range after expiration to the total expiratory area; the ratio of the peak expiratory flow rate to the peak inspiratory flow rate; the ratio of the instantaneous flow rate at 25% expiratory volume to the total expiratory volume; the ratio of the instantaneous flow rate at 50% expiratory volume to the instantaneous flow rate at 50% inspiratory volume; the ratio of the instantaneous flow rate at 75% expiratory volume to the instantaneous flow rate at 75% inspiratory volume; the ratio of the instantaneous flow rate at 25% expiratory volume to the highest expiratory flow rate; the ratio of the instantaneous flow rate at 25% expiratory volume to the instantaneous flow rate at 75% expiratory volume; and the ratio of the curvature of the curve to the right of the highest point of expiratory flow rate on the flow-volume loop curve to the average curvature of the post-expiratory phase segment.

3. The method for auxiliary judgment of the three-branch classification of infant asthma as described in claim 1, characterized in that, The method for obtaining the curve morphology features includes: A geometric fitting model of the complete two-dimensional flow-volume loop during tidal breathing is performed to extract the shape of the curve; the origin A of the curve is taken as the origin of the coordinate system, and the maximum value of the ordinate of the curve during the breathing stage is taken as the highest point C of the curve. Using the highest point as the fulcrum, the curve is divided into the left side and the right side of the highest breathing point. Two-dimensional straight line fitting is performed on the curves of the left and right sides respectively to obtain the optimal point B on the left and the optimal point D on the right. The origin A is connected to the optimal point B on the left, the optimal point B on the left to the highest point C, the highest point C to the optimal point D on the right, the optimal point D on the right to the projection E of point D on the x-axis, and point E to point A, thus obtaining the pentagonal feature of the tidal breathing curve. Based on the characteristics of the pentagon, calculate the distance to BC, the angle of ∠BCD, the angle of ∠CAE, the angle of ∠ABD, the angle of ∠ACD, the slope of line CD, the root mean square error of line CD and the curve to the right of point C, the mean absolute error of line CD and the curve to the right of point C, the area of ​​the pentagon, and the ratio of the area to the perimeter of the pentagon.

4. The method for auxiliary judgment of the three-branch classification of infant asthma as described in claim 3, characterized in that, A two-dimensional straight line fit is performed on the curve on the left to obtain the optimal point B on the left, including: Using the left side of the highest point as the index range, traverse every point on the curve as a candidate point; For any candidate point, calculate the slope of the line from the origin to that candidate point to obtain the equation of the first interval line. Calculate the slope of the line from that candidate point to the highest point to obtain the equation of the second interval line. Thus, obtain the first interval line equation and the second interval line equation for each candidate point. For any candidate point, a first predicted value is obtained by predicting the data points on the curve between the origin and the candidate point based on the first interval straight line equation. A second predicted value is obtained by predicting the data points on the curve between the highest point and the candidate point based on the second interval straight line equation. A first root mean square error is calculated based on all first predicted values, and a second root mean square error is calculated based on all second predicted values. A weighted linear combination error is calculated based on the first and second root mean square errors, thus obtaining the weighted linear combination error for each candidate point. The candidate point with the smallest weighted linear combination error is selected as the optimal point B on the left.

5. The method for auxiliary judgment of the three-branch classification of infant asthma as described in claim 3, characterized in that, Two-dimensional straight line fitting is performed on the curve on the right to obtain the optimal point D on the right. This includes: extracting data points on the curve on the right starting from the highest point C of the curve, fitting a linear function using the least squares method to obtain the slope and intercept representing the trend at the end of the respiratory process, and obtaining the fitted straight line; inputting the rightmost x-coordinate of the curve to the fitted linear function to obtain the extrapolated position of the fitted straight line on the rightmost side of the curve, and taking this extrapolated position as the optimal point D on the right.

6. The method for auxiliary judgment of the three-branch classification of infant asthma as described in claim 1, characterized in that, The process of extracting heterogeneous features and obtaining grouped log-likelihood ratios from preprocessed tidal breathing data and demographic data includes: Gaussian discriminant analysis was performed on the characteristics of traditional tidal breathing test data to calculate the log-likelihood ratio, thus obtaining the first log-likelihood ratio. The log-likelihood ratio of higher-order derived coupling features is calculated by a class-conditional multivariate Student-t model that combines a class-conditional model with a multivariate Student-t distribution, thus obtaining the second log-likelihood ratio. Using the curve morphology features as input to the conditional probability principal component analysis model, the log-likelihood ratio is calculated to obtain the third log-likelihood ratio. Using demographic data features as input to the Bayesian logistic regression model, the log-likelihood ratio is calculated to obtain the fourth log-likelihood ratio.

7. The method for auxiliary judgment of the three-branch classification of infant asthma as described in claim 6, characterized in that, The method of calculating the log-likelihood ratio of higher-order derived coupling features by combining a class-conditional model with a multivariate Student-t distribution includes: before obtaining the second log-likelihood ratio based on the class-conditional multivariate Student-t model, performing a monotonically invertible transformation on the mathematical properties of each feature in the higher-order derived coupling features, so that all features are mapped to the real number domain (-∞, +∞), and then inputting them into the class-conditional multivariate Student-t model.

8. The method for auxiliary judgment of the three-branch classification of infant asthma as described in claim 1, characterized in that, The hierarchical Bayesian fusion of each log-likelihood ratio to obtain the fusion result includes: A linear Bayesian discriminant structure is used to hierarchically fuse the log-likelihood ratios to obtain the log-posterior ratio; Based on the sigmoid function, the log-posterior ratio is processed to obtain the posterior probability; The posterior probability is used as the fusion result.

9. The method for auxiliary judgment of the three-branch classification of infant asthma as described in claim 1, characterized in that, The aforementioned three-branch classification-based asthma risk assessment based on fusion results includes: The relationship between the fusion result and a preset judgment threshold is determined. The preset judgment threshold includes a first judgment threshold and a second judgment threshold, and the first judgment threshold is greater than the second judgment threshold. If the fusion result is greater than or equal to the first judgment threshold, the asthma risk is determined to be high. If the fusion result is less than or equal to the second judgment threshold, the asthma risk is determined to be low. If the fusion result is less than the first judgment threshold but greater than the second judgment threshold, the asthma risk is determined to be in an uncertain region and further examination and judgment are required.

10. A computer-readable storage medium, characterized in that, The medium stores a program that can be loaded and executed by a processor as described in any one of claims 1 to 9, which is an auxiliary method for the three-branch classification of infant asthma.

Images