CPAP pressure prediction method based on multi-dimensional physiological indicators

Abstract

The application discloses a continuous positive airway pressure (CPAP) pressure prediction method based on multi-dimensional physiological indexes. The method first collects polysomnography (PSG) data and anthropometric indexes of an obstructive sleep apnea (OSA) patient, and obtains two types of data based on the PSG data: (1) quantifying OSA inner type parameters by using PUPbeta automatic analysis software; (2) obtaining PSG derived parameters by using artificial analysis data; after completing data preprocessing, a two-way stepwise regression method based on Akaike Information Criterion (AIC) is used to screen core features, and then eight kinds of machine learning models, including stepwise linear regression, ridge regression, Lasso regression, decision tree, random forest, gradient boosting regression, support vector regression and XGBoost regression, are constructed and trained, the optimal model is screened through model performance evaluation, and CPAP pressure prediction is realized.

Description

Technical Field

[0001] This invention relates to the field of medical device technology, and more specifically, to a method for predicting CPAP stress based on multidimensional physiological indicators. Background Technology

[0002] Obstructive sleep apnea (OSA) is a prevalent chronic sleep-disordered breathing disease worldwide. Its core pathological feature is repeated upper airway collapse during sleep, leading to apnea or hypoventilation events. It can cause serious complications such as daytime sleepiness, hypertension, and cardiovascular disease, and usually requires long-term, standardized treatment intervention.

[0003] Continuous positive airway pressure (CPAP) is currently the preferred treatment for obstructive sleep apnea (OSA). It maintains airway patency by continuously delivering a constant positive pressure to the upper airway, and its effectiveness is highly dependent on the precise setting of CPAP pressure parameters. However, traditional CPAP pressure determination primarily relies on manual pressure titration procedures in sleep laboratories. This process typically requires a specialist to operate polysomnography (PSG) equipment and gradually adjust the pressure throughout the patient's sleep to determine the optimal therapeutic pressure value. This method suffers from numerous drawbacks, including complex procedures, high time and labor costs, significant consumption of medical resources, and limited clinical application.

[0004] To reduce reliance on manual titration, existing studies have attempted to construct CPAP pressure prediction models or empirical equations based on clinical indicators. However, most existing prediction models are based solely on demographic characteristics and conventional PSG-derived indicators, such as the Apnea-Hypopnea Index (AHI) and Oxygen Desaturation Index (ODI). Their predictive performance and explanatory power are limited, with determination coefficients (R²) typically ranging from 0.30 to 0.50, making it difficult to meet the clinical needs of precision and personalized treatment.

[0005] In recent years, research on the pathophysiological mechanisms of OSA has gradually formed the PALM framework. This framework characterizes the pathogenesis of OSA from multiple dimensions, namely, passive upper airway collapsibility, arousal threshold (AT), loop gain (LG), and muscle responsiveness. Intrinsic parameters quantified from PSG data can directly reflect these core pathophysiological mechanisms, providing a new theoretical basis and technical approach for CPAP stress prediction. However, current research has not yet achieved a systematic integration of OSA pathophysiological intrinsic parameters with traditional clinical and PSG indicators. Furthermore, it lacks efficient feature screening mechanisms for high-dimensional features and multi-model comparison and optimization strategies, resulting in significant shortcomings in the clinical usability, stability, and predictive accuracy of existing prediction models.

[0006] Therefore, constructing a CPAP stress prediction method that integrates pathophysiological endogenous information, possesses efficient feature screening capabilities, and balances prediction accuracy with individualized characteristics has become a key technical problem that urgently needs to be solved in the clinical treatment of OSA.

[0007] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of the present invention, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0008] The purpose of this invention is to address the common problems of insufficient explanatory power and weak correlation with pathophysiological mechanisms when using traditional PSG-derived indicators to predict CPAP pressure in OSA patients. This invention provides a prediction method that integrates pathophysiological intrinsic parameters, anthropometric indicators, and PSG-derived indicators to achieve mechanism-driven accurate estimation of individual CPAP pressure, reduce reliance on traditional manual titration processes, and improve the efficiency and consistency of clinical treatment.

[0009] In a first aspect of the invention, a method for predicting pressure during continuous positive airway pressure (CPAP) in patients with obstructive sleep apnea (OSA) based on multidimensional physiological indicators is provided, comprising: (I) Data acquisition steps, including: (Ia) Collect anthropometric indicators of the patient to be predicted, including at least the body mass index (BMI). (Ib) Collect polysomnography (PSG) data from the patient to be predicted; and obtain OSA intrinsic type parameters and PSG derived parameters based on the PSG data, wherein, The OSA intrinsic parameters include at least: LG1, Vpassive, and Vmin, which are calculated using PUPbeta automatic analysis software. The PSG derived parameters include at least: AI, RE_OD_Mean, Ap_Dur_Long, and RR_ArI, which are obtained through manual analysis of PSG data; (II) CPAP stress prediction steps: Input the feature values ​​of at least the above 8 features of the patient to be predicted, namely BMI, LG1, Vpassive, Vmin, AI, RE_OD_Mean, Ap_Dur_Long, and RR_ArI, into the prediction model and output the corresponding CPAP stress prediction results. The prediction model is obtained by training a regression model, and the regression model training method includes: (1) Data acquisition steps, including: ① Collect anthropometric data of OSA patients, including BMI, neck circumference (NC), chest circumference (CC), and abdominal circumference (AC). ② Collect polysomnography (PSG) data from OSA patients; and obtain OSA intrinsic type parameters (calculated using PUPbeta automated analysis software) and PSG derived parameters (obtained through manual analysis of PSG data) based on the PSG data; among which, The OSA intrinsic parameters include: LG1, LGn, ArThres, Vpassive, and Vmin; The PSG derived parameters include, but are not limited to: AHI, AHI_REM, AHI_NREM, AI, AI_Ob, ODI_3, ODI_4, SpO2_Min, RE_OD_Mean, Ap_Dur_Mean, Ap_Dur_Long, Hyp_Dur_Mean, Hyp_Dur_Long, CT90, RR_ArI; ③ Provides a target value for Continuous Positive Airway Pressure (CPAP), which is determined by standardized manual titration and used as a reference value for model training and evaluation; (2) Data preprocessing steps: The data obtained in step (1) is processed for missing values, data type conversion and outlier filtering, and the data is divided into training set and validation set according to a preset ratio; (4) Model training steps: Use the training set and validation set in (2) to train and validate the regression model to obtain the prediction model.

[0010] In another preferred embodiment, the regression model includes, but is not limited to: stepwise linear regression, ridge regression, Lasso regression, Elastic Net regression, partial least squares regression (PLSR), decision tree regression, random forest regression, gradient boosting regression, AdaBoost regression, support vector regression (SVR), K-nearest neighbor regression (KNNRegression), eXtreme gradient boosting (XGBoost), LightGradient Boosting Machine (LightGBM), CatBoost regression, deep neural network regression, convolutional neural network (CNN) regression, recurrent neural network (RNN) regression, long short-term memory (LSTM) regression, Transformer regression, and ensemble or combination models of the above models.

[0011] In another preferred embodiment, in step (4): ① The standardization process uses centering and scaling methods to transform features based on the mean and standard deviation of the training set; ② The parameters of the gradient boosting regression model are set as follows: distribution type is Gaussian distribution, number of trees is 500, interaction depth is 3, and learning rate is 0.01.

[0012] In another preferred embodiment, in step (4): ① The standardization process uses centering and scaling methods to transform features based on the mean and standard deviation of the training set; ③ The parameters of the XGBoost regression model are set as follows: objective function is squared error regression, learning rate is 0.01, maximum tree depth is 3, and subsample ratio is 0.8.

[0013] In another preferred embodiment, in step (4): ① The standardization process uses centering and scaling methods to transform features based on the mean and standard deviation of the training set; ② The parameters of the gradient boosting regression model are set as follows: distribution type is Gaussian distribution, number of trees is 500, interaction depth is 3, and learning rate is 0.01; ③ The parameters of the XGBoost regression model are set as follows: objective function is squared error regression, learning rate is 0.01, maximum tree depth is 3, and subsample ratio is 0.8.

[0014] In another preferred embodiment, the model training method further includes: (3) Feature screening step: This step is located between step (2) and step (4). Based on the training set in step (2), a full linear regression model is constructed, and the bidirectional stepwise regression method based on the Akaike Information Criterion (AIC) is used to screen the core features that are significantly related to CPAP stress.

[0015] In another preferred embodiment, in the model training step (4), the core features obtained by screening are subjected to feature scaling processing, and machine learning models are trained respectively. The machine learning models are selected from the following group: stepwise linear regression, ridge regression, Lasso regression, decision tree regression, random forest regression, gradient boosting regression, support vector regression (SVR), and extreme gradient boosting (XGBoost).

[0016] In another preferred embodiment, the model training method further includes: (5) Model evaluation and selection: On the validation set, the predictive performance of each model is evaluated using the coefficient of determination (R²), mean square error (MSE), and / or root mean square error (RMSE), and the model with the best predictive performance is selected.

[0017] In another preferred example, the optimal model selected in step (5) is the gradient boosting regression model, whose validation set R²≥0.57 and RMSE≤1.37.

[0018] In another preferred embodiment, the feature values ​​of at least eight features of the patient to be predicted, namely BMI, LG1, Vpassive, Vmin, AI, RE_OD_Mean, Ap_Dur_Long, and RR_ArI, are input into the optimal model, and the corresponding CPAP pressure prediction results are output.

[0019] In another preferred embodiment, the prediction model is a validated model, and the validation criteria for the clinical application of its output CPAP pressure prediction value are: the prediction results are validated by overnight manual titration, and the titration pressure can eliminate more than 90% of obstructive respiratory events (including apnea, hypoventilation and respiratory effort-related awakening).

[0020] In another preferred embodiment, the patient described in step (1) must meet the following inclusion criteria: (a1) The age range is 18 to 70 years old; (b1) Meets the diagnostic criteria for OSA in the International Classification of Sleep Disorders, Third Edition (ICSD-3); (c1) Exclude patients with severe respiratory diseases, neuromuscular diseases, comorbid insomnia and sleep apnea (COMISA), and other sleep disorders, including rapid eye movement sleep behavior disorder (RBD), narcolepsy, and / or idiopathic hypersomnia.

[0021] In another preferred embodiment, the multidimensional physiological indicators include three categories: (a2) Anthropometric indicators: including at least Body Mass Index (BMI); (b2) OSA internal parameters: including at least loop gain (LG1), passive upper airway collapsibility (Vpassive), and minimum ventilation (Vmin). (c2) PSG derived parameters: including at least the Apnea Index (AI), Respiratory Event-Related Mean Oxygen Desaturation (RE_OD_Mean), Longest Apnea Duration (Ap_Dur_Long), and Respiratory-Related Arousal Index (RR_ArI).

[0022] In another preferred embodiment, the (b2) OSA intrinsic parameter in the multidimensional physiological indicators also includes the arousal threshold (AT).

[0023] In another preferred embodiment, the collection of PSG data in steps (I) and (1) must conform to the standards of the American Academy of Sleep Medicine's Manual of Interpretation of Sleep and Related Events, and must include at least the following physiological signals: (a3) Electroencephalogram (EEG): Frontal, central and occipital leads; (b3) Electrooculogram (EOG) and electromyography (EMG): used for sleep staging and wakefulness detection; (c3) Nasal pressure sensor and chest and abdominal respiratory plethysmography: used for monitoring airflow and respiratory effort; (d3) Pulse Oximeter: Used for recording peripheral oxygen saturation (SpO2), with a sampling frequency ≥1 Hz.

[0024] In another preferred embodiment, in step (2) of the data preprocessing step, the missing value handling adopts the predicted mean matching method, the data type conversion converts the character data to the numeric type, and the outlier filtering deletes rows with all missing values ​​and invalid data.

[0025] In another preferred embodiment, the model training method includes the following steps: (1) Data collection Two types of data (clinical data and monitoring data) were collected from OSA patients aged 18-70 years. Patients with severe respiratory diseases, neuromuscular diseases, and specific sleep disorders (COMISA, RBD, narcolepsy, and idiopathic hypersomnia) were excluded. ① Clinical data: including anthropometric measurements, at least BMI and neck circumference; ② Monitoring data: PSG records were acquired for two consecutive nights. The first night was used to extract OSA internal parameters and PSG-derived parameters, and the second night was used for manual titration to obtain the minimum effective pressure of CPAP. The intrinsic parameters are extracted using the PUPBeta platform and include at least the ring gain (LG1), the airway passive collapse index (Vpassive), and the minimum ventilation (Vmin). The PSG derived parameters include: AHI, AHI_REM, AHI_NREM, AI, AI_Ob, ODI_3, ODI_4, SpO2_Min, RE_OD_Mean, Ap_Dur_Mean, Ap_Dur_Long, Hyp_Dur_Mean, Hyp_Dur_Long, CT90, RR_ArI; (2) Data preprocessing (a) Missing value handling: Missing values ​​are imputed for continuous variables using the predicted mean matching method; (b) Data cleaning: Delete sample rows with all missing values ​​and convert character variables to numeric variables; (c) Dataset partitioning: The training set and validation set are randomly partitioned in an 8:2 ratio; (3) Feature Filtering A full linear regression model with CPAP as the dependent variable and all candidate indicators as independent variables was constructed based on the training set. The core features that make significant contributions to CPAP stress prediction were screened using the bidirectional stepwise regression method based on the AIC criterion. Finally, eight core feature variables were determined: LG1, Vpassive, Vmin, BMI, AI, RE_OD_Mean, Ap_Dur_Long, and RR_ArI. (4) Model training (a) Feature standardization: Only the core features obtained from the screening are centered and scaled, and the mean and standard deviation of the training set are saved for subsequent prediction stages; (b) Multi-model construction: Based on the standardized core features, the following 8 prediction models are trained respectively: i) Stepwise linear regression: A linear regression model is constructed based on selected features. The optimal feature combination is obtained through stepwise variable selection, and an interpretable prediction equation is output. ii) Ridge regression: By introducing an L2 regularization term, the model overfitting and multicollinearity are suppressed, the model's generalization ability is improved, and cross-validation is used to determine the optimal regularization parameter; iii) Lasso regression: Feature compression and variable selection are achieved by introducing an L1 regularization term, and the optimal λ value is selected by 5-fold cross-validation; iv) Decision tree regression: Construct a regression tree model based on the feature splitting criterion to achieve nonlinear relationship modeling; v) Random Forest Regression: A regression model is built based on the ensemble of multiple decision trees, with the optimal number of trees being 500, and variable importance assessment is enabled; vi) Gradient boosting regression: The gradient boosting framework is used to gradually optimize the residuals. The preferred distribution type is Gaussian distribution, the preferred number of trees is 500, the preferred interaction depth is 3, and the preferred learning rate is 0.01. vii) Support Vector Regression: A nonlinear regression model is constructed by selecting the Radial Basis Function (RBF) kernel function and trained based on standardized features; viii) XGBoost Regression: The regression model is constructed using the extreme gradient boosting framework. The objective function is preferably squared error regression, the learning rate is preferably 0.01, the maximum tree depth is preferably 3, and the subsample ratio is preferably 0.8. The above eight prediction models can be used individually or combined using weighted fusion, stacking, or voting fusion methods. (5) Model evaluation and selection The performance of each model was evaluated using the coefficient of determination (R²), mean squared error (MSE), and / or root mean square error (RMSE) on both the training and validation sets. The prediction accuracy and generalization ability of the models were compared comprehensively. The gradient boosting regression model performed best, with an R² of 0.733 and an RMSE of 1.096 on the training set and an R² of 0.572 and an RMSE of 1.367 on the validation set, achieving a good balance between prediction accuracy and stability.

[0026] In another preferred embodiment, the (II) CPAP stress prediction step includes: standardizing the eight core features of the patient to be predicted according to the mean and standard deviation corresponding to the training set, and then inputting them into the prediction model (e.g., a trained gradient boosting regression model) to output the CPAP stress prediction value.

[0027] In another preferred embodiment, in application scenarios where model interpretability needs to be emphasized, the eight core features of the patient to be predicted are standardized according to the mean and standard deviation corresponding to the training set, and then input into the prediction model (e.g., a trained stepwise linear regression model or a trained Lasso regression model), and the CPAP stress value is directly calculated using its prediction equation at the original scale.

[0028] In a second aspect of the invention, a system for predicting CPAP stress in OSA patients is provided, comprising: Data acquisition module: Collects anthropometric indicators and polysomnography (PSG) data of the patient to be predicted, wherein the anthropometric indicators include at least BMI; Data preprocessing module: Based on the PSG data collected by the data acquisition module, OSA intrinsic parameters and PSG derived parameters are obtained. The OSA intrinsic parameters include at least LG1, Vpassive, and Vmin, which are calculated by PUPbeta automatic analysis software. The PSG derived parameters include at least AI, RE_OD_Mean, Ap_Dur_Long, and RR_ArI, which are obtained by manual analysis of PSG data. Prediction Module: Input the feature values ​​of at least the above 8 features of the patient to be predicted, namely BMI, LG1, Vpassive, Vmin, AI, RE_OD_Mean, Ap_Dur_Long, and RR_ArI, into the prediction module, and output the corresponding CPAP pressure prediction results.

[0029] In a third aspect of the invention, a device for predicting CPAP pressure in OSA patients is provided, comprising: processor; Memory, which stores computer-readable instructions; The processor is configured to perform the method according to the first aspect of the invention by executing the computer-readable instructions.

[0030] In a fourth aspect of the invention, a computer-readable storage medium is provided storing computer-readable instructions that, when executed by a processor, implement the method according to a first aspect of the invention.

[0031] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description

[0032] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention. It is obvious that the drawings described below are merely some embodiments of the invention, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0033] Figure 1 The overall flowchart of the CPAP stress prediction method in this embodiment of the invention is shown. The process mainly includes four stages: data acquisition, data processing, feature selection, and model training.

[0034] Figure 2The scatter plot shows the CPAP stress prediction results of the gradient boosting regression model of this invention on the validation set. The horizontal axis represents the actual CPAP stress value, and the vertical axis represents the CPAP stress value predicted by the model. Blue scatter points represent the correspondence between the actual and predicted values ​​for each subject sample; the red dashed line represents the ideal prediction reference line (i.e., the predicted value equals the actual value); and the green solid line represents the model fitting trend line. Detailed Implementation

[0035] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the embodiments set forth herein; rather, they are provided so that the invention will be thorough and complete, and the concept of the exemplary embodiments will be fully conveyed to those skilled in the art.

[0036] Unless otherwise defined, all terms and phrases used herein include their meanings as they have in the art, unless explicitly stated otherwise or clearly indicated from the context of their use. While any methods and materials similar to or equivalent to those described herein may be used in carrying out or testing the invention, specific methods and materials are now described.

[0037] Through extensive and in-depth research, the inventors have designed for the first time a method for predicting continuous positive airway pressure (CPAP) in patients with obstructive sleep apnea (OSA) based on multi-dimensional physiological indicators (hereinafter referred to as the "CPAP pressure prediction method based on multi-dimensional physiological indicators"). This method is applicable to the rapid and accurate setting of pressure parameters of CPAP devices in clinical OSA treatment and belongs to the field of sleep-disordered breathing treatment technology.

[0038] Specifically, this invention designs a method for predicting continuous positive airway pressure (CPAP) in patients with obstructive sleep apnea (OSA) by combining pathophysiological intrinsic indices and anthropometric parameters. This method integrates OSA intrinsic indices based on pathophysiological mechanisms with traditional clinical parameters, improves model computational efficiency through feature selection, and enhances the stability and accuracy of CPAP pressure prediction through a multi-model comparison strategy. The gradient boosting regression model performed best on the validation set, with an R² of 0.572 and an RMSE of 1.367. This method effectively reduces the time and labor costs of traditional CPAP pressure titration and holds promise for providing OSA patients with a personalized and efficient CPAP pressure prediction solution, demonstrating clinical application potential. Based on this, this invention was completed.

[0039] the term

[0040] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0041] As used herein, the terms “containing” or “including (comprise)” can be open-ended, semi-closed, or closed. In other words, the terms also include “consistently made of” or “composed of”.

[0042] OSA: Obstructive Sleep Apnea. During sleep, due to anatomical narrowing of the pharynx and collapse of soft tissues, airflow through the mouth and nose is completely or partially interrupted, but breathing efforts through the chest and abdomen still continue.

[0043] CPAP: is a physical therapy technique that uses a ventilator to generate continuous positive airway pressure to maintain an open upper airway.

[0044] OSA respiratory endotype (Obstructive Sleep Apnea Respiratory Endotype) reflects the quantitative characteristics of the core pathophysiological mechanisms of OSA. It is derived based on the PALM framework (airway collapseability, arousal threshold, ventilation control stability, and muscle responsiveness) and can be extracted from PSG data through an automated analysis platform (such as PUP Beta). Core indicators include loop gain (LG1 / LGn), passive airway collapseability (Vpassive), minimum tidal volume (Vmin), and arousal threshold (ArThres), which are directly related to the physiological mechanisms of disease occurrence and development.

[0045] LG1: Loop Gain at 1 cycle / min (LG1), used to characterize the feedback sensitivity of the respiratory control system without the cyclic delay effect.

[0046] LGn: Overall Loop Gain (LGn) is used to characterize the overall feedback gain level of a respiratory control system when the cyclic delay effect is included.

[0047] ArThres: Arousal Threshold refers to the level of ventilation drive required to induce sleep-wakefulness. It is expressed as a percentage of normal ventilation level and is used to reflect a subject's arousal sensitivity to respiratory stimuli.

[0048] Vpassive: Passive Airway Collapsibility refers to the ventilation level under normal ventilation-driven conditions. It is used to reflect the degree of upper airway collapse in passive conditions. The lower the value, the more likely the airway is to collapse.

[0049] Vmin: Minimum Ventilation (Vmin) refers to the ventilation level under the lowest ventilatory drive condition, and is used to reflect the patency of the upper airway under the lowest muscle tone condition.

[0050] BMI: Body Mass Index, the ratio of weight to the square of height (kg / m²).

[0051] NC: Neck Circumference, the horizontal circumference of the neck at the Adam's apple node.

[0052] CC: Chest Circumference, measured horizontally around the ribcage, passing under the lower angles of both scapulae, around the nipple level (male) / at the fourth intercostal space at the base of the breast (female).

[0053] AC: Abdominal Circumference, measured horizontally around the abdomen at the midpoint of the line connecting the lower edge of the costal arch and the upper edge of the iliac crest.

[0054] AHI: Apnea-Hypopnea Index, the total number of apnea and hypopnea events during sleep per hour.

[0055] AHI_REM: Apnea-Hypopnea Index during Rapid Eye Movement (REM) sleep, the total number of apnea and hypopnea events per hour during the rapid eye movement (REM) sleep stage.

[0056] AHI_NREM: Non-rapid eye movement sleep apnea-hypopnea index (NREM-AHI), the total number of apnea and hypopnea events per hour during non-rapid eye movement (NREM) sleep.

[0057] AI: Apnea Index, the number of all apnea events (obstructive + mixed + central) occurring per hour during sleep.

[0058] AI_Ob: Obstructive Apnea Index, the number of obstructive apnea events occurring during sleep per hour.

[0059] ODI_3: 3% Oxygen Desaturation Index, the number of events per hour in which blood oxygen saturation drops by ≥3% during sleep.

[0060] ODI_4: 4% Oxygen Desaturation Index, the number of events per hour in which blood oxygen saturation drops by ≥4% during sleep.

[0061] SpO2_Min: Minimum Oxygen Saturation, the lowest level of oxygen saturation detected during sleep.

[0062] RE-OD-Mean: Mean Respiratory Event-Related Oxygen Desaturation, the average decrease in blood oxygen saturation associated with all respiratory events (apnea + hypoventilation).

[0063] Ap_Dur_Mean: Mean Apnea Duration, the average duration of all apnea events (obstructive + mixed + central).

[0064] Ap-Dur-Long: Longest Apnea Duration, the longest duration of a single apnea event.

[0065] Hyp_Dur_Mean: Mean Hypopnea Duration, the average duration of all hypoventilation events.

[0066] Hyp_Dur_Long: Longest Hypopnea Duration, the longest duration of a single hypopnea event.

[0067] CT90: Cumulative Time below 90% OxygenSaturation, which is the percentage of total sleep time during which blood oxygen saturation is below 90%.

[0068] RR-ArI: Respiratory-Related Arousal Index, the number of awakenings related to respiratory events during sleep per hour.

[0069] PSG (Polysomnography) is a detection technique that analyzes sleep structure and related physiological and behavioral changes by simultaneously monitoring multiple parameters, including electroencephalography (EEG), electromyography (EMG), electrooculography (EOG), nasal and oral airflow, chest and abdominal respiratory movements, blood oxygen saturation, electrocardiogram (ECG), snoring, and exhaled carbon dioxide partial pressure. It is an important method for diagnosing sleep disorders.

[0070] AIC Regression: Bidirectional Stepwise Regression is a feature selection method based on the Akaike Information Criterion (AIC). Through a bidirectional iterative process of "forward introduction" (incorporating significantly relevant variables into the model) and "backward elimination" (removing redundant or insignificant variables), it selects the core feature set with the best explanatory power for the dependent variable (such as CPAP stress) from the candidate variable set, balancing model fit and simplicity, and reducing the interference of redundant variables on the prediction effect.

[0071] GBR (Gradient Boosting Regression) is an ensemble learning algorithm that iteratively builds a series of weak regressors (such as decision trees). Each iteration corrects the prediction error (residual) of the previous model and uses a gradient descent strategy to minimize the loss function (in this invention, the squared loss corresponding to a Gaussian distribution). By accumulating the prediction results of the weak models, a strong model is formed. The core parameters include the number of trees (n.trees), interaction depth (interaction.depth), and learning rate (shrinkage). It features high prediction accuracy and strong generalization ability and is suitable for nonlinear prediction scenarios of CPAP pressure.

[0072] RMSE: Root Mean Square Error is a quantitative indicator that measures the degree of deviation between the model's predicted values ​​and the actual values. It is calculated as the square root of the mean of the sum of squared prediction errors, with the same units as the dependent variable. The smaller the value, the higher the accuracy of the model's prediction.

[0073] R2: Coefficient of Determination, reflects the model's ability to explain the variation of the dependent variable. The value ranges from 0 to 1. The closer R2 is to 1, the higher the proportion of the variation of the dependent variable that the model can explain, and the better the prediction effect. It is divided into training set R2 (reflecting the model's fitting ability) and validation set R2 (reflecting the model's generalization ability).

[0074] Manual analysis of PSG data to obtain PSG derived parameters

[0075] In this invention, PSG-derived parameters are obtained through manual analysis of PSG data. An example of the specific method is as follows: PSG-derived parameters were obtained from the first night's PSG monitoring data. All records were scored according to the American Academy of Sleep Medicine's Manual of Interpretation of Sleep and Related Events: Rules, Terminology, and Technical Specifications (3rd Edition) (AASM 3.0). Apnea was defined as a ≥90% drop in the oral and nasal thermal sensor signal from the pre-event baseline, lasting at least 10 seconds; based on the presence of respiratory effort in the chest and abdominal respiratory signals, it was classified as obstructive, central, or mixed apnea according to AASM standards. Hypopnea was defined as a ≥30% drop in the nasal pressure signal from the baseline, lasting at least 10 seconds, accompanied by a ≥3% decrease in blood oxygen saturation or awakening. The Apnea Index (AI) was the number of apneas per hour of sleep. The Hypopnea Index (HI) was the number of hypopneas per hour of sleep. The Oxygen Depression Index included the Oxygen Depression ≥3% Index (ODI-3) and the Oxygen Depression ≥4% Index (ODI-4), defined as the number of times blood oxygen saturation decreased by ≥3% and ≥4% per hour of sleep, respectively. The lowest blood oxygen saturation was the lowest blood oxygen value recorded during the entire sleep period. The cumulative duration of oxygen saturation < 90% (CT90) is defined as the total duration (in seconds) during sleep when oxygen saturation remains below 90%. Respiratory-related arousal refers to arousal accompanied by apnea or hypopnea; the Respiratory-Related Arousal Index (RR-ArI) is defined as the number of such events occurring per hour of sleep. The mean oxygen saturation drop associated with respiratory events (RE-OD-Mean) is a derived indicator from routine polysomnography, representing the average magnitude of oxygen saturation drop associated with apnea and hypopnea events. Sleep staging and arousal scores are based on electroencephalography (EEG), electrooculography (EOG), and electromyography (EMG) signals recorded by polysomnography, and are performed according to the AASM version 3.0 standard.

[0076] A method for predicting continuous positive airway pressure in patients with obstructive sleep apnea based on multidimensional physiological indicators.

[0077] This invention provides a method for predicting the pressure of continuous positive airway pressure (CPAP) in patients with obstructive sleep apnea based on multi-dimensional physiological indicators. The method includes: (I) Data acquisition steps, including: (Ia) Collect anthropometric indicators of the patient to be predicted, wherein the anthropometric indicators include at least BMI; (Ib) Collect polysomnography (PSG) data from the patient to be predicted; and obtain OSA intrinsic type parameters and PSG derived parameters based on the PSG data, wherein, The OSA intrinsic parameters include at least: LG1, Vpassive, and Vmin, which are calculated using PUPbeta automatic analysis software. The PSG derived parameters include at least: AI, RE_OD_Mean, Ap_Dur_Long, and RR_ArI, which are obtained through manual analysis of PSG data; (II) CPAP stress prediction steps: Input the feature values ​​of at least the above 8 features of the patient to be predicted, namely BMI, LG1, Vpassive, Vmin, AI, RE_OD_Mean, Ap_Dur_Long, and RR_ArI, into the prediction model and output the corresponding CPAP stress prediction results. The prediction model is obtained by training a regression model, and the regression model training method includes: ① Collect anthropometric parameters of patients with obstructive sleep apnea (OSA), including BMI, NC, CC, and AC; ② Collect polysomnography (PSG) data from patients with obstructive sleep apnea (OSA); and obtain OSA intrinsic type parameters (calculated using PUPbeta automated analysis software) and PSG derived parameters (obtained through manual analysis of PSG data) based on the PSG data; among which, The OSA intrinsic parameters include: LG1, LGn, ArThres, Vpassive, and Vmin; The PSG derived parameters include: AHI, AHI_REM, AHI_NREM, AI, AI_Ob, ODI_3, ODI_4, SpO2_Min, RE_OD_Mean, Ap_Dur_Mean, Ap_Dur_Long, Hyp_Dur_Mean, Hyp_Dur_Long, CT90, RR_ArI; ③ Provides a target value for Continuous Positive Airway Pressure (CPAP), which is determined by standardized manual titration and used as a reference value for model training and evaluation; (2) Data preprocessing steps: The data obtained in step (1) is processed for missing values, data type conversion and outlier filtering, and the data is divided into training set and validation set according to a preset ratio; (4) Model training steps: Use the training set and validation set in (2) to train and validate the regression model to obtain the prediction model.

[0078] In one specific embodiment of the present invention, the pressure prediction method flow is as follows: Figure 1 As shown, the prediction model training method includes the following steps: (1) Data collection Two types of data (i.e., clinical data and monitoring data) were collected from OSA patients aged 18-70 years, excluding patients with severe cardiopulmonary diseases, neuromuscular diseases and other sleep disorders. ① Clinical data: including anthropometric measurements, at least BMI and neck circumference; ② Monitoring data: PSG records were acquired for two consecutive nights. The first night was used to extract OSA intrinsic parameters and PSG derived parameters, and the second night was used to manually titrate the CPAP target pressure. The intrinsic parameters are extracted using the PUPBeta platform and include at least the ring gain (LG1), the airway passive collapse index (Vpassive), and the minimum ventilation (Vmin). The PSG derived parameters include: AHI, AHI_REM, AHI_NREM, AI, AI_Ob, ODI_3, ODI_4, SpO2_Min, RE_OD_Mean, Ap_Dur_Mean, Ap_Dur_Long, Hyp_Dur_Mean, Hyp_Dur_Long, CT90, RR_ArI; (2) Data preprocessing (a) Missing value handling: Missing values ​​are imputed for continuous variables using the predicted mean matching method; (b) Data cleaning: Delete sample rows with all missing values ​​and convert character variables to numeric variables; (c) Dataset partitioning: The training set and validation set are randomly partitioned in an 8:2 ratio; (3) Feature Filtering A full linear regression model with CPAP as the dependent variable and all candidate indicators as independent variables was constructed based on the training set. The core features that make significant contributions to CPAP stress prediction were screened using the bidirectional stepwise regression method based on the AIC criterion. Finally, eight core feature variables were determined: LG1, Vpassive, Vmin, BMI, AI, RE_OD_Mean, Ap_Dur_Long, and RR_ArI. (4) Model training (a) Feature standardization: Only the core features obtained from the screening are centered and scaled, and the mean and standard deviation of the training set are saved for subsequent prediction stages; (b) Multi-model construction: Based on the standardized core features, the following 8 prediction models are trained respectively: i) Stepwise linear regression: Constructs a linear model based on selected features and outputs an interpretable prediction equation; ii) Ridge regression: By introducing an L2 regularization term, the model overfitting and multicollinearity are suppressed, the model's generalization ability is improved, and cross-validation is used to determine the optimal regularization parameter; iii) Lasso regression: 5-fold cross-validation is used to select the optimal λ value, with alpha set to 0 / 1 respectively; iv) Decision tree regression: A regression tree was constructed using analysis of variance. v) Random forest regression: 500 trees, variable importance assessment enabled; vi) Gradient boosting regression: Gaussian distribution, 500 trees, interaction depth 3, learning rate 0.01; vii) Support Vector Regression: Using radial basis function kernel function, trained based on standardized features; viii) XGBoost Regression: The objective function is squared error regression, with a learning rate of 0.01, a maximum tree depth of 3, and a subsample ratio of 0.8. (5) Model evaluation and selection The performance of each model was evaluated using the coefficient of determination (R²), mean squared error (MSE), and / or root mean square error (RMSE) on both the training and validation sets. The predictive accuracy and generalization ability of the models were then compared comprehensively. The gradient boosting regression model performed best, with an R² of 0.733 and an RMSE of 1.096 on the training set, and an R² of 0.572 and an RMSE of 1.367 on the validation set. Figure 2This achieves a better balance between prediction accuracy and stability.

[0079] Figure 2 The results showed that the model predictions were generally in good agreement with the actual values. The validation set determination coefficient (R²) was 0.572, the root mean square error (RMSE) was 1.367, and the mean absolute error (MAE) was 1.178. This indicates that the constructed model can accurately predict the CPAP pressure required by the subjects, demonstrating that the feature system and prediction model constructed in this invention have good clinical prediction capabilities and application value.

[0080] In one specific embodiment of the present invention, the (II) CPAP stress prediction step includes: standardizing the eight core features of the patient to be predicted according to the mean and standard deviation corresponding to the training set, and then inputting them into the prediction model (e.g., a trained gradient boosting regression model) to output the CPAP stress prediction value.

[0081] In another specific embodiment of the present invention, in application scenarios where model interpretability needs to be emphasized, the eight core features of the patient to be predicted are standardized according to the mean and standard deviation corresponding to the training set, and then input into the prediction model (e.g., a trained stepwise linear regression model or a trained Lasso regression model), and the CPAP pressure value is directly calculated using its prediction equation at the original scale.

[0082] The main advantages of this invention include: (1) Mechanism innovation: For the first time, intrinsic parameters that can directly characterize the pathophysiological mechanism of OSA (such as loop gain LG1, passive airway collapse Vpassive, and minimum ventilation Vmin) are introduced into the CPAP pressure prediction model, breaking through the traditional empirical modeling paradigm that only relies on PSG-derived indicators. It explains the individual differences in CPAP pressure requirements from two core mechanism dimensions: ventilation control stability and upper airway mechanical characteristics. (2) Prediction accuracy: By integrating internal type indicators, traditional anthropometric indicators and PSG-derived parameters, and combining feature screening and multi-model comparison optimization strategies, a high-performance prediction model was constructed. The optimal model achieved a determination coefficient (R²) of 0.572 on the validation set, which is significantly better than the existing CPAP prediction models based on conventional clinical indicators (R² is mostly below 0.50), thus improving the stability and generalization ability of the prediction results. (3) Clinical applicability: No additional invasive examinations or special experimental procedures are required. The intrinsic parameters can be extracted and pressure prediction can be completed solely based on routine polysomnography (PSG) data, which significantly reduces the time and labor costs of manual titration. It is suitable for patients with mild to moderate-severe OSA (AHI < 50 times / hour) and has the feasibility of direct application in the routine CPAP treatment process in clinical practice.

[0083] The present invention will be further explained below with reference to specific embodiments.

[0084] It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Experimental methods in the following embodiments, unless otherwise specified, are generally performed under conventional conditions as described in *Molecular Cloning: A Laboratory Guide* by Sambrook J. et al. (translated by Huang Peitang et al., Beijing: Science Press, 2002), or according to the manufacturer's recommendations (e.g., product instructions). Unless otherwise stated, percentages and parts are by weight. Unless otherwise specified, the experimental materials and reagents used in the following embodiments are commercially available or can be prepared according to literature methods.

[0085] Example 1

[0086] (1) Subjects: 114 OSA patients diagnosed at Huashan Hospital affiliated with Fudan University from January 2022 to August 2025 were selected, including 15 mild cases, 38 moderate cases, and 61 severe cases; 90 were male (78.9%) and 24 were female (21.1%), with a median age of 47.5 years. All patients met the following criteria: (a) Inclusion criteria: 18-70 years old, diagnosed with OSA by PSG, completed two nights of laboratory PSG recording and manual CPAP titration, and the PSG data quality met the requirements for OSA respiratory pattern extraction; (b) Exclusion criteria: patients with severe respiratory diseases, neuromuscular diseases, other sleep disorders (such as RBD), comorbid insomnia (COMISA), or those whose diagnosis of nighttime head-up affects AHI assessment.

[0087] (2) Data acquisition and preprocessing: (a) Data acquisition: ①PSG monitoring: Standard clinical leads were used to record signals such as electroencephalogram (EEG), electrooculogram (EOG), electromyogram (EMG), electrocardiogram (ECG), nasal pressure airflow, chest and abdominal respiratory effort, and pulse oxygen saturation, and scored according to the AASM 3.0 standard; ②OSA intrinsic parameters extraction: The PSG data of the first night were processed through the PUPBeta platform to extract indicators such as LG1, Vpassive, and Vmin, and the median of NREM sleep period was taken as the final value; ③ PSG-derived parameters were obtained based on the PSG monitoring data from the first night. All records were scored according to the *American Academy of Sleep Medicine Manual for the Interpretation of Sleep and Related Events: Rules, Terminology, and Technical Specifications (3rd Edition)* (AASM 3.0). Apnea was defined as a ≥90% drop in the oral and nasal thermal sensor signal from the pre-event baseline, lasting at least 10 seconds; based on the presence of respiratory effort in the chest and abdominal respiratory signals, it was classified as obstructive, central, or mixed apnea according to AASM standards. Hypopnea was defined as a ≥30% drop in nasal pressure signal from the baseline, lasting at least 10 seconds, accompanied by a ≥3% decrease in blood oxygen saturation or awakening. The Apnea Index (AI) is the number of apneas per hour of sleep. The Hypopnea Index (HI) is the number of hypopneas per hour of sleep. Blood oxygen saturation drop indices include the ≥3% drop index (ODI-3) and the ≥4% drop index (ODI-4), defined as the number of times blood oxygen saturation drops by ≥3% and ≥4% per hour of sleep, respectively. The lowest blood oxygen saturation was the lowest blood oxygen value recorded during the entire sleep period. The cumulative duration of blood oxygen saturation < 90% (CT90) was defined as the total duration (in seconds) during sleep when blood oxygen saturation remained below 90%. Respiratory-related arousal refers to arousal accompanied by apnea or hypopnea; the Respiratory-Related Arousal Index (RR-ArI) is defined as the number of such events occurring per hour of sleep. The mean decrease in blood oxygen saturation associated with respiratory events (RE-OD-Mean) is a derived indicator from routine polysomnography, representing the average decrease in blood oxygen saturation associated with apnea and hypopnea events. Sleep staging and arousal scores were based on electroencephalogram (EEG), electrooculogram (EOG), and electromyography (EMG) signals recorded by polysomnography, and were performed according to the AASM version 3.0 standard. ④ Anthropometry indicators: Measure BMI, neck circumference, chest circumference, abdominal circumference, etc.; ⑤ CPAP target pressure: determined by standardized manual titration on the second night, with breathing events marked by technicians with more than 10 years of experience, and verified by sleep physicians certified by RPSGT.

[0088] (b) Data preprocessing: ① Candidate Indicator Set: Contains 25 variables (CPAP is the dependent variable, and the remaining 24 are candidate independent variables), specifically LG1, LGn, ArThres, Vpassive, Vmin, BMI, NC, CC, AC, AHI, AHI_REM, AHI_NREM, AI, AI_Ob, ODI_3, ODI_4, SpO2_Min, RE_OD_Mean, Ap_Dur_Mean, Ap_Dur_Long, Hyp_Dur_Mean, Hyp_Dur_Long, CT90, RR_ArI; ② Missing value handling: Missing values ​​are filled using the predicted mean matching method, and rows with completely missing values ​​are deleted; ③ Dataset partitioning: n=114, training set n=91 (80%), validation set n=23 (20%), random seed 81 to ensure reproducible grouping.

[0089] (3) Feature selection: Based on the training set, a full linear model is constructed, and the stepAIC function (MASS package) is used to perform bidirectional stepwise regression. Eight core features are selected: LG1, Vpassive, Vmin, BMI, AI, RE_OD_Mean, Ap_Dur_Long, RR_ArI. Redundant variables are excluded to improve model efficiency.

[0090] (4) Model training and evaluation: (a) Feature standardization: The preProcess function of the caret package is used to center and scale the 8 core features. The mean and standard deviation of the training set are used to standardize the validation set and prediction data. (b) Model training: Eight models were trained using R software (version 4.5.1) and corresponding packages (openxlsx, MASS, glmnet, etc.); (c) Performance Evaluation: The performance of all models is shown in Table 1 below: Table 1 Model Performance Comparison

[0091] As shown in Table 1 above, the gradient boosting regression model performs best on the validation set, balancing accuracy and generalization ability, and is therefore selected as the optimal prediction model. Figure 2 ).

[0092] Figure 2 The results showed that the model predictions were generally in good agreement with the actual values. The validation set determination coefficient (R²) was 0.572, the root mean square error (RMSE) was 1.367, and the mean absolute error (MAE) was 1.178. This indicates that the constructed model can accurately predict the CPAP pressure required by the subjects, demonstrating that the feature system and prediction model constructed in this invention have good clinical prediction capabilities and application value.

[0093] (5) Prediction Example: A patient with moderate OSA was selected, with the following core feature values: LG1=0.8, Vpassive=85, Vmin=60, BMI=28.5, AI=15.2, RE_OD_Mean=4.8, Ap_Dur_Long=25.3, RR_ArI=18.6. After standardization of the training set mean and standard deviation, the value was input into the gradient boosting regression model, and the output CPAP pressure prediction value was 7.2 cmH2O. The error between the predicted value and the manually titrated value (7.0 cmH2O) was only 0.2 cmH2O, which verified the accuracy of the model.

[0094] Other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of the invention are indicated by the appended claims.

Claims

1. A method for predicting pressure of continuous positive airway pressure for obstructive sleep apnea patients based on multi-dimensional physiological indicators, characterized in that, include: (I) Data acquisition steps, including: (Ia) Collect anthropometric indicators of the patient to be predicted, wherein the anthropometric indicators include at least body mass index (BMI); (Ib) Collect polysomnography (PSG) data from the patient to be predicted; and obtain OSA intrinsic type parameters and PSG derived parameters based on the PSG data, wherein, The OSA intrinsic parameters include at least: LG1, Vpassive, and Vmin, which are calculated using PUPbeta automatic analysis software. The PSG derived parameters include at least: AI, RE_OD_Mean, Ap_Dur_Long, and RR_ArI, which are obtained through manual analysis of PSG data; (II) CPAP stress prediction steps: Input the feature values ​​of at least the above 8 features of the patient to be predicted, namely BMI, LG1, Vpassive, Vmin, AI, RE_OD_Mean, Ap_Dur_Long, and RR_ArI, into the prediction model and output the corresponding CPAP stress prediction results. The prediction model is obtained by training a regression model, and the regression model training method includes: (1) Data acquisition steps, including: ① Collect anthropometric data of OSA patients, including: body mass index (BMI), neck circumference (NC), chest circumference (CC), and abdominal circumference (AC); ② Collect polysomnography (PSG) data from OSA patients; and obtain OSA intrinsic type parameters and PSG derived parameters based on the PSG data; wherein, The OSA intrinsic parameters include: LG1, LGn, ArThres, Vpassive, and Vmin; The PSG derived parameters include, but are not limited to: AHI, AHI_REM, AHI_NREM, AI, AI_Ob, ODI_3, ODI_4, SpO2_Min, RE_OD_Mean, Ap_Dur_Mean, Ap_Dur_Long, Hyp_Dur_Mean, Hyp_Dur_Long, CT90, RR_ArI; ③ Provides a target value for continuous positive airway pressure ventilation, which is determined by standardized manual titration and used as a reference value for model training and evaluation; (2) Data preprocessing steps: The data obtained in step (1) is processed for missing values, data type conversion and outlier filtering, and the data is divided into training set and validation set according to a preset ratio; (4) Model training steps: Use the training set and validation set in (2) to train and validate the regression model to obtain the prediction model.

2. The method according to claim 1, characterized in that, The model training method also includes: (3) Feature screening step: This step is located between step (2) and step (4). Based on the training set in step (2), a full linear regression model is constructed, and the bidirectional stepwise regression method based on the Akaike Information Criterion is used to screen the core features that are significantly related to CPAP stress.

3. The method according to claim 2, characterized in that, In the model training step (4), the core features obtained by screening are subjected to feature scaling processing, and machine learning models are trained respectively. The machine learning models are selected from the following group: stepwise linear regression, ridge regression, Lasso regression, decision tree regression, random forest regression, gradient boosting regression, support vector regression, and extreme gradient boosting regression.

4. The method according to claim 3, characterized in that, The model training method also includes: (5) Model evaluation and selection: On the validation set, the predictive performance of each model is evaluated using the coefficient of determination R², mean square error MSE, and / or root mean square error RMSE, and the model with the best predictive performance is selected.

5. The method according to claim 1, characterized in that, The multidimensional physiological indicators include three categories: (a2) Anthropometric indicators: including at least body mass index (BMI); (b2) OSA internal parameters: including at least loop gain LG1, passive airway collapseability Vpassive, and minimum ventilation Vmin; (c2) PSG derived parameters: including at least the apnea index AI, respiratory event-related mean oxygen saturation decrease RE_OD_Mean, longest apnea duration Ap_Dur_Long, and respiratory-related arousal index RR_ArI.

6. The method according to claim 1, characterized in that, The collection of PSG data in steps (I) and (1) must comply with the standards of the American Academy of Sleep Medicine's Manual of Interpretation of Sleep and Related Events, and must include at least the following physiological signals: (a3) Electroencephalogram (EEG): Frontal, central, and occipital leads; (b3) Electrooculography and mandibular electromyography: used for sleep staging and wakefulness detection; (c3) Nasal pressure sensor and chest and abdominal respiratory plethysmography: used for airflow and respiratory effort monitoring; (d3) Pulse oxygen saturation monitor: used for recording blood oxygen saturation, with a sampling frequency ≥1 Hz.

7. The method according to claim 1, characterized in that, The model training method includes the following steps: (1) Data collection Data were collected from two categories of OSA patients aged 18-70 years, excluding those with severe respiratory diseases, neuromuscular diseases, and specific sleep disorders, including COMISA, RBD, narcolepsy, and idiopathic hypersomnia. ① Clinical data: including anthropometric measurements, at least BMI and neck circumference; ② Monitoring data: PSG records were acquired for two consecutive nights. The first night was used to extract OSA internal parameters and PSG-derived parameters, and the second night was used for manual titration to obtain the minimum effective pressure of CPAP. The intrinsic parameters were extracted using the PUPBeta platform and include at least the loop gain LG1, the airway passive collapse index Vpassive, and the minimum ventilation volume Vmin. The PSG derived parameters include: AHI, AHI_REM, AHI_NREM, AI, AI_Ob, ODI_3, ODI_4, SpO2_Min, RE_OD_Mean, Ap_Dur_Mean, Ap_Dur_Long, Hyp_Dur_Mean, Hyp_Dur_Long, CT90, RR_ArI; (2) Data preprocessing (a) Missing value handling: Missing values ​​are imputed for continuous variables using the predicted mean matching method; (b) Data cleaning: Delete sample rows with all missing values ​​and convert character variables to numeric variables; (c) Dataset partitioning: The training set and validation set are randomly partitioned in an 8:2 ratio; (3) Feature Filtering A full linear regression model with CPAP as the dependent variable and all candidate indicators as independent variables was constructed based on the training set. The core features that make significant contributions to CPAP stress prediction were screened using the bidirectional stepwise regression method based on the AIC criterion. Finally, eight core feature variables were determined: LG1, Vpassive, Vmin, BMI, AI, RE_OD_Mean, Ap_Dur_Long, and RR_ArI. (4) Model training (a) Feature standardization: Only the core features obtained from the screening are centered and scaled, and the mean and standard deviation of the training set are saved for subsequent prediction stages; (b) Multi-model construction: Based on the standardized core features, the following 8 prediction models are trained respectively: i) Stepwise linear regression: A linear regression model is constructed based on selected features. The optimal feature combination is obtained through stepwise variable selection, and an interpretable prediction equation is output. ii) Ridge regression: By introducing an L2 regularization term, the model overfitting and multicollinearity are suppressed, the model's generalization ability is improved, and cross-validation is used to determine the optimal regularization parameter; iii) Lasso regression: Feature compression and variable selection are achieved by introducing an L1 regularization term, and the optimal λ value is selected by 5-fold cross-validation; iv) Decision tree regression: Construct a regression tree model based on the feature splitting criterion to achieve nonlinear relationship modeling; v) Random Forest Regression: A regression model is constructed based on the ensemble of multiple decision trees, with the optimal number of trees being 500, and variable importance assessment is enabled; vi) Gradient boosting regression: The gradient boosting framework is used to gradually optimize the residuals. The preferred distribution type is Gaussian distribution, the preferred number of trees is 500, the preferred interaction depth is 3, and the preferred learning rate is 0.

01. vii) Support Vector Regression: Construct a nonlinear regression model by selecting the radial basis kernel function and train it based on standardized features; viii) XGBoost Regression: The regression model is constructed using the extreme gradient boosting framework. The objective function is preferably squared error regression, the learning rate is preferably 0.01, the maximum tree depth is preferably 3, and the subsample ratio is preferably 0.

8. The above eight prediction models can be used individually or combined using weighted fusion, stacked fusion, or voting fusion methods. (5) Model evaluation and selection The performance of each model was evaluated using the coefficient of determination (R²), mean squared error (MSE), and / or root mean square error (RMSE) on both the training and validation sets. The prediction accuracy and generalization ability of the models were compared comprehensively. The gradient boosting regression model performed best, with an R² of 0.733 and an RMSE of 1.096 on the training set and an R² of 0.572 and an RMSE of 1.367 on the validation set, achieving a good balance between prediction accuracy and stability.

8. The method according to claim 1, characterized in that, The (II) CPAP stress prediction step includes: standardizing the eight core features of the patient to be predicted according to the mean and standard deviation corresponding to the training set, inputting them into the prediction model, and outputting the CPAP stress prediction value.

9. A system for predicting CPAP stress in OSA patients, characterized in that, include: Data acquisition module: Collects anthropometric indicators and polysomnography data of the patient to be predicted, wherein the anthropometric indicators include at least BMI; Data preprocessing module: Based on the PSG data collected by the data acquisition module, OSA intrinsic parameters and PSG derived parameters are obtained. The OSA intrinsic parameters include at least LG1, Vpassive, and Vmin, which are calculated by PUPbeta automatic analysis software. The PSG derived parameters include at least AI, RE_OD_Mean, Ap_Dur_Long, and RR_ArI, which are obtained by manual analysis of PSG data. Prediction Module: Input the feature values ​​of at least the above 8 features of the patient to be predicted, namely BMI, LG1, Vpassive, Vmin, AI, RE_OD_Mean, Ap_Dur_Long, and RR_ArI, into the prediction module, and output the corresponding CPAP pressure prediction results.

10. A device for predicting CPAP pressure in OSA patients, characterized in that, include: processor; Memory, which stores computer-readable instructions; The processor is configured to perform the method according to any one of claims 1 to 8 by executing the computer-readable instructions.

11. A computer-readable storage medium storing computer-readable instructions, characterized in that, When the computer-readable instructions are executed by a processor, they implement the method according to any one of claims 1 to 8.

Images