Construction method of hypoxemia prediction model during painless gastrointestinal endoscopy

Abstract

The invention relates to a construction method of a hypoxemia prediction model during painless gastrointestinal endoscopy. The method belongs to the technical field of medical model construction. In order to solve the problem that there is no reliable related prediction model for hypoxemia with a large sample size at present, the method comprises the steps that prediction factors related to hypoxemia risks are screened and verified through a logistic regression model of forward stepwise regression, prediction variables are determined, based on the prediction variables, an HAPPY-12K model is constructed through the logistic regression model, and the prediction factors related to the hypoxemia risks are determined through the HAPPY-12K model. And the HAPPY-12K model is verified and evaluated by adopting sample data of multiple centers and sample data of different races, so that the external validity of the model is improved. The model provided by the invention is helpful to predict and evaluate the hypoxemia of a patient during painless gastrointestinal endoscopy, and has better robustness, accuracy and clinical general applicability.

Description

Technical Field

[0001] The present invention relates to the technical field of medical model construction, and particularly to a method for constructing a hypoxemia prediction model during painless gastroscopy and colonoscopy examinations. Background Art

[0002] In recent years, the demand for gastroscopy and colonoscopy examinations has been increasing, whether for symptomatic patients or cancer screening. With the development of comfort healthcare, painless gastroscopy has emerged. Although sedative drugs during endoscopic procedures improve patient comfort and facilitate endoscopic performance, problems related to airway management cannot be ignored. Among them, the occurrence of hypoxemia deserves particular attention. Transient hypoxemia is considered a sign of an increased risk of cardiovascular complications. Even in pediatric populations with a low incidence of ischemic heart disease, hypoxemia has been identified as a sentinel event for future critical events during endoscopic examinations. Prolonged hypoxemia can lead to myocardial ischemia, malignant arrhythmias, permanent damage to the nervous system, and even death. Hypoxemia remains the most common cause of death, especially during anesthesia induction. To prevent the occurrence of adverse events, it is necessary to promptly and effectively manage hypoxemia, such as suspending the gastroscopy examination, using a face mask for ventilation, and in some cases, intubation may be required. However, the use of assisted ventilation devices is often accompanied by related complications and an increase in medical costs. Pulse oximetry is very suitable for monitoring fast-paced procedures such as outpatient painless gastroscopy, but it has a lag in assessing hypoxia. This means that the displayed oxygen saturation level may not immediately reflect the current physiological state. In some patients, it may be too late to perform effective intervention when a low pulse oximetry reading is detected. With the increase in the number of patients and comorbidities, it is necessary to comprehensively and thoroughly evaluate the risk factors for hypoxemia during painless gastroscopy.

[0003] Existing studies have reported that the incidence of hypoxemia during digestive endoscopy ranges from 5.3% to 50%. This difference may be due to different definitions of hypoxemia, study populations, and sample sizes. However, it is undeniable that hypoxemia is a very common event and can easily cause serious complications during painless gastroscopy. In painless gastroscopy, the lack of airway protection and the unique body position required for the examination pose challenges to airway management. Anticipated and unanticipated difficult airways, as well as obstructive sleep apnea-hypopnea syndrome (OSAHS), significantly increase the risk of airway obstruction and hypoxemia.

[0004] Although painless gastroscopy provides a more comfortable examination experience for patients, the risk of hypoxemia cannot be ignored when sedatives or anesthetic drugs are used in a non-intubated state. Therefore, early identification of high-risk populations for hypoxemia, formulation of personalized sedation plans, or diversion to anesthesia in the operating room not only improves the safety of patient examinations but also saves medical resources. To this end, constructing a risk factor prediction model for hypoxemia has important clinical significance. Summary of the Invention

[0005] Aiming at the defects existing in the above-mentioned prior art, the present invention aims to provide a method for constructing a hypoxemia prediction model during painless gastroscopy and colonoscopy, so as to solve the problems existing in the above-mentioned background technology section.

[0006] To achieve the above object, the present invention adopts the following technical solutions:

[0007] A method for constructing a hypoxemia prediction model during painless gastroscopy and colonoscopy, comprising the following steps:

[0008] Step 1: Use a logistic regression model with forward stepwise regression to screen and verify the predictive factors related to the risk of hypoxemia, and determine the predictive variables;

[0009] Step 2: Based on the determined predictive variables, use a logistic regression model to construct the HAPPY-12K model;

[0010] Step 3: Verify and evaluate the predictive performance of the HAPPY-12K model.

[0011] As a further preferred solution of the above technical solution: in Step 1, the predictive variables include main effect predictive factors and interaction effect predictive factors. The steps of screening and verifying the predictive factors related to the risk of hypoxemia include:

[0012] Step 11: Collect clinical data and establish a sample set, and divide the sample set into a training set and multiple validation sets;

[0013] Step 12: Use the forward stepwise regression method to screen the predictive variables in the training set, and verify the predictive variables through the validation sets within a single center to determine the main effect predictive factors related to the risk of hypoxemia;

[0014] Step 13: Use the determined main effect predictive factors as covariates, continue to use the forward stepwise regression method to screen for interactions, and verify using the validation sets within a single center to determine the interaction effect predictive factors related to the risk of hypoxemia.

[0015] A further preferred solution is as follows: In Step 12, the main effect predictors determined to be related to the risk of hypoxemia include short and thick neck, BMI, Mallampati grade, limited jaw protrusion, pre-operative mean arterial pressure, history of snoring, and thyromental distance.

[0016] A further preferred solution is as follows: In Step 13, the interaction effect predictor determined to be related to the risk of hypoxemia is tongue fat × thyromental distance.

[0017] As a further preferred solution to the above technical solution: In Step 2, after determining the main effect predictors and interaction effect predictors related to the risk of hypoxemia, a multi-factor logistic regression model is constructed. By maximizing the likelihood function, the estimated values of the regression coefficients of the main effect predictors and interaction effect predictors in the training set are obtained. The regression coefficients reflect the influence degrees of the main effect predictors and interaction effect predictors on the result analogy, thereby constructing the HAPPY-12K model. The expression of the constructed HAPPY-12K model is as follows:

[0018] HAPPY-12K_Score = -9.9521 + 1.2866×short_and_thick_neck + 0.1645×bmi

[0019] + 0.7046×mallampati_grade + 1.9627×limited_jaw_protrusion

[0020] + 0.0210×pre-operative_mean_arterial_pressure

[0021] + 0.5527×history_of_snoring + 1.7891×thyromental_distance

[0022] + 0.3922×tongue_fat - 2.7220×thyromental_distance×tongue_fat

[0023] Where: short_and_thick_neck represents a short and thick neck; mallampati_grade represents the Mallampati classification; limited_jaw_protrusion represents micrognathia; pre-operative_mean_arterial_pressure represents the pre-operative mean arterial pressure; history_of_snoring represents a history of snoring; thyromental_distance represents the thyromental distance; tongue_fat represents a large tongue.

[0024] As a further preferred solution to the above technical solution: In step Step 3, verify and evaluate the prediction performance of the HAPPY-12K model, including the following steps:

[0025] S31: Verify the AUC, ROC, EO ratio, and calibration curve of the HAPPY-12K model through a training set and multiple validation sets to determine the robustness of the HAPPY-12K model;

[0026] S32: Establish an ROC curve based on the best HAPPY-12K model, compare the AUC determined by the ROC curve with the AUC of other models, and obtain the prediction performance of the HAPPY-12K model and other models for hypoxemia;

[0027] S33: Conduct a decision curve analysis on the HAPPY-12K model, and then evaluate the HAPPY-12K model.

[0028] Decision curve analysis is a method for directly evaluating the clinical value of a prediction model, and it calculates the clinical net benefit (NB) of allocating intervention measures based on the model. Where: The calculation method of NB is: TP and FP are the number of true positives and false positives respectively, n is the sample size of the study, and P t is the acceptable level of the predicted probability of adverse events that are considered necessary for intervention.

[0029] The clinical value can also be represented by the net reduction (NR) of the intervention, that is, the number of unnecessary interventions compared with the "universal intervention" strategy under the guidance of the HAPPY-12K model. Where: The calculation expression of NR is:

[0030]

[0031] In the formula, NB Model is the NB of the HAPPY-12K model, and NB All is the NB of the "universal intervention" strategy.

[0032] In addition, the present invention also provides an electronic device, which includes at least one processor and a memory communicatively connected to the processor; wherein, the memory stores instructions executable by the processor, and when the instructions are executed by the processor, the processor is enabled to execute the method described above.

[0033] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0034] 1. Based on a large amount of sample data, the present invention uses a logistic regression model to screen and verify the predictive factors related to hypoxemia, constructs the HAPPY-12K model for predicting hypoxemia during painless gastroscopy using the determined predictive factors, and trains the model using datasets from multiple centers and different ethnic groups to ensure the robustness and accuracy of the model and improve the general applicability of the model.

[0035] 2. By comparing the predictive values of different models for hypoxemia, the present invention helps doctors provide personalized treatment strategies for high-risk groups, effectively allocate resources, and thus improve the quality and efficiency of medical services. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 is a flowchart of a method for constructing a hypoxemia prediction model during painless gastrointestinal endoscopy according to the present invention;

[0037] Figure 2 is a drawing of the ROC curve of different models using the training set according to the present invention;

[0038] Figure 3 is a drawing of the ROC curve of different models using validation set 1 according to the present invention;

[0039] Figure 4 is a drawing of the ROC curve of different models using validation set 2 according to the present invention;

[0040] Figure 5 is a drawing of the ROC curve of different models using validation set 3 according to the present invention;

[0041] Figure 6 is a drawing of the ROC curve of different models using validation set 4 according to the present invention;

[0042] Figure 7 is a drawing of the ROC curve of different models using validation set 5 according to the present invention;

[0043] Figure 8 are the results such as AUC, sensitivity, specificity, accuracy, etc. of different models at the optimal threshold of the training set;

[0044] Figure 9Results such as AUC, sensitivity, specificity, and accuracy of different models at the optimal threshold of Validation Set 1;

[0045] Figure 10 Results such as AUC, sensitivity, specificity, and accuracy of different models at the optimal threshold of Validation Set 2;

[0046] Figure 11 Results such as AUC, sensitivity, specificity, and accuracy of different models at the optimal threshold of Validation Set 3;

[0047] Figure 12 Results such as AUC, sensitivity, specificity, and accuracy of different models at the optimal threshold of Validation Set 4;

[0048] Figure 13 Results such as AUC, sensitivity, specificity, and accuracy of different models at the optimal threshold of Validation Set 5;

[0049] Figure 14 Results such as AUC, sensitivity, specificity, and accuracy of different models at the optimal threshold of the combined data. Detailed implementation manners

[0050] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Apparently, the described embodiments are some, but not all, of the embodiments of the present invention.

[0051] Refer to Figures 1-14 , the present invention proposes a method for constructing a hypoxemia prediction model during painless gastroscopy and colonoscopy, and uses sample data from multiple centers to verify the stability of the constructed prediction model. The specific steps for establishing this prediction model are as follows:

[0052] Step 1: Use a logistic regression model with forward stepwise regression to screen and verify the predictive factors related to the risk of hypoxemia, and determine the predictive variables;

[0053] Step 2: Based on the determined predictive variables, construct the HAPPY-12K model using a logistic regression model;

[0054] Step 3: Verify and evaluate the predictive performance of the HAPPY-12K model.

[0055] Specifically, the clinical data of patients who underwent gastrointestinal endoscopy in multiple centers from May 2020 to November 2023 were collected as a sample set. In the present invention, the clinical data of patients from 7 centers including the First Affiliated Hospital of Nanjing Medical University (Jiangsu Provincial People's Hospital), Nanjing First Hospital, Sir Run Run Shaw Hospital Affiliated to Nanjing Medical University, Changzhou Second People's Hospital, People's Hospital of Kizilsu Kirgiz Autonomous Prefecture in Xinjiang, Northern Jiangsu People's Hospital, and Taizhou People's Hospital in Jiangsu were collected. The first four hospitals represent the southern region of Jiangsu, and the last two hospitals represent the northern region of Jiangsu. All participants in these six hospitals are Han ethnicity, and all participants in the People's Hospital of Kizilsu Kirgiz Autonomous Prefecture in Xinjiang are Uyghur ethnicity. The research of the present invention has been approved by the hospital ethics committee (Clinical trials: ChiCTR2000032801, ChiCTR2300074128). The main criterion for judgment is whether hypoxemia occurs, which is defined as an oxygen saturation (SpO2) below 95% for a duration exceeding 10 seconds, while the judgment criterion of Nanjing First Hospital is defined as SpO2 < 95%.

[0056] (I) Inclusion and exclusion criteria for patients:

[0057] Eligible participants must have a comprehensive preoperative assessment, laboratory tests, complete medical records, and relevant medical history records. The included patients are aged ≥ 18 years, with an American Society of Anesthesiologists (ASA) physical status classification score of I-III, and scheduled for gastrointestinal endoscopy. The exclusion criteria are: (i) Patients with a history of severe respiratory insufficiency. (ii) Individuals with severe important organ dysfunction. (iii) Patients with mental disorders. (iv) Those known to be allergic to propofol or emulsifiers. (v) Patients with acute upper gastrointestinal bleeding.

[0058] (II) Patient admission examinations and anesthesia:

[0059] After all patients entered the outpatient operating room, they all maintained a left lateral position with their knees bent to facilitate gastroscopy and colonoscopy. Standard monitors were used to record vital signs: non-invasive blood pressure monitoring (NIBP), electrocardiogram monitoring (ECG), and pulse oximetry (SpO2). All patients were provided with oxygen supplementation via a double nasal catheter (5 - 7 L / min for 3 minutes). In the above-mentioned 7 centers of the present invention, it is a common practice to sedate patients receiving propofol alone or in combination with low-dose opioids and / or etomidate, with an induction dose of propofol of 1 - 2 mg / kg (the induction dose for obese patients was given according to their ideal body weight). After the anesthetic depth reached a level where the modified Observer's Assessment of Alertness / Sedation score was 1 point, the endoscopist began the examination. During the operation, propofol was added according to the patient's pain response, with a supplementary dose of 10 - 20 mg. In cases where SPO2 < 95% persisted for more than 10 seconds, the anesthesiologist provided necessary airway support, including jaw thrust, mask ventilation, insertion of an oropharyngeal / nasopharyngeal airway, and tracheal intubation if necessary.

[0060] (III) Medical records and relevant medical history of patients:

[0061] Recorded patient variables included gender, age, ASA grade, height, weight, neck circumference, snoring history, and some significant underlying diseases (hypertension, diabetes, coronary heart disease, asthma, chronic obstructive pulmonary disease, upper respiratory tract infection), electrocardiogram diagnosis (normal or abnormal), chest CT diagnosis (pulmonary inflammation, etc., baseline heart rate, systolic blood pressure, and pulse oxygen saturation).

[0062] 1. Items related to the STOP - Bang questionnaire: snoring, fatigue, observable apnea, hypertension, BMI, age, neck circumference, male. The total score ranges from 0 to 8 points.

[0063] 2. Items related to the NoSAS questionnaire: neck circumference > 40 cm, 4 points; BMI between 25 - 30, 3 points; BMI ≥ 30, 5 points; snoring, 2 points; age > 55 years old, 4 points; male, 2 points. If the threshold is 8 points or above, it indicates a high risk of OSA.

[0064] 3. The modified Mallampati class (MMP) includes four categories: Class I, everything visible (tonsillar pillars); Class II, the uvula is fully visible; Class III, only the soft palate and the base of the uvula are visible; and Class IV, the soft palate is not visible.

[0065] 4. Difficult mask ventilation (DMV): Presence of two or more of the following factors: age > 55 years old, BMI > 26 kg / m², presence of a beard, missing teeth, snoring history.

[0066] 5. Difficult intubation (DI): Presence of one or more of the following factors: Mallampati classification > II, limited protrusion of the mandible, limited distance to the thyroid gland (<6.5 cm), limited mouth opening (<30 mm).

[0067] (IV) Conditions of the included patients:

[0068] According to the above requirements, the present invention included a total of 11,957 participants, including 5,888 males and 6,069 females, with an average age of 53.43 ± 14.11. All the included patients were divided into a training set (n = 2,518) and five validation sets (n = 9,439). In the training set (the first batch of data from the First Affiliated Hospital of Nanjing Medical University), among the 2,518 participants, 261 experienced hypoxemia after recruitment. In validation set 1 (the second batch of data from the First Affiliated Hospital of Nanjing Medical University), among the 2,327 participants, there were 242 hypoxemia events. In validation set 2 (the third batch of data from the First Affiliated Hospital of Nanjing Medical University), among the 2,179 participants, 227 hypoxemia events occurred. In validation set 3 (hospitals from four centers in southern Jiangsu), among the 4,227 participants, there were 816 hypoxemia events. In validation set 4 (hospitals from two centers in northern Jiangsu), among the 587 participants, 167 experienced hypoxemia. In validation set 5 (People's Hospital of Korla Prefecture, Xinjiang), there were 119 participants, among whom 10 had hypoxemia events.

[0069] In the training set from the First Affiliated Hospital of Nanjing Medical University, 32 variables were screened using the forward stepwise regression method of the logistic regression model to determine the predictive factors related to the risk of hypoxemia. Then, the selected predictive factors were verified using the internal validation set (i.e., validation set 1) from the First Affiliated Hospital of Nanjing Medical University, and finally, seven main effect predictive factors related to the risk of hypoxemia with consistent directions were determined, as shown in Table 1.

[0070] Table 1 Main effect predictive factors related to the risk of hypoxemia

[0071]

[0072] Then, using the above - determined seven main effect predictive factors as covariates, the forward stepwise regression method of the logistic regression model was continued to screen all possible interactions. These potential interactions were further verified using validation set 1, and finally, a pair of interaction effect predictive factors related to the risk of hypoxemia with consistent directions were determined, as shown in Table 2. The predictive factors were finally retained if they met the following conditions: (i) P ≤ 0.05 in the screening stage; (ii) P ≤ 0.05 in the validation stage; (iii) the direction of influence was consistent in both stages.

[0073] Table 2 Main effect predictors and interaction effect predictors related to the risk of hypoxemia

[0074]

[0075] Finally, a logistic regression model was constructed in the training set with the identified 7 main effect predictors and a pair of interaction effect predictors. The 7 main effect predictors included short and thick neck, BMI, Mallampati classification, limited jaw protrusion, preoperative mean arterial pressure, history of snoring, and thyromental distance. The pair of interaction effect predictors was tongue fat × thyromental distance.

[0076] After determining the main effect predictors and interaction effect predictors related to the risk of hypoxemia, a multivariate logistic regression model was constructed. The estimated values of the regression coefficients of the main effect predictors and interaction effect predictors under the training set were obtained by maximizing the likelihood function. The regression coefficients reflected the influence degree of the main effect predictors and interaction effect predictors on the result analogy, thus constructing the HAPPY-12K model, and its expression is as follows:

[0077] HAPPY-12K_Score = -9.9521 + 1.2866 × short_and_thick_neck + 0.1645 × bmi

[0078] + 0.7046 × mallampati_grade + 1.9627 × limited_jaw_protrusion

[0079] + 0.0210 × pre-operative_mean_arterial_pressure

[0080] + 0.5527 × history_of_snoring + 1.7891 × thyromental_distance

[0081] + 0.3922 × tongue_fat - 2.7220 × thyromental_distance × tongue_fat

[0082] In the formula: short_and_thick_neck represents short and thick neck; mallampati_grade represents Mallampati classification; limited_jaw_protrusion represents limited jaw protrusion; pre-operative_mean_arterial_pressure represents preoperative mean arterial pressure; history_of_snoring represents history of snoring; thyromental_distance represents thyromental distance; tongue_fat represents tongue fat.

[0083] Compared with non - hypoxemic patients, hypoxemic patients have more obesity signs (short and thick neck, excessive tongue fat, BMI) (P < 0.05 value increases). Obese patients are one of the most challenging patient groups in airway management. BMI is an important risk factor for anesthesia - and sedation - related adverse events, and BMI is significantly correlated with the frequency of hypoxic attacks. There are several main reasons why obese patients are prone to hypoxia during painless gastroscopy. First, many obese patients have a reduced mouth opening, increased submental fat and pharyngeal tissue, and increased tongue fat. All these factors can lead to airway stenosis, which is consistent with the independent risk factors found in the present invention. In addition, fat deposition around the neck and chest regions can lead to limited neck mobility, manifested as limited thyroid distance with a short and thick neck (short and thick neck), and increased MMP classification. In obese patients, the increase in neck circumference and thyroid distance, as well as the increase in Mallampati classification, may increase the risk of complications during surgical sedation. In addition, due to the accumulation of fat on the chest wall, the lung compliance, functional residual capacity, total lung volume, total lung capacity, and expiratory reserve volume of obese patients are all reduced, resulting in a rapid decline in SPO2 in obese patients during short - term apnea or hypoventilation. In fact, an increased neck circumference (short and thick neck) indicates excessive palatal and pharyngeal soft tissues, snoring during sleep, and exacerbates oropharyngeal collapse under sedation. Therefore, patients with a history of snoring may be more prone to hypoxemia under anesthesia, which is consistent with the independent factor in the present invention that identifies a history of snoring as a risk factor for hypoxemia.

[0084] An ROC curve was established using the HAPPY - 12K model, and the predictive ability of the model for hypoxemia was judged according to the area under the ROC curve. In the training set of the present invention, the incidence of hypoxemia was 10.37% (261 vs 2518). In validation sets 1 - 5, the incidences of hypoxemia were 10.39% (242 vs 2327), 10.42% (227 vs 2179), 19.30% (816 vs 4227), 28.45% (167 vs 587), and 8.40% (10 vs 119) respectively. The results showed that all 8 predictive factors identified in the present invention related to the risk of hypoxemia had a high negative predictive value (0.9608), the sensitivity of the HAPPY - 12K model was 71.7%, the specificity was 80.4%, and the area under the ROC was 0.830 (0.8015 - 0.8583).

[0085] Considering the risk factors related to sleep apnea syndrome, mask ventilation, and endotracheal intubation, the HAPPY-12K model of the present invention was evaluated and compared with five other existing models. These models include the STOP-Bang questionnaire, the NoSAS score, difficult mask ventilation (DMV), difficult intubation (DI), and NoSAS combined with the modified Mallampati classification (NoSAS+MMP). In addition, model establishment based only on variables without obtaining scores was also considered, namely the NoSAS*, DMV*, and DI* models. These three models include one or several variables related to the basic, diagnostic, and monitoring aspects, such as Figures 2-14 As shown, the discriminatory ability and calibration ability of the above existing models and the HAPPY-12K model of the present invention are compared below.

[0086] (I) Comparison of discriminatory ability:

[0087] The discriminatory ability of the model prediction is represented by the receiver operating characteristic (ROC) curve of the subjects. The area under the curve (AUC) (determined by the "pROC" package in r), and the 95% confidence interval (CI) of the AUC is calculated through 2000 stratified bootstrap replications. In addition, sensitivity, specificity, positive predictive value, negative predictive value, and accuracy are used to evaluate the discriminatory ability of the HAPPY-12K of the present invention.

[0088] As Figure 2 、 Figure 8 shown, in the training dataset, the AUC of the HAPPY-12K model is 0.830 (95% CI: 0.802-0.858), the negative predictive value is 0.9608, the sensitivity is 0.717, and the specificity is 0.804. The AUC of the NoSAS+MMP model is 0.813 (95% CI: 0.7842-0.8427), and the negative predictive value and specificity are 0.9552 and 0.8321 respectively. The AUC of the NoSAS model is 0.795 (95% CI: 0.7645-0.8256). The discrimination rate of the NoSAS* model is 0.803 (95% CI: 0.7723-0.8327), the negative predictive value is 0.9511, and the specificity drops to 0.826. The AUC of the STOP-Bang model is 0.753 (95% CI: 0.7174-0.788), the high negative predictive value is 0.9416, and the specificity is as high as 0.8861. Therefore, high-risk patients with sleep apnea syndrome, especially those with a higher Mallampati score, may face a higher risk of hypoxemia during propofol anesthesia.

[0089] Difficult mask ventilation and difficult tracheal intubation are risk factors for hypoxemia during general anesthesia induction. Difficult mask ventilation is an important predictor of airway complications. In the training set, the AUC of the DMV model was 0.745 (95% CI: 0.7136 - 0.7768). The negative predictive value of DMV was relatively high, at 0.9574, but the specificity was only 0.6172. While the AUC of the DMV* model was 0.780 (95% CI: 0.7494 - 0.8112), and the specificity of the DMV* model could reach 0.8401. It can be seen that the AUC, specificity, and negative predictive value of the DMV model and the DMV* model are all lower than those of the HAPPY-12K model. This may be because the DMV model does not include factors related to the imbalance of the internal structure of the pharynx, such as tongue fat, thyromental distance, and mouth opening, but these indicators are all potential risk factors for DMV and help to determine the degree of DMV.

[0090] In addition to the factors related to DMV, the predictive value of difficult tracheal intubation for hypoxemia-related indicators was also studied. The discrimination rate of the DEI model was 0.673 (95% CI: 0.6406 - 0.7056). The AUC of the DEI* model was 0.680 (95% CI: 0.647 - 0.7133), the negative predictive value was 0.943, but the specificity was relatively low, only 0.6464.

[0091] In addition, similar results were obtained in the validation sets of the above existing models. The AUCs of the HAPPY-12K model in validation sets 1 to 5 were: 0.861, 0.856, 0.826, 0.775, and 0.941, respectively. Referring to Figures 3-7 、 Figures 9-13 as shown, the AUCs of the combined data were 0.841 as seen Figure 14 。

[0092] (2) Calibration ability of the HAPPY-12K model of the present invention:

[0093] To evaluate the calibration ability of the model, the expected number of hypoxemia events was compared with the number of actually observed hypoxemia events, and the expected-observed ratio (i.e., EO ratio) was calculated. Calibration was evaluated by plotting the ratio of the mean predicted event count to the mean observed event count (calibration curve). In the training set, the HAPPY-12K model showed good consistency (EO ratio = 1.000, 95% CI: 0.897 - 1.130); in validation sets 1 and 2, the HAPPY-12K model showed the best calibration (EO ratio = 0.962, 95% CI: 0.860 - 1.093; EO ratio = 0.962, 95% CI: 0.857 - 1.097), validation set 4 yielded good results (EO ratio = 0.728, 95% CI: 0.646 - 0.836), while validation sets 3 and 5 did not show obvious calibration.

[0094] Obesity is often accompanied by hypertension, metabolic syndrome, and obstructive sleep apnea. It is estimated that at least 75% of the incidence of hypertension is directly related to obesity. Combining the prediction results of the present invention, it was found that in the hypoxemia group, 37.5% of the patients had a definite history of hypertension, and 61.22% of the hypertensive patients had a BMI > 26. Through multivariate analysis, high MAP_pre was associated with the incidence of hypoxemia. Compared with non-obese patients, the circulating blood volume, cardiac output, and post-obesity load of obese patients were all significantly increased. In particular, patients with sleep apnea syndrome are exposed to repeated nocturnal hypoxemia, with increased peripheral chemoreflex sensitivity, high sympathetic drive, and an increased risk of hypertension. Sustained nocturnal hypoxemia and obesity can remodel the myocardium, reduce oxygen supply, activate cardiac sympathetic nerves, and damage cardiac function, ultimately potentially leading to myocardial hypertrophy and chronic heart disease. At the same time, the above conditions in turn can limit the physical activity of obese patients, thereby potentially leading to further weight gain and an increased likelihood of developing hypertension. Obesity, hypoxemia, and hypertension interact with each other and promote each other's development.

[0095] The above has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. What is described in the above embodiments and the specification only illustrates the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements all fall within the scope of the present invention claimed. The scope of the present invention claimed is defined by the appended claims and their equivalents.

Claims

1. A method for constructing a prediction model of hypoxemia during painless gastroenteroscopy, characterized in that It includes the following steps: Step 1: Use a logistic regression model with forward stepwise regression to screen and validate predictors related to the risk of hypoxemia, and determine the predictive variables; Step 2: Based on the determined predictive variables, construct the HAPPY-12K model using a logistic regression model; Step 3: Validate and evaluate the predictive performance of the HAPPY-12K model.

2. The construction method of a hypoxemia prediction model during painless gastroscopy and colonoscopy according to claim 1, characterized in that, In Step 1, the predictive variables include main effect predictors and interaction effect predictors. The steps to screen and validate predictors related to the risk of hypoxemia include: Step 11: Collect clinical data and establish a sample set, and divide the sample set into a training set and multiple validation sets; Step 12: Use the forward stepwise regression method to screen predictive variables in the training set, and validate the predictive variables through the validation sets within a single center to determine the main effect predictors related to the risk of hypoxemia; Step 13: Use the determined main effect predictors as covariates, continue to use the forward stepwise regression method to screen for interactions, and use the validation sets within a single center for validation to determine the interaction effect predictors related to the risk of hypoxemia.

3. The construction method of a hypoxemia prediction model during painless gastroscopy and colonoscopy according to claim 2, wherein, In Step 12, the determined main effect predictors related to the risk of hypoxemia include short and thick neck, BMI, Mallampati grade, limited jaw protrusion, pre-operative mean arterial pressure, history of snoring, and thyromental distance.

4. The method for constructing a hypoxemia prediction model during painless gastroscopy and colonoscopy according to claim 3, wherein, In Step 13, the determined interaction effect predictor related to the risk of hypoxemia is tongue fat × thyromental distance.

5. The construction method of a hypoxemia prediction model during painless gastroenteroscopy according to claim 4, wherein, In Step 2, after determining the main effect predictors and interaction effect predictors related to the risk of hypoxemia, construct a multi-factor logistic regression model, and obtain the estimated values of the regression coefficients of the main effect predictors and interaction effect predictors under the training set by maximizing the likelihood function. This regression coefficient reflects the degree of influence of the main effect predictors and interaction effect predictors on the result analogy, thereby constructing the HAPPY-12K model. The expression of the constructed HAPPY-12K model is as follows: HAPPY-12K_Score = -9.9521 + 1.2866×short_and_thick_neck + 0.1645×bmi + 0.7046×mallampati_grade + 1.9627×limited_jaw_protrusion + 0.0210×pre-operative_mean_arterial_pressure + 0.5527×history_of_snoring + 1.7891×thyromental_distance + 0.3922×tongue_fat - 2.7220×thyromental_distance×tongue_fat Where: short_and_thick_neck represents a short and thick neck; mallampati_grade represents the Mallampati classification; limited_jaw_protrusion represents micrognathia; pre-operative_mean_arterial_pressure represents the pre-operative mean arterial pressure; history_of_snoring represents a history of snoring; thyromental_distance represents the thyromental distance; tongue_fat represents macroglossia.

6. The construction method of a hypoxemia prediction model during painless gastroscopy and colonoscopy according to claim 5, characterized in that, In Step 3, the prediction performance of the HAPPY-12K model is verified and evaluated, including the following steps: S31: Verify the AUC, ROC, EO ratio, and calibration curve of the HAPPY-12K model through a training set and multiple validation sets to determine the robustness of the HAPPY-12K model; S32: Establish an ROC curve based on the best HAPPY-12K model, compare the AUC determined by the ROC curve with the AUC of other models, and obtain the prediction performance of the HAPPY-12K model and other models for hypoxemia; S33: Conduct a decision curve analysis on the HAPPY-12K model, and then evaluate the HAPPY-12K model.

7. An electronic device, characterized in that: The electronic device includes at least one processor and a memory communicatively connected to the processor; wherein, the memory stores instructions executable by the processor, and when the instructions are executed by the processor, the processor is enabled to execute the method according to any one of claims 1-6.

Images