Intelligent electronic bronchoscopy patient hypoxemia risk prediction system
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-15
- Publication Date
- 2026-08-14
AI Technical Summary
[0003]针对上述情况，为克服现有技术的缺陷，本发明提供了智能化电子支气管镜检查患者低血氧发生风险预测系统，针对一般低血氧发生风险预测系统存在对临床风险区分度不足，对漏诊的临床代价量化不足，易受术中监测噪声影响，进而导致风险预测可靠性差的问题，本方案通过定义临床风险代价，重点强化对漏诊场景的惩罚权重，针对性强化内镜漏诊惩罚；并引入临界风险区间放大机制，聚焦内镜易漏诊样本；将梯度幅值与临床风险强绑定，实现高危强更、低危微调；引入平滑符号函数适配内镜监测噪声环境；进而提高低血氧发生风险预测可靠性；针对一般低血氧发生风险预测系统存在未耦合内镜场景的临界风险特性，无法适配内镜个体化临床需求，进而导致风险预测效果差的问题，本方案通过临界风险区间放大项结合增强损失的差值平方项，双重放大电子支气管镜术中临界风险区间的损失差，有效降低临床漏检率；引入风险平滑因子与风险导向指数构建风险导向损失，实现惩罚力度灵活可调；梯度大小直接由电子支气管镜低血氧风险偏差决定，优先规避漏诊高危风险；进而提高低血氧发生风险预测效果
[0019]（1）针对一般低血氧发生风险预测系统存在对临床风险区分度不足，对漏诊的临床代价量化不足，易受术中监测噪声影响，进而导致风险预测可靠性差的问题，本方案通过定义临床风险代价，重点强化对漏诊场景的惩罚权重，针对性强化内镜漏诊惩罚；并引入临界风险区间放大机制，聚焦内镜易漏诊样本；将梯度幅值与临床风险强绑定，实现高危强更、低危微调；引入平滑符号函数适配内镜监测噪声环境；进而提高低血氧发生风险预测可靠性。
Smart Images
Figure CN122575747A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of data processing, specifically to an intelligent system for predicting the risk of hypoxemia in patients undergoing electronic bronchoscopy. Background Technology
[0002] Hypoxemia risk prediction systems are intelligent systems that collect relevant data on human physiological indicators, process the data, and use models to predict the likelihood of hypoxemia events. However, general hypoxemia risk prediction systems suffer from insufficient differentiation of clinical risks, inadequate quantification of the clinical costs of missed diagnoses, and susceptibility to intraoperative monitoring noise, leading to poor reliability in risk prediction. Furthermore, these systems are often not coupled with the critical risk characteristics of endoscopic scenarios, failing to adapt to individualized clinical needs in endoscopy, resulting in poor risk prediction performance. Summary of the Invention
[0003] To address the above issues and overcome the shortcomings of existing technologies, this invention provides an intelligent system for predicting the risk of hypoxemia in patients undergoing electronic bronchoscopy. Addressing the problems of insufficient clinical risk differentiation, inadequate quantification of the clinical cost of missed diagnoses, and susceptibility to intraoperative monitoring noise, leading to poor prediction reliability, this solution defines clinical risk costs, emphasizes the penalty weighting for missed diagnoses, and specifically strengthens the penalty for endoscopic missed diagnoses. It also introduces a critical risk range amplification mechanism to focus on samples prone to endoscopic missed diagnoses; strongly binds gradient amplitude to clinical risk, enabling stronger adjustments for high-risk cases and fine-tuning for low-risk cases; and introduces a smoothing sign function to adapt to endoscopic monitoring noise loops. This approach aims to improve the reliability of hypoxemia risk prediction. Addressing the issue that general hypoxemia risk prediction systems often lack coupling with endoscopic scenarios, failing to meet individualized clinical needs and resulting in poor prediction performance, this solution combines a critical risk interval amplification term with the squared difference term of the enhancement loss to double-amplify the loss difference within the critical risk interval during electronic bronchoscopy, effectively reducing the clinical missed detection rate. Furthermore, it introduces a risk smoothing factor and a risk-oriented index to construct a risk-oriented loss, allowing for flexible adjustment of the penalty intensity. The gradient magnitude is directly determined by the hypoxemia risk deviation during electronic bronchoscopy, prioritizing the avoidance of high-risk missed diagnoses, thereby improving the overall effectiveness of hypoxemia risk prediction.
[0004] The technical solution adopted by the present invention is as follows: The intelligent electronic bronchoscopy patient hypoxia risk prediction system provided by the present invention includes a monitoring data acquisition module, a hypoxia risk cost definition module, a risk-oriented constraint design module, a risk-oriented loss construction module, a risk gradient stabilization calculation module, a hypoxia risk prediction model design module, and a hypoxia risk prediction module.
[0005] The monitoring data acquisition module acquires historical hypoxemia test data and preprocesses it to obtain a monitoring dataset.
[0006] The hypoxia risk cost definition module defines the hypoxia risk deviation cost by combining a logarithmic difference structure with a critical risk interval amplification mechanism.
[0007] The risk-oriented constraint design module designs gradient constraint rules based on the risk cost of low blood oxygenation.
[0008] The risk-oriented loss construction module constructs a risk-oriented loss baseline and an enhancement loss based on gradient constraints, and introduces a risk smoothing factor and a risk-oriented index to construct the risk-oriented loss.
[0009] The risk gradient stabilization calculation module obtains the updated gradient by differentiating the risk-oriented loss with the clinical smoothing factor.
[0010] The fully connected deep network built by the hypoxemia risk prediction model design module uses the monitoring dataset as input and risk-oriented loss as the optimization objective to design a hypoxemia risk prediction model.
[0011] The hypoxemia risk prediction module is based on a trained hypoxemia risk prediction model to achieve real-time hypoxemia risk prediction.
[0012] Furthermore, the hypoxia risk cost definition module defines clinical risk costs to reflect the degree of deviation between the model's predicted risk and the actual hypoxia event, and introduces a critical risk interval amplification mechanism to focus on critical risk samples.
[0013] Furthermore, the risk-oriented constraint design module designs gradient constraint rules in reverse, forcing the model gradient to be proportional to the risk cost.
[0014] Furthermore, the risk-oriented loss construction module, based on the guiding gradient constraint, first constructs a risk-oriented loss baseline, then introduces a difference square term to amplify the loss difference in the critical interval, expanding it into an enhanced risk-oriented loss; and introduces a risk smoothing factor, which is equivalent to the natural logarithm in forward calculation, and the gradient is always 1 during backpropagation. The risk-oriented index is used to adjust the risk deviation penalty intensity to establish the final risk-oriented loss.
[0015] Furthermore, the risk gradient stabilization calculation module calculates the derivative of the risk-oriented loss with respect to the clinical smoothing factor to obtain the final update gradient.
[0016] Furthermore, the hypoxemia risk prediction model design module constructs a fully connected deep network, takes the monitoring dataset as input, uses risk-oriented loss as the optimization objective, and employs mini-batch stochastic gradient descent for backpropagation to update the network parameters, thereby obtaining the hypoxemia risk prediction model.
[0017] Furthermore, the hypoxemia risk prediction module acquires hypoxemia examination data in real time based on the trained hypoxemia risk prediction model, and inputs it into the hypoxemia risk prediction model after preprocessing to obtain the current patient's hypoxemia risk probability; and performs clinical risk stratification and early warning based on the predicted probability.
[0018] The beneficial effects achieved by the present invention using the above solution are as follows:
[0019] (1) In view of the problems that general hypoxemia risk prediction systems have insufficient differentiation of clinical risks, insufficient quantification of the clinical cost of missed diagnosis, and are easily affected by intraoperative monitoring noise, resulting in poor reliability of risk prediction, this solution defines the clinical risk cost, focuses on strengthening the penalty weight of missed diagnosis scenarios, and specifically strengthens the penalty for missed endoscopic diagnoses; introduces a critical risk interval amplification mechanism to focus on samples that are prone to missed endoscopic diagnoses; strongly binds the gradient amplitude with clinical risk to achieve strong adjustment for high risk and fine adjustment for low risk; introduces a smooth sign function to adapt to the noise environment of endoscopic monitoring; thereby improving the reliability of hypoxemia risk prediction.
[0020] (2) In view of the problem that the general hypoxemia risk prediction system has the critical risk characteristics that are not coupled with the endoscopy scenario, and cannot adapt to the individualized clinical needs of endoscopy, thus resulting in poor risk prediction effect, this solution combines the critical risk interval amplification term with the difference square term of the enhancement loss to double amplify the loss difference of the critical risk interval during electronic bronchoscopy, effectively reducing the clinical missed detection rate; introduces risk smoothing factor and risk guidance index to construct risk guidance loss, so as to realize flexible adjustment of the penalty intensity; the gradient size is directly determined by the hypoxemia risk deviation of electronic bronchoscopy, giving priority to avoiding high-risk missed diagnosis; thereby improving the prediction effect of hypoxemia risk. Attached Figure Description
[0021] Figure 1 This is a flowchart illustrating the intelligent electronic bronchoscopy patient hypoxemia risk prediction system provided by the present invention.
[0022] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation
[0023] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.
[0024] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.
[0025] Example 1, see Figure 1 The intelligent electronic bronchoscopy patient hypoxia risk prediction system provided by the present invention includes a monitoring data acquisition module, a hypoxia risk cost definition module, a risk-oriented constraint design module, a risk-oriented loss construction module, a risk gradient stabilization calculation module, a hypoxia risk prediction model design module, and a hypoxia risk prediction module.
[0026] The monitoring data acquisition module acquires historical hypoxemia test data and preprocesses it to obtain a monitoring dataset.
[0027] The hypoxia risk cost definition module defines the hypoxia risk deviation cost by combining a logarithmic difference structure with a critical risk interval amplification mechanism.
[0028] The risk-oriented constraint design module designs gradient constraint rules based on the risk cost of low blood oxygenation.
[0029] The risk-oriented loss construction module constructs a risk-oriented loss baseline and an enhancement loss based on gradient constraints, and introduces a risk smoothing factor and a risk-oriented index to construct the risk-oriented loss.
[0030] The risk gradient stabilization calculation module obtains the updated gradient by differentiating the risk-oriented loss with the clinical smoothing factor.
[0031] The fully connected deep network built by the hypoxemia risk prediction model design module uses the monitoring dataset as input and risk-oriented loss as the optimization objective to design a hypoxemia risk prediction model.
[0032] The hypoxemia risk prediction module is based on a trained hypoxemia risk prediction model to achieve real-time hypoxemia risk prediction.
[0033] Example 2, see Figure 1This embodiment is based on the above embodiment. The monitoring data acquisition module acquires historical hypoxemia examination data: it acquires the patient's physiological indicators and operational characteristics related to hypoxemia events throughout the entire process of electronic bronchoscopy, including preoperative SpO2, heart rate, blood pressure, ASA classification, degree of airway stenosis, anesthesia method, oxygenation method, operation duration, and insertion depth. The acquired data is preprocessed: missing values are filled with the median, outliers are removed using the 3σ criterion, all continuous features are Z-score standardized to eliminate dimensional differences, and discrete features are categorically coded. Sample labeling is performed: a hypoxemia event is defined as SpO2 < 90% within 5 minutes during or after the operation and lasting ≥ 30 seconds, and is marked as y = 1 if the condition is met, otherwise y = 0; thus, a monitoring dataset is obtained.
[0034] Example 3, see Figure 1 This embodiment is based on the above embodiment. The hypoxemia risk cost definition module defines clinical risk costs, focusing on strengthening the penalty weight for missed diagnoses (hypoxemia actually occurs but the model predicts a very low probability), reflecting the degree of deviation between the model's predicted risk and the actual hypoxemia event, ensuring that the optimization direction is consistent with the clinical risk judgment, and introducing a critical risk interval amplification mechanism, expressed as: ; ; where y is the patient's actual hypoxemia occurrence label, 1 for occurrence and 0 for non-occurrence; It is the probability of low blood oxygenation output by the model; It is the cost of low blood oxygen risk deviation, which measures the degree of difference between predicted risk and actual clinical events; is the critical amplification factor, ranging from 2 to 5, used to amplify the cost difference in the critical risk range; k is the critical decision threshold, ranging from 0.4 to 0.6. It is a critical risk interval amplification term; the logarithmic difference structure can amplify the difference in low probability intervals. When the actual event y=1 but the predicted risk is very low, the difference will increase significantly, thereby explicitly quantifying the high clinical cost of missed diagnosis and avoiding insufficient punishment for high-risk missed diagnosis; through the critical risk interval amplification mechanism, we focus on the critical risk samples that are easy to miss in electronic bronchoscopy and improve the model's ability to identify critical hypoxia events.
[0035] Example 4, see Figure 1 This embodiment is based on the above embodiment. The risk-oriented constraint design module, starting from the safety of gradient updates, designs gradient constraint rules in reverse, forcing the model gradient to be proportional to the risk cost. This ensures that the model generates large gradients and strong parameter updates when there is a high risk of misjudgment, and small gradients and fine-tuning when there is a low risk of bias. The gradient direction strictly follows clinical safety logic: if... For situations that underestimate risk and represent the most dangerous clinical scenario, the gradient should be positive to promote... If the gradient increases, the corresponding risk gradient should be negative, and a small adjustment should be made accordingly.
[0036] Represented as: ;in, It is a gradient constraint direction term oriented towards low blood oxygen risk; It is a sign smoothing function. z is the demonstration parameter. It is the smoothing kurtosis coefficient, with a value of 10~20, used to control the smoothing degree and adapt to the noise characteristics of electronic bronchoscopy monitoring data; by locking the gradient direction through the smoothing sign function and controlling the gradient amplitude in combination with risk cost, the gradient update strength is strongly bound to the clinical risk level, which is used to distinguish the clinical difference between high-risk missed diagnosis and low-risk misdiagnosis, while effectively suppressing gradient oscillation caused by monitoring noise and improving the training stability of the model on monitoring data of patients undergoing electronic bronchoscopy.
[0037] By performing the above operations, this solution addresses the problems of general hypoxemia risk prediction systems, such as insufficient differentiation of clinical risks, inadequate quantification of the clinical cost of missed diagnoses, and susceptibility to intraoperative monitoring noise, leading to poor reliability in risk prediction. This solution defines clinical risk costs, focuses on strengthening the penalty weight for missed diagnoses, and specifically enhances the penalty for endoscopic missed diagnoses. It also introduces a critical risk range amplification mechanism to focus on samples prone to endoscopic missed diagnoses; strongly binds gradient amplitudes to clinical risks, enabling stronger adjustments for high-risk cases and fine-tuning for low-risk cases; and introduces a smoothing sign function to adapt to the noise environment of endoscopic monitoring. These measures ultimately improve the reliability of hypoxemia risk prediction.
[0038] Example 5, see Figure 1 This embodiment is based on the above embodiment. The risk-oriented loss construction module, based on the guided gradient constraint, back-derives the loss function used for network training. First, a risk-oriented loss baseline is constructed, and then it is expanded to enhance the risk-oriented loss, strengthening the identification ability of critical risk samples and samples close to the low blood oxygen threshold. This significantly improves the predictive sensitivity and recall rate for high-risk patients, as expressed as: ; ;in, It is a risk-oriented loss baseline; This enhances risk-oriented loss; the loss baseline ensures the basic risk fitting ability, and the enhanced loss introduces the difference squared term to amplify the loss difference in the critical interval, making the model pay more attention to intermediate samples that are prone to hypoxia but whose risk assessment is ambiguous, thus reducing the clinical false negative rate; to adapt to different clinical risk sensitivity requirements, a risk smoothing factor is introduced to establish the final risk-oriented loss, achieving flexible and adjustable loss index, expressed as: ; ; ;
[0039] in, is the risk smoothing factor, which is equivalent to the natural logarithm in forward calculation and has a gradient of 1 during backpropagation; p is the risk orientation index, which takes values from 1 to 4. It comes at the cost of low blood oxygen levels; It is a low blood oxygen risk-oriented loss; the penalty for high-risk bias is adjusted by the risk-oriented index p. The larger the p, the stronger the penalty for missed diagnosis. It can be flexibly configured according to clinical tolerance.
[0040] Example 6, see Figure 1 This embodiment is based on the above embodiment. The risk gradient stabilization calculation module calculates the derivative of the risk-oriented loss with respect to the clinical smoothing factor to obtain a stable, smooth, and clinically interpretable final update gradient. This ensures that the gradient magnitude is strictly proportional to the clinical risk bias, avoiding gradient oscillations and numerical instability, as expressed as: ;in, It updates the gradient for low blood oxygen risk after stabilization; it eliminates the gradient scaling effect caused by the logarithmic function by using a clinical smoothing factor, so that the gradient magnitude is determined only by the clinical risk bias, thereby improving training stability while ensuring the correct gradient direction.
[0041] Furthermore, the hypoxemia risk prediction model design module constructs a lightweight, deployable, fully connected deep network. Using the monitoring dataset as input and risk-oriented loss as the optimization objective, it employs mini-batch stochastic gradient descent for backpropagation to update network parameters, thus obtaining the hypoxemia risk prediction model. During training, the gradient is entirely driven by clinical risk costs, significantly improving the model's accuracy in identifying high-risk hypoxemia samples and its timely warning. The model inference and parameter updates are represented as follows: ; Where x is the preprocessed monitoring data; It is the mapping function of the risk prediction model for hypoxemia. It is the set of learnable parameters of the model; This is the learning rate, with a value of 10. -4 ~5×10 -3 ; These are the updated model parameters;
[0042] The model employs a fully connected deep neural network with the following structure: Input layer → Hidden layer 1 (64 neurons, ReLU activation) → Hidden layer 2 (32 neurons, ReLU activation) → Hidden layer 3 (16 neurons, ReLU activation) → Output layer (1 neuron, Sigmoid activation). The hidden layers uniformly use the ReLU activation function to mitigate gradient vanishing and accelerate convergence. The output layer uses Sigmoid activation to map the output results to the [0,1] probability interval, corresponding to the probability of hypoxemia. Numerical protection is applied to the predicted probabilities after the model's output layer. ; It is a smoothing term, take 10.-6 .
[0043] By performing the above operations, this solution addresses the problem that general hypoxemia risk prediction systems suffer from poor prediction performance due to their uncoupled critical risk characteristics in endoscopic scenarios, making them unsuitable for individualized clinical needs in endoscopy. Instead, it addresses this by combining a critical risk interval amplification term with the squared difference term of the enhancement loss to double-amplify the loss difference in the critical risk interval during electronic bronchoscopy, effectively reducing the clinical missed detection rate. Furthermore, it introduces a risk smoothing factor and a risk-oriented index to construct a risk-oriented loss, allowing for flexible adjustment of the penalty intensity. The gradient magnitude is directly determined by the hypoxemia risk deviation in electronic bronchoscopy, prioritizing the avoidance of high-risk missed diagnoses, thereby improving the prediction effect of hypoxemia risk.
[0044] Example 7, see Figure 1 This embodiment is based on the above embodiment. The hypoxemia risk prediction module acquires hypoxemia examination data in real time based on the trained hypoxemia risk prediction model. After preprocessing, the data is input into the hypoxemia risk prediction model to obtain the current patient's hypoxemia risk probability. Clinical risk stratification and early warning are then performed based on the predicted probability: if... If deemed low risk, routine monitoring will be implemented; if If determined to be of medium risk, increase the frequency of intraoperative monitoring; if If the risk is deemed high, prepare for high-flow oxygen administration in advance and report to management personnel.
[0045] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.
[0046] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.
Claims
1. An intelligent system for predicting the risk of hypoxemia in patients undergoing electronic bronchoscopy, characterized in that: The system includes a monitoring data acquisition module, a hypoxemia risk cost definition module, a risk-oriented constraint design module, a risk-oriented loss construction module, a risk gradient stabilization calculation module, a hypoxemia occurrence risk prediction model design module, and a hypoxemia occurrence risk prediction module. The monitoring data acquisition module acquires historical hypoxemia test data and preprocesses it to obtain a monitoring dataset. The hypoxia risk cost definition module defines the hypoxia risk deviation cost by combining a logarithmic difference structure with a critical risk interval amplification mechanism. The risk-oriented constraint design module designs gradient constraint rules based on the risk cost of low blood oxygenation. The risk-oriented loss construction module constructs a risk-oriented loss baseline and an enhancement loss based on gradient constraints, and introduces a risk smoothing factor and a risk-oriented index to construct the risk-oriented loss. The risk gradient stabilization calculation module obtains the updated gradient by differentiating the risk-oriented loss with the clinical smoothing factor. The fully connected deep network built by the hypoxemia risk prediction model design module uses the monitoring dataset as input and risk-oriented loss as the optimization objective to design a hypoxemia risk prediction model. The hypoxemia risk prediction module is based on a trained hypoxemia risk prediction model to achieve real-time hypoxemia risk prediction.
2. The intelligent electronic bronchoscopy patient hypoxemia risk prediction system according to claim 1, characterized in that: The hypoxia risk cost definition module defines clinical risk costs to reflect the degree of deviation between the model's predicted risk and the actual hypoxia event, and introduces a critical risk interval amplification mechanism to focus on critical risk samples.
3. The intelligent electronic bronchoscopy patient hypoxemia risk prediction system according to claim 1, characterized in that: The risk-oriented constraint design module reverse-engineers gradient constraint rules, forcing the model gradient to be proportional to the risk cost.
4. The intelligent electronic bronchoscopy patient hypoxemia risk prediction system according to claim 1, characterized in that: The risk-oriented loss construction module is based on the guiding gradient constraint. First, a risk-oriented loss baseline is constructed, and then the difference square term is introduced to amplify the loss difference in the critical interval, thus expanding it into an enhanced risk-oriented loss. Furthermore, a risk smoothing factor is introduced, which is equivalent to the natural logarithm in forward calculation and has a gradient that is always 1 during backpropagation. The risk-oriented index is used to adjust the penalty for risk deviation and establish the final risk-oriented loss.
5. The intelligent electronic bronchoscopy patient hypoxemia risk prediction system according to claim 1, characterized in that: The hypoxemia risk prediction model design module constructs a fully connected deep network, takes the monitoring dataset as input, uses risk-oriented loss as the optimization objective, and uses mini-batch stochastic gradient descent for backpropagation to update the network parameters, thus obtaining the hypoxemia risk prediction model.
6. The intelligent electronic bronchoscopy patient hypoxemia risk prediction system according to claim 5, characterized in that: The hypoxemia risk prediction module is based on a trained hypoxemia risk prediction model. It acquires hypoxemia examination data in real time, preprocesses it, and inputs it into the hypoxemia risk prediction model to obtain the current patient's hypoxemia risk probability. Clinical risk stratification and early warning are performed based on the predicted probability.