An integrated machine learning prediction method for postoperative complication nursing risk of lung cancer

Abstract

The application discloses a kind of lung cancer postoperative complication nursing risk integrated machine learning prediction method, specifically related to medical care and health risk prediction field, for solving the problem of lack of effective comprehensive modeling for the risk assessment of existing lung cancer postoperative nursing on complication.The method is by synchronously collecting thoracic drainage pressure waveform and ventilator tidal volume waveform in postoperative nursing cycle, dynamic mutual information change trajectory under multiple time scales is constructed to determine that air leakage activity enters stable platform period;In platform period, trigger controlled negative pressure step excitation test, extract key features reflecting system stability and gas leakage dynamic characteristics;Combined with feature intersection, nursing stage time clipping and the anatomical influence weight of lobectomy position, the risk characteristics are modulated, and input into pre-trained neural network, to realize the dynamic prediction of postoperative complication risk, and provide technical support for fine nursing decision.

Description

Technical Field

[0001] This invention relates to the field of medical care and health prediction technology, and more specifically, to an integrated machine learning method for predicting the risk of postoperative complications in lung cancer patients. Background Technology

[0002] Following surgical treatments such as lobectomy, postoperative complications in lung cancer patients remain a significant factor affecting their prognosis and hospital safety. In clinical practice, patients often require various nursing interventions postoperatively, including chest drainage, mechanical ventilation, and continuous vital sign monitoring. The duration of air leakage, changes in respiratory mechanics, and the operational status of the drainage system are all closely related to lung re-expansion, chest cavity healing, and the occurrence of complications. However, current postoperative care primarily relies on the experience of nursing staff and discrete time-point examination results, with limited utilization of continuous dynamic information such as chest drainage pressure waveforms and tidal volume changes. This makes it difficult to reflect the evolution of air leakage and potential risks in a timely and objective manner.

[0003] Furthermore, differences in anatomical structure, respiratory function, and drainage pathways at different lobectomy sites can lead to significant inconsistencies in the risk indication significance of the same monitoring indicators among different patients. However, existing assessment methods typically do not systematically model these differences. While some studies have attempted to introduce machine learning methods to predict postoperative complications, these methods mostly remain at the level of static indicators or single time window analysis, lacking a comprehensive characterization of the evolutionary characteristics of nursing stages, controlled nursing operation responses, and the coupling relationships of multi-source physiological signals. This makes it difficult to meet the actual needs of refined and phased nursing decision-making.

[0004] Therefore, there is a need for a technical solution that can dynamically predict the risk of complications by combining continuous physiological signals, nursing procedures, and individual anatomical differences in postoperative care scenarios, in order to improve the safety and foresight of postoperative care for lung cancer. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide an integrated machine learning prediction method for the nursing risk of postoperative complications in lung cancer to address the problems raised in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] An integrated machine learning method for predicting the risk of postoperative complications in lung cancer patients includes the following steps:

[0008] S1. During the current postoperative care period of the patient, simultaneously collect the pressure waveform sequence of the chest drainage tube and the tidal volume waveform sequence of the ventilator;

[0009] S2. Calculate the mutual information values ​​of the pressure waveform sequence and the tidal volume waveform sequence of the ventilator at multiple time scales, and establish the dynamic mutual information change trajectory.

[0010] S3. When the dynamic mutual information change trajectory converges to the preset stable range, it is determined that the current patient's leakage activity has entered the plateau period.

[0011] S4. When the patient's air leakage activity enters a plateau phase, control the chest drainage device to perform a negative pressure stepwise excitation test, and extract the power spectral entropy value of the drainage pressure waveform and the amplitude envelope attenuation coefficient of the gas flow waveform respectively.

[0012] S5. Combine the power spectral entropy value, amplitude envelope attenuation coefficient and the convergence speed of the dynamic mutual information change trajectory to generate a risk feature vector.

[0013] S6. Map the current location of the patient's surgical lobectomy to a set of anatomical influence weight parameters, modulate the contribution of the risk feature vector based on nursing stage time clipping, and input the modulated risk feature vector into a pre-trained neural network to predict the risk of postoperative complications.

[0014] As a further aspect of the present invention, in step S1, the simultaneous acquisition of the pressure waveform sequence of the chest drainage tube and the tidal volume waveform sequence of the ventilator specifically includes:

[0015] Simultaneously acquire the analog voltage signal from the chest drainage tube pressure sensor and the raw signal output by the ventilator. Perform analog-to-digital conversion on the analog voltage signal from the pressure sensor to obtain a digital pressure waveform sequence. At the same time, extract the digital tidal volume waveform sequence corresponding to the raw signal output by the ventilator.

[0016] The synchronized pressure waveform digital sequence and tidal volume waveform sequence are divided by a sliding window to generate a series of time-aligned waveform segment pairs. Invalid waveform segment pairs with signal interruption or interference are filtered out, and the remaining continuous and valid waveform segment pairs are merged back into the pressure waveform sequence and tidal volume waveform sequence.

[0017] As a further aspect of the present invention, in step S2, calculating the mutual information values ​​of the pressure waveform sequence and the tidal volume waveform sequence of the ventilator at multiple time scales, and establishing a dynamic mutual information change trajectory specifically includes:

[0018] Set a set of window sizes with different time lengths that include multiple time scales, and for each time scale, slide to extract a pair of synchronized waveform segments of the corresponding time length from the time-aligned pressure waveform sequence and tidal volume waveform sequence;

[0019] For each extracted waveform segment pair, kernel density estimation is used to calculate the distribution density of pressure waveform data points and tidal volume waveform data points, which are respectively used as the marginal probability distribution of pressure and the marginal probability distribution of tidal volume. At the same time, the distribution density of the joint occurrence of the two data points is calculated as the joint probability distribution. Based on the log ratio of the joint probability distribution and the marginal probability distribution, the mutual information value of the segment at the corresponding time scale is calculated.

[0020] The mutual information of time series across all time scales is aggregated into a multidimensional mutual information value matrix, forming a dynamic trajectory of mutual information change.

[0021] As a further aspect of the present invention, in S3, determining that the current patient's air leakage activity has entered a plateau phase when the dynamic mutual information change trajectory converges to a preset stable range specifically includes:

[0022] The preset stable interval includes the moving average and moving standard deviation stability thresholds corresponding to mutual information sequences at different scales. For each mutual information time series representing a single time scale in the dynamic mutual information change trajectory, the moving average and moving standard deviation within the current nursing cycle are calculated.

[0023] When the mutual information sequence at all time scales is within its corresponding stable interval, the dynamic mutual information change trajectory is determined to converge, generating a judgment signal that the current patient's leakage activity has entered a plateau phase.

[0024] As a further aspect of the present invention, in step S4, extracting the power spectral entropy value of the drainage pressure waveform and the amplitude envelope attenuation coefficient of the gas flow waveform specifically includes:

[0025] The negative pressure setting value of the chest drainage device is adjusted according to the preset step-like pressure change sequence. When the negative pressure setting value is maintained at each step level, the real-time pressure waveform of the chest drainage tube pressure sensor and the real-time gas flow waveform of the drainage device gas flow meter are collected.

[0026] The real-time pressure waveform at each step level is converted into power spectral density through Fourier transform, and the corresponding information entropy is calculated as the power spectral entropy value.

[0027] Simultaneously, a Hilbert transform is performed on the real-time gas flow waveform to obtain the amplitude envelope. An exponential function is then fitted to the amplitude envelope during the set decay segment of the step maintenance period. The coefficient of the resulting exponential term is determined as the amplitude envelope decay coefficient characterizing the gas leakage rate change.

[0028] As a further aspect of the present invention, in step S5, the convergence speed of the power spectral entropy value, the amplitude envelope attenuation coefficient, and the dynamic mutual information change trajectory is combined using feature cross-combination to generate a risk feature vector, specifically including:

[0029] Based on the moving average sequence of mutual information sequences at each time scale before convergence in the dynamic mutual information change trajectory, the absolute value of the linear regression slope is calculated as the convergence speed of the dynamic mutual information change trajectory.

[0030] The power spectral entropy, amplitude envelope attenuation coefficient, and convergence speed are multiplied pairwise to construct a set of interactive feature terms. These terms are then concatenated with the three original features (power spectral entropy, amplitude envelope attenuation coefficient, and convergence speed) to form a risk feature vector.

[0031] As a further aspect of the present invention, in S6, mapping the current surgical lobectomy location of the patient to a set of anatomical influence weight parameters and modulating the contribution of the risk feature vector based on nursing stage time trimming specifically includes:

[0032] Based on the location of lobectomy in the current patient's surgical record, and according to the differences in functional impact, a set of anatomical impact weight parameters is set for the risk feature vector;

[0033] Based on the nursing focus corresponding to different postoperative care stages, a binary mask vector of feature contribution is configured for each stage, and feature clipping is performed on each feature in the risk feature vector based on the time dimension of the nursing stage.

[0034] The risk feature vector after feature trimming is multiplied element-wise with the anatomical influence weight parameter to finally output the modulated and trimmed multi-stage risk feature vector.

[0035] As a further aspect of the present invention, step S6, which involves inputting the modulated risk feature vector into a pre-trained neural network to predict the risk of postoperative complications, specifically includes:

[0036] A dataset of historical patient samples is extracted. The modulated multi-stage risk feature vector of each patient sample is used as input, and the clinical statistical record of whether a preset specific complication occurs within the postoperative nursing week is used as the supervision label. A neural network model is established and trained based on the historical patient sample dataset. The neural network model is a feedforward network structure with at least one hidden layer.

[0037] The modulated multi-stage risk feature vector of the current patient is input into the trained neural network model, which outputs the predicted probability value of the current patient developing postoperative complications at multiple time points in the future.

[0038] The technical effects and advantages of the integrated machine learning prediction method for the nursing risk of postoperative complications in lung cancer of this invention are as follows:

[0039] This invention systematically integrates multi-source continuous physiological signals generated during lung cancer postoperative care, enabling refined and dynamic assessment of postoperative complication risks and effectively improving the accuracy and foresight of nursing risk prediction.

[0040] By simultaneously analyzing the multi-timescale coupling relationship between chest drainage pressure waveform and ventilator tidal volume waveform throughout the nursing cycle, this study objectively characterizes the dynamic correlation between air leakage activity and respiratory status, thereby more accurately identifying the plateau phase of air leakage activity and avoiding subjective errors caused by traditional experience-based judgment. The convergence speed of dynamic mutual information change trajectories is introduced, incorporating the coupled evolution process into risk modeling, achieving a synergistic characterization of static and dynamic evolutionary features. Furthermore, by combining the impact of surgical lobectomy location on anatomical functional differences and the time-based trimming mechanism of risk feature participation at different nursing stages, the risk assessment results better reflect the physiological and operational condition changes in real nursing scenarios. Finally, through a pre-trained neural network, the modulated multi-stage risk features are comprehensively analyzed, outputting time-directed complication risk prediction results, providing nursing staff with more targeted early warning information, and thus improving the scientific rigor and safety of postoperative nursing decisions. Attached Figure Description

[0041] Figure 1 This is a schematic diagram of an integrated machine learning method for predicting the risk of postoperative complications in lung cancer patients according to the present invention. Detailed Implementation

[0042] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0043] Example 1

[0044] Figure 1 This invention presents an integrated machine learning method for predicting the risk of postoperative complications in lung cancer patients, comprising the following steps:

[0045] S1. During the current postoperative care period of the patient, simultaneously collect the pressure waveform sequence of the chest drainage tube and the tidal volume waveform sequence of the ventilator;

[0046] S2. Calculate the mutual information values ​​of the pressure waveform sequence and the tidal volume waveform sequence of the ventilator at multiple time scales, and establish the dynamic mutual information change trajectory.

[0047] S3. When the dynamic mutual information change trajectory converges to the preset stable range, it is determined that the current patient's leakage activity has entered the plateau period.

[0048] S4. When the patient's air leakage activity enters a plateau phase, control the chest drainage device to perform a negative pressure stepwise excitation test, and extract the power spectral entropy value of the drainage pressure waveform and the amplitude envelope attenuation coefficient of the gas flow waveform respectively.

[0049] S5. Combine the power spectral entropy value, amplitude envelope attenuation coefficient and the convergence speed of the dynamic mutual information change trajectory to generate a risk feature vector.

[0050] S6. Map the current location of the patient's surgical lobectomy to a set of anatomical influence weight parameters, modulate the contribution of the risk feature vector based on nursing stage time clipping, and input the modulated risk feature vector into a pre-trained neural network to predict the risk of postoperative complications.

[0051] In S1, the pressure waveform sequence of the chest drainage tube and the tidal volume waveform sequence of the ventilator are acquired simultaneously.

[0052] A pressure sensor is placed proximal to the chest drainage tube at the patient's postoperative bedside. This pressure sensor outputs a continuous analog voltage signal proportional to the instantaneous pressure change within the drainage tube. The pressure sensor is selected based on a range covering the commonly used negative pressure range for clinical drainage, for example, covering... A linear output sensor with an output voltage range of 40 cmH2O to +10 cmH2O, for example, 0–5V, is used to ensure sufficient resolution of minute pressure fluctuations. Simultaneously, a raw signal stream consistent with the respiratory cycle is extracted from the ventilator's signal output interface. This raw signal contains tidal volume data reflecting changes in inspiratory and expiratory phases during a single breath. The ventilator output signal itself is in digital form and has a clear timestamp. Subsequently, the two signals are aligned using a unified time synchronization mechanism, which can be implemented using a unified sampling clock or timestamp-based calibration, ensuring a one-to-one correspondence between the pressure signal and the tidal volume signal on the same time axis. For the analog voltage signal output by the pressure sensor, an analog-to-digital converter is used for sampling and quantization. The sampling frequency is set according to the nursing monitoring accuracy requirements, for example, 100Hz, to balance the ability to capture rapid pressure changes within the respiratory cycle with data volume control. After analog-to-digital conversion, a continuous digital pressure waveform sequence is obtained. At the same time, the tidal volume numerical sequence corresponding to the timestamp is parsed and extracted from the original signal output by the ventilator, thus forming a digital tidal volume waveform sequence that is strictly time-corresponding to the pressure waveform sequence.

[0053] After synchronous acquisition and digitization of the pressure waveform sequence and tidal volume waveform sequence, a sliding window segmentation operation is performed on the two types of waveform data along a unified time axis. Specifically, based on the temporal characteristics of changes in respiratory rhythm and drainage status during postoperative care, the time length of the sliding window is set, for example, a fixed value within the range of 5 to 20 seconds. In this embodiment, 10 seconds can be used as the window length. Simultaneously, a window step size is set, for example, 1 second or 2 seconds, to ensure a certain overlap between adjacent waveform segments and avoid missing key transient information. During the sliding process, data within the corresponding time range of the pressure waveform sequence and tidal volume waveform sequence are simultaneously captured at each window position, forming a pair of waveform segments with strictly aligned time. Subsequently, an effectiveness test is performed on each pair of waveform segments. The effectiveness test is based on the continuity and integrity of the signal, for example, detecting whether there are signal interruptions, saturation, abnormal abrupt changes, or obvious deviations from physiological rhythms within the segment. Abnormal waveform segments detected due to sensor detachment, signal loss, or external interference are directly marked as invalid and discarded. Waveform segments with continuous waveforms, stable amplitude changes, and consistent with normal respiratory rhythms are marked as valid. After completing the full-sequence sliding detection, all valid waveform segments are reassembled according to their original time sequence to reconstruct continuous pressure and tidal volume waveform sequences. During the splicing process, the original time sequence of each segment is preserved to ensure that the reconstructed waveform sequence maintains continuity and consistency in the time dimension.

[0054] In step S2, the mutual information values ​​of the pressure waveform sequence and the tidal volume waveform sequence of the ventilator are calculated at multiple time scales to establish a dynamic mutual information change trajectory.

[0055] For the pressure and tidal volume waveform sequences that have already undergone time alignment processing, a set of window sizes with multiple time lengths is set based on the temporal characteristics of respiratory rhythm changes and leakage evolution in postoperative nursing scenarios. This is used to characterize the correlation between the two types of waveforms from different observation scales. The time scales are set using a discrete hierarchical strategy, rather than continuous variation, to ensure the controllability and stability of the calculation process. For example, in this embodiment, the window length can be set to four scales: 5 seconds, 10 seconds, 20 seconds, and 40 seconds. Shorter time scales are used to capture rapid changes within a single or a few respiratory cycles, while longer time scales are used to reflect the overall coupling trend within the nursing stage. After setting the window sets, for each time scale, a sliding capture operation is simultaneously performed on the pressure and tidal volume waveform sequences along a unified time axis. The sliding step size is set according to the nursing monitoring accuracy requirements, such as 1 second or 2 seconds, to ensure appropriate overlap between adjacent waveform segments, thereby avoiding the loss of key information due to window boundaries. During the sliding process, each window movement simultaneously extracts data from the pressure waveform sequence and the tidal volume waveform sequence within the same start and end time range, forming a pair of synchronized waveform segments with strictly consistent time.

[0056] After acquiring synchronous waveform segment pairs at different time scales, statistical modeling is performed on the distribution characteristics of pressure waveform data points and tidal volume waveform data points for each extracted waveform segment pair. Specifically, kernel density estimation is used to continuously describe the distribution of data points within a segment, avoiding the instability issues caused by simple histogram statistics with a limited sample size. For data points in the pressure waveform segment, a continuous probability density distribution is obtained through kernel density estimation based on their value distribution within that time window, serving as the corresponding pressure marginal probability distribution for that segment. Similarly, the same kernel density estimation process is performed on the data points in the tidal volume waveform segment to obtain the tidal volume marginal probability distribution. Furthermore, based on the synchronous occurrence relationship of the two types of waveform data points at the same time position, their joint distribution is modeled, i.e., the probability density distributions of pressure data points and tidal volume data points appearing in pairs within the same segment are statistically analyzed to form a joint probability distribution. After constructing the marginal probability distribution and the joint probability distribution, the mutual information value of waveform segments at this time scale is calculated based on the logarithmic ratio between the joint probability distribution and the corresponding marginal probability distribution. This quantifies the correlation strength between pressure changes and tidal volume changes at this observation scale and time position. This mutual information value can comprehensively reflect the statistical dependence of the two types of signals, without relying on the assumption of a linear relationship.

[0057] Following the time scale dimension and the order of time progression, the mutual information time series obtained at each time scale are aligned and organized to maintain consistency on the time axis and form a parallel structure in the scale dimension. This method combines mutual information values ​​corresponding to the same time position at different time scales into a multidimensional data structure. One dimension distinguishes different time scales, and the other reflects the temporal evolution sequence within the nursing cycle. This multidimensional mutual information value matrix can simultaneously express the dynamic characteristics of mutual information changes over time and the distribution differences of mutual information at different observation scales, thus forming a complete dynamic mutual information change trajectory. This dynamic mutual information change trajectory is not subjected to additional dimensionality reduction or compression processing but is directly input as a leak status determination and feature extraction. Analysis of this trajectory can intuitively reflect the overall evolution trend of the correlation strength between pressure waveforms and tidal volume waveforms during the nursing process and the stability changes at different time scales.

[0058] In step S3, when the dynamic mutual information change trajectory converges to a preset stable range, it is determined that the current patient's air leakage activity has entered a plateau phase.

[0059] For the constructed dynamic mutual information change trajectory, the trajectory data is split according to the time scale dimension, so that each mutual information time series corresponds to the evolution of the correlation strength between the pressure waveform and the tidal volume waveform at a single time scale. For each mutual information time series, moving average statistical calculation is continuously performed along the time axis throughout the entire postoperative care cycle of the current patient. The moving average statistical calculation uses a fixed-length statistical window, the length of which is set according to the stability requirements of state changes in the nursing scenario, for example, a fixed value of 30 seconds, 60 seconds, or 120 seconds. 60 seconds is selected as the statistical window length to balance the smoothing ability for short-term fluctuations and the response speed to trend changes. As the statistical window slides along the time axis, the moving average of each mutual information value sequence within the window is calculated to characterize the central trend of the mutual information level at that time scale; simultaneously, the moving standard deviation of the mutual information values ​​is calculated within the same statistical window to characterize the stability of the mutual information fluctuation amplitude. After completing the calculation of the moving average and moving standard deviation, a corresponding stability threshold interval is set for each time scale based on historical nursing data or clinical experience rules. The stability threshold interval is set using a dual-threshold structure, simultaneously limiting the range of variation of the moving average and the upper limit of the moving standard deviation. For example, it is stipulated that the variation of the moving average within a certain number of consecutive statistical windows shall not exceed 10% of its long-term average, and the moving standard deviation shall be lower than a preset upper limit value. The upper limit value is determined based on the statistical results of historical patient data, taking the 75th percentile of the mutual information standard deviation distribution in the same patient group as a reference. A clear and executable stability interval determination rule is constructed for each time-scale mutual information sequence.

[0060] After constructing the stable intervals corresponding to the mutual information sequences at each time scale, a convergence determination operation is performed on the overall dynamic mutual information change trajectory. At the current nursing time point, the latest statistical window results corresponding to each time scale mutual information time series are compared one by one to determine whether its moving average and moving standard deviation simultaneously fall within the pre-set stable interval range for that time scale. This determination process adopts a parallel consistency determination mechanism, rather than a single-scale independent determination. That is, only when the mutual information time series at all time scales meet their respective stable interval conditions within the same time period are the overall dynamic mutual information change trajectory considered to have entered a convergent state. This consistency determination method can effectively avoid misjudgment problems caused by occasional stability at a single time scale, ensuring that the convergence determination reflects the objective state of stable coupling between pressure waveforms and tidal volume waveforms at multiple observation scales. During the determination process, to avoid the interference of instantaneous noise or short-term stable intervals on the results, a duration constraint is introduced into the consistency determination results. That is, the above-mentioned stable conditions are required to be continuously met within a certain number of consecutive statistical windows. For example, it is set that the stable interval conditions are met for 3 to 5 consecutive statistical windows before the convergence of the dynamic mutual information change trajectory is finally confirmed. The specific value of this duration constraint can be set according to the accuracy requirements of nursing monitoring. In this embodiment, three consecutive statistical windows are used as the confirmation condition. Once the above-mentioned multi-scale consistency and persistence conditions are met, a judgment signal indicating that the current patient's air leakage activity has entered a plateau phase is immediately generated, and this judgment signal is used as one of the triggering conditions for subsequent negative pressure step ladder provocation tests.

[0061] In step S4, the power spectral entropy value of the drainage pressure waveform and the amplitude envelope attenuation coefficient of the gas flow waveform are extracted respectively.

[0062] Once the patient's air leakage activity is determined to have plateaued, a negative pressure step activation operation is performed on the chest drainage device. The step pressure changes are not adjusted continuously, but rather using a discrete stepwise approach to adjust the negative pressure setpoints step by step, ensuring a stable drainage system state at each pressure level. The step levels are determined based on commonly used negative pressure ranges for chest drainage in clinical practice; for example, the negative pressure steps can be set sequentially as follows: 5cmH2O, 10cmH2O, 15cmH2O and The system uses four levels of 20cmH2O, with the number of steps fixed at 3 to 6. This example uses 4 levels. Each step level corresponds to a fixed maintenance period to ensure the drainage system reaches a relatively stable state under this negative pressure condition. The maintenance period is set via time control, for example, between 15 and 30 seconds. In this example, 20 seconds is used as the single-step maintenance time. During step switching, the negative pressure setpoint changes sequentially, and the system immediately enters the maintenance phase after each switch. While the negative pressure setpoint is maintained at the current step level, real-time pressure waveform data output by the chest drainage tube pressure sensor and real-time gas flow waveform data output by the drainage device gas flow meter are simultaneously collected. To avoid signal interference during step switching, a transient segment and a stable acquisition segment are clearly defined within each step maintenance period. For example, the first 3 seconds of the step maintenance period are considered a transient buffer segment and are not included in subsequent analysis; the remaining time is considered the effective acquisition segment.

[0063] After acquiring real-time pressure waveforms at each negative pressure level, frequency domain analysis and power spectral entropy calculation are performed sequentially for the stable acquisition segment of the pressure waveform corresponding to each level. The acquired pressure waveform data is preprocessed to eliminate the influence of DC components and slow drift. Mean removal is used to make the waveform fluctuate around the zero axis, thus ensuring the stability of the frequency domain analysis results. Subsequently, a frequency domain transformation operation is performed on the preprocessed pressure waveform, converting the time-domain pressure waveform into the corresponding frequency distribution form. During the frequency domain transformation, the waveform is segmented according to a predetermined sampling frequency to ensure that the spectral resolution meets the analysis requirements of the dynamic changes in the respiratory and drainage systems. After the frequency domain transformation, the power spectral density distribution of the pressure waveform at the current level is constructed based on the frequency domain amplitude information. This power spectral density reflects the distribution of pressure waveform energy at different frequency components. To facilitate subsequent entropy calculation, the power spectral density is normalized so that the sum of the power proportions of each frequency component is a fixed value, thereby eliminating the influence of overall energy differences at different levels on the entropy value. After normalization, the power proportions of each frequency component are statistically analyzed according to the principle of information entropy calculation, resulting in a power spectral entropy value that characterizes the complexity of the pressure waveform spectral distribution at that level. This power spectral entropy value numerically reflects the dispersion and uncertainty of the pressure waveform's spectral distribution under the current negative pressure condition. By obtaining the power spectral entropy values ​​at different levels, a set of frequency domain characteristic parameters corresponding to changes in negative pressure are formed.

[0064] For the real-time gas flow waveforms obtained at each negative pressure level, amplitude envelope extraction and attenuation feature calculation are performed sequentially. The gas flow waveforms are preprocessed to remove obvious abnormal peaks and measurement noise, ensuring a continuous and smooth overall trend. Subsequently, a Hilbert transform is performed on the preprocessed gas flow waveforms to extract the corresponding amplitude envelope from the original flow signal. This amplitude envelope describes the outer contour characteristics of the gas flow rate over time. After amplitude envelope extraction, the envelope is divided into time segments within each negative pressure level maintenance period to define the start and end positions of the attenuation analysis segment. The attenuation analysis segment is determined based on the level maintenance time; for example, the middle to later part of the level maintenance period can be used as the attenuation analysis segment. In this embodiment, the time interval from 5 seconds after the start of the level maintenance period to 5 seconds before its end can be used as the attenuation analysis segment to avoid transient fluctuations caused by level switching. After determining the attenuation analysis segment, an exponential function fitting operation is performed on the amplitude envelope within this time interval to obtain the exponential coefficient describing the attenuation trend of the envelope over time. The exponential coefficient is used to quantitatively characterize the rate at which the gas flow rate decreases over time under the current negative pressure step conditions. Its magnitude reflects the dynamic characteristics of the gas leakage rate change. By performing the above envelope extraction and exponential fitting operations on each negative pressure step level, a set of amplitude envelope attenuation coefficients corresponding to the negative pressure step level can be obtained, thus forming a set of time-domain characteristic parameters reflecting the dynamic response characteristics of gas leakage.

[0065] In step S5, the power spectral entropy value, amplitude envelope attenuation coefficient, and convergence speed of dynamic mutual information change trajectory are combined by feature cross-combination to generate a risk feature vector.

[0066] For the constructed dynamic mutual information change trajectory, the mutual information time series at each time scale is first extracted according to the time scale dimension, and the evolution process before entering the stable interval is analyzed. The pre-convergence stage refers to the time interval during which the mutual information time series has not yet met the stable interval determination criteria, but has shown an overall transition towards a stable state, or defined as the time interval from the start of the test to the convergence determination. This stage reflects the real physiological process of the relationship between postoperative air leakage activity and breathing pattern transitioning from unstable to stable. For the mutual information time series within this stage, its moving average sequence is calculated along the time axis to smooth short-term fluctuations and highlight the overall change trend. Based on this, trend fitting processing is performed on the moving average sequence, and the slope of the fitted change trend is used to quantify the speed characteristic of mutual information change over time, and its absolute value is taken as the convergence speed index of the dynamic mutual information change trajectory. Clinically, the convergence rate reflects the speed at which the coupling relationship between thoracic drainage and ventilator support stabilizes. A higher convergence rate indicates a rapid change in the coupling relationship between early postoperative air leakage and respiratory mechanics, suggesting a drastic dynamic adjustment process during pleural healing, lung re-expansion, or drainage status adjustment. Conversely, a lower and more gradual convergence rate indicates a slower stabilization of the coupling relationship, reflecting a more coordinated state between the patient's respiratory and drainage systems postoperatively. Since postoperative complications are often closely related to insufficient thoracic environment stability and imbalances in respiratory and drainage coordination, this convergence rate serves as an important quantitative characteristic for characterizing the stability of the postoperative recovery process, providing a direct indicator of physiological dynamics for complication risk assessment.

[0067] After calculating the convergence rate, the power spectral entropy value, amplitude envelope attenuation coefficient, and convergence rate are incorporated into the risk feature construction process. The power spectral entropy value originates from the spectral distribution characteristics of the drainage pressure waveform under negative pressure step excitation conditions. Its magnitude reflects the degree of energy dispersion of the pressure waveform at different frequency components, corresponding to the stability and complexity of the thoracic drainage system under controlled negative pressure in nursing scenarios. A high power spectral entropy value indicates dispersed spectral components of the pressure waveform and complex system response, usually related to unstable drainage paths, uneven lung tissue rebound, or disturbances in the thoracic environment, all of which may increase the risk of postoperative complications. The amplitude envelope attenuation coefficient originates from the attenuation characteristics of the gas flow waveform during the maintenance period of the negative pressure step. It reflects the rate of decrease in gas leakage over time, corresponding to the pleural rupture closure speed, lung tissue re-expansion efficiency, and leakage improvement in actual nursing care. A smaller attenuation coefficient indicates a slower decrease in gas flow, suggesting a longer duration of leakage or insufficient improvement, which is clinically highly correlated with complications such as pneumothorax and persistent leakage. To further characterize the synergistic relationship among the aforementioned features, the power spectral entropy, amplitude envelope attenuation coefficient, and convergence speed are multiplied pairwise to construct interactive feature terms. This allows the features to reflect not only single-dimensional information but also the coupling effect between system stability, leakage dynamics, and coupling evolution speed. Finally, the three original features and the interactive feature terms are concatenated in a fixed order to form a well-structured and semantically clear risk feature vector. This risk feature vector numerically integrates multiple key nursing dimensions, such as the stability of the thoracic drainage system, the improvement of gas leakage, and the evolution speed of the respiratory-drainage coupling relationship. It can comprehensively reflect the potential risk status of patients during postoperative recovery and is therefore suitable as input features for postoperative complication prediction models.

[0068] In step S6, the current location of the patient's surgical lobectomy is mapped to a set of anatomical influence weight parameters, and the contribution modulation of the risk feature vector is performed based on nursing stage time trimming.

[0069] The specific location of the lung lobectomy is retrieved from the patient's surgical record. This location includes one or more of the left upper lobe, left lower lobe, right upper lobe, right middle lobe, or right lower lobe. Different lung lobes play objectively different roles in respiratory function, ventilation-perfusion ratio, and thoracic drainage pathways. These differences directly affect the duration of postoperative air leakage, lung re-expansion rate, and the probability of complications. For example, after right upper lobe resection, due to the higher position of the residual lung tissue and the longer drainage path, postoperative gas retention and drainage obstruction are more likely. Conversely, after lower lobe resection, the drainage path is relatively smooth due to gravity, but the impact on respiratory mechanics is more significant. Therefore, in this embodiment, different anatomical influence weights are assigned to the risk characteristics corresponding to different resection locations based on the differences in the lung lobe's contribution to respiratory mechanics and the complexity of the drainage path. The weight parameter set is set using a discrete grading method, for example, dividing the anatomical influence into three levels: high, medium, and low, and mapping each level to a different weight value range. In this embodiment, the following settings can be configured: resection of the right upper lobe or multiple lobes corresponds to a higher anatomical impact weight; resection of the left upper lobe or right middle lobe corresponds to a medium anatomical impact weight; and resection of a single left lower lobe or right lower lobe corresponds to a lower anatomical impact weight. This set of weight parameters is not applied to a single feature, but rather to the entire risk feature vector, reflecting the difference in the indicative significance of the same risk feature for complication risk under different anatomical conditions.

[0070] Further considering the phased characteristics of postoperative care, a time-based trimming operation is performed on the risk feature vector based on the care phase. Postoperative care can typically be divided into several phases with clear nursing focuses, such as the early postoperative monitoring phase, the drainage adjustment phase, and the recovery observation phase. The focus of risk differs significantly between these phases. For example, in the early postoperative monitoring phase, the focus is on whether air leakage persists and whether the drainage system is stable. At this stage, the features composed of the pressure waveform power spectrum entropy value and the gas flow amplitude envelope attenuation coefficient have high reference value for risk assessment. In the recovery observation phase, the focus shifts more towards whether the coupling relationship between respiration and drainage is stable. At this stage, the trend characteristics reflected by the convergence speed of the dynamic mutual information change trajectory are more meaningful. Based on these differences in nursing focus, a binary mask vector of feature contribution is configured for each care phase during implementation. The length of this mask vector is consistent with the dimension of the risk feature vector, and its elements only take two values: "participating" or "not participating." For example, in the early postoperative stage, the mask positions corresponding to the power spectral entropy value and amplitude envelope attenuation coefficient are set to valid, while the other positions are set to invalid; in the recovery stage, the convergence speed-related features are set to valid, while the other features are temporarily masked.

[0071] After the nursing stage feature trimming is completed, element-level modulation is performed on the trimmed risk feature vector and the aforementioned set of anatomical influence weight parameters. Specifically, element-level multiplication refers to directly multiplying the value of each feature component in the risk feature vector with its corresponding anatomical influence weight, thereby amplifying or compressing the contribution of that feature under the current patient anatomical conditions at the numerical level. This modulation process does not change the dimensional structure of the risk feature vector, but only adjusts the relative weight relationships of its components, ensuring that the same set of features can reflect different risk indication intensities under different lobectomy scenarios. After element-level multiplication, a risk feature vector that has undergone both nursing stage trimming and anatomical influence modulation is obtained. Structurally, this vector can be represented as a set of multiple feature sub-vectors arranged in the order of nursing stages, forming a multi-nursing stage risk feature vector. In the actual output format, this multi-nursing stage risk feature vector can be organized into a feature sequence arranged in chronological order, where each time node corresponds to a trimmed and modulated feature vector instance. This format preserves the temporal relationship between different nursing stages and fully reflects the modulation results of anatomical differences on feature contributions. Through the above processing, the final output multi-stage nursing risk feature vector can simultaneously encode the patient's surgical anatomical conditions, nursing stage focus, and multi-source physiological characteristic information in numerical expression, making it more realistically reflect the patient's potential health risk status during the postoperative recovery process.

[0072] In step S6, the modulated risk feature vector is input into a pre-trained neural network to predict the risk of postoperative complications.

[0073] A dataset of historical lung cancer postoperative patients was compiled from hospital information and follow-up records. Each patient sample included a modulated multi-stage risk feature vector generated during the postoperative care period, as well as clinical statistical records of whether pre-defined specific complications occurred during the postoperative care period. The pre-defined specific complications were precisely defined as events requiring high-priority warning in nursing practice, such as one or more of the following: "persistent air leakage exceeding a pre-defined duration," "pneumothorax requiring additional intervention," and "atelectasis requiring bronchoscopic suctioning." These were judged and recorded according to a unified standard during the sample compilation phase. For the input data of each patient sample, the multi-stage risk feature vector was organized in a fixed order, for example, arranging the early postoperative stage, drainage adjustment stage, and recovery observation stage sequentially. Each stage corresponded to a feature sub-vector that had undergone binary masking and anatomical weight modulation. The sub-vectors of each stage were then concatenated sequentially to form a single input vector, ensuring consistency in the input dimension across different patient samples. A feedforward neural network model is then established. The model structure adopts a fully connected architecture with at least one hidden layer. In this embodiment, a two-hidden-layer structure is used: the first hidden layer is used to perform nonlinear mapping on the input vector to extract combined pattern features, and the second hidden layer is used to further compress and aggregate cross-stage information. The number of nodes in each hidden layer is set according to the input dimension. For example, when the input vector dimension is in the tens of dimensions, the number of nodes in the first hidden layer is set to 64 and the number of nodes in the second hidden layer is set to 32. Activation units are added after the hidden layers to enhance the nonlinear expressive power. The output layer is configured according to the label definition. If the complication is binary, the output layer is configured with a single probability output node and outputs the probability of occurrence. If the complication contains multiple categories or multiple tasks are performed in parallel, the output layer is configured with the corresponding number of output nodes and outputs the probability of occurrence of each complication. The training process is performed using supervised learning. The historical samples are first divided into training and validation sets proportionally, for example, in an 8:2 ratio. The input vectors are then numerically normalized to the same scale to eliminate differences in the dimensions of different features. During training, training set samples are input in batches, and an iterative training strategy with a fixed number of rounds is adopted. For example, the training rounds are set to 50 rounds and the batch size is set to 32. After each round of training, the prediction error is calculated on the validation set and the optimal model parameters are recorded. Finally, the parameters of the trained neural network model are obtained, forming a pre-trained neural network model for predicting the risk of postoperative complications.

[0074] After training the neural network model and fixing its parameters, the trained model is used to predict the postoperative complication risk of the current patient. Specifically, for the multi-stage risk feature vector obtained during the current patient's postoperative care period, anatomical influence weight modulation and binary masking of care stages are first performed according to rules consistent with historical samples to ensure that the dimensional order, stage arrangement order, and numerical scale of the input vector remain consistent with the training stage. Subsequently, the modulated multi-stage risk feature vector is fed into the trained feedforward neural network model for inference calculation. The model performs nonlinear mapping of the input features and cross-stage information aggregation in the hidden layer, and generates complication risk prediction results in the output layer. To achieve risk prediction for "multiple future time points," in this embodiment, the prediction time points are pre-discretized and defined as several fixed time nodes, such as the 6th, 12th, 24th, and 48th hours postoperatively. Corresponding probability output nodes are configured in the model output layer according to the time point dimension, allowing the model to output the probability of complication occurrence at multiple time points in a single inference. The output results represent the risk level of the current patient developing the target complication at the corresponding future time point in the form of probability values, and these probability values ​​are arranged in chronological order to form a risk prediction sequence. This risk prediction sequence is used in the nursing setting to reflect the trend of risk changes over time, and nursing staff implement phased monitoring and intervention arrangements accordingly. Through this reasoning process, the current patient's multi-stage nursing characteristics are stably mapped into multi-time-point risk probability outputs, and the definition of the output time points and the organization of input features are consistent with the training phase, thereby ensuring that the prediction results have a clear computational path and a repeatable implementation method.

[0075] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. The semiconductor medium can be a solid-state drive.

[0076] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0077] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and modules described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0078] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0079] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0080] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0081] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0082] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0083] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An integrated machine learning method for predicting the risk of postoperative complications in lung cancer surgery, characterized in that, Includes the following steps: S1. During the current postoperative care period of the patient, simultaneously collect the pressure waveform sequence of the chest drainage tube and the tidal volume waveform sequence of the ventilator. S2. Calculate the mutual information values ​​of the pressure waveform sequence and the tidal volume waveform sequence of the ventilator at multiple time scales, and establish the dynamic mutual information change trajectory. S3. When the dynamic mutual information change trajectory converges to the preset stable range, it is determined that the current patient's leakage activity has entered the plateau period. S4. When the patient's air leakage activity enters a plateau phase, control the chest drainage device to perform a negative pressure stepwise excitation test, and extract the power spectral entropy value of the drainage pressure waveform and the amplitude envelope attenuation coefficient of the gas flow waveform respectively. S5. Combine the power spectral entropy value, amplitude envelope attenuation coefficient and the convergence speed of the dynamic mutual information change trajectory to generate a risk feature vector. S6. Map the current location of the patient's surgical lobectomy to a set of anatomical influence weight parameters, modulate the contribution of the risk feature vector based on nursing stage time clipping, and input the modulated risk feature vector into a pre-trained neural network to predict the risk of postoperative complications. In step S4, the extraction of the power spectral entropy value of the drainage pressure waveform and the amplitude envelope attenuation coefficient of the gas flow waveform specifically includes: The negative pressure setting value of the chest drainage device is adjusted according to the preset step-like pressure change sequence. When the negative pressure setting value is maintained at each step level, the real-time pressure waveform of the chest drainage tube pressure sensor and the real-time gas flow waveform of the drainage device gas flow meter are collected. The real-time pressure waveform at each step level is converted into power spectral density through Fourier transform, and the corresponding information entropy is calculated as the power spectral entropy value. Simultaneously, Hilbert transform is performed on the real-time gas flow waveform to obtain the amplitude envelope. During the set decay segment of the step maintenance period, the amplitude envelope is fitted with an exponential function to obtain the coefficient of the exponential term corresponding to the fitting result, which is used as the amplitude envelope decay coefficient characterizing the gas leakage rate change characteristics. In step S5, the risk feature vector is generated by combining the power spectral entropy value, the amplitude envelope attenuation coefficient, and the convergence speed of the dynamic mutual information change trajectory through feature cross-combination. Specifically, this includes: Based on the moving average sequence of mutual information sequences at each time scale before convergence in the dynamic mutual information change trajectory, the absolute value of the linear regression slope is calculated as the convergence speed of the dynamic mutual information change trajectory. The power spectral entropy, amplitude envelope attenuation coefficient, and convergence speed are multiplied pairwise to construct a set of interactive feature terms. These terms are then concatenated with the three original features (power spectral entropy, amplitude envelope attenuation coefficient, and convergence speed) to form a risk feature vector.

2. The integrated machine learning prediction method for the nursing risk of postoperative complications in lung cancer as described in claim 1, characterized in that, In step S1, the simultaneous acquisition of the pressure waveform sequence of the chest drainage tube and the tidal volume waveform sequence of the ventilator specifically includes: Simultaneously acquire the analog voltage signal from the chest drainage tube pressure sensor and the raw signal output by the ventilator. Perform analog-to-digital conversion on the analog voltage signal from the pressure sensor to obtain a digital pressure waveform sequence. At the same time, extract the digital tidal volume waveform sequence corresponding to the raw signal output by the ventilator. The synchronized pressure waveform digital sequence and tidal volume waveform sequence are divided by a sliding window to generate a series of time-aligned waveform segment pairs. Invalid waveform segment pairs with signal interruption or interference are filtered out, and the remaining continuous and valid waveform segment pairs are merged back into the pressure waveform sequence and tidal volume waveform sequence.

3. The integrated machine learning prediction method for the nursing risk of postoperative complications in lung cancer as described in claim 1, characterized in that, In step S2, calculating the mutual information values ​​between the pressure waveform sequence and the tidal volume waveform sequence of the ventilator at multiple time scales, and establishing a dynamic mutual information change trajectory specifically includes: Set a set of window sizes with different time lengths that include multiple time scales, and for each time scale, slide to extract synchronous waveform segments of the corresponding time length from the time-aligned pressure waveform sequence and tidal volume waveform sequence; For each extracted waveform segment pair, kernel density estimation is used to calculate the distribution density of pressure waveform data points and tidal volume waveform data points, which are respectively used as the pressure marginal probability distribution and the tidal volume marginal probability distribution. At the same time, the distribution density of the joint occurrence of the two data points is calculated as the joint probability distribution. Based on the log ratio of the joint probability distribution and the marginal probability distribution, the mutual information value of the segment at the corresponding time scale is calculated. The mutual information of time series across all time scales is aggregated into a multidimensional mutual information value matrix, forming a dynamic trajectory of mutual information change.

4. The integrated machine learning prediction method for the nursing risk of postoperative complications in lung cancer as described in claim 1, characterized in that, In step S3, determining that the current patient's leakage activity has entered a plateau phase when the dynamic mutual information change trajectory converges to a preset stable range specifically includes: The preset stable interval includes the moving average and moving standard deviation stability thresholds corresponding to mutual information sequences at different scales. For each mutual information time series representing a single time scale in the dynamic mutual information change trajectory, the moving average and moving standard deviation within the current nursing cycle are calculated. When the mutual information sequence at all time scales is within its corresponding stable interval, the dynamic mutual information change trajectory is determined to converge, generating a judgment signal that the current patient's leakage activity has entered a plateau phase.

5. The integrated machine learning prediction method for the nursing risk of postoperative complications in lung cancer as described in claim 1, characterized in that, In step S6, mapping the current surgical lobectomy location of the patient to a set of anatomical influence weight parameters and modulating the contribution of the risk feature vector based on nursing stage time trimming specifically includes: Based on the location of lobectomy in the current patient's surgical record, and according to the differences in functional impact, a set of anatomical impact weight parameters is set for the risk feature vector; Based on the nursing focus corresponding to different postoperative care stages, a binary mask vector of feature contribution is configured for each stage, and feature clipping is performed on each feature in the risk feature vector based on the time dimension of the nursing stage. The risk feature vector after feature trimming is multiplied element-wise with the anatomical influence weight parameter to finally output the modulated and trimmed multi-stage risk feature vector.

6. The integrated machine learning prediction method for the nursing risk of postoperative complications in lung cancer according to claim 1, characterized in that, In step S6, inputting the modulated risk feature vector into a pre-trained neural network to predict the risk of postoperative complications specifically includes: A dataset of historical patient samples is extracted. The modulated multi-stage risk feature vector of each patient sample is used as input, and the clinical statistical record of whether a preset specific complication occurs within the postoperative nursing week is used as the supervision label. A neural network model is established and trained based on the historical patient sample dataset. The neural network model is a feedforward network structure with at least one hidden layer. The modulated multi-stage risk feature vector of the current patient is input into the trained neural network model, which outputs the predicted probability value of the current patient developing postoperative complications at multiple time points in the future.

Images