Recognition method for drainage liquid infection risk after orthopedics department operation

Abstract

The invention relates to an orthopedic postoperative drainage liquid infection risk identification method. Drainage liquid is regarded as system output jointly driven by postoperative local immune regulation and wound recovery; capturing a nonlinear time sequence curve; first-order derived trend modeling is carried out; abrupt change after the change rate is stable is analyzed; establishing a postoperative recovery normal path framework, and forming a standard recovery trend library by using drainage fluid dynamic parameters of non-infected cases; comparing the Euclidean offset degree and directional change between the drainage parameter behavior vector of the current patient and the standard trajectory; the change is converted into a visual recovery offset map, and infection prediction data is measured from a multi-day trend angle; constructing and identifying local abnormal signal aggregation; forming a framework of infection event transition to obtain clinical infection nodes of qualitative change caused by quantitative change; according to the method, early-stage accurate recognition can be realized, the intervention opportunity can be obviously advanced, single threshold misjudgment is avoided, the recognition accuracy and specificity are improved, the model adaptability is improved, and the recognition interpretability is enhanced.

Description

Technical Field

[0001] The present invention relates to a method for identifying the infection risk of drainage fluid after orthopedic surgery. Background Art

[0002] Currently, in clinical practice, the methods used to identify the risk of postoperative drainage fluid infection mainly include routine observation, physical sign monitoring, laboratory indicator analysis, bacterial culture, and early warning models that partially rely on static biochemical indicators. These methods have played an auxiliary role in the diagnosis of postoperative infection to a certain extent, but there are still a series of shortcomings and technical drawbacks that need to be addressed. In particular, when facing complex, early, and atypical postoperative infections, they show obvious lag, one-sidedness, and a high misjudgment rate. First, the current mainstream method relies on manual observation of changes in the color, properties, and volume of drainage fluid, such as darkening of color, turbidity of drainage, and sudden decrease or increase in drainage volume. This method relies entirely on the experience and judgment of medical staff, is highly subjective, lacks continuity and quantitative standards, and is prone to missed judgments and delayed identification at night or when staff is insufficient. Second, many hospitals still rely on daily sampling of drainage fluid to be sent to the laboratory for routine bacterial culture and blood index analysis, such as changes in total white blood cell count, C-reactive protein (CRP), procalcitonin (PCT), etc. Although these indicators do have a certain trend of change during the infection process, their physiological response lags. Usually, obvious abnormalities only appear when the infection has developed to a certain extent. They are not suitable as early warning signals of infection, especially in some latent infections and peri-prosthetic microbial infiltration inflammation. The relevant indicators may not change significantly even 24 hours after the infection occurs, resulting in missing the optimal intervention window.

[0003] Furthermore, although bacterial culture is considered the gold standard, it has complex procedures and a long processing time (typically 24–72 hours). It is prone to false negatives due to factors such as delayed delivery, sample contamination, and low colony density. Furthermore, it fails to reflect local dynamic changes in real time, making it more of a diagnostic tool than a predictive tool. Regarding model-based approaches, some studies have attempted to use machine learning methods such as statistical regression, support vector machines, and random forests to classify and predict postoperative infection. However, these models generally rely on single-point postoperative data inputs, such as drainage fluid pH and white blood cell count on the first postoperative day, for binary classification training. These models lack the ability to perceive the evolution of physiological states, meaning they cannot capture the entire process of a metric shifting from normal to abnormal, resulting in a profound lack of understanding of the temporal dimension. Furthermore, most of these models are black-box models, lacking clinical interpretability. They cannot clearly identify the parameter behind the current infection risk score shift, how long the shift has occurred, and in what direction. This directly prevents physicians from making credible decisions based on model results, and they may even completely abandon the model when conflicting results arise, severely limiting the practical application of intelligent approaches.

[0004] All of these current methods share a common problem: overreliance on a single threshold judgment logic. In real-world clinical practice, a single warning value is often set as a trigger, such as pH <6.8, a flow rate decrease >50%, or CRP >100 mg / L. However, this rigid threshold is highly susceptible to individual variability. Different age, surgical procedure, surgical area, underlying medical conditions, and even the use of postoperative drainage tubes and drainage methods can lead to significant variations in the applicable range of the threshold. This results in either a large number of false positives (incorrectly diagnosed as infections, but actually experiencing normal fluctuations) or a large number of false negatives (true infections that do not meet the abnormal value criteria). Furthermore, these methods generally fail to identify synergistic changes between multiple indicators, unable to determine whether a simultaneous directional reversal of two indicators represents a greater risk, or whether a sustained, slow shift in a parameter warrants more caution than a sudden, short-term change. Consequently, their ability to discern whether quantitative changes have accumulated into qualitative changes is severely limited. In terms of data collection, most hospitals currently rely on manual recording, intermittent sampling, and analysis, which prevents the formation of high-frequency, continuous time-series data streams. This directly limits the training and deployment of advanced models based on trend modeling and dynamic recognition, and also prevents recognition systems from capturing the most critical signal: changing physiological trends. Even some devices that support continuous recording of data such as drainage volume and conductivity lack integrated data interfaces and standardized indicator mapping mechanisms, resulting in severe data fragmentation and making unified modeling and analysis difficult. Summary of the Invention

[0005] The purpose of the present invention is to provide a method for identifying the risk of drainage fluid infection after orthopedic surgery, thereby solving some of the drawbacks and shortcomings pointed out in the background art.

[0006] The present invention solves the above-mentioned technical problems by adopting the following technical solutions: considering drainage fluid as a system output driven by postoperative local immune regulation and trauma recovery; capturing nonlinear time series curves including viscosity, color turbidity, flow rate, and pH indicators; and performing first-order derivative trend modeling on the curves to analyze the rate of change, including accelerated pH drop or sudden changes after stable flow rate;

[0007] Establish a framework for postoperative recovery pathways, using the dynamic parameters of drainage fluid from non-infected cases to form a standard recovery trend library; compare the Euclidean deviation and directional changes between the current patient's drainage parameter behavior vector and the standard trajectory;

[0008] The changes are converted into a visual recovery offset map, and the infection prediction data is measured from a multi-day trend perspective; then a framework is constructed to identify the transition from local abnormal signal aggregation to infection event formation, and the clinical infection node where quantitative change leads to qualitative change is obtained.

[0009] Furthermore, the nonlinear timing curve construction method includes:

[0010] A multi-parameter in-situ sensing module is integrated into the postoperative drainage channel to continuously collect the physical and chemical parameters of the drainage fluid. The in-situ sensing module collects these parameters at a minute-level granularity and timestamps each data point before inputting it into an edge computing unit. The edge computing unit then establishes a nonlinear time series behavior curve for each parameter and extracts its rate of change, acceleration, and abnormal inflection points.

[0011] The behavior curves of each parameter are compared with the standard postoperative recovery model to identify the degree of deviation and trend direction; when multiple parameters are identified to show a synchronous deviation trend within the same time window, it is judged as an early signal of infection risk transition and the infection risk score result is output.

[0012] Furthermore, the viscosity is measured by a micro-oscillation cantilever beam sensing structure, and the vibration attenuation characteristics of the cantilever beam are used to characterize the viscous behavior of the drainage fluid; the color turbidity is obtained by an optical detection module arranged in the drainage channel, and the optical detection module includes a channel for measuring direct absorbed light and a channel for detecting 90-degree scattered light, so as to simultaneously evaluate color and turbidity.

[0013] Furthermore, the flow rate and pH are acquired through synchronous acquisition, and their time phase difference is analyzed through behavioral curve analysis to determine whether local environmental acidification is accompanied by a sudden change in flow rate. The standard recovery model is constructed based on the time series behavioral data of various drainage fluid parameters in normal patients without infection after surgery, and is used as a control trajectory for subsequent individual identification.

[0014] This approach uses postoperative drainage fluid as a continuous system output dynamically controlled by local immune regulation and tissue repair. By collecting multiple key indicators (flow rate and pH), a behavioral phase shift function model is constructed to determine whether the current patient status deviates from the standard non-infection recovery path, thereby achieving dynamic identification and early warning of infection risks.

[0015] Unlike traditional methods that rely on static threshold judgment, a self-created integral function is introduced to simultaneously examine the relationship between the flow rate change rate and the pH directional deviation; the function is designed as follows:

[0016]

[0017] in:

[0018] Ψ(t) represents the overall intensity score of the deviation of the drainage behavior of the current patient from the standard recovery trajectory within a certain time window; the upper and lower limits of the integral t0 and t1 represent the start and end time of the current calculation window, which are usually set to 4 or 8 consecutive hours after surgery; V(t) represents the flow rate of the drainage fluid at time t, and its derivative is Indicates how fast the flow rate changes and is a key parameter for measuring the dynamic stability of the liquid;

[0019] φ(t) represents the direction of the pH trend of the drainage fluid, and its dominant phase is obtained by fitting continuous pH data (for example, continuous acidification can be converted to 180 degrees); θ0(t) represents the trend phase of pH in the standard non-infectious recovery model at that moment, and its data are derived from the recovery curve template established for a large number of non-infected patients;

[0020] The difference between the two, φ(t)-θ0(t), represents the deviation angle between the current pH behavior trend and the standard path, which is converted into the degree of directional consistency through the sine function; if the trend is consistent, the value approaches 0, and if the trend is completely opposite, the value approaches ±1;

[0021] In general, the integral calculation of this function reflects two characteristics: first, whether there is a sudden change in flow rate (i.e. is negative or changes sharply); second, whether the change trend of pH deviates significantly from the standard trajectory; when the two occur simultaneously, the absolute value of the product term increases and the integral result Ψ(t) rises significantly, indicating the risk of infectious transition in the local environment.

[0022] Furthermore, the edge computing unit includes an adaptive behavior feature extraction module for dynamically updating the baseline behavior curve of individual patients; the infection risk score result is output to the clinical terminal according to the risk level, and a clinical intervention prompt is triggered when the score value reaches a preset warning threshold.

[0023] Furthermore, the construction of the comparison method of the Euclidean deviation and the directional change includes:

[0024] S1. Construct a postoperative recovery path model, including collecting multi-parameter time series data of drainage fluid from non-infected patients and classifying and clustering them according to surgical type, location, and individual characteristics to form a standard recovery trajectory library;

[0025] S2. Acquire multiple continuous parameters of the postoperative drainage fluid of the current target patient within a set time window to form a multidimensional behavior vector;

[0026] S3, aligning the current patient behavior vector with the corresponding template in the standard recovery trajectory, and performing point-by-point comparison based on temporal consistency;

[0027] S4. Calculate the Euclidean deviation between the current behavior vector and the standard trajectory to quantify the overall degree of deviation; simultaneously calculate the trend direction change of each parameter, identify whether there is a trend in the opposite direction of the standard trajectory, and mark it as a directional outlier;

[0028] S5. When the Euclidean deviation exceeds the preset deviation threshold and multiple parameters show directional outliers at the same time, it is determined that the patient has a transition trend in the risk of postoperative infection and an early warning signal is output.

[0029] Furthermore, the standard recovery trajectory library generates multiple representative recovery curve templates through cluster analysis to adapt to different surgical procedures, surgical areas and patient physiological background differences; the drainage fluid parameter collection time window is 4 to 24 consecutive hours, and is resampled at a uniform time interval for time alignment with the standard trajectory.

[0030] Furthermore, the behavior vector is a unified vector structure formed by splicing multiple indicators according to timestamps; the directional outlier determination includes comparing whether the rising or falling trend of the current indicator is opposite to the trend direction of the corresponding time period of the standard trajectory, and recording the outlier event when the trend is reversed;

[0031] In the above scheme, the key drainage fluid indicators of flow rate, pH, and turbidity are constructed into a unified time vector structure, and a custom directional function is used to determine whether their changing trends in each time period are synchronized or opposite to the standard recovery trajectory, thereby identifying behavioral deviation events.

[0032] In order to achieve this goal, a newly constructed trend direction outlier function is introduced, which is expressed as follows:

[0033]

[0034] This formula determines the trend direction of all n key behavioral indicators at the current time point t and outputs a total outlier score Δ(t). The meaning of each variable is as follows:

[0035] Δ(t): represents the outlier strength score of the trend direction between the current patient and the standard recovery model at time point t. A larger value indicates a more severe trend reversal and a higher risk of infection. n: the number of key drainage fluid parameters monitored, such as flow rate, pH, turbidity, and temperature. σ i (t): The trend direction of the current i-th indicator at time point t, defined as a sign function, which takes +1 if it continues to rise, -1 if it falls, and 0 if it remains unchanged. η i (t): The absolute value of the rate of change of the current i-th indicator in this period, that is, the strength of its trend significance, which is used to weight the behavioral impact of different indicators. ξ i (t): The theoretical trend direction of the indicator at time t in the standard recovery trajectory, with the same value as σ i (t), that is, +1, -1 or 0.

[0036] The bracketed term [1-ξ i (t)] is introduced to emphasize that when the standard trend is rising (i.e. ξi (t)=+1), while the current is decreasing (σ i (t) = -1), the product is negative; this term approaches zero when the direction is the same, while the total outlier is amplified when the direction is opposite. Therefore, this function Δ(t) essentially focuses on the opposite deviation of the trend direction and combines it with the indicator weight to determine whether the deviation is severe enough.

[0037] Furthermore, the Euclidean deviation and the directional outlier result are used jointly for risk identification. Only when both meet the warning conditions at the same time, it is determined that there is a tendency for early infection transition.

[0038] Beneficial effects of the present invention:

[0039] Through continuous dynamic monitoring and nonlinear trend modeling of key physiological parameters in drainage fluid (such as pH, flow rate, turbidity, viscosity, etc.), the system not only identifies whether the indicators are abnormal, but also determines whether the changing trends indicate signs of potential infection transitions. Compared with traditional methods that rely on body temperature, redness, swelling, or laboratory culture results, the warning time is an average of 24-48 hours earlier, creating a valuable intervention window for clinicians. By introducing a standard recovery trajectory library and performing hierarchical cluster matching based on factors such as surgical type, anatomical site, age, and physiological background, each patient will receive a unique set of postoperative health trajectory control templates. This, combined with the adaptive behavioral benchmark adjustment mechanism in the edge computing unit, dynamically calibrates the individual recovery path.

[0040] All monitoring indicators are uniformly constructed as time behavior vectors and calculated through trend-oriented functions, integral models and outlier scoring mechanisms. Each warning is accompanied by a visual trend chart, indicator direction comparison and risk score curve. Doctors can clearly see which parameters begin to deviate and when, and whether the deviation direction is dangerous, which facilitates medical decision-making and solves the problem of unexplainability of black box models. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 This is a flow chart of the method for identifying the risk of drainage fluid infection after orthopedic surgery of the present invention.

[0042] Figure 2 This is a flow chart of the method for constructing a nonlinear timing curve of the present invention.

[0043] Figure 3 A flow chart is constructed for the present invention's method for comparing Euclidean offset and directional change.

[0044] Figure 4 This is an embodiment diagram of the peroneus longus tendon passing through the drainage wound of the present invention. DETAILED DESCRIPTION

[0045] The following is a detailed description of the specific embodiments of the present invention with reference to the accompanying drawings.

[0046] Combined with attachment Figure 1The present method for identifying infection risk in postoperative drainage fluid in orthopedics considers drainage fluid as a continuously generated system output variable controlled by the combined action of multiple physiological mechanisms. The state of drainage fluid is a comprehensive expression of multiple factors, including the local postoperative immune response, tissue damage and repair processes, cytokine release, and microbial intervention. The key to this method is to view drainage fluid as an externally observable expression of system behavior, rather than simply a biochemical waste product. This allows changes in its behavior to reveal potential physiological abnormalities, particularly early signs of infection. The system requires continuous acquisition of multiple core drainage fluid parameters, particularly viscosity, color, turbidity, flow rate, and pH. These parameters are not randomly sampled; rather, they are confirmed through medical research to often undergo abnormal changes before infection occurs. For example, pH typically decreases continuously (localized acidification) in the early stages of bacterial infection, while flow rate can increase or decrease sharply for a short period of time due to increased inflammatory exudate. Color and turbidity reflect changes in components in the fluid, such as proteins, cells, and pathogens, while viscosity is related to contents such as cellulose and inflammatory exudate. In terms of data acquisition, the system relies on an integrated intelligent drainage module embedded with a micro-MEMS sensor array, enabling real-time measurement of the physicochemical parameters of the drainage fluid without compromising drainage patency. Sensor data is collected every minute, forming a high-density time series. All data is initially cleaned and cached by an edge processing chip. Raw data preprocessing includes sensor self-calibration (drift removal), missing value interpolation (based on nearest neighbor time periods), outlier filtering (based on standard deviation multiples), and time normalization (aligning all data to a unified time axis). After data preparation, the modeling phase begins. This approach, rather than employing traditional static thresholding, models and analyzes nonlinear behavior curves. Specifically, each drainage fluid parameter is treated as a continuous function with respect to time, and its first-order derivative—the rate of change at each moment—is calculated to identify trending events. For example, if the first-order derivative of the pH curve suddenly changes from a slow decline to a rapid decline, the system identifies accelerated acidification. Similarly, if the flow rate curve is stable for a long period of time but the derivative becomes large or negative and steep at a certain time, this is identified as a sudden flow rate fluctuation. This processing approach enables the system to not only identify abnormal values, but also detect signals that behavior is deviating from its normal trajectory. Based on this, the entire model uses a sequence analysis algorithm based on a sliding time window. Each behavioral indicator forms a separate trend scoring sub-model, which is combined into a unified infection risk scoring network. In terms of technical implementation, the model structure is comparable to a shallow LSTM or GRU unit, with a simple structure that retains the characteristics of short-term series variation. During training, the system uses a historical dataset of orthopedic postoperative drainage fluid, divided into non-infected and infected groups, and annotated with the windows where infection occurred in each time series.The model aims to identify these windows and, through supervised learning (cross-entropy loss function), continuously adjusts trend-sensitive parameters to ensure that the final infection prediction output is highly consistent with the true label. The training process uses a rolling time window mechanism to enhance the model's adaptability to characteristics of different time periods; label smoothing and data augmentation strategies are also introduced to improve the model's robustness to mild trend shifts. Ultimately, the system no longer relies on whether a single parameter has crossed a certain threshold, but instead identifies it through the synergy of trend mutations across multiple indicators. This involves determining whether multiple indicators exhibit a combination of direction reversals, acceleration anomalies, and synchronized fluctuations within a similar timeframe.

[0047] By establishing a framework for postoperative recovery pathways and constructing a standardized recovery trend library comprised of dynamic drainage fluid parameters from a large number of non-infected cases, this approach can be used to measure behavioral deviations in current postoperative patients during their actual recovery process, thereby enabling dynamic monitoring and early identification of potential infection risks. The entire approach is based on the premise that the occurrence of infection often causes local physiological states to deviate from the normal recovery trajectory. This deviation is not always manifested as a single indicator exceeding the standard, but rather as multiple physiological parameters gradually deviating from the normal evolution path over time. Therefore, the first step in this approach is to construct a library of standard trajectories. This involves extracting continuous drainage fluid parameters, including pH, flow rate, color turbidity, viscosity, and other indicators, from a large number of previously uninfected patients after surgery. All data are collected hourly from the first to fifth postoperative day to form a high-frequency time series. To ensure consistency, these data are first normalized (e.g., using Z-score or min-max scaling) to remove individual baseline differences and indicator dimension differences. Each indicator is then resampled (to ensure consistent time series length for each case), outliers are processed (e.g., to eliminate data drift caused by collector blockage or special patient interventions), and hierarchical cluster analysis is performed based on the patient's surgical procedure type (e.g., joint replacement, spinal fixation, fracture fixation), surgical site (e.g., hip, knee, lumbar), and underlying pathological factors (e.g., diabetes, presence of implants). Similar recovery trends are grouped together to generate multiple standard recovery trajectory templates. Each template is a multidimensional, multi-time-point behavioral matrix that serves as a reference for the recovery path. During the patient's postoperative drainage process, the system continuously collects various parameters of the drainage fluid and generates a behavior vector slice in real time within a set time window (for example, 6 hours, 12 hours, and 24 hours after surgery). This vector structure is composed of the continuous values ​​of all indicators spliced ​​according to timestamps, forming a multidimensional time series vector. The system then aligns this vector point by point with the standard template trajectory that matches the corresponding surgical procedure and patient background. In a point-to-point manner, it calculates the Euclidean distance between the behavior vector and the standard trajectory, that is, it calculates the overall offset degree of all sampling points in the multi-parameter space. This process can be understood as the quantification of the distance between the current trajectory and the healthy trajectory in the time series multidimensional space. However, this geometric shift isn't enough to reveal the true risk, so the system further introduces trend direction analysis: for each parameter curve, its local trend direction is extracted within the current time period (for example, the difference between the three-point sliding average is calculated to determine whether it is rising, falling, or flat). This is then compared with the trend direction of the corresponding time period in the standard trajectory. A consistent trend direction is scored as 0, while a completely opposite trend direction is scored as 1. Finally, the directional differences of all parameters within the time window are weighted and summed to obtain a directional outlier score. This score is used to assess whether the current physiological behavior is heading in the wrong direction, not just the distance.The core of the entire recognition model is the fusion modeling of the aforementioned Euclidean deviation and trend directional outlier scores. In the algorithm implementation, this component is modeled using a shallow graphical neural network structure to incorporate the interdependencies between different parameters. For example, in the early stages of infection, a sudden change in flow rate and a drop in pH occur simultaneously; this synergistic change is an important pattern signal. The model is trained using a labeled historical dataset of real postoperative infection and non-infection cases. Through supervised learning, the model is trained to identify which deviation combinations represent infection risk. A focal loss function is introduced during training to enhance the model's sensitivity to early deviation features, and cross-validation is used to optimize the model parameters. The final output is an infection risk score, which is based not only on whether a specific indicator is abnormal, but also on a dynamic judgment based on the spatial distance and directional differences between the patient's current overall physiological trajectory and the standard healthy trajectory.

[0048] To determine whether the immediate state of the drainage fluid is abnormal, the system tracks data over multiple days, transforming the temporal trends of all key parameters into a visual recovery offset map. This allows for the identification of clusters of local abnormal signals and, based on this, a transition framework for infection events is constructed, capturing the point where quantitative change leads to qualitative change, i.e., the turning point where early infection mutates into a clinical infection state. Regarding data acquisition, the system uses a microsensor array embedded in the postoperative drainage pathway to continuously collect multiple physiological parameters of the drainage fluid, including but not limited to pH, flow rate, turbidity, color, and viscosity. The acquisition cycle is set to every minute to every ten minutes, forming a continuous time series data stream. Each time series data item is timestamped with a high-precision timestamp, and the system performs preliminary data processing in real time at the edge computing end. This primarily includes time series alignment, outlier removal (such as extreme jump points and physical interference), data normalization (using zero-mean unit variance or interval mapping), and sliding smoothing (such as local weighted regression or Savitzky-Golay filtering) to ensure that the analyzed trend signals are stable and comparable. In terms of data structure, the system extracts key trend features of each indicator during each monitoring cycle (typically every 12 or 24 hours), such as local growth rate, decline rate, peak occurrence time, and magnitude of change. These features are embedded in a two-dimensional space to construct a recovery offset map. The horizontal axis of this map represents time (hours or days after surgery), and the vertical axis represents the offset strength or degree of outlier trend direction for each parameter. The map uses heat, color gradients, or markers to indicate the degree of deviation of parameter behavior from the standard trajectory. This visualization not only provides an intuitive reference for physicians but also serves as important structural feature data input for subsequent model training, including deep trend recognition. The core innovation of this method lies in constructing an infection transition identification framework. Unlike traditional rule-based models, this framework uses a dynamic evolutionary mechanism to identify the process by which potential abnormal behaviors aggregate and evolve into infection events. In terms of algorithmic design, the system continuously monitors hotspots in the recovery offset map, particularly for sudden changes, reversals, or accelerations that occur simultaneously in multiple parameters at similar time points. When the offset magnitudes or directions of multiple indicators within a time window are abnormal and clustered within the same local time interval, the system identifies this area as an abnormal aggregation zone. These abnormal clusters are then fed into a two-stage model. The first stage uses a graph convolutional network (GCN) to construct a graph structure based on the temporal proximity between abnormal events, extracting the degree of aggregation and structural features between signals. The second stage uses a recursive structure (such as a Bi-GRU or Transformer encoder) to capture the evolution of abnormalities and assess whether they are transitioning to a systemic state, that is, from a subclinical abnormal state to a true infection risk. The model uses historical real-world clinical case data as training samples, with each case annotated with whether the infection occurred, the time of occurrence, and the path of parameter abnormality evolution.Before training, all case data is first converted into a unified format for a recovered offset map using the aforementioned process. Then, abnormal clusters are extracted and annotated (e.g., the time shift from the clinical diagnosis of infection to the cluster). Training utilizes a multi-task loss function, optimizing the accuracy of identifying transition events (e.g., using binary cross-entropy) while constraining the offset error of predicted time nodes (e.g., using a smoothed L1 loss or temporal regression loss). During training, the model uses a dynamic learning rate and early stopping strategy to prevent overfitting, and incorporates a reweighting mechanism for difficult examples to improve the ability to identify atypical infection pathways. The system ultimately outputs a dynamic risk score curve and a time estimate for predicted transition points. This information can be fed back to healthcare professionals in real time for early warning and can also be used to generate a disease progression diagram for subsequent decision-making. The key advantage of this approach is that it no longer relies on static judgments about whether a single item is abnormal on a given day. Instead, it builds a mechanism to identify how multiple subclinical abnormalities gradually aggregate over time and trigger risk mutations. This allows for high-probability prediction and early warning before actual clinical diagnosis of infection.

[0049] Example 1:

[0050] In this example, a 65-year-old patient, Mr. Zhang, underwent left knee arthroplasty and was installed with an intelligent drainage monitoring system on day 0. The system includes a built-in multi-parameter in-situ sensing module, including a solid-state pH electrode, a micro-thermal film flow sensor, a near-infrared optical turbidity probe, and a viscosity microbeam vibration detection unit. The system collects drainage fluid parameters every 5 minutes, continuously recording flow rate (mL / h), pH value (dimensionless), turbidity (relative units), and viscosity (cP). All data is timestamped and sent to the bedside edge computing unit for real-time processing. The observation period is the first 8 hours of the first postoperative day, and a total of 96 time points are collected. The edge computing module cleans the raw data (eliminating transient jumps) and normalizes it (using Z-score normalization), then constructs a nonlinear time series curve for each parameter. For example, Mr. Zhang's pH value began to decline continuously from 7.21 in the first hour after surgery, and dropped to 6.96 in the fifth hour. During this process, the first derivative of the pH behavior curve (i.e., the rate of change) gradually decreased from -0.02 to -0.07, and the second derivative of the curve also showed an obvious negative inflection point, indicating that acidification was accelerating; the flow rate remained stable in the first 4 hours after surgery (about 28 to 30 mL per hour), but suddenly dropped to 21 mL / h from the fifth hour, accompanied by fluctuations. The first derivative of the flow rate curve suddenly changed from close to 0 to -9, and fluctuated repeatedly in the next 2 hours, and its acceleration showed an abnormal reversal pattern; the turbidity and viscosity curves also showed brief abnormalities in the 5th to 6th hour: the turbidity changed from the original linear decline to a steady increase, and the viscosity increased from 1.2 cP to 1.8 cP. The system synchronously compares the behavioral curves of the above four parameters with the standard recovery model - this model was established based on 102 non-infected patients of the same age as Mr. Zhang, who underwent the same surgical procedure and had no history of infection. The standard model shows that the pH should slowly decrease on the first day and then stabilize, the flow rate should slowly decrease but not drop suddenly after the fifth hour, and the turbidity and viscosity should gradually decrease without an upper inflection point. The system uses the Euclidean distance method to calculate the average behavioral deviation between Mr. Zhang's current data and the standard trajectory. The result is 2.63 (the average deviation of the standard group is between 0.5 and 1.2), which is significantly higher. At the same time, trend direction analysis shows that three of the four indicators are in the opposite direction of the standard trajectory in the 5th to 6th hour period (for example, the standard pH stabilizes but the actual decline accelerates). The system marks this window as a directional outlier. Since more than three indicators deviated synchronously during this time period and the trend direction reversed, the system aggregated and marked this behavior as a potential infection transition node based on the built-in infection transition prediction model. Combined with the previous historical data of physiological parameters and the hot spot area formed in the recovery offset map (i.e., a concentrated outbreak of multiple offset points), the final output infection risk score was 78 / 100, which was much higher than the set threshold of 65. The system immediately sent an early warning to the attending physician and nurse's mobile terminal, and recommended bacterial culture of the drainage fluid, rechecking of white blood cell count and C-reactive protein, and early intervention.The results showed that Mr. Zhang was diagnosed with Staphylococcus aureus infection on the third day. However, because the system captured the abnormal trend on the first day, the medical team began to adjust the antibiotic and drainage plan on the second day, avoiding the risk of infection spread and reoperation.

[0051] Viscosity measurement utilizes a micro-oscillating cantilever beam sensor structure, a micro-MEMS device. Essentially, when a cantilever structure oscillates at a controllable frequency and is immersed in drainage fluid, the fluid's viscosity influences the cantilever's vibration amplitude decay curve. As the viscosity of the drainage fluid increases, the damping force on the cantilever increases, manifesting as a faster decay rate for the oscillation signal. By monitoring the energy dissipation rate during the oscillation cycle and its nonlinear decay over time, the system constructs a highly sensitive damping-time response curve, from which the actual viscosity of the fluid can be inferred. The entire vibration response signal is controlled collaboratively by a chip-level integrated accelerometer and drive electrodes, achieving an accuracy of 0.01 cP and supporting a sampling frequency of 5 Hz. In actual deployment, the system continuously measures data with a sampling period of one minute. The data is de-noised and filtered by the sensor chip's internal pre-processing unit before being written to the edge computing system for viscosity behavior curve generation and derivative analysis. Meanwhile, color turbidity is acquired by an optical detection module located within the drainage channel. This module consists of two optical pathways: a direct channel, which assesses the forward absorption of light at different wavelengths (e.g., 470 nm, 525 nm, and 630 nm) by the drainage fluid. This path is sensitive to changes in drainage fluid color (e.g., red, pale yellow, and pink) across the RGB dimensions. A 90-degree side scatter channel, specifically designed to measure the scattering intensity of incident light by particles in the fluid (e.g., cell debris, lipid droplets, bacteria, and exudate proteins), is highly sensitive to turbidity. The system utilizes a synchronized emission and acquisition mechanism, ensuring simultaneous measurement in both channels. This generates a complete pair of color-turbidity optical vectors within each sampling period. After acquisition, the data first enters the edge module for spectroscopic analysis and intensity normalization (to eliminate nonphysiological factors such as light source attenuation and optical path contamination). It is then assembled into a time series with subsecond intervals and fed into the nonlinear curve modeling module. In Mr. Zhang's case, the data continuously recorded by the optical module showed that the drainage fluid had a stable color in the absorption peak λ1 = 525nm region (light yellow) for the first four hours after surgery, and the scattered light intensity remained in the range of 40-43 (low turbidity). However, within the fifth to sixth hour, the direct absorption suddenly shifted to the red, the λ1 peak intensity dropped to 32, and the turbidity scattering channel intensity rose to 62, showing obvious characteristics of protein and red blood cell leakage and inflammatory cell accumulation. These data are immediately fed into the model. The system first maps the original absorption-scattering pairs into a color concentration matrix (analog vectors obtained by training the Kohonen self-organizing map network). Then, combined with the accelerated pH drop and the sudden drop in flow rate in the same time period, an abnormal superposition is formed in the multi-indicator trend fusion module, which is pushed to the next stage of the infection transition identification network.In terms of model structure, the system utilizes a segmented deep fusion neural network. The underlying input is a sequence of nonlinear derivatives and trend classification labels for each indicator. A multi-head attention mechanism is employed to extract interplay between different parameters within a specific time window, such as the temporal synergistic burst of a pH drop and turbidity increase. During training, the system uses a large dataset of postoperative drainage fluid from a hospital (a complete parameter set containing nearly 1,400 cases labeled with infection / non-infection). Each sample is a continuous time series segment labeled with infection and the hourly offset before infection. The loss function is a dual objective function (infection classification error + time offset prediction error). Training uses the Adam optimizer with dropout and regularization to mitigate overfitting. The resulting model accurately identifies abnormal clustering points 12 to 48 hours before infection, with an average prediction time error of less than 3 hours. Therefore, the abnormal turbidity and sudden increase in viscosity that occurred in Mr. Zhang's fluid 5 hours after surgery, as early signs of a localized infection, were accurately captured and aggregated into trend-based events. Combined with the trained model's transition recognition capabilities, a high-risk score was ultimately assigned, triggering a proactive alert notification.

[0052] Mr. Zhang was continuously monitored from 9:00 a.m. on the first day after surgery. The system began to collect the drainage fluid flow rate V(t) and pH value every 5 minutes. In a 4-hour time window [t0, t1] = [9:00, 13:00], a total of 48 time point data sequences were obtained.

[0053] In the data preprocessing stage, all the data collected at the points are first linearly interpolated to fill a small number of gaps, and the noise is filtered using wavelet denoising (db4 wavelet, 3-layer decomposition). After obtaining a smooth curve, the velocity derivative at each time point is calculated. and the dominant phase φ(t) of the pH fitting curve. For example, around 13:00, 5 hours after surgery, Mr. Zhang's flow rate dropped from the stable value V(t)≈30mL / h in the first 30 minutes to 21mL / h. (Unit: mL / h 2 ), which is a significant sudden drop; at the same time, the pH value dropped from 7.18 to 6.95. In the sliding fitting curve, the dominant phase φ(t) reached about 180° around 13:00 (indicating continued acidification), while according to the standard recovery template, the pH should have stabilized 5 hours after the operation, and the reference phase of the standard model θ0(t) was about 30° (i.e., the pH maintained a slight recovery or remained flat).

[0054] Substitute the above measured data into the formula:

[0055]

[0056] The numerical approximation is performed using discrete integration (trapezoidal method), with a sampling interval of 5 minutes, a lower limit of integration of t0 = 9:00, and an upper limit of t1 = 13:00. In the data segment around 13:00, the following measured values ​​are set:

[0057] φ(t)=180°, θ0(t)=30°, then sin(150°)≈0.5

[0058] The single product is: |-18·0.5|=9

[0059] Assume that the anomaly lasts for 30 minutes (6 points), the integral contribution of this interval is 9 0.5 = 4.5, the width of each point is 5 minutes (0.083 hours), and the total contribution is approximately 6 × 4.5 × 0.083 = 2.24

[0060] The system also calculates the normal behavior segments at other time points. In the normal segment, Or φ(t)≈θ0(t), so the product term tends to 0 and contributes very little to the integral value.

[0061] Therefore, the integration over the entire time window results in:

[0062]

[0063] According to the statistics of model training data, the average value of Ψ(t) for standard recovered patients ranges from 0.3 to 1.1; while the Ψ(t) for confirmed infection cases is mostly concentrated between 2.0 and 4.5. Therefore, the Ψ(t) = 2.6 calculated by Mr. Zhang during this period is already in the high infection risk range.

[0064] The range of parameters involved in the model includes: Normal fluctuations are usually between -5 and +5, while those exceeding ±10 are considered mutations. After the angle difference of φ(t)-θ0(t) is converted into radians, its sin value is generally within ±0.3, indicating a consistent trend, while those exceeding ±0.6 are considered to be trend reversals. Therefore, the maximum value of the product can theoretically reach 20 (such as extreme mutation + trend reversal). In this system, a value exceeding 8 is considered a strong anomaly.

[0065] The model structure adopts a lightweight convolution-attention fusion network, and the input features are the The model uses φ(t), θ0(t), raw flow velocity, pH value, and their second-order derivative features. The first layer consists of two 1D-CNNs to extract behavioral change rate features. The second layer uses a multi-head self-attention mechanism to extract trend aggregation patterns in the temporal phase difference sequence. The final layer is a fully connected classifier that outputs a risk score. Training was performed using real case data (n=1187, 284 positive samples). The training and validation accuracy reached 92.4%, with an average early prediction window of 38 hours. During training, the function integral value Ψ(t) is used as a continuous risk label for each sample, guiding the model to learn to extract risk transition signals from nonlinear offsets.

[0066] Finally, 5 hours after the operation, the system successfully identified the synergistic reaction of Mr. Zhang's decreased local immune system control ability, accumulation of tissue metabolites and mutation of exudate behavior. Through the behavioral phase integral model, it issued an early warning of Staphylococcus aureus infection 36 hours in advance. After culture verification and antibiotic intervention, the patient successfully avoided a second surgery in the surgical area.

[0067] During the first 24 hours after Mr. Zhang's surgery, the system collected nearly 300 complete time series data points, covering five dimensions: pH, flow rate, turbidity, viscosity, and color. After collection, each data point was first quickly verified, including edge filters to remove ultra-physiological abnormal points (such as interference with a flow rate of 0 or above 100 mL / h), noise signal processing (based on time domain smoothing or frequency domain noise suppression), and then the standardization module performed dynamic Z-score normalization. This process not only considers group statistical data, but also integrates Mr. Zhang's personal preoperative physiological data and the average of the stable period within the first 6 hours after surgery to ensure that the standardized curve is both globally comparable and retains individual trend characteristics. Based on this, the adaptive behavioral feature extraction module built into the edge computing unit was activated. The system used Mr. Zhang's early stable behavioral segment on the first postoperative day as his initial individual behavioral baseline template. It then reanalyzed the nonlinear behavioral curve for the past six consecutive hours (i.e., a sliding window) every hour, extracting features such as first-order derivatives, second-order derivatives, trend extremes, inflection points, and acceleration curvature. These features were then compared for trend similarity with the individual's initial template. Multi-channel alignment was also performed with a standard template for the group matching the procedure. The system introduced a fusion loss function to balance the deviation scores between the true individual behavioral curve and the group recovery model. Specifically, when a metric of Mr. Zhang consistently deviated from the group average trend but remained consistent with his previous trajectory, the system reduced the sensitivity of his risk score. Conversely, when a sudden behavioral change or a reversal of trend occurred, the system immediately increased the weight of that behavior in the risk score. The key to this mechanism is that the model's behavioral feature extraction no longer relies on static template matching, but instead updates the individual recovery trajectory hourly, fine-tuning the baseline model to achieve a truly adaptive prediction system. The entire prediction and scoring module is based on a lightweight stacked residual network (Residual Temporal CNN + LSTM). The network structure is optimized and deployed on the edge computing unit to ensure low power consumption and high-speed reasoning capabilities. The model input is the behavior vector of each indicator in the past 6 hours (including its original value, derivative, directional difference, pH phase, flow gradient and other 28-dimensional features), and outputs a dynamic infection risk score R(t). The score ranges from 0 to 100. The model constructs a scoring mechanism through the sigmoid layer and standardized regression, and sets the threshold T warn=65 to determine whether the clinical intervention process needs to be triggered. The system scored R(t) = 62 at the 30th hour, indicating a moderate deviation; at the 34th hour, it rose to R(t) = 71, crossing the warning threshold, and immediately triggered the built-in clinical intervention prompt module. The system pushed alarm information synchronously through the nurse terminal and the doctor's mobile device, accompanied by auxiliary information such as the indicator trend chart for the past 6 hours, the current behavior deviation intensity score, and the combination of indicator synergy abnormalities (such as continuous pH decrease + drastic flow rate fluctuation + turbidity trend reversal). It also recommended immediate implementation of the following countermeasures: drainage fluid bacterial culture, local B-ultrasound assessment of the presence of abscess cavities, and consideration of early adjustment of empirical antibiotics. This prompt was adopted by the clinical team and handled in advance. Mr. Zhang was subsequently diagnosed with Staphylococcus aureus infection on the third day, but due to sufficient early warning, the infection was controlled in situ and did not affect the prosthesis, avoiding reoperation and functional impairment. The training process of the risk scoring system is based on a multi-stage transfer learning framework. The basic network is first pre-trained on a national public dataset of drainage fluid parameters (n≈4500). It is then fine-tuned on a set of postoperative infection cases (n=983, of which 231 are positive) after being labeled and cleaned at our hospital to improve the model's sensitivity to local behavioral variations. The training loss function is the risk score error + alarm time offset penalty + false positive penalty coefficient. The optimizer is AdamW, and the learning rate scheduler is dynamically adjusted within the interval [1e -4 ,5e -6 ], during the training process, each epoch dynamically updates the residual gating weight of the model for the individual behavior trajectory, so that the prediction results can be adjusted in real time according to the patient's performance.

[0068] In summary, the system's edge computing architecture and adaptive behavior modeling module not only complete continuous, multi-parameter, high-precision data processing, but also, through deep individualized curve management, dynamic scoring output and closed-loop intervention prompt mechanism, constructs a highly robust intelligent infection risk identification method that is truly suitable for orthopedic postoperative scenarios. It is not only technically feasible, but also highly practical clinically.

[0069] Example 2:

[0070] Based on Example 1, during the 8 hours from 2:00 a.m. to 10:00 a.m. on the first day after surgery, the system continuously collected four key indicators of Mr. Zhang's drainage fluid: pH, flow rate, turbidity, and viscosity. The sampling frequency was once every 10 minutes, totaling 48 sets of data points.

[0071] In step S1, the system first selected 187 samples that met the requirements of Mr. Zhang's surgery type (left knee replacement), location (knee), age group (60-70 years old), and whether an implant was used (yes) based on the historical data of 1,026 non-infected patients in the hospital's existing infection monitoring database. The drainage fluid parameters of each patient in the first 48 hours after surgery were extracted to form a four-dimensional time series data set. The data was aligned with the standard time axis (for example, the first hour after surgery, the 12th hour after surgery, etc.), and the K-means clustering method was used to group similar recovery trajectories, ultimately forming 6 standard recovery template curves, each template representing a typical recovery path. Mr. Zhang was matched to the third type of template, with the corresponding characteristics of a steady pH drop, a gradual decrease in flow rate, a slow drop in turbidity, and a viscosity close to normal.

[0072] Entering the S2 stage, the system extracts the continuous behavior data of Mr. Zhang’s four parameters in the current monitoring window (2 a.m. to 10 a.m.), and splices them into a multidimensional behavior vector X = {x1, x2, ..., X 48}, each vector element represents a set of parameters at a certain moment (every 10 minutes), such as [pH(t), V(t), C(t), η(t)], where C represents turbidity and η is viscosity.

[0073] In S3, the system compares Mr. Zhang's behavior vector with the standard vector of the matching template on the same time axis. Perform point-by-point alignment and ensure time consistency to ensure that Mr. Zhang's data at the 6th hour after surgery is only compared with the data at the 6th hour after surgery in the template.

[0074] S4 is the calculation core of this method. The system first calculates the Euclidean offset between the current behavior vector and the standard template, which is in the form of:

[0075]

[0076] Substituting Mr. Zhang's data, assuming the mean deviation of each point at each moment is 0.35 and the square of the deviation is 0.1225, the total of the 48 points is approximately 5.88. Taking the square root, we get the Euclidean distance D ≈ 2.42. This value, on average for the non-infected recovery trajectory, ranges from 1.0 to 1.6, so a value of 2.42 significantly deviates from the population norm.

[0077] The system also determines the trend direction for each parameter set separately by calculating the sign of the sliding first-order derivative of each indicator within the current window (+1 for rising, -1 for falling, and 0 for stationary). The system then compares the sign of the derivative of the same indicator in the template. When the directions are opposite (for example, the pH is currently falling while the standard should be rising), it is recorded as a directional outlier. Mr. Zhang's pH, flow rate, and turbidity all showed a direction opposite to the standard during this period, with only the viscosity trend being consistent. The system recorded three-quarters of the indicators as directional outliers, and this trend persisted for at least three consecutive sampling points (i.e., the trend remained opposite for 30 minutes or more).

[0078] In S5, the system compares the Euclidean deviation D = 2.42 with the internal threshold T D =2.0, the result is exceeded; the number of directional outlier indicators is ≥3, and the continuous duration is ≥3 points (30 minutes), which meets the abnormal aggregation conditions. The system determines that there is an infection jump risk in the current period and immediately generates an early warning signal R(t) = 83 / 100, which far exceeds the trigger threshold T R =65.

[0079] The clinical response mechanism was triggered synchronously, and the system pushed the scoring curve, trend graph, and parameter deviation radar chart to the nurse station terminal. When the doctor on duty received the push, Mr. Zhang had no visible drainage abnormalities or fever. Microbial testing of the drainage fluid was performed 24 hours in advance, and antimicrobial pretreatment was initiated. Mr. Zhang was ultimately confirmed to be positive for infection on the third postoperative day, but timely inflammation control prevented the spread of prosthetic infection. Mr. Zhang did not require further surgery and was successfully transferred to the rehabilitation department on the seventh postoperative day.

[0080] This whole process example shows that the five stages from standard recovery path modeling, patient individual behavior vector generation and comparison, Euclidean distance calculation and trend direction analysis to warning signal output are highly engineered and clinically feasible. The adjustable range of parameters that the core model relies on is: Euclidean threshold setting range T D =1.8-2.2, the trend outlier ratio is determined when ≥50% of the indicators are in the opposite direction for more than three consecutive time points. The scoring result R(t) is output according to the combination logic and weighting function, and the scoring range is 0-100, where 65 is the high-sensitivity trigger threshold.

[0081] During Mr. Zhang's postoperative infection risk monitoring process, in order to accurately identify the trajectory of his drainage fluid behavior and make personalized trend deviation judgments, the system adopted a standard recovery trajectory library construction and time alignment analysis technology route. Through cluster analysis, a diverse recovery trajectory template was constructed, and the patient's drainage fluid parameters collected in real time were mapped to the standard template to achieve dynamic comparison with the expected recovery path, thereby identifying the precursors of infection. First, in the data acquisition stage, the Hospital Infection Risk Assessment Center sorted out the inpatient surgery data of the past three years and selected a total of 1,412 orthopedic patients who had no postoperative infection, covering nine common surgical procedures such as hip replacement, knee replacement, spinal fusion, and humeral internal fixation. The patients were labeled and sorted according to the surgical area (such as hip, knee, spine), surgical method (minimally invasive or open), underlying disease (such as diabetes, hypertension), preoperative CRP value and age group. Each patient included at least 48 hours of continuous drainage fluid data after surgery. The collected parameters include pH, flow rate, turbidity, viscosity and color spectrum absorption value. The sampling period is 10 minutes, and the timestamp is recorded. During the data preprocessing phase, the system first performed a temporal integrity check, eliminating cases with more than 15% missing segments or continuous abnormal jumps, leaving 1186 valid samples. All data sequences were then linearly interpolated to correct small missing segments and uniformly resampled, with time intervals of equal length (10, 15, and 30 minutes) selected to ensure that all patients had the same temporal dimension after alignment. For example, if a 24-hour window was selected and sampling was performed every 15 minutes, each curve would contain 96 sampling points. All parameters were uniformly converted into a matrix structure, resulting in a 5×96 parameter behavior matrix for each patient. Cluster analysis employed the K-means DTW (Dynamic Time Warping) (DTW) distance metric for time series. Because different behavioral patterns have similar morphology over time but differ in displacement or rate, traditional Euclidean distance is insufficient to reflect the inherent trend of the sequence. DTW allows for nonlinear temporal alignment. The system initially set the number of clusters, K, to 10. Clustering effectiveness was evaluated using the silhouette coefficient and Davies-Bouldin index, with optimal discrimination achieved at K = 6. Each cluster center represents a standard recovery trajectory template and is marked with the characteristics of the population it is suitable for. For example, the pH value first decreases and then increases after surgery - suitable for young patients who recover quickly from knee replacement surgery, or the flow rate decreases slowly but the turbidity increases slightly in the early stage - suitable for elderly hip implant patients, etc.

[0082] When Mr. Zhang entered the postoperative monitoring system, the system first matched Template 3 from the standard trajectory library based on his surgical procedure (open left knee replacement), age (65 years), and underlying medical conditions (hypertension, impaired glucose tolerance). This template represents a recovery pattern characterized by a steady decline in pH over the 24 hours following surgery, a slow decline in flow rate after the eighth hour, a rapid decrease in turbidity after the 12th hour, and a stable viscosity pattern. The system collected Mr. Zhang's drainage fluid data in real time, sampling every 10 minutes. The analysis window spanned 12 consecutive hours, from the fourth to the 16th hour after surgery. Data for the five parameters were collected at 72 time points, forming a 5×72-dimensional observation matrix. Since the standard template has a 5×96 structure, the system linearly interpolated Mr. Zhang's data to align its temporal dimensions with the template, resulting in a 5×96 current behavior matrix. A time alignment algorithm then synchronously compared the trend orientation, numerical offset, and gradient changes at corresponding time points between the two models, providing input features for subsequent Euclidean offset calculation, trend direction determination, and integral scoring modeling.

[0083] The trend direction outlier function model, which is of key significance in the present invention, is further introduced to accurately identify whether there is a trend reversal behavior in the local area of ​​postoperative physiological indicators, and to infer whether there is a behavioral trajectory deviation event based on this, as an early signal of infection transition. From the first to the second morning after Mr. Zhang's operation, the drainage system continuously collected three key parameters: pH, flow rate, and turbidity. The time window length was 24 hours, the sampling frequency was every 10 minutes, and a total of 144 time points. Each time point formed a set of three-dimensional behavior vectors. To achieve unified modeling, the system sequentially spliced ​​the pH, flow rate, and turbidity at each time point into a vector x(t) = [x1(t), x2(t), x3(t)], and finally constructed a 3×144 behavior matrix, and structurally aligned it with the matching standard template trajectory x0(t) to form a parallel sequence for trend comparison.

[0084] During the processing, the system first calculates the trend direction σ of each indicator on the sliding window (3 points, i.e. 30 minutes) i (t), if the current point is greater than the center point of the previous window, it is judged to be rising (+1), if it is less than, it is falling (-1), and if it is basically unchanged (the absolute value of the change is less than the physiological noise threshold ε, set to 0.02), it is 0. Taking Mr. Zhang from the 20th hour to the 21st hour after surgery as an example (i.e., from 5:00 to 6:00 the next morning), the pH value dropped from 7.18 to 6.94, σ1(t) = -1; the flow rate dropped from 26 mL / h to 18 mL / h, σ2(t) = -1; and the turbidity increased from 49 to 61, σ3(t) = +1. Change rate intensity η i(t) are: pH decreases by 0.24 (η1=0.24), flow rate decreases to 8 mL / h (η2=0.32 after normalization), and turbidity increases by 12 units (η3=0.27).

[0085] The system compares the recovery template 3 matched by Mr. Zhang, where the standard trend at the 20th hour is: pH enters a stable stage (ξ1(t) = 0), the flow rate slowly decreases but the change is very small (ξ2(t) = 0), and the turbidity continues to decrease (ξ3(t) = -1). Substituting the trend direction outlier function of the present invention into it:

[0086]

[0087] Perform calculations. Now n = 3, substitute:

[0088] pH term: σ1=-1, η1=0.24, ξ1=0→-1×0.24×(1-0)=-0.24

[0089] Velocity term: σ2 = -1, η2 = 0.32, ξ2 = 0 → -1 × 0.32 × (1 - 0) = -0.32

[0090] Turbidity term: σ3 = +1, η3 = 0.27, ξ3 = -1 → +1 × 0.27 × (1 - (-1)) = +0.54

[0091] The calculation shows that:

[0092] Δ(t)=-0.24-0.32+0.54=-0.02

[0093] Although Mr. Zhang's flow rate and pH both dropped significantly during this period, and the turbidity changed in the opposite direction (the system template expected it to drop but it actually rose), the positive and negative offsets canceled each other out, and the overall trend outlier score was only -0.02, which was within the neutral offset range set by the system (the low risk threshold is |Δ(t)|<0.15), and the system did not trigger a risk warning. However, in the next hour (hour 21-22), the pH continued to drop to 6.85, σ1 = -1, η1 = 0.09; the flow rate further dropped to 12mL / h, σ2 = -1, η2 = 0.4; the turbidity rose rapidly to 75, σ3 = +1, η3 = 0.36, while the three trends of the standard template during this period were still ξ1 = 0, ξ2 = 0, ξ3 = -1. Substituting into the formula, we can get:

[0094] pH term: -1×0.09×(1-0)=-0.09

[0095] Flow rate term: -1×0.4×(1-0)=-0.4

[0096] Turbidity term: 1×0.36×(1-(-1))=+0.72

[0097] sum:

[0098] Δ(t)=-0.09-0.4+0.72=+0.23

[0099] Because the absolute value exceeds the system's warning threshold (the system defines a high-risk trend reversal when |Δ(t)| > 0.20), the system records this time period as a directional outlier peak, forming a continuous reverse chain with the previous trend, marked as a behavioral drift stability zone. If the outlier score Δ(t) exceeds the high-risk threshold for three consecutive periods (>30 minutes), the system immediately pushes a warning flag, triggering behavioral deviation map visualization and doctor-side notification.

[0100] The core advantage of this method is that its function structure does not need to judge whether a specific value is out of bounds, but rather detects whether physiological behavior is accelerating in the wrong direction. The coefficient range is specifically set as follows: i ∈{-1,0,+1},η i After normalization, the distribution range is [0.0-1.0], where scores below 0.05 are not counted as outlier scores and are considered physiological fluctuations; i The value range is the same as σ i The system alarm threshold is set to |Δ(t)|>0.20 by default and can be dynamically adjusted between 0.15 and 0.30 according to the patient's condition to balance sensitivity and specificity.

[0101] During the training phase of this function model, cross-validation was performed using annotated postoperative data (n=203 positive infection cases, n=742 non-infection cases). Through precision optimization and threshold parameter adjustment, the time offset between the outlier function and the system's final infection diagnosis was kept within an average of 10 hours. Mr. Zhang was ultimately confirmed to have a Staphylococcus aureus infection by culture on the morning of the third postoperative day, but the system successfully alerted him in the early morning of the second day. The continuous trend reversal segments identified by this function were pushed to the attending physician's terminal, preventing the spread of local infection and controlling symptoms.

[0102] During the postoperative monitoring process, the system constructed Mr. Zhang's behavior vector in real time, and spliced ​​the various indicators of his drainage fluid (such as pH, flow rate, turbidity, viscosity, etc.) into a multi-dimensional time series vector according to the timestamp in each sampling cycle. It also aligned them point by point using a standard recovery template that matched his surgical procedure, anatomical site, and physiological background. Based on this, the Euclidean deviation was calculated to quantify the absolute deviation of Mr. Zhang's overall indicator trajectory in the current time window.

[0103] Between the 26th and 28th hours after Mr. Zhang's surgery, the system measured a cumulative Euclidean deviation of his behavior vector from the standard template of 2.36. This value is usually no more than 1.4 for patients with this procedure in the standard recovery state. The system records that this value exceeds the warning threshold (the threshold interval is set to 2.0±0.1). However, the present invention does not immediately issue an alarm, but instead synchronously enters the directional outlier determination logic. The trend direction outlier function module performs trend symbol analysis and standard direction comparison on the three key indicators of current pH, flow rate, and turbidity. The phenomenon of trend direction reversal is identified for three consecutive time periods. For example, the pH should tend to be stable in the standard trajectory but is currently accelerating downward, the flow rate should be stable but is currently rapidly declining, and the turbidity rebounds from the downward trend. The directional outlier score of each parameter continuously exceeds 0.25 (the system-set directional abnormality score threshold is 0.20) and lasts for more than 30 minutes. At this point, the system activates the joint discrimination rule: if the current Euclidean deviation is greater than 2.0 and the directional outlier scores of at least two key parameters exceed the high-risk threshold within a continuous time window, the system will regard the behavior segment as a critical state of infection risk transition.

[0104] In Mr. Zhang's case, the system first identified through behavioral deviation vectors that his overall physiological state had significantly deviated from the standard recovery trajectory. It then used directional outlier analysis to confirm that these differences were not deviations that were still moving in a positive direction, but deviations in the wrong direction. This double confirmation logic greatly reduced the risk of false positives and avoided false positives due to postoperative fluctuations or parameter noise. The system output a risk score of 87 / 100 based on the joint scoring model, and triggered an early warning mechanism after the score exceeded the joint threshold (set to 80). The patient was identified as a potential infection transition individual and a push notification was generated on the doctor's terminal: including a behavioral trajectory diagram for the past 6 hours, a trend turning highlight prompt, a dynamic map of Euclidean deviations, an indicator risk radar chart, and other visual data, along with personalized intervention recommendations, such as immediate bacterial culture of the drainage fluid, local B-ultrasound examination of periprosthetic fluid accumulation, and evaluation of whether to adjust antimicrobial medications.

[0105] The underlying model of the entire joint judgment mechanism is a hierarchical decision network structure. The first layer is a trend recognition network (directional outlier function), the second layer is a behavioral deviation distance scoring network (Euclidean distance discrimination), and the third layer is a joint logistic regression unit or weighted average fusion layer. The model training process uses a labeled postoperative monitoring dataset for supervised learning. Each sample data structure includes four categories: Euclidean deviation, outlier score, infection onset time, intervention time, and whether the final outcome is infection. The model goal is to maximize prediction accuracy and early recognition window. Training uses ten-fold cross-validation and an early stopping mechanism to control overfitting. Ultimately, the model maintains a specificity of over 95% in uninfected samples and achieves an average early recognition time of 36 hours in samples with infection outcomes.

[0106] This mechanism played a key role in Mr. Zhang's treatment. Since the system captured the dual signals of trajectory deviation and trend reversal 32 hours before the infection occurred and pushed an early warning in time, the doctor started empirical antibiotic treatment as soon as the local inflammatory state was confirmed, successfully controlling the spread of infection and avoiding prosthesis replacement and long-term recovery stagnation.

[0107] Example 3:

[0108] Combined with attachment Figure 4 Patient information: male, 34 years old, with left tibia and fibula fracture, who had undergone plate fixation for 18 months. Due to repeated leakage from the soft tissue fistula, the implant was to be removed. Figure 4 a) After the plate was removed, the peroneus longus tendon was exposed (black thick arrow), with surrounding scars and hyperplasia of old sinus tissue. Figure 4 b): Complete synovectomy + debridement and irrigation of necrotic soft tissue to form a spindle-shaped defect of approximately 10 cm × 2 cm.

[0109] Using 1-0 nylon interrupted-release technique, add a new suture segment of 1-2 finger length every 48 hours at the dressing table in the ward ( Figure 4 c, 4d). After each suture, a Jackson-Pratt (JP) negative pressure drainage tube was placed subcutaneously ( Figure 4 e) and insert the miniature multi-parameter in-situ sensing kit of the present invention into the proximal end of the JP tube. The kit can continuously record the following information in minutes:

[0110] 1. Flow rate (micro-thermal film sensor);

[0111] 2. pH (solid-state glass electrode);

[0112] 3. Turbidity / color (dual-channel optical probe);

[0113] 4. Viscosity (micro-oscillating cantilever beam).

[0114] Monitoring begins immediately after surgery. All data is normalized and denoised in real time by a bedside edge computing unit, and then integrated into the patient's individual behavior vector. The system automatically matches a standard restoration template 4 (the cluster center of 147 non-infected patients of the same age group and surgical procedure) for the lower limb implant removal and staged closure scenario.

[0115] The Euclidean deviation D(t) and the trend direction outlier score S(t) are cyclically calculated within a 6-h sliding window for early warning.

[0116] Key monitoring nodes and model outputs:

[0117]

[0118]

[0119] The triggering decision logic and thresholds are derived from the patented Euclidean offset + directional outlier joint discrimination framework, which can output an early warning an average of 24-48 hours in advance of the clinical diagnosis of infection.

[0120] A potential infection trend was detected within 24 hours of the second suture, approximately 20 hours earlier than the bacterial culture results. The alarm simultaneously sent a trend heat map and risk radar chart to the mobile phone, allowing the doctor to adjust the antibiotics immediately, avoiding the need for suture removal and repeated debridement. Even with repeated shortening and disassembly of the drainage pathway, the sensor module maintained a high-frequency continuous data stream, demonstrating the portability and robustness of the technical solution. The wound healed immediately, with no reinfection of the implant; the patient was able to walk with weight after 6 weeks.

Claims

1. A method for identifying the risk of drainage fluid infection after orthopedic surgery, characterized by The following steps are involved: The drainage fluid is considered as a system output driven by postoperative local immune regulation and trauma recovery; nonlinear time series curves including viscosity, color turbidity, flow rate, and pH indicators are captured; first-order derivative trend modeling is performed to analyze the rate of change, including accelerated pH drop or sudden changes after flow rate stabilization; Establish a framework for postoperative recovery pathways, using the dynamic parameters of drainage fluid from non-infected cases to form a standard recovery trend library; compare the Euclidean deviation and directional changes between the current patient's drainage parameter behavior vector and the standard trajectory; The changes are converted into a visual recovery offset map, and the infection prediction data is measured from a multi-day trend perspective; then a framework is constructed to identify the transition from local abnormal signal aggregation to infection event formation, and the clinical infection node where quantitative change leads to qualitative change is obtained.

2. The method for identifying the risk of infection of drainage fluid after orthopedic surgery according to claim 1, characterized in that The nonlinear timing curve construction method includes: A multi-parameter in-situ sensing module is integrated into the postoperative drainage channel to continuously collect the physical and chemical parameters of the drainage fluid. The in-situ sensing module collects these parameters at a minute-level granularity and timestamps each data point before inputting it into an edge computing unit. The edge computing unit then establishes a nonlinear time series behavior curve for each parameter and extracts its rate of change, acceleration, and abnormal inflection points. The behavior curves of each parameter are compared with the standard postoperative recovery model to identify the degree of deviation and trend direction; when multiple parameters are identified to show a synchronous deviation trend within the same time window, it is judged as an early signal of infection risk transition and the infection risk score result is output.

3. The method for identifying the risk of drainage fluid infection after orthopedic surgery according to claim 2, characterized in that The viscosity is measured by a micro-oscillating cantilever beam sensing structure, and the vibration attenuation characteristics of the cantilever beam are used to characterize the viscous behavior of the drainage fluid; the color turbidity is obtained by an optical detection module arranged in the drainage channel. The optical detection module includes a channel for measuring directly absorbed light and a channel for detecting 90-degree scattered light to simultaneously evaluate color and turbidity.

4. The method for identifying infection risk of drainage fluid after orthopedic surgery according to claim 3, characterized in that The flow rate and pH are acquired synchronously, and their time phase difference is analyzed through behavioral curve analysis to determine whether local environmental acidification is accompanied by a sudden change in flow rate; the standard recovery model is constructed based on the time series behavioral data of various drainage fluid parameters in normal patients under non-infectious conditions after surgery, and is used as a control trajectory for subsequent individual identification.

5. The method for identifying infection risk of drainage fluid after orthopedic surgery according to claim 4, characterized in that The edge computing unit includes an adaptive behavior feature extraction module for dynamically updating the baseline behavior curve of individual patients; the infection risk score result is output to the clinical terminal according to the risk level, and a clinical intervention prompt is triggered when the score value reaches a preset warning threshold.

6. The method for identifying infection risk of drainage fluid after orthopedic surgery according to claim 1, characterized in that The construction of the comparison method of the Euclidean deviation and the directional change includes: S1. Construct a postoperative recovery path model, including collecting multi-parameter time series data of drainage fluid from non-infected patients and classifying and clustering them according to surgical type, location, and individual characteristics to form a standard recovery trajectory library; S2. Acquire multiple continuous parameters of the postoperative drainage fluid of the current target patient within a set time window to form a multidimensional behavior vector; S3, aligning the current patient behavior vector with the corresponding template in the standard recovery trajectory, and performing point-by-point comparison based on temporal consistency; S4. Calculate the Euclidean deviation between the current behavior vector and the standard trajectory to quantify the overall degree of deviation; simultaneously calculate the trend direction change of each parameter, identify whether there is a trend in the opposite direction of the standard trajectory, and mark it as a directional outlier; S5. When the Euclidean deviation exceeds the preset deviation threshold and multiple parameters show directional outliers at the same time, it is determined that the patient has a transition trend in the risk of postoperative infection and an early warning signal is output.

7. The method for identifying infection risk of drainage fluid after orthopedic surgery according to claim 6, characterized in that The standard recovery trajectory library generates multiple representative recovery curve templates through cluster analysis to adapt to different surgical procedures, surgical areas and patient physiological background differences; the drainage fluid parameter collection time window is continuous 4 hours to 24 hours, and is resampled at a uniform time interval for time alignment with the standard trajectory.

8. The method for identifying infection risk of drainage fluid after orthopedic surgery according to claim 7, characterized in that The behavior vector is a unified vector structure formed by splicing multiple indicators according to timestamps; the directional outlier determination includes comparing whether the rising or falling trend of the current indicator is opposite to the trend direction of the corresponding time period of the standard trajectory, and recording the outlier event when the trend is reversed.

9. The method for identifying infection risk of drainage fluid after orthopedic surgery according to claim 8, characterized in that The Euclidean deviation and the directional outlier result are used together for risk identification. Only when both meet the warning conditions at the same time, it is determined that there is a tendency for early infection transition.

Images