An artificial intelligence peritonitis monitoring device, system and method

Abstract

The application relates to the technical field of medical monitoring and discloses an artificial intelligence peritonitis monitoring device, system and method, which comprises a closed darkroom, a high-transparency pipeline, a simulated natural light source, a color sensor, a parallel light source, a light sensor, an opaque partition and an intelligent chip. The closed darkroom is divided into an upper color monitoring area and a lower turbidity monitoring area through the opaque partition. The high-transparency pipeline vertically penetrates through the opaque partition and is connected with a peritoneal dialysis pipeline drainage end and a waste liquid barrel. The intelligent chip is connected with the color sensor and the light sensor and is internally provided with an artificial intelligence algorithm. The intelligent chip carries out noise suppression through the construction of a four-dimensional Kalman filter state vector containing RGB three primary color components and turbidity, carries out peritonitis risk determination based on fusion data of dialysate optical characteristics and patient information, and has a peritonitis identification sensitivity of greater than or equal to 95%, a specificity of greater than or equal to 99% and an accuracy of greater than or equal to 98%. The application solves the problem of single detection dimension and provides precise support for early warning.

Description

An artificial intelligence peritonitis monitoring device, system and method Technical Field

[0001] This invention relates to the field of medical monitoring technology, specifically to an artificial intelligence peritonitis monitoring device, system, and method. Background Technology

[0002] Peritoneal dialysis (PD) is the core renal replacement therapy for patients with end-stage renal disease (ESRD). Due to its advantages such as ease of operation, home-based care, minimal impact on the cardiovascular system, and low consumption of medical resources, it has been widely adopted globally, especially suitable for elderly, children, and patients in remote areas. Peritonitis, the most common and serious complication of peritoneal dialysis, occurs at an annual rate of 0.2-0.6 times per person per year. It not only leads to peritoneal dysfunction and dialysis failure but can also cause systemic infection and sepsis, making it a significant cause of hospitalization and death, and a major reason for patients to be transferred to hemodialysis. Therefore, early detection and timely intervention of peritonitis are crucial for ensuring patient prognosis.

[0003] Currently, the clinical diagnosis of peritonitis mainly relies on the "symptom observation + laboratory testing" model: a preliminary judgment is made based on the patient's complaints of abdominal pain, fever, and other symptoms, followed by confirmation through bacterial culture and etiological testing in a hospital laboratory. In addition, some technologies attempt to assist detection using optical means. For example, prior art document 1 (US17201569) discloses a dialysate testing system that relies on a sensor card to detect specific biomarkers, requiring the use of a peritoneal dialysis circulation device; prior art document 2 (US20170136166A1) proposes detecting dialysate turbidity and color using a portable detection device, outputting only a single predicted value; and prior art document 3 (CN109475678B) discloses a detection method that combines photos taken with symptom queries using a smartphone.

[0004] The existing technologies have significant limitations: First, the detection scenarios are limited. The system in Comparative Document 1 needs to be integrated into the circulator, and Comparative Document 3 relies on smartphones and storage devices, neither of which achieves seamless integration with dialysis tubing or truly automated home monitoring. Second, the detection dimensions are limited. Comparative Document 1 focuses on biomarkers, Comparative Document 2 only integrates turbidity and color, and Comparative Document 3 relies on visual images and subjective symptoms, none of which incorporate individual patient information and objective vital signs. Third, the data processing is simplistic. Comparative Document 2 uses dual-algorithm switching, and Comparative Document 3 only calculates the average brightness of the image without employing intelligent multi-feature fusion analysis, resulting in insufficient accuracy. Fourth, the linkage is insufficient. The existing technologies have not constructed a complete "monitoring-early warning-patient response" closed loop, resulting in poor timeliness of intervention.

[0005] Therefore, there is an urgent need to develop a new device, system, and method for monitoring peritonitis. Summary of the Invention

[0006] The present invention aims to provide an artificial intelligence peritonitis monitoring device, system and method to solve the problem that the existing peritonitis detection cycle is long and cannot achieve early warning.

[0007] To solve the above problems, the present invention adopts the following technical solution:

[0008] Option 1: An artificial intelligence peritonitis monitoring device, comprising a sealed darkroom, highly transparent tubing, simulated natural light source, color sensor, parallel light source, light sensor, opaque partition, and smart chip;

[0009] The enclosed darkroom is divided into an upper color monitoring area and a lower turbidity monitoring area by an opaque partition. The highly transparent pipeline vertically penetrates the opaque partition and connects the peritoneal dialysis tubing drainage end to the waste liquid tank.

[0010] Within the color monitoring area, a simulated natural light source and a color sensor are respectively placed on both sides of a highly transparent pipeline. The light passes through the dialysate in the pipeline and is received by the color sensor to obtain the RGB three primary color components.

[0011] Within the turbidity monitoring area, a parallel light source is placed on one side of the high-transparency pipeline, and a light sensor is placed inside the pipeline with its detection direction perpendicular to the direction of the parallel light, used to receive scattered light to calculate turbidity;

[0012] The smart chip is connected to a color sensor and a light sensor respectively, and has a built-in artificial intelligence algorithm. The smart chip suppresses noise by constructing a 4D Kalman filter state vector containing RGB primary color components and turbidity. It determines the risk of peritonitis based on the fusion data of dialysate optical characteristics and patient information. The peritonitis identification sensitivity is ≥95%, specificity is ≥99%, and accuracy is ≥98%. The smart chip adopts an "event-driven" wireless data transmission mode, uploading complete data only when the risk of peritonitis is detected, and uploading a data summary under normal conditions.

[0013] Beneficial effects: By constructing a stable detection environment through a closed darkroom and opaque partitions, light interference is avoided; by integrating dual-dimensional detection of color and turbidity, combined with artificial intelligence algorithms based on multi-source data fusion, the problem of traditional detection relying on hospitals and having a single dimension is solved, providing accurate data support for early warning.

[0014] The core of this invention's "event-driven" data transmission mode is to upload complete data only when a risk of peritonitis is detected, and upload a data summary under normal conditions. Essentially, it dynamically adjusts the data transmission content based on the monitoring results: prioritizing diagnostic integrity under risk conditions, and prioritizing reducing power consumption and transmission costs under normal conditions, thus adapting to the battery life and practicality requirements of home monitoring.

[0015] In this invention, the "4D Kalman filter state vector" refers to integrating the four key optical parameters of the dialysate—the red (R), green (G), and blue (B) primary color components and turbidity (T)—into a single vector (denoted as X=[R,G,B,T]). T This is used as the input to the filtering algorithm. The state equation of the 4D Kalman filter is: The observation equation is ;in, Let t be the state vector. A is a 4×4 state transition matrix, and H is a 4×4 observation matrix. , These are process noise and observation noise (variance range 0.01-1.0). Four independent optical detection metrics (R, G, B, T) are used, each dynamically changing over time (e.g., fluctuations in optical characteristics during dialysate flow). By constructing state and observation equations, measurement noise (e.g., light source fluctuations, sensor errors, liquid flow interference) is suppressed in real time, improving the smoothness of optical feature data by 60% and providing stable input for subsequent feature extraction and model determination.

[0016] Furthermore, the color temperature range of the simulated natural light source is 5500K-6500K, the sampling frequency of the color sensor is 1Hz, and the output accuracy is 8-bit RGB component data; the parallel light source is an infrared LED with a wavelength of 860nm, and the distance between the light sensor and the parallel light source is 2cm-3cm.

[0017] Beneficial effects: By limiting parameters such as light source color temperature, sampling frequency, and sensor spacing, the accuracy of RGB data and turbidity detection can be ensured, and early trends of dialysate yellowing and turbidity can be accurately captured, thereby improving detection sensitivity.

[0018] Furthermore, the intelligent chip adopts the nRF52832 chip, including a 16-bit precision data sampling module, an MCU microprocessor, an embedded artificial intelligence algorithm device, and a wireless communication module. The embedded artificial intelligence algorithm device is a random forest model or a multilayer feedforward neural network.

[0019] Beneficial effects: By adopting the highly integrated nRF52832 chip and high-precision sampling module, a balance is achieved between low power consumption and high-fidelity data processing, which is suitable for the battery life requirements of home devices, and the embedded intelligent algorithm ensures real-time and accurate analysis.

[0020] Furthermore, the high-transparency pipeline is made of quartz glass, with an inner diameter of 8mm-10mm, and is a single-use structure with standard Luer connectors at both ends.

[0021] Beneficial effects: The disposable quartz glass tubing design avoids cross-infection, and the standard Luer connectors ensure compatibility with existing dialysis equipment, lowering the threshold for clinical promotion.

[0022] Option 2: An artificial intelligence peritonitis monitoring system, comprising any of the aforementioned artificial intelligence peritonitis monitoring devices, a cloud server, a medical care terminal, and a patient terminal;

[0023] The monitoring device interacts with a cloud server via the wireless communication module of the smart chip. The cloud server communicates bidirectionally with both the medical staff and the patient. The patient terminal includes a patient information collection module and a built-in early warning module. The patient information collection module collects 15 items of patient information, and the built-in early warning module receives early warning signals from the smart chip and alerts the patient through multiple methods. The smart chip of the monitoring device determines the risk of peritonitis based on the fusion data of the optical characteristics of the dialysis fluid and the patient information.

[0024] Beneficial effects: By constructing a "device-cloud-patient" linkage architecture, objective testing data and individual patient information are integrated to form a complete monitoring closed loop, solving the problems of insufficient linkage and delayed intervention in existing technologies.

[0025] Furthermore, the patient information collection module collects five basic information items, including gender and age, and ten symptom and sign information items, including body temperature and heart rate. The accuracy of body temperature collection is 0.1℃, and the error of heart rate collection does not exceed 2 beats / minute.

[0026] Beneficial effects: It clarifies the accuracy of patient information collection, ensures the reliability of individual data, provides precise input for individualized analysis of artificial intelligence algorithms, and improves the pertinence of risk assessment.

[0027] Furthermore, the built-in early warning module includes an LED light, a buzzer, and an LCD screen. When the monitoring device determines that there is a risk of peritonitis, the LED light flashes at a frequency of 1Hz, the buzzer emits a warning sound at 2-second intervals, and the LCD screen displays the early warning information simultaneously.

[0028] Beneficial effects: The multi-mode early warning design avoids the problem of single prompts being ignored. The combination of sound and light with text guidance helps patients quickly identify risks and take the right measures, thus improving the efficiency of early warning response.

[0029] Furthermore, the cloud server adopts a distributed storage architecture, supporting concurrent data transmission from no less than 1,000 monitoring devices with a data transmission latency of no more than 500ms.

[0030] Beneficial effects: The distributed storage and low-latency transmission design ensures that the system has the ability to be applied on a large scale, meets the needs of medical institutions for centralized management of batches of patients, and improves the efficiency of medical resource utilization.

[0031] Option 3: An artificial intelligence-based peritonitis monitoring method, applied to the system described in Option 2, includes the following steps:

[0032] S1: The color sensor and light sensor of the monitoring device collect RGB data of the dialysate and turbidity-related scattered light data respectively, with a sampling frequency of 1Hz;

[0033] S2: Perform Kalman filtering on the collected data to construct a 4-dimensional state vector containing RGB primary color components and turbidity to achieve noise suppression;

[0034] S3: Extract 24 optical features within a 60-second time window and fuse them with 15 patient information items to form a 39-dimensional feature vector;

[0035] S4: The feature vector is analyzed by artificial intelligence algorithm, and the peritonitis risk assessment result is output based on the fusion data of the optical characteristics of dialysis fluid and patient information;

[0036] S5: If a high-risk condition is identified, an alert will be triggered and the results will be transmitted to the cloud, medical staff, and patients.

[0037] Beneficial effects: Through the complete process of "collection-filtering-feature fusion-intelligent analysis-early warning", dynamic monitoring and trend analysis are realized. The 39-dimensional feature vector solves the problem of low accuracy of single indicator analysis and provides a scientific basis for early warning.

[0038] In step S3, the 60-second time window uses a sliding window method to extract features in real time. After feature extraction, Min-Max normalization is performed, and the normalization formula is as follows: .

[0039] Furthermore, the artificial intelligence algorithm in step S4 is a random forest model, which consists of 200 decision trees. When splitting a node, 6 features are randomly selected, and the Gini index is used as the splitting criterion. The model has a sensitivity of ≥95%, a specificity of ≥99%, and an accuracy of ≥98% for the identification of peritonitis.

[0040] Beneficial effects: Optimizing the parameters of the random forest model ensures the model's generalization ability and recognition accuracy. The accuracy rate of over 98% significantly reduces the risk of missed diagnoses and misdiagnoses, improving diagnostic reliability compared to traditional methods.

[0041] The advantages of this invention are:

[0042] 1. Outstanding early warning capability: By monitoring early changes in the optical properties of dialysis fluid and combining intelligent algorithms, signs of peritonitis can be detected more than 3 days before the appearance of clinical symptoms, providing an early warning at least 7 days earlier than traditional bacterial culture.

[0043] 2. Continuous home monitoring: The device is small and highly integrated, and can be directly connected to the home peritoneal dialysis tubing without any additional operation, enabling 24-hour continuous monitoring, which is especially suitable for patients with limited mobility.

[0044] 3. Multi-dimensional and precise analysis: Integrating 24 optical features and 15 patient information items, a 39-dimensional feature vector is constructed, and the intelligent algorithm achieves an accuracy rate of over 98%.

[0045] 4. Non-invasive, safe and convenient: No blood collection or contact with the patient's body is required. Disposable tubing avoids cross-contamination. The operation is simple and patients can complete the monitoring at home.

[0046] 5. Highly efficient doctor-patient collaboration: Early warning information is synchronized to both the patient and medical staff in real time, allowing medical staff to provide remote guidance, reducing unnecessary outpatient visits, and achieving precise diagnosis and treatment.

[0047] 6. Strong compatibility and scalability: It can seamlessly connect with existing mainstream peritoneal dialysis equipment, support algorithm upgrades and the addition of new monitoring functions, and can be connected to hospital information systems.

[0048] 7. Significantly reduced medical costs: Early intervention reduces the peritonitis-related hospitalization rate by more than 45%, saving each high-risk patient 10,000 to 20,000 yuan in medical expenses annually.

[0049] Furthermore, this invention utilizes changes in the turbidity and color of peritoneal dialysis fluid as the earliest and most sensitive objective indicators of peritonitis, and patient-specific information (such as age and dialysis duration) and vital signs (such as body temperature and heart rate) significantly enhance risk assessment. By integrating optical sensing, multi-dimensional data fusion, and intelligent algorithms, this invention addresses the core problems of existing technologies—limited application scenarios, single-dimensionality analysis, low accuracy, and delayed response—filling the technological gap in non-invasive continuous home monitoring and possessing irreplaceable clinical value in improving the safety of peritoneal dialysis. Attached Figure Description

[0050] Figure 1 is a schematic diagram of the peritonitis monitoring device of the present invention.

[0051] Figure 2 is a schematic diagram of the connection structure of the peritonitis monitoring system in this invention.

[0052] Figure 3 is a schematic diagram of the hardware structure of the smart chip in this invention.

[0053] Figure 4 is a schematic diagram of the connection structure of the smart chip in an embodiment of the present invention.

[0054] Figure 5 is a flowchart of the algorithm according to an embodiment of the present invention.

[0055] The reference numerals in the accompanying drawings include: 1. Enclosed darkroom; 2. Highly transparent conduit; 3. Simulated natural light source; 4. Color sensor; 5. Opaque partition; 6. Light sensor; 7. Parallel light source; 8. Smart chip. Detailed Implementation

[0056] The following detailed description illustrates the specific implementation method:

[0057] As shown in Figure 1, a schematic diagram of the peritonitis monitoring device, a highly transparent tube 2 runs from top to bottom through the center of the sealed dark chamber 1. On the top surface of the sealed dark chamber 1, the drainage end of the peritoneal dialysis tubing is connected to the top opening of the highly transparent tube 2. On the bottom surface of the sealed dark chamber 1, a peritoneal dialysis waste container is connected to the bottom opening of the highly transparent tube 2. An opaque partition 5 is installed in the middle of the sealed dark chamber 1, dividing the transparent dark chamber into an upper space and a lower space. The highly transparent tube 2 passes through the opaque partition 5, with its top end connected to the drainage end of the peritoneal dialysis tubing and its bottom end connected to the peritoneal dialysis waste container.

[0058] In the upper space of the enclosed darkroom 1, a simulated natural light source 3 is set on the left side of the high transparency pipe 2 in the enclosed darkroom 1, and a color sensor 4 is set on the right side of the high transparency pipe 2. The light from the simulated natural light source 3 passes through the high transparency pipe 2 and is transmitted to the color sensor 4.

[0059] In the lower space of the enclosed darkroom 1, a parallel light source 7 is installed on the left side of the highly transparent pipe 2. A light sensor is installed inside the highly transparent pipe 2 in the lower space. The light source generated by the parallel light source 7 is transmitted to the right side through the highly transparent pipe 2, and at the same time, the light source of the parallel light source 7 is transmitted to the light sensor.

[0060] As shown in Figure 2, the peritonitis monitoring system includes a peritonitis monitoring device, a cloud server, a medical staff terminal, and a patient terminal. When the system is working, the peritonitis monitoring device transmits signals to the cloud server. The cloud server then communicates with both the medical staff terminal and the patient terminal. The patient terminal includes a patient information collection module and a built-in early warning module.

[0061] As shown in Figure 1, the peritonitis monitoring device consists of a sealed dark room 1, a highly transparent pipeline 2, a simulated natural light source 3, a color sensor 4, a parallel light source 7, a light sensor, an opaque partition 5, and a smart chip 8.

[0062] (1) The sealed darkroom 1 is the outermost layer of the device, used to isolate external light and ensure the stability of the internal detection environment.

[0063] (2) A square high-transparency tube 2 runs through the middle of the dark room. The top end of the high-transparency tube 2 is connected to the drainage end of the peritoneal dialysis tube, and the bottom end is connected to the peritoneal dialysis waste tank. The peritoneal dialysis drainage fluid flows from the top end of the high-transparency tube 2 to the bottom end.

[0064] (3) The upper part of the sealed darkroom 1 is the liquid color monitoring area, and the lower part is the liquid turbidity monitoring area. The two areas are separated by an opaque partition 5 to avoid mutual interference.

[0065] (4) The liquid color monitoring area includes a simulated natural light source 3 and a color sensor 4. The light from the light source passes through a highly transparent pipe 2 and the liquid flowing through it, and reaches the color sensor 4 to obtain the components of the three primary colors: red (R), green (G), and blue (B).

[0066] (5) The liquid turbidity monitoring area includes a monochromatic parallel light source 7 and a light sensor. The detection direction of the light sensor is perpendicular to the direction of the parallel light. The parallel light passes through the highly transparent pipe 2 and the liquid flowing through it, and is scattered by suspended particles in the liquid. The scattered light is detected by the light sensor. Then, the liquid turbidity is calculated based on the intensity of the scattered light.

[0067] The principle of liquid turbidity calculation:

[0068] When a beam of light shines on a dialysate containing suspended particles, the particles scatter the light. By measuring the intensity of the scattered light at a 90-degree angle to the incident light, the number and volume of particles in the solution can be calculated. The formula for the intensity of the scattered light is:

[0069]

[0070] in, It is the intensity of scattered light. It is the intensity of the incident light. It is the number of particles per unit volume. It refers to the volume of a single particle. It is the wavelength of the incident light. It is a constant.

[0071] Turbidity With the intensity of scattered light Positive correlation:

[0072]

[0073] Constants are obtained through instrument calibration. , The turbidity can then be calculated. .

[0074] (6) The intelligent chip 8 device includes a sensor interface, a data sampling module, a memory, a microprocessor, an embedded artificial intelligence algorithm device, a wireless data communication device, and a power management module.

[0075] The sensor interface acquires data from color sensor 4 and light sensor 4.

[0076] The data sampling module is used to perform analog-to-digital conversion sampling on color sensor 4 and light sensor.

[0077] The memory device is used to temporarily store the acquired information and calculation results.

[0078] Microprocessors handle computational and flow control tasks. They are implemented using a System-on-a-Chip (SoC) and include a processor, memory, and peripheral circuitry. A single chip can perform multiple functions such as data acquisition, conversion, storage, processing, and input / output. Microprocessors can be MCUs, Application-Specific Integrated Circuits (ASICs), or Field-Programmable Gate Arrays (FPGAs).

[0079] The embedded artificial intelligence algorithm device is an artificial neural network model, which is embedded in a chip or stored in an on-chip flash memory. It is used to process patient information, fluid color and turbidity data in real time, perform data analysis and calculation, and detect peritonitis events.

[0080] The wireless data communication device is based on an "event-driven" mode to transmit data. Once the smart chip 8 detects a peritonitis event, it will record the event and time, and transmit the peritonitis event results to the cloud server through a wireless transmission module (WIFI, 4G, BLE, etc.). At the same time, it will send a warning signal to the patient through the built-in warning module on the patient's end.

[0081] The power management module includes an external power supply and a lithium battery, and supplies power to the system through relevant connectors;

[0082] The hardware structure of the intelligent chip 8 is shown in Figure 3: The low-power microprocessor is connected to the data sampling module and the general peripheral interface GPIOUARTSPI via a bus. The data sampling module is connected to the sensor interface, which is connected to the color sensor 4 and the light sensor respectively. The general peripheral interface GPIOUARTSPI is connected to the patient-side LED, buzzer and LCD display respectively.

[0083] The patient-side module includes a patient information collection module and a built-in early warning module.

[0084] The collected patient information includes basic information, symptoms, and signs. Basic patient information includes gender, age, height, weight, and duration of peritoneal dialysis treatment. Symptoms and signs include body temperature, heart rate, blood pressure, abdominal pain, abdominal distension, nausea, vomiting, constipation, and loss of appetite. The collected patient information is input into eight modules of the smart chip.

[0085] The built-in warning module includes LEDs, a buzzer, and an LCD screen. It receives warning information from the smart chip 8 and can send peritonitis warning signals to patients through light, sound, and image display.

[0086] The cloud server receives and stores the warning information output by the smart chip 8, and then transmits the warning information to the medical staff.

[0087] The medical staff can visualize the early warning information, which can be viewed and confirmed by medical staff to take further action, such as adjusting the treatment plan or notifying the patient to go to the hospital for treatment.

[0088] The advantages of this invention are:

[0089] 1. Early warning significantly shortens the time to detection of peritonitis.

[0090] This invention enables early warning of peritonitis by real-time monitoring of changes in the optical properties (color and turbidity) of peritoneal dialysis fluid and combining this with artificial intelligence algorithms. Compared to traditional methods that rely on symptoms and bacterial culture, this approach can detect peritonitis before clinical symptoms appear, significantly shortening diagnosis and intervention time and reducing the risk of complications and death.

[0091] 2. Dynamic and continuous monitoring enhances safety.

[0092] This system can dynamically and continuously monitor the risk of peritonitis in peritoneal dialysis patients, avoiding delays in treatment due to missed diagnoses, misdiagnoses, or patient negligence, and greatly improving the safety and compliance of peritoneal dialysis.

[0093] 3. Non-invasive and convenient, suitable for home and community use.

[0094] This device can be integrated into dialysis tubing or configured independently, requiring no additional blood collection or complex operations. Patients can achieve automatic monitoring and early warning at home, reducing their burden and improving their quality of life. It is suitable for widespread adoption in families and communities.

[0095] 4. Intelligent analysis and personalized risk assessment

[0096] The system integrates basic patient information, symptoms, signs, and turbidity data, and uses artificial intelligence algorithms for multi-dimensional analysis to achieve individualized peritonitis risk assessment, resulting in more accurate early warnings and reduced false alarms and missed alarms.

[0097] 5. Remote collaboration to optimize patient management

[0098] Early warning information can be uploaded to the cloud server in real time, allowing medical staff to remotely view the patient's status, provide timely guidance for adjusting treatment plans, realize telemedicine and intelligent management, and improve the efficiency of medical resource utilization.

[0099] 6. Highly scalable and easy to integrate and upgrade

[0100] The system adopts a modular design, which facilitates integration with existing peritoneal dialysis equipment. It can also be expanded to include more physiological parameter monitoring and intelligent analysis functions according to clinical needs, and has good scalability and industrialization prospects.

[0101] 7. Reduce medical costs and decrease hospitalization rates

[0102] Early detection and intervention of peritonitis can effectively reduce hospitalizations, referrals, and changes in dialysis methods caused by peritonitis, thereby reducing medical costs and alleviating the burden on society and families.

[0103] The specific implementation process is as follows:

[0104] Example 1

[0105] This embodiment is a hardware implementation of a peritonitis monitoring device.

[0106] 1.1 Component Selection and Assembly

[0107] Enclosed Darkroom 1: Made of ABS engineering plastic, measuring 15cm×10cm×20cm, with black flocked cloth pasted on the inner wall, and pre-installed pipe interfaces at the top and bottom.

[0108] High-transparency pipe 2: made of quartz glass, with an inner diameter of 8mm, an outer diameter of 10mm, a length of 18cm, and standard Luer connectors at both ends.

[0109] Light source and sensor: The simulated natural light source 3 uses an SMD5050 LED module (color temperature 6000K); the color sensor 4 uses a high-sensitivity RGB sensor TCS34725 chip, which is set at a position 180° from the light source; the parallel light source 7 is an 860nm infrared LED (IR333-A); the light sensor uses an S1133 photodiode, which is set at a 90° angle from the light source.

[0110] Opaque partition 5: Made of acrylic sheet, 2mm thick, with a round hole in the center matching the high transparency pipe 2, and a sealing strip on the edge.

[0111] Smart Chip 8: As shown in Figure 4, the core chip is the nRF52832, integrating an ARM Cortex-M4 core. The data sampling module uses the ADS1115 chip (16-bit ADC), and the wireless communication module supports BLE5.0 and WiFi dual-mode. Smart Chip 8 is connected to the color sensor 4 and the light sensor through the first and second pins of connector J7, to the buzzer through the first and second pins of connector J10, to the LED light through the third and fourth pins of J10, and to the LCD screen through the fifth and sixth pins of J10.

[0112] 1.2 Assembly Process

[0113] An opaque partition 5 is fixed in the middle of the enclosed darkroom 1 (8cm from the top), and a highly transparent pipe 2 passes through the round hole in the partition and is fixed. In the upper color monitoring area, a simulated natural light source 3 and a color sensor 4 are fixed on both sides of the pipe (5cm apart). In the lower turbidity monitoring area, a parallel light source 7 is fixed on the left side of the pipe (3cm from the partition), and a light sensor is fixed inside the pipe by a bracket, forming a 90° angle with the light from the parallel light source 7. A smart chip 8 is fixed at the bottom of the darkroom and connected to each sensor and the power module via an FPC cable.

[0114] 1.3 Technical Effects

[0115] The device can effectively isolate external light interference, the color sensor 4 outputs RGB component error of less than 2%, turbidity detection range of 0-500 NTU, accuracy of 1 NTU; it can work continuously for 8 hours with 3.7V lithium battery power supply, meeting the monitoring needs of a single peritoneal dialysis session; the modular design makes the failure rate less than 0.5%.

[0116] Example 2

[0117] This embodiment demonstrates the implementation of the random forest model algorithm for smart chips.

[0118] 2.1 Data Preprocessing Module

[0119] As shown in the algorithm flowchart in Figure 5: Data preprocessing involves sequentially acquiring sensor data, Kalman filtering, and feature extraction. Then, patient basic information is collected and fused together for feature fusion and data normalization. Simultaneously, a prediction model is built using a training dataset, the training results obtained from offline model training, and the data normalization results. Finally, the results are output. Specifically, the sensor acquires data at a frequency of 1Hz, which is then transmitted to the nRF52832 after analog-to-digital conversion via the ADS1115 chip. In the Kalman filtering module, both the state transition matrix A and the observation matrix H are 4×4 identity matrices, and qR, qG, qB, and qT in the process noise covariance matrix Q are all set to 0.1. Using a 60-second time window, six features (max, min, mean, etc.) are extracted from each indicator and concatenated with 15 patient-coded data points to form a 39-dimensional feature vector, which is then normalized to the [0,1] interval using Min-Max.

[0120] Sensor data acquisition

[0121] The RGB color sensor samples colors at a frequency of 1Hz, and the red, green, and blue color components are denoted as follows: t is the sampling time.

[0122] The light sensor samples the light intensity value at a frequency of 1 Hz, converts it into a turbidity value using an internal calibration formula, and denoted as . t is the sampling time.

[0123] Filtering of sensor data

[0124] Because sensor data contains noise, in order to suppress noise, smooth data, estimate the state in real time, and enhance robustness, this invention performs joint Kalman filtering on the collected color and turbidity data.

[0125] State-space modeling:

[0126] Define the system's state vector as follows:

[0127]

[0128] in: The red component at time t, Let the green component be at time t. The blue component at time t, Let t be the turbidity at time t.

[0129] The observation vector is:

[0130]

[0131] in, This represents the actual observation value of the sensor at time t.

[0132] State transition equation:

[0133] Assume that the system state changes over time and can be described by a linear model:

[0134]

[0135] in, The state transition matrix is ​​a 4×4 identity matrix. To observe the noise, , Here is the process noise covariance matrix:

[0136]

[0137] in, The process noise variances for red, green, blue, and turbidity are initially set to small values ​​(e.g., 0.01-1.0), which can be adjusted based on experimental data.

[0138] Observation equation:

[0139]

[0140] in, The observation matrix is ​​a 4×4 identity matrix; To observe the noise, R is the observation noise covariance matrix:

[0141]

[0142] in, The observation noise variances for red, green, blue, and turbidity are respectively obtained through experimental observation or sensor manuals.

[0143] State prediction:

[0144]

[0145] Based on the optimal estimate from the previous time step, predict the state at the current time step. Initialize to initial observations or empirical values.

[0146] Covariance prediction:

[0147]

[0148] Predict the uncertainty of the current state estimate. The P-state estimate covariance matrix reflects the reliability of the current estimate. Initialization can be performed by multiplying an identity matrix by a positive constant.

[0149] Update steps:

[0150] Kalman gain calculation:

[0151]

[0152] We measure the reliability of predictions and observations to determine how to weight and merge them.

[0153] Status Update:

[0154]

[0155] The predicted values ​​are corrected using the observed values ​​to obtain the optimal estimate for the current moment.

[0156] Covariance update:

[0157]

[0158] Update the uncertainty of the current state estimate.

[0159] Feature extraction

[0160] The time window was set to 60 seconds, and sampling was performed at 1 Hz. Each window contained 60 sets of data, all of which underwent the aforementioned Kalman filtering process. The data for the three color components and the turbidity window are as follows:

[0161]

[0162]

[0163]

[0164]

[0165] right Feature extraction was performed separately, and the extracted features included: maximum value (max), minimum value (min), arithmetic mean (mean), standard deviation (STD), median (median), and interquartile range (IQR). This resulted in four feature sequences, each containing six features, for a total of 24 features.

[0166]

[0167]

[0168]

[0169]

[0170] Patient information collection and feature fusion

[0171] The patient's basic information and symptom / sign information collected from the patient's end are coded. The coding rules are shown in Table 1:

[0172] Table 1

[0173]

[0174] The data from 15 patients and 24 features were concatenated to form the final feature vector, which contains 39 features:

[0175] ]

[0176] Data normalization

[0177] The concatenated feature matrix is ​​then normalized using Min-Max based on the features, as shown in the following formula:

[0178]

[0179] 2.2 Deployment of the Random Forest Model

[0180] The model was trained offline using Python's scikit-learn library. The training dataset contained data from 1000 patients (700 healthy individuals and 300 cases of peritonitis) and 500 animal experiments, divided into training and validation sets in a 7:3 ratio. Model parameters: 200 decision trees, with 6 features randomly selected during node splits, a minimum split sample size of 2, and class weights calculated as wc = nsamples / (K). (nc) Automatic adjustment. After training is complete, it is stored in the on-chip flash memory of the nRF52832.

[0181] The random forest classification model consists of 200 decision trees as the base learner. Each decision tree generates a subset of training data by randomly sampling from the training dataset using the bootstrap sampling method. At each node split, 6 features are randomly selected from all features for optimal splitting. The minimum number of split samples per node is set to 2. Each decision tree is trained independently. Furthermore, the weights of each category are automatically adjusted to balance the influence of each category during the model training process. The binary classification warning label (with peritonitis and without peritonitis) is predicted using the majority voting method.

[0182] During training, each decision tree uses the Gini index as the splitting criterion when splitting nodes. The Gini index of node t is defined as follows:

[0183]

[0184] in, Let K be the proportion of samples of class k in node t, where K is the total number of classes (2 in this embodiment). Each split selects the feature and split point that minimizes the weighted average Gini index.

[0185] Weight category adjustment settings are as follows:

[0186]

[0187] in, K is the total number of samples, and K is the number of categories. denoted as the number of samples in class c.

[0188] The model's output category is determined by majority voting:

[0189]

[0190] in, Let C be the predicted category of the output, C be the set of all possible categories (either peritonitis occurred or peritonitis did not occur), and I(.) be the indicator function. If the value is 1, then the value is 0; otherwise, the value is 0.

[0191] During the prediction process, the 39 normalized input features are used to predict the binary classification warning labels of the samples to be detected using the majority voting method consistent with the training module.

[0192] 2.3 Technical Effects

[0193] The model achieves an accuracy of 98.2%, sensitivity of 95.5%, and specificity of 99.1% on the validation set; the single-sample inference time is only 80ms, and the compressed model size is approximately 1.5MB, making it suitable for the storage and computing capabilities of embedded chips.

[0194] Example 3

[0195] This embodiment demonstrates the integrated application of a peritonitis monitoring system.

[0196] 3.1 System Composition and Communication Protocol

[0197] Cloud server: Alibaba Cloud ECS instance (4 cores, 8GB memory), Ubuntu 20.04 operating system, MySQL 8.0 database, using MQTT protocol for communication.

[0198] Patient-side: Android and iOS apps, supporting manual input or Bluetooth connection to smart bracelets to collect patient information, with built-in LED, buzzer and display warning modules.

[0199] Healthcare professionals: The web-based client is deployed on the hospital server and supports doctor account login. Doctors can view real-time patient data, historical trends, and alert records.

[0200] 3.2 Workflow

[0201] Referring to Figure 2, when a patient is undergoing peritoneal dialysis, the monitoring device is connected to the drainage tubing. After being turned on, the device automatically completes a self-test and connects to the cloud.

[0202] The device collects data in real time and completes feature extraction and model inference every 60 seconds. Under normal conditions, it only uploads summary information, while under risk conditions, it immediately triggers a local warning and uploads complete data.

[0203] The complete data comprises all the information supporting accurate diagnosis and traceability monitoring by medical staff, covering "raw data + processed data + judgment results," specifically including:

[0204] 1. Raw sensor data: All raw sampling data within a 60-second time window (1Hz sampling, a total of 60 sets of RGB three primary color component data + 60 sets of turbidity-related scattered light data).

[0205] 2. Preprocessed data: Smoothed data after 4D Kalman filtering, 24 optical features (6 features including max, min, and mean values ​​for R / G / B / turbidity respectively);

[0206] 3. Fusion data: 39-dimensional complete feature vector (24 optical features + 15 patient information items, such as age, body temperature, abdominal pain symptoms, etc.);

[0207] 4. Determine relevant data: the output of the artificial intelligence model (peritonitis risk probability / binary classification label), and key parameters of the model inference process (such as the voting results of the random forest and the probability output of the neural network).

[0208] 5. Auxiliary information: data acquisition timestamp, device operating status parameters (such as battery level, sensor calibration status).

[0209] The data digest is a "condensed information package" under normal conditions, retaining only core status identifiers and excluding redundant raw data. Specifically, it includes:

[0210] 1. Status assessment result: Only the conclusion of "no risk of peritonitis" is clearly stated (no probability details are required);

[0211] 2. Key feature statistics: Core summary values ​​of 24 optical features (such as the mean of R / G / B components, mean turbidity, no complete feature set required);

[0212] 3. Basic status information: data acquisition timestamp, verification indicators that the device is working properly (e.g., no sensor faults, sufficient battery power);

[0213] 4. Simplify patient information: Only retain basic information related to long-term trends (such as dialysis duration), without the need for real-time vital signs data (such as single temperature and heart rate).

[0214] In this embodiment, the size of the data digest is typically only 1 / 10 to 1 / 5 of the complete data, which can significantly reduce wireless transmission power consumption and cloud storage pressure, enabling the device to achieve long battery life when powered by lithium batteries. The complete data is used for "precise diagnosis and traceability", while the data digest is used for "status reporting and trend tracking". This design not only meets the needs of clinical diagnosis, but also adapts to the actual use scenario of home monitoring. It is a balanced innovation of function and power consumption (different from the "real-time full transmission" of the comparison file).

[0215] After receiving the warning information, the cloud server simultaneously pushes it to the patient's app and the medical staff's app;

[0216] Patients can view early warning guidelines through the app, and medical staff can use the data to determine whether the patient needs to seek medical attention. Once a diagnosis is made, a treatment plan will be issued.

[0217] 3.3 Technical Effects

[0218] The system's data transmission latency is consistently within 300ms, and it can support concurrent access from 500 devices. In a trial involving 20 patients in a hospital, it successfully provided early warning for 3 cases of early peritonitis, with the warning time being an average of 3 days earlier than the onset of clinical symptoms. None of the patients developed severe illness.

[0219] Example 4

[0220] This embodiment is an alternative implementation of the neural network model.

[0221] In this embodiment, the random forest model in embodiment 2 is replaced with a multilayer feedforward neural network, while other hardware and software processes remain unchanged.

[0222] 4.1 Model Structure and Training

[0223] The model employs a 3-layer structure: an input layer with 39 nodes, 3 hidden layers (512 neurons per layer), and an output layer with 1 node (using the sigmoid activation function). Training is performed using the TensorFlow 2.0 framework with a learning rate of 0.001, an L2 regularization coefficient of 0.01, an Adam optimizer, and a cross-entropy loss function. An early stopping strategy is used (the model stops if the validation set loss does not decrease after 5 epochs).

[0224] A multi-layer feedforward neural network is used for classification. The model consists of a 39-node input layer, 3 hidden layers, and 1 output layer. The output layer has 1 node, and the activation function is the sigmoid function, which outputs binary classification probabilities. The multi-layer feedforward neural network uses stochastic gradient descent (SGD) for backpropagation, with a learning rate of 0.001. L2 regularization (weight decay) is used to prevent overfitting. The categorical cross-entropy loss function is calculated, and the model converges and training stops when the loss function tends to plateau.

[0225] Each hidden layer contains 512 neurons, and the activation function is the sigmoid function.

[0226]

[0227] The output of the neuron is:

[0228]

[0229] in, This is the output of the i-th neuron in the previous layer. As weight, For bias, For activation functions;

[0230] The training loss function uses the Categorical Cross-Entropy loss function:

[0231]

[0232] in, The label is the actual sample label (0 or 1). Predict probabilities for the model.

[0233] During the prediction process, the 39 input features of the sample to be detected are normalized and then input into the neural network. After calculation by each layer, the output is a probability value (ranging from 0 to 1). A judgment threshold is set (e.g., 0.5). If the output probability is greater than or equal to the threshold, it is judged as a positive class (1, peritonitis has occurred); otherwise, it is judged as a negative class (0, peritonitis has not occurred).

[0234] 4.2 Technical Effects

[0235] The neural network model achieves an accuracy of 98.5%, sensitivity of 95.6%, and specificity of 99.2% on the same validation set, which is a slight improvement over the random forest model. The model size is approximately 2MB, and the single-sample inference time is 120ms, making it suitable for scenarios with higher accuracy requirements.

[0236] In summary, compared to existing technologies, this invention is non-obvious:

[0237] (i) The non-obviousness of the technical approach: the leap from "single-function detection" to "multi-dimensional closed-loop monitoring"

[0238] The core idea of ​​existing technologies is "single metric detection in a single scenario":

[0239] Comparative document 1 (US17201569) relies on sensor cards to detect specific biomarkers and requires the use of a circulatory system. It is limited to hospital or specific equipment scenarios and does not involve the fusion of individual patient information.

[0240] Comparison document 2 (US20170136166A1) only detects turbidity and color, uses a simple dual-algorithm switching, and does not build an intelligent analysis model or a doctor-patient linkage mechanism;

[0241] Comparison document 3 (CN109475678B) relies on smartphone photography and subjective symptom queries, and is greatly affected by the shooting environment and operating procedures, thus failing to achieve automated continuous monitoring.

[0242] This invention breaks through the limitations of existing technologies in terms of scenarios, dimensions, and functions, and proposes an overall approach of "automated home monitoring + multi-source data fusion + intelligent analysis + doctor-patient closed loop":

[0243] In response to the characteristics of home-based peritoneal dialysis treatment, the detection device is seamlessly integrated with the dialysis tubing. Through the miniaturization of optical modules and the embedded deployment of algorithms, true continuous home monitoring is achieved. This is not simply about reducing the size of existing equipment, but rather a cross-scenario innovation that comprehensively considers safety, convenience, and cost.

[0244] Existing technologies have not recognized the synergistic gain effect of individual patient information and optical features. This invention, through verification with a large amount of clinical data, integrates 15 patient information items with 24 optical features to construct a 39-dimensional feature vector, breaking the traditional understanding of "relying solely on dialysate indicators". It is an innovative discovery based on medical mechanisms and data mining.

[0245] By constructing a closed loop of "device-cloud-patient terminal", the problem of "disconnection between monitoring and intervention" in existing technologies is solved, enabling real-time transmission and rapid response of early warning information, and improving the efficiency of medical services.

[0246] (II) Non-obviousness of key technologies: Collaborative innovation of optical detection and intelligent algorithms

[0247] 1. Innovative Design of a Two-Dimensional Optical Inspection System

[0248] While prior art document 2 (US20170136166A1) mentions turbidity and colorimetry detection, it fails to address the issue of mutual interference between the two. Prior art document 3 (CN109475678B) relies on visual images, resulting in low detection accuracy. The innovation of this invention lies in:

[0249] By dividing the enclosed darkroom into independent areas using opaque partitions, physical isolation between color and turbidity detection is achieved, avoiding interference from light sources. This design requires multiple experimental verifications of the partition's position, material, and sealing method, and is not a conventional choice for those skilled in the art.

[0250] The turbidity detection uses an 860nm infrared light source combined with a vertical detection direction. The wavelength is selected based on the spectral characteristics of the dialysate components, and the vertical direction maximizes the capture of scattered light. Compared with the general light source and detection angle of Comparative Document 2, the detection accuracy is improved by 40%, which is an optimization and innovation based on physical mechanisms.

[0251] 2. Breakthroughs in data processing and intelligent algorithms

[0252] Comparison file 2 uses a simple dual-algorithm switching, and comparison file 3 only performs image brightness averaging calculation; neither involves complex data processing or intelligent models. The innovation of this invention lies in:

[0253] A 4D Kalman filter state vector containing RGB and turbidity was constructed, and the noise matrix was optimized for the flow characteristics of dialysate, improving data smoothness by 60%. This is not a simple application of the filtering algorithm, but an adaptive innovation of the algorithm.

[0254] A fusion scheme of "60-second time window + 39-dimensional features" is proposed. The time window is determined based on the change cycle of early peritonitis indicators, and the features are selected through mutual information entropy analysis. Compared with simple data splicing of comparison files, the model inference efficiency is improved by 30%.

[0255] Quantizing and deploying random forest and neural network models to embedded chips resolves the contradiction between "complex models and chip resource limitations." There are no similar engineering implementations of models for peritonitis monitoring in the comparative documents, which require the integration of knowledge across algorithm and hardware domains.

[0256] 3. System Integration and Engineering Implementation

[0257] None of the comparison documents 1-3 achieved complete system linkage. The innovation of this invention lies in:

[0258] The design employs an "event-driven" data transmission mode, uploading complete data only when a risk is detected and summary information is uploaded under normal conditions. This reduces device power consumption by 50%, adapting to home battery life requirements. Compared to real-time full transmission of comparison files, this represents a balanced innovation between functionality and power consumption.

[0259] The distributed cloud architecture supports concurrent access from thousands of devices, with data transmission latency controlled within 500ms, meeting the needs of large-scale clinical applications. However, its engineering implementation is more challenging than existing technologies.

[0260] (iii) Non-obviousness of technical effects

[0261] Existing technologies generally have an accuracy rate of less than 85% and can only provide early warnings a few hours in advance. However, this invention, through the aforementioned innovations, achieves an accuracy rate of over 98%, provides early warnings 3 days earlier, and reduces hospitalization rates by 45%. These effects are not simply the sum of existing technologies, but rather the result of the synergistic effect of various innovations, exceeding the conventional expectations of those skilled in the art.

[0262] This invention surpasses the existing technology level in terms of technical concept, key technology and technical effect. It solves the core problems of "limited scenarios, low accuracy, late warning and poor linkage" in traditional peritonitis monitoring. The integration and innovation of its technical solution cannot be easily thought of or achieved by those skilled in the art based on the existing technology. It has outstanding substantive features and significant progress, and meets the requirements of inventiveness.

[0263] The above descriptions are merely embodiments of the present invention, and common knowledge such as specific technical solutions and / or characteristics are not described in detail here. It should be noted that those skilled in the art can make various modifications and improvements without departing from the technical solutions of the present invention, and these should also be considered within the scope of protection of the present invention. These modifications and improvements will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. An artificial intelligence peritonitis monitoring device, characterized in that, The system includes a sealed darkroom, highly transparent tubing, a simulated natural light source, a color sensor, a parallel light source, a light sensor, an opaque partition, and a smart chip. The sealed darkroom is divided into an upper color monitoring area and a lower turbidity monitoring area by the opaque partition. The highly transparent tubing vertically penetrates the opaque partition and connects the peritoneal dialysis tubing drainage end to the waste liquid tank. Within the color monitoring area, the simulated natural light source and the color sensor are respectively located on both sides of the highly transparent tubing. Light passing through the dialysate in the tubing is received by the color sensor to obtain the RGB primary color components. Within the turbidity monitoring area, the parallel light source is located on one side of the highly transparent tubing, and the light sensor is located on the tubing. The internal detection direction is perpendicular to the direction of parallel light, used to receive scattered light to calculate turbidity; the smart chip is connected to the color sensor and the light sensor respectively, and has built-in artificial intelligence algorithm. The smart chip suppresses noise by constructing a 4D Kalman filter state vector containing RGB three primary color components and turbidity. It judges the risk of peritonitis based on the fusion data of dialysate optical characteristics and patient information. The peritonitis identification sensitivity is ≥95%, specificity is ≥99%, and accuracy is ≥98%; the smart chip adopts an "event-driven" wireless data transmission mode, uploading complete data only when peritonitis risk is detected, and uploading data summary under normal conditions.

2. The apparatus according to claim 1, characterized in that, The simulated natural light source has a color temperature range of 5500K-6500K, the color sensor has a sampling frequency of 1Hz, and outputs 8-bit RGB component data with an accuracy of 8 bits; the parallel light source is an infrared LED with a wavelength of 860nm, and the distance between the light sensor and the parallel light source is 2cm-3cm.

3. The apparatus according to claim 1, characterized in that, The intelligent chip uses the nRF52832 chip, which includes a 16-bit precision data sampling module, an MCU microprocessor, an embedded artificial intelligence algorithm device, and a wireless communication module. The embedded artificial intelligence algorithm device is a random forest model or a multilayer feedforward neural network.

4. The apparatus according to claim 1, characterized in that, The highly transparent pipeline is made of quartz glass with an inner diameter of 8mm-10mm. It is a single-use structure with standard Luer connectors at both ends.

5. An artificial intelligence peritonitis monitoring system, characterized in that, The device includes the artificial intelligence peritonitis monitoring device as described in any one of claims 1-4, a cloud server, a medical care terminal, and a patient terminal; the monitoring device interacts with the cloud server via the wireless communication module of the smart chip, and the cloud server communicates bidirectionally with both the medical care terminal and the patient terminal; the patient terminal includes a patient information collection module and a built-in early warning module, the patient information collection module is used to collect 15 items of patient information, and the built-in early warning module receives early warning signals from the smart chip and alerts the patient in multiple ways; the smart chip of the monitoring device determines the risk of peritonitis based on the fusion data of the optical characteristics of dialysis fluid and patient information.

6. The system according to claim 5, characterized in that, The patient information collection module collects five basic information items, including gender and age, and ten symptom and sign information items, including body temperature and heart rate. The accuracy of body temperature collection is 0.1℃, and the error of heart rate collection does not exceed 2 beats / minute.

7. The system according to claim 5, characterized in that, The built-in early warning module includes an LED light, a buzzer, and an LCD screen. When the monitoring device determines that there is a risk of peritonitis, the LED light flashes at a frequency of 1Hz, the buzzer emits a warning sound at 2-second intervals, and the LCD screen displays the early warning information simultaneously.

8. The system according to claim 5, characterized in that, The cloud server adopts a distributed storage architecture, supporting concurrent data transmission from no less than 1,000 monitoring devices with a data transmission latency of no more than 500ms.

9. An artificial intelligence-based method for monitoring peritonitis, characterized in that, The system described in claim 5 comprises the following steps: S1: RGB data of the dialysate and turbidity-related scattered light data are collected by the color sensor and light sensor of the monitoring device, respectively, with a sampling frequency of 1Hz; S2: Kalman filtering is performed on the collected data to construct a 4-dimensional state vector containing the RGB primary color components and turbidity to achieve noise suppression; S3: 24 optical features within a 60-second time window are extracted and fused with 15 patient information items to form a 39-dimensional feature vector; S4: The feature vector is analyzed by an artificial intelligence algorithm, and the peritonitis risk assessment result is output based on the fused data of dialysate optical features and patient information; S5: If a high risk is determined, an early warning is triggered and the result is transmitted to the cloud, the medical staff terminal, and the patient terminal.

10. The method according to claim 9, characterized in that, The artificial intelligence algorithm in step S4 is a random forest model, which consists of 200 decision trees. When splitting a node, 6 features are randomly selected and the Gini index is used as the splitting criterion. The model has a sensitivity of ≥95%, a specificity of ≥99%, and an accuracy of ≥98% for the identification of peritonitis.

Images