Stoma base plate based on three-dimensional liquid diode, stoma detection system and stoma detection algorithm

Abstract

The invention relates to a stoma base plate based on three-dimensional liquid diodes. The stoma base plate is composed of a liquid transmission layer and an intelligent information acquisition layer. The liquid transmission layer comprises a vertical liquid diode layer which is arranged on the bottom layer close to the skin surface and used for achieving preliminary absorption of sweat or interstitial fluid and one-way transmission in the direction perpendicular to the skin surface; the horizontal liquid diode layer is arranged on the upper layer of the vertical liquid diode and is used for receiving sweat or tissue fluid transmitted by the vertical liquid diode and unidirectionally guiding the sweat or tissue fluid to an edge outlet along the horizontal direction and finally discharging the sweat or tissue fluid; the intelligent information acquisition layer comprises a miniature thermocouple and a miniature borborygmus sensor and covers the upper layer of the horizontal liquid diode layer or the lower layer of the vertical liquid diode layer, and an electrode part of the intelligent information acquisition layer is coupled with sweat.

Description

Technical Field

[0001] The present invention belongs to the technical field of stoma detection, and in particular relates to a stoma chassis, a stoma detection system, and an algorithm based on a three-dimensional liquid diode. Background Art

[0002] Stoma trays are important medical devices in stoma care, used to protect the skin around the stoma, connect to the ostomy bag, and prevent leakage of excreta. However, existing stoma trays or dressings generally suffer from insufficient drainage when dealing with sweat or tissue fluid on the skin surface. Although some products have water absorption functions, they lack effective directional drainage capabilities, causing sweat or tissue fluid to remain on the skin surface and unable to be continuously transported and discharged along the preset path. This can cause long-term irritation to the epidermis, resulting in unstable tray connections and even complications in the skin around the stoma.

[0003] Skin-contact or textile-based wearable electronic devices have poor air permeability, and sweat easily accumulates at the interface between the skin and the device, causing discomfort to the user and reduced accuracy of signal monitoring over long periods of time. Physiological signals near the stoma are mostly collected using sensors that are attached to the skin. However, existing commercial electrodes require a coupling agent to collect electrical or acoustic signals (acoustic and electrical signals need to be transmitted by liquid on the skin surface). In this case, sweat in a specified direction is beneficial for coupling, while sweat in non-specified directions can easily damage the sensor.

[0004] Currently, Yu Xinge's team at the City University of Hong Kong, China, has proposed a three-dimensional liquid diode solution for soft, integrated permeable electronics. They have constructed a three-dimensional liquid diode (3D LD) with a bionic microstructure, which enables sweat to be continuously transported and discharged along a preset path, while preventing sweat from flowing back, and allowing wearable electronic devices to be directly integrated above it. Specifically, the vertical liquid diode (VLD) absorbs sweat from the skin surface in a unidirectional direction (perpendicular to the skin); the horizontal liquid diode (HLD) discharges the sweat absorbed by the VLD in a unidirectional direction (horizontal to the skin) to the edge outlet. The theoretical maximum sweating rate of 3D LD is 11.6ml / (cm2·min), which is 4,000 times higher than the physiological sweat rate during exercise. Even under sweating conditions, it has excellent skin-friendliness, user comfort and stable signal reading behavior. Summary of the Invention

[0005] The purpose of the present invention is to provide, based on the above-mentioned three-dimensional liquid diode solution, a stoma chassis designed to solve the above-mentioned problems, which can directionally guide skin sweat, and the chassis can better collect electrophysiological signals near the stoma (sweat in the specified direction is beneficial to coupling, avoiding sweat in non-specified directions from damaging the sensor).

[0006] In order to achieve the above objectives, the technical solution adopted by the present invention is to design a stoma baseplate based on a three-dimensional liquid diode, which is composed of a liquid transmission layer and an intelligent information collection layer:

[0007] The liquid transmission layer is composed of the two types of liquid diodes mentioned above, including a vertical liquid diode (VLD) layer, which is arranged on the bottom layer close to the skin surface, and is used to achieve initial absorption of sweat or tissue fluid and unidirectional transmission perpendicular to the skin surface; a horizontal liquid diode (HLD) layer, which is arranged on the upper layer of the vertical liquid diode (VLD), and is used to receive sweat or tissue fluid transmitted by the vertical liquid diode, and guide it unidirectionally in the horizontal direction to the edge outlet and finally discharge it.

[0008] The intelligent information collection layer, which covers the upper layer of the horizontal liquid diode (HLD) layer or the lower layer of the vertical liquid diode (VLD) layer, that is, between the skin and the vertical liquid diode (VLD) layer, is mainly a sensor with electrodes, and its electrodes are coupled with sweat. In other words, sweat passes through the electrodes as it is discharged, forming a coupling.

[0009] Furthermore, an edge liquid collection layer is provided at the edge of the horizontal liquid diode (HLD) layer, which uses a polymer absorbent resin (SAP) composite activated carbon and silver ions to absorb liquid and resist bacteria, preventing the backflow of sweat or tissue fluid, thereby keeping the skin breathable and dry during wearing.

[0010] A stoma detection system is based on the above-mentioned multi-layer stoma chassis, and integrates sensors in the intelligent information collection layer:

[0011] Micro-thermocouples monitor skin temperature (accuracy ±0.1°C). Placed between the skin and the vertical liquid diode (VLD) layer, the electrodes couple to the sweat and detect the temperature of sweat on the skin's surface. This temperature, secreted by sweat glands, represents the subcutaneous temperature, making it more suitable for detecting stoma inflammation. (Stomal inflammation sometimes occurs not on the skin's surface, but 1-2 cm below. Measuring only the skin's surface temperature cannot effectively detect the subcutaneous / inflammatory temperature.)

[0012] Miniature bowel sound sensor: It uses a MEMS microphone with a frequency range of 100-600Hz. The electrodes are coupled to sweat to collect water sound signals from the sweat. Similarly, the sound signal is transmitted more effectively in sweat, avoiding the side effects of poor contact between the microphone and the skin. At the same time, the volume of the liquid in the ostomy bag will change, causing the liquid to emit sounds of different frequencies (the higher the water level, the higher the frequency), but this sound cannot be heard by the user or will be ignored by the ears. At the same time, sounds of different frequencies can also reflect the concentration and texture of the contents, thereby indirectly judging the extent to which the liquid in the ostomy bag is prone to overflow and odor (more viscous liquids are considered to be more likely to overflow and odor). The miniature bowel sound sensor can also detect sounds emitted by the liquid in the ostomy bag that are not noticed or heard by the user, and combined with AI intelligent judgment, it reminds whether the ostomy bag needs to be replaced.

[0013] The stoma is often obscured by the stoma base and pouch, making it difficult for the user and medical staff to see the stoma's condition in real time. Therefore, it is necessary to install several sensors on the stoma base to facilitate the timely detection of stoma abnormalities. Therefore, a series of abnormality indicators are needed to facilitate automatic alarms by the intelligent stoma detection system.

[0014] Intelligent detection system / sensor alarm clinical indicators:

[0015] The normal bowel sound is 4-5 times / minute; more than 10 times / minute is active bowel sound, which is seen in acute enteritis and gastrointestinal bleeding; accompanied by loud, high-pitched, metallic sound is hyperactive bowel sound, which is seen in mechanical intestinal obstruction; bowel sounds less than normal are weakened bowel sounds, which can also be caused by senile constipation, peritonitis, electrolyte imbalance, low gastrointestinal motility, and persistent obstruction; if no bowel sounds are heard after 3-5 minutes of continuous auscultation, and there are still no bowel sounds after stimulating the abdomen, it is bowel sound disappearance; therefore, auscultation is required to last at least 3-5 minutes, which is seen in acute peritonitis and paralytic intestinal obstruction.

[0016] Skin temperature (ST) refers to the temperature of the outermost layer of the human body. Normal skin temperature is between 33.5°C and 36.9°C. Skin temperature is closely related to heart activity and sweating, and can fluctuate depending on stress and anxiety levels. Localized skin infections such as cellulitis, erysipelas, furuncles, and carbuncles can cause localized inflammation, resulting in redness, swelling, and fever.

[0017] When the skin is inflamed, its surface temperature will increase significantly, the thickness of the damaged part of the epidermis will become thinner, and the hardness and elasticity will also change, causing changes in the volume of bowel sounds. Therefore, a skin temperature sensor and a bowel sound collection sensor are set above the liquid transmission layer to regularly collect temperature and sound data. Combined with the infrared liquid level sensor in the ostomy bag, the volume of the bag contents is monitored to remind patients to check their skin health and replace the stoma base plate and ostomy bag in time.

[0018] Normal abdominal skin temperature is approximately 36°C, and inflammation may raise it by 1-2°C. Therefore, the temperature deviation score (TDS) can be set as:

[0019]

[0020] Where Tcurrent is the current skin temperature; Tbaseline is the baseline temperature (e.g., 36°C); and Tthreshold is the inflammation threshold temperature (e.g., 37.5°C). If the current skin temperature is lower than the baseline temperature, resulting in a negative TDS calculation result, TDS is always set to 0.

[0021] Normal bowel sounds are "gas-passing" or "intermittent gurgling," with a high and short pitch. Their intensity should be close to a whisper or weaker, within the 20-40dB range. If the skin becomes inflamed and becomes thinner or harder, the volume of the monitored bowel sounds may increase. The frequency range of normal bowel sounds is concentrated in the 100-600Hz band. The bowel sound change score (BSCS) is:

[0022]

[0023] Wherein, dB is the bowel sound volume obtained by monitoring close to the abdominal wall, dBcurrent is the current bowel sound volume, dBbaseline is the baseline bowel sound volume, for example, 40dB, and dBthreshold is the inflammation threshold bowel sound volume, for example, 50dB. fc is the center value of the bowel sound frequency range, fc,current is the center value of the current bowel sound frequency range, fc,baseline is the center value of the baseline bowel sound frequency range, for example, 300Hz, and fc,threshold is the center value of the inflammation threshold bowel sound frequency range, for example, 400Hz.

[0024] Taking the above two scores into consideration and assigning weights w1 and w2 to them respectively (considering that skin temperature changes can more directly reflect the degree of inflammation, the values can be w1 = 0.6 and w2 = 0.4, which can also be adjusted according to clinical practice) to obtain the ostomy skin health score (OSHS):

[0025] OSHS=w1·TDS+w2·BSCS

[0026] To avoid false alarms caused by instantaneous fluctuations, the average of the last N measurements is used (for example, 6 measurements in 24 hours). If OSHSavg ≥ 1.0, an alarm is triggered:

[0027]

[0028] Beneficial effects of the present invention:

[0029] The multi-layered stoma chassis structure directs the removal of sweat, a major problem with the stoma. By leveraging the conductivity and acoustic transmission of sweat, which can cause stoma damage, it is coupled to sensor electrodes to better collect physiological signals near the stoma, facilitating 24-hour automated stoma monitoring.

[0030] The health scoring formulas (TDS, BSCS, OSHS) have clear mathematical expressions, reasonable parameter definitions, and threshold settings based on extensive clinical experience.

[0031] Multimodal health monitoring: Integrates temperature, bowel sound, and liquid level sensors, combined with a dynamic scoring algorithm, to achieve early warning of inflammation, breaking through the limitations of single parameter monitoring.

[0032] Scientific nature of the scoring formula:

[0033] The normalization design of TDS and BSCS was reasonable, and the weight distribution (0.6:0.4) was consistent with the direct indicative effect of temperature on inflammation.

[0034] The sliding window average (OSHS_avg) effectively reduces false alarms caused by instantaneous fluctuations. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1This is an implementation of Example 1. The sensor electrodes in the figure are positioned between the skin and the vertical liquid diode (VLD) layer. The VLD layer first guides sweat from the skin into the HLD layer, and then from the center of the HLD layer into the periphery of the HLD layer. The periphery of the HLD layer can be configured in a variety of ways, such as providing a water reservoir to absorb and store liquid. Alternatively, the HLD layer can be configured to facilitate coupling with the electrodes, coupling sweat to the electrodes at the periphery of the HLD layer. Alternatively, multiple coupling slots (for sweat introduction) can be provided to facilitate insertion of electrodes from various sensors into the coupling slots. DETAILED DESCRIPTION

[0036] The present invention will be further described below with reference to the embodiments of the present invention and the accompanying drawings.

[0037] Example 1

[0038] like Figure 1 The overall shape is a circle with a hole in the middle, which is similar to the size and shape of a common stoma base plate.

[0039] The vertical liquid diode layer is close to the skin layer and uses ordinary polyester fabric as the base. It is first immersed in a hydrophobic slurry (a fluorinated silane hydrophobic agent reduces the surface energy of the fabric, nano-silica / titanium dioxide particles form a microstructure to increase surface roughness, and a thickener enhances viscosity and adhesion) to make it super-hydrophobic. Then, a certain area of the surface is treated with plasma to convert the surface of the area into hydrophilic, so that the liquid can move vertically upward in this area. As can be seen from the figure, the sweat discharged by the sweat glands on the surface of the skin is actively discharged upward by the vertical liquid diode layer. According to the theory of microdynamics, the sweating rate is 11.6ml / (cm2·min), which is 4000 times higher than the physiological sweating rate during exercise. It has excellent skin friendliness and user comfort even under sweating conditions.

[0040] The horizontal liquid diode layer covers the upper surface of the vertical liquid diode layer and is away from the skin surface. It is made of medical-grade silicon-based materials such as polydimethylsiloxane (PDMS). Microcolumns are etched on the surface of PDMS. The microcolumns have a sparse arrangement in the middle and dense arrangement on the outside. The pressure difference between the two interfaces can guide sweat droplets to move from the side with wider microcolumn spacing to the narrower side. Figure 1 It can be seen that the sweat transmitted by the vertical liquid diode layer is transported to the periphery and discharged by the horizontal liquid diode layer.

[0041] The stoma chassis is provided with an edge liquid collection layer ( Figure 1 (left) Made of a polymer absorbent polymer (SAP) material, it boasts strong liquid absorption and storage capabilities. It also incorporates activated carbon and silver ion composite materials to absorb odors and provide antibacterial properties. Sweat transported by the horizontal liquid diode layer is absorbed and stored by the peripheral liquid collection layer.

[0042] Likewise Figure 1 , electrode grooves can also be set around the horizontal liquid diode layer ( Figure 1 (right side) The electrode slot connects to the sweat transported by the horizontal liquid diode. The sensor electrodes can be inserted into the electrode slot, thus achieving the purpose of coupling the electrodes and sweat. This sweat, which would otherwise damage the sensor, has the effect of enhancing the electrode coupling due to the electrode slot structure, making the sensor collection effect better.

[0043] Intelligent information collection layer: It uses ordinary polyester fabric as the base, integrates micro thermocouples and micro wireless biological sound collection sensors on its surface, is powered by a built-in lithium battery, and is connected to a mobile phone app via Bluetooth. Figure 1 In the embodiment, it can be set at the bottom of the vertical liquid diode layer and in the sandwich layer on the skin surface, or it can be set at the top of the horizontal liquid diode layer.

[0044] The layers are bonded with silicone adhesive, or ultrasonic welding / hot pressing are used for bonding, or small magnets are used for magnetic connection to facilitate quick disassembly and replacement of a part.

[0045] The surface in contact with the skin is covered with a hydrogel adhesive, which maintains effective adhesion to the skin while not affecting the drainage of liquid on the skin surface.

[0046] Data acquisition: MEMS temperature sensor + piezoelectric acoustic sensor (also compatible with pH sensor), sampling rate: 100Hz±5%;

[0047] Edge computing: ARM Cortex-M7 + CMSIS-NN acceleration, inference latency: <3ms

[0048] Core model: Dual LSTM (64-32) + attention mechanism, AUC: 0.93±0.02

[0049] Feedback interface: HL7 medical data standard + DICOM image interface, compatible with hospital information system.

[0050] Example 2

[0051] Individualized stoma health monitoring algorithm based on machine learning + manual scoring.

[0052] 1. Data collection and annotation

[0053] Physician Rating Mechanism:

[0054] Each time the system triggers an alarm (level one or level two), the physician will conduct a clinical assessment of the patient's peristomal skin condition and give a quantitative score (such as 0-1 points, 0 for normal and 1 for severe inflammation).

[0055] If there is no alarm, monitoring data from some time points can be randomly sampled regularly (such as weekly) and supplemented by doctors to ensure data balance.

[0056] Annotation data format:

[0057] Input features: current and historical monitoring data (temperature, bowel sound volume, frequency center value, liquid level, TDS, BSCS, OSHS and their statistics).

[0058] Target variable: Physician rating (continuous value or categorical label).

[0059] 2. Feature Engineering

[0060] Time series feature extraction:

[0061] Sliding window statistics: Extract the mean, variance, maximum value, trend slope, etc. of the monitoring data within N hours before the alarm.

[0062] Event correlation characteristics: alarm frequency, alarm level, interval between adjacent alarms, etc.

[0063] Integration of individual characteristics:

[0064] Patient baseline data: age, weight, stoma type, previous complications, etc. (need to be entered through the App).

[0065] Environmental factors: season, activity intensity (collected with the assistance of acceleration sensors).

[0066] 3. Model selection and training

[0067] Algorithm selection:

[0068] Regression tasks (predicting physician scores): gradient boosting trees (such as XGBoost, LightGBM), support vector regression (SVR).

[0069] Classification task (inflammation grade classification): Random Forest, Logistic Regression, Shallow Neural Network.

[0070] Incremental learning design:

[0071] Using an online learning framework (such as the River library), the model parameters are updated after each new physician's score, gradually adapting to individual differences.

[0072] Initially, a pre-trained model (based on historical population data) is used, and then fine-tuned using patient data.

[0073] 4. Model Evaluation and Feedback Mechanism

[0074] Evaluation Metrics:

[0075] Regression task: mean square error (MSE), mean absolute error (MAE).

[0076] Classification task: accuracy, AUC-ROC curve, F1 score.

[0077] Feedback closed loop design:

[0078] The system regularly generates model performance reports to prompt physicians to review abnormal prediction cases.

[0079] Physicians can manually modify the model's prediction results, and the modified data will be automatically added to the training set, triggering model retraining.

[0080] 5. Clinical deployment and validation

[0081] Pilot Testing:

[0082] A small patient population (e.g., 50 cases) was selected and 3 months of data were collected to verify the consistency between the model predictions and the physician scores.

[0083] Compare the false alarm rate and missed alarm rate of the traditional OSHS algorithm and the machine learning model.

[0084] Enhanced interpretability:

[0085] Use SHAP values (Shapley Additive Explanations) to visualize feature contributions and help physicians understand the model's decision logic.

[0086] Example 3, based on Example 2, adds a staged optimization process.

[0087] 1. Data Preprocessing and Feature Engineering

[0088] Multi-dimensional data fusion:

[0089] In addition to temperature, sound intensity, and frequency, it also integrates auxiliary sensor data such as humidity and pH value

[0090] Extract time series features (sliding window mean, rate of change, extreme value distribution)

[0091] Constructing composite features: quadratic terms and interaction terms of TDS / BSCS

[0092] Personalized baseline calibration:

[0093]

[0094] 2. Dynamic Threshold Optimization Model

[0095] Use a two-layer LSTM network to implement time series modeling:

[0096]

[0097]

[0098] 3. Adaptive Weight Adjustment Mechanism

[0099] Attention Mechanism:

[0100] from keras.self_attention import SeqSelfAttention

[0101] attn_model=Sequential()

[0102] attn_model.add(LSTM(64,return_sequences=True))

[0103] attn_model.add(SeqSelfAttention(attention_activation='sigmoid')) dynamic allocation of feature importance:

[0104]

[0105] 4. Decision Optimization Module

[0106] Fuzzy logic control:

[0107] import skfuzzy as fuzz

[0108] #Define fuzzy sets: temperature, sound, alarm level

[0109] temp_universe=np.arange(35,39,0.1)

[0110] sound_universe=np.arange(30,60,1)

[0111] alarm_level = fuzz.trimf(np.arange(0,1.1,0.1),[0.7,0.9,1.0]) Reinforcement learning threshold optimization:

[0112]

[0113] 5. System Integration and Verification

[0114] Online Learning Framework:

[0115]

[0116] Evaluation Metrics:

[0117] Time-weighted alarm accuracy (TWAR)

[0118] Clinical false positive cost function:

[0119] Cost=0.8*FN+0.2*FP#focus on reducing the false negative rate

[0120] VI. Clinical Deployment Plan

[0121] Edge computing optimization:

[0122] Model quantization using TensorFlow Lite

[0123] Deploy a lightweight inference engine:

[0124] void run_inference(float*input,float*output){

[0125] / / Use the ARM CMSIS-NN library to accelerate inference

[0126] }

[0127] This solution improves upon the shortcomings of traditional fixed-threshold methods by employing dynamic feature weighting, personalized baseline calibration, and fuzzy decision-making techniques. Experiments show that it can reduce the false positive rate by 37% while maintaining a 98% recall rate on a simulated dataset.

Claims

1. A stoma baseplate based on a three-dimensional liquid diode, characterized in that: It consists of a liquid transmission layer and an intelligent information collection layer: The liquid transport layer includes: A vertical liquid diode layer is provided at the bottom layer close to the skin surface, and is used to achieve the initial absorption of sweat or tissue fluid and unidirectional transmission perpendicular to the skin surface; The horizontal liquid diode layer is arranged on the upper layer of the vertical liquid diode and away from the skin surface, and is used to receive the sweat or tissue fluid transmitted by the vertical liquid diode and guide it in a unidirectional manner in the horizontal direction to the edge outlet and finally discharge it; An edge liquid collection layer is set around the horizontal liquid diode layer, which is made of a polymer water-absorbent resin (SAP) material and has a strong ability to absorb and store liquid; An electrode groove is provided around the horizontal liquid diode layer, the electrode groove is connected to the sweat transported by the horizontal liquid diode, and the sensor electrode is inserted into the electrode groove; The intelligent information collection layer includes a micro-thermocouple and a micro-bowel sound sensor, which is covered on the upper layer of the horizontal liquid diode layer or the lower layer of the vertical liquid diode layer, and its electrodes are inserted into the electrode slot.

2. An intelligent stoma detection system, characterized in that: The sensor data as claimed in claim 1 is collected, and the alarm algorithm is as follows: the temperature deviation score is set to: Wherein, Tcurrent is the current skin temperature; Tbaseline is the baseline temperature; Tthreshold is the inflammation threshold temperature. When the current skin temperature is lower than the baseline temperature and the TDS calculation result is negative, TDS=0. Bowel sound change scores were set as: Wherein, dB is the bowel sound intensity obtained by monitoring close to the abdominal wall, dBcurrent is the current bowel sound intensity, dBbaseline is the baseline bowel sound intensity, dBthreshold is the inflammation threshold bowel sound intensity, fc is the center value of the bowel sound frequency range, fc,current is the center value of the current bowel sound frequency range, and fc,baseline is the center value of the baseline bowel sound frequency range. fc,threshold is the center value of the inflammation threshold bowel sound frequency range; Taking the above two scores into consideration and assigning weights w1 and w2 to them, w1 = 0.6 and w2 = 0.4, we can obtain the health score of the skin around the stoma: OSHS=w1·TDS+w2·BSCS. If OSHS≥1.0, an alarm is triggered.

3. A stoma monitoring algorithm based on machine learning, characterized in that: include: Multimodal sensor module: used to collect real-time data on temperature, sound intensity, frequency, humidity and pH value of the skin around the stoma; Edge computing module: deployed on each monitoring device to perform data preprocessing, feature extraction, and lightweight anomaly detection, wherein feature extraction includes calculating the temperature deviation score TDS and the bowel sound change score BSCS as described in claim 2; Dynamic threshold optimization module: This module uses a time series model that combines a two-layer LSTM network with an attention mechanism. The input is the TDS and BSCS feature sequence of the last N measurements, and the output is the health risk probability value. Federated learning module: Aggregates encrypted data from multiple monitoring devices, optimizes global model parameters through a spatiotemporal attention mechanism, and differentially distributes the data to each edge node; Feedback closed-loop module: Receives clinical verification results and patient physiological feedback data to trigger online model fine-tuning and threshold adaptive adjustment.

4. A stoma monitoring algorithm based on machine learning as claimed in claim 3, characterized in that: The data preprocessing of the edge computing module includes: The temperature time series data is lossily compressed using the Swing Door algorithm, and the compression deviation threshold is set to ±5% of the baseline value; Dynamic time warping (DTW) is used to align bowel sound frequency sequences collected by multiple devices.

5. The stoma monitoring algorithm based on machine learning according to claim 3, characterized in that: The federated learning module performs the following operations: Adopt multi-party secure computing (MPC) protocol to encrypt and aggregate gradient data; Extract the geographical distribution features of device clusters through the spatiotemporal graph convolutional network ST-GCN; Use the concept drift detection algorithm Page-Hinkley test to trigger local model retraining.

6. The stoma monitoring algorithm based on machine learning according to claim 3, characterized in that: Add manual scoring process, including the following steps, 1. Data collection and annotation Physician Rating Mechanism: Each time the system triggers an alarm, a physician will conduct a clinical assessment of the patient's peristomal skin condition and provide a quantitative score; If there is no alarm, monitoring data from some time points can be randomly sampled regularly and supplemented by doctors to ensure data balance; Annotation data format: Input features: current and historical monitoring data, such as temperature, bowel sound intensity, frequency center value, liquid level, TDS, BSCS, OSHS and their statistics as described in claim 2; Target variable: Physician rating; 2. Feature Engineering Time series feature extraction: Sliding window statistics: extract the mean, variance, maximum value, and trend slope of monitoring data within N hours before the alarm; Event correlation characteristics: alarm frequency, alarm level, and interval between adjacent alarms; Integration of individual characteristics: Patient baseline data: age, weight, stoma type, previous complications, etc.; Environmental factors: season, activity intensity; 3. Model selection and training Algorithm selection: Regression task or classification task; Incremental learning design: Using an online learning framework, the model parameters are updated after each new physician's rating, gradually adapting to individual differences; Initially use a pre-trained model and then fine-tune it using patient data; 4. Model Evaluation and Feedback Mechanism Evaluation Metrics: Regression task: mean square error, mean absolute error; Classification tasks: accuracy, AUC-ROC curve, F1 score; Feedback closed loop design: The system regularly generates model performance reports, prompting physicians to review abnormal prediction cases; Physicians can manually modify the model's prediction results, and the modified data will be automatically added to the training set, triggering model retraining.

7. The stoma monitoring algorithm based on machine learning according to claim 3, characterized in that: Add a phased optimization process, including the following steps:

1. Data preprocessing and feature engineering Multi-dimensional data fusion: In addition to temperature, sound intensity, and frequency, it also integrates auxiliary sensor data such as humidity and pH value Extract time series features, sliding window mean, rate of change, and extreme value distribution; Construct composite features: quadratic terms and interaction terms of TDS / BSCS; Personalized baseline calibration; 2. Dynamic Threshold Optimization Model Use a two-layer LSTM network to implement time series modeling:

3. Adaptive weight adjustment mechanism Attention mechanism; Dynamic allocation of feature importance; 4. Decision Optimization Module Fuzzy logic control; Reinforcement learning threshold optimization; 5. System integration and verification Online learning framework; Evaluation metrics; Time-weighted alarm accuracy; Clinical false alarm cost function; 6. Clinical deployment plan Edge computing optimization: Use TensorFlow Lite for model quantization; Deploy a lightweight inference engine.

8. The intelligent stoma detection system according to claim 2, characterized in that: To avoid false alarms caused by instantaneous fluctuations, the average of the last N measurements is used. If OSHSavg ≥ 1.0, an alarm is triggered:

Images