An ostomy complication early prediction and intelligent auxiliary nursing system based on artificial intelligence
The intelligent nursing system, which utilizes multimodal data acquisition and edge computing, addresses the issues of single-modality, delayed response, and privacy risks in existing stoma care technologies. It enables early prediction and automated intervention of stoma complications, significantly reducing the incidence of complications.
Patent Information
- Application Number
- CN202611080108.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-21
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies in stoma care rely on a single image modality, ignoring information such as the physical and chemical parameters of excretion and the patient's activity level. Cloud processing suffers from delays and privacy risks, lacks interpretability, cannot respond to complications in real time, and can only identify abnormalities that have occurred rather than make early predictions, resulting in passive intervention methods.
The system employs a multimodal data acquisition module to simultaneously acquire stoma images, excretion physicochemical parameters, and patient vital signs. An edge computing module extracts temporal features in real time and generates interpretable feature representations. A cloud center prediction module is used to predict the probability of future complications. A closed-loop intervention module executes tiered nursing instructions, including automated interventions using augmented reality devices and smart ostomy bags.
It enables early prediction of complications and automatic execution of personalized nursing interventions while protecting privacy and providing real-time response, significantly reducing the incidence and severity of complications and improving the intelligence and precision of nursing care.
Smart Images

Figure CN123120282A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of medical care and artificial intelligence, specifically to an artificial intelligence-based system for early prediction and intelligent assisted care of stoma complications. Background Technology
[0002] Comorostomy is an important postoperative drainage diversion procedure for patients with colorectal cancer, bladder cancer, and other diseases. Hundreds of thousands of new comorbidities are diagnosed globally each year, with a postoperative complication rate as high as 30%-60%, mainly including leakage, infection, skin detachment, and parastomal hernia. Early identification and timely intervention of complications are crucial for improving patients' quality of life.
[0003] In existing technologies, stoma care mainly relies on patient self-observation and regular follow-up visits. Some solutions attempt to introduce image recognition technology, using smartphones to take photos of the stoma and upload them to the cloud for classification and judgment. However, such solutions have obvious shortcomings: (1) they rely on only a single image modality, ignoring information closely related to complications, such as excretion physicochemical parameters and patient activity levels; (2) cloud processing has delays and privacy risks, and cannot respond in real time; (3) they lack interpretability, and patients and nurses find it difficult to understand the basis of the model's judgment; (4) their predictive ability is limited, and most can only identify abnormalities that have occurred rather than predict them in the early stages; (5) the intervention methods are passive, only providing information prompts and unable to link nursing devices to perform automated interventions.
[0004] Therefore, there is an urgent need for an intelligent nursing system that can integrate multimodal data, perform real-time analysis at the edge, possess interpretability and early prediction capabilities, and proactively implement tiered interventions. Summary of the Invention
[0005] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide an artificial intelligence-based system for early prediction and intelligent assisted care of stoma complications.
[0006] To achieve the aforementioned objective, the technical solution of the present invention is implemented as follows: an artificial intelligence-based early prediction and intelligent assisted nursing system for stoma complications, comprising: The multimodal data acquisition module is used to simultaneously acquire image data of the stoma and surrounding tissues, as well as physical or chemical parameter data of excrement.
[0007] The edge computing module, deployed at the patient end, is used to extract temporal features from multimodal data and generate a feature representation of the current state and an explanation of the contribution of this feature representation to various complication types through an interpretable heterogeneous data fusion model.
[0008] The cloud center prediction module is used to receive feature representations and predict the probability of at least one stoma complication occurring within a preset time window based on a time-series generation model.
[0009] The closed-loop intervention module is used to automatically generate and execute graded nursing instructions based on probability and contribution interpretation. The graded nursing instructions include at least sensory guidance instructions for the patient or physical drive instructions for the stoma care device.
[0010] Based on the above system, the present invention also provides a method for early prediction and intelligent assisted nursing of stoma complications based on the same system architecture, including: synchronously collecting multimodal data; extracting temporal features at the edge and generating interpretable feature representations; sending the feature representations to the cloud and predicting the probability of future complications based on the temporal generation model; and generating and executing graded nursing instructions based on the probability and contribution interpretation.
[0011] Preferably, the graded nursing instructions generated by the closed-loop intervention module include: nursing operation guidance instructions superimposed and displayed through an augmented reality device, instructions to release active substances through a smart ostomy bag, or instructions to automatically coordinate external nursing resources.
[0012] Preferably, the interpretable heterogeneous data fusion model in the edge computing module adopts a graph neural network based on an attention mechanism. The attention coefficients of this network are dynamically calculated based on the temporal correlation between the data of each modality and the pre-trained prior correlation matrix.
[0013] Preferably, the time-series generation model in the cloud center prediction module is a conditional diffusion model or a conditional generative adversarial network. This model uses historical multimodal fusion features as conditions to predict the feature representation of future complication events through reverse denoising or adversarial generation.
[0014] Preferably, the multimodal data acquisition module further includes wearable sensors for detecting the patient's vital signs or physical activity level, and the multimodal data further includes vital signs or physical activity level data.
[0015] Preferably, the edge computing module employs a lightweight temporal feature extraction network derived from knowledge distillation, which adapts the model size and inference power consumption to the resource constraints of edge devices.
[0016] Preferably, the cloud center prediction module also includes a federated learning aggregator, which aggregates model gradient updates uploaded by multiple edge computing modules to globally update the prediction model while protecting patient data from leaving the local machine.
[0017] Preferably, the closed-loop intervention module further includes a counterfactual interpretation submodule, which outputs the predicted change in complication risk after changing at least one interventionable feature and generates nursing recommendations with minimal intervention cost.
[0018] The beneficial effects of this invention are reflected in: This invention synchronously acquires multidimensional data such as stoma images, excretion physicochemical parameters, patient vital signs, and activity levels through a multimodal data acquisition module. Temporal features are extracted in real-time by an edge computing module on the patient's end, and an interpretable heterogeneous data fusion model is used to generate a feature representation and contribution explanation for the current state. This allows patients and caregivers to intuitively understand the basis of the model's judgments while protecting privacy, reducing bandwidth, and ensuring real-time response. Subsequently, this feature representation is transmitted to the cloud center, where a temporal generative model (such as a conditional diffusion model or generative adversarial network) predicts the probability distribution and severity of complications such as leakage, infection, skin detachment, and parastomal hernia within the next 24-72 hours. This overcomes the lag of traditional methods that can only identify abnormalities that have already occurred, achieving true early detection. Early warning is issued; then, the closed-loop intervention module automatically executes graded instructions based on the predicted risk level and interpretable path—personalized nursing guidance is displayed overlaid through augmented reality devices when the risk is low; physical intervention is achieved by releasing active substances through a smart ostomy bag when the risk is medium; and community nursing resources are automatically coordinated and patient data is synchronized when the risk is high—forming a complete closed loop of "perception-decision-execution-feedback," significantly reducing the incidence and severity of complications; at the same time, by leveraging a federated learning aggregator to collaboratively update the predictive model across multiple centers without aggregating raw data, and by having the counterfactual interpretation submodule output nursing suggestions with minimal intervention cost, privacy protection, continuous model optimization, and personalized nursing decisions are further balanced, improving the overall intelligence, precision, and proactivity of ostomy care. Attached Figure Description
[0019] In the attached diagram: Figure 1 This is a schematic diagram of the overall architecture of the system of the present invention.
[0020] Figure 2 This is a schematic diagram of the structure of an interpretable heterogeneous data fusion model in the edge computing module.
[0021] Figure 3 This is a schematic diagram illustrating the prediction principle of the time-series diffusion model in the cloud center prediction module.
[0022] Figure 4 This is a flowchart of the hierarchical instruction execution process for the closed-loop intervention module.
[0023] Figure 5 Example of the output interface for the counterfactual explanation submodule. Detailed Implementation
[0024] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. Obviously, the described embodiments are only a part of the embodiments of the invention, and not all of them. Unless otherwise specified, the embodiments and features described in this application can be combined with each other. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.
[0025] It should be noted that if the embodiments of the invention involve directional indicators (such as up, down, left, right, front, back, etc.), the directional indicators are only used to explain the relative positional relationship and movement of the components in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indicators will also change accordingly.
[0026] Furthermore, "multiple" refers to two or more. Additionally, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of a person skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the invention.
[0027] Example 1: System Overall Architecture like Figure 1 As shown, the stoma complication early prediction and intelligent assisted nursing system of this embodiment includes the following four core modules: Multimodal data acquisition module (100): includes a near-field spectral imaging unit (101) for capturing hyperspectral images of the stoma and surrounding skin area at regular intervals or when the patient changes the ostomy bag daily; a microfluidic biosensor array (102) integrated into the ostomy bag for real-time detection of pH, glucose concentration, occult blood index, and proteolytic enzyme activity in excrement; a wearable inertial measurement unit and a flexible piezoelectric film sensor (103) for capturing abdominal electromyographic signals and body movement; and a patient terminal (104) for inputting subjective discomfort levels and dietary records. All data are acquired in a time-synchronized manner with timestamp alignment.
[0028] Edge computing module (200): Deployed in the edge gateway device (such as a smartphone or dedicated handheld terminal) carried by the patient. This module includes: a lightweight temporal feature extractor (201), which adopts a temporal convolutional network and gated recurrent unit structure, and compresses the large model in the cloud to less than 2MB through knowledge distillation; and an interpretable heterogeneous data fusion model (202), which is built based on a graph attention network, uses the feature sequences of each modality as nodes and the temporal correlation between modalities as edges, dynamically calculates the contribution weight of each modality to different complication types, and outputs a fused feature vector and an interpretable heatmap.
[0029] Cloud-based prediction module (300): Deployed on a remote cloud server or hospital private cloud. This module includes: a temporal diffusion predictor (301), which uses the fused feature vector sequence uploaded from the edge as a condition to predict the probability distribution and severity of complications such as leakage, infection, skin peeling, and parastomal hernia within the next 24-72 hours through a reverse denoising process; and a federated learning aggregator (302), which receives local model gradient updates (after differential privacy scrambling) uploaded from multi-center edge nodes and globally updates the prediction model.
[0030] Closed-loop intervention module (400): Automatically executes graded instructions based on predicted risk level and interpretable heat map. Level 1 instruction (low risk): Displays the optimal dressing change posture guide and operation verification prompts in real time overlaid on the patient's stoma area via augmented reality projection device (401). Level 2 instruction (medium risk): Links with the intelligent ostomy bag (402) to activate its internal controllable release microneedle array, releasing antibacterial enzymes or healing-promoting factors at a predetermined dose. Level 3 instruction (high risk): Generates structured nursing events, automatically schedules community nursing visits, and synchronizes the current patient's multimodal risk data to the visiting nurse's terminal (403).
[0031] Example 2: Specific Implementation of an Explainable Heterogeneous Data Fusion Model like Figure 2 As shown, the interpretable heterogeneous data fusion model in this embodiment employs a graph attention network (GAT). Let's assume that K modal feature vectors are collected within a time window. Construct a fully connected graph, where node i has the following characteristics: Attention coefficient The calculation is as follows: in The introduced prior bias term is generated from a "complication-modality" association matrix pre-trained on large-scale real-world data. For example, for infectious complications, the prior weight of occult blood in feces is higher; for skin abrasion, the prior weight of image texture features is higher. This bias term effectively solves the problem of fusion reliability during the cold start phase or when a certain modality's data is temporarily missing. The final fused feature vector x^ is obtained through weighted summation, and the attention weight matrix is output as an interpretability heatmap.
[0032] Example 3: Early prediction of complications based on a time-series diffusion model like Figure 3As shown, the temporal diffusion model in the cloud center prediction module adopts a conditional generative adversarial diffusion architecture. Traditional diffusion models generate new samples by progressively adding noise to the data (forward process) and then learning inverse denoising. This embodiment transforms it into conditional prediction: using the multimodal fusion feature sequence c of the preceding time step as a condition, it predicts the feature representation y of complication events within the future time window.
[0033] Specifically, the training phase involves analyzing real-life complication precursor feature sequences. Gradual noise addition Train a denoising network Predict the added noise. Inference stage: from pure noise Initially, condition c guides the gradual denoising process to obtain the predicted values. Then, the decoder maps it to the complication type and severity level.
[0034] Introducing a discriminator network to distinguish generated... With the real The generator is trained adversarially to further improve the temporal consistency of predictions and the sensitivity of early warnings.
[0035] Example 4: Augmented Reality and Microneedle Array Driving in Closed-Loop Intervention This embodiment describes the specific execution device of the closed-loop intervention module.
[0036] Augmented reality projection device: This device includes a miniature structured light projector and a scene camera. It first performs a 3D reconstruction of the skin around the stoma, identifying topological changes such as wrinkles and swelling. The posture guide line in the primary command is not a fixed path, but rather its projection position is dynamically adjusted based on the real-time 3D topology, guiding the patient to fit the stoma bag base plate at the optimal angle and pressure, avoiding leakage due to improper fit.
[0037] The microneedle array of the smart ostomy bag consists of a thermally responsive shape-memory polymer substrate and soluble needle tips encapsulated with active drugs (such as antibacterial enzymes and healing-promoting factors). An edge computing module outputs pulse-width modulation (PWM) signals with different duty cycles based on the predicted risk level, locally heating the polymer substrate and causing it to shrink due to a phase change. Different temperatures correspond to different needle tip exposure depths, thus achieving multi-level (low / medium / high) controlled drug release, rather than a simple all-or-nothing release.
[0038] Example 5: Counterfactual Explanation and Minimal Intervention Recommendation In this embodiment, the closed-loop intervention module further includes a counterfactual interpretation submodule. Given the current patient state characteristic x and the predicted risk probability... This submodule answers the question: "If a modifiable characteristic (such as daily duration of bed rest, glucose intake, or frequency of bed changes) is altered, how will the risk of future complications change?" Specific implementation method: Incorporate the modifiable feature vectors... Separating from complete features involves training a differentiable causal model (such as a structural causal model or a counterfactual generative network) to target features. Intervention (set to) With other features fixed, recalculate the predicted probability. The output states, "After changing feature i to the target value, the risk is reduced." It also recommends a set of actions with minimal intervention cost (e.g., "reducing bed rest time by 2 hours per day can reduce the risk of leakage by 18%), presented in text or graphic form on the patient's terminal.
[0039] Example 6: Federated Learning-Driven Global Model Update In this embodiment, to protect patient privacy, the cloud center does not collect raw multimodal data or feature representations, but instead employs federated learning. Each edge computing node completes model training locally, calculates the gradient update of the model parameters, adds local differential privacy noise (such as a Laplace mechanism) to this gradient, and then uploads it to the federated learning aggregator in the cloud center. The aggregator performs a weighted average of the gradients from multiple nodes to obtain the global model update, which is then distributed back to each edge node. Throughout the entire process, the raw data never leaves the patient's local device, meeting medical data compliance requirements, while model performance continuously improves as the number of participating nodes increases.
[0040] Industrial Applicability: The system and method provided by this invention can be integrated into smart ostomy bags, wearable devices, and mobile terminals, making them suitable for hospitals, community nursing centers, and home settings. The various components of the system are already feasible: mature products are available for multispectral imaging modules, microfluidic sensors, wearable IMUs, edge computing chips (such as ESP32-S3 or Raspberry Pi Zero), and cloud service infrastructure. The effectiveness of temporal diffusion models and graph attention networks in image and temporal data domains has been verified by a large body of literature. Therefore, this invention possesses clear industrial applicability.
[0041] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
[0042] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.
[0043] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.
Claims
1. An artificial intelligence-based system for early prediction and intelligent assisted nursing of stoma complications, characterized in that, include: The multimodal data acquisition module is used to simultaneously acquire image data of the stoma and surrounding tissues, as well as physical or chemical parameter data of the excrement; An edge computing module, deployed at the patient's end, is used to extract temporal features from the multimodal data and generate a feature representation of the current state and an explanation of the contribution of the feature representation to each complication type through an interpretable heterogeneous data fusion model. The cloud center prediction module is used to receive the feature representation and predict the probability of at least one stoma complication occurring within a future preset time window based on a time series generation model. as well as The closed-loop intervention module is used to automatically generate and execute graded nursing instructions based on the probability and contribution interpretation. The graded nursing instructions include at least sensory guidance instructions for the patient or physical drive instructions for the stoma care device.
2. The system according to claim 1, characterized in that, The hierarchical nursing instructions generated by the closed-loop intervention module include: nursing operation guidance instructions superimposed on the display via augmented reality devices, instructions to release active substances through linkage with the intelligent ostomy bag, or instructions to automatically coordinate external nursing resources.
3. The system according to claim 1, characterized in that, The interpretable heterogeneous data fusion model in the edge computing module adopts a graph neural network based on an attention mechanism. The attention coefficients of this network are dynamically calculated based on the temporal correlation between the data of each modality and the pre-trained prior correlation matrix.
4. The system according to claim 1, characterized in that, The time-series generation model in the cloud center prediction module is a conditional diffusion model or a conditional generative adversarial network. This model uses historical multimodal fusion features as conditions to predict the feature representation of future complication events through reverse denoising or adversarial generation.
5. The system according to claim 1, characterized in that, The multimodal data acquisition module also includes wearable sensors for detecting the patient's vital signs or physical activity level, and the multimodal data further includes the vital signs or physical activity level data.
6. The system according to claim 1, characterized in that, The edge computing module employs a lightweight temporal feature extraction network derived from knowledge distillation, enabling the model size and inference power consumption to adapt to the resource constraints of edge devices.
7. The system according to claim 1, characterized in that, The cloud center prediction module also includes a federated learning aggregator, which aggregates model gradient updates uploaded by multiple edge computing modules to globally update the prediction model while protecting patient data from leaving the local machine.
8. The system according to claim 1, characterized in that, The closed-loop intervention module also includes a counterfactual interpretation submodule, which outputs the predicted change in complication risk after changing at least one interventionable feature and generates nursing recommendations with minimal intervention cost.
9. A method for early prediction and intelligent assisted nursing of stoma complications based on the system described in any one of claims 1 to 8, characterized in that, include: Simultaneously collect multimodal data from ostomy patients, including at least image data and physical or chemical parameter data of excrement; Temporal features are extracted from the multimodal data at the edge, and the current state feature representation and contribution explanation are generated through interpretable heterogeneous data fusion. The feature representation is sent to the cloud, and the probability of future complications is predicted based on a time-series generation model; Based on the probability and contribution interpretation, graded care instructions are generated and executed, which include at least sensory guidance instructions or physical drive instructions.
