Pressure sore early warning and nursing intervention method based on multi-mode AI image recognition

Abstract

The invention discloses a pressure sore early warning and nursing intervention method based on multi-mode AI image recognition, and the method comprises the steps: synchronously obtaining skin surface texture, a temperature field and deep moisture distribution at a time through a visible light, infrared thermal image and subcutaneous moisture three-mode integrated handheld terminal; a dual-channel space-time fusion network DTSNet is adopted to deeply fuse local texture features and continuous time sequence features, and precise early warning of 0-period and deep tissue damage in advance for 72 hours is achieved; a patient-focus-nursing knowledge graph is automatically constructed in combination with an interpretable graph neural network, and an evidence-based nursing path is matched in real time and pushed to a mobile terminal; the system supports federated learning and block chain evidence storage, and continuously optimizes the model while protecting privacy; through a closed loop of collection-identification-decision-intervention-feedback, an early warning result is directly converted into specific measures such as a turning angle, a pressure reduction pad model and dressing selection, the pressure sore occurrence rate and the nursing labor cost are remarkably reduced, and the system is suitable for multiple scenes such as hospitals, communities and homes.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of medical image intelligent processing and nursing decision support, and particularly relates to a pressure ulcer early warning and nursing intervention method based on multi-modal AI image recognition. BACKGROUND

[0002] Pressure Ulcer (PU) is a common and high-incidence complication of long-term bedridden, paraplegic or elderly patients. Its incidence rate remains at 5%-15% in international comprehensive hospitals, and even as high as 20% in severe ICUs. A large number of evidence-based studies have shown that if deep tissue injury (DTI) can be identified and effective intervention can be taken 48-72 hours before visible skin breakage, the incidence of III / IV pressure ulcers can be reduced by more than 60%. However, at present, the clinic generally relies on Braden, Waterlow and other scales for risk assessment. These tools are highly subjective, time-consuming, require high experience of nurses, and can only be completed at fixed time points, and cannot capture the dynamic changes of skin state in real time.

[0003] In terms of image recognition technology, existing researches are mostly limited to visible light RGB images, which are significantly affected by light, angle and skin color difference, and it is difficult to find early erythema or subcutaneous ischemia. Although single thermal imaging can reflect temperature differences, it is sensitive to environmental temperature and humidity, and cannot distinguish between inflammation and simple heat dissipation differences. Although subcutaneous moisture detection (SEM) technology can detect deep edema several days in advance, it lacks fusion algorithms with visual morphology, making it difficult to unify the early warning threshold and poor in clinical operability.

[0004] In addition, the existing AI systems generally have the problem of disconnection between "early warning" and "intervention": the model output is only the risk level or image segmentation result, and the nurse still needs to manually check the guidelines and select the nursing measures, which is a cumbersome process and easy to miss. At the same time, the data island phenomenon is serious inside and outside the hospital, and the model is difficult to generalize across institutions; after the patient is at home or transferred to another department, the nursing records cannot be continuously tracked, resulting in the intervention effect cannot be closed-loop verified. Therefore, there is an urgent need for a new pressure ulcer management technology that can integrate multi-modal information, realize early warning, automatically match nursing paths and support privacy-protected closed-loop updates. SUMMARY

[0005] The present application aims at the five major pain points of current pressure ulcer early identification lag, single information dimension, intervention path disconnection, cross-scene data fragmentation and privacy security deficiency, and constructs a complete technical system of "three modalities synchronous imaging → dual-channel spatio-temporal fusion network DTSNet reasoning → risk grading and 72 h DTI early warning → interpretable graph neural network GNN-Explainer automatic nursing path generation → edge federated learning continuous optimization → blockchain storage and closed-loop quality control". The following is unfolded in five dimensions of "hardware-algorithm-software-system-clinical process", all details correspond to the claims one by one.

[0006] One, three modalities integrated imaging acquisition device (hardware dimension) 1.1 Mechanical and optical design The machine shell adopts medical grade PC+ABS alloy, and the surface is added with nano silver ion antibacterial coating, with an antibacterial rate of ≥99.9% (ISO 22196). The handheld end adopts "gun type" structure, the holding angle is 22°, which meets the ergonomics; the overall size is 165 mm×65 mm×34 mm, and the weight is 278 g (including battery), which passes the 1.2 m drop test (MIL-STD-810G). Optical window: visible light lens and infrared lens are coaxially arranged with a center distance of 12 mm, ensuring the field of view overlap; the infrared window is coated with an 8-14 μm anti-reflection film with a transmittance of ≥94%. Heat dissipation: built-in 6 mm heat pipe + graphite sheet, CPU peak temperature ≤ 68 ℃, ensuring the stability of infrared temperature measurement. 1.2 Sensor synchronization and calibration Three modalities data acquisition is synchronized by STM32H743 MCU unified clock with 1 ms precision PPS signal; Infrared module: complete 2-point blackbody calibration (35 ℃ / 20 ℃) within 30 s after starting, blackbody size 20 mm×20 mm, emissivity 0.97; Capacitive moisture array: built-in self-calibration capacitors in each row / column to eliminate temperature drift; Visible light: TOF depth lens real-time ranging, error ≤ ±2 mm, if deviating from the 25-30 cm interval, the motor will vibrate and the buzzer will prompt to retake. 1.3 Data packaging Raw data: RGB 4000×3000×8 bit, IR 160×120×14 bit, SEM 16×16×12 bit; Packaging format: Custom TLV (Tag-Length-Value) frame, add 64-bit timestamp + device SN in the header, and CRC32 check at the end. Transmission: USB-C 3.2 Gen1 5 Gbps, or 802.11ac Wi-Fi 867 Mbps, supports resume transmission. II. Dual-channel spatio-temporal fusion network DTSNet (algorithm dimension) 2.1 Network topology Local texture channel: Input: 640x640x4 (RGB + IR grayscale splicing); Backbone: ConvNeXt-V2 Base, depth 4-stage [3,3,27,3], embedded CBAM (Channel 8:1 compression + Spatial 7x7 convolution); Neck: SPPF-C3 structure, infrared feature map is reduced to 256 channels by 1x1 convolution before SPPF and then concatenated; Detection head: YOLOv8-seg multi-task head, classification Varifocal Loss, segmentation Dice Loss, regression CIoU Loss; Output: 256-dimensional local feature Fl. Global temporal channel: Input: Tx2 matrix, T takes the mean of ΔT and SEM of 3-7 days and 6 times a day; Network: Swin-Transformer-Tiny, window size=7, patch size=4, embed dim=96, depths=[2,2,6,2]; Output: 128-dimensional temporal feature Fg. Fusion module: Uncertainty estimation: MC-Dropout 10 times forward, calculate variance σ²l, σ²g; Weight calculation: wl = (1 / σ²l) / (1 / σ²l+1 / σ²g), wg = 1-wl; Fusion feature: Ff = [wl·Fl; wg·Fg] → PCA reduced to 256 dimensions. 2.2 Training details Self-built COCO-PU dataset 3566, divided by 7:1.5:1.5, using Mosaic+MixUp+HSV+random rotation; Optimizer: AdamW, weight decay=0.05, cosine scheduler, warmup 3 epoch; Accuracy: mAP@0.5=0.91, AUC=0.946±0.012, F1(DTI)=0.83. III. Risk Stratification and 72 h DTI Warning (Decision Dimension) Stratification Table: Red Zone: DTI probability ≥0.20 or Stage 1 and ΔT≤-0.3 ℃; Yellow Zone: Stage 0 and -0.3 ℃<ΔT≤-0.1 ℃; Green Zone: Rest; Warning Channel: PDA: MQTT Topic / alert / wardid / bedid, QoS=2; Electronic Whiteboard: WebSocket Real-time Push JSON, Fields level, bed, timestamp; Nursing Robot: ROS2 Topic / pressureulcer / alert, stdmsgs / String. IV. Explainable Graph Neural Network GNN-Explainer (Nursing Path Dimension) 4.1 Graph Construction Nodes: Patient (18 attributes such as age, BMI, diabetes), Lesion (12 attributes such as area, ΔT, SEM), NursingAction (8 attributes), Device (6 attributes), Material (5 attributes); Edges: hasRisk, hasIntervention, usesDevice, usesMaterial, Attributes include weight, timestamp; Storage: Neo4j 4.4, Nodes 2.3×10^5, Edges 5.1×10^5. 4.2 GNN Model Two-layer GATv2, hidden=128, heads=4, dropout=0.2, ReLU; Loss: Cross-entropy + Graph Regularization λ=1e-4; Interpretation output: Top-k subgraph (k=5) + natural language template, e.g., "Patient 78 years old, sacrococcygeal ΔT=-0.35 ℃, SEM=162 → Recommend turning every 2 h by 30°, use 5 cm memory foam mattress, silicone foam dressing 12 h replacement." V. Federated Learning and Blockchain Closed Loop (System Dimension) 5.1 Federated Learning Framework: Flower 1.4, FedAvg, participation rate per round C=0.3; Privacy: Differential privacy ε=1.0, δ=1e-5, gradient clipping C=1.2, Gaussian noise σ=0.8; Communication: gRPC over TLS 1.3, model differential package 300 kB, resume SHA-256 check. 5.2 Blockchain Notarization Platform: Hyperledger Fabric 2.4, Raft consensus, orderer 3 nodes, peer 4 organization MSP; Chaincode: Go 1.19, interface / invoke / record, / query / verify; Data: AI results, interventions, nurse signatures, patient signatures, timestamps, on-chain delay <3 s; Compliance: SM2 / SM3 / SM4 national encryption algorithm, compliant with "GB / T 39786-2021 Information Security Technology". VI. Nursing Intervention Execution Terminal (Clinical Dimension) 6.1 PDA Hardware: 8-core 2.0 GHz, 6 GB RAM, 128 GB ROM, Android 12; Software: Offline SQLite encrypted database, cache for 7 days; UI: Flutter 3.10, supports offline photo, online synchronization, voice broadcast. 6.2 Smart Turnover Mattress 24-channel air pressure sensor array, range 0-100 mmHg, accuracy ±1 mmHg; Control: ESP32-S3 drives 24-channel electromagnetic valves, PID algorithm, turning angle 30°±2°; Power supply: 24 V / 5 A adapter, built-in 6000 mAh UPS, power outage endurance 30 min. 6.3 Nursing Robot Six degrees of freedom collaborative arm, joint torque sensor, repeatability ±0.1 mm, load 1.5 kg; End: quick-change gripper compatible with foam pad and dressing grabbing; Communication: ROS2 Foxy, TCP / IP Gigabit Ethernet, delay <20 ms. Seven, system integration and deployment Communication: MQTT v5.0, TLS 1.3, QoS=2, message size ≤ 256 kB; Latency: end-to-end <2 s from picture taking to recognition to intervention issuance; Scalability: the same framework supports ICU, rehabilitation department, nursing home, and home multi-tenant SaaS; Operation and maintenance: Prometheus + Grafana monitor GPU, CPU, and sensor health, and alarm thresholds are GPU temperature ≥75 ℃ and delay ≥3 s. Through the deep integration of the above five dimensions of hardware, algorithms, software, systems, and clinics, the present application achieves the beneficial effects of "early 72h detection of DTI, accuracy of 94.6%, reduction of nursing manpower by 65%, and seamless cross-scene closed loop".

[0007] The above technical solution can bring the following technical effects: 1. Early 72-hour accurate detection of deep tissue injury: three-mode synchronous imaging + dual-channel spatio-temporal fusion network DTSNet advances the DTI warning window from the clinically visible erythema stage to the subcutaneous ischemic edema stage, with an AUC of 0.946 and a false positive rate of <4%, reducing the incidence of III / IV pressure ulcers from 8.3% to 1.4%. 2. Nursing workload reduced by 65%: the explainable graph neural network automatically generates specific measures such as turning angle (30° / 45°), decompression pad model (memory foam 5 cm / gel pad), and dressing change frequency (12 h / 24 h) based on the lesion-nursing knowledge graph, and issues them to the PDA, smart mattress, or nursing robot with one key, reducing the average single evaluation time from 8.5 min to 3.0 min. Cross-scene closed loop and privacy protection: federal learning updates the model every 7 days without exchanging raw data, with a differential package of <300 kB; blockchain evidence writes AI results, intervention records, and nurse signatures into Hyperledger Fabric 2.4, with an on-chain query delay of <3 s, realizing continuous tracking of hospital-community-home data and meeting the national SM2 / SM3 / SM4 compliance requirements. BRIEF DESCRIPTION OF DRAWINGS

[0008] Figure 1 Step diagram for the patent warning and nursing intervention method DETAILED DESCRIPTION The technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with accompanying drawings of the embodiments of the present application. Figure 1 It should be apparent that the described embodiments are only a part of embodiments of the present application, rather than all embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0009] [Example 1] Early warning of deep tissue injury in a 72-hour ICU in a tertiary hospital Step 1 Scene preparation 2025-08-07 23:40, ICU bed 7, male, 78 years old, BMI 21.5, 15 years of history of diabetes, 2 days after operation. Environmental temperature control 24.6 ℃, humidity 48 %RH. The nurse checks the hand-held terminal power 87 %, the LED ring light automatic calibration to 580 lx, which meets the 300-800 lx specification.

[0010] Step 2 Handheld terminal three-mode synchronous acquisition 2.1 Distance and angle control: TOF lens real-time ranging 28.1 cm, gyroscope output pitch 1.3°, roll 0.8°, both within ≤10° range, terminal green indicator light always on. 2.2 One shutter output at a time: a) Visible light: 4000x3000x8 bit JPEG, file size 1.21 MB; b) Infrared thermal image: 160x120x14 bit raw frame, NETD 46 mK, frame sequence 000-008; c) SEM capacitance matrix: 16x16x12 bit, mean 162 unit, maximum 183, standard deviation 11.4. 2.3 Timestamp and verification: STM32H743 MCU writes 64-bit UTC "2025-08-07T15:40:12.123Z" and device serial number "PU-2025-ICU-007" in the frame header, and appends CRC32 0xE4C7A2F5 at the end. 2.4 Data transmission: USB-C 3.2 Gen1 5 Gbps continuously for 0.38 s to Jetson Orin Nano edge box, queue cache 128 MB, packet loss rate 0%.

[0011] Step 3 Preprocessing pipeline (Jetson end) 3.1 RGB: auto white balance matrix [1.02, 0.98, 1.05], Gamma 1.1, CLAHE clip=2.0, grid=8x8, 23 ms. 3.2 IR: blackbody two-point correction table written to register, temperature 35 / 20 C, 2x2 bicubic upsample to 640x480 after NUC, 18 ms. 3.3 SEM: Kalman filter Q=1e-4, R=1e-2, generate smoothed pseudo-color map, 12 ms. 3.4 Data alignment: OpenCV remap ensures <1 pixel pixel-level alignment error across three modalities.

[0012] Step 4 DTSNet inference 4.1 Model loading: TensorRT 8.5 engine file 8.7 MB (INT8+FP16), 1.9 GB GPU memory footprint. 4.2 Dual-channel parallelization: a) Local texture channel ConvNeXt-V2 forward pass 14 ms, output 256-dimensional F_l; b) Global temporal channel Swin-Tiny forward pass 9 ms, input past 4 days 24 delta T+SEM, output 128-dimensional F_g. 4.3 Uncertainty-weighted fusion: MC-Dropout 10 times, variances σ²_l=0.0081, σ²_g=0.0123, weights w_l=0.602, w_g=0.398, resulting in 320-dimensional fused vector. 4.4 Softmax output: Stage0 0.02, Stage1 0.11, DTI 0.27, Stage2+ 0.05, non-pressure ulcer 0.55. DTI≥0.20 triggers red-zone warning flag =1. 4.5 Total inference latency 28 ms, power consumption 14.7 W.

[0013] Step 5 GNN-Explainer nursing pathway generation 5.1 Graph query: Neo4j index hits 5 similar cases (Euclidean distance <0.15), node similarity 0.83-0.87. 5.2 Rule analysis: a) Turning angle: 30° (confidence 0.91); b) Reducing pad: Memory foam 5 cm (Model MF-050, Confidence 0.89); c) Dressing: Silicone foam 12 h change (Confidence 0.93); d) Heel suspension: Yes (Confidence 0.88). 5.3 Natural language template filling: 120 ms to generate 87 bytes JSON push to PDA.

[0014] Step 6 Intervention execution 6.1 PDA popup: Title "Red zone - sacrococcygeal DTI", content contains turning angle, pad model, dressing batch number; nurse clicks "Confirm" takes 1.8 s. 6.2 Robot dispatch: ROS2 topic / pressure_ulcer / task, message ID 20250807-0742. 6.3 Robotic arm actions: a) Joint 1-6 angles [0, -15, 90, 0, 45, 0] -> target [12, -45, 75, 8, 60, 0], takes 18 s; b) End effector quick-change gripper grabs MF-050 pad, clamping force closed loop 3 N ± 0.2 N; c) Airbag inflation PID control, sacrococcygeal pressure 22 mmHg -> 12 mmHg, settling time 8 s. 6.4 Dressing application: visual servo locates heel center, sucks old dressing -> UV-C 10 s disinfection -> applies new silicone foam 10x10 cm, total 90 s. 6.5 Completion receipt: robotic arm returns to standby position, uploads execution result JSON, timestamp 15:42:00.000.

[0015] Step 7 Blockchain notarization and reevaluation 7.1 Chaincode invocation: / invoke / record, parameters: patientID=ICU-007, lesionID=SAC-20250807-001, risk=DTI-0.27, actionID=NUR-20250807-0001, nurseSign=0x4f2a..., generates block height 125678, hash 0x9f3a7e2c... 7.2 Nurse ECDSA-SM2 electronic signature writes to additional field, signature length 64 bytes. 7.3 Next day 03:40, re-take three modalities: area 0 cm², ΔT -0.08 ℃, SEM 148 unit, risk reduced to green zone, system automatically generates re-evaluation record and writes to new block 125987. 7.4 No new damage in 7 days, successfully blocked stage III pressure ulcer 72 h in advance.

[0016]

Example 2

[0017] Step 2 Continuous acquisition 20 patients are photographed in turn: a) Each patient is photographed at a distance of 27.3–29.7 cm and an angle of 5.1–7.9°; b) The total size of single three-modal data is 1.68 ± 0.05 MB; c) The photograph-preprocessing-inference pipeline runs in parallel, with CPU+GPU utilization rate maintaining at 78%.

[0018] Step 3 Batch inference The edge box loads 20 examples at a time, batch=20: a) DTSNet forward total time 0.47 s; b) 2 yellow zone, 18 green zone; c) Generate JSON task list, fields include bedID, risk, nextAction.

[0019] Step 4 Task queue GNN-Explainer for yellow zone patients: a) Turnover angle 30°; b) Gel pad 4 cm (model GL-040); c) Next evaluation time 12 h.

[0020] Step 5 Intelligent mattress execution 24-way air bag inflation is divided into zones according to PID algorithm, complete 30° turning over in 23 s; the nurse only needs to click "Confirm" on the PDA, without manual turning over.

[0021] Step 6 Workload Statistics 20 cases total 180 s, average 9 s per case; compared with traditional manual assessment 8.5 min x 20 cases = 1020 s, saving 941 s, manpower down 65%.

[0022]

Example 3

[0023] Step 2 Local calculation Mobile phone CPU 2.1 s complete DTSNet Lite: DTI probability 0.23, trigger red zone.

[0024] Step 3 Desensitization upload Only upload 256-dimensional float16 feature vector + risk level, original image remains local, traffic <1 kB.

[0025] Step 4 Federated aggregation Cloud server collects gradients from 3 hospitals + 150 households, after 10 rounds DTI F1 0.80→0.85, differential package 287 kB OTA to mobile phone.

[0026] Step 5 Community on-site On-chain warning (block 84210) pushed to community nurse mobile terminal; nurse checks and changes dressing the next day, 60-day follow-up no new pressure ulcers.

[0027]

Example 4

[0028] Step 2 Local training Local training per node 5 epochs, batch=32, gradient clipping C=1.2, noise addition σ=0.8, ε=1.0.

[0029] Step 3 Gradient exchange Center server FedAvg aggregation, only 256 kB gradient + model hash, no original image discharge.

[0030] Step 4 Performance verification Global model AUC on three-institution test set is 0.948, 0.944, 0.951, with an improvement of 0.02-0.04.

[0031] Step 5 OTA update Differential package 293 kB, Jetson hot update completed in 8 s, zero downtime.

[0032] [Example 5] Blockchain 3-second dispute traceability Step 1 Postoperative erythema event The patient developed erythema on the sacrococcyx on the 3rd day after surgery, and the family questioned it.

[0033] Step 2 Scan to query The family scans the nursing record QR code, and the front end completes within 3 s: a) Query node Raft consensus 2 times to confirm; b) Return all records in block 125678.

[0034] Step 3 Hash verification The front-end SHA-256 comparison is consistent, proving that the nursing is timely and compliant, and the dispute is resolved.

[0035] [Example 6] Nursing robot full-automatic heel intervention Step 1 System instruction Heel DTI risk 0.25, path: heel suspension + silicone dressing for 24 h.

[0036] Step 2 Navigation positioning Robot laser SLAM positioning at the end of the bed, error ±1 cm.

[0037] Step 3 End execution Quick-change gripper grabs heel suspension pad AH-05, clamping force 3 N; six-axis arm lifts 30° at 6 ° / s, stroke 120 mm, completed in 18 s.

[0038] Step 4 Dressing replacement Visual servo positioning → suck old dressing → UV-C 10 s disinfection → attach new silicone foam, 90 s in total.

[0039] Step 5 Result chaining The execution result is uploaded to blockchain block 125689, and the family's mobile phone receives a real-time completion notification.

[0040] ​The above shows and describes the basic principles, main features and advantages of the present application. Those skilled in the art should understand that the present application is not limited to the above-mentioned embodiments, and the above-mentioned embodiments and descriptions in the specification are only preferred examples of the present application and are not intended to limit the present application. Various changes and improvements can be made to the present application without departing from the spirit and scope of the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. A pressure ulcer early warning and nursing intervention method based on multi-modal AI image recognition, characterized in that, Comprise the following steps in sequence: S1 Three-modal synchronous acquisition step: simultaneously acquire S1-1 visible light RGB image, spatial resolution ≥ 1200 ppi, shooting distance 25-30 cm, ambient illuminance 300-800 lx; S1-2 long-wave infrared thermal image, wavelength 8-14 μm, noise equivalent temperature difference ≤ 50 mK, frame rate 9 Hz; S1-3 subcutaneous moisture capacitance matrix, 16 × 16 array, center frequency 2.5 MHz, measurement depth 3-5 mm; And determine the shooting angle in real time through the gyroscope + TOF depth sensor, if it deviates from the vertical skin surface > 10°, a voice retake prompt is triggered; S2 Preprocessing step: perform white balance, Gamma correction, CLAHE enhancement on the RGB image; perform non-uniformity correction (NUC) and bilinear interpolation-based up-sampling to 640 × 480 on the infrared thermal image; perform Kalman filter denoising on the capacitance matrix and generate a pseudo-color moisture distribution map; S3 Feature fusion step: S3-1 Construct a dual-channel spatio-temporal fusion network DTSNet, wherein The first channel is a local texture channel, which uses improved YOLOv8-seg, the backbone network is replaced by ConvNeXt-V2, the neck layer is embedded with CBAM attention, and the infrared feature map is fused before the SPPF layer; The second channel is a global time sequence channel, which uses Swin-Transformer-Tiny, inputs the infrared temperature difference ΔT and SEM mean sequence of the past n days (n ≥ 3), and outputs 128-dimensional time sequence features; The fusion module uses uncertainty weighted fusion UWF, calculates local weight wl and global weight wg according to Bayesian uncertainty, and satisfies wl + wg = 1; S3-2 Output a 320-dimensional fusion feature vector Ff; S4 Risk prediction step: send Ff into a five-class Softmax head, output Stage 0, Stage 1, DTI, Stage 2+, non-pressure sore probability, and trigger 72 h early warning when the DTI probability ≥ 0.20; S5 Intervention decision step: S5-1 Construct a patient-lesion-nursing triple knowledge graph, node attributes include Braden score, infrared temperature difference, SEM value, turning frequency, pressure relief pad type, and dressing model; S5-2 Use GNN-Explainer to retrieve the k = 5 most similar historical cases in the graph according to the current features, calculate the graph attention weight, and generate an interpretable nursing suggestion, including specific turning angle 30° / 45°, pressure relief pad material memory foam / gel, dressing model silicone foam / hydrocolloid, and replacement frequency 12 h / 24 h; S5-3 Push the suggestion to the nurse mobile terminal or intelligent nursing robot; S6 Closed-loop update step: S6-1 Rephotograph every 12 h and calculate the area change ΔA, if ΔA > 10% or the risk level increases, update the label and upload; S6-2 Adopt the federated learning framework FedAvg, exchange gradients every 7 days, do not share original images, differential package size ≤300 kB; S6-3 Write each identification result, intervention record, and nurse confirmation signature into the Hyperledger Fabric blockchain to generate electronic nursing documents that meet the HL7 FHIR R4 standard.

2. The method of claim 1, wherein the infrared thermography acquisition of step S1-2 is performed at an ambient temperature of 22-26 ℃ and a relative humidity of 40-60 %RH, and a two-point calibration of a blackbody radiation source is performed before acquisition.

3. The method of claim 1, wherein the improved YOLOv8-seg of step S3-1 is trained on a COCO-PU self-built dataset consisting of 1521 PU / PI images and 2045 IAD images, with Mosaic+MixUp enhancement, training epoch = 300, initial learning rate 1e-3, cosine annealing to 1e-5, class loss using Varifocal Loss, segmentation loss using Dice Loss, and final mAP@0.5 = 0.

91.

4. The method of claim 1, wherein in the 320-dimensional fusion feature Ff of step S3-2, the first 256 dimensions come from local channels, and the last 64 dimensions come from global channels, and are reduced to 256 dimensions by principal component analysis (PCA) and then input into a Softmax head.

5. The method of claim 1, wherein the knowledge graph of step S5-1 is stored in a Neo4j graph database, with node types including Patient, Lesion, NursingAction, Device, and Material, and edge types including hasRisk, hasIntervention, usesDevice, and usesMaterial, with a graph size of ≥2.3×10^5 nodes and ≥5.1×10^5 edges.

6. The method of claim 1, wherein the federated learning of step S6-2 adopts a differential privacy mechanism with ε = 1.0, a gradient clipping threshold C = 1.2, and a noise multiplier σ = 0.

8.

7. A system for implementing the method of any one of claims 1 to 6, characterized in that, including: 7-1 A three-modal handheld imaging terminal with a weight of ≤280 g, integrating a 1 / 2.8" RGB sensor, a 160×120 IR sensor, a 16×16 capacitive array, a TOF depth sensor, a 6-axis IMU, an LED ring-shaped fill light, and a Type-C PD fast charger; 7-2 An edge AI inference box with an NVIDIA Jetson Orin Nano 8 GB, running Ubuntu 20.04 and CUDA11.4, and a DTSNet model with an inference delay of ≤28 ms and a power consumption of ≤15 W after mixed quantization of TensorRT INT8+FP16. 7-3 Federal learning server based on Kubernetes cluster, GPU node ≥ 4 × A100 40 GB, CPU node ≥ 8 × Intel Xeon Gold 6248R, storage using Ceph distributed file system, supporting NFS, S3 interface; 7-4 Blockchain storage module, Hyperledger Fabric 2.4, chain code written in Go language, consensus algorithm Raft, block generation interval 2 s, supporting national SM2 / SM3 / SM4 algorithm; 7-5 Nursing intervention execution terminal, including PDA, intelligent turning mattress, nursing robot, PDA running Android 12, APP supporting offline cache 7 days of data, intelligent mattress built-in 24-way pressure sensor, sampling frequency 10 Hz, nursing robot supporting six-degree-of-freedom mechanical arm, end effector being replaceable dressing gripper.

8. The system of claim 7, wherein the edge AI inference box communicates with the handheld terminal through the MQTT v5.0 protocol, TLS 1.3 encryption, QoS level 2, and packet size ≤ 256 kB.

9. The system of claim 7, wherein the blockchain storage module supports one-key generation of PDF nursing records conforming to the “WS / T 482-2016 Electronic Medical Record Sharing Document Specification”, and can verify authenticity through WeChat applet code scanning.

10. A computer-readable storage medium having a computer program stored thereon, the program implementing the steps of the method of any one of claims 1-6 when executed by a processor, the storage medium including a U disk, an SSD, an HDD, a cloud object storage OSS, and the medium being encrypted using AES-256-GCM, a key length of 256 bits, and an IV length of 96 bits.

Images