Flexible periorbital pupil real-time monitoring method and system based on near-infrared eyelid penetration

Abstract

The invention discloses a flexible periorbital pupil real-time monitoring method and system based on near-infrared eyelid penetration. The system comprises the flexible intelligent eye pad and a central processing platform. The eye pad adopts a hollow annular design, is attached to the skin of an eye socket through a medical pressure-sensitive adhesive and is prevented from being in direct contact with an eyeball, and a near-infrared light source, an image sensor, a microprocessor, a wireless communication module and a battery are integrated in the eye pad. The method comprises the steps that near-infrared light penetrates through eyelids to image pupils in an eye closing state; the eye opening / closing state is automatically identified, and the monitoring mode is seamlessly switched; the image sequence is analyzed in real time, and physiological indexes such as pupil diameter, symmetry and light reflex parameters are extracted; constructing a parameter trend chart, and identifying an abnormal mode by using a time sequence algorithm; and performing hierarchical early warning based on fusion judgment of a threshold rule and a machine learning model. According to the invention, continuous, non-invasive and real-time monitoring of pupils of critical patients is realized, and the problems of monitoring blind areas, strong subjectivity and incapability of monitoring eye-closed patients in the prior art are solved.

Description

Technical Field

[0001] This invention relates to the fields of medical devices and medical information technology, specifically to a method and system for real-time monitoring of the flexible periorbital pupil based on near-infrared penetrating eyelids. Background Technology

[0002] Pupil diameter, symmetry, and pupillary light reflex are key indicators for assessing central nervous system function, particularly the function of the midbrain and oculomotor nerves. In acute neurological injuries, such as intracranial hemorrhage, stroke, and post-neurosurgery and intensive care, pupillary changes are among the earliest and most sensitive signs of increased intracranial pressure (ICP) and brain herniation. Timely detection of abnormal pupillary changes is crucial for saving lives and improving prognosis. Currently, the standard clinical method for pupillary monitoring relies on intermittent manual observation and measurement by healthcare professionals using a handheld pupillary ruler and pen-style flashlight. This method has several inherent limitations:

[0003] 1. Monitoring blind spots: 24-hour continuous monitoring is not possible, and critical pupil change thresholds may be missed during measurement intervals, leading to delays in diagnosis and rescue.

[0004] 2. Highly subjective: The measurement results heavily depend on the observer's experience, judgment, and ambient lighting, lacking objective and quantifiable data, and there may be significant differences between different observers.

[0005] 3. Heavy workload: Frequent timed measurements take up a lot of medical staff's working time and increase their workload.

[0006] 4. Ineffective for patients with closed eyes: Traditional methods are difficult or impossible to implement for patients who are unable to open their eyes due to eyelid edema, sedation, coma, or the use of muscle relaxants. Forcibly opening the eyelids for measurement is not only difficult to perform, but may also cause corneal damage, increased intraocular pressure, or patient agitation, and continuous monitoring cannot be achieved.

[0007] To overcome the limitations of manual measurements, some automated devices, such as handheld automatic pupillometers, have emerged on the market. These devices can provide quantified pupillary parameters, improving the objectivity of a single measurement. However, they are essentially still intermittent measurement tools and do not solve the core problem of "monitoring blind spots." Furthermore, they still require the patient to open their eyes during use, limiting their practicality for agitated or uncooperative ICU patients.

[0008] Other studies have attempted to use cameras fixed to the headboard or headrest for continuous pupil monitoring. However, these methods typically require patients to maintain a specific position and keep their eyes open, which is unreliable in the noisy and dynamic environment of a real ICU and also cannot handle patients with closed eyes. Currently, there are some related patents, such as eye-covering monitoring devices or devices that require opening the eyelids. However, these solutions often have problems such as poor comfort, potential pressure on the eyeball, interference with the patient's natural physiological state (e.g., hindering eyelid closure), or risks of infection (e.g., contact with the cornea).

[0009] Therefore, there is an urgent need in this field for a revolutionary technology and product that can truly achieve safe, non-intrusive, continuous, and real-time pupil monitoring of patients with open and closed eyes, in order to fill a key gap in existing clinical practice and provide a reliable safety barrier for the monitoring of critically ill neurological patients. Summary of the Invention

[0010] The technical problem to be solved by the present invention is to provide a flexible periorbital pupil real-time monitoring method and system based on near-infrared penetrating eyelids, so as to solve the problem that the existing technology cannot perform continuous, non-invasive, real-time pupil monitoring of patients in a naturally closed eye state.

[0011] To address the aforementioned technical problems, embodiments of the present invention provide the following technical solution: a flexible periorbital pupil real-time monitoring method based on near-infrared penetrating eyelids, comprising the following steps:

[0012] The flexible smart eye patch device is worn on the skin around both eyes of the patient. The eye patch device has a hollow structure to ensure no contact with the eyeball and is fixed by built-in medical pressure-sensitive adhesive. After the system is powered on, it completes self-test and establishes a wireless communication connection with an external terminal.

[0013] The eye patch device emits low-intensity near-infrared light with a wavelength of 700-1000nm through a near-infrared light source module. At the same time, the eye patch continuously collects images through a near-infrared image sensor module integrated in the same eye patch. The images are direct images when the patient's eyes are open and transmitted images after the near-infrared light penetrates the eyelids when the eyes are closed.

[0014] The acquired image sequences are analyzed in real time by a microprocessor or a remote central processing platform, automatically identifying whether the patient's eyes are open or closed, and seamlessly switching between direct viewing mode and near-infrared penetration mode accordingly to ensure the continuity and accuracy of monitoring data under different physiological conditions.

[0015] On the central processing platform, an image processing algorithm is run to extract one or more pupil physiological parameters from the image sequence;

[0016] The extracted pupil physiological parameters are stored in chronological order, and a time series trend graph is constructed. The trend graph is analyzed using a time series data analysis algorithm to identify abnormal change patterns in the pupil parameters.

[0017] Based on preset threshold rules or the output of machine learning models, the abnormal change patterns are judged; when the first-level alarm conditions are met, the highest-level emergency alarm is immediately triggered; when the second-level alarm conditions are met, an early warning alarm is triggered; and the alarm information is pushed to medical staff through the system interface and mobile terminals.

[0018] The pupil parameter data, trend charts, and alarm event records generated during the monitoring process are automatically integrated into the hospital's electronic medical record system, and monitoring reports are generated for clinical review and auditing.

[0019] Preferably, the pupillary physiological parameters include pupil diameter, pupil area, left and right pupil symmetry index, pupillary light reflex latency, pupillary light reflex contraction speed, pupillary light reflex contraction amplitude, and neuro-pupil index.

[0020] Preferably, the step of performing real-time analysis of the acquired image sequence to automatically identify the patient's open or closed eye state specifically includes:

[0021] The acquired image sequence is preprocessed and the region of interest is standardized. The original near-infrared image is filtered for noise reduction and contrast enhancement. The standardized eye region is located and cropped.

[0022] Multi-dimensional feature extraction and fusion analysis were performed to extract texture features and grayscale statistical features in parallel from standardized images, as well as the uniform texture and low contrast characteristics of eyelid skin in the closed state.

[0023] State decisions based on rules or deep learning classifiers input the extracted features into a preset threshold rule model or a pre-trained lightweight convolutional neural network classifier, and output the probability judgment of whether the eyes are open or closed.

[0024] Preferably, in the image preprocessing and region of interest normalization steps, a cascaded classifier based on Haar features or a semantic segmentation network based on U-Net is used to quickly locate and segment the eyeball or eyelid region.

[0025] Preferably, the image processing algorithm is used to extract one or more pupil physiological parameters from the image sequence, specifically including:

[0026] For near-infrared transmission images with eyes closed, an image segmentation algorithm based on region growing or level set evolution is used, combined with the extremely low gray value of the pupil region relative to the iris region, to segment out the pupil boundary.

[0027] For a direct view image with eyes open, an algorithm based on circular Hough transform or edge detection combined with ellipse fitting is used to extract the pupil boundary.

[0028] Based on the segmented pupil boundaries, the pupil diameter and area are calculated; the pupil symmetry index is calculated by comparing the left and right eye image sequences.

[0029] When the near-infrared light source module emits standardized light pulse stimulation, the dynamic changes in pupil size in the image sequence before and after stimulation are analyzed, the light reflex latency, contraction speed, and contraction amplitude are calculated, and the neural pupillary index is calculated by combining these dynamic parameters.

[0030] Preferably, the analysis of the trend graph using time-series data analysis algorithms includes using a sliding window-based statistical process control method, a change point detection algorithm, or a long short-term memory network model to identify gradual or abrupt abnormal trends in pupil parameters.

[0031] Preferably, the step of judging the abnormal change pattern based on a preset threshold rule or the output of a machine learning model specifically includes:

[0032] Construct a comprehensive feature vector that includes instantaneous values ​​of pupil parameters, dynamic trend statistics, and time series model output;

[0033] Parallel execution of logical judgments based on preset threshold rules and probabilistic inferences based on machine learning classification models;

[0034] A decision fusion strategy is adopted to integrate the rule judgment results and model inference results to generate the final alarm level judgment.

[0035] Preferably, the first-level alarm condition is that the diameter of one pupil increases sharply beyond a first threshold within a preset time window and is accompanied by the disappearance of light reflection; the second-level alarm condition is that the bilateral pupil asymmetry index is continuously higher than a second threshold for a preset duration, and / or the moving average of the light reflection velocity is continuously lower than a third threshold for a preset duration.

[0036] This invention also proposes a flexible periorbital pupil real-time monitoring system based on near-infrared penetrating eyelids for implementing the above method, comprising:

[0037] At least two flexible smart eye patch devices are used to be worn around the patient's left and right eye sockets, respectively;

[0038] A central processing platform is wirelessly connected to the flexible smart eye patch device;

[0039] The flexible smart eye patch device includes:

[0040] The flexible patch body is ring-shaped with a hollow area in the middle to expose the eyeball. Its bottom surface is coated with medical pressure-sensitive adhesive for adhesion to the skin of the eye socket.

[0041] A near-infrared light source module is embedded in the patch body and is used to emit near-infrared light with a wavelength of 700-1000nm;

[0042] A near-infrared image sensor module is embedded in the patch body for acquiring images of the eye;

[0043] The microprocessor and control module, embedded in the patch body, is used to control the switching and pulses of the near-infrared light source and to perform preliminary processing of image data.

[0044] A wireless communication module, embedded in the patch body, is used to interact with the central processing platform for data exchange.

[0045] The power module is embedded in the patch body and supplies power to the various electronic modules inside the eye patch.

[0046] The central processing platform includes:

[0047] The data receiving unit is used to receive image data or processed data from the flexible smart eye patch device;

[0048] The state recognition and mode switching unit is used to automatically recognize the open / closed eye state based on image data and control the switching of monitoring modes.

[0049] The pupil parameter analysis unit is used to run image processing algorithms to extract pupil physiological parameters from image sequences.

[0050] The trend analysis and early warning unit is used to store pupil parameters, construct trend graphs, analyze abnormal patterns, and trigger graded alarms.

[0051] The data integration and reporting unit is used to integrate data and alarm records into the electronic medical record system and generate monitoring reports.

[0052] Preferably, the flexible smart eye patch device also integrates a motion sensor for monitoring eye movements; the central processing platform also uses the motion sensor data to assist in assessing the state of neural function.

[0053] The beneficial effects of the above-described technical solution of the present invention are as follows:

[0054] 1. This invention solves the problem of "monitoring blind spots" in clinical pupil monitoring. Through continuous wear of a flexible smart eye patch, it achieves 24-hour uninterrupted real-time monitoring, constructing a reliable neural function safety net. It can capture instantaneous changes in the pupil, gaining valuable time for rescue. Simultaneously, its non-intrusive design avoids interference with patients from frequent operations, significantly improving the quality of intensive care and freeing medical staff from the heavy workload of timed measurements.

[0055] 2. This invention utilizes the penetrating properties of near-infrared light into eyelid tissue to achieve high-quality pupil imaging even when the patient's eyes are naturally closed. This makes continuous monitoring of sedated, comatose, or eyelid-edema patients a reality, eliminating the need to forcibly open the eyelids, perfectly conforming to physiological conditions and being highly humane. Simultaneously, its hollow, non-contact physical design fundamentally eliminates the risk of corneal damage and intraocular infection.

[0056] 3. This invention utilizes advanced image processing algorithms to output objective, quantified, multi-dimensional pupil parameters, eliminating subjective errors inherent in manual measurement. Simultaneously, based on trend analysis of time-series data and an intelligent early warning mechanism, it can identify early abnormal changes that are difficult for the human eye to detect, enabling proactive warnings and allowing clinical intervention to commence before the crisis fully manifests. Attached Figure Description

[0057] Figure 1 This is a flowchart of the flexible periorbital pupil real-time monitoring method based on near-infrared penetrating eyelids according to the present invention;

[0058] Figure 2 This is a schematic diagram of the flexible smart eye patch structure of the flexible periorbital pupil real-time monitoring system based on near-infrared penetrating eyelids according to the present invention.

[0059] Figure 3 This diagram shows the functional modules of the central processing platform of the flexible periorbital pupil real-time monitoring system based on near-infrared penetrating eyelids according to the present invention. Detailed Implementation

[0060] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0061] like Figure 1 As shown, this invention provides a flexible periorbital pupil real-time monitoring method based on near-infrared penetrating eyelids. The method includes the following steps:

[0062] Step S100: Device Wearing and Initialization. The flexible smart eye patch device is worn on the skin around both eye sockets. The eye patch device has a hollow structure to ensure no contact with the eyeballs during wear. The eye patch is comfortably and firmly fixed to the skin around the eye sockets via medical pressure-sensitive adhesive at its base. After the system is powered on, the built-in module of the eye patch completes a self-test and establishes a stable communication connection with an external central processing platform (such as a bedside monitor, central station, or mobile terminal) via a wireless communication module (such as Bluetooth, Zigbee, or Wi-Fi).

[0063] Step S200: Near-infrared imaging and image acquisition. A near-infrared light source module integrated within the eye patch emits low-intensity near-infrared light with a wavelength of 700-1000nm to illuminate the eye area. This wavelength of light has good penetrability to human tissues (especially the eyelids) and is safe for the eyes. Simultaneously, a near-infrared image sensor module (such as a CMOS or CCD sensor) integrated within the same eye patch continuously acquires images of the eye. When the patient's eyes are open, the sensor acquires an image of the eye under direct vision; when the patient's eyes are closed, the near-infrared light penetrates the eyelid tissue, and the sensor acquires a transmitted image of the pupil area behind the eyelid.

[0064] Step S300: Status Recognition and Mode Switching. The acquired image sequence is analyzed in real time by a microprocessor (located within the eye patch) or a remote central processing platform to automatically identify whether the patient's eyes are open or closed. Based on the recognition result, the system seamlessly switches between "high-precision direct-view mode" and "near-infrared penetration mode." In direct-view mode, natural light or ambient light can be fully utilized for more accurate pupil measurement; in penetration mode, imaging relies entirely on near-infrared light. This intelligent switching ensures the continuity and accuracy of monitoring data under different physiological states.

[0065] Step S400: Pupil Parameter Extraction. On the central processing platform, advanced image processing algorithms are run to precisely extract one or more key pupil physiological parameters from the image sequence acquired and pattern-recognized in step S200. These parameters include, but are not limited to: pupil diameter, pupil area, left and right pupil symmetry indicators (such as size difference ratio), pupillary light reflex latency (the time from light stimulation to the onset of pupillary constriction), pupillary light reflex constriction velocity, pupillary light reflex constriction amplitude, and a comprehensive neuro-pupil index.

[0066] Step S500: Trend Analysis and Anomaly Pattern Recognition. The extracted pupil physiological parameters are stored in a database in chronological order, and time-series trend graphs are constructed (e.g., pupil diameter variation curve over time). These trend graphs are analyzed using time-series data analysis algorithms (such as sliding window statistics, change point detection, or LSTM-based time-series models) to identify abnormal change patterns in pupil parameters, such as progressive pupil dilation, increased asymmetry, or decreased light reflex.

[0067] Step S600: Tiered Early Warning and Information Push. Based on preset threshold rules or the output of a machine learning model trained on clinical data, the abnormal change patterns identified in step S500 are judged. Specifically:

[0068] Step S601: Construct the input feature vector

[0069] The features representing abnormal pupil parameter variation patterns output by the time-series data analysis unit are organized into a structured feature vector. This vector includes:

[0070] Static parameters: instantaneous values ​​of parameters such as pupil diameter, difference between left and right pupil diameters, and light reflection velocity at the current moment.

[0071] Dynamic trend parameters: statistics calculated based on a sliding window, such as the moving average, trend slope, and standard deviation of pupil diameter over the past 10 minutes; as well as the magnitude and confidence level of change output by the change point detection algorithm.

[0072] Temporal model output: If models such as LSTM are used, it includes the state vector output by the hidden layer of the model or the predicted value of the pupil state in the next stage.

[0073] This feature vector integrates the instantaneous state of the parameters, short-term historical trends, and model prediction information, providing a comprehensive data foundation for subsequent judgments.

[0074] Step S602: Parallel execution of threshold rule judgment and machine learning model inference

[0075] The system executes two decision paths in parallel to improve the comprehensiveness and robustness of the response:

[0076] Path A: Threshold rule judgment

[0077] The specific parameters in the feature vector constructed in step S601 are compared with a preset, configurable threshold rule base. The rule base includes at least:

[0078] Level 1 Alarm Rule: For example, if (instantaneous pupil diameter on one side > baseline value + 2mm) and (instantaneous light reflection velocity on that side = 0), then a Level 1 alarm is triggered.

[0079] Level 2 Alarm Rules: For example, a Level 2 alarm is triggered if (the moving average of the difference between the left and right pupil diameters > 0.5 mm for 10 minutes) or (the moving average of the light reflection velocity < 30% of the baseline velocity for 15 minutes).

[0080] Path B: Machine Learning Model Inference

[0081] The same feature vector is input into a pre-trained multi-class machine learning model (such as a gradient boosting decision tree or support vector machine). This model is trained on a large amount of clinical data (including normal, early abnormal, and critical abnormal states) and can output the probability of belonging to "normal state", "secondary alert state", or "primary alert state".

[0082] Step S603: Generate final alarm decision based on decision fusion strategy

[0083] By combining the results from both paths, a final, unified alarm decision is generated:

[0084] High confidence trigger: If both path A (threshold rule) and path B (machine learning model) determine that the same level of alarm should be triggered (e.g., both are determined to be level 1 alarms), the system will immediately trigger that level of alarm with high confidence.

[0085] Model Priority or Review Trigger: If path A does not trigger an alarm, but the model of path B judges it to be in a certain alarm state with a high probability (e.g., probability > 85%), the system will trigger an alarm of that level, but it can be marked as "requires review" or prompted in the push message as "based on AI model prediction".

[0086] Rule priority triggering: If the path B model judges it as normal or low-level alarm, but the hard rules of path A (especially the first-level alarm rule) are triggered, the system will prioritize following the rules and immediately trigger the corresponding alarm to ensure the bottom line of clinical safety.

[0087] No alarm: The system maintains a no alarm state only when both paths are judged to be in normal condition.

[0088] Step S604: Trigger the alarm and generate alarm context information

[0089] Based on the final decision in step S603, the system performs the following operations:

[0090] Trigger Alarm: Invoke the alarm management service to trigger an alarm of the corresponding level (Level 1 / Level 2) and push it through preset channels (system interface, mobile terminal).

[0091] Context Generation: Simultaneously, the system automatically generates a structured alarm context report, including: a snapshot of the key parameter data that triggered the alarm, the specific rule entry or model prediction probability that led to the trigger, and links to relevant time-series trend image segments. This report is pushed along with the alarm and provides detailed information for subsequent data integration steps in S700.

[0092] When the Level 1 alarm criteria are met (e.g., typical characteristics of brain herniation: a rapid increase in pupil diameter exceeding a threshold within a short period accompanied by loss of light reflex), the system immediately triggers the highest level emergency alarm (red alarm). When the Level 2 alarm criteria are met (e.g., the algorithm identifies abnormal trend changes, such as a continuous increase in pupil asymmetry and a progressive slowing of light reflex velocity, but not reaching the Level 1 alarm criteria), the system triggers an early warning alarm (yellow alarm). These alarm messages are pushed to relevant personnel in real time through the system interface (e.g., monitor screen) and mobile terminals (e.g., medical staff's PDAs or mobile phones).

[0093] Step S700: Data Integration and Report Generation. All pupil parameter data, trend charts, and triggered alarm event records generated throughout the monitoring process are automatically integrated into the hospital's Electronic Medical Record (EMR) system via standard interfaces (such as HL7 and FHIR). Simultaneously, the system can automatically generate structured monitoring reports for clinical review, auditing, and verification.

[0094] This invention provides a flexible periorbital pupil real-time monitoring system based on near-infrared penetrating eyelids for implementing the above-described method. The system includes:

[0095] At least two flexible smart eye patch devices 1 are used to be worn around the left and right eye sockets of the patient, respectively.

[0096] A central processing platform 2 is wirelessly connected to the flexible smart eye patch device.

[0097] like Figure 2 As shown, the flexible smart eye patch device includes:

[0098] The flexible patch body 101 is typically made of biocompatible medical-grade silicone or polyurethane film and is ring-shaped (e.g., donut-shaped or miniature swimming goggles). It has a hollow area in the center to fully expose the eyeball and cornea, achieving zero contact and zero pressure. Its bottom surface is coated with medical pressure-sensitive adhesive for comfortable and secure adhesion to the skin around the eye socket. This pressure-sensitive adhesive is preferably a hypoallergenic hydrocolloid or silicone gel pressure-sensitive adhesive with moisture-permeable and breathable properties to protect skin health.

[0099] Near-infrared light source module 102: Embedded in the solid ring of the patch body, it is usually composed of multiple low-power near-infrared LEDs (light-emitting diodes) arranged in a ring or a specific array to emit low-intensity near-infrared light with a wavelength of 700-1000nm. This wavelength of light can safely penetrate eyelid tissue to provide illumination for closed-eye monitoring.

[0100] Near-infrared image sensor module 103: Also embedded in the patch body, located inside the near-infrared light source ring or at a specific position, used to receive near-infrared light reflected from the eyelid, iris and pupil areas to capture eye images.

[0101] Microprocessor and control module 104: Embedded within the patch body, it can be a microcontroller (MCU) or a system-on-a-chip (SoC). It is responsible for controlling the switching of the near-infrared light source, modulating the pulse frequency (for light reflection testing), performing preliminary processing (such as compression and formatting) on ​​the raw data acquired by the image sensor, and coordinating the work of various modules within the eye patch.

[0102] Wireless communication module 105: Embedded in the patch body, such as a Bluetooth Low Energy (BLE) module or a Wi-Fi module, responsible for wirelessly transmitting data processed by the microprocessor or raw image data to the central processing platform and receiving instructions from the platform.

[0103] Power module 106: Embedded in the patch body, it is usually a miniature, rechargeable or disposable button battery or flexible battery that provides power to all electronic modules in the eye patch.

[0104] Optional modules: Motion sensors, such as miniature accelerometers and gyroscopes, are used to monitor eye movements and eyelid tremors, which can serve as auxiliary information for assessing neural function.

[0105] The core technology of this flexible smart eye patch utilizes near-infrared light to penetrate the eyelids and image the pupil in a closed state. After the near-infrared light penetrates the eyelids, the iris tissue reflects a significant amount of light, while the pupil (as an opening) absorbs most of the light and reflects very little. By capturing this difference in reflection, the image sensor can clearly distinguish the boundary of the pupil behind the eyelids, thereby accurately calculating its diameter. By controlling the near-infrared light source to emit standardized light pulses, the system can stimulate pupil contraction and quantify the latency, contraction speed, and amplitude of light reflection by analyzing a series of images.

[0106] like Figure 3 As shown, the central processing platform 2 can be a dedicated hardware device, a software platform on a server, or a cloud platform, and includes:

[0107] Data receiving unit 201: Responsible for receiving image data or preprocessed data packets from the left and right eye patch devices.

[0108] State Recognition and Mode Switching Unit 202: Runs an open / closed eye state recognition algorithm to automatically identify open / closed eye states and seamlessly switch between direct viewing mode and near-infrared penetration mode accordingly. This ensures the continuity and accuracy of monitoring data under different physiological states and issues commands to control the working mode of the eye patch or subsequent data processing procedures. Specifically, the open / closed eye state recognition algorithm performs image preprocessing and region of interest standardization on the acquired image sequence. This is achieved by using a cascaded classifier based on Haar features or a semantic segmentation network based on U-Net for rapid localization and segmentation of the eyeball or eyelid region. Then, the acquired raw near-infrared image is filtered for noise reduction and contrast enhancement, and the standardized eye region is located and cropped. Multi-dimensional feature extraction and fusion analysis are performed, extracting texture features and grayscale statistical features in parallel from the standardized image, as well as the uniform texture and low contrast characteristics of the eyelid skin in the closed eye state. Based on a rule-based or deep learning classifier, the extracted features are input into a preset threshold rule model or a pre-trained lightweight convolutional neural network classifier, outputting a probability judgment of whether the eyes are open or closed.

[0109] Pupil Parameter Analysis Unit 203: Runs sophisticated image processing algorithms to extract various physiological pupil parameters from image sequences with high precision. Specifically: For near-infrared transmission images with eyes closed, it employs image segmentation algorithms based on region growing or level set evolution, combined with the extremely low grayscale value of the pupil region relative to the iris region, to segment the pupil boundary; for direct-view images with eyes open, it uses algorithms based on circular Hough transform or edge detection combined with ellipse fitting to extract the pupil boundary; based on the segmented pupil boundary, it calculates the pupil diameter and area; by comparing the left and right eye image sequences, it calculates the pupil symmetry index; when the near-infrared light source module emits standardized light pulse stimulation, it analyzes the dynamic changes in pupil size in the image sequences before and after stimulation, calculates the light reflection latency, contraction speed, and contraction amplitude, and integrates these dynamic parameters to calculate the neural pupillary index.

[0110] Trend Analysis and Early Warning Unit 204: Responsible for storing historical pupil data, constructing visual trend charts, and running sequence analysis algorithms to identify anomalies, including using sliding window-based statistical process control methods, change point detection algorithms, or long short-term memory network models to identify gradual or abrupt abnormal trends in pupil parameters. It triggers tiered alarms according to preset rules. The preset rules involve constructing a comprehensive feature vector containing instantaneous pupil parameter values, dynamic trend statistics, and time series model outputs; parallel execution of logical judgments based on preset threshold rules and probabilistic inferences based on machine learning classification models; and employing a decision fusion strategy to integrate the rule judgment results and model inference results to generate the final alarm level judgment. The final alarm judgment is divided into Level 1 and Level 2 alarm conditions. The Level 1 alarm condition is the detection of a unilateral pupil diameter that sharply increases beyond the first threshold within a preset time window, accompanied by the disappearance of light reflex. The Level 2 alarm condition is the detection of a bilateral pupil asymmetry index that remains above the second threshold for a preset duration, and / or a moving average of the light reflex velocity that remains below the third threshold for a preset duration.

[0111] Data Integration and Reporting Unit 205: Responsible for interacting structured data with the hospital information system and generating standardized reports.

[0112] The principles of the present invention are explained below with reference to specific embodiments:

[0113] Example 1: Structure of Flexible Smart Eye Patch

[0114] The flexible smart eye patch device in this embodiment is shaped like a flexible ring, similar to a miniature swimming goggle. Its core feature is its hollow design: the hollow area in the center faces the user's eyeball and cornea during wear, ensuring no contact between the device and the sensitive surface of the eye. The solid part of the eye patch is made of highly biocompatible medical-grade silicone, with a thickness controlled at 1-3 mm, and a soft texture that conforms to the curvature of the orbital bone. Its bottom surface is coated with a layer of hypoallergenic, moisture-permeable, and breathable medical pressure-sensitive adhesive to comfortably and firmly adhere the eye patch to the skin around the user's eye socket.

[0115] like Figure 2As shown, multiple functional modules are integrated within the solid ring of the eye patch using miniaturized and flexible electronics technology. The near-infrared light source module consists of six 850nm wavelength micro-LEDs arranged in a ring at equal intervals, emitting uniform, low-intensity near-infrared light towards the eye. The near-infrared image sensor module (such as a miniature global shutter CMOS sensor) is located inside the light source ring, its optical axis roughly aligned with the center of the pupil. The microprocessor and control module, along with the wireless communication module (such as a BLE5.0 module), are integrated on a flexible circuit board, responsible for controlling the LED drive (including generating standardized light pulses for light reflection testing), reading sensor data, and performing preliminary processing and packaging. A rechargeable miniature lithium polymer battery powers all modules. All electronic components are encapsulated in medical-grade silicone to ensure safety upon contact with the human body and the durability of the device. The eye patch is preferably a single-use product to eliminate the risk of cross-infection.

[0116] Example 2: Workflow of the monitoring system

[0117] The system workflow in this embodiment is as follows:

[0118] Wearing and Connection. Healthcare professionals apply the two flexible smart eye patches to the patient's eye sockets, ensuring the hollow area is aligned with the eyeballs. After powering on, the eye patches complete a self-test and establish a Bluetooth connection with the bedside central processing platform (such as a ruggedized tablet).

[0119] Image acquisition. The near-infrared LEDs within the eye patch operate in a low-duty-cycle pulse mode to reduce power consumption and thermal effects. The near-infrared image sensor continuously acquires images of the eye at a certain frame rate (e.g., 15-30 fps).

[0120] Data transmission and status recognition. The acquired image data is initially compressed by the microprocessor and then transmitted to the central processing platform via a wireless module. The platform runs a status recognition algorithm (see Example 3) to determine whether the patient's eyes are currently open or closed.

[0121] Mode switching. Based on the status recognition result, the system automatically switches the working mode. If the eyes are open, it enters "direct viewing mode", which can perform higher precision measurements by combining ambient light; if the eyes are closed, it remains in "near-infrared penetration mode".

[0122] Pupil parameter analysis. The platform calls the pupil parameter extraction algorithm (see Example 4) to calculate pupil diameter, symmetry, light reflection parameters, etc. in real time from the image sequence of the current mode.

[0123] Trend analysis and early warning judgment. The calculated parameters are stored in the database and the trend chart is updated in real time. The system continuously runs the trend analysis algorithm and compares it with the preset alarm threshold (see Example 5 for details).

[0124] Alarm Triggering and Push Notifications. If the alarm conditions are met, the system immediately triggers the corresponding alarm level (Level 1 Red or Level 2 Yellow), displays the alarm information on the tablet screen, and simultaneously pushes the alarm to the central monitoring screen at the nursing station and the mobile handheld terminal of the on-duty nurse.

[0125] Data logging and integration. All data, including raw images (optionally stored), parameters, trend charts, and alarm events, are logged. After monitoring concludes or periodically, the system automatically generates reports and writes key data into the hospital's electronic medical record system via the HL7 interface.

[0126] Example 3: State Recognition and Mode Switching Method

[0127] The state recognition and mode switching method in this embodiment specifically includes:

[0128] Step S301: Image Preprocessing and ROI Normalization. For each received near-infrared image frame, Gaussian filtering is first performed to reduce noise, followed by histogram equalization to enhance contrast. Next, a pre-trained cascaded classifier based on Haar features (or a more advanced U-Net segmentation network) is used to quickly locate the eye region, and a normalized image patch of fixed size is cropped out as the region of interest (ROI) for subsequent processing.

[0129] Step S302: Multi-dimensional feature extraction. Two types of features are extracted in parallel from the standardized ROI:

[0130] Texture features: For example, calculating the Local Binary Pattern (LBP) histogram of an image. When the eyes are open, the iris and pupil regions exhibit rich texture structures, and the LBP histogram distribution is relatively dispersed; while when the eyes are closed, the eyelid skin surface is relatively smooth and uniform, and the LBP histogram distribution is more concentrated.

[0131] Gray-scale statistical characteristics: Calculate the mean, standard deviation, and entropy of the gray-scale of the ROI image. When the eyes are open, the dark color of the pupil contrasts highly with the iris and sclera, resulting in a relatively high gray-scale standard deviation and entropy. When the eyes are closed, near-infrared light is uniformly scattered through the eyelids, resulting in low image contrast and lower gray-scale standard deviation and entropy.

[0132] Step S303: State Decision. The extracted texture and grayscale feature vectors are input into a pre-trained lightweight convolutional neural network (e.g., a simple 3-layer CNN) or a threshold-based logistic regression classifier. The classifier outputs a probability value between 0 and 1, representing the probability that the current image is in a "closed eyes" state. A threshold (e.g., 0.5) is set; if the probability is greater than the threshold, the image is judged as "open eyes," otherwise it is judged as "closed eyes."

[0133] Step S304: Mode Switching. The system maintains a brief historical state buffer (e.g., the state of the last 5 frames). Mode switching is only performed when the judgment results of multiple consecutive frames (e.g., 3 frames) are all in the same state, in order to avoid frequent mode jumps caused by single-frame misjudgment or blinking, and to ensure the stability of monitoring.

[0134] Example 4: Pupil Parameter Extraction Method

[0135] The pupil parameter extraction method in this embodiment has different focuses depending on the different modes:

[0136] (1) For near-infrared transmission mode (closed eyes) images:

[0137] First, image enhancement. To further highlight the pupil area, top-hat transformation can be used to correct for uneven illumination, and contrast-limited adaptive histogram equalization (CLAHE) can be used to enhance local details.

[0138] Secondly, pupil segmentation. Since the pupil region absorbs most of the light in near-infrared transmission images, appearing as an extremely dark and well-connected area, this invention preferably uses a region growing algorithm for segmentation. Starting with the darkest pixel or multiple seed points in the image, region growing is performed based on pixel grayscale similarity until a clear boundary is encountered (corresponding to the edge of the iris and pupil).

[0139] Boundary fitting and parameter calculation are performed again. Edge extraction is performed on the segmented binarized pupil region. Due to eyelid occlusion, the pupil outline may be incomplete; therefore, an ellipse fitting algorithm (such as least squares) is used to reconstruct the complete pupil boundary ellipse. Based on the fitted ellipse, the pupil diameter (which can be the average of the major and minor axes) and area can be calculated.

[0140] (2) For images in direct viewing mode (open eyes):

[0141] First, the iris and pupil are located. In images with the eyes open, the pupil typically appears as a clear circle or a near-circular dark area. A circular Hough transform can be used to directly detect the circular boundary of the pupil. Canny edge detection can be performed first, and then contours that match the pupil size and circularity can be fitted with an ellipse.

[0142] Next, parameter calculations are performed. Based on the located pupil boundaries, the pupil diameter and area are calculated.

[0143] The light reflection parameters are calculated again. When the system controls the near-infrared light source to emit a standardized light pulse stimulus, the timestamp of the stimulus start is recorded. The change curve of pupil size in the image sequence before and after the stimulus (e.g., 2-3 seconds) is analyzed.

[0144] The latency of the pupillary light reflex is defined as the time elapsed from the onset of light stimulation until the first detectable narrowing of the pupil diameter.

[0145] Light-reflective contraction rate: defined as the maximum rate of change of pupil diameter (mm / s) during the initial stage of pupil contraction (e.g., the first 200 ms).

[0146] Constriction amplitude of pupillary light reflex: defined as the percentage or absolute value of the change in pupil diameter before and after stimulation.

[0147] The neuro-pupil index is a quantitative indicator that integrates multiple dynamic parameters mentioned above. It can be used to fuse multiple parameters into a single score through a predefined formula or machine learning model to comprehensively assess pupillary motor function.

[0148] Finally, symmetry is calculated. The pupil diameters are calculated by simultaneously comparing the image sequences from the left and right eye patches, and the difference or ratio of the left and right pupil diameters is used as an indicator of pupil symmetry.

[0149] Example 5: Early Warning Mechanism

[0150] The graded early warning mechanism in this embodiment is as follows:

[0151] The system continuously monitors parameters such as pupil diameter, symmetry, and light reflection speed from both eyes.

[0152] Level 2 Alert (Yellow, Early Warning) Judgment:

[0153] Condition 1: The moving average difference between the diameters of the left and right pupils is greater than 0.5 mm for more than 10 consecutive minutes.

[0154] Condition 2: The moving average of the pupillary light reflex contraction velocity of one or both pupils is below 30% of the baseline velocity for more than 15 consecutive minutes.

[0155] If either condition 1 or condition 2 is met, a Level 2 alert is triggered. This alert alerts healthcare professionals to pay close attention, as it may indicate a slow increase in intracranial pressure or early deterioration of neurological function.

[0156] Level 1 Alert (Red, Emergency Alert) Judgment:

[0157] Conditions: The diameter of one pupil increases rapidly by more than 2 mm within 5 minutes, and during this process, the pupillary light reflex of that side completely disappears (i.e., the pupil does not constrict under light stimulation).

[0158] If this condition is met, a Level 1 alarm will be triggered immediately. This alarm strongly suggests the possibility of acute brain herniation, requiring immediate emergency intervention by medical personnel.

[0159] Once an alarm is triggered, it will be recorded in the system's alarm log and pushed to medical staff.

[0160] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for real-time monitoring of the periorbital pupil based on near-infrared penetrating eyelids, characterized in that, Includes the following steps: The flexible smart eye patch device is worn on the skin around both eyes of the patient. The eye patch device has a hollow structure to ensure no contact with the eyeball and is fixed by built-in medical pressure-sensitive adhesive. After the system is powered on, it completes self-test and establishes a wireless communication connection with an external terminal. The eye patch device emits low-intensity near-infrared light with a wavelength of 700-1000nm through a near-infrared light source module. At the same time, the eye patch continuously collects images through a near-infrared image sensor module integrated in the same eye patch. The images are direct images when the patient's eyes are open and transmitted images after the near-infrared light penetrates the eyelids when the eyes are closed. The acquired image sequences are analyzed in real time by a microprocessor or a remote central processing platform, automatically identifying whether the patient's eyes are open or closed, and seamlessly switching between direct viewing mode and near-infrared penetration mode accordingly. On the central processing platform, an image processing algorithm is run to extract one or more pupil physiological parameters from the image sequence; The extracted pupil physiological parameters are stored in chronological order, and a time series trend chart is constructed. The trend graph is analyzed using time-series data analysis algorithms to identify abnormal change patterns in pupil parameters; Based on preset threshold rules or the output of machine learning models, the abnormal change patterns are judged; when the conditions for a Level 1 alarm are met, the highest level of emergency alarm is immediately triggered. When the conditions for a Level 2 alarm are met, an early warning alarm is triggered. The alarm information will be pushed to medical staff through the system interface and mobile terminals; The pupil parameter data, trend charts, and alarm event records generated during the monitoring process are automatically integrated into the hospital's electronic medical record system, and monitoring reports are generated for clinical review and auditing.

2. The method for real-time monitoring of the flexible periorbital pupil based on near-infrared penetrating eyelids according to claim 1, characterized in that, The pupillary physiological parameters include pupil diameter, pupil area, left and right pupil symmetry index, pupillary light reflex latency, pupillary light reflex contraction speed, pupillary light reflex contraction amplitude, and neuro-pupil index.

3. The method for real-time monitoring of the flexible periorbital pupil based on near-infrared penetrating eyelids according to claim 1, characterized in that, The process of real-time analysis of the acquired image sequences to automatically identify the patient's open or closed eye state specifically includes: The acquired image sequence is preprocessed and the region of interest is standardized. The original near-infrared image is filtered for noise reduction and contrast enhancement. The standardized eye region is located and cropped. Multi-dimensional feature extraction and fusion analysis were performed to extract texture features and grayscale statistical features in parallel from standardized images, as well as the uniform texture and low contrast characteristics of eyelid skin in the closed state. State decisions based on rules or deep learning classifiers input the extracted features into a preset threshold rule model or a pre-trained lightweight convolutional neural network classifier, and output the probability judgment of whether the eyes are open or closed.

4. The method for real-time monitoring of the flexible periorbital pupil based on near-infrared penetrating eyelids according to claim 3, characterized in that, In the image preprocessing and region of interest normalization steps, a cascaded classifier based on Haar features or a semantic segmentation network based on U-Net is used to quickly locate and segment the eyeball or eyelid region.

5. The method for real-time monitoring of the flexible periorbital pupil based on near-infrared penetrating eyelids according to claim 1, characterized in that, The image processing algorithm extracts one or more pupil physiological parameters from the image sequence, specifically including: For near-infrared transmission images with eyes closed, an image segmentation algorithm based on region growing or level set evolution is used, combined with the extremely low gray value of the pupil region relative to the iris region, to segment out the pupil boundary. For a direct view image with eyes open, an algorithm based on circular Hough transform or edge detection combined with ellipse fitting is used to extract the pupil boundary. Based on the segmented pupil boundaries, the pupil diameter and area are calculated; the pupil symmetry index is calculated by comparing the left and right eye image sequences. When the near-infrared light source module emits standardized light pulse stimulation, the dynamic changes in pupil size in the image sequence before and after stimulation are analyzed, the light reflex latency, contraction speed, and contraction amplitude are calculated, and the neural pupillary index is calculated by combining these dynamic parameters.

6. The method for real-time monitoring of the flexible periorbital pupil based on near-infrared penetrating eyelids according to claim 1, characterized in that, The analysis of trend graphs using time-series data analysis algorithms includes using a sliding window-based statistical process control method, a change point detection algorithm, or a long short-term memory network model to identify gradual or abrupt abnormal trends in pupil parameters.

7. The method for real-time monitoring of the flexible periorbital pupil based on near-infrared penetrating eyelids according to claim 1, characterized in that, The judgment of the abnormal change pattern based on the preset threshold rule or the output of the machine learning model specifically includes: Construct a comprehensive feature vector that includes instantaneous values ​​of pupil parameters, dynamic trend statistics, and time series model output; Parallel execution of logical judgments based on preset threshold rules and probabilistic inferences based on machine learning classification models; A decision fusion strategy is adopted to integrate the rule judgment results and model inference results to generate the final alarm level judgment.

8. The method for real-time monitoring of the flexible periorbital pupil based on near-infrared penetrating eyelids according to claim 1, characterized in that, The first-level alarm condition is that the diameter of one pupil increases sharply beyond the first threshold within a preset time window and is accompanied by the disappearance of light reflection; the second-level alarm condition is that the bilateral pupil asymmetry index is continuously higher than the second threshold for a preset duration, and / or the moving average of the light reflection velocity is continuously lower than the third threshold for a preset duration.

9. A flexible periorbital pupil real-time monitoring system based on near-infrared penetrating eyelid for implementing the method of any one of claims 1-8, characterized in that, include: At least two flexible smart eye patch devices are used to be worn around the patient's left and right eye sockets, respectively; A central processing platform is wirelessly connected to the flexible smart eye patch device; The flexible smart eye patch device includes: The flexible patch body is ring-shaped with a hollow area in the middle to expose the eyeball. Its bottom surface is coated with medical pressure-sensitive adhesive for adhesion to the skin of the eye socket. A near-infrared light source module is embedded in the patch body and is used to emit near-infrared light with a wavelength of 700-1000nm; A near-infrared image sensor module is embedded in the patch body for acquiring images of the eye; The microprocessor and control module, embedded in the patch body, is used to control the switching and pulses of the near-infrared light source and to perform preliminary processing of image data. A wireless communication module, embedded in the patch body, is used to interact with the central processing platform for data exchange. The power module is embedded in the patch body and supplies power to the various electronic modules inside the eye patch. The central processing platform includes: The data receiving unit is used to receive image data or processed data from the flexible smart eye patch device; The state recognition and mode switching unit is used to automatically recognize the open / closed eye state based on image data and control the switching of monitoring modes. The pupil parameter analysis unit is used to run image processing algorithms to extract pupil physiological parameters from image sequences. The trend analysis and early warning unit is used to store pupil parameters, construct trend graphs, analyze abnormal patterns, and trigger graded alarms. The data integration and reporting unit is used to integrate data and alarm records into the electronic medical record system and generate monitoring reports.

10. The flexible periorbital pupil real-time monitoring system based on near-infrared penetrating eyelids according to claim 9, characterized in that, The flexible smart eye patch device also integrates a motion sensor for monitoring eye movements; the central processing platform also uses the motion sensor data to assist in assessing the state of neural function.

Images