Intelligent intestinal tract monitoring method and system based on multispectral sensing technology

Abstract

The invention discloses an intelligent intestinal tract monitoring method and system based on a multispectral sensing technology. The method comprises the following steps: acquiring spectral data of excrement through a multispectral sensor array integrated on a toilet seat; carrying out data processing by adopting segmented trapezoidal architecture mapping and a Spiking neural network, and extracting a light transmittance characteristic matrix; a nonlinear attention mechanism is realized based on a Hopfield network, and a light transmittance dynamic curve is generated; performing intestinal preparation degree evaluation by using a sparse MLP in combination with a motif structure optimization algorithm; and remote monitoring and intelligent feedback are realized through a 5G network. Objective quantitative evaluation of the intestinal preparation process is achieved, the preparation efficiency and quality of enteroscopy and intestinal surgery are improved, and the technical problems that traditional intestinal preparation evaluation is high in subjectivity and lacks real-time monitoring are solved.

Description

Technical Field

[0001] The present invention relates to the field of medical monitoring technology, and in particular to an intelligent intestinal monitoring method and system based on multispectral sensing technology, which is particularly suitable for intestinal preparation quality assessment before colonoscopy and intestinal surgery. Background Art

[0002] Colonoscopy and intestinal surgery are important tools for diagnosing and treating digestive system diseases. Their effectiveness and safety largely depend on the adequacy of preoperative bowel preparation. Traditional bowel preparation assessment relies primarily on the patient's subjective description and the medical staff's judgment, lacking objective and quantitative evaluation criteria and methods.

[0003] Currently, commonly used bowel preparation methods include oral laxatives and specific dietary restrictions. In clinical practice, physicians often provide patients with written instructions, instructing them to take laxatives before the examination and observe their stool until clear fluid is excreted. However, this approach relies heavily on patient self-judgment and compliance, frequently leading to inadequate bowel preparation.

[0004] A more advanced method uses a vision system to photograph fecal matter and analyze its color and turbidity. This technology captures fecal matter images using a camera installed in the restroom and uses computer vision algorithms to analyze their visual features to determine the degree of bowel preparation. However, this method presents privacy concerns, and image quality varies significantly under different lighting conditions, limiting analysis accuracy.

[0005] The main problems with existing technologies include: first, the lack of non-invasive, objective and quantitative real-time bowel preparation assessment methods, which often leads to delays or cancellations of examinations due to insufficient preparation; second, there is information asymmetry between doctors and patients. Doctors cannot understand the patient's bowel preparation progress in real time and can only confirm whether the preparation is sufficient after the patient arrives at the hospital, resulting in a waste of medical resources and a decline in patient experience. Summary of the Invention

[0006] The purpose of the present invention is to overcome the shortcomings of the existing technology and provide an intelligent intestinal monitoring method and system based on multispectral sensing technology to achieve objective quantitative evaluation of the intestinal preparation process and improve the preparation efficiency and quality of colonoscopy and intestinal surgery.

[0007] To achieve the above object, the technical solution adopted by the present invention is:

[0008] The present invention provides an intelligent intestinal monitoring method based on multispectral sensing technology, comprising: activating a multispectral sensor array integrated in a toilet seat based on a trigger signal of a toilet usage status sensor, and simultaneously collecting spectral reflection and transmission data within a wavelength range of 450nm-1100nm at a sampling frequency of 120Hz to generate a multidimensional spectral data stream; normalizing the multidimensional spectral data stream to obtain processed data, and mapping the processed data to a segmented ladder computing architecture, performing parallel noise reduction and feature enhancement processing through a Spiking neural network, and extracting a fecal transmittance feature matrix; based on the fecal transmittance, The system uses a feature matrix to construct a Hopfield network model and implement a nonlinear attention mechanism, adaptively focusing on time periods and wavelength ranges with significant transmittance changes, generating a dynamic curve of fecal transmittance and key change points; constructing feature vectors for the transmittance dynamic curve and key change points, and processing them using a sparse MLP model combined with a modular structure optimization algorithm to generate a bowel preparation quality assessment result that meets the BOSTON scoring standard; transmitting the bowel preparation quality assessment result to the doctor-side application via the 5G network, and pushing personalized guidance suggestions to the patient based on the bowel preparation quality assessment result, realizing remote monitoring and intelligent feedback.

[0009] Preferably, the trigger signal based on the toilet usage status sensor activates the multispectral sensor array integrated in the toilet seat, and simultaneously collects spectral reflection and transmission data in the wavelength range of 450nm-1100nm at a sampling frequency of 120Hz to generate a multidimensional spectral data stream, including: based on the trigger signal detected by the toilet usage status sensor, activating the multispectral sensor array comprising 6 narrow-band light sources of different wavelengths and corresponding photodetectors in the annular area inside the toilet seat to form an annular sensor network; establishing a timing synchronization mechanism for light source emission and detector reception for the annular sensor network to ensure that the emission timing of each wavelength light source accurately matches the detector receiving window; the multispectral sensor array controlled by the timing synchronization mechanism simultaneously collects spectral reflection and transmission data of 450nm, 550nm, 650nm, 850nm, 950nm and 1100nm wavelengths at a sampling frequency of 120Hz, and records the light intensity change data of each band; the light intensity change data of each band is timestamped and packaged, and transmitted to the local processing unit via the low-power Bluetooth protocol to generate the multidimensional spectral data stream.

[0010] Preferably, the multidimensional spectral data stream is normalized to obtain processed data, the processed data is mapped to a segmented ladder computing architecture, and parallel noise reduction and feature enhancement processing is performed through a spiking neural network to extract a fecal transmittance feature matrix, including: zero-point calibration and ambient light interference elimination of the spectral data of each band in the multidimensional spectral data stream, calculation of relative intensity and differential data, and obtaining normalized spectral data; dividing the frequency domain analysis results of the normalized spectral data into three processing levels: high frequency band, mid-frequency band, and low frequency band, and mapping them to a segmented ladder computing architecture; deploying a spiking neural network of a leaky integrate-and-fire neuron model on the segmented ladder architecture, using fast response parameters and a local connection mode for the high frequency band, a balanced time constant and a sparse connection structure for the mid-frequency band, and a long time constant and a global connection mode for the low frequency band; through multi-level parallel processing of the spiking neural network, calculating the transmittance index of the feces in each band, and constructing the fecal transmittance feature matrix containing three-dimensional information of time, wavelength, and transmittance.

[0011] Preferably, based on the feces transmittance characteristic matrix, a Hopfield network model is constructed and a nonlinear attention mechanism is implemented to adaptively focus on time periods and wavelength intervals where transmittance changes significantly, and generate a dynamic curve of feces transmittance and key change points, including: based on the feces transmittance characteristic matrix, the continuous time series change data of feces transmittance are reconstructed through a 20-second time window sliding technique and weighted average processing; an energy function containing time-dependent terms, wavelength-related terms and stability terms is constructed for the continuous time series change data, and the Hopfield network model is trained using a contrast divergence algorithm; based on the Hopfield network model, a nonlinear transformation function of transmittance change, local volatility and information entropy is designed, and the comprehensive attention weight of time attention and wavelength attention is calculated; the continuous time series change data is weighted averaged by the comprehensive attention weight, and a smooth and continuous dynamic curve of the transmittance is generated by applying the B-spline interpolation technique, and inflection points, stable platforms and rapid change points are identified as the key change points through differential geometry and cluster analysis.

[0012] Preferably, the transmittance dynamic curve and key change points are subjected to feature vector construction, and a sparse MLP model is used in combination with a modular structure optimization algorithm for processing to generate an intestinal preparation quality assessment result that meets the BOSTON scoring standard, including: based on the transmittance dynamic curve and key change points, multidimensional feature indicators including the final transmittance value, transmittance rise rate, number of stable platforms, curve fluctuation index and similarity with the standard pattern are extracted to construct a feature vector; a sparse MLP model containing 4 hidden layers is designed for the feature vector, and a sparse connection structure is established by adopting feature grouping connection, residual jump connection and dynamic pruning strategy; the sparse MLP model is divided into a transparency module, a stability module and a timing module using modular theory, the internal structure of each module is optimized by information bottleneck theory, and the dependency relationship between modules is determined based on a directed acyclic graph to obtain an optimized sparse MLP model; the segmented scores and total score probability distribution of the right colon segment, the transverse colon segment and the left colon segment are calculated by the optimized sparse MLP model, and the confidence assessment and time series score are combined to generate the intestinal preparation quality assessment result that meets the BOSTON scoring standard.

[0013] Preferably, based on the trigger signal detected by the toilet usage status sensor, a multispectral sensor array comprising 6 narrow-band light sources of different wavelengths and corresponding photodetectors is activated in the annular area inside the toilet seat to form an annular sensor network, including: based on the toilet usage status sensor, detecting the state change signal of the user using the toilet and generating a trigger signal; wherein the toilet usage status sensor includes a pressure sensor, an infrared human body sensor and a sound detector; performing signal conditioning and digital processing on the trigger signal, judging the trigger condition by a microcontroller and generating a sensor array activation instruction; based on the sensor array activation instruction, Twelve evenly distributed locations in the annular area simultaneously activate LED narrowband light sources and silicon photodiode detectors, where the wavelengths of the light sources are 450nm, 550nm, 650nm, 850nm, 950nm and 1100nm, respectively, with each wavelength corresponding to two light source-detector pairs. An optical transmission channel is established through the light source-detector pairs to pass through the excrement, and each silicon photodiode detector synchronously receives the transmitted light signal of the corresponding wavelength, forming the annular sensor network. A self-test is performed on the annular sensor network to verify the working status of each sensor in the annular sensor network and the patency of the optical path, thereby ensuring the reliability of data collection.

[0014] Preferably, the spectral data of each band in the multidimensional spectral data stream are zero-point calibrated and ambient light interference is eliminated, and relative intensity and differential data are calculated to obtain normalized spectral data, including: based on the reference values ​​of each wavelength collected in the no-excrement state, the original readings of each wavelength in the multidimensional spectral data stream are zero-point calibrated and the relative intensity value is calculated; the adjacent time window difference technology is used for the relative intensity value to eliminate the influence of slowly changing ambient light by taking advantage of the characteristic that the ambient light changes little in a short time, and obtain differential data; amplitude normalization is performed based on the differential data, and the data of all wavelengths are mapped to a standardized numerical interval to generate standardized spectral data; fast Fourier transform is performed on the standardized spectral data to convert the time domain signal into a frequency domain representation, and the signal is divided into a high frequency band, a medium frequency band and a low frequency band according to the frequency characteristics; the high frequency band, the medium frequency band and the low frequency band are denoised by a combination of Gaussian filtering and median filtering to retain the effective signal characteristics while suppressing random noise, and obtain the normalized spectral data.

[0015] Preferably, the method of reconstructing the continuous time-series variation data of the feces transmittance based on the feces transmittance characteristic matrix through a 20-second time window sliding technique and weighted averaging processing includes: defining a time window of 20 seconds in length based on the feces transmittance characteristic matrix, sliding along the time axis with a step size of 2 seconds, and extracting the transmittance data of each time point within the window; applying a Gaussian kernel weighting function with the window center as the highest weight to the transmittance data, calculating the weighted average transmittance value, eliminating the influence of random noise, and obtaining discrete transmittance data points after weighted averaging; A higher sampling density and weight coefficient are used for processing the absorption peaks of hemoglobin, bilirubin and water corresponding to the key wavelengths of 550nm, 650nm and 850nm; the discrete transmittance data points after the weighted average are interpolated using a cubic spline interpolation algorithm to generate a continuous transmittance time series function with a time resolution of 200ms; the absorption coefficient and optical path length calibration parameters of each wavelength are calculated based on the Beer-Lambert law, and the physical parameter correction of the continuous transmittance time series function is performed to obtain the continuous time series change data.

[0016] Preferably, based on the transmittance dynamic curve and key change points, multidimensional feature indicators including the final transmittance value, transmittance rising rate, number of stable platforms, curve fluctuation index and similarity with the standard pattern are extracted to construct a feature vector, including: extracting final state characteristics based on the transmittance dynamic curve, including the final transmittance value, the final transmittance stable time and the wavelength difference index, reflecting the final clarity of the excrement; calculating process characteristics of the transmittance dynamic curve, including the transmittance rising rate, the number and duration of stable platforms, and the curve fluctuation index, reflecting the dynamic change characteristics of the excretion process; extracting time characteristics based on the key change points, including the time to first reach the threshold, the inflection point distribution density and the overall duration, reflecting the time distribution pattern of intestinal preparation; performing shape encoding on the transmittance dynamic curve through Fourier transform and principal component analysis, calculating the similarity matrix with the predefined standard pattern library, and obtaining pattern characteristics; normalizing and dimensionally integrating the final state characteristics, process characteristics, time characteristics and pattern characteristics to construct the feature vector.

[0017] The present invention also provides an intelligent intestinal monitoring system based on multispectral sensing technology, including: a multispectral sensor module, which is used to activate the multispectral sensor array integrated in the toilet seat based on the trigger signal of the toilet usage status sensor, and simultaneously collect spectral reflection and transmission data in the wavelength range of 450nm-1100nm at a sampling frequency of 120Hz to generate a multidimensional spectral data stream; a data processing module, which is used to normalize the multidimensional spectral data stream to obtain processed data, and map the processed data to a segmented ladder computing architecture, perform parallel noise reduction and feature enhancement processing through a Spiking neural network, and extract the excrement transmittance feature matrix; an intelligent analysis module, which is used to analyze the excrement transmittance feature matrix based on the The fecal transmittance feature matrix is ​​used to construct a Hopfield network model and implement a nonlinear attention mechanism, adaptively focusing on time periods and wavelength ranges where transmittance changes significantly, and generating a dynamic curve of fecal transmittance and key change points; an evaluation calculation module is used to construct feature vectors of the transmittance dynamic curve and key change points, and a sparse MLP model combined with a module structure optimization algorithm is used to process them to generate a bowel preparation quality evaluation result that meets the BOSTON scoring standard; a communication module is used to transmit the bowel preparation quality evaluation result to the doctor-side application via the 5G network, and push personalized guidance suggestions to the patient based on the bowel preparation quality evaluation result, thereby realizing remote monitoring and intelligent feedback.

[0018] The beneficial effects of the present invention are:

[0019] 1. It achieves objective quantitative evaluation of the bowel preparation process, avoiding the uncertainty of traditional methods that rely on patients' subjective judgment, and improving the efficiency and quality of preparation for colonoscopy and intestinal surgery;

[0020] 2. Through multispectral sensing technology, richer spectral information is obtained than traditional visual systems, improving the accuracy of identifying fecal characteristics while avoiding privacy issues;

[0021] 3. The innovative use of segmented ladder architecture mapping and spiking neural network technology enables real-time noise reduction and feature extraction of complex biological signals, improving the system's computational efficiency and accuracy.

[0022] 4. A nonlinear attention framework based on the Hopfield network can adaptively focus on time periods and wavelength ranges with significant transmittance changes, significantly improving the ability to capture the dynamic changes in fecal transmittance.

[0023] 5. Through the sparse MLP module-based structural optimization algorithm, efficient feature mapping and score conversion are achieved, significantly improving the accuracy and interpretability of bowel preparation assessment;

[0024] 6. It realizes real-time information sharing and remote guidance between doctors and patients, solves the problem of information asymmetry between doctors and patients, and improves the efficiency of medical resource utilization and patient experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0026] Figure 1 Schematic diagram of the process of the intelligent intestinal monitoring method based on multispectral sensing technology of the present invention;

[0027] Figure 2 This is a schematic diagram of the deployment of the multispectral sensor array of the present invention;

[0028] Figure 3 This is a schematic diagram of the Spiking neural network structure of the present invention;

[0029] Figure 4 This is an example diagram of the transmittance dynamic curve and key change points of the present invention;

[0030] Figure 5 This is a schematic diagram of the structure of the intelligent intestinal monitoring system based on multispectral sensing technology of the present invention. DETAILED DESCRIPTION

[0031] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0032] like Figure 1 As shown, the present invention provides an intelligent intestinal monitoring method based on multispectral sensing technology, comprising the following steps:

[0033] Step S1: Based on the trigger signal of the toilet usage status sensor, the multispectral sensor array integrated in the toilet seat is activated, and the spectral reflection and transmission data in the wavelength range of 450nm-1100nm are simultaneously collected at a sampling frequency of 120Hz to generate a multi-dimensional spectral data stream. This step is the basis of data acquisition of the present invention. By deploying the multispectral sensor array in the annular area inside the toilet seat, contactless and real-time monitoring of the spectral characteristics of excrement is achieved. When the user uses the toilet, the built-in status sensor will automatically trigger the work and activate the multispectral sensor array to start data collection. The use of multi-wavelength narrow-band light sources covers the spectral range from visible light to near-infrared, which can fully capture the spectral characteristics of excrement. The high sampling frequency ensures the accurate capture of the rapid changes in the state of excrement, providing high-quality raw data for subsequent analysis.

[0034] Step S2: Normalize the multidimensional spectral data stream to obtain processed data, and map the processed data to a segmented ladder computing architecture. Using a spiking neural network, parallel noise reduction and feature enhancement are performed to extract the fecal transmittance feature matrix. This step is a key step in data preprocessing and feature extraction. First, the raw spectral data is normalized to eliminate ambient light interference and sensor differences to ensure data comparability. An innovative segmented ladder computing architecture is then employed to perform hierarchical processing based on the data's frequency characteristics, efficiently allocating computing resources. On this basis, a bio-inspired spiking neural network is deployed for signal processing. This network simulates the pulse emission mechanism of neurons in the human brain and is particularly suitable for processing time series data. Through multi-level parallel processing, a feature matrix containing three-dimensional information of time, wavelength, and transmittance is ultimately extracted, laying the foundation for subsequent analysis.

[0035] Step S3: Based on the fecal transmittance characteristic matrix, a Hopfield network model is constructed and a nonlinear attention mechanism is implemented to adaptively focus on time periods and wavelength intervals where transmittance changes significantly, generating a dynamic curve of fecal transmittance and key change points. This step utilizes the pattern recognition and dynamic characteristic capture capabilities of modern Hopfield networks to conduct an in-depth analysis of the transmittance characteristic matrix. First, continuous time-series change data is reconstructed using a time window sliding technique. Then, an energy function containing time-dependent terms, wavelength-related terms, and stability terms is designed to construct a Hopfield network model. Based on this model, a nonlinear attention mechanism is implemented that can automatically identify and focus on key moments and key wavelengths where transmittance changes significantly, filter out redundant information, and improve analysis efficiency. Ultimately, a smooth and continuous dynamic transmittance curve is generated, and key change points such as inflection points and stable platforms are identified, providing an intuitive basis for the quantitative assessment of the degree of intestinal preparation.

[0036] Step S4: Construct feature vectors for the dynamic transmittance curve and key change points, and use a sparse MLP model combined with a modular structure optimization algorithm for processing to generate an intestinal preparation quality assessment result that meets the BOSTON scoring standard. This step converts the dynamic transmittance features into clinically usable assessment results. First, multidimensional feature indicators are extracted from the dynamic transmittance curve, including final state features, process features, time features, and pattern features, to construct a comprehensive feature vector. Then, a multi-layer perceptron model with a sparse connection structure is designed, and strategies such as feature grouping connection, residual jump connection, and dynamic pruning are used to improve the efficiency of the model. The network structure is optimized through modular theory, the model is divided into functional modules, and the internal structure of the modules and the relationship between modules are optimized. Finally, the intestinal preparation quality assessment results that meet the BOSTON scoring standard widely used in clinical practice are output, including the scores of each segment of the colon and the overall score, providing doctors with an objective and quantitative evaluation basis.

[0037] Step S5: The intestinal preparation quality assessment results are transmitted to the doctor-side application via the 5G network, and personalized guidance suggestions are pushed to the patient based on the intestinal preparation quality assessment results, realizing remote monitoring and intelligent feedback. This step realizes the secure transmission and intelligent application of the assessment results. First, the assessment results are encrypted and transmitted to the cloud server at high speed via the 5G network to ensure data security and real-time performance. The doctor-side application receives the decrypted assessment results and displays the historical curve, current status and predicted completion time in the form of intuitive charts to help doctors remotely monitor the progress of the patient's intestinal preparation. At the same time, based on the assessment results and historical data, combined with the medical knowledge base, personalized guidance suggestions are pushed to the patient, such as adjusting the laxative dosage, changing the diet or adjusting the preparation time. A two-way interactive channel is also established between doctors and patients, enabling doctors to adjust the guidance plan in real time based on the monitoring results, greatly improving the efficiency and quality of intestinal preparation.

[0038] like Figure 2 As shown, in step S1, based on the trigger signal of the toilet usage status sensor, the multispectral sensor array integrated in the toilet seat is activated to simultaneously collect spectral reflection and transmission data in the wavelength range of 450nm-1100nm at a sampling frequency of 120Hz to generate a multidimensional spectral data stream, including:

[0039] Step S1.1: Based on the trigger signal detected by the toilet usage status sensor, a multispectral sensor array consisting of six narrowband light sources with different wavelengths and corresponding photodetectors is activated in the circular area inside the toilet seat, forming a ring-shaped sensor network. This sub-step implements the automatic activation of the intelligent sensor network. A multispectral sensor array is integrated into the circular area inside the toilet seat, consisting of six narrowband light sources with different wavelengths and their corresponding photodetectors. When a user uses the toilet, the built-in toilet usage status sensors (such as pressure sensors, infrared human body sensors, and sound detectors) detect the change in status and generate a trigger signal. Upon receiving the trigger signal, the sensor array is automatically activated, forming a complete ring-shaped sensor network. This design ensures that the system operates only when necessary, saving energy, avoiding unnecessary data collection, and protecting user privacy.

[0040] Step S1.2: Establish a timing synchronization mechanism for light source emission and detector reception for the ring sensor network to ensure that the emission timing of each wavelength light source accurately matches the detector receiving window. This sub-step solves the synchronization problem of simultaneous multi-wavelength acquisition. A precise timing control mechanism is used to allocate specific working time windows to each wavelength light source and detector to ensure that the light source emission accurately matches the corresponding detector receiving window. This timing synchronization mechanism uses microsecond-level precise control to effectively avoid interference between different wavelengths and improve the signal-to-noise ratio. At the same time, a dynamic adjustment function is also implemented, which can adaptively adjust the emission intensity and receiving sensitivity of each wavelength according to environmental conditions and excrement characteristics, further optimizing data acquisition quality.

[0041] Step S1.3: Using the multispectral sensor array controlled by the timing synchronization mechanism, spectral reflectance and transmittance data at wavelengths of 450 nm, 550 nm, 650 nm, 850 nm, 950 nm, and 1100 nm are simultaneously collected at a sampling frequency of 120 Hz, recording the intensity variation data for each wavelength. This substep enables high-frequency acquisition of multi-wavelength spectral data. The six specific wavelengths selected have important physiological significance: 450 nm primarily corresponds to the absorption characteristics of bilirubin, 550 nm to hemoglobin, 650 nm to stercobilin, 850 nm to water molecules, and 950 nm and 1100 nm to various organic substances. The high sampling frequency of 120 Hz enables the capture of rapid changes in fecal status, particularly the dynamic changes in fecal transparency under the effects of laxatives. Simultaneously recording the intensity variations of both reflected and transmitted light for each wavelength provides comprehensive information on the optical properties of the feces, laying the foundation for subsequent transmittance calculation and feature extraction.

[0042] Step S1.4: Timestamp and package the light intensity change data of each band, and transmit it to the local processing unit via the low-power Bluetooth protocol to generate the multi-dimensional spectral data stream. This sub-step completes the marking, packaging and transmission of the original data. Accurately timestamp the collected light intensity change data of each band to ensure the time sequence and traceability of the data. The marked data is then structured and packaged, and an efficient data compression algorithm is used to reduce the transmission burden. The data transmission adopts the low-power Bluetooth 5.0 protocol, which not only ensures that the transmission rate meets the real-time requirements of the 120Hz sampling frequency, but also minimizes energy consumption and extends the battery life of the device. After receiving the transmitted data packet, the local processing unit performs a preliminary integrity check and error correction, and finally forms a structured multi-dimensional spectral data stream to prepare for subsequent data processing and analysis.

[0043] Furthermore, in step S1.1, based on the trigger signal detected by the toilet usage status sensor, a multispectral sensor array comprising six narrowband light sources of different wavelengths and corresponding photodetectors is activated in the annular area inside the toilet seat, forming an annular sensor network, including:

[0044] Step S1.1.1: Based on the toilet usage status sensor, detect the status change signal of the user using the toilet and generate a trigger signal; wherein the toilet usage status sensor includes a pressure sensor, an infrared human body sensor and a sound detector. This sub-step implements an intelligent trigger mechanism. A high-precision pressure sensor is installed between the toilet seat and the base, which can detect the pressure changes generated when the user sits down; an infrared human body sensor is installed on the inner wall of the toilet, which can detect the changes in thermal radiation when excrement enters the toilet; at the same time, a sound detector is installed around the toilet, which can identify the characteristic sounds during the excretion process. These three sensors form a multimodal fusion detection network. Through the comprehensive analysis of the three sensor signals, it can accurately determine whether the user is using the toilet and whether excrement has entered the toilet. This multimodal fusion detection method greatly improves the accuracy of the trigger, effectively avoids false triggering and missed triggering, and also takes into account the differences in usage habits of different users.

[0045] Step S1.1.2: The trigger signal is conditioned and digitized. The microcontroller determines the trigger condition and generates a sensor array activation command. This substep completes the trigger signal processing and decision-making. First, the raw analog signals from the pressure sensor, infrared human body sensor, and sound detector are conditioned, including amplification, filtering, and stabilization, to eliminate environmental interference and random noise. The conditioned signals are then converted to digital signals using a high-precision ADC (analog-to-digital converter) and input into a low-power microcontroller. The microcontroller then runs a trigger determination algorithm based on a Bayesian decision model, comprehensively analyzing the digital signals from the three sensors and calculating the probability distribution of the usage state. When the probability of the usage state exceeds a preset threshold (typically set at 95%), the microcontroller generates a sensor array activation command, initiating multispectral sensing. This probabilistic trigger determination mechanism ensures high sensitivity while avoiding frequent false triggers, optimizing energy consumption and user experience.

[0046] Step S1.1.3: Based on the sensor array activation instruction, the LED narrowband light source and silicon photodiode detector are simultaneously activated at 12 evenly distributed positions in the annular area inside the toilet seat, where the wavelengths of the light sources are 450nm, 550nm, 650nm, 850nm, 950nm and 1100nm, respectively, and each wavelength corresponds to two light source-detector pairs. This sub-step achieves the precise deployment and activation of the multispectral sensor array. LED narrowband light sources and silicon photodiode detectors are installed at 12 evenly distributed positions in the annular area inside the toilet seat. Each wavelength corresponds to two light source-detector pairs, which are distributed in symmetrical positions in the annular area, ensuring all-round coverage of excrement. The LED light source uses narrowband filter technology to ensure the purity of the emission spectrum, and the full width at half maximum (FWHM) is controlled within ±5nm. The silicon photodiode detector is equipped with a corresponding narrowband filter, which only receives light signals of specific wavelengths, effectively suppressing interference from ambient light and other wavelengths. Upon receiving an activation command, the microcontroller simultaneously activates all LED light sources through a precise current control circuit and activates the corresponding detector receiving circuits, forming a complete multi-spectral sensing network. This circular distribution design ensures accurate spectral data is obtained regardless of the position of excrement in the toilet.

[0047] Step S1.1.4: The light source-detector pairs are used to establish an optical transmission channel through the excreta. Each silicon photodiode detector synchronously receives the transmitted light signal of the corresponding wavelength, forming the ring sensing network. This substep establishes a stable optical measurement channel. Each light source-detector pair is precisely aligned, forming a direct optical transmission channel. When excreta enters the toilet, these light paths pass through the excreta. During propagation, the light signal is selectively absorbed and scattered by excreta components, carrying rich information about its composition. The silicon photodiode detectors feature a high-sensitivity design with a signal-to-noise ratio exceeding 60dB, enabling them to accurately capture even subtle changes in the transmitted light signal. Using synchronous detection technology, each detector only collects data during the period when the light source of the corresponding wavelength is emitting, effectively avoiding crosstalk. The current signal output by the detector is converted to a voltage signal by a transimpedance amplifier and then to a digital signal by a 16-bit high-precision ADC, preserving the subtle signal variations. This high-precision optical path design and signal acquisition scheme ensures accurate measurement of the light transmission characteristics of excreta at different wavelengths, providing reliable data for subsequent analysis.

[0048] Step S1.1.5: Perform a self-test on the ring sensor network to verify the operating status of each sensor and the optical path, ensuring the reliability of data collection. This sub-step implements the system's self-diagnostic function. After each activation, a comprehensive self-test begins, consisting of three key steps: First, electrical parameter testing, measuring the drive current and operating voltage of each LED light source to ensure they are within the design range; second, optical path patency verification, collecting baseline light intensity data before excrement enters and comparing it with stored standard values ​​to determine whether the optical path is blocked by foreign objects or optical components are contaminated; third, detector response testing, measuring the dark current and photoelectric response of each detector to ensure that its sensitivity meets requirements. If any anomalies are detected during the self-test, a detailed error log will be recorded, and the type of anomaly will determine whether to continue operation or issue a maintenance reminder. Only after all test items pass will the system enter the formal data collection state. This rigorous self-test mechanism greatly improves the system's reliability, ensures the quality of each data collection, and reduces false positives caused by hardware issues.

[0049] In step S2, the multidimensional spectral data stream is normalized to obtain processed data, the processed data is mapped to a segmented ladder computing architecture, and parallel noise reduction and feature enhancement processing is performed through a spiking neural network to extract the fecal transmittance feature matrix, including:

[0050] Step S2.1: Zero-point calibration and ambient light interference elimination are performed on the spectral data of each band in the multidimensional spectral data stream. Relative intensity and differential data are calculated to obtain normalized spectral data. This substep standardizes the raw data. First, zero-point calibration is performed. Using the baseline light intensity value B0(λ) collected in the absence of feces as a reference point, the relative intensity R'(λ) = R(λ) / B0(λ) is calculated for the raw reading R(λ) at each wavelength λ to eliminate sensitivity differences between sensors. R'(λ) represents the ratio of light intensity relative to the baseline value. It is a dimensionless normalized value used to calibrate for system measurement errors and sensor characteristic differences. Its value is typically between 0 and 1, with 1 indicating maximum transmittance (complete transparency). Next, a technique called adjacent time window differencing is used to address ambient light interference. Taking advantage of the fact that ambient light changes little over a short period of time, differential data D(λ,t) = R'(λ,t) - R'(λ,t-Δt) is calculated, effectively removing the influence of slowly varying ambient light. The processed data is then amplitude-normalized, mapping all wavelengths to the interval [0, 1] to produce normalized spectral data N(λ, t), where t represents the time variable in seconds, representing the time at which the spectral data was collected. This multi-step normalization process ensures comparability of data collected under different environmental conditions and improves robustness in changing environments. This approach is particularly effective in eliminating external interference and ensuring data quality, particularly in the variable lighting conditions of a home environment.

[0051] Step S2.2: The frequency domain analysis results of the normalized spectral data are divided into three processing levels: high frequency band, medium frequency band, and low frequency band, and mapped to a piecewise ladder computing architecture. This sub-step realizes multi-scale decomposition and mapping of the data. First, a fast Fourier transform (FFT) is performed on the normalized spectral data N(λ,t) to convert the time domain signal into a frequency domain representation F(λ,f), where f represents the frequency variable in Hertz (Hz), reflecting the intensity distribution of different frequency components in the signal, and is a quantitative representation of the periodic changes of the spectral data in the time dimension. Then, according to the frequency characteristics, the data is divided into three frequency bands: the high frequency band (>10Hz) mainly contains noise and fast-changing information; the medium frequency band (1-10Hz) contains the main characteristic information of the changes in the state of excrement; and the low frequency band (<1Hz) contains slowly changing background information. An innovative piecewise ladder computing architecture is adopted, and its design is inspired by the information processing mode of the human visual cortex. In this architecture, the bottom layer (high frequencies) utilizes a large number of simple processing units for parallel processing, the middle layer (intermediate frequencies) utilizes a moderate number of more complex processing units, and the top layer (low frequencies) utilizes a smaller number of highly complex processing units. This architectural design aligns computing resource allocation with signal characteristics, giving high-frequency signals more parallel processing resources while lower-frequency signals receive more complex processing capabilities, achieving an optimal balance between computing efficiency and processing accuracy.

[0052] Step S2.3: Deploy a Spiking neural network of the Leaky Integrate-and-Fire neuron model on the segmented ladder architecture, adopt fast response parameters and local connection mode for the high frequency band, adopt a balanced time constant and sparse connection structure for the mid-frequency band, and adopt a long time constant and global connection mode for the low frequency band. This sub-step implements a biologically inspired neural network processing mechanism. A Spiking neural network (SNN) is deployed on the segmented ladder architecture, which is a third-generation neural network that simulates the working mode of biological neurons. SNN adopts the Leaky Integrate-and-Fire (LIF) neuron model, and its membrane potential dynamic equation is τ(dV / dt) = -(VV rest ) + RI(t), where τ is the membrane time constant, which indicates how quickly the neuron responds to input changes; V is the neuron membrane potential; V rest is the resting potential, i.e., the stable potential when there is no external input; R is the membrane resistance; I(t) is the external input current, representing the signal strength received by the neuron. When the membrane potential V exceeds the threshold V th When the signal is received, the neuron fires a pulse and resets. At different frequency band levels, SNN uses different parameter configurations and connection topologies: the high-frequency layer uses fast-response parameter settings (τ is small, about 5ms) and local connection mode to achieve fast parallel noise reduction; the medium-frequency layer uses a balanced time constant (τ is about 20ms) and a sparse connection structure, focusing on extracting the changing characteristics of the light transmittance of excrement; the low-frequency layer uses a long time constant (τ is large, about 50ms) and a global connection mode to capture slowly changing background trends. Cross-level inhibitory connections are also implemented, allowing high-level processing units to modulate the activities of low-level units, forming a top-down attention mechanism. This multi-level SNN architecture can simultaneously process signal components with different time scales and frequency characteristics, significantly improving the ability to extract spectral features of excrement.

[0053] The spiking neural network was trained using a supervised learning method based on time encoding. A training set was constructed using spectral data from standard samples of varying concentrations and compositions. During training, the network's time constant, τ, was set to 5ms for the high-frequency layer, 20ms for the mid-frequency layer, and 50ms for the low-frequency layer. The model updated its weights using a combination of spike timing-dependent plasticity (STDP) and error backpropagation. The Adam optimizer was used, with an initial learning rate of 0.001 and a cosine annealing strategy. The model was trained for 100 epochs on 5,000 standard samples and achieved a signal recovery accuracy exceeding 95% on 1,000 independent validation samples.

[0054] Step S2.4: Through the multi-level parallel processing of the Spiking neural network, the transmittance index of the excrement in each band is calculated, and the excrement transmittance feature matrix containing three-dimensional information of time, wavelength and transmittance is constructed. This sub-step completes the extraction of transmittance features and matrix construction. After the SNN processing is completed, the enhanced and noise-reduced multi-band spectral signal is obtained. The transmittance is calculated based on the improved Beer-Lambert law. For wavelength λ, incident light intensity I0(λ) and transmitted light intensity I(λ), the transmittance T(λ) is calculated as T(λ) = I(λ) / I0(λ) = , where α(λ) is the absorption coefficient of excrement at that wavelength, and d is the optical path length. The optical path length is first calibrated based on the relative position of the sensors. Then, the above formula is applied to the SNN-processed signals at each wavelength to calculate the real-time transmittance T(λ,t). To enhance feature representation, a series of derived features are also calculated, including the wavelength difference index (WDI) = T(λ1) / T(λ2), which reflects the ratio of transmittance at different wavelengths; the transmittance variation rate (TCV) = dT(λ,t) / dt, which reflects the rate of transmittance change over time; and the transmittance stability index (TSI), which reflects the degree of transmittance fluctuation within a short time window. All these features are organized into a three-dimensional feature matrix M with the dimensions [time × wavelength × feature type]. This structured feature matrix preserves the correlation between time, wavelength, and multiple features, providing a comprehensive data foundation for subsequent dynamic curve analysis. In this way, a highly informative transmittance feature matrix is ​​extracted from the raw multispectral data, achieving data dimensionality reduction and information enhancement.

[0055] Furthermore, in step S2.1, zero-point calibration and ambient light interference elimination are performed on the spectral data of each band in the multidimensional spectral data stream, and relative intensity and differential data are calculated to obtain normalized spectral data, including:

[0056] Step S2.1.1: Based on the baseline values ​​for each wavelength collected in the no-fecal-discharge state, the raw readings for each wavelength in the multidimensional spectral data stream are zero-calibrated to calculate relative intensity values. This substep calibrates inter-sensor differences. During the initialization phase before each use, baseline light intensity values ​​(B0(λ)) for each wavelength are collected in the toilet bowl in a no-fecal-discharge state as calibration reference points. These baseline light intensity values ​​reflect the characteristic differences of each light source-detector pair under ideal light transmission conditions. When actual measurement begins, the relative intensity is calculated using the formula R'(λ) = R(λ) / B0(λ) for the raw readings R(λ) at each wavelength λ. This ratio calculation method effectively eliminates errors caused by factors such as differences in luminous intensity between different light sources, differences in detector sensitivity, and minor changes in optical path geometry. This dynamic calibration method ensures continuous measurement accuracy, especially for issues such as LED light source aging and detector performance drift that may occur during long-term use. The baseline value library is regularly updated to track and record performance trends of each sensor, providing a basis for equipment maintenance.

[0057] Step S2.1.2: Apply the adjacent time window differencing technique to the relative intensity values, leveraging the fact that ambient light changes slowly over short periods of time to eliminate the effects of slowly varying ambient light, thereby generating differential data. This substep addresses the issue of ambient light interference. Despite the use of narrowband filters, ambient light (particularly sunlight and fluorescent lighting) can still affect measurement results through scattering and other means. Taking advantage of the generally slow variations in ambient light, the adjacent time window differencing technique is employed. Specifically, a time window of length Δt (typically 0.5 seconds) is defined, and differential data D(λ,t) = R'(λ,t) - R'(λ,t-Δt) is calculated. Because fecal matter typically changes much faster than ambient light, this differencing method effectively preserves rapid signal changes caused by feces while suppressing slow background drift caused by ambient light. An adaptive window length algorithm is also employed to dynamically adjust Δt based on the rate of signal change. Shorter windows are used for faster signal changes, while longer windows are used for slower signal changes to enhance interference rejection. This flexible differential processing strategy enables the system to adapt to various complex ambient light conditions and ensure measurement stability.

[0058] Step S2.1.3: Based on the differential data, amplitude normalization is performed, and the data of all wavelengths are mapped to a standardized numerical range to generate standardized spectral data. This sub-step realizes a unified representation of the data. Due to differences in physical properties, the original amplitude ranges of optical signals of different wavelengths may vary greatly. Direct comparison and processing of these heterogeneous data will cause high-amplitude signals to dominate the analysis results. To solve this problem, the differential data is amplitude normalized, and the improved Min-Max normalization method is used to map the data of each wavelength λ to the [0,1] interval: N(λ,t) = (D(λ,t) - D min (λ)) / (D max (λ) - D min (λ)), where D min (λ) and D max (λ) represents the typical minimum and maximum values ​​for each wavelength, determined in advance through extensive historical data statistics. A dynamic boundary adjustment mechanism has also been designed. When data outside the preset range is detected, the boundary values ​​are automatically updated and the normalization results are recalculated, ensuring that even abnormal data is properly processed. This normalization process makes data from different wavelengths comparable, facilitating subsequent multi-wavelength joint analysis and facilitating efficient training and inference of algorithms such as neural networks.

[0059] Step S2.1.4: Perform a fast Fourier transform on the standardized spectral data to convert the time-domain signal into a frequency-domain representation. The signal is then divided into high-frequency, mid-frequency, and low-frequency bands based on their frequency characteristics. This substep implements frequency-domain decomposition of the signal. A fast Fourier transform (FFT) algorithm is applied to the standardized time-series data at each wavelength to convert it into a frequency-domain representation, F(λ, f), revealing the various frequency components within the signal. Based on extensive clinical data analysis and signal characteristics research, the spectrum is divided into three key frequency bands: the high-frequency band (>10 Hz) primarily contains random noise, electrical interference, and rapid physiological fluctuations; the mid-frequency band (1-10 Hz) contains the primary characteristic information of changes in fecal status and is the core frequency band for assessing bowel readiness; and the low-frequency band (<1 Hz) contains slowly varying background trends and baseline drift. A detailed analysis of the spectral characteristics of different wavelengths revealed that certain wavelengths (e.g., 650 nm) exhibit more prominent characteristics in the mid-frequency band, while others (e.g., 850 nm) contain important information in the low-frequency band. Based on these findings, the optimal frequency band division boundaries were customized for each wavelength, further improving the accuracy of signal decomposition. This frequency domain analysis provides a scientific basis for subsequent segmented ladder architecture mapping, ensuring that computing resources are focused on the frequency bands with the highest information content.

[0060] Step S2.1.5: De-noise the high-frequency band, the mid-frequency band, and the low-frequency band using a combination of Gaussian and median filtering, preserving valid signal features while suppressing random noise, thereby obtaining the normalized spectral data. This substep completes the frequency-domain noise reduction process. Customized filtering strategies are employed based on the characteristics of different frequency bands: For the high-frequency band, Gaussian filtering is primarily applied to smooth random spike noise, with a filter kernel width of 3 sampling points to preserve rapidly varying valid signals. For the mid-frequency band, a cascade combination of Gaussian and median filtering is employed, first using a median filter (with a window size of 5 sampling points) to remove outliers, followed by a Gaussian filter for smoothing, maximizing the preservation of characteristic information. For the low-frequency band, median filtering is primarily applied with a window size of 9 sampling points to effectively suppress long-period interference. An adaptive filter parameter adjustment mechanism is also implemented, dynamically adjusting the filter strength based on real-time signal-to-noise ratio evaluation results, increasing the filter strength when noise is high and reducing the filter strength when the signal is clear to preserve more detail. After these meticulous noise reduction processes, the signal quality of each frequency band is significantly improved, with noise levels reduced by approximately 85% while retaining over 95% of the effective signal features. Finally, the processed signals of each frequency band are recombined to generate high-quality normalized spectral data, laying a solid foundation for subsequent feature extraction.

[0061] like Figure 3 As shown in step S2.3, the Spiking neural network adopts the Leaky Integrate-and-Fire neuron model, and its membrane potential dynamic equation is: τ(dV / dt) = -(VV rest ) + RI(t), where τ is the membrane time constant, which indicates how quickly the neuron responds to input changes; V is the neuron membrane potential; V rest is the resting potential, i.e., the stable potential when there is no external input; R is the membrane resistance; I(t) is the external input current, representing the signal strength received by the neuron. When the membrane potential V exceeds the threshold V th When the neurons fire, they reset. The network employs different parameter configurations and connection topologies within a segmented ladder architecture: The high-frequency layer uses fast-response parameter settings (small τ) and a local connection pattern to achieve rapid parallel noise reduction; the mid-frequency layer employs a balanced time constant and sparse connection structure to focus on extracting the changing characteristics of fecal transmittance; and the low-frequency layer employs a long time constant (large τ) and a global connection pattern to capture slowly changing background trends.

[0062] like Figure 4 As shown, in step S3, based on the feces transmittance feature matrix, a Hopfield network model is constructed and a nonlinear attention mechanism is implemented to adaptively focus on time periods and wavelength intervals where transmittance changes significantly, and generate a dynamic curve of feces transmittance and key change points, including:

[0063] Step S3.1: Based on the fecal transmittance characteristic matrix, a 20-second time window sliding technique and weighted averaging are used to reconstruct the continuous time series variation of fecal transmittance. This substep achieves smooth reconstruction of the time series data. First, the time and wavelength dimensions are extracted from the transmittance characteristic matrix M to form a two-dimensional time series representation. To address discontinuities and random fluctuations that may occur during the sampling process, a sliding time window technique is used for reconstruction. Specifically, a 20-second time window is defined, sliding along the time axis in 2-second steps. For each time point within the window, transmittance data for all wavelengths are extracted. A weighted average is calculated using a Gaussian kernel weighting function G(t-t0) = exp(-(t-t0)² / 2σ²), with the center of the window as the highest weight. Here, σ is set to 1 / 4 of the window length (i.e., 5 seconds) to ensure a smooth transition. This weighted averaging effectively eliminates the effects of random noise and instantaneous fluctuations while preserving the main trends in transmittance variation. For key wavelengths (such as 550nm, 650nm, and 850nm, corresponding to the absorption peaks of hemoglobin, bilirubin, and water, respectively), a higher sampling density and weighting coefficient are used to ensure that the characteristics of these wavelengths are fully expressed. Through this time window reconstruction, continuous transmittance time series data T'(t,λ) with high temporal resolution (up to 200ms) is generated, providing structured input for the Hopfield network.

[0064] Step S3.2: Construct an energy function containing time dependency, wavelength correlation and stability terms for the continuous time series change data, and use the contrast divergence algorithm to train the Hopfield network model. This sub-step establishes a network model based on the energy function. The modern Hopfield network is used as the core modeling tool. This network is a modern improved version of the classic Hopfield network with stronger pattern storage capacity and dynamic characteristics capture ability. A special energy function E(v) is designed, which contains three key components: time dependency E t (v) Capture the transfer relationship of transmittance at adjacent time points and model the continuity and trend of transmittance over time; wavelength-related term E λ (v) Modeling the mutual constraints between transmittances at different wavelengths to reflect the spectral characteristics of fecal components; the stability term E s (v) Encourages the transmittance to stabilize over time, which is consistent with the clinical expectation that the feces will eventually become clear and stable during bowel preparation. The complete energy function is expressed as E(v) = α·E t (v) +β·E λ (v) + γ·E s(v), where α, β, and γ are weight coefficients for balancing the contributions of each factor, determined through grid search optimization. The network is trained using the contrastive divergence algorithm, utilizing historical patient data from different bowel preparation stages to construct a priori knowledge base. After training, the network is able to recover dynamic transmittance patterns consistent with physical laws and clinical characteristics from noisy transmittance data through the principle of energy minimization, providing a theoretical basis and computational framework for implementing nonlinear attention mechanisms.

[0065] Step S3.3: Based on the Hopfield network model, nonlinear transformation functions of transmittance change, local fluctuation and information entropy are designed to calculate the comprehensive attention weight of temporal attention and wavelength attention. This sub-step implements an adaptive attention mechanism. Based on the Hopfield network, an innovative nonlinear attention mechanism is implemented to adaptively focus on time periods and wavelength intervals with significant transmittance changes. The core of the attention weight function A(t,λ) is the nonlinear transformation function f, which comprehensively considers three key indicators: the transmittance change ΔT(t,λ) reflects the absolute change in transmittance; the local fluctuation σ(t,λ) represents the degree of fluctuation within a short time window; and the information entropy I(t,λ) quantifies the uncertainty and information content of the signal. The nonlinear transformation function adopts an improved Softmax form: A(t,λ) = softmax(f(ΔT(t,λ), σ(t,λ),I(t,λ))), where the f function is implemented through a deep neural network and optimized through training with a large amount of clinical data. First, the temporal attention A is calculated. t (t), focus on the time period with the largest rate of change of transmittance; then calculate the wavelength attention A λ (λ), focusing on the wavelength that is most sensitive to the characteristics of excrement; finally, calculate the comprehensive attention A(t,λ) = A t (t)·A λ (λ), achieving joint attention in both time and wavelength dimensions. A dynamic thresholding technique is also introduced to adaptively adjust the attention trigger threshold θ(t) = μ + k·σ(t) based on historical data distribution, where μ is the historical mean, σ(t) is the time-varying standard deviation, and k is an adjustable coefficient. This nonlinear attention mechanism accurately locates the critical moments and wavelengths of transmittance changes, filtering out redundant information and concentrating computing resources on the data areas with the greatest diagnostic value.

[0066] Step S3.4: Perform weighted averaging on the continuous temporal variation data using the comprehensive attention weight, apply B-spline interpolation technology to generate a smooth and continuous dynamic curve of the transmittance, and identify inflection points, stable platforms, and rapid change points as the key change points through differential geometry and cluster analysis. This sub-step completes the dynamic curve generation and key point identification. Based on the comprehensive attention weight A(t,λ), perform weighted averaging on the continuous temporal variation data T'(t,λ), and calculate the attention weighted average transmittance:

[0067]

[0068] This weighted average ensures that the high-attention area has a dominant role in the curve shape, highlighting the changing features with diagnostic value. To enhance the interpretability and smoothness of the curve, the cubic B-spline interpolation technique is applied to generate a smooth and continuous transmittance dynamic curve. (t), while retaining the key features of the original data. After the curve is generated, the following types of feature points are automatically identified: Inflection Points are points where the second-order derivative of the curve is zero, representing the turning point of the transmittance change rate; Stable Plateaus are intervals where the transmittance change rate is close to zero and lasts for a certain period of time (usually ≥30 seconds); Rapid Change Points are points where the transmittance change rate exceeds a preset threshold (usually 3 times the average change rate); Target Achievement Point is the time point when the transmittance first reaches the target value (usually 90% of the final stable value). Differential geometry methods are used to calculate the first-order derivative of the curve (d / dt) and the second derivative (d² / dt²), an extreme value detection algorithm is used to identify inflection points; a sliding window variance analysis is used to identify stable platform; and threshold comparison is used to determine points of rapid change and target achievement. These key change points, along with the complete transmittance dynamic curve, provide a comprehensive quantitative basis for the quantitative assessment of bowel preparation.

[0069] Furthermore, in step S3.1, based on the feces transmittance characteristic matrix, continuous time series variation data of feces transmittance is reconstructed through a 20-second time window sliding technique and weighted average processing, including:

[0070] Step S3.1.1: Based on the fecal transmittance characteristic matrix, a time window of 20 seconds is defined, and the time axis is slid in steps of 2 seconds to extract the transmittance data at each time point in the window. This sub-step realizes the structured extraction of time series data. First, two-dimensional slices of the time and wavelength dimensions are extracted from the three-dimensional transmittance characteristic matrix M to form wavelength-time transmittance data T(t,λ), where T(t,λ) represents the fecal transmittance value measured at time point t and wavelength λ. Taking into account the possible unevenness and random missing of the original sampling, the sliding window technology is used for data reorganization. Specifically, a time window W(t0,Δt) of 20 seconds is defined, where t0 is the center time point of the window and Δt is the half-length of the window (10 seconds). The window slides along the time axis with a fixed step size of 2 seconds, covering the entire monitoring process. For each window position, the wavelength-time transmittance data T(t,λ) for all time points t∈[t0-Δt, t0+Δt] within the window are extracted to form a local dataset D(t0) = {T(t,λ)| t∈[t0-Δt, t0+Δt], λ∈Λ}, where Λ represents the set of all monitored wavelengths. A 20-second window length is the optimal value determined through extensive clinical testing, balancing the needs of suppressing short-term fluctuations and preserving long-term trends. A 2-second sliding step ensures sufficient temporal resolution while avoiding excessive consumption of computational resources. This sliding window technique provides a structured data foundation for subsequent weighted averaging, enabling effective processing of noise and outliers while preserving temporal continuity.

[0071] Step S3.1.2: Apply a Gaussian kernel weighting function with the center of the window as the highest weight to the transmittance data, calculate the weighted average transmittance value, eliminate the influence of random noise, and obtain the discrete transmittance data points after weighted averaging. This sub-step implements position-based weighted averaging. Apply a Gaussian kernel weighting function to the transmittance data in each time window for averaging. The Gaussian kernel function is defined as G(t,t0) = exp(-(t-t0)² / 2σ²), where t0 is the time point at the center of the window, and σ is the parameter that controls the width of the weight distribution, which is set to 1 / 2 of the half-length of the window (i.e. 5 seconds). This design allows data points close to the center of the window to obtain higher weights, while the weights of data points away from the center gradually decrease, forming a smooth bell-shaped weight distribution. For each wavelength λ in the window, calculate the weighted average transmittance ( t0,λ) = ∑ t G(t, t0)·T(t,λ) / ∑ tG(t, t0). This weighted averaging process has multiple advantages: first, it can effectively suppress random noise and short-term fluctuations. During the averaging process, the noise cancels each other out while the signal accumulates and strengthens. Second, the smooth weight distribution ensures the continuity between the processed data points, avoiding mutations and unnatural jumps. Third, the window overlap design (the step size is smaller than the window length) further enhances the coherence between adjacent processed points. After this process, a series of discrete transmittance data points after weighted averaging are obtained. (t0,λ), although these data points are discrete, they have significantly reduced the impact of noise and laid the foundation for subsequent interpolation processing.

[0072] Step S3.1.3: Targeting the absorption peaks of hemoglobin, bilirubin, and water at key wavelengths of 550nm, 650nm, and 850nm, higher sampling density and weighting coefficients are used for processing. This substep achieves refined processing of key wavelengths. Special attention is paid to specific wavelengths that correspond to characteristic absorption peaks of key components in excreta: 550nm corresponds to the primary absorption peak of hemoglobin, a key indicator for assessing blood content in excreta; 650nm corresponds to the characteristic absorption of bilirubin, reflecting the presence of hepatobiliary metabolites; and 850nm corresponds to the near-infrared absorption band of water molecules, indicating the water content and dilution of excreta. A differentiated processing strategy was employed for these three key wavelengths: first, the sampling density was increased by reducing the original 2-second sliding step size to 0.5 seconds to achieve higher temporal resolution. Second, the weighting coefficients were increased by introducing a wavelength-specific weighting factor w(λ) based on the Gaussian kernel function. The modified weighting function is G'(t,t0,λ) = w(λ)·G(t,t0), where w(550nm) = 1.5, w(650nm) = 1.8, and w(850nm) = 1.3. These weighting coefficients were determined through extensive clinical data analysis and feature importance assessment. Furthermore, a more rigorous outlier detection and processing mechanism was applied to these key wavelengths, using an improved Z-score method to identify and replace anomalous data points. This refined processing of key wavelengths significantly improved sensitivity to changes in fecal composition, particularly the precise tracking of decreases in bilirubin and hemoglobin levels during bowel preparation, providing a key basis for accurate assessment of bowel preparation.

[0073] Step S3.1.4: Interpolate the weighted averaged discrete transmittance data points using a cubic spline interpolation algorithm to generate a continuous transmittance time series function with a time resolution of 200ms. This sub-step achieves the generation of a high-resolution continuous curve. (t0,λ) applies the cubic spline interpolation algorithm to construct a continuous transmittance time series function (t,λ). Cubic spline interpolation is a piecewise polynomial interpolation method that not only ensures the matching of function values ​​at the interpolation points, but also ensures the continuity of the first-order derivatives and second-order derivatives, producing a natural and smooth curve. The specific algorithm used is the cubic spline interpolation with natural boundary conditions, that is, it is assumed that the second-order derivatives at both ends of the curve are zero. This boundary condition is suitable for the physical properties of transmittance changes. The interpolation process is carried out in two steps: first, construct a piecewise cubic polynomial function Si(t) =ai + bi(t-ti) + ci(t-ti)² + di(t-ti)³, where i represents the interval between adjacent data points; then, determine the coefficients ai, bi, ci, di by solving the linear equations, so that the values ​​of the interpolation function at the data points, the first-order derivatives, and the second-order derivatives all meet the continuity conditions. After the interpolation is completed, the continuous function is interpolated with a uniform time step of 200ms. (t,λ) is sampled to generate high-resolution transmittance time series data. This high-resolution sampling enables the capture of rapidly changing details, especially at critical moments when the excrement state changes significantly. Furthermore, the smoothing properties of cubic spline interpolation suppress potential overfitting, avoid unnatural oscillations and spikes, and ensure the physical plausibility of the curve.

[0074] Step S3.1.5: Based on the Beer-Lambert law, calculate the absorption coefficient and optical path length calibration parameters for each wavelength, perform physical parameter correction on the continuous transmittance time series function, and obtain the continuous time series change data. This sub-step completes the calibration and correction of physical parameters. Based on optical theory, the physical parameter correction of the continuous transmittance time series function is performed to ensure the physical significance and comparability of the data. First, apply the Beer-Lambert law I = I0· , where I is the transmitted light intensity, I0 is the incident light intensity, α is the absorption coefficient, and d is the light path length. Transmittance T = I / I0 = It is directly related to the absorption coefficient and the optical path length. The baseline absorption coefficient α0(λ) for each wavelength λ is determined by standard sample testing, and then the effective optical path length d(t) is estimated based on the toilet geometry and the excrement distribution model. Considering the dynamic changes in liquid volume and distribution during excretion, an adaptive optical path model is used to dynamically adjust d(t) based on the real-time data from the pressure sensor and liquid level sensor. The corrected transmittance is calculated as Tcorr(t,λ) = , where d0 is the standardized reference optical path length. In addition, the effect of temperature on the absorption coefficient is taken into account, and a temperature correction factor fT(T) = 1 + kT·(T-T0) is introduced, where T is the real-time measured excrement temperature, T0 is the reference temperature, and kT is the temperature coefficient. The final corrected transmittance is Tfinal(t,λ) = Tcorr(t,λ)·fT(T). These physical parameter corrections ensure the comparability of measurement data under different conditions, enabling accurate reflection of the true optical property changes of excrement without being disturbed by changes in environmental conditions and measurement conditions. The fully corrected continuous time-series change data provides high-quality input for subsequent Hopfield network modeling.

[0075] like Figure 4 As shown, in step S3.4, the following types of feature points are automatically identified: inflection points (points where the second derivative of the curve is zero, representing the turning point of the transmittance change rate); stable plateaus (stable plateaus), where the transmittance change rate approaches zero and persists for a certain period of time; rapid change points (rapid change points), where the transmittance change rate exceeds a threshold; and target achievement points (target achievement points), when the transmittance first reaches the target value. A combination of differential geometry and cluster analysis is used to identify these feature points, providing a quantitative basis for subsequent bowel preparation assessment.

[0076] In step S4, a feature vector is constructed for the transmittance dynamic curve and key change points, and a sparse MLP model combined with a modular structure optimization algorithm is used for processing to generate a bowel preparation quality assessment result that meets the BOSTON scoring standard, including:

[0077] Step S4.1: Based on the transmittance dynamic curve and key change points, multi-dimensional feature indicators including the final transmittance value, transmittance rising rate, number of stable platforms, curve fluctuation index and similarity with the standard pattern are extracted to construct a feature vector. This sub-step realizes comprehensive feature extraction and construction. Four types of core features are extracted from the transmittance dynamic curve and key change points: final state features, process features, time features and pattern features. The final state features reflect the final clarity of the excrement, including the final transmittance value FT = T(tend), which represents the transmittance level at the end of monitoring; the final transmittance stabilization time ST, which is defined as the duration of the transmittance change rate below the threshold; the wavelength difference index WDI = T(tend,λ1) / T(tend,λ2), which reflects the ratio relationship of transmittance at different wavelengths, with special attention to the 550nm / 850nm ratio, which is particularly sensitive to the relative content of hemoglobin and water. The process characteristics reflect the dynamic change characteristics of the excretion process, including the transmittance rise rate TR = (T(tend) - T(tstart)) / Δt, which represents the overall transmittance change rate; the number and duration of stable platforms SPn and SPd, which represent the interval characteristics of the transmittance curve where the change rate is close to zero; and the curve fluctuation index is:

[0078]

[0079] Where T(ti) represents the actual transmittance value at time point ti, T smooth (ti) represents the smoothed transmittance value at the same time point, and n represents the total number of sampling points within the evaluation interval. Temporal features reflect the temporal distribution pattern of bowel preparation, including the first threshold time (FTT), defined as the moment when transmittance first exceeds 70% of the target value; the inflection point distribution density (IPD), representing the number of inflection points per unit time; and the total duration (TD), representing the total duration from the start of the preparation to reaching a stable state. Pattern features are derived by comparing the transmittance dynamic curve with predefined standard patterns. First, the shape of the transmittance dynamic curve is encoded using Fourier transform and principal component analysis. Then, a similarity matrix (SM) with the standard pattern library is calculated, where each element (SMi) represents the similarity score with the i-th standard pattern. Finally, all features are normalized and dimensionally integrated to construct a unified feature vector (F = [FT, ST, WDI, TR, SPn, SPd, FI, FTT, IPD, TD, SM1, ..., SMk]). This vector comprehensively captures all aspects of transmittance dynamics and provides rich input information for the subsequent scoring model.

[0080] Step S4.2: A sparse MLP model with 4 hidden layers is designed for the feature vector, and a sparse connection structure is established by adopting feature group connection, residual skip connection and dynamic pruning strategy. This sub-step realizes the design and construction of an efficient neural network. A multi-layer perceptron (MLP) is used as the basic model architecture, but a number of innovative designs are introduced to improve efficiency and performance. The overall structure of the network includes: an input layer, the dimension of which is consistent with the dimension of the feature vector, and receives normalized features; a hidden layer, which includes 4 hidden layers, and the number of nodes decreases layer by layer (128-64-32-16), forming a funnel-shaped structure; an output layer, with 9 nodes, corresponding to the 9 scores of the BOSTON scoring standard (0-9 points). Unlike traditional MLP, it adopts three sparse connection strategies: Feature Grouping, which groups related features into groups, and each group of features is only connected to specific hidden layer neurons. For example, the final state features are mainly connected to the first group of hidden neurons, the process features are connected to the second group, and so on. This grouping connection reduces the number of connections by about 60%; Residual Skip Connections, which increase cross-layer connections and allow information to be passed directly from low layers to high layers, alleviating the gradient vanishing problem of deep networks while providing multi-scale feature fusion capabilities; Dynamic Pruning, which gradually removes connections with weights close to zero during training. Ultimately, only about 10-20% of connections are retained in the network, significantly reducing computational complexity and storage requirements. This sparse MLP design not only improves computational efficiency, but also enhances the generalization and interpretability of the model, making it particularly suitable for the high requirements for model transparency and reliability in medical scenarios.

[0081] Step S4.3: Using modularity theory, the sparse MLP model is divided into a transparency module, a stability module, and a timing module. The internal structure of each module is optimized using the information bottleneck theory, and the dependencies between modules are determined based on a directed acyclic graph to obtain the optimized sparse MLP model. This sub-step implements structural optimization based on modularity theory. Modularity theory is introduced to guide network structure design, breaking down complex evaluation tasks into functional modules. Specifically, the sparse MLP model is divided into three core functional modules: a transparency module, which mainly processes features related to fecal clarity, including the final transmittance value and wavelength difference index; a stability module, which focuses on the stability characteristics of the excretion process, including the number of stable platforms and the curve fluctuation index; and a timing module, which focuses on temporal distribution characteristics, including the time to first reach the threshold and the overall duration. The internal structure of each module is optimized based on the information bottleneck theory, which guides the maximum compression of redundant information while retaining task-related information. The optimization process includes: determining the minimum sufficient statistics for each module, that is, the minimum feature set that can retain sufficient discriminative information; designing the optimal information flow path to enable efficient information transmission within the module; and adjusting the number of neurons and connection strength to form the optimal information compression-expression balance in each module. The dependencies between modules are represented by a directed acyclic graph (DAG), which clarifies the information flow path: transparency module → stability module → timing module → output layer, while allowing specific cross-module connections to capture complex feature interactions. This structural optimization based on modular theory significantly improves the interpretability and robustness of the model, enabling the generation of more reliable evaluation results while making it easier for doctors to understand the evaluation process and basis.

[0082] Among them, the training data of the sparse MLP model includes 800 cases of clinical annotated data, each of which contains a dynamic curve of light transmittance and a bowel preparation score jointly assessed by three professional physicians according to the BOSTON standard. The dataset is divided into training set, validation set and test set in a ratio of 7:1:2. The model adopts the cross-entropy loss function and uses a weighted approach to deal with the problem of class imbalance. The training adopts a small batch gradient descent method with a batch size of 32, an initial learning rate of 0.005, and L1 regularization to promote sparsity. The model performs early stopping on the validation set and eventually achieves an overall accuracy of 87% on the test set. Among them, the classification accuracy for critical state (BOSTON score 4-5 points) is 82%, the sensitivity is 85%, and the specificity is 89%, which is better than the 75% accuracy of the traditional visual assessment method.

[0083] Step S4.4: The segmented scores and total score probability distributions of the right colon, transverse colon, and left colon segments are calculated using the optimized sparse MLP model, and the confidence assessment and time series score are combined to generate the intestinal preparation quality assessment results that meet the BOSTON scoring standard. This sub-step completes the generation of the final scoring result. The eigenvectors are processed by the optimized sparse MLP model to generate intestinal preparation quality assessment results that meet the BOSTON scoring standard widely used in clinical practice. The BOSTON score is a standard assessment tool for intestinal preparation quality before colonoscopy, which is divided into 0-9 points. The higher the score, the more adequate the intestinal preparation. First, the segmented scores of the three main segments of the colon are calculated: the right colon segment (right hemicolon) score SR, the transverse colon segment (transverse colon) score ST, and the left colon segment (left hemicolon and sigmoid colon) score SL. The score range for each segment is 0-3 points. The calculation process adopts a probabilistic output method, that is, the probability distribution P(S=s) is calculated for each score value (0-3), and the score with the highest probability is finally taken as the score for that segment. The total score STotal = SR+ST+SL and its probability distribution are also calculated. To enhance the reliability of the assessment, two auxiliary indicators were introduced: confidence assessment C = f(Pmax, H), where Pmax is the maximum probability value and H is the entropy of the probability distribution. A higher confidence level indicates a more reliable assessment result; and time series score TS, calculated by direct analysis of the transmittance time series, serves as a supplementary validation of the MLP model score. The final assessment results include: segmented scores and their probability distribution, total scores and their probability distribution, confidence assessment, and clinical recommendations based on the scores. When the total score is ≥6 and each segment is ≥2 points, the bowel preparation is considered adequate and suitable for colonoscopy. When the total score is <6 or any segment is <2 points, the bowel preparation is considered inadequate and further preparation is required. This comprehensive assessment provides physicians with an objective and quantitative basis to help them make decisions about whether to proceed with colonoscopy, significantly improving the efficiency and quality of bowel preparation.

[0084] Furthermore, in step S4.1, based on the transmittance dynamic curve and key change points, multi-dimensional feature indicators including the final transmittance value, transmittance rising rate, number of stable platforms, curve fluctuation index and similarity with the standard pattern are extracted to construct a feature vector, including:

[0085] Step S4.1.1: Extract final-state characteristics based on the transmittance dynamic curve, including the final transmittance value, the final transmittance stabilization time, and the wavelength difference index, reflecting the final fecal clarity. This substep quantitatively characterizes the final fecal clarity. First, extract the final transmittance value, FT = T(tend), from the transmittance dynamic curve. This value directly reflects the fecal clarity at the end of monitoring and ideally approaches the baseline transmittance of pure water. To ensure measurement stability, the final transmittance is calculated as the average value within a 5-minute window at the end of monitoring, rather than the instantaneous value at a single time point. This averaging effectively eliminates random fluctuations in the final-state measurement. The final transmittance stabilization time ST is also calculated, defined as the duration during which the transmittance change rate |dT / dt| remains below a threshold value, θstable (typically set at 0.5% / minute). Stabilization time is a key indicator for assessing bowel preparation completion; a longer stability time indicates that fecal clarity has been achieved, rather than a temporary improvement. In addition, the wavelength difference index (WDI) is calculated, which includes three key ratios: WDI1 = T(tend, 550nm) / T(tend, 850nm), which reflects the relative content of hemoglobin and water; WDI2 = T(tend, 650nm) / T(tend, 850nm), which reflects the relative content of bilirubin and water; and WDI3 = T(tend, 450nm) / T(tend, 650nm), which reflects the ratio of different pigment components. These wavelength difference indices can capture subtle compositional differences that are difficult to distinguish with the naked eye and are particularly sensitive to residual blood and bile pigments. Combining these final state features into a vector (Fterminal = [FT, ST, WDI1, WDI2, WDI3]) comprehensively characterizes the final clarity of feces, providing a direct basis for assessing the adequacy of bowel preparation.

[0086] Step S4.1.2: Calculate the process characteristics of the transmittance dynamic curve, including the transmittance rise rate, the number and duration of stable platforms, and the curve fluctuation index, to reflect the dynamic change characteristics of the excretion process. This sub-step realizes the feature extraction of the dynamic process of excretion. First, calculate the overall transmittance rise rate TR = (T(tend) - T(tstart)) / Δt, where Δt is the total monitoring time. This indicator reflects the average rate of improvement in the clarity of excrement. In order to capture more detailed change characteristics, the segmented rise rate TRi = (T(ti+1) - T(ti)) / (ti+1 - ti) is also calculated, and the entire process is divided into 3-5 key stages. The rise rate of each stage reflects the efficiency of intestinal preparation in different periods. The stable platform in the transmittance curve is identified by the local minimum detection algorithm, which is defined as the transmittance change rate |dT / dt| continuously lower than the threshold θ plateauThe interval of τmin (usually set to 1% / minute) and the duration of τmin (usually set to 2 minutes) is recorded. The number of stable platforms SPn and the duration of each platform SPdi are recorded. These indicators reflect the stage characteristics of the excretion process and usually correspond to the time window when different batches of laxatives take effect. The curve fluctuation index is also calculated:

[0087]

[0088] Where T smooth The smoothed curve is obtained by applying a Savitzky-Golay filter (window length 5% of the monitoring time, polynomial order 3). The fluctuation index quantifies the degree of irregularity in the curve. A higher fluctuation index generally indicates an unstable excretion process, which may be associated with poor patient compliance or bowel dysfunction. Combining these process features into the vector Fprocess = [TR, TR1, ..., TRk, SPn, SPd1, ..., SPdm, FI] comprehensively captures the dynamic characteristics of the excretion process, providing an important basis for evaluating the quality and efficiency of bowel preparation.

[0089] Step S4.1.3: Extract temporal features based on the key change points, including the time to first threshold, inflection point distribution density, and overall duration, to reflect the temporal distribution pattern of bowel preparation. This substep extracts temporal features. First, the time to first threshold (FTT) is calculated. This is defined as the moment when the transmittance first exceeds the target value, typically set at 70% of the final stable transmittance. This time marks the point at which bowel preparation begins to take significant effect and is an important indicator for assessing the onset of laxative efficacy. Multiple threshold times are also calculated, including FTT50 (reaching 50% of the target value), FTT90 (reaching 90% of the target value), and FTT95 (reaching 95% of the target value), to construct a complete time-response curve. Based on the previously identified inflection points (points where the second derivative of the curve is zero), the inflection point distribution density (IPD) is calculated as Ninflection / Ttotal, where Ninflection is the total number of inflection points and Ttotal is the total monitoring time. The inflection point distribution density reflects the frequency of state changes during the excretion process; a higher density generally indicates greater fluctuations in the excretion process. We further analyzed the temporal distribution patterns of inflection points, calculating the mean μinterval and standard deviation σinterval of the inflection point intervals, as well as the distribution characteristics of the inflection point amplitude (the absolute value of the first-order derivative at the inflection point). We also calculated the overall duration (TD), defined as the total time from the first significant change (the rate of change in transmittance exceeding three standard deviations of the baseline fluctuation) to the attainment of a steady state (meeting the definition of stable duration). This metric reflects the time efficiency of the entire bowel preparation process. These temporal features were combined into the vector Ftime = [FTT, FTT50, FTT90, FTT95, IPD, μinterval, σinterval, TD], which comprehensively captures the temporal distribution characteristics of bowel preparation and provides a quantitative basis for evaluating the time efficiency and patterns of the preparation process.

[0090] Step S4.1.4: Shape encoding of the dynamic transmittance curve is performed using Fourier transform and principal component analysis. A similarity matrix with a predefined standard pattern library is calculated to obtain pattern features. This substep implements pattern matching analysis of the curve shape. The dynamic transmittance curve is first normalized, including time normalization (mapping curves of varying lengths to a uniform [0,1] time interval) and amplitude normalization (mapping transmittance values ​​to the [0,1] interval). A fast Fourier transform (FFT) is then applied to convert the time-domain curve into a frequency-domain representation, retaining the amplitude and phase information of the first k (typically k = 10) principal frequency components to form the spectral feature vector Ffreq. Principal component analysis (PCA) is also applied to reduce the dimensionality of the normalized curve, extracting principal components that explain 90% of the variance (typically 3-5 principal components) to form the shape feature vector F_shape. Maintain a predefined standard pattern library L = {M1, M2, ..., Mp}, containing p typical bowel preparation process patterns, each with its own spectral and shape characteristics. These standard patterns are derived through cluster analysis of extensive clinical data and represent different types of bowel preparation dynamics, such as "rapid monotonic rise," "step rise," "fluctuating rise," and "delayed response." Calculate the similarity between the current curve and each standard pattern to form a similarity matrix SM, where the element calculation formula is:

[0091]

[0092] The similarity calculation uses a weighted combination of cosine similarity and dynamic time warping (DTW) distance. The most similar patterns and their similarity values ​​provide a qualitative judgment of the type of bowel preparation process currently being performed, while the complete similarity matrix serves as the pattern feature vector Fpattern = [SM1, SM2, ..., SM p ], providing high-level feature input based on pattern recognition for subsequent scoring models.

[0093] Step S4.1.5: Normalize and dimensionally integrate the final state, process, time, and pattern features to construct the feature vector. This substep integrates and standardizes multidimensional features. First, normalize each feature to ensure that features of different dimensions and ranges have equal weight in subsequent analysis. For continuous numerical features, the Z-score normalization method is used: Z = (x - μ) / σ, where μ and σ are the historical mean and standard deviation of the feature, respectively. This normalization approximates the feature value distribution to a standard normal distribution N(0,1). For features with clear upper and lower bounds, the Min-Max normalization method is used: X' = (x - xmin) / (xmax - xmin), mapping the feature values ​​to the interval [0,1]. Feature importance is also assessed using a random forest-based feature importance analysis method to calculate each feature's contribution to the final score. Feature selection is performed based on the importance scores, retaining the top 85% of features and removing redundant and low-information features. For highly correlated feature pairs (correlation coefficient |r| > 0.85), only the more important one is retained, further reducing feature dimensionality. Finally, the filtered final state feature Fterminal, process feature Fprocess, time feature Ftime, and pattern feature Fpattern are integrated into a unified feature vector F = [Fterminal, Fprocess, Ftime, Fpattern]. This combined feature vector typically contains 20-30 dimensions and comprehensively captures all aspects of the dynamic changes in transmittance, providing rich and refined input information for the sparse MLP model. A confidence weight wi is also assigned to each feature to reflect the reliability of the feature measurement. High-confidence features are given higher weights in subsequent models.

[0094] In step S4.2, the overall architecture of the sparse MLP model includes: an input layer with the same dimension as the feature vector, which receives normalized features; a hidden layer consisting of four hidden layers, with the number of nodes in each layer decreasing layer by layer (e.g., 128-64-32-16); and an output layer with nine nodes, corresponding to the nine scores of the BOSTON scoring system (0-9). Unlike traditional MLPs, this approach employs three sparse connection strategies: Feature Grouping, which groups related features into groups, connecting each group of features only to specific hidden layer neurons; Residual Skip Connections, which add cross-layer connections, allowing information to be passed directly from lower layers to higher layers; and Dynamic Pruning, which gradually removes connections with near-zero weights during training, ultimately retaining only approximately 10-20% of the connections in the network.

[0095] In step S5, the bowel preparation quality assessment result is transmitted to the doctor's application via the 5G network, and personalized guidance suggestions are pushed to the patient based on the bowel preparation quality assessment result, realizing remote monitoring and intelligent feedback, including:

[0096] Step S5.1: Receive the bowel preparation quality assessment results generated in step S4.4, encrypt the data, and securely transmit them to the cloud server via the 5G network. This substep ensures secure transmission of the assessment results. First, the bowel preparation quality assessment results generated by the assessment calculation module are received, including the segmented scores, total scores, confidence ratings, and time series data. Considering the sensitivity of medical data and privacy protection requirements, multi-layered data security measures are implemented. First, the data is de-identified, removing direct identifiers and retaining only the hash value of the patient ID as an associated identifier. The assessment results are then encrypted using the AES-256 encryption algorithm, and the key is securely distributed using asymmetric encryption. The encrypted data is encapsulated in the standard HL7 FHIR format to ensure compatibility with medical information. Leveraging the high bandwidth and low latency of the 5G network, a secure communication channel is established via the TLS 1.3 protocol to transmit the encrypted data to the cloud server. Segmented checksums and resumable transmission are used during the transmission process to ensure data integrity in the event of network instability. It also enables real-time status monitoring of data transmission, recording transmission start time, completion time, transmission rate, and verification results, ensuring that each assessment result reaches the cloud server securely, completely, and promptly. This multi-layered secure transmission mechanism not only protects patient privacy but also ensures the integrity and reliability of medical data during transmission, complying with healthcare data security regulations such as HIPAA.

[0097] Step S5.2: After receiving the encrypted data and decrypting it, the cloud server pushes the real-time assessment results to the doctor-side application, constructing a monitoring interface that includes historical curves, current status, and predicted completion time. This sub-step implements a visual monitoring interface for the doctor. After receiving the encrypted data, the cloud server first performs authentication and access control checks to confirm the data reception authority. The server then decrypts the data using the securely stored key to restore the original assessment results. The decrypted data is then associated with the historical records in the patient's electronic medical record (EMR) to construct a complete patient bowel preparation profile. The doctor-side application establishes a real-time communication connection with the cloud server via the WebSocket protocol to receive the pushed assessment results. The application features an intuitive monitoring interface with three core components: a historical curve panel displays the complete history of the transmittance dynamic curve, marking key change points and important time nodes, and supporting zooming and time window selection; a current status panel, which displays the latest BOSTON score (total and segmented), confidence indicators, and key characteristic values ​​in a dashboard format, using color coding (red-yellow-green) to visually indicate the degree of readiness; and a forecast panel, which predicts the remaining time to reach a full readiness state (BOSTON score ≥ 6) based on historical data and current trends, and displays the predicted confidence interval. The interface also provides contextual information such as patient basic information, scheduled colonoscopy time, and laxative use history. Physicians can view detailed assessment reports through the interface, including specific values ​​and reference ranges for each characteristic indicator, as well as generated analysis recommendations. This real-time, intuitive monitoring interface enables physicians to remotely monitor patients' bowel preparation progress, allowing them to adjust examination plans or provide additional guidance in a timely manner, significantly improving the efficiency of medical resources.

[0098] Step S5.3: Based on the assessment results and historical data, combined with the medical knowledge base, personalized guidance recommendations are generated for the patient's current bowel preparation status. This substep implements intelligent patient guidance. The current assessment results are first analyzed in conjunction with the patient's historical preparation data and demographic characteristics (such as age, gender, body mass index, and previous bowel surgery). The built-in medical knowledge base is then accessed, which contains bowel preparation guidelines, laxative usage guidelines, solutions to common problems, and personalized recommendations for special populations (such as the elderly and patients with renal insufficiency). Based on comprehensive analysis, targeted guidance is generated, encompassing four main categories: Preparation status assessment, which clearly informs patients of the current adequacy of their bowel preparation and explains the assessment results in plain language; follow-up action recommendations, which provide specific action guidelines based on the current status, such as "continue taking laxatives as planned," "increase fluid intake," "extend preparation time," or "preparation is sufficient, discontinue laxatives"; precautions, which provide specific precautions and solutions for special circumstances identified during the assessment (such as large transmittance fluctuations or poor stability); and schedule adjustments, which advise patients to adjust the remaining laxative schedule or coordinate their examination with the hospital based on the predicted completion time. Natural language generation technology is used to convert these recommendations into clearly structured, easy-to-understand text messages, with language complexity adjusted to suit the patient's education level and comprehension ability. Special circumstances (such as abnormal or low-confidence assessment results) are flagged for physician review to ensure the safety and appropriateness of the recommendations. This personalized guidance significantly improves patient compliance and bowel preparation effectiveness, reducing examination cancellations and rescheduling due to inadequate preparation.

[0099] Step S5.4: Establish a secure, two-way interactive channel between the physician and patient, allowing the physician to adjust the guidance plan based on monitoring results and receive patient feedback. This substep enables real-time interaction between physicians and patients. A secure, two-way communication channel is established between the physician application and the patient mobile app, using end-to-end encryption to ensure secure communication. Physicians can use this channel to perform various interactive operations: view detailed assessment data, including the raw transmittance curve and complete feature analysis; send personalized guidance messages, including text, image, and voice instructions; adjust automatically generated recommendations to modify or supplement the guidance plan based on the patient's specific situation; and set reminders and alerts to automatically notify the physician when the patient's status reaches specific conditions (such as transmittance meeting standards or abnormal fluctuations). Patients can use the mobile app to: receive and confirm the physician's guidance recommendations; submit questions and concerns, including text descriptions and image uploads; record the time and dosage of laxatives taken, as well as fluid intake; and report any discomfort or side effects, such as nausea, abdominal pain, or dizziness. Structured question and answer templates are also provided to guide patients in providing standardized feedback, facilitating rapid physician assessment. To improve communication efficiency, intelligent message classification and prioritization are implemented, prioritizing urgent issues (such as severe discomfort symptoms) to physicians. Complete communication records are also maintained and archived as part of the medical documentation. This two-way interactive mechanism significantly improves the efficiency of doctor-patient communication, enabling physicians to promptly adjust guidance plans based on patients' actual situations and feedback. It also enhances patients' sense of engagement and security, improving the overall medical experience and the quality of bowel preparation.

[0100] like Figure 5 As shown, the present invention also provides an intelligent intestinal monitoring system based on multispectral sensing technology, comprising:

[0101] The multispectral sensor module 10 is used to activate the multispectral sensor array integrated in the toilet seat based on the trigger signal of the toilet usage status sensor, and simultaneously collect spectral reflection and transmission data in the wavelength range of 450nm-1100nm at a sampling frequency of 120Hz to generate a multidimensional spectral data stream;

[0102] a data processing module 20 for normalizing the multidimensional spectral data stream to obtain processed data, mapping the processed data to a segmented ladder computing architecture, performing parallel noise reduction and feature enhancement processing via a spiking neural network, and extracting a fecal transmittance feature matrix;

[0103] An intelligent analysis module 30 is configured to construct a Hopfield network model and implement a nonlinear attention mechanism based on the feces transmittance characteristic matrix, adaptively focusing on time periods and wavelength intervals where transmittance changes significantly, and generating a dynamic curve of feces transmittance and key change points;

[0104] An evaluation calculation module 40 is used to construct a feature vector of the transmittance dynamic curve and key change points, and process them using a sparse MLP model combined with a modular structure optimization algorithm to generate a bowel preparation quality evaluation result that meets the BOSTON scoring standard;

[0105] The communication module 50 is used to transmit the intestinal preparation quality assessment result to the doctor-side application through the 5G network, and push personalized guidance suggestions to the patient based on the intestinal preparation quality assessment result to achieve remote monitoring and intelligent feedback.

[0106] The working principle of the present invention is:

[0107] First, the multispectral sensor module 10 collects spectral reflection and transmission data of excrement through the multispectral sensor array integrated in the toilet seat to generate a multidimensional spectral data stream. When the user uses the toilet, the toilet usage status sensor detects the state change, generates a trigger signal, and activates the multispectral sensor array in the annular area inside the toilet seat. The array contains 6 narrow-band light sources with different wavelengths (450nm-1100nm) and corresponding photodetectors to form a ring sensing network. The system establishes a timing synchronization mechanism for light source emission and detector reception, simultaneously collects spectral data of different wavelengths at a sampling frequency of 120Hz, and transmits the data to the local processing unit through the low-power Bluetooth protocol.

[0108] Next, the data processing module 20 normalizes the multidimensional spectral data stream and uses a segmented ladder architecture mapping and spiking neural network for data preprocessing and noise reduction to extract the fecal transmittance feature matrix. Specifically, the system first performs zero-point calibration on the spectral data of each band and eliminates ambient light interference to obtain normalized spectral data. The data is then divided into three processing levels: high frequency band, mid-frequency band, and low frequency band, and mapped to a segmented ladder computing architecture. A spiking neural network based on the Leaky Integrate-and-Fire neuron model is deployed on this architecture, applying different processing strategies to different frequency bands. Ultimately, the transmittance index of the feces in each band is calculated, constructing a feature matrix containing three-dimensional information: time, wavelength, and transmittance.

[0109] Then, based on the transmittance feature matrix, the intelligent analysis module 30 constructs a Hopfield network model and implements a nonlinear attention mechanism to generate a dynamic curve of the fecal transmittance and key change points. The system reconstructs the continuous time-series change data of transmittance through a 20-second time window sliding technique and weighted averaging. An energy function containing time-dependent, wavelength-related, and stability terms is constructed for this data, and the Hopfield network model is trained using a contrast divergence algorithm. Based on this model, the system designs a nonlinear transformation function, calculates the comprehensive attention weight, and generates a smooth and continuous dynamic transmittance curve through weighted averaging and B-spline interpolation technology. The system also identifies inflection points, stable platforms, and rapid change points on the curve as key change points.

[0110] Next, the evaluation and calculation module 40 constructs feature vectors for the transmittance dynamic curve and key change points, and uses a sparse MLP model combined with a modular structure optimization algorithm to quantitatively evaluate the degree of intestinal preparation. The system extracts multidimensional feature indicators including the final transmittance value, transmittance increase rate, and the number of stable platforms to construct a feature vector. Then, a sparse MLP model with 4 hidden layers is designed, and a sparse connection structure is established using feature grouping connection, residual jump connection, and dynamic pruning strategy. The model structure is optimized using modular theory, and finally the segmented scores and total scores of the right, transverse, and left segments of the colon are calculated to generate intestinal preparation quality assessment results that meet the BOSTON scoring standard.

[0111] Finally, the communication module 50 transmits the bowel preparation quality assessment results to the doctor's application in real time via the 5G network, and pushes personalized guidance recommendations to the patient based on the assessment results. The system encrypts the assessment results and securely transmits them to the cloud server via the 5G network. After decryption, the server pushes the real-time assessment results to the doctor and constructs a monitoring interface. Simultaneously, based on the assessment results and historical data, combined with the medical knowledge base, personalized guidance recommendations tailored to the patient's current condition are generated. The system also establishes a two-way interactive channel between doctors and patients, allowing doctors to adjust guidance plans based on monitoring results and receive feedback from patients.

[0112] Through the above working principle, the present invention realizes the objective quantitative evaluation of the intestinal preparation process, solves the technical problems of traditional intestinal preparation evaluation being highly subjective and lacking real-time monitoring, and improves the preparation efficiency and quality of colonoscopy and intestinal surgery.

[0113] The technical innovations of the present invention are mainly reflected in the following aspects:

[0114] 1. Segmented ladder architecture mapping and scheduling Spiking neural network technology: Innovatively maps multispectral sensor data into a ladder computing architecture based on frequency characteristics, and uses a spiking neural network for efficient parallel processing, achieving real-time noise reduction and feature extraction of complex biological signals.

[0115] 2. Nonlinear attention framework based on modern Hopfield network: Combining the improved Hopfield network with the nonlinear attention mechanism, it adaptively focuses on time periods and wavelength ranges with significant transmittance changes, greatly improving the ability to capture the dynamic changes in fecal transmittance.

[0116] 3. Sparse MLP module-based structural optimization algorithm: The module theory is used to guide the structural design of the MLP model. By optimizing the inter-layer connection pattern and weight distribution, efficient feature mapping and score conversion are achieved, significantly improving the accuracy and interpretability of bowel preparation assessment.

[0117] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. An intelligent intestinal monitoring method based on multispectral sensing technology, characterized in that: include: Based on the trigger signal of the toilet usage status sensor, the multispectral sensor array integrated in the toilet seat is activated to simultaneously collect spectral reflection and transmission data in the wavelength range of 450nm-1100nm at a sampling frequency of 120Hz, generating a multidimensional spectral data stream; Normalizing the multidimensional spectral data stream to obtain processed data, mapping the processed data to a segmented ladder computing architecture, performing parallel noise reduction and feature enhancement processing through a spiking neural network, and extracting a fecal transmittance feature matrix; Based on the feces transmittance feature matrix, a Hopfield network model is constructed and a nonlinear attention mechanism is implemented to adaptively focus on time periods and wavelength intervals where transmittance changes significantly, generating a dynamic curve of feces transmittance and key change points; The transmittance dynamic curve and key change points are subjected to feature vector construction, and are processed using a sparse MLP model combined with a modular structure optimization algorithm to generate a bowel preparation quality assessment result that meets the BOSTON scoring standard; The intestinal preparation quality assessment result is transmitted to the doctor-side application via the 5G network, and personalized guidance suggestions are pushed to the patient based on the intestinal preparation quality assessment result to achieve remote monitoring and intelligent feedback.

2. The method according to claim 1, characterized in that The trigger signal based on the toilet usage status sensor activates the multispectral sensor array integrated in the toilet seat, and simultaneously collects spectral reflection and transmission data in the wavelength range of 450nm-1100nm at a sampling frequency of 120Hz to generate a multidimensional spectral data stream, including: Based on the trigger signal detected by the toilet usage status sensor, a multispectral sensor array consisting of six narrow-band light sources with different wavelengths and corresponding photodetectors is activated in the circular area inside the toilet seat to form a ring-shaped sensing network; Establishing a timing synchronization mechanism for light source emission and detector reception in the ring sensor network to ensure that the emission timing of each wavelength light source accurately matches the detector receiving window; The multispectral sensor array controlled by the timing synchronization mechanism simultaneously collects spectral reflectance and transmittance data at wavelengths of 450nm, 550nm, 650nm, 850nm, 950nm and 1100nm at a sampling frequency of 120Hz, and records the light intensity change data of each band; The light intensity change data of each band is timestamped and packaged, and transmitted to a local processing unit via a low-power Bluetooth protocol to generate the multi-dimensional spectral data stream.

3. The method according to claim 1, characterized in that Normalizing the multidimensional spectral data stream to obtain processed data, mapping the processed data to a segmented ladder computing architecture, performing parallel noise reduction and feature enhancement processing through a spiking neural network, and extracting a fecal transmittance feature matrix, including: Performing zero-point calibration and ambient light interference elimination on the spectral data of each band in the multidimensional spectral data stream, calculating relative intensity and differential data, and obtaining normalized spectral data; The frequency domain analysis results of the normalized spectral data are divided into three processing levels: high frequency band, mid frequency band and low frequency band, and mapped to a segmented ladder computing architecture; Deploying a Spiking neural network of the Leaky Integrate-and-Fire neuron model on the segmented ladder architecture, using fast response parameters and a local connection mode for high frequency bands, using a balanced time constant and a sparse connection structure for mid-frequency bands, and using a long time constant and a global connection mode for low frequency bands; Through the multi-level parallel processing of the Spiking neural network, the light transmittance index of the excrement in each band is calculated, and the light transmittance characteristic matrix of the excrement containing three-dimensional information of time, wavelength and light transmittance is constructed.

4. The method according to claim 1, wherein Based on the fecal transmittance feature matrix, a Hopfield network model is constructed and a nonlinear attention mechanism is implemented to adaptively focus on time periods and wavelength ranges where transmittance changes significantly. A dynamic curve of fecal transmittance and key change points are generated, including: Based on the feces transmittance characteristic matrix, the continuous time series change data of the feces transmittance is reconstructed through a 20-second time window sliding technology and weighted average processing; An energy function including a time-dependent term, a wavelength-related term, and a stability term is constructed for the continuous time-series change data, and a Hopfield network model is trained using a contrastive divergence algorithm; Based on the Hopfield network model, a nonlinear transformation function of transmittance variation, local fluctuation and information entropy is designed to calculate the comprehensive attention weight of time attention and wavelength attention; The continuous time-series change data is weightedly averaged by the comprehensive attention weight, and the B-spline interpolation technology is applied to generate a smooth and continuous transmittance dynamic curve, and the inflection point, stable platform and rapid change point are identified as the key change points through differential geometry and cluster analysis.

5. The method according to claim 1, wherein The transmittance dynamic curve and key change points are constructed with feature vectors, and processed using a sparse MLP model combined with a modular structure optimization algorithm to generate a bowel preparation quality assessment result that meets the BOSTON scoring standard, including: Based on the transmittance dynamic curve and key change points, multi-dimensional feature indicators including the final transmittance value, transmittance rising rate, number of stable platforms, curve fluctuation index and similarity with the standard pattern are extracted to construct a feature vector; A sparse MLP model with four hidden layers is designed for the feature vector, and a sparse connection structure is established by adopting feature grouping connection, residual skip connection and dynamic pruning strategy; The sparse MLP model is divided into a transparency module, a stability module, and a timing module using modular sequence theory. The internal structure of each module is optimized using information bottleneck theory, and the dependencies between modules are determined based on a directed acyclic graph to obtain an optimized sparse MLP model. The segmented scores and total score probability distribution of the right colon segment, transverse colon segment and left colon segment are calculated by the optimized sparse MLP model, and the intestinal preparation quality assessment results that meet the BOSTON scoring standard are generated by combining confidence assessment and time series scoring.

6. The method according to claim 2, characterized in that Based on the trigger signal detected by the toilet usage status sensor, a multispectral sensor array consisting of six narrow-band light sources with different wavelengths and corresponding photodetectors is activated in the circular area inside the toilet seat, forming a ring-shaped sensing network, including: Based on the toilet usage status sensor, a state change signal of the user using the toilet is detected and a trigger signal is generated; wherein the toilet usage status sensor includes a pressure sensor, an infrared human body sensor and a sound detector; Performing signal conditioning and digital processing on the trigger signal, determining the trigger condition through a microcontroller and generating a sensor array activation instruction; Based on the sensor array activation command, a narrow-band LED light source and a silicon photodiode detector are simultaneously activated at 12 evenly distributed locations in the annular area inside the toilet seat. The wavelengths of the light sources are 450nm, 550nm, 650nm, 850nm, 950nm, and 1100nm, respectively, with each wavelength corresponding to two light source-detector pairs. An optical transmission channel is established through the light source-detector pair to pass through the excrement, and each silicon photodiode detector synchronously receives the transmitted light signal of the corresponding wavelength, thereby forming the ring sensing network; A self-test is performed on the ring sensor network to verify the working status of each sensor in the ring sensor network and the smoothness of the optical path, thereby ensuring the reliability of data collection.

7. The method according to claim 3, characterized in that The zero-point calibration and ambient light interference elimination of the spectral data of each band in the multidimensional spectral data stream are performed, and relative intensity and differential data are calculated to obtain normalized spectral data, including: Based on the reference values ​​of each wavelength collected in the absence of excrement, zero-point calibration is performed on the original readings of each wavelength in the multidimensional spectral data stream to calculate the relative intensity value; Adopting adjacent time window difference technology for the relative intensity value, eliminating the influence of slowly varying ambient light by taking advantage of the fact that ambient light changes little in a short period of time, and obtaining differential data; Performing amplitude normalization processing based on the differential data, mapping data of all wavelengths to a standardized numerical interval, and generating standardized spectral data; Performing a fast Fourier transform on the standardized spectral data to convert the time domain signal into a frequency domain representation, and dividing the signal into a high frequency band, a mid-frequency band, and a low frequency band according to the frequency characteristics; The high frequency band, the medium frequency band and the low frequency band are subjected to noise reduction processing by combining Gaussian filtering and median filtering, thereby retaining effective signal features while suppressing random noise, and obtaining the normalized spectral data.

8. The method according to claim 4, characterized in that The method of reconstructing the continuous time series variation data of the excrement transmittance based on the excrement transmittance characteristic matrix by using a 20-second time window sliding technique and weighted average processing includes: Based on the feces transmittance characteristic matrix, a time window of 20 seconds is defined, and the transmittance data of each time point within the window is extracted by sliding along the time axis with a step length of 2 seconds; Applying a Gaussian kernel weighting function with the window center as the highest weight to the transmittance data, calculating a weighted average transmittance value, eliminating the influence of random noise, and obtaining discrete transmittance data points after weighted average; A higher sampling density and weight coefficient are used for processing the absorption peaks of hemoglobin, bilirubin and water at the key wavelengths of 550nm, 650nm and 850nm; Interpolating the weighted averaged discrete transmittance data points using a cubic spline interpolation algorithm to generate a continuous transmittance time series function with a time resolution of 200 ms; The absorption coefficient of each wavelength and the optical path length calibration parameter are calculated based on the Beer-Lambert law, and the physical parameter correction is performed on the continuous transmittance time series function to obtain the continuous time series change data.

9. The method according to claim 5, characterized in that Based on the transmittance dynamic curve and key change points, multi-dimensional feature indicators including the final transmittance value, transmittance rising rate, number of stable platforms, curve fluctuation index and similarity with the standard pattern are extracted to construct a feature vector, including: Extracting final state characteristics based on the transmittance dynamic curve, including the final transmittance value, the final transmittance stable time, and the wavelength difference index, to reflect the final clarity of the excreta; Calculating process characteristics of the transmittance dynamic curve, including transmittance rising rate, number and duration of stable platforms, and curve fluctuation index, to reflect the dynamic change characteristics of the excretion process; Extracting time features based on the key change points, including the time to first reach the threshold, the distribution density of inflection points, and the overall duration, to reflect the time distribution pattern of bowel preparation; Performing shape coding on the transmittance dynamic curve by Fourier transform and principal component analysis, calculating a similarity matrix with a predefined standard pattern library, and obtaining pattern features; The final state features, process features, time features and pattern features are normalized and dimensionally integrated to construct the feature vector.

10. An intelligent intestinal monitoring system based on multispectral sensing technology, characterized in that: include: The multispectral sensor module is used to activate the multispectral sensor array integrated in the toilet seat based on the trigger signal of the toilet usage status sensor. It simultaneously collects spectral reflection and transmission data in the wavelength range of 450nm-1100nm at a sampling frequency of 120Hz to generate a multidimensional spectral data stream; a data processing module for normalizing the multidimensional spectral data stream to obtain processed data, mapping the processed data to a segmented ladder computing architecture, performing parallel noise reduction and feature enhancement processing through a spiking neural network, and extracting a fecal transmittance feature matrix; An intelligent analysis module is used to construct a Hopfield network model and implement a nonlinear attention mechanism based on the feces transmittance feature matrix, adaptively focusing on time periods and wavelength intervals where transmittance changes significantly, and generating a dynamic curve of feces transmittance and key change points; An evaluation calculation module is used to construct feature vectors of the transmittance dynamic curve and key change points, and process them using a sparse MLP model combined with a modular structure optimization algorithm to generate a bowel preparation quality evaluation result that meets the BOSTON scoring standard; The communication module is used to transmit the intestinal preparation quality assessment results to the doctor-side application through the 5G network, and push personalized guidance suggestions to the patient based on the intestinal preparation quality assessment results to achieve remote monitoring and intelligent feedback.

Images