A machine learning-based liver disease thyroid function abnormality early warning system

Abstract

The application belongs to the field of data model processing, and discloses a liver disease thyroid function abnormality early warning system based on machine learning; data is collected, and the collected data is marked through prior processing; potential noise data is subjected to self-cleaning processing to obtain verification data; a causal path is pre-constructed according to a false alarm signal scene, a liver-thyroid cross-organ causal network is formed by comprehensively integrating the causal path, and time sequence analysis of the verification data is matched with the causal path to form a generated causal feature vector; a denoising engine is deployed to re-optimize the causal feature vector, a fusion feature vector is obtained, and an edge device is updated through global learning; the contribution of the fused feature vector is quantified, hierarchical early warning is performed through judgment, and liver disease thyroid function abnormality early warning based on machine learning is realized.

Description

Technical Field

[0001] This invention relates to the field of data model processing, and more specifically, to a machine learning-based early warning system for liver disease and thyroid dysfunction. Background Technology

[0002] Unlike hospitals, home environments are not standardized, and patients' daily activities (such as dietary habits, mood swings, and changes in home environment) introduce a large amount of noisy data. For example, a patient's intake of iodine-rich foods may temporarily increase serum iodine levels, indirectly affecting thyroid hormone levels. However, the system lacks a real-time filtering mechanism and cannot distinguish whether this is a short-term fluctuation induced by diet or a true abnormal signal related to liver fibrosis. Similarly, stress responses experienced by patients during high-pressure workdays (such as anxiety or insomnia) may lead to abnormally high heart rate and activity levels. This data is captured by wearable devices and uploaded to the cloud, but the model misinterprets it as signs of thyroid dysfunction (such as hyperthyroidism-related sympathetic nerve excitation) rather than simply the influence of environmental stress.

[0003] Wearable devices primarily rely on sensors to collect physiological indicators, but in home settings, this data is easily affected by external variables. For example, smoke or high temperatures encountered by patients while cooking at home can cause sensor bias, resulting in an artificially inflated heart rate reading of 5-10%, which in turn affects the model's analysis of liver disease-thyroid interaction factors.

[0004] These interfering data are directly fed into machine learning models, causing a shift in the distribution of input features. For example, in a typical case, a patient ingested high-iodine foods after a weekend dinner party. The system detected a transient abnormality in TSH-related indicators (which was actually a normal physiological response), but the model, based on historical training data (mostly from hospital-controlled environments), classified it as high-risk, triggering a false positive alert. This not only increases unnecessary anxiety for patients (such as frequently receiving "potential thyroid abnormality" notifications) but may also dilute the effectiveness of alerts on the doctor's end, resulting in wasted resources and delayed responses to genuine abnormalities.

[0005] Therefore, machine learning-based early warning systems for liver disease and thyroid dysfunction need to be designed and innovated to meet practical needs. Summary of the Invention

[0006] To overcome the above-mentioned deficiencies of the prior art and to achieve the above objectives, the present invention provides the following technical solution: a machine learning-based early warning system for liver disease and thyroid dysfunction, comprising: a multimodal data acquisition module: used to acquire data, and to mark the acquired data through prior processing, and to perform self-cleaning processing on the data marked as potential noise, and to obtain verification data;

[0007] Causal analysis module: Based on the false alarm signal scenario, pre-construct causal paths, integrate causal paths to form a liver-thyroid cross-organ causal network, and match the time series analysis of the verification data with the causal paths to generate causal feature vectors;

[0008] Federated optimization module: Deploys a denoising engine to further optimize the causal feature vector, obtains a fused feature vector, and updates the edge devices through global learning;

[0009] Early warning module: Used to measure the contribution of the fused feature vector and issue graded early warnings based on the judgment.

[0010] Preferably, the method for acquiring verification data includes: collecting data, marking the collected data through prior processing, and performing self-cleaning processing on data marked as potential noise.

[0011] Users wear wearable devices to measure physiological and physical data. Physiological data includes heart rate and blood oxygen data per minute, while physical data includes exercise acceleration and exercise time. The collected data is stored at a uniform frequency.

[0012] Deploy an array of environmental sensors to measure environmental data in real time, including temperature, humidity, air quality, sound decibels, and light intensity;

[0013] Input your food diary, mood score, and self-reported symptoms via a mobile app;

[0014] Whenever a set of data is collected, it is tested using physiological knowledge. Data that fails the test is deleted. For data that passes the test, if it does not conform to the normal range in physiological knowledge, it is marked as potential noise data and stored for a 5-minute window of data for self-cleaning. Otherwise, it is used normally to obtain verification data.

[0015] Preferably, the self-cleaning treatment method includes:

[0016] First, extract the heart rate from the potentially noisy data and correct it;

[0017] For other data, if the light intensity is greater than the intensity threshold, the blood oxygenation error is determined to be caused by light interference and corrected using a predetermined rule; otherwise, it is deleted directly.

[0018] The verification data is then decomposed into multiple sub-signal components, including low-frequency components and high-frequency / abnormal components. Components marked as noise are removed; and the remaining components are reconstructed to generate the final denoised data.

[0019] Preferably, the method for extracting heart rate from potential noise data and correcting it includes:

[0020] If the measured temperature is greater than the temperature threshold and the device's measured fit is greater than the upper fit threshold, then the heart rate error is determined to be caused by the thermal conduction effect. The correction method is: heart rate correction value = measured heart rate - k1 × (measured temperature - temperature threshold). High-frequency components are separated first and marked as high-temperature interference noise.

[0021] If the acceleration is greater than the acceleration threshold and the duration of exercise is greater than the duration threshold, then the heart rate error is determined to be caused by motion artifacts. The correction method is: heart rate correction value = measured heart rate - k2 × acceleration. Abnormal spike components are separated first and marked as motion artifact noise.

[0022] If the device fit is less than the lower fit threshold, the heart rate error is determined to be caused by poor sensor contact. The data for that time period is marked as unreliable and discarded.

[0023] If the ambient temperature is greater than the temperature threshold and the activity level is greater than the activity threshold, then the heart rate error is determined to be the superposition effect of high temperature and exercise. The correction method is: heart rate correction value = measured heart rate - [k1 × (ambient temperature - 30°C) + k2 × exercise acceleration]. High frequency and abnormal peak components are separated first and marked as high temperature-exercise composite interference noise.

[0024] Among them, the fit of the equipment is determined by humidity.

[0025] Preferably, the method for obtaining the causal feature vector includes:

[0026] A liver-thyroid cross-organ causal network was constructed and used as prior knowledge. It was initialized as a directed acyclic graph, and initial weights were assigned to the edges of each path.

[0027] Perform time series analysis on the validation data to detect temporal patterns between variables; match the results of the time series analysis with paths in the causal network;

[0028] For a successfully matched path, the strength of the causal effect is quantified, and a causal feature vector is generated, which is then labeled using the path number.

[0029] The causal network is dynamically adjusted based on the user's private data.

[0030] Preferably, the liver-thyroid trans-organ causal network includes path 1, path 2, path 3, path 4, path 5, path 6, path 7 and path 8;

[0031] The path is as follows: Path 1: Dietary log → Temporary TSH fluctuation → False alarm signal; Path 2: Mood score → Abnormal heart rate → False alarm signal; Path 3: Temperature → Falsely high heart rate → False alarm signal; Path 4: Exercise acceleration / exercise time → Abnormal heart rate → False alarm signal; Path 5: Light intensity → Falsely low blood oxygen → False alarm signal; Path 6: Humidity → Device fit → Abnormal heart rate / Falsely low blood oxygen → False alarm signal; Path 7: Air quality / sound decibels → Abnormal heart rate → False alarm signal; Path 8: Self-reported symptoms → Abnormal heart rate → False alarm signal; Other situations output a real alarm signal.

[0032] In this context, → represents an edge, and / represents or.

[0033] Preferably, the method for further optimizing the causal feature vector by deploying the denoising engine to obtain the fused feature vector includes:

[0034] Deploy a denoising engine in the cloud, take the causal feature vector as input, and obtain the denoised causal feature vector by simulating the noise injection and reverse generation process;

[0035] The denoised causal feature vector is fused with the verification data to obtain the fused feature vector.

[0036] Preferably, the method for updating edge devices through global learning includes:

[0037] The discovery and quantification of causal paths running on each user edge device are used as a small causal discovery model, and the gradients of the small causal discovery model parameters and the denoising engine parameters are uploaded to the cloud.

[0038] The cloud performs global learning and updates the data, then distributes it to the user's edge devices.

[0039] Preferably, the method of quantifying the contribution of the fused feature vector and then issuing a graded early warning based on the determination includes:

[0040] Analyze the fused feature vectors and output the contribution of each early warning decision;

[0041] When the contribution of liver-related factors is greater than expected and the contribution of environmental interference is less than expected, a Level 1 alarm is triggered, and an immediate medical attention reminder is displayed.

[0042] When the contribution of environmental interference exceeds the threshold, a level 2 warning is issued and a retest reminder is displayed; when the contribution of liver-related factors is similar to that of environmental interference (e.g., the deviation is within 5%), the case is transferred to the doctor for review.

[0043] Preferably, the doctor's terminal displays abnormal causal network paths of the patient and provides alarm attribution analysis.

[0044] The technical effects and advantages of the machine learning-based early warning system for liver disease and thyroid dysfunction of this invention are as follows:

[0045] The multimodal data acquisition module comprehensively collects physiological, physical, environmental, and behavioral data. Multimodal fusion technology is used to calculate cross-modal weights and remove environmental noise. This provides a more comprehensive reflection of the user's health status and environmental interference, avoiding misjudgments based on a single factor.

[0046] Through a causal analysis module, a pre-constructed liver-thyroid cross-organ causal network is built, and the causal path weights are dynamically adjusted, enabling accurate differentiation between short-term disturbances and long-term pathological signals. Compared to traditional machine learning models (such as correlation-based prediction models), causal inference can reveal the causal relationships between variables, thereby improving the accuracy and interpretability of predictions. Dynamic adjustment of causal path weights enables personalized adaptation.

[0047] By employing edge-side self-cleaning and a small causal discovery model, combined with a cloud-based denoising engine and federated learning, efficient edge-cloud collaboration is achieved. This reduces the cloud computing burden, improves real-time performance, and enhances the overall model performance through the cloud-based denoising engine and federated learning.

[0048] The early warning module outputs the contribution of each early warning decision and categorizes warnings based on contribution, providing highly interpretable alert information. Contribution analysis provides transparent decision-making basis, and the tiered early warning strategy avoids excessive alerts, reducing user anxiety and wasting doctors' resources. Attached Figure Description

[0049] Figure 1 This is a schematic diagram of the structure of a machine learning-based early warning system for liver disease and thyroid dysfunction according to the present invention.

[0050] Figure 2 This is a schematic diagram illustrating the steps of a machine learning-based early warning system for liver disease and thyroid dysfunction according to the present invention. Detailed Implementation

[0051] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0052] Example 1

[0053] Please see Figure 1 and Figure 2As shown in this embodiment, a machine learning-based early warning system for liver disease and thyroid dysfunction includes:

[0054] In scenarios where liver fibrosis patients use wearable devices (such as smart bracelets or mobile apps) connected to machine learning-based early warning systems at home, environmental interference leading to noisy data becomes particularly prominent. This problem stems from various uncontrollable factors in patients' daily lives, which can unintentionally contaminate the input data, causing the system model to have biased predictions of thyroid dysfunction. Specifically:

[0055] Unlike hospitals, home environments are not standardized, and patients' daily activities (such as dietary habits, mood swings, and changes in home environment) introduce a large amount of noisy data. For example, a patient's intake of iodine-rich foods (such as seafood or iodized salt) may temporarily increase serum iodine levels, indirectly affecting thyroid hormone indicators (such as TSH or T4 values). However, the system lacks a real-time filtering mechanism and cannot distinguish whether this is a short-term fluctuation induced by diet or a real abnormal signal related to liver fibrosis. Similarly, stress responses experienced by patients during high-pressure workdays (such as anxiety or insomnia) may lead to abnormally high heart rate and activity levels. This data is captured by wearable devices and uploaded to the cloud, but the model misinterprets it as a sign of thyroid dysfunction (such as hyperthyroidism-related sympathetic nerve excitation) rather than simply the influence of environmental stress.

[0056] Wearable devices primarily rely on sensors to collect physiological indicators (such as heart rate sampling per minute and activity levels recorded via accelerometers). However, in home settings, these data are easily affected by external variables. For example, smoke or high temperatures encountered by patients while cooking at home may cause sensor bias, resulting in an artificially inflated heart rate reading of 5-10%, which in turn affects the model's analysis of liver disease-thyroid interaction factors (such as the correlation between heart rate fluctuations and liver enzyme levels).

[0057] These interfering data are directly fed into machine learning models, causing a shift in the distribution of input features. For example, in a typical case, a patient ingested high-iodine foods after a weekend dinner party. The system detected a transient abnormality in TSH-related indicators (which was actually a normal physiological response), but the model, based on historical training data (mostly from hospital-controlled environments), classified it as high-risk, leading to a false positive alarm. Overall, this noise results in an alarm false alarm rate as high as 50%, with dietary interference accounting for 30%, stress-related interference for 15%, and other environmental factors (such as seasonal allergies or household pollutants) for 5%. This not only increases unnecessary anxiety for patients (such as frequently receiving "potential thyroid abnormality" notifications) but may also dilute the effectiveness of alarms on the doctor's end, resulting in wasted resources and delayed responses to genuine abnormalities.

[0058] This problem stems from the fact that the dynamic and heterogeneous nature of the home environment was not fully considered during the system design, resulting in insufficient robustness of the model to noise, which in turn affects the reliability and accuracy of the entire early warning process.

[0059] Users wear wearable devices (such as smart bracelets) to measure physiological and physical data. Physiological data includes heart rate and blood oxygen data per minute (measured via photoplethysmography pulse wave signals integrated in the wearable device), and physical data includes exercise acceleration and exercise time (measured via an accelerometer integrated in the wearable device). The collected data is stored at a uniform frequency; ensuring that the data is stored in a local buffer at a uniform frequency (e.g., once per second) to handle brief signal interruptions.

[0060] Deploy an array of environmental sensors to measure environmental data in real time, including temperature (such as measuring the high-temperature environment of cooking through temperature sensors), humidity, air quality (such as other data affecting air quality, such as particulate matter and volatile organic compounds), sound decibels (such as measuring the sound of the range hood during cooking by combining a microphone), and light intensity.

[0061] Input your food log (e.g., using AI to recognize food images), mood rating (e.g., self-assessment based on a mood scale), and self-reported symptoms (e.g., fatigue, anxiety) via a mobile app.

[0062] Whenever a set of data is collected (such as heart rate sampling value, ambient temperature and dietary input), it is tested using physiological knowledge. Data that fails the test (such as heart rate > 200 beats / minute) is deleted. For data that passes the test, if it does not conform to the normal range in physiological knowledge, it is marked as potential noise data and stored for a 5-minute window of data for self-cleaning. Otherwise, it is used normally to obtain verification data.

[0063] Physiological knowledge is derived from pre-trained physical constraint models downloaded from local storage or the cloud. These models predefine biologically reasonable ranges and upper / lower limits (e.g., normal resting heart rate in adults is typically 60-100 beats per minute, with an upper limit of no more than 200 beats per minute; normal blood oxygen saturation is between 95-100%; and acceleration during exercise does not exceed human limits, such as 5 meters squared per second). These ranges are based on medical research data (e.g., the European Association for the Study of the Liver (EASL-ALEH) Clinical Practice Guidelines: Non-invasive testing for assessing the severity and prognosis of liver disease) and physical equations (e.g., models of the effect of temperature on sensor accuracy, such as heart rate sensor deviation reaching 5% at high temperatures).

[0064] Ensure data conforms to realistic constraints before being processed, avoiding contamination from obviously invalid or extreme values ​​(such as impossible heart rate readings). Implement early intervention mechanisms to achieve data "self-cleaning" and improve overall noise robustness.

[0065] Self-cleaning methods include: real-time calibration of sensor data at the edge to remove noise caused by environmental interference (such as high temperature and smoke); heart rate is one of the most common and direct physiological indicators in wearable devices, and it is easily affected by environmental interference (such as high temperature and exercise) and pathological factors (such as hyperthyroidism and sympathetic nerve excitation). Therefore, in the early warning scenario of thyroid dysfunction in patients with liver fibrosis, abnormal heart rate (such as tachycardia) may be an important manifestation of hyperthyroidism or sympathetic nerve excitation, thus the accuracy of heart rate data is crucial to predictive ability.

[0066] First, the heart rate is extracted from the potentially noisy data and corrected. Although the corrected physiological signal has reduced some obvious errors, it may still contain complex noise components (such as high-frequency fluctuations and abnormal spikes). These noise components may be caused by a combination of factors, such as the residual effects of environmental interference and the random noise of the sensor itself.

[0067] For other data, if the light intensity is greater than the intensity threshold, the blood oxygenation error is determined to be caused by light interference, and a predetermined rule is used for correction [e.g., blood oxygenation correction value = measured blood oxygen + k3 × (light intensity - 1000 lux), where k3 is the light interference coefficient (calibrated experimentally)]. Otherwise, it is directly deleted. Then, the verification data is decomposed into multiple sub-signal components, including low-frequency components (usually representing real physiological signals) and high-frequency / abnormal components (usually representing noise), and the components marked as noise are removed.

[0068] The retained components are then reconstructed to generate the final denoised data;

[0069] The goal of reconstructing the retained components is to recombine them into a complete time-series signal, restoring it to the same sampling frequency and time axis as the original signal. This is achieved by superimposing the separated sub-signal components along the time axis, filling in missing or discarded values ​​through linear interpolation or resampling to generate a denoised signal.

[0070] Methods for extracting and correcting heart rate from potentially noisy data include:

[0071] If the measured temperature is greater than the temperature threshold and the device's measured fit is greater than the upper fit threshold, then the heart rate error is determined to be caused by the thermal conduction effect. The artificially high heart rate is expected to be 5%-10% of the actual heart rate. The correction method is: heart rate correction value = measured heart rate - k1 × (measured temperature - temperature threshold). High-frequency components are separated first and marked as high-temperature interference noise.

[0072] If the acceleration is greater than the acceleration threshold and the duration of exercise is greater than the time threshold, then the heart rate error is determined to be caused by motion artifacts. The artificially high heart rate is expected to be 10%-15% of the actual heart rate. The correction method is: heart rate correction value = measured heart rate - k2 × acceleration. Abnormal peak components are separated first and marked as motion artifact noise.

[0073] If the device fit is less than the lower fit threshold, the heart rate error is determined to be caused by poor sensor contact, which manifests as intermittent data loss or abnormal fluctuations. The data in this time period is marked as unreliable and discarded.

[0074] If the ambient temperature is greater than the temperature threshold and the activity level is greater than the activity threshold, then the heart rate error is determined to be the superposition effect of high temperature and exercise. The artificially high heart rate is expected to be 15%-20% of the actual heart rate. The correction method is: heart rate correction value = measured heart rate - [k1×(ambient temperature-30°C)+k2×exercise acceleration]. High-frequency and abnormal peak components are separated first and marked as high temperature-exercise composite interference noise.

[0075] The device fit is determined by humidity. Poor sensor contact can be caused by skin moisture. For example, if the humidity is <50%, the device fit is 90% (assuming dry skin in low humidity results in higher fit); if the humidity is between 50% and 80%, the device fit is 90% - 2 × (humidity - 50%) (assuming a linear decrease in fit as humidity increases); if the humidity is >80%, the device fit is 30% (assuming significantly lower fit due to moist skin in high humidity). Alternatively, a mathematical formula can be designed to represent the numerical conversion between humidity and device fit.

[0076] k1 represents the thermal conductivity coefficient, with units of "heart rate change / temperature change" (e.g., beats / minute / °C). k2 represents the motion artifact coefficient, with units of "heart rate change / acceleration" (e.g., beats / minute / g). k1 and k2 are usually obtained through experimental calibration. For example, in a laboratory environment, the effects of different temperatures (e.g., 25°C, 30°C, 35°C) on the heart rate sensor are simulated, the amplitude of the artificial heart rate rise is measured, and the value of k1 is fitted. The effects of different exercise intensities (e.g., accelerations of 1g, 2g, 3g) on ​​the heart rate sensor are simulated, the amplitude of the artificial heart rate rise is measured, and the value of k2 is fitted.

[0077] By identifying causal pathways, it is possible to determine whether an abnormality (such as increased heart rate or decreased blood oxygen) is caused by short-term disturbances (such as high temperature or exercise) or by long-term pathological conditions (such as thyroid abnormalities), which helps in subsequent causal discovery and feature extraction.

[0078] A liver-thyroid cross-organ causal network was constructed and used as prior knowledge. It was initialized as a directed acyclic graph, and an initial weight was assigned to the edge of each path (e.g., the initial weight was set to 1). These weights will be dynamically adjusted in the subsequent causal discovery process to reflect the strength of personalized causal effects.

[0079] Perform time series analysis on the validation data to detect temporal patterns between variables; for example, analyze whether "TSH fluctuates temporarily within 2 hours after a high-iodine diet is recorded in the diet log", or whether "heart rate becomes abnormal after mood score increases".

[0080] The results of time series analysis are matched with paths in the causal network; for example, if the data shows "the heart rate is falsely high after the measured temperature is greater than the temperature threshold", then path 3 is matched (temperature → falsely high heart rate → false alarm signal).

[0081] For successfully matched paths, the strength of the causal effect is quantified using statistical or causal inference methods, and a causal feature vector is generated and labeled using the path number. The quantification of the strength of the causal effect can be based on the correlation coefficient or regression coefficient. For example, if path 1 (dietary log → temporary TSH fluctuation) appears frequently in the data, and the correlation coefficient between TSH fluctuation and high iodine diet is 0.8, then the strength of the causal effect of this path is set to 0.8.

[0082] The causal network is dynamically adjusted based on user's private data. For example, if a user has never recorded a high-iodine diet, the weight of path 1 may be reduced or removed. If a user is frequently in a high-temperature environment and experiences frequent false high heart rate, the weight of path 3 may be increased. The weight reflects the strength of the causal relationship between variables in the path, used to determine whether an abnormality is caused by short-term disturbances or long-term pathological signals. For example, if the overall weight (sum of all edge weights) of path 3 (temperature → false high heart rate → false alarm signal) is high, while the weights of other paths are low, the system is more inclined to attribute the heart rate abnormality to high-temperature disturbances rather than thyroid abnormalities. The higher the path weight, the more likely the system is to consider that path as the main cause of the abnormality, thus generating corresponding alarm signals (false or real), or prioritizing its retrieval during subsequent path matching.

[0083] Example:

[0084] The corrected data for the user within a 5-minute window is as follows:

[0085] Dietary log: Record high-iodine diet.

[0086] Temperature: 35°C (exceeding the temperature threshold).

[0087] Motion acceleration: 2.5g (exceeding the acceleration threshold).

[0088] Heart rate: 110 beats / minute (abnormal).

[0089] Blood oxygen: 98% (normal).

[0090] Mood score: 8 (high stress).

[0091] Humidity: 60% (equipment fit = 90% - 2 × (60% - 50%) = 70%).

[0092] Light intensity: 800 lux (not exceeding the intensity threshold).

[0093] Air quality: Normal.

[0094] Sound volume: Normal.

[0095] Self-reported symptoms: None.

[0096] Using the causal network (path 1 to path 8) as prior knowledge, initialize a directed acyclic graph (DAG).

[0097] Time series analysis:

[0098] The analysis focused on whether TSH levels experienced temporary fluctuations after a high-iodine diet was recorded in a diet log.

[0099] Analyze whether the heart rate is artificially high after the temperature exceeds the temperature threshold.

[0100] Analyze whether the heart rate is abnormal after the acceleration of exercise exceeds the acceleration threshold.

[0101] Analyze whether heart rate is abnormal after mood score increases.

[0102] Path matching:

[0103] Matching path 1 (diet log → temporary TSH fluctuations → false alarm signals).

[0104] Matching path 2 (emotional score → abnormal heart rate → false alarm signal).

[0105] Matching path 3 (temperature → artificially high heart rate → false alarm signal).

[0106] Matching path 4 (motion acceleration / exercise time → abnormal heart rate → false alarm signal).

[0107] Quantification of causal effects:

[0108] Pathway 1: The causal effect strength is 0.8 (due to the high correlation between TSH fluctuations and high iodine diet).

[0109] Path 2: The causal effect strength is 0.6 (due to the correlation between abnormal heart rate and mood score).

[0110] Pathway 3: The causal effect strength is 0.7 (due to the correlation between falsely high heart rate and temperature).

[0111] Path 4: The causal effect strength is 0.5 (due to the correlation between abnormal heart rate and exercise).

[0112] Path 5: The causal effect strength is 0.0 (due to normal blood oxygenation).

[0113] Path 6: The causal effect strength is 0.0 (due to humidity not causing abnormal fit).

[0114] Path 7: The causal effect strength is 0.0 (due to normal air quality and sound decibels).

[0115] Path 8: The causal effect strength is 0.0 (due to no related symptom reports).

[0116] Causal feature vector generation:

[0117] Causal feature vector: [0.8,0.6,0.7,0.5,0.0,0.0,0.0,0.0].

[0118] Store causal feature vectors in a local buffer at a uniform frequency (e.g., once per second);

[0119] Based on medical knowledge, a liver-thyroid cross-organ causal network is established, including pathways 1, 2, 3, 4, 5, 6, 7, and 8.

[0120] The pathways are as follows: Path 1: Dietary log → Temporary TSH fluctuation → False alarm signal; Path 2: Mood score → Abnormal heart rate → False alarm signal; Path 3: Temperature → Falsely high heart rate → False alarm signal; Path 4: Exercise acceleration / exercise time → Abnormal heart rate → False alarm signal; Path 5: Light intensity → Falsely low blood oxygen → False alarm signal; Path 6: Humidity → Equipment fit → Abnormal heart rate / Falsely low blood oxygen → False alarm signal; Path 7: Air quality / sound decibels → Abnormal heart rate → False alarm signal; Path 8: Self-reported symptoms → Abnormal heart rate → False alarm signal; Other situations output a real alarm signal. In other words, anything outside the above pathways is considered a thyroid dysfunction.

[0121] Where → represents an edge, and / represents or;

[0122] Path 1 represents the causal chain of short-term disturbances caused by a high-iodine diet. Consuming high-iodine foods (such as seafood) can lead to short-term fluctuations in thyroid-stimulating hormone (TSH) levels. These short-term fluctuations can be misinterpreted as thyroid abnormalities, triggering false alarm signals.

[0123] Path 2 represents the causal chain of short-term disturbances caused by emotional stress. High emotional stress (such as anxiety) in users can lead to abnormally high heart rates. This heart rate abnormality caused by emotional stress can be misdiagnosed as a thyroid abnormality, thus triggering false alarm signals.

[0124] Path 3 represents the causal chain of short-term disturbances caused by high-temperature environments. Ambient temperatures exceeding a certain threshold can lead to falsely high heart rate readings from heart rate sensors. This falsely high heart rate caused by high temperatures can be misinterpreted as a thyroid abnormality, thus triggering false alarm signals.

[0125] Path 4 represents the causal chain of short-term disturbances caused by exercise. Exercise acceleration exceeding an acceleration threshold leads to an abnormally high heart rate. Exercise duration exceeding a time threshold further exacerbates the abnormal heart rate. This exercise-induced abnormal heart rate can be misinterpreted by the system as a thyroid abnormality, thus triggering false alarm signals.

[0126] Path 5 represents the causal chain of short-term disturbances caused by light intensity. Light intensity exceeding a threshold can lead to falsely low readings from the blood oxygen sensor. This falsely low blood oxygen level caused by light intensity can be misinterpreted by the system as hypoxemia, thus triggering a false alarm signal.

[0127] Path 6 represents the causal chain of short-term interference caused by humidity through device fit. Increased humidity leads to decreased device fit (e.g., humidity > 80%, device fit = 30%). Device fit below the lower fit threshold causes abnormal heart rate sensor measurements (e.g., intermittent data loss or abnormal fluctuations). This results in falsely low blood oxygen sensor measurements. These humidity-induced anomalies are misinterpreted by the system as thyroid abnormalities or hypoxemia, thus triggering false alarm signals.

[0128] Path 7 represents the causal chain of short-term disturbances caused by air quality and sound decibels. Decreased air quality (such as excessive particulate matter or volatile organic compounds) may cause breathing difficulties in users, leading to abnormal heart rates. High-decibel sounds (such as the sound of a range hood) may increase user stress, also causing abnormal heart rates. These heart rate abnormalities caused by air quality or sound decibels may be misinterpreted by the system as thyroid abnormalities, thus triggering false alarm signals.

[0129] Path 8 represents a causal chain of short-term disturbances caused by self-reported symptoms. User-reported symptoms (such as fatigue and anxiety) may cause abnormally high heart rates. This abnormal heart rate caused by self-reported symptoms may be misinterpreted by the system as a thyroid abnormality, thus triggering false alarm signals.

[0130] Deploy a denoising engine in the cloud, take the causal feature vector as input, and further remove any residual noise in the causal feature vector by simulating noise injection and reverse generation processes to obtain the denoised causal feature vector.

[0131] The denoised causal feature vector is fused with the validation data (using multimodal fusion techniques) to obtain a fused feature vector. The goal of fusion is to calculate cross-modal weights and remove environmental noise. For example, if an abnormally elevated heart rate is associated with a temperature exceeding a temperature threshold, the weight of the abnormal heart rate in the prediction is reduced, suggesting it may be due to high-temperature interference rather than thyroid abnormalities.

[0132] The temperature threshold is either set by expert experience or determined by laboratory testing.

[0133] The noise injection process is simulated by progressively adding noise to the causal feature vectors, generating a noise distribution sequence. For example, the path weight sequence is progressively "diffused" into a pure noise state to simulate dietary or stress interference. A trained denoising network is then used to progressively restore the "clean" causal feature vector distribution. The denoising network integrates medical research data as prior knowledge to ensure the model distinguishes between short-term disturbances (such as diet-induced TSH fluctuations) and long-term pathological signals (such as liver disease-related thyroid abnormalities). The noise scheduling parameters are adaptively adjusted based on the environmental context (such as temperature deviations). For example, in high-temperature environments, the noise scheduling parameters are increased to simulate stronger noise interference.

[0134] A conditional diffusion model is used to generate realistic noise data. Input environmental parameters (e.g., "temperature 30°C + oil fume environment"), and output the corresponding causal feature vector noise pattern. A dual-channel discriminator is constructed: Channel 1: distinguishes between real causal feature vectors and generated noise. Channel 2: distinguishes between liver disease-related thyroid abnormalities and environmental interference. The generator aims to synthesize the "worst-case noise" that can deceive the dual discriminator, forcing the main model to improve its robustness.

[0135] Output: The denoised causal feature vector, which serves as the input for subsequent steps.

[0136] The discovery and quantification of causal paths running on each user's edge device are treated as a small causal discovery model. This small causal discovery model is a tool used to analyze data, verify whether a causal path applies to a specific user, and quantify the strength of the path's causal effect. For example, a model for a patient living alone can be independently optimized to suit their specific home environment (such as high frequency of high-temperature cooking).

[0137] Specifically, the small-scale causal discovery model takes validation data as input and verifies whether paths in the causal network hold true in the user data through time series analysis, path matching, and causal effect quantification, and calculates the weight (causal effect strength) of each path. The output is a causal feature vector.

[0138] Upload the gradients (not the raw data) of the parameters of the small causal discovery model and the parameters of the denoising engine to the cloud;

[0139] The cloud-based system updates the global model after learning from multiple users, aggregating model parameters and gradients to update the global causal discovery and denoising models. The updated global model is then personalized based on the user's causal feature vectors and distributed to the user's edge devices, achieving personalized noise adaptation.

[0140] Global learning typically uses weighted averaging (such as FedAvg) or other aggregation methods, and the distribution method is to adjust the model parameters based on the user's causal feature vector.

[0141] Analyze the fused feature vectors and output the contribution of each early warning decision; for example:

[0142] Elevated TSH levels: 82% probability of dietary interference, 15% probability of liver-related causes.

[0143] Abnormal heart rate: 70% probability of interference from high temperature, 20% probability of interference from emotional stress, and 7% probability of suspected hyperthyroidism.

[0144] When the contribution of liver-related factors is greater than expected and the contribution of environmental interference is less than expected, a Level 1 alarm (red) is triggered, and an immediate medical attention reminder is displayed.

[0145] When the contribution of environmental interference exceeds the threshold, a level 2 warning (yellow) is issued and a retest reminder is displayed; when the contribution of liver-related factors is similar to that of environmental interference (e.g., the deviation is within 5%), the case is transferred to a doctor for review.

[0146] Example: Extracting path weights associated with heart rate abnormalities based on data and causal feature vectors:

[0147] Path 2 (Emotional score → Abnormal heart rate → False alarm signal): Weight = 0.6.

[0148] Path 3 (Temperature → Falsely High Heart Rate → False Alarm Signal): Weight = 0.7.

[0149] Path 4 (Exercise acceleration / exercise time → abnormal heart rate → false alarm signal): weight = 0.5.

[0150] Other pathways (such as pathways 1, 5, 6, 7, and 8) have a weight of 0 because they are not related to heart rate abnormalities (e.g., pathway 1 involves TSH fluctuations, and pathway 5 involves hypoxemia).

[0151] The contribution of environmental disturbance is determined by the weights of path 2, path 3 and path 4.

[0152] Total weight = Path 2 weight + Path 3 weight + Path 4 weight = 0.6 + 0.7 + 0.5 = 1.8.

[0153] Environmental interference contribution of each path = weight of that path / total weight × 100%.

[0154] The contribution of Path 2 (emotional stress) is approximately 33.33% (0.6 / 1.8 × 100%).

[0155] The contribution of path 3 (high temperature interference) is approximately 38.89% (0.7 / 1.8×100%).

[0156] The contribution of path 4 (motion interference) is approximately 27.78% (0.5 / 1.8 × 100%).

[0157] Total environmental disturbance contribution = Path 2 + Path 3 + Path 4 = 33.33% + 38.89% + 27.78% = 100%.

[0158] Since the heart rate abnormalities are entirely explained by pathways 2, 3, and 4 (i.e., short-term disturbances), there is no evidence of long-term pathological signals (such as thyroid abnormalities).

[0159] Therefore, the contribution from liver origin is 0%.

[0160] Probability of high temperature interference (path 3): 38.89%.

[0161] Probability of emotional stress (path 2): 33.33%.

[0162] Probability of motion interference (path 4): 27.78%.

[0163] Probability of suspected hyperthyroidism (hepatic origin): 0%.

[0164] When the contribution of liver-related factors is greater than 60% and the contribution of environmental interference is less than 5%, a Level 1 alarm is triggered, and a "Seek Medical Attention Immediately" reminder is displayed.

[0165] When the contribution of environmental interference is greater than 30%, a level 2 warning will be issued and a "retest" reminder will be displayed.

[0166] Physician review: When the deviation between the contribution of liver-derived factors and the contribution of environmental interference is within 5%, the case is forwarded to a physician for review.

[0167] Hepatic contribution = 0%, environmental interference contribution = 100%, bias = 100% > 5% (not satisfied).

[0168] A level 2 warning (yellow) is issued, along with a "Retest" reminder. The user receives the message: "Your abnormal heart rate (110 beats / minute) may be caused by high temperature interference (contribution 38.89%), emotional stress (contribution 33.33%), or exercise interference (contribution 27.78%). Retesting is recommended."

[0169] The doctor's interface displays abnormal causal network paths for patients (e.g., "temperature → artificially high heart rate → false alarm signal") and provides alarm attribution analysis (e.g., "high temperature interference contributes 70%)". The doctor's feedback (e.g., confirming thyroid abnormalities or environmental interference) is used to update the path weights of the causal network and the global model parameters, forming a closed-loop learning mechanism.

[0170] Example 2

[0171] This embodiment discloses an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the operation mode of the aforementioned machine learning-based early warning system for liver disease and thyroid dysfunction.

[0172] Since the electronic device described in this embodiment is the electronic device used to implement the machine learning-based early warning system for liver disease and thyroid dysfunction described in this application, those skilled in the art can understand the specific implementation methods and various variations of the electronic device in this embodiment based on the machine learning-based early warning system for liver disease and thyroid dysfunction described in this application. Therefore, how the electronic device implements the method in this application will not be described in detail here. Any electronic device used by those skilled in the art to implement the machine learning-based early warning system for liver disease and thyroid dysfunction described in this application falls within the scope of protection of this application.

[0173] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0174] The above description is merely a preferred embodiment of the present invention, and the scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for users of ordinary technical skills, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A machine learning-based early warning system for liver disease and thyroid dysfunction, characterized in that, include: Multimodal data acquisition module: used to collect data, and to mark the collected data through prior processing. For data marked as potential noise, self-cleaning processing is performed to obtain verification data; Causal analysis module: Based on the false alarm signal scenario, pre-construct causal paths, integrate causal paths to form a liver-thyroid cross-organ causal network, and match the time series analysis of the verification data with the causal paths to generate causal feature vectors; Federated optimization module: Deploys a denoising engine to further optimize the causal feature vector, obtains a fused feature vector, and updates the edge devices through global learning; Early warning module: used to measure the contribution of the fused feature vector and issue graded early warnings based on the judgment. The method for obtaining the causal feature vector includes: A liver-thyroid cross-organ causal network was constructed and used as prior knowledge. It was initialized as a directed acyclic graph, and initial weights were assigned to the edges of each path. Perform time series analysis on the validation data to detect temporal patterns between variables; match the results of the time series analysis with paths in the causal network; For a successfully matched path, the strength of the causal effect is quantified, and a causal feature vector is generated, which is then labeled using the path number. The causal network is dynamically adjusted based on the user's private data; The method for further optimizing the causal feature vector by deploying the denoising engine to obtain the fused feature vector includes: Deploy a denoising engine in the cloud, take the causal feature vector as input, and obtain the denoised causal feature vector by simulating the noise injection and reverse generation process; The denoised causal feature vector is fused with the verification data to obtain the fused feature vector; The method for updating edge devices through global learning includes: The discovery and quantification of causal paths running on each user edge device are used as a small causal discovery model, and the gradients of the small causal discovery model parameters and the denoising engine parameters are uploaded to the cloud. The cloud performs global learning and updates the data, then distributes it to the user's edge devices.

2. The early warning system for liver disease and thyroid dysfunction based on machine learning according to claim 1, characterized in that, The method for acquiring verification data includes: collecting data, marking the collected data through prior processing, and performing self-cleaning processing on data marked as potential noise. Users wear wearable devices to measure physiological and physical data. Physiological data includes heart rate and blood oxygen data per minute, while physical data includes exercise acceleration and exercise time. The collected data is stored at a uniform frequency. Deploy an array of environmental sensors to measure environmental data in real time, including temperature, humidity, air quality, sound decibels, and light intensity; Enter your food diary, mood score, and self-reported symptoms via the mobile app; Whenever a set of data is collected, it is tested using physiological knowledge. Data that fails the test is deleted. For data that passes the test, if it does not conform to the normal range in physiological knowledge, it is marked as potential noise data and stored for a 5-minute window of data for self-cleaning. Otherwise, it is used normally to obtain verification data.

3. The early warning system for liver disease and thyroid dysfunction based on machine learning according to claim 2, characterized in that, The self-cleaning treatment method includes: First, extract the heart rate from the potentially noisy data and correct it; For other data, if the light intensity is greater than the intensity threshold, the blood oxygenation error is determined to be caused by light interference and corrected using a predetermined rule; otherwise, it is deleted directly. The verification data is then decomposed into multiple sub-signal components, including low-frequency components and high-frequency / abnormal components. Components marked as noise are removed; and the remaining components are reconstructed to generate the final denoised data.

4. The early warning system for liver disease and thyroid dysfunction based on machine learning according to claim 3, characterized in that, The method for extracting and correcting heart rate from potentially noisy data includes: If the measured temperature is greater than the temperature threshold and the device's measured fit is greater than the upper fit threshold, then the heart rate error is determined to be caused by the thermal conduction effect. The correction method is: heart rate correction value = measured heart rate - k1 × (measured temperature - temperature threshold). High-frequency components are separated first and marked as high-temperature interference noise. If the acceleration is greater than the acceleration threshold and the duration of exercise is greater than the duration threshold, then the heart rate error is determined to be caused by motion artifacts. The correction method is: heart rate correction value = measured heart rate - k2 × acceleration. Abnormal spike components are separated first and marked as motion artifact noise. If the device fit is less than the lower fit threshold, the heart rate error is determined to be caused by poor sensor contact. The data in that 5-minute window is marked as unreliable and discarded. If the ambient temperature is greater than the temperature threshold and the activity level is greater than the activity threshold, then the heart rate error is determined to be the superposition effect of high temperature and exercise. The correction method is: heart rate correction value = measured heart rate - [k1 × (ambient temperature - 30°C) + k2 × exercise acceleration]. High frequency and abnormal peak components are separated first and marked as high temperature-exercise composite interference noise. Among them, the fit of the equipment is determined by humidity, and k1 and k2 represent the thermal conductivity coefficients.

5. A machine learning-based early warning system for liver disease and thyroid dysfunction according to claim 4, characterized in that, The liver-thyroid cross-organ causal network includes path 1, path 2, path 3, path 4, path 5, path 6, path 7 and path 8; The path is as follows: Path 1: Dietary log → Temporary TSH fluctuation → False alarm signal; Path 2: Mood score → Abnormal heart rate → False alarm signal; Path 3: Temperature → Falsely high heart rate → False alarm signal; Path 4: Exercise acceleration / exercise time → Abnormal heart rate → False alarm signal; Path 5: Light intensity → Falsely low blood oxygen → False alarm signal; Path 6: Humidity → Device fit → Abnormal heart rate / Falsely low blood oxygen → False alarm signal; Path 7: Air quality / sound decibels → Abnormal heart rate → False alarm signal; Path 8: Self-reported symptoms → Abnormal heart rate → False alarm signal; Other situations output a real alarm signal. In this context, → represents an edge, and / represents or.

6. A machine learning-based early warning system for liver disease and thyroid dysfunction according to claim 5, characterized in that, The method of quantifying the contribution of the fused feature vector and then issuing a graded early warning based on the judgment includes: Analyze the fused feature vectors and output the contribution of each early warning decision; When the contribution of liver-related factors is greater than expected and the contribution of environmental interference is less than expected, a Level 1 alarm is triggered, and an immediate medical attention reminder is displayed. When the contribution of environmental interference exceeds the threshold, a level 2 warning is issued and a retest reminder is displayed; when the contribution of liver-related factors is similar to that of environmental interference, the case is transferred to the doctor for review.

7. A machine learning-based early warning system for liver disease and thyroid dysfunction according to claim 6, characterized in that, The doctor's terminal displays abnormal causal network paths of patients and provides alarm attribution analysis.

Images