Lung cancer risk early warning system based on remote four diagnosis information

Abstract

The invention relates to the technical field of medical data processing, in particular to a lung cancer risk early warning system based on remote four diagnosis information, which comprises an acquisition module, a generation module, a calculation module, an early warning module and a correction module. According to the method, through multi-modal time sequence fusion and causal map driven closed-loop regulation and control, the more stable, interpretable and adaptive early warning capability for the lung cancer risk is realized, heterogeneous evidences from observation, listening and interrogation are aggregated into a dynamic alarm index according to time and causal paths, instantaneous noise is filtered by using trend consistency and adversarial test, and the early warning effect is improved. Therefore, false alarms caused by short-time fluctuation are remarkably reduced while the sensitivity to real abnormity is kept; the method achieves balance among false alarm reduction, missed diagnosis risk reduction, positive prediction value improvement and user sampling burden optimization, and effectively solves the problem of insufficient early warning accuracy caused by risk identification lag due to dependence on single static medical record information.

Description

Technical Field

[0001] This invention relates to the field of medical data processing technology, and in particular to a lung cancer risk early warning system based on remote four diagnostic methods. Background Technology

[0002] With the rapid development of medical technology and the popularization of telemedicine applications, the clinical demand for dynamic monitoring of high-risk groups is constantly increasing. However, traditional lung cancer screening methods rely on regular imaging examinations and single indicator assessments, which have problems such as low monitoring frequency, insufficient data dimensions, and untimely response. It is difficult to identify early risk signals of individuals in a timely manner and carry out precise intervention, which brings serious challenges to clinical risk management and decision support.

[0003] Chinese Patent Application Publication No. CN120072296A discloses a lung cancer risk early warning system based on multi-factor supporting information. The system includes: a baseline information acquisition unit for acquiring the baseline information of the target patient, including name, contact number, ID number, and names of several hospitals previously visited; a hospital passive terminal for extracting the hospital names from the baseline information and transmitting the baseline information to a corresponding docking sub-terminal; a docking sub-terminal including a medical record database and a sub-terminal data management unit for retrieving the target patient's past medical records from the medical record database after receiving the transmitted baseline information; a sub-terminal data management unit for encrypting and encoding the past medical records according to a preset encryption rule to obtain the target patient's coded past medical records; and a restoration and display unit for restoring the coded past medical records and displaying them.

[0004] Therefore, the aforementioned system has the following problems: it is based solely on static medical record data and lacks continuous monitoring and analysis of patients' dynamic physiological and behavioral indicators; data encryption only applies to the medical record transmission stage, lacking end-to-end data security and privacy protection strategies; and it does not introduce individualized abnormal pattern mining and causal reasoning mechanisms, thus failing to provide accurate and dynamic early warning of lung cancer risk. Summary of the Invention

[0005] To address this, the present invention provides a lung cancer risk early warning system based on remote four diagnostic methods information. This system overcomes the problem of insufficient early warning accuracy caused by the reliance on a single static medical record information in the prior art through dynamic modeling that integrates real-time monitoring data from the four diagnostic methods, time-series representation learning, and causal reasoning algorithms.

[0006] To achieve the above objectives, the present invention provides a lung cancer risk early warning system based on remote four diagnostic methods, comprising:

[0007] The acquisition module is used to acquire the original time-series data of inspection, auscultation, inquiry and palpation collected by remote terminals and wearable devices according to a preset acquisition period, and to extract lung cancer index data from the original time-series data.

[0008] A generation module, which is connected to the acquisition module, is used to perform self-supervised representation learning on the lung cancer index data within a preset rolling time window to generate individual temporal embedding vectors, and to perform unsupervised clustering and frequent subsequence mining based on the individual temporal embedding vectors to generate several abnormal pattern templates.

[0009] A calculation module, connected to the generation module, is used to calculate a dynamic alarm index based on the current preset rolling time window, the abnormal mode template, and the preset medical causal graph.

[0010] The early warning module, which is connected to the calculation module, is used to generate observational early warnings, enhanced sampling early warnings, or clinical recommendation early warnings based on the dynamic alarm index and the preset alarm range, and to perform evidence robustness verification before triggering the clinical recommendation early warning, and to issue a structured lung cancer early warning event form and a clinical confirmation instruction when the evidence robustness verification is passed.

[0011] The correction module, which is connected to the early warning module, the calculation module and the acquisition module respectively, is used to calculate the deviation parameter based on the receipt of the received structured lung cancer early warning event form and clinical confirmation data, and to adjust the preset alarm range or the preset acquisition cycle or the graph edge weight in the preset medical causal graph based on the deviation parameter.

[0012] Furthermore, the acquisition module includes:

[0013] The data cleaning unit is used to perform interpolation, denoising, and standardization preprocessing on the original time-series stream data to obtain preprocessed data;

[0014] A feature extraction unit, which is connected to the data cleaning unit, is used to extract the lung cancer indicator data from the preprocessed data;

[0015] The lung cancer indicator data include the nocturnal oxygen saturation drop index, the proportion of breath sounds and wheezing energy, the number of cough events per unit time, the weekly rate of weight loss, the asynchronous index of chest and abdominal movements, and the pulse wave conduction time.

[0016] Furthermore, the generation module is used to maximize the similarity of embedding vectors of adjacent data augmented views within the preset scrolling time window through a preset temporal contrastive learning model to generate the individual temporal embedding vector, and to minimize the similarity of embedding vectors of non-adjacent data augmented views within the preset scrolling time window through the preset temporal contrastive learning model to generate the individual temporal embedding vector.

[0017] Furthermore, the generation module is also used to cluster all the individual time-series embedding vectors using a density-based preset clustering algorithm, and to use a preset prefix projection algorithm to mine the frequent abnormal subsequences common to the user indicator data within the abnormal category clusters obtained by clustering, so as to generate several abnormal pattern templates.

[0018] Furthermore, the computing module includes:

[0019] A matching degree calculation unit is used to calculate the matching degree between the lung cancer indicator data of the current preset rolling time window and each of the abnormal pattern templates;

[0020] An index calculation unit, connected to the matching degree calculation unit, is used to calculate the dynamic alarm index based on the matching degree and the preset medical causal graph.

[0021] The preset medical causal graph is a directed weighted graph structure, including several nodes representing clinical concepts related to lung cancer and several edges representing the causal relationships between these clinical concepts.

[0022] Furthermore, the early warning module includes:

[0023] The early warning unit is used to compare the dynamic alarm index with the preset alarm range, and to generate the observation warning, the enhanced sampling warning, or the clinical recommendation warning based on the comparison result.

[0024] A verification unit, connected to the early warning unit, is used to check whether the dynamic alarm index of the current preset rolling time window is consistent with the change trend of the dynamic alarm index of the preset number of preset rolling time windows in the past within a preset verification time before generating the clinical recommendation early warning, and to issue the structured lung cancer early warning event form and the clinical confirmation instruction when the change trend is consistent.

[0025] Furthermore, the structured lung cancer early warning event sheet is a machine-readable data structure, whose data fields include a unique early warning identifier, a user identifier, a dynamic alarm index, a list of key lung cancer indicator data that triggers the early warning, time series segments of abnormal data, and the abnormal pattern template ID and medical causal graph path on which it is based.

[0026] Furthermore, the correction module is used to obtain receipts of all structured lung cancer early warning event forms and clinical confirmation data within a preset historical correction period, and to calculate performance deviation parameters and causal path deviation parameters based on the clinical confirmation data, and to adjust the preset alarm range or the preset collection period according to the performance deviation parameters, and to adjust the graph edge weights according to the causal path deviation parameters.

[0027] Furthermore, the correction module is used to count the total number of clinical suggestion warning events in the clinical confirmation data, the number of clinical suggestion warning events that are confirmed as false positives, and the number of data collection missing events, and to calculate the system false alarm rate and user data missing rate. Also, for each causal path in the preset medical causal graph, it counts the number of all triggered clinical suggestion warning events and the number of events that are confirmed as true positives, and calculates the accuracy of each causal path.

[0028] Furthermore, the correction module is also used to increase the preset alarm range when the system false alarm rate is greater than the preset false alarm rate threshold, and to decrease the preset collection cycle when the user data missing rate is greater than the preset missing rate threshold, and to decrease the corresponding graph edge weight when the accuracy is continuously less than the preset accuracy threshold, and to increase the graph edge weight when the accuracy is greater than the preset accuracy threshold and the graph edge weight is less than the preset edge weight upper limit.

[0029] Among these, the priority of false positive rate correction is higher than that of missing rate correction, and the priority of missing rate correction is higher than that of accuracy correction.

[0030] Compared with existing technologies, the beneficial effects of this invention are that, through multimodal temporal fusion and closed-loop regulation driven by causal graphs, a more robust, interpretable, and adaptive early warning capability for lung cancer risk is achieved. On the one hand, the system aggregates heterogeneous evidence from observation, auscultation, inquiry, and palpation into a dynamic alarm index according to time and causal path, and uses trend consistency and adversarial testing to filter transient noise, thereby significantly reducing false alarms caused by short-term fluctuations while maintaining sensitivity to real anomalies. On the other hand, based on three types of deviation parameters calculated from clinical feedback—false alarm rate, data missing rate, and path accuracy—a hierarchical nonlinear judgment strategy is adopted to adjust the alarm threshold, sampling frequency, and graph edge weights respectively. For example, when the false alarm rate is high, the trigger threshold is relaxed to improve specificity; when data is significantly missing, the sampling period is shortened to fill in key time sequences; and when the clinical confirmation rate of a certain path is low, the weight of that path is reduced to avoid amplifying erroneous causality. This inter-parameter linkage preserves the system's sensitivity to short-term anomalies while continuously calibrating causal weights through historical feedback, prompting the model output to converge on the balance between statistical consistency and physiological relevance. Simultaneously, event slips and contribution factor logs provide traceable explanatory evidence for manual review and posterior annotation. Overall, this system achieves a balance between reducing false alarms, lowering the risk of missed diagnoses, improving positive predictive values, and optimizing user sampling burden. It facilitates reliable and auditable clinical decision support in community screening, postoperative follow-up, and high-risk population monitoring, effectively addressing the problem of insufficient early warning accuracy caused by reliance on single, static medical record information leading to delayed risk identification.

[0031] Furthermore, by performing data cleaning such as interpolation, denoising, and standardization at the end-user side, and by specifically extracting lung cancer indicators such as the nocturnal oxygen saturation index, the proportion of wheezing energy in breath sounds, the number of coughs per unit time, the weekly rate of weight loss, the asynchrony index of chest and abdominal movements, and pulse wave transit time, the system can perform cross-validation and causal association judgment on weak signals from different channels after time-series alignment and characterization. For example, a continuous decrease in the nocturnal oxygen saturation index and asynchrony of chest and abdominal movements jointly indicate ventilation / gas exchange mismatch or focal obstruction; the high-frequency co-occurrence of wheezing energy in breath sounds and cough events suggests airway involvement or local stenosis; continuous weight loss amplifies the pathological significance of the above physiological abnormalities; and short-term changes in pulse wave transit time can reflect the systemic effects of sympathetic activation or hypoxic stress. Cleaning and feature extraction not only improve the temporal integrity and reliability of signals (reducing false positives caused by artifacts and missing data), but also enable the model to identify the physiological correlations between multimodal features and use them for weight adjustment. This significantly reduces false positives caused by single-channel noise while improving the early detection rate of abnormalities, ultimately achieving earlier, more interpretable and operable early warning outputs for suspected lung lesions, which facilitates subsequent enhanced sampling and clinical review.

[0032] Furthermore, by performing time-series comparative learning on multimodal lung cancer indicator sequences, various physiological signals such as nocturnal oxygen saturation decline, wheezing, cough frequency, weight loss rate, asynchrony of chest and abdominal movements, and pulse wave conduction time are uniformly represented. The generated individual time-series embedding vectors can reflect the collaborative characteristics and potential abnormal patterns of indicators over time, thereby achieving sensitive capture of abnormal lung function. Furthermore, by comparing rolling time windows and enhanced data views, the model can distinguish between short-term fluctuations and persistent changes, enabling the dynamic warning index to accurately reflect individual lung cancer risk. At the same time, it provides a reliable basis for the early warning module and correction module, realizing dynamic adjustment of early warning thresholds, collection cycles, and causal graph edge weights. This allows the system to capture anomalies while reducing the probability of false alarms and missed alarms, improving the overall prediction accuracy and response timeliness.

[0033] Furthermore, by employing density-based clustering and prefix projection algorithms, it is possible to effectively distinguish between normal fluctuations and potential abnormal patterns in individual lung cancer indicators, and identify recurring multimodal abnormal subsequences within specific time windows, such as the combined pattern of decreased blood oxygenation at night with increased cough frequency and abnormal weight loss rates. This approach incorporates the temporal correlation and mutual influence of multiple parameters, including changes in lung function, respiratory characteristics, and medical history indicators, into the analysis, thereby forming a reliable abnormal pattern template. This provides an accurate basis for calculating dynamic warning indices, enabling early and sensitive monitoring of lung cancer risk, and ensuring that the generated warnings are highly consistent with actual individual physiological trends, facilitating clinical intervention and continuous health management.

[0034] Furthermore, through the synergistic effect of the matching degree calculation unit and the index calculation unit, this system can quantitatively compare real-time collected lung cancer indicator data with historical abnormal patterns, and integrate the mutual influence between different indicators through the node and edge weighting of the causal graph to generate a dynamic warning index. This method can not only reflect the independent changes of each lung cancer-related parameter, but also capture the correlation effects between multimodal indicators. For example, a decrease in blood oxygen at night may cause heart rate fluctuations, and abnormal cough frequency may affect the coordination of chest and abdominal movements, thus forming a more accurate risk assessment. Ultimately, the dynamic warning index can comprehensively analyze the degree of abnormality and causal relationship of various physiological signals, providing timely and accurate risk warnings and a reliable basis for clinical intervention and personalized health management.

[0035] Furthermore, by comparing the dynamic alarm index with the preset alarm range, a tiered warning system for the user's health status is achieved. When the index is below the lower limit of the range, an observation warning is generated; in the middle range, an enhanced sampling warning is generated; and above the upper limit of the range, a clinical recommendation warning is generated. Before issuing a clinical recommendation warning, the verification unit compares the trend of the dynamic alarm index over several consecutive rolling time windows to ensure that the current abnormal changes are consistent with the past trends before issuing the warning event order and clinical confirmation instruction. This balances sensitivity and stability, allowing the rate of change, amplitude, and temporal relationship of different indicators to be reasonably mapped to alarm intensity and decision execution, improving the accuracy of warnings and reducing the risk of false alarms caused by occasional fluctuations.

[0036] Furthermore, the structured lung cancer early warning event sheet integrates the unique early warning identifier, user information, dynamic alarm index, major lung cancer indicators and their abnormal time series segments, as well as the triggering abnormal pattern template and the corresponding medical causal graph path into a machine-readable data structure through unified data fields. This enables the system to accurately determine abnormal patterns based on the magnitude, frequency, and temporal characteristics of indicator changes, automatically associate them with the corresponding causal nodes and paths, and efficiently track and reproduce the early warning logic in subsequent analysis or clinical intervention. This achieves precise mapping and dynamic feedback between indicator changes, abnormal patterns, and alarm decisions, improving the interpretability, accuracy, and reliability of early warnings.

[0037] Furthermore, by acquiring all structured lung cancer early warning event records, their receipts, and clinical confirmation data from past historical correction periods, the system's false alarm rate, user data missing rate, and accuracy of each causal path are analyzed to calculate performance deviation parameters and causal path deviation parameters. Based on the performance deviation parameters, the system can dynamically adjust the preset alarm range or collection period to ensure that alarms are neither overly sensitive leading to false alarms nor overlooking important anomalies. Based on the causal path deviation parameters, the system adjusts the edge weights in the medical causal graph to make the association between abnormal pattern templates and clinical concept nodes more accurate. This achieves continuous optimization and coordination among the dynamic alarm index, alarm range, collection period, and causal path edge weights, improving the overall reliability and response accuracy of the early warning system.

[0038] Furthermore, by quantitatively calculating the system's false alarm rate, user data missing rate, and the accuracy of each causal path, a multi-dimensional evaluation of the early warning performance is achieved. Based on these indicators, the preset alarm range, collection cycle, and graph edge weights are dynamically adjusted, so that a closed-loop feedback relationship is formed between abnormal changes in lung cancer indicators, data collection integrity, and causal graph signals. This effectively reduces false alarms and missed alarms while ensuring sensitivity, thereby improving the reliability and scientific nature of clinical early warning.

[0039] Furthermore, the early warning system is optimized through a tiered adaptive adjustment strategy: when the false alarm rate exceeds a set threshold, the maximum value of the preset alarm range is multiplied by a first adjustment coefficient to increase it, thereby reducing the false alarm probability; if the user data missing rate exceeds the threshold, the collection cycle is multiplied by a second adjustment coefficient to shorten it, thereby improving data integrity and continuity; when the causal path accuracy is consistently below the threshold, the corresponding graph edge weights are multiplied by a third adjustment coefficient to decrease them, thereby improving early warning sensitivity; conversely, when the accuracy is above the threshold and the graph edge weights have not reached their upper limit, they are multiplied by a fourth adjustment coefficient to increase them, thereby optimizing risk response capabilities. This adjustment mechanism dynamically corrects various parameters according to the priority order of false alarm rate, missing rate, and accuracy, achieving a coordinated balance between accuracy, sensitivity, and data integrity in the early warning system, thereby improving the reliability and practicality of lung cancer risk prediction. Attached Figure Description

[0040] Figure 1 This is a schematic diagram of the lung cancer risk early warning system based on remote four diagnostic methods in this embodiment;

[0041] Figure 2 This embodiment generates a judgment logic diagram for the early warning unit;

[0042] Figure 3 This embodiment demonstrates the logic diagram for the verification unit to issue structured lung cancer early warning event forms and clinical confirmation instructions.

[0043] Figure 4 This is a logic diagram for determining the edge weights of the graph in the correction module of this embodiment. Detailed Implementation

[0044] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0045] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0046] Please see Figure 1 The diagram shown is a schematic of a lung cancer risk early warning system based on remote four diagnostic methods in this embodiment. This embodiment provides a lung cancer risk early warning system based on remote four diagnostic methods, including:

[0047] The acquisition module is used to acquire the original time-series data of inspection, auscultation, inquiry and palpation collected by remote terminals and wearable devices according to a preset acquisition period, and to extract lung cancer index data from the original time-series data.

[0048] A generation module, which is connected to the acquisition module, is used to perform self-supervised representation learning on the lung cancer index data within a preset rolling time window to generate individual temporal embedding vectors, and to perform unsupervised clustering and frequent subsequence mining based on the individual temporal embedding vectors to generate several abnormal pattern templates.

[0049] A calculation module, connected to the generation module, is used to calculate a dynamic alarm index based on the current preset rolling time window, the abnormal mode template, and the preset medical causal graph.

[0050] The early warning module, which is connected to the calculation module, is used to generate observational early warnings, enhanced sampling early warnings, or clinical recommendation early warnings based on the dynamic alarm index and the preset alarm range, and to perform evidence robustness verification before triggering the clinical recommendation early warning, and to issue a structured lung cancer early warning event form and a clinical confirmation instruction when the evidence robustness verification is passed.

[0051] The correction module, which is connected to the early warning module, the calculation module and the acquisition module respectively, is used to calculate the deviation parameter based on the receipt of the received structured lung cancer early warning event form and clinical confirmation data, and to adjust the preset alarm range or the preset acquisition cycle or the graph edge weight in the preset medical causal graph based on the deviation parameter.

[0052] In this embodiment, the acquisition module obtains lung cancer indicator data through multi-terminal and multi-sensor collaborative acquisition and preprocessing on the device side: visual inspection is acquired by a front-facing or chest camera (resolution ≥1280×720, frame rate ≈30fps, ambient light ≥100lux), and human / thoracic ROI cropping, white balance and brightness normalization are performed on the device side, and chest displacement and visual cough events are derived using optical flow or keypoint methods; auscultation is acquired by a microphone array (sampling rate 16kHz or 44.1kHz, 16bit), and then processed... Features such as MFCC (Multi-Functional Calibration) with a bandpass filter of 50-8kHz, a frame length of 25ms, and a step size of 10ms are extracted and recorded from a VOC sensor (optional, sampling interval 5-15min, sensitivity at the ppb level); palpation is performed using a finger clip / wristband PPG and SpO2 module (PPG sampling 100-200Hz, SpO2 (blood oxygen saturation index) update 1Hz, resolution 0.1%), and synchronized with an IMU (50-100Hz) to detect motion artifacts; medical history taking is performed using a structured questionnaire or voice input. The system collects smoking history, cough duration, hemoptysis, past medical history, and weight (preferably automatically reported by a home electronic scale, with an accuracy of ±0.1kg, weekly reporting by default; if only self-reported data is used, the data source is marked). When the difference between the electronic scale and the self-report is >0.5kg, the system automatically marks the inconsistency and triggers a retest prompt. The system prioritizes the scale report in the model and uses the self-report as supplementary evidence. On-device quality control is performed (image quality Q_img≥0.8, audio SNR≥12dB, PPG motion artifact rate<10%). Only summarized feature vectors or statistics are uploaded (if necessary, ≤10s of original audio and video clips are uploaded for verification with user authorization). All data is time-stamped with UTC (ISO-8601, NTP time synchronization), encrypted on-device, and transmitted to the server via TLS1.2 / 1.3. Static storage uses AES-256 and records the device ID, collection location, and environmental metadata for quality control and re-collection guidance. This ensures the consistency and availability of key indicators such as weight with other diagnostic channels during the consultation while protecting privacy.

[0053] The preset acquisition cycle refers to the system collecting, summarizing, and reporting all four diagnostic modalities using a uniform time step. For example, information such as nighttime blood oxygen saturation, breath sounds, cough frequency, chest movement, and weight changes are all recorded and uploaded according to the same time rhythm. This cycle depends on the timeliness of lung cancer risk monitoring, device power consumption, and data bandwidth limitations, and is usually set between 5 and 15 minutes. In this embodiment, it is set to 5 minutes, which can ensure monitoring accuracy while taking into account transmission stability and energy consumption control. The preset rolling time window refers to a historical data window of a fixed time length used for calculation and modeling, which consists of data from multiple consecutive acquisition cycles. This time window depends on the disease progression rate, the need to smooth abnormal fluctuations, and the sequence length required for model training, and is usually set between 7 and 14 days. In this embodiment, it is set to 14 days, which can stably capture persistent abnormalities in lung cancer-related physiology and symptoms, and effectively reduce short-term interference. The preset alarm range refers to a graded threshold range defined based on a dynamic alarm index, used to determine the risk level and trigger corresponding early warning measures. This range depends on the balance between sensitivity and specificity, the risk distribution characteristics of different populations, and the ability to allocate clinical intervention resources. It is usually set between [0.4-0.8]. In this embodiment, it is set to [0.5, 0.7], which can effectively reduce the false alarm rate while ensuring early identification capability and improve the clinical reference value of the early warning results.

[0054] By employing multimodal temporal fusion and closed-loop regulation driven by causal graphs, a more robust, interpretable, and adaptive early warning capability for lung cancer risk is achieved. On one hand, the system aggregates heterogeneous evidence from clinical observation, auscultation, inquiry, and palpation into a dynamic alarm index based on time and causal path. It uses trend consistency and adversarial testing to filter transient noise, thereby significantly reducing false alarms caused by short-term fluctuations while maintaining sensitivity to real anomalies. On the other hand, based on three types of deviation parameters calculated from clinical feedback—false alarm rate, data missing rate, and path accuracy—a hierarchical nonlinear judgment strategy is used to adjust the alarm threshold, sampling frequency, and graph edge weights respectively. For example, when the false alarm rate is high, the trigger threshold is relaxed to improve specificity; when data is significantly missing, the sampling period is shortened to fill in key time series; and when the clinical confirmation rate of a certain path is low, the weight of that path is reduced to avoid amplifying erroneous causality. This linkage between parameters not only preserves the system's sensitive response to short-term anomalies but also continuously calibrates causal weights through historical feedback, prompting the model output to converge between statistical consistency and physiological relevance. At the same time, event slips and contribution factor logs provide traceable explanatory evidence for manual review and posterior annotation. Overall, this system strikes a balance between reducing false alarms, lowering the risk of missed diagnoses, improving positive predictive values, and optimizing the user's sampling burden. It facilitates reliable and auditable clinical decision support in community screening, postoperative follow-up, and high-risk population monitoring, effectively solving the problem of insufficient early warning accuracy caused by the reliance on single static medical record information, which leads to delayed risk identification.

[0055] Specifically, the acquisition module includes:

[0056] The data cleaning unit is used to perform interpolation, denoising, and standardization preprocessing on the original time-series stream data to obtain preprocessed data;

[0057] A feature extraction unit, which is connected to the data cleaning unit, is used to extract the lung cancer indicator data from the preprocessed data;

[0058] The lung cancer indicator data include the nocturnal oxygen saturation drop index, the proportion of breath sounds and wheezing energy, the number of cough events per unit time, the weekly rate of weight loss, the asynchronous index of chest and abdominal movements, and the pulse wave conduction time.

[0059] By performing data cleaning such as interpolation, denoising, and standardization at the end-user side, and by specifically extracting lung cancer indicators such as the nocturnal oxygen saturation index, the proportion of wheezing energy in breath sounds, the number of coughs per unit time, the weekly rate of weight loss, the asynchrony index of chest and abdominal movements, and pulse wave transit time, the system can perform cross-validation and causal association judgment on weak signals from different channels after time-series alignment and characterization. For example, a continuous decrease in the nocturnal oxygen saturation index and asynchrony of chest and abdominal movements jointly indicate ventilation / gas exchange mismatch or focal obstruction; the high-frequency co-occurrence of wheezing energy in breath sounds and cough events suggests airway involvement or local stenosis; continuous weight loss amplifies the pathological significance of the above physiological abnormalities; and short-term changes in pulse wave transit time can reflect the systemic effects of sympathetic activation or hypoxic stress. Cleaning and feature extraction not only improve the temporal integrity and reliability of signals (reducing false positives caused by artifacts and missing data), but also enable the model to identify the physiological correlations between multimodal features and use them for weight adjustment. This significantly reduces false positives caused by single-channel noise while improving the early detection rate of abnormalities, ultimately achieving earlier, more interpretable and operable early warning outputs for suspected lung lesions, which facilitates subsequent enhanced sampling and clinical review.

[0060] Specifically, the generation module is used to maximize the similarity of the embedding vectors of adjacent data augmented views within the preset scrolling time window through a preset temporal contrastive learning model to generate the individual temporal embedding vector, and to minimize the similarity of the embedding vectors of non-adjacent data augmented views within the preset scrolling time window through the preset temporal contrastive learning model to generate the individual temporal embedding vector.

[0061] In this embodiment, the generation module employs a contrastive learning model based on a temporal Transformer structure. The model includes an input layer, three multi-head self-attention encoding layers, and one fully connected embedding layer, with an output embedding vector dimension of 128. The input data is a multimodal lung cancer indicator sequence arranged according to a preset 5-minute acquisition cycle, including six categories of indicators: nocturnal oxygen saturation decline index, energy ratio of breath sounds and wheezing, number of cough events per unit time, weekly weight loss rate, asynchronicity index of chest and abdominal movements, and pulse wave conduction time. During the training phase, the model performs multimodal data augmentation operations on the input sequence, including random time axis jitter ±10%, amplitude scaling from 0.8 to 1.2 times, random Gaussian noise perturbation (standard deviation 0.01), and interval masking (maximum 2 time slices). After enhancing the sequence input model, the InfoNCE contrastive loss function is used to maximize the cosine similarity of the view embedding vectors corresponding to adjacent rolling time windows (window size of 60 minutes, sliding step size of 5 minutes) and minimize the embedding similarity of non-adjacent time windows, thereby generating embedding vectors that can represent the multimodal temporal dynamic features of an individual. During training, the Adam optimizer is used with an initial learning rate of 0.001, a batch size of 64, and 100 iterations to achieve model convergence. The resulting stable temporal embedding vectors are used for subsequent unsupervised clustering and frequent subsequence mining to identify abnormal pattern templates related to lung cancer.

[0062] By performing time-series comparative learning on multimodal lung cancer indicator sequences, various physiological signals such as nocturnal oxygen saturation decline, wheezing, cough frequency, weight loss rate, asynchrony of chest and abdominal movements, and pulse wave conduction time are uniformly represented. The generated individual time-series embedding vectors can reflect the collaborative characteristics and potential abnormal patterns of indicators over time, thereby achieving sensitive capture of abnormal lung function. Furthermore, through comparison of rolling time windows and enhanced data views, the model can distinguish between short-term fluctuations and persistent changes, enabling the dynamic warning index to accurately reflect individual lung cancer risk. At the same time, it provides a reliable basis for the early warning module and correction module, realizing dynamic adjustment of early warning thresholds, collection cycles, and causal graph edge weights. This allows the system to capture anomalies while reducing the probability of false alarms and missed alarms, improving the overall prediction accuracy and response timeliness.

[0063] Specifically, the generation module is further configured to cluster all the individual time-series embedding vectors using a density-based preset clustering algorithm, and to mine the common frequent abnormal subsequences in the user indicator data within the abnormal category clusters obtained by clustering using a preset prefix projection algorithm, so as to generate several abnormal pattern templates.

[0064] In this embodiment, the generation module first inputs the individual time-series embedding vectors obtained by all users within the rolling time window into a density-based clustering algorithm. By analyzing the density distribution of the vectors in the multidimensional feature space, time-series embedding vectors with similar abnormal features are grouped into the same cluster, thereby distinguishing between normal fluctuations and potential abnormal patterns. Subsequently, a preset prefix projection algorithm is applied to the user lung cancer index data sequence within each abnormal cluster to perform frequent subsequence mining. The system automatically identifies specific index change patterns that recur within the cluster, such as a persistent decrease in blood oxygen at night accompanied by an increase in the frequency of coughing or an abnormal rate of weight loss. Finally, several abnormal pattern templates are generated. These templates can capture both individual-specific abnormal features and reflect representative trends of abnormal lung function in the group, providing an accurate basis for subsequent calculation of dynamic alarm indices and triggering early warnings.

[0065] Prefix projection algorithm is a sequence pattern mining method used to efficiently discover frequently occurring subsequences from a large number of time series or event sequences. In this embodiment, the algorithm constructs a prefix tree structure for each lung cancer indicator data sequence within an anomaly cluster in chronological order, projects each prefix sequentially, counts its frequency of occurrence within the cluster, and removes subsequences below a preset threshold, thereby quickly identifying key recurring anomaly patterns within the cluster. This method can efficiently extract joint anomaly features of multimodal indicators, providing reliable data support for generating anomaly pattern templates.

[0066] By employing density-based clustering and prefix projection algorithms, this study effectively distinguishes between normal fluctuations and potential abnormal patterns in individual lung cancer indicators. It also identifies recurring multimodal abnormal subsequences within specific time windows, such as the combined pattern of decreased nocturnal blood oxygenation with increased cough frequency and abnormal weight loss rates. This approach incorporates the temporal correlations and interactions of various parameters, including changes in lung function, respiratory characteristics, and medical history indicators, into the analysis. This results in reliable abnormal pattern templates, providing accurate data for dynamic warning index calculations, enabling early and sensitive monitoring of lung cancer risk, and ensuring that the generated warnings are highly consistent with actual individual physiological trends, facilitating clinical intervention and continuous health management.

[0067] Specifically, the computing module includes:

[0068] A matching degree calculation unit is used to calculate the matching degree between the lung cancer indicator data of the current preset rolling time window and each of the abnormal pattern templates;

[0069] An index calculation unit, connected to the matching degree calculation unit, is used to calculate the dynamic alarm index based on the matching degree and the preset medical causal graph.

[0070] The preset medical causal graph is a directed weighted graph structure, including several nodes representing lung cancer-related clinical concepts and several edges representing the causal relationships between lung cancer-related clinical concepts. Here, Q = Σ[Pi × Σ(aij × Lj)], where Q is the dynamic alarm index, Pi is the matching degree between the current preset rolling time window and the i-th abnormal pattern template, aij is the graph edge weight between the i-th abnormal pattern template and the j-th downstream lung cancer-related clinical concept node directly associated with it, and Lj is the preset clinical concept risk value corresponding to the j-th lung cancer-related clinical concept node.

[0071] In this embodiment, the matching degree calculation unit compares the lung cancer indicator data collected within the current preset rolling time window with the feature sequences of each abnormal pattern template point by point, and calculates the similarity score by combining the weight coefficients of the indicators and the time series change trend. Specifically, for multi-dimensional indicators such as the nighttime blood oxygen saturation index, the proportion of breath sound and wheezing energy, the number of cough events per unit time, the weekly rate of weight loss, the asynchronous index of chest and abdominal movements, and the pulse wave conduction time, the system first normalizes each indicator, and then uses cosine similarity or dynamic time warping algorithms to calculate the similarity with the abnormal pattern template, and finally obtains a matching degree value that reflects the degree of matching between the current data and each abnormal pattern template, which is used for the subsequent generation of dynamic alarm index.

[0072] In this embodiment, the edge weights of the graph are preset weighted values ​​used to represent the causal strength between the abnormal pattern template and downstream lung cancer-related clinical concepts. Typically, these edge weights are set between 0.1 and 1.0. In this embodiment, the initial value is set to 0.5, and it can be dynamically adjusted based on historical clinical confirmation data and false positives to reflect the importance and reliability of each causal path in lung cancer risk prediction.

[0073] By leveraging the synergistic effect of the matching degree calculation unit and the index calculation unit, this system can quantitatively compare real-time collected lung cancer indicator data with historical abnormal patterns. Furthermore, it integrates the mutual influences between different indicators through node and edge weighting of the causal graph to generate a dynamic warning index. This method not only reflects the independent changes of various lung cancer-related parameters but also captures the correlation effects between multimodal indicators. For example, a decrease in blood oxygenation at night may cause heart rate fluctuations, and abnormal cough frequency may affect the coordination of chest and abdominal movements, thus forming a more accurate risk assessment. Ultimately, the dynamic warning index can comprehensively analyze the degree of abnormality and causal relationships of various physiological signals, providing timely and accurate risk warnings and offering a reliable basis for clinical intervention and personalized health management.

[0074] Please see Figure 2 As shown, this is a logic diagram for the early warning unit to generate an early warning in this embodiment. In this embodiment, the early warning module includes:

[0075] The early warning unit is used to compare the dynamic alarm index with the preset alarm range, and to generate the observation warning, the enhanced sampling warning, or the clinical recommendation warning based on the comparison result.

[0076] Please see Figure 3 As shown, this is a logic diagram for the verification unit to issue a structured lung cancer early warning event form and a clinical confirmation instruction in this embodiment. In this embodiment, the verification unit, which is connected to the early warning unit, is used to check whether the dynamic alarm index of the current preset rolling time window is consistent with the changing trend of the dynamic alarm index of the preset number of preset rolling time windows in the past within a preset verification time before generating the clinical recommendation early warning, and to issue the structured lung cancer early warning event form and the clinical confirmation instruction when the changing trend is consistent.

[0077] In this embodiment, when the dynamic alarm index is less than the minimum value of the preset alarm range, the warning unit generates an observation warning; when the dynamic alarm index is greater than or equal to the minimum value of the preset alarm range and less than or equal to the maximum value of the preset alarm range, the warning unit generates an enhanced sampling warning; when the dynamic alarm index is greater than the maximum value of the preset alarm range, the warning unit generates a clinical suggestion warning.

[0078] In this embodiment, the verification unit uses four consecutive preset rolling time windows as reference windows to monitor the changes in the dynamic alarm index of the current rolling time window in real time, and calculates the average trend deviation between the current index and the index of each rolling time window within the reference window. When the deviation between the current index and the reference trend is less than a preset trend deviation threshold, the trend is determined to be consistent, and the verification unit triggers the issuance of a structured lung cancer early warning event form and generates a clinical confirmation instruction. This ensures that the clinical recommendation warning is only issued when the continuous trend verification is successful, thereby reducing the risk of false alarms caused by occasional fluctuations.

[0079] A preset trend deviation threshold is used to determine whether the dynamic alarm index of the current rolling time window is consistent with the index change trend in the reference window. This threshold is usually set between 0.1 and 0.2, and in this embodiment, it is set to 0.15. By comparing whether the deviation between the current index and the average trend of the reference window is less than this threshold, it is determined whether the trends are consistent, thereby ensuring that clinical recommendation warnings are only issued when the continuous trend verification is passed, avoiding false alarms caused by a single abnormal fluctuation.

[0080] By comparing the dynamic alarm index with the preset alarm range, a tiered warning system for the user's health status is achieved. When the index is below the lower limit of the range, an observation warning is generated; in the middle range, an enhanced sampling warning is generated; and above the upper limit of the range, a clinical recommendation warning is generated. Before issuing a clinical recommendation warning, the verification unit compares the trend of the dynamic alarm index over several consecutive rolling time windows to ensure that the current abnormal changes are consistent with the past trends before issuing the warning event order and clinical confirmation instruction. This balances sensitivity and stability, allowing the rate of change, amplitude, and temporal relationship of different indicators to be reasonably mapped to alarm intensity and decision execution, improving the accuracy of warnings and reducing the risk of false alarms caused by occasional fluctuations.

[0081] Specifically, the structured lung cancer early warning event sheet is a machine-readable data structure whose data fields include a unique early warning identifier, a user identifier, a dynamic alarm index, a list of key lung cancer indicator data that triggers the early warning, time series segments of abnormal data, and the abnormal pattern template ID and medical causal graph path on which it is based.

[0082] The structured lung cancer early warning event form integrates a unique early warning identifier, user information, dynamic alarm index, major lung cancer indicators and their abnormal time series segments, as well as triggering abnormal pattern templates and corresponding medical causal graph paths into a machine-readable data structure through unified data fields. This enables the system to accurately determine abnormal patterns based on the magnitude, frequency, and temporal characteristics of indicator changes, automatically associate them with corresponding causal nodes and paths, and efficiently track and reproduce the early warning logic in subsequent analysis or clinical intervention. This achieves precise mapping and dynamic feedback between indicator changes, abnormal patterns, and alarm decisions, improving the interpretability, accuracy, and reliability of early warnings.

[0083] Specifically, the correction module is used to obtain receipts of all structured lung cancer early warning event forms and clinical confirmation data within a preset historical correction period, calculate performance deviation parameters and causal path deviation parameters based on the clinical confirmation data, adjust the preset alarm range or the preset collection period according to the performance deviation parameters, and adjust the graph edge weights according to the causal path deviation parameters.

[0084] The preset historical correction period refers to the length of the past time frame referenced by the system when calculating deviation parameters and correcting alarms. It is a continuous time window looking backward from the current time, used to collect receipts and clinical confirmation data from all past structured lung cancer warning event forms. This allows for the assessment of the system's false alarm rate, user data missing rate, and the accuracy of each causal path, generating performance deviation parameters and causal path deviation parameters. Typically, this period is set between 20 and 40 days, depending on the warning system's response requirements and data update frequency. In this embodiment, it is set to 30 days, ensuring sufficient data to reflect recent changes in users' health status and providing a reliable basis for calculating performance deviation parameters and causal path deviation parameters.

[0085] By acquiring all structured lung cancer early warning event records, their receipts, and clinical confirmation data from past historical correction periods, the system's false alarm rate, user data missing rate, and accuracy of each causal path are analyzed to calculate performance deviation parameters and causal path deviation parameters. Based on the performance deviation parameters, the system can dynamically adjust the preset alarm range or collection period to ensure that alarms are neither overly sensitive leading to false alarms nor miss important anomalies. Based on the causal path deviation parameters, the system adjusts the edge weights in the medical causal graph to make the association between abnormal pattern templates and clinical concept nodes more accurate. This achieves continuous optimization and coordination among the dynamic alarm index, alarm range, collection period, and causal path edge weights, improving the overall reliability and response accuracy of the early warning system.

[0086] Specifically, the correction module is used to count the total number of clinical suggestion warning events in the clinical confirmation data, the number of clinical suggestion warning events that are confirmed as false positives, and the number of data collection missing events, and to calculate the system false alarm rate and user data missing rate. In addition, for each causal path in the preset medical causal graph, it counts the number of all triggered clinical suggestion warning events and the number of events that are confirmed as true positives, and calculates the accuracy of each causal path.

[0087] In this embodiment, the correction module calculates the false positive rate by dividing the number of events confirmed as false positives by the total number of clinical suggestion warning events within a preset historical correction period; it also calculates the user data missing rate by dividing the number of missing data collection events by the preset number of collections; and for each causal path in the preset medical causal graph, it calculates the accuracy rate of the path by dividing the number of events confirmed as true positives in the clinical suggestion warning events triggered by that path by the total number of events triggered by that path. This provides a precise quantitative basis for subsequently adjusting the preset alarm range, collection period, and graph edge weights.

[0088] By quantitatively calculating the system's false alarm rate, user data missing rate, and accuracy of each causal path, a multi-dimensional evaluation of early warning performance is achieved. Based on these indicators, the preset alarm range, collection cycle, and graph edge weights are dynamically adjusted, so that a closed-loop feedback relationship is formed between abnormal changes in lung cancer indicators, data collection integrity, and causal graph signals. This effectively reduces false alarms and missed alarms while ensuring sensitivity, thereby improving the reliability and scientific nature of clinical early warning.

[0089] Please see Figure 4 As shown, this is the logic diagram for adjusting the graph edge weights by the correction module in this embodiment. In this embodiment, the correction module is further used to: increase the preset alarm range by multiplying the maximum value of the preset alarm range by a preset first adjustment coefficient when the false alarm rate of the system is greater than the preset false alarm rate threshold; decrease the preset collection period by multiplying the preset collection period by a preset second adjustment coefficient when the user data missing rate is greater than the preset missing rate threshold; decrease the corresponding graph edge weight by multiplying the graph edge weight by a preset third adjustment coefficient when the accuracy is continuously less than the preset accuracy threshold; and increase the graph edge weight by multiplying the graph edge weight by a preset fourth adjustment coefficient when the accuracy is greater than the preset accuracy threshold and the graph edge weight is less than the preset edge weight upper limit.

[0090] Among these, the priority of false positive rate correction is higher than that of missing rate correction, and the priority of missing rate correction is higher than that of accuracy correction.

[0091] The preset false alarm rate threshold is used to determine the system's tolerance for false alarms. It depends on the historical false alarm statistics of early warning events and is typically set between 0.05 and 0.2. In this embodiment, it is set to 0.1, which can reduce the interference of occasional false alarms on clinical decision-making while ensuring sensitivity. The preset first adjustment coefficient is used to increase the maximum value of the preset alarm range. It depends on the degree to which the false alarm rate deviates from the threshold and the system's response sensitivity. It is typically set between 1.0 and 1.5. In this embodiment, it is set to 1.2, which can appropriately expand the alarm range to reduce the probability of false alarms. The preset missing rate threshold is used to determine the standard for the completeness of user data collection. It depends on the proportion of missing data in historical data collection and is typically set between 0.05 and 0.3. In this embodiment, it is set to 0.15, which can ensure data continuity while avoiding the impact of short-term missing data on early warning judgment. The preset second adjustment coefficient is used to shorten the collection cycle. It depends on the degree of missing rate exceeding the standard and the data acquisition capability. It is typically set between 0.7 and 0.95. In this embodiment, it is set to 0.85, which can increase the data sampling frequency and improve the timeliness of early warnings. The preset accuracy threshold is used to determine the standard for predicting the performance of the medical causal graph. It depends on the historical clinical confirmation data and... The early warning matching condition is typically set between 0.7 and 0.95; in this embodiment, it is set to 0.8 to ensure that causal path correction is triggered only when predictive ability declines. A preset third adjustment coefficient is used to reduce the edge weights of the graph, depending on the extent to which the accuracy falls below a threshold; it is typically set between 0.7 and 0.95; in this embodiment, it is set to 0.85 to reduce the impact of unreliable paths on the dynamic alarm index and improve the system's judgment accuracy. A preset accuracy threshold is used to determine whether the causal path is reliable, depending on the clinically confirmed true positive rate; it is typically set between 0.7 and 0.95. In the example, the value is set to 0.9, which supports increasing the edge weight of the graph when the accuracy is too high. The preset upper limit of edge weight is used to limit the maximum value of the edge weight in the causal graph. It depends on the maximum sensitivity requirement of the system and is usually set between 1.0 and 2.0. In this embodiment, it is set to 1.5, which can prevent a single path from having too much impact on the dynamic alarm index. The preset fourth adjustment coefficient is used to increase the edge weight of the graph. It depends on the extent to which the accuracy exceeds the threshold and is usually set between 1.05 and 1.3. In this embodiment, it is set to 1.1, which can enhance the contribution of reliable paths to the dynamic alarm index and improve the early warning response capability.

[0092] The early warning system is optimized through a tiered adaptive adjustment strategy: when the false alarm rate exceeds a set threshold, the maximum value of the preset alarm range is multiplied by a first adjustment coefficient to increase it, thereby reducing the false alarm probability; if the user data missing rate exceeds the threshold, the collection cycle is multiplied by a second adjustment coefficient to shorten it, thereby improving data integrity and continuity; when the causal path accuracy is consistently below the threshold, the corresponding graph edge weights are multiplied by a third adjustment coefficient to decrease them, thereby improving early warning sensitivity; conversely, when the accuracy is above the threshold and the graph edge weights have not reached their upper limit, they are multiplied by a fourth adjustment coefficient to increase them, thereby optimizing risk response capabilities. This adjustment mechanism dynamically corrects various parameters according to the priority order of false alarm rate, missing rate, and accuracy, achieving a coordinated balance between accuracy, sensitivity, and data integrity in the early warning system, thus improving the reliability and practicality of lung cancer risk prediction.

[0093] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A lung cancer risk early warning system based on remote four-diagnosis information, characterized in that, The method comprises the following steps: An acquisition module is configured to acquire original time series flow data of pulse diagnosis, olfactory diagnosis, interrogation diagnosis and palpation diagnosis collected by a remote terminal and a wearable device according to a preset collection period, and extract lung cancer index data from the original time series flow data; A generation module, connected with the acquisition module, is configured to perform self-supervised representation learning on the lung cancer index data within a preset rolling time window to generate individual time series embedding vectors, and perform unsupervised clustering and frequent subsequence mining based on the individual time series embedding vectors to generate a plurality of abnormal pattern templates; A calculation module, connected with the generation module, is configured to calculate a dynamic alarm index according to the current preset rolling time window, the abnormal pattern templates and a preset medical causal graph; An early warning module, connected with the calculation module, is configured to generate an observation early warning, a reinforced sampling early warning or a clinical recommendation early warning according to the dynamic alarm index and a preset alarm range, perform evidence robustness verification before triggering the clinical recommendation early warning, and issue a structured lung cancer early warning event sheet and a clinical confirmation instruction when the evidence robustness verification is passed; A correction module, connected with the early warning module, the calculation module and the acquisition module respectively, is configured to calculate a deviation parameter according to the return and clinical confirmation data of the structured lung cancer early warning event sheet received, and adjust the preset alarm range, the preset collection period or the edge weight of the preset medical causal graph in the causal graph according to the deviation parameter. 2.The lung cancer risk early warning system based on remote four-diagnosis information according to claim 1, characterized in that, The acquisition module comprises: A data cleaning unit is configured to perform interpolation, denoising and standardization preprocessing on the original time series flow data to obtain preprocessed data; A feature extraction unit, connected with the data cleaning unit, is configured to extract the lung cancer index data from the preprocessed data; The lung cancer index data comprises a nocturnal oxygen saturation drop index, a wheezing sound energy proportion, a number of cough events per unit time, a body weight weekly drop rate, a chest and abdomen motion asynchrony index and a pulse wave conduction time. 3.The lung cancer risk early warning system based on remote four-diagnosis information according to claim 2, characterized in that, The generation module is configured to maximize the embedding vector similarity of data augmentation views in adjacent preset rolling time windows by a preset time series contrast learning model to generate the individual time series embedding vectors, and minimize the embedding vector similarity of data augmentation views in non-adjacent preset rolling time windows by the preset time series contrast learning model to generate the individual time series embedding vectors. 4.The lung cancer risk early warning system based on remote four-diagnosis information according to claim 3, characterized in that, The generation module is further configured to cluster all the individual time series embedding vectors by a preset clustering algorithm based on density, and mine frequent abnormal sub-sequences common in user index data in abnormal category clusters obtained by clustering using a preset prefix projection algorithm to generate a plurality of abnormal pattern templates. 5.The lung cancer risk early warning system based on remote four-diagnosis information according to claim 4, characterized in that, The calculation module comprises: A matching degree calculation unit is configured to calculate the matching degree of the lung cancer index data of the current preset rolling time window and each abnormal pattern template; An index calculation unit, connected with the matching degree calculation unit, is configured to calculate the dynamic alarm index according to the matching degree and the preset medical causal graph. The preset medical causal graph is a directed and weighted graph structure, including a plurality of nodes representing lung cancer related clinical concepts and a plurality of edges representing the causality between the lung cancer related clinical concepts. 6.The lung cancer risk early warning system based on remote four-diagnosis information according to claim 5, characterized in that, The early warning module comprises: The early warning unit is configured to compare the dynamic warning index with the preset warning range, and generate the observation early warning, the intensive sampling early warning or the clinical recommendation early warning according to the comparison result. The verification unit is connected with the early warning unit, and is configured to check whether the dynamic warning index of the preset rolling time window is consistent with the change trend of the dynamic warning indexes of a preset number of preset rolling time windows in the past within a preset verification time period before the clinical recommendation early warning is generated, and issue the structured lung cancer early warning event sheet and the clinical confirmation instruction when the change trend is consistent.

7. The lung cancer risk early warning system based on remote four-diagnosis information according to claim 6, characterized in that, The structured lung cancer early warning event sheet is a machine-readable data structure, and the data fields thereof include an early warning unique identifier, a user identifier, a dynamic warning index, a list of main lung cancer index data triggering the early warning, a time sequence segment of abnormal data, an abnormal mode template ID and a medical causal graph path. 8.The lung cancer risk early warning system based on remote four-diagnosis information according to claim 7, characterized in that, The correction module is configured to obtain the feedback and the clinical confirmation data of all the structured lung cancer early warning event sheets in a preset historical correction period, calculate a performance bias parameter and a causal path bias parameter based on the clinical confirmation data, adjust the preset warning range or the preset sampling period according to the performance bias parameter, and adjust the graph edge weight according to the causal path bias parameter. 9.The lung cancer risk early warning system based on remote four-diagnosis information according to claim 8, characterized in that, The correction module is configured to count the total number of the clinical recommendation early warning events, the number of the clinical recommendation early warning events confirmed as false positives and the number of data sampling missing events in the clinical confirmation data, calculate a system false positive rate and a user data missing rate, and count the number of all the clinical recommendation early warning events triggered for each causal path in the preset medical causal graph and the number of the events confirmed as true positives, and calculate the accuracy rate of each causal path. 10.The lung cancer risk early warning system based on remote four-diagnosis information according to claim 9, wherein, The correction module is further configured to increase the preset warning range when the system false positive rate is greater than a preset false positive rate threshold, decrease the preset sampling period when the user data missing rate is greater than a preset missing rate threshold, decrease the graph edge weight when the accuracy rate is continuously less than a preset accuracy rate threshold, and increase the graph edge weight when the accuracy rate is greater than a preset accuracy rate threshold and the graph edge weight is less than a preset upper limit of the edge weight. The priority of the false positive rate correction is higher than that of the missing rate correction, and the priority of the missing rate correction is higher than that of the accuracy rate correction.

Images