A physiological and pathological risk prediction and explanation output method, device, system and storage medium based on individualized dynamic health baseline and multi-day time series trend window

Abstract

The application discloses a physiological and pathological risk prediction and explanation output method, device and system based on individual dynamic health baseline and multi-day time series trend window and a storage medium, and belongs to the field of health monitoring and intelligent early warning. The application establishes an individual dynamic health baseline, extracts multi-dimensional trend features from a multi-day sliding window and performs credibility correction, generates a personalized adaptive threshold for abnormality determination and outputs an abnormal confidence vector, inputs the trend features and the abnormal confidence vector into a prediction model for risk prediction, and generates operable intervention suggestions through SHAP explainability analysis and SOP knowledge graph mapping. The application integrates five technical features of individual baseline, trend features, abnormal confidence vector, SHAP explanation and SOP knowledge graph, solves problems of early warning lag, individualization loss, one-sided feature analysis, prediction disconnection with SOP, and realizes a closed loop of monitoring-prediction-explanation-intervention.

Description

Technical Field

[0001] This invention belongs to the field of smart elderly care and time-series intelligent health monitoring technology, specifically involving a method, device, system and storage medium for pathological risk prediction and interpretable monitoring of elderly people in elderly care institutions. Background Technology

[0002] As my country's population continues to age, institutional care has become one of the core care scenarios for the elderly who are frail, disabled, and suffer from multiple diseases. Compared with the general population, elderly people living in nursing homes generally have significant differences in their physiological baseline, high rates of chronic disease comorbidities, and strong insidious disease progression, which places higher demands on the accuracy, foresight, and feasibility of health monitoring.

[0003] Current mainstream health monitoring systems for elderly care institutions are mostly based on IoT-based vital sign collection and rely on fixed threshold alarms as their core mechanism. However, they suffer from four systemic technical deficiencies in practical clinical care implementation, making it difficult to meet the needs of the elderly for refined health management: 2.1 The abnormal early warning mechanism is lagging behind, and the ability to capture progressive pathological risks is insufficient. Current monitoring systems generally use single-point static thresholds to determine abnormalities, triggering alarms only when physiological parameters exceed preset critical values. This mechanism can only identify acute abnormal events that have already occurred and cannot capture the gradual evolution of pathological risks during the compensatory and incubation periods. For example, the slow rise in body temperature in the early stages of respiratory infections, and fluid retention and compensatory increase in heart rate before the deterioration of heart failure, will all show a trend of deviation lasting for several hours to several days before reaching the clinical criteria for fever or emergency.

[0004] To address this issue, the industry has attempted improvements such as daily change rate alerts and daily machine learning risk scoring, but all of these have inherent limitations: daily change rates are easily affected by short-term physiological fluctuations and measurement noise, resulting in a high false alarm rate; traditional machine learning early warning models still use daily static vital signs as the core input, failing to fully explore long-term time-series evolution patterns, and the early warning window for risks is very limited, making it impossible to achieve true proactive intervention.

[0005] 2.2 Lack of individualized baseline adaptation and poor accuracy of unified judgment criteria The normal physiological range of the elderly is affected by multiple factors such as age, underlying diseases, long-term medication, and physical condition, resulting in significant individual differences. For example, the baseline resting heart rate of people with chronic heart disease can be maintained at 90-100 beats / min for a long time, which is significantly higher than the normal range of healthy elderly people; the baseline blood oxygen saturation of elderly people with COPD is also generally lower than the general standard.

[0006] To address this issue, existing technologies have developed various solutions, including population grouping thresholds, simple historical mean baselines, and coarse-grained adaptive thresholds. However, none of these solutions achieve truly individualized and accurate assessments: population grouping is based solely on coarse-grained divisions of age and gender, leaving significant individual differences within each group; simple historical mean baselines fail to perform steady-state screening of the data and do not differentiate between different physiological periods of day and night, rendering the baseline inadequate for reference; and coarse-grained adaptive thresholds passively adjust the range based solely on recent data, lacking data quality and risk assessment mechanisms. This can easily lead to the misjudgment of progressive pathological abnormalities as normal baseline drift, thereby masking risk signals and causing underreporting.

[0007] 2.3 The time-series features are too singular, and the patterns of risk evolution over multiple days are not fully explored. The monitoring data of vital signs and behaviors over several consecutive days contain multidimensional evolutionary patterns of the occurrence and development of pathological risks, but the existing system's utilization of time-series data is clearly insufficient.

[0008] Currently, the mainstream technical approaches can be divided into three categories: The first category only provides visualization of time series curves, relying entirely on nurses to manually interpret trends, which is extremely inefficient and cannot be scaled up for automatic early warning; the second category only extracts basic trend features such as slope and mean difference, which is single-dimensional and cannot characterize complex risk patterns such as parameter fluctuation disorders, circadian rhythm disruption, and multi-sign synergistic changes, and has weak resistance to data gaps and environmental noise; the third category uses end-to-end deep learning models such as LSTM and TCN to directly process raw time series data, but the internal logic of the model is uninterpretable, making it difficult to gain the trust of clinical nurses. At the same time, high-quality labeled data is scarce in elderly care scenarios, and the cost of model training and deployment is extremely high, making it difficult to promote on a large scale.

[0009] 2.4 The interpretability of the early warning output is weak, and it is not linked to standardized nursing procedures. Most current risk prediction systems only output a risk probability score or risk level, which is a typical "black box" output. Nursing staff cannot know the core basis of the warning, making it difficult to verify the rationality of the warning or quickly locate the root cause of the abnormality, resulting in low system adoption rate and trust.

[0010] To address the interpretability issue, existing solutions fall into three categories: risk scoring with accompanying trend charts, fixed-rule expert systems, and post-event explainable AI algorithms. However, none of these solutions form a complete closed loop from early warning to intervention. The accompanying charts only provide data display without offering causal explanations or treatment guidelines, requiring caregivers to make their own judgments. Fixed-rule expert systems generate early warnings based on general clinical rules, resulting in high rule base maintenance costs, an inability to adapt to individual differences and progressively ambiguous risks, and very limited coverage scenarios. Post-event explainable AI (such as SHAP and LIME) only outputs a list of feature importance, requiring professional knowledge to interpret the results. Furthermore, they do not establish a mapping relationship with standardized nursing operation procedures (SOPs) in elderly care institutions, making them unable to be directly converted into actionable nursing actions, resulting in a significant gap between early warning and intervention.

[0011] In summary, existing technologies cannot simultaneously achieve the four goals of "proactive risk warning, individualized and accurate judgment, multi-dimensional temporal feature mining, and interpretable closed-loop intervention," and are difficult to adapt to the complex actual care scenarios in elderly care institutions. A targeted technical solution is urgently needed to address these issues. Summary of the Invention

[0012] In view of this, and addressing the significant individual differences and coexisting multiple diseases among the elderly in nursing homes, this invention proposes a method for predicting and interpreting pathological risks in the elderly based on individual baselines and multi-day trend windows. This method aims to overcome common technical shortcomings of existing health monitoring systems, such as high false alarm rates and delayed identification of progressive risks. The pathological risk predictions mentioned in this paper are all supplementary suggestions and do not constitute medical diagnosis or treatment recommendations. The final clinical decision is made by medical staff. Specifically, the following core technical issues need to be addressed: 1. In complex elderly care scenarios, the data from multi-source heterogeneous sensors are characterized by strong noise and uneven confidence levels, leading to insufficient reliability in the extraction of time-series features; 2. The traditional fixed threshold judgment mechanism is rigid and cannot adapt to the differences in the physiological baseline of different elderly people, nor can it keep up with the dynamic evolution of individual physiological state. 3. The long-term evolution of progressive pathological risks such as the compensatory stage of heart failure and the latent infection stage is hidden, and traditional single-point detection mechanisms cannot capture early risk signals in advance; 4. The risk prediction conclusions lack interpretable evidence, cannot be effectively linked with the standardized nursing procedures (SOP) of elderly care institutions, and are difficult to directly guide the implementation of nursing interventions.

[0013] The core technical concept of this invention is as follows: using a time-segmented dynamic individual health baseline as a personalized reference benchmark, extracting multi-dimensional temporal trend features through a multi-day sliding window to capture the progressive pathological evolution pattern, combining adaptive personalized thresholds to improve the accuracy of abnormality judgment, and finally seamlessly connecting the prediction results with nursing SOPs through interpretability analysis to construct a closed-loop intelligent monitoring system that integrates "monitoring-prediction-interpretation-intervention".

[0014] The core difference between this invention and existing technologies lies in the fact that existing health anomaly detection schemes are mostly based on fixed thresholds or single-day static data for judgment, and there is no systematic solution that deeply couples five major technical features: individual dynamic baseline, multi-dimensional multi-day trend features, anomaly confidence vector, SHAP interpretable parsing, and SOP knowledge graph mapping. This invention organically integrates these five features through targeted technical design, with each module supporting each other and working together to achieve synergistic effects far exceeding the simple superposition of single technical features.

[0015] The specific technical solution is as follows: A method for predicting and interpreting physiological and pathological risks based on an individualized dynamic health baseline and a multi-day time-series trend window includes the following steps: Step 1: Input data acquisition and preprocessing. Obtain multi-source heterogeneous data from the health profiling system or IoT sensors of elderly care institutions, form a feature vector F(t) according to a unified time window t (5-60 minutes), and calculate the data confidence C_i(t) and data missing rate M_i(t) for each feature. Step 2: Individualized dynamic health baseline modeling, establishing a dynamic and segmented personal health baseline for each elderly person and each feature, including time period / scenario segmentation, steady-state sample screening, baseline parameter calculation and dynamic adaptive update mechanism, wherein the update rate α is determined by grid search on the validation set; Step 3: Multi-day sliding window trend feature extraction. For the prediction time t, take the past N days (3-14 days) as the sliding window W_N(t). Within this window, extract at least one of the following trend features: drift features, slope features, fluctuation features, persistent anomaly features, periodic disruption features, and multi-source collaborative features. Then, perform confidence correlation correction on all extracted trend features based on the average confidence score (mean(C_i)) and missing rate (mean(M_i)) of the data within the window. The specific correction formula is as follows: ;

[0016] In addition to the standard processing procedure under normal data quality conditions, this invention also sets up a supplementary handling branch for abnormal scenarios of sliding window data failure: when the data missing rate in the sliding window is higher than the preset failure threshold or the data confidence is lower than the preset failure threshold, the following methods can be used for handling: (1) Do not perform numerical correction on the multi-dimensional trend features, do not generate abnormal confidence vectors, and concatenate the original multi-dimensional trend features with the data confidence and data missing rate in the corresponding sliding window into a joint feature vector, directly input it into the prediction model, and execute the subsequent interpretation output steps; in order to avoid false negatives in high-risk states due to direct reduction of feature values; (2) Trigger parallel data anomaly and equipment inspection warnings, and terminate the subsequent anomaly judgment, risk prediction and interpretation output process corresponding to this sliding window.

[0017] Step 4: Personalized Threshold Adaptation and Anomaly Detection. This step abandons fixed thresholds and achieves individualized and dynamic anomaly detection. For feature i in time period p, its high and low anomaly detection thresholds are dynamically generated by individual baseline parameters: θ_high(i,p) = μ(i,p) + k·MAD(i,p), θ_low(i,p) = μ(i,p) - k·MAD(i,p), where the coefficient k is determined through table lookup or cross-validation optimization. When an anomaly is detected, this step also generates an anomaly confidence vector. This vector contains confidence values ​​for various anomaly types, corresponds to timestamps, and is combined into an anomaly confidence vector, which is then stored in the evaluation result table.

[0018] The anomaly confidence vector generated in step four is composed of the anomaly confidence scores corresponding to each of the multiple anomaly monitoring dimensions, used to quantify the degree of anomaly risk across multiple dimensions; each anomaly monitoring dimension is matched with an independent data source. For each anomaly monitoring scenario, the corresponding anomaly confidence score is calculated. The specific calculation paradigm is as follows: features are extracted from the original monitoring signals of the corresponding scenario to obtain an initial judgment result; then, the initial judgment result is corrected by combining associated auxiliary monitoring data, and the final anomaly confidence score corresponding to that dimension is output. The final anomaly confidence scores of all dimensions are combined to form the aforementioned anomaly confidence vector. The anomaly confidence score includes at least one of the following: cardiac anomalous confidence score, respiratory anomalous confidence score, fall anomalous confidence score, and resting anomalous confidence score. The confidence level for cardiac aberration is calculated by extracting the skewness, kurtosis, and high-frequency energy ratio of the original electrocardiogram or pulse waveform to calculate the signal quality score, and simultaneously calculating the deviation rate of the current instantaneous heart rate relative to the individualized dynamic baseline; when the signal quality score meets the preset threshold and the heart rate change conforms to the electrophysiological law, the confidence level for cardiac aberration approaches 1; The confidence level for respiratory abnormality is calculated by extracting the periodic features of the respiratory waveform through an autocorrelation function, obtaining an initial confidence level for respiratory abnormality based on the periodic features, and then attenuating and correcting the initial confidence level by combining body motion sensor data; when the accelerometer of the body motion sensor detects violent turning or large-amplitude movement, the confidence level of the respiratory signal is linearly reduced according to the intensity of the movement. The confidence level of the fall anomaly is determined based on the peak intensity of the sum of multi-axis acceleration vectors and the static persistence characteristics after the fall. The body's static time and weak kinetic energy after the fall impact are evaluated through a logistic regression model. If a regular step count signal is detected after the impact, the confidence level of the fall anomaly is corrected to an extremely low value. The static anomaly confidence level adopts multi-source sensor cross-validation logic, which integrates smart mattress pressure sensing data, room infrared sensor position signals, and wearable device body motion data; when the multi-source signals conflict in spatial location and behavioral logic, the confidence level corresponding to the current behavioral state is reduced.

[0019] Step 5: Predictive model construction and training. Construct a predictive model to predict the probability p_k(t,T) of a specific type of pathological risk occurring within the future time window T (24-72 hours). This step uses the trend features extracted in step 3 (after confidence correction) and the abnormal confidence vector generated in step 4 as input features. Step Six: Generate interpretable output and map it to SOPs. Transform model predictions into information that nursing staff can directly understand and execute. Use SHAP values ​​to analyze the contribution of each trend feature to the current risk prediction. Construct a knowledge graph of risk type-symptom / evidence-SOP actions. Automatically match the corresponding standardized intervention process in the nursing SOPs of elderly care institutions and generate an actionable list of review and intervention suggestions to achieve human-machine collaboration and closed-loop intervention. In this process, individual baseline modeling in step two provides a personalized reference system for trend feature extraction in step three; trend feature extraction in step three provides input for SHAP interpretation in step six; the anomaly confidence vector generated in step four provides key input features for the prediction model in step five; the prediction model in step five performs risk prediction based on the outputs of steps three and four; SHAP interpretation in step six provides interpretability based on the results of steps two, three, and four; and the SOP knowledge graph in step six generates intervention suggestions based on the Top-K contribution features generated by SHAP interpretation. Each step is interdependent and works synergistically to achieve a closed loop of "prediction-interpretation-intervention".

[0020] It also includes a device for predicting and interpreting pathological risks in elderly people in nursing homes based on individual baselines and multi-day trend windows.

[0021] It also includes a system for predicting and interpreting pathological risks in elderly people in nursing homes based on individual baselines and multi-day trend windows, which adopts an edge-cloud collaborative architecture.

[0022] It also includes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method of any one of claims 1 to 7.

[0023] Beneficial effects In the method described in this invention, the five technical features—individual baseline, trend characteristics, anomaly confidence vector, SHAP interpretation, and SOP knowledge graph—are not simply superimposed, but rather interdependent and synergistic, working together to achieve a closed loop of "prediction-intervention." (1) Individual baselines provide personalized references for trend characteristics. Individual baseline modeling (step two) provides a personalized reference frame for trend feature extraction (step three), ensuring that the trend features extracted are changes relative to the individual baseline, rather than general changes. Without an individual baseline, the trend features are general and do not consider individual differences, which will significantly reduce prediction accuracy.

[0024] (2) Trend features capture dynamic changes and provide input for SHAP interpretation. Trend feature extraction (Step 3) captures the dynamic changes in pathological risk, providing input for SHAP interpretation (Step 6). Without trend features, SHAP interpretation can only be based on static daily data and cannot explain the gradual evolution of risk.

[0025] (3) SHAP interpretation is based on individual baseline and trend characteristics, providing interpretability. SHAP interpretation (step six), based on individual baselines (step two) and trend characteristics (step three), provides interpretability, telling caregivers why an alert was issued. Without individual baselines and trend characteristics, SHAP interpretation is meaningless because the input to the object it interprets (the predictive model) is incomplete.

[0026] (4) Anomaly confidence vector quantification quantifies the degree of anomaly in multiple dimensions, providing key inputs for the prediction model. The anomaly confidence vector generation (step four) quantifies the degree of abnormality in the elderly across multiple dimensions (heartbeat, respiration, falls, stillness, etc.), providing key input features for the prediction model (step five). There is a close technical dependency between steps four and five: without the anomaly confidence vector generated in step four, the prediction model in step five would lack multi-dimensional anomaly quantification information and would be unable to accurately capture complex risk patterns.

[0027] (5) The SOP knowledge graph generates intervention suggestions based on the SHAP interpretation, thus achieving a closed loop. The SOP knowledge graph (step six) automatically matches the corresponding standardized intervention procedures in the nursing SOPs of elderly care institutions with the Top-K contribution features generated from the SHAP interpretation (step six), generating a list of actionable intervention recommendations. Without the SHAP interpretation, the SOP knowledge graph would not know which feature to intervene in.

[0028] (6) Validation of synergistic effect: This system integrates five core units for collaborative computation: individual dynamic health baseline, multi-dimensional trend features, anomaly confidence vector, SHAP interpretable analysis, and SOP medical knowledge graph. Compared with single-module or partial-module combinations, integrating all five units simultaneously optimizes the recognition model's AUC and F1 score, significantly reducing the false alarm rate for anomaly warnings. Experimental results demonstrate a mutually reinforcing coupling relationship among the five units, resulting in a synergistic gain in overall recognition performance exceeding the sum of the individual unit effects.

[0029] (7) In addition, during cold start, the baseline of the population is quickly initialized by matching age and chronic disease labels. At the same time, the baseline update is paused or frozen according to the pathological risk threshold and the unidirectional drift trend of the indicators. This shortens the system start-up waiting time and avoids the pollution of the health baseline by pathological deterioration data, thus continuously ensuring the accuracy of abnormal judgment and risk prediction. Attached Figure Description

[0030] Figure 1 Flowchart of a pathological risk prediction method based on individualized baseline and SOP mapping; Figure 1 The flowchart shown describes the complete processing chain from data acquisition to interpretable output, which includes six key steps: input data acquisition and preprocessing, individualized dynamic health baseline modeling, multi-day sliding window trend feature extraction, personalized threshold adaptation and anomaly detection, early prediction model construction and training, interpretable output generation and SOP mapping. Detailed Implementation

[0031] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Those skilled in the art can make various modifications, equivalent substitutions, or improvements to the following embodiments based on the disclosure of the present invention, all of which fall within the protection scope of the present invention.

[0032] Example 1: General Implementation Steps Step 1: Input Data Acquisition and Preprocessing Multi-source heterogeneous data is acquired from the health profiling system of elderly care institutions or directly from IoT sensors and nursing records. The data is divided into a feature vector F(t) according to a unified time window t (preferably 5-60 minutes, more preferably 15-30 minutes). Here, i is the feature index.

[0033] The acquired characteristics include the following four categories: physical signs (such as resting heart rate HR_r, heart rate variability HRV, core body temperature, blood oxygen saturation SpO2, respiratory rate RR, etc.); sleep behavior (such as number of times getting out of bed, number of times waking up at night, total sleep duration, daytime steps or activity equivalents, etc.); environmental factors (such as room temperature and humidity, CO2 concentration, noise level, etc.); and nursing records (such as medication adherence, food intake, water intake, urine output, chief complaints recorded by caregivers, etc.).

[0034] For each feature F_i(t), its data confidence C_i(t) and data missing rate M_i(t) are calculated simultaneously. All data are then standardized and imputed for missing values ​​before proceeding to the next module.

[0035] Step 2: Individualized Dynamic Health Baseline Modeling This step establishes a dynamic, segmented personal health "baseline" for each elderly person and each characteristic, which is the cornerstone of individualized judgment.

[0036] (1) Time period / scene segmentation: Divide the day into several time periods with relatively stable physiological states. For example, it can be divided into nighttime sleep period (22:00-06:00), daytime rest period (06:00-12:00, 14:00-18:00), and daytime activity period (12:00-14:00, 18:00-22:00).

[0037] (2) Steady-state sample screening: For the current prediction time t, back D days (e.g., D=30) are used to screen "steady-state data windows" in each time period. Screening conditions include: no acute health event records in the time period, missing data rate of the feature is lower than the threshold (e.g., <20%), and mean confidence level of the data is higher than the threshold (e.g., >0.8).

[0038] (3) Baseline parameter calculation: For each feature i and time period p, the baseline center value μ(i,p) (using the median, which is more robust to outliers), the baseline fluctuation measure MAD(i,p) (median absolute deviation, which is more robust to outliers), and the individualized normal range band [L(i,p), U(i,p)] are calculated based on the steady-state sample set.

[0039] (4) Dynamic adaptive update mechanism: When the steady-state sample quality is good for K consecutive days (e.g., K=7), baseline update is initiated. The baseline center value is updated using an exponential moving average. μ_new(i,p) = (1-α) · μ_old(i,p) + α · μ_recent(i,p) Where μ_new(i,p) represents the updated baseline center value, μ_old(i,p) represents the baseline center value before the update, and μ_recent(i,p) represents the baseline center value obtained from the most recent K consecutive days of high-quality steady-state samples; the update rate α is determined by optimization on the validation set through grid search, α∈{0.01, 0.02, …, 0.2}, with AUC (Area Under the ROC Curve) as the optimization objective. In a preferred embodiment, the infection risk prediction α=0.08, and the heart failure deterioration risk prediction α=0.12. The actual parameter values ​​are a range, determined according to the scenario.

[0040] The update rate α is also controlled by the data quality of the most recent period (e.g., 5 days): if the mean confidence level of the data in the most recent 5 days is below a threshold (e.g., <0.85), then α is reduced (e.g., α→α / 2), making the baseline update slower; if the data quality is good and the mean confidence level of the data in the most recent 5 days is greater than 0.9, then α is appropriately increased (e.g., α→3α / 2); if the mean confidence level of the data in the most recent 5 days is between 0.85 and 0.9, then α remains unchanged.

[0041] When performing baseline adaptive updates, if the pathological risk probability output by the prediction model is higher than the preset warning threshold, or if the system detects a continuous unidirectional drift trend, the baseline update mechanism is frozen to prevent pathological slow deterioration from contaminating the healthy baseline.

[0042] During the cold start phase, if the monitored subject has no historical steady-state data, a general health baseline matching the age and chronic disease label of the population is used as the initial baseline. Furthermore, during the baseline adaptive update process, if the probability of pathological risk output by the prediction model exceeds the safety threshold, the update of the baseline center value is paused until the pathological risk returns to the safe range.

[0043] Step 3: Extracting Trend Features from a Multi-Day Sliding Window For the prediction time t, take the past N days (N is configurable, such as 3, 7, or 14 days, preferably 7 days) as a sliding window W_N(t). Within this window, form the time series sequence for each feature i. Extract at least one trend feature from the following categories: Drift characteristics: Calculate the difference between each point in the sequence and the individual baseline μ(i,p) of the corresponding time period, and then extract the mean, cumulative sum (equivalent to AUC), and maximum duration of the difference sequence.

[0044] Slope-related features: Perform linear regression on the sequence to extract the slope and coefficient of determination R², which reflect the long-term trend and stability of parameter changes.

[0045] Fluctuation characteristics: Calculate the variance, coefficient of variation (CV), kurtosis, and sample entropy of the sequence to reflect the stability or disorder of the indicator.

[0046] Persistent abnormality characteristics: The number of days, proportion, and maximum consecutive abnormal period when the values ​​in the statistical sequence exceed the individual's normal range [L(i,p), U(i,p)].

[0047] Periodic disruption characteristics: Compare the baseline deviation between daytime and nighttime segments, calculate diurnal rhythm indicators (such as nighttime / daytime heart rate ratio), and assess their deviation from historical steady-state values.

[0048] Multi-source co-variant features: Calculate the correlation or degree of co-variant change of related features (such as body temperature and heart rate, activity level and heart rate) within a window.

[0049] Credibility Correction: For all extracted trend features, confidence is adjusted based on the average confidence and missing rate of the data within the window, using the following formula: trend_corrected = trend_raw × (1 - mean(M_i)) × mean(C_i) in: trend_corrected The corrected trend characteristic value. trend_raw These are the original trend characteristic values ​​(such as slope, variance, etc.). mean(M_i) The average missing rate (between 0 and 1) of the data within the window. mean(C_i) The average confidence level (between 0 and 1) of the data within the window. It is important to note that the extracted trend features can be uniformly converted into "deviation" or "relative drift" during calculation. For physiological indicators with a non-zero baseline value (such as the nighttime / daytime heart rate ratio), the trend feature trend_raw should be defined as the difference between the indicator and the historical steady-state baseline value (i.e., trend_raw = actual ratio - steady-state ratio) to achieve individualized center alignment of the features.

[0050] Based on the standard procedure of numerically correcting trend features according to the confidence level and missing rate of data within the sliding window when the data quality is normal, this embodiment considers that when the data quality significantly deteriorates, directly correcting the trend features by numerical attenuation may artificially suppress the true physiological deviation, easily misjudging low fluctuations caused by missing data or signal interruption as stable physiological states, posing a risk of underreporting of severe illness. Therefore, this embodiment sets up a tiered handling branch for abnormal scenarios of sliding window data failure.

[0051] When the average data missing rate within the sliding window exceeds the preset failure threshold, or the average data confidence level falls below the preset failure threshold, the following two handling methods can be selected based on the degree of data failure and business needs: The first approach is a downgraded prediction mode: This mode does not apply numerical weighting corrections to multi-dimensional trend features, preserving the original deviation values ​​of the trend features. It also skips the anomaly detection step based on individual dynamic health baselines, does not generate anomaly confidence vectors, and avoids introducing noise into the prediction model from abnormal results derived from low-quality data. The original multi-dimensional trend features are concatenated with the data confidence and missing data rate within the corresponding sliding window to form a joint feature vector, which is directly input into the prediction model for pathological risk prediction. The subsequent SHAP feature contribution analysis and standardized intervention process are then executed normally for output matching. This approach employs the core logic of feature decoupling and joint input, transforming data quality indicators into independent feature dimensions, decoupling them from the original physiological trend information. Taking a Gradient Boosting Tree (GBDT) type machine learning model as an example, after the joint feature vector is input into the model, the splitting node logic of the tree model allows the model to adaptively perceive the data quality of each dimension, dynamically reducing the decision weights of low-quality feature dimensions. Simultaneously, it calls other related vital signs with higher confidence across dimensions to participate in risk calculation. Even with local data gaps, this approach preserves true physiological deviation signals, avoids underreporting, and maximizes the continuity and accuracy of risk monitoring.

[0052] The second mode is alarm termination: This triggers parallel data anomaly alarms and equipment check alerts. The alert information is simultaneously pushed to the operation and maintenance management terminal and the nursing execution terminal, indicating that the current monitoring window data is insufficiently valid and requiring investigation of sensor device status, network connectivity, or device usage. Under this approach, the entire process of anomaly judgment, pathological risk prediction, and interpretation corresponding to this sliding window is terminated, and no risk probability results are output to avoid unreliable data misleading nursing decisions.

[0053] Taking a specific scenario as an example: Suppose an elderly person's original temperature trend slope over the past 7 days was 0.07℃ / day. If a sensor detaches, the average missing rate within the window spikes to 0.85, and the average confidence level drops to 0.1. If a conventional numerical attenuation correction method is used, the temperature slope will be incorrectly corrected to near 0, leading to a misjudgment of a stable, healthy state and a potential risk of missed fever reports. In this solution, the system does not manually attenuate the temperature slope but directly generates a joint feature vector [0.07, 0.85, 0.1], achieving multi-dimensional control and anomaly interception through the aforementioned hierarchical logic.

[0054] The two methods described above together constitute a tiered data fault tolerance mechanism, prioritizing monitoring continuity in minor failure scenarios and ensuring result reliability in severe failure scenarios, while covering both health risk monitoring and closed-loop equipment operation and maintenance needs.

[0055] Step 4: Personalized Threshold Adaptation and Anomaly Detection This invention abandons fixed thresholds and achieves individualized and dynamic anomaly detection.

[0056] (1) Personalized threshold generation: For feature i in time period p, its high and low anomaly judgment thresholds are dynamically generated by individual baseline parameters: Upper limit threshold: θ_high(i,p) = μ(i,p) + k·MAD(i,p), Lower threshold: θ_low(i,p) = μ(i,p) - k·MAD(i,p), Here, the coefficient k is an adjustable parameter. The initial value of k can be set according to the clinical significance of the feature (e.g., body temperature k=1.5, heart rate k=2.0), and can be adaptively adjusted according to the elderly person's chronic disease label, age, and recent steady-state data quality.

[0057] (2) Adaptive adjustment method for k value: The k value can be determined by one of the following methods; Method 1 (Lookup Table Method): Establish a lookup table mapping individual elderly person labels to k-value amplification factors. The final k-value = k_base × amplification factor 1 × amplification factor 2 × … . The lookup table is as follows:

[0058] Method 2 (Cross-validation optimization): Cross-validate the k values ​​on the historical labeled dataset and select the k value that results in the highest F1 score. The range of k values ​​is: {0.5, 1.0, 1.5, 2.0, 2.5, 3.0}.

[0059] In a preferred embodiment, the infection risk prediction k=1.5 (baseline k=1.0, chronic disease label=none, age=82 years → amplification factor 1.5), and the heart failure worsening risk prediction k=1.2 (baseline k=1.0, chronic disease label=chronic heart failure → amplification factor 1.2). The actual parameter values ​​are a range, determined according to the scenario.

[0060] (3) Threshold bandwidth dynamic adjustment: When the baseline model detects a slow and stable drift in the physiological state of the elderly and there is no other risk evidence, the system can allow the threshold bandwidth to shift with the baseline; conversely, when other risk evidence increases, the system can tighten the k value to improve the sensitivity to anomalies.

[0061] (4) Anomaly Confidence Vector Generation and Output: The core output of this step is the anomaly confidence vector, which will serve as one of the key input features of the prediction model in step five. For each anomaly type (abnormal heartbeat, abnormal breathing, abnormal fall, abnormal stillness, etc.), its anomaly confidence value (between 0 and 1) is calculated; finally, the anomaly confidence value is matched with the original timestamp to form the anomaly confidence vector auxiliary sensing signal: Confidence_Vector(t) = [C_cardiac(t), C_respiratory(t), C_fall(t), C_static(t), …] For different types of physiological and behavioral abnormalities, the system uses different multi-dimensional feature fusion algorithms to calculate the confidence component (value range 0-1). For example, C_cardiac(t) represents the confidence level for cardiac aberration, evaluated based on the Signal Quality Index (SQI) of the heart rate signal. The algorithm extracts the skewness, kurtosis, and high-frequency energy ratio of the original ECG or pulse waveform to calculate the signal quality score; simultaneously, it compares the current instantaneous heart rate with an individualized dynamic baseline. If the deviation rate is high and the heart rate change conforms to electrophysiological laws, the confidence level is close to 1.

[0062] C_respiratory(t) represents the confidence level for respiratory abnormalities. The periodicity of the respiratory waveform is detected using the autocorrelation function (ACR), and attenuation correction is performed in conjunction with body motion sensor data. If the body motion sensor (accelerometer) detects that the elderly person is in a state of violent turning over or large-scale activity, the confidence level of the respiratory signal is linearly reduced according to the intensity of the movement to eliminate false alarms of respiratory rate caused by body motion artifacts.

[0063] C_fall(t) represents the fall anomaly confidence level, determined based on the peak intensity of the sum of multiaxial acceleration vectors (SVM) combined with the "static persistence characteristics" after the fall. The algorithm uses a logistic regression model to evaluate the body's resting time and weak kinetic energy after the fall impact. If a regular step count signal is detected immediately after the impact, the confidence level of the fall anomaly is adjusted to an extremely low value (e.g., below 0.1). C_static(t) represents the confidence level for static anomalies. A multi-source sensor cross-validation logic is employed. It integrates the pressure sensors from the smart mattress, the position signals from the room's infrared (PIR) sensor, and the body motion data from the wearable device. If a conflict arises between the multi-source signals in terms of spatial location and behavioral logic (e.g., the infrared sensor indicates the elderly person is in the living room, while the mattress indicates the elderly person is in place), the confidence level of the current behavioral state is reduced.

[0064] The system concatenates the independent confidence components obtained above according to a preset dimensional order and strictly aligns them with the timestamps of the original feature extraction to generate an abnormal confidence vector Confidence_Vector(t), which is stored in the evaluation result table and directly passed to step five as the input feature of the prediction model.

[0065] Technical Results: The anomaly confidence vector quantifies the degree of anomaly in the elderly across multiple dimensions, providing rich risk signals for the predictive model. Experiments show that using the anomaly confidence vector as one of the input features of the predictive model improves both AUC and F1 scores to varying degrees compared to using only trend features.

[0066] Step 5: Predictive Model Building and Training The goal of constructing a predictive model is to predict the probability p_k(t,T) of a specific type of pathological risk (such as infection risk, heart failure worsening risk, dehydration risk, delirium risk) occurring within a future time window T (e.g., 24 hours, 48 ​​hours, 72 hours).

[0067] This step is technically dependent on step four: the prediction model in this step directly uses the anomaly confidence vector output from step four as one of the key input features. Together with the trend features extracted in step three (after confidence correction), they constitute the complete input feature vector of the prediction model. The input feature vector is constructed as follows: Feature_Vector(t) = [Trend_Features(t)∪Confidence_Vector(t)∪Context_Features(t)] Where Trend_Features(t) is the trend feature extracted in step three (after confidence correction), Confidence_Vector(t) is the abnormal confidence vector generated in step four, and Context_Features(t) is the context feature (such as age, chronic disease label, current time period, etc.).

[0068] Model selection: Supports multiple model architectures.

[0069] (1) Rule + Probability Fusion Model: The trend features extracted in step three and the abnormal confidence vector in step four are mapped to probability components of each risk category through preset clinical rules, and then the results are combined and weighted for output. This approach has strong interpretability. Example of the rules is as follows: Rule 1: Body temperature trend slope > 0.05°C / day AND heart rate abnormality confidence level > 0.7 AND duration ≥ 3 days → Infection risk probability +0.3. Rule 2: Cumulative weight gain over 3 days > 2kg AND resting heart rate trend slope > 1.5 beats / min / day AND respiratory abnormality confidence level > 0.6 → Heart failure worsening risk probability +0.4. Rule 3: Decrease in daytime activity > 30% AND increase in nighttime awakenings > 3 times / day AND resting abnormality confidence level > 0.5 → Delirium risk probability +0.25. (2) Machine learning models: such as Gradient Boosting Tree (GBDT) and Logistic Regression. The input is a complete feature vector (trend features + anomaly confidence vector + context features), and the output is the risk probability.

[0070] (3) Temporal deep learning models: such as Long Short-Term Memory Network (LSTM) and Temporal Convolutional Network (TCN). They can directly process raw or multi-layer feature sequences and model complex temporal dependencies.

[0071] Model training and updates: Supervised learning is performed using historical data (including feature sequences and labeled risk events). The model can employ an online learning strategy, continuously fine-tuning and updating it with new data.

[0072] Step Six: Generate Interpretable Output and Map SOP This step transforms model predictions into information that caregivers can directly understand and execute, which is key to achieving human-machine collaboration and closed-loop intervention.

[0073] 1. Contribution Analysis: Using model-independent interpretability methods (such as SHAP values) or the model's built-in feature importance mechanism, calculate the contribution of each trend feature to the current risk prediction. Select the Top-K features with the largest contributions.

[0074] Scientific Basis: According to recent scientific research (Nature Scientific Reports 2025, “Personalized health monitoring using explainable AI”), SHAP values ​​can quantify the contribution of each feature to the model’s predictions and provide global and local interpretability, significantly improving caregivers’ trust in and adoption of AI systems.

[0075] 2. Evidence Chain Construction: For each Top-K feature, trace its specific performance within the sliding window W_N(t), display its numerical change curve, and highlight the time period exceeding the individualization threshold.

[0076] 3. Natural Language Interpretation Generation: Convert Top-K features and their supporting evidence into concise natural language statements.

[0077] Example: "Increased risk of infection (probability 85%). Main reasons: 1. Body temperature has been slowly rising over the past 3 days and has been close to the upper limit for 10 consecutive hours. 2. Resting heart rate at night is 15% higher than baseline and changes in tandem with the increase in body temperature. 3. Daytime activity level is 30% lower than baseline." 4. SOP Action Mapping and Suggestion Output: Construct a knowledge graph of risk type-symptom / evidence-SOP action. Based on the predicted risk type and Top-K driving characteristics, automatically match the corresponding standardized intervention process in the nursing SOP of elderly care institutions, and generate an actionable review and intervention suggestion list.

[0078] The SOP knowledge graph construction method employs a triplet structure. <Subject, Predicate, Object> Entity types include: risk types (infection risk, heart failure worsening risk, dehydration risk, delirium risk); symptoms / evidence (elevated body temperature, increased heart rate, decreased activity level, fragmented sleep, etc.); SOP actions (retake body temperature, measure blood oxygen saturation, inquire about symptoms, increase monitoring frequency, notify the doctor, etc.). Relationship types include: has_symptom (Risk type → Symptoms / Evidence) triggers_action (Symptoms / evidence → SOP actions) requires_evidence (SOP Action → Evidence Requirements). The inference rule uses the IF-THEN rule. An example of the inference rule is as follows: IF Risk type = Infection risk AND Temperature trend slope > 0.05°C / day AND Confidence level of abnormal heartbeat > 0.7 AND Duration ≥ 3 days THEN Recommended action 1 = Recheck body temperature (2 hours apart), Recommended action 2 = Measure blood oxygen saturation, Recommended action 3 = Ask if there is any discomfort such as sore throat / cough; IF Risk Type = Worsening Heart Failure AND Cumulative Weight Gain over 3 Days > 2 kg AND Resting Heart Rate Trend Slope > 1.5 beats / min / day AND Confidence Level of Abnormal Respiration > 0.6 THEN Recommended Action 1 = Measure your weight daily (on an empty stomach in the morning); Recommended Action 2 = Assess the condition of lower limb edema; Recommended Action 3 = Inform your doctor to assess whether the dosage of diuretics and digitalis needs to be adjusted. Knowledge Graph Example: The constructed knowledge graph includes: 4 risk type nodes (infection risk, heart failure worsening risk, dehydration risk, delirium risk), approximately 30 symptom / evidence nodes (elevated body temperature, increased heart rate, decreased activity level, fragmented sleep, weight gain, shortness of breath, etc.), approximately 50 SOP action nodes (retest body temperature, measure blood oxygen, inquire about symptoms, increase monitoring, notify the doctor, adjust body position, restrict sodium intake, etc.), and a total of approximately 200 triplets.

[0079] Technical Results: In some simulation experiments, after using SOP knowledge graph mapping, the average response time of nursing staff from receiving warnings to starting intervention, as well as the incidence of adverse events (such as confirmed infection and acute exacerbation of heart failure), all decreased to varying degrees.

[0080] Example 2: Infection Risk Prediction In this embodiment, the technical solution of the present invention is described in detail using the prediction of infection risk in the elderly as an example.

[0081] Step 1: Data acquisition and preprocessing.

[0082] Multi-source heterogeneous data of elderly person A (male, 82 years old, suffering from chronic heart failure) over the past 30 days were obtained from a health profiling system of a nursing home. Features included: resting heart rate (HR_r), core body temperature, blood oxygen saturation (SpO2), number of nighttime awakenings, daytime steps, and water intake. The data were aggregated hourly (Δt=60 minutes) to form a feature vector F(t).

[0083] Step 2: Individualized dynamic health baseline modeling.

[0084] The day was divided into a nighttime sleep period (22:00-06:00) and a daytime activity period (06:00-22:00). For resting heart rate characteristics, steady-state samples from the past 30 days were selected from the nighttime sleep period (no records of acute health events, missing data rate <20%, mean confidence level >0.8). The baseline central value μ was calculated to be 92 beats / min, MAD = 8 beats / min, and the individualized normal range band was [76, 108] beats / min.

[0085] The baseline update rate α was optimized on the validation set using a grid search, and the α with the highest AUC was selected as 0.08.

[0086] Step 3: Extract trend features from a multi-day sliding window.

[0087] The past 7 days were used as a sliding window. Trend features of body temperature were extracted: drift features (mean difference from baseline was 0.35°C, cumulative sum was 2.45°C / day), slope features (slope = 0.07°C / day, R² = 0.73), and fluctuation features (CV = 1.8%). Simultaneously, trend features of resting heart rate were extracted: mean difference from baseline was 8.5 beats / min, and the nighttime / daytime heart rate ratio was 1.18 (historical steady-state value was 1.05).

[0088] The extracted trend features are corrected for confidence: Assuming the mean missing rate within the window is mean(M_i)=0.15 and the mean confidence level is mean(C_i)=0.9, then the corrected body temperature slope is 0.07×(1-0.15)×0.9=0.05355.

[0089] Step 4: Personalized threshold adaptation and anomaly detection, and generation of anomaly confidence vector.

[0090] For body temperature characteristics, setting k=1.5 (k_base=1.0 for elderly A without chronic disease labels, age 82 → amplification factor 1.5), then: - Upper threshold: θ_high = 36.8 + 1.5×0.2 = 37.1°C - Lower threshold: θ_low = 36.8 - 1.5×0.2 = 36.5°C If a body temperature exceeds the upper limit threshold for three consecutive days, it is considered abnormal.

[0091] Generate anomaly confidence vector: Calculate the confidence values ​​for abnormal heartbeat (0.72), abnormal respiration (0.68), abnormal fall (0.15), and abnormal rest (0.25). Combine these into anomaly confidence vector: Confidence_Vector(t) = [0.72, 0.68, 0.15, 0.25].

[0092] Step 5: Pre-prediction model construction and training, using the anomaly confidence vector from Step 4 as input features.

[0093] The Gradient Boosting Tree (GBDT) model is used, with the input being a complete feature vector (the trend feature vector extracted in step three (after confidence correction) plus the anomaly confidence vector generated in step four plus contextual features), and the output being the probability of infection risk within the next 48 hours. Simulated cross-validation results show significant improvements in AUC, F1, and false positive rate.

[0094] Step 6: Generate interpretable output and map SOPs.

[0095] The model predicts that elderly person A has an 87% chance of developing an infection within the next 48 hours. Using SHAP analysis, the top-3 contributing features are: body temperature trend slope (contribution 0.32), confidence level of abnormal heartbeats (contribution 0.28), and the degree of deviation of nighttime resting heart rate from baseline (contribution 0.22).

[0096] The generated natural language explanation reads: "Increased risk of infection (probability 87%). Main reasons: 1. Body temperature has been slowly rising over the past 3 days and has been close to the upper limit threshold for 10 consecutive hours. 2. The confidence level of abnormal heartbeats is 0.72, indicating a decrease in heart rate variability. 3. Resting heart rate at night is 15% higher than baseline, and this change is correlated with the increase in body temperature. 4. Daytime activity level is 30% lower than baseline." Simultaneously, based on the SOP knowledge base, an intervention recommendation list is generated: 1. Recheck body temperature (2-hour interval) 2. Measure and record blood oxygen saturation 3. Inquire whether the elderly person has discomfort such as sore throat or cough 4. Increase the frequency of fluid intake monitoring 5. Notify the on-duty doctor for assessment if necessary. Predicted outcome: The elderly patient developed a fever (38.1°C) 48 hours later and was diagnosed with a respiratory infection by a doctor. Thanks to the early warning, caregivers took timely intervention measures, and the condition was brought under control promptly.

[0097] Example 3: Prediction of the Risk of Heart Failure Worsening In this embodiment, the technical solution of the present invention is described in detail using the example of predicting the risk of worsening heart failure in the elderly.

[0098] Step 1: Data acquisition and preprocessing.

[0099] Multi-source heterogeneous data from the past 14 days were obtained from a health profiling system for elderly patient B (female, 79 years old, suffering from congestive heart failure). Features included: resting heart rate (HR_r), heart rate variability (HRV), respiratory rate (RR), number of nighttime awakenings, daytime steps, and weight. The data were aggregated hourly to form a feature vector F(t).

[0100] Step 2: Individualized dynamic health baseline modeling.

[0101] For resting heart rate characteristics, the baseline central value was calculated to be μ = 88 bpm and MAD = 7 bpm during the nighttime sleep period. For weight characteristics, the baseline central value was calculated to be μ = 65.5 kg and MAD = 0.8 kg during the morning measurement period.

[0102] The baseline update rate α was optimized through grid search, and the α=0.12 with the highest AUC was selected.

[0103] Step 3: Extract trend features from a multi-day sliding window.

[0104] The past 7 days were used as a sliding window. Trend characteristics of resting heart rate were extracted: slope = 2.1 beats / min / day, cumulative difference from baseline = 14.7 beats / min / day. Trend characteristics of weight were extracted: slope = 0.15 kg / day, cumulative difference from baseline = 1.05 kg / day. Trend characteristics of nighttime awakenings were extracted: slope = 0.8 times / day, reflecting decreased sleep quality.

[0105] The extracted trend features are adjusted for confidence: Assuming the mean missing rate within the window is mean(M_i)=0.12 and the mean confidence level is mean(C_i)=0.92, then the adjusted weight slope is 0.15×(1-0.12)×0.92=0.12144. Step 4: Personalized threshold adaptation and anomaly detection, and generation of anomaly confidence vector.

[0106] For weight characteristics, setting k=1.2 (elderly person B's chronic disease label = chronic heart failure → amplification factor 1.2), then: - Upper threshold: θ_high = 65.5 + 1.2×0.8 = 66.46kg - Lower threshold: θ_low = 65.5 - 1.2×0.8 = 64.54kg If a person's weight exceeds the upper limit for three consecutive days, it is considered abnormal (indicating fluid retention).

[0107] Generate anomaly confidence vector: Calculate the confidence values ​​for abnormal heartbeat (0.65), abnormal respiration (0.78), abnormal fall (0.12), and abnormal rest (0.35). Combine them into anomaly confidence vector: Confidence_Vector(t) = [0.65, 0.78, 0.12, 0.35].

[0108] Step 5: Pre-prediction model construction and training, using the anomaly confidence vector from Step 4 as input features.

[0109] A rule-based and probability-based fusion model is used. The rule includes: "Cumulative weight gain over 3 days > 2 kg AND resting heart rate trend slope > 1.5 beats / min / day AND respiratory abnormality confidence > 0.6 → probability of heart failure worsening + 0.4". The combined weighted output risk probability is then used.

[0110] Cross-validation results showed that AUC, F1, and false positive rate all improved to varying degrees.

[0111] Step 6: Generate interpretable output and map SOPs.

[0112] The model predicts an 82% probability of heart failure worsening in elderly person B within the next 72 hours. The generated natural language explanation is: "Increased risk of heart failure worsening (probability 82%). Main reasons: 1. Weight has been increasing over the past 3 days, accumulating an increase of 1.2 kg, exceeding the individualized threshold. 2. Abnormal breathing confidence level reaches 0.78, indicating increased respiratory rate and shallow breathing. 3. Resting heart rate shows a continuous upward trend, and nighttime HRV decreases, suggesting impaired autonomic nervous system function. 4. Increased nighttime awakenings and decreased sleep quality, possibly related to nighttime dyspnea." Simultaneously, by matching the SOP knowledge base, an intervention recommendation list is generated: 1. Measure weight daily (fasting in the morning) 2. Assess lower extremity edema 3. Adjust sitting / lying position 4. Restrict sodium intake 5. Inform the doctor to assess whether the diuretic and digitalis dosages need to be adjusted Predicted outcome: The elderly patient developed shortness of breath 60 hours later, and was assessed by a doctor as having worsening heart failure (NYHA classification progressing from II to III). Due to early warning and timely intervention, emergency hospitalization was avoided.

[0113] Example 4: Dehydration Risk Prediction In this embodiment, the technical solution of the present invention is described in detail using the example of predicting the risk of dehydration in the elderly.

[0114] Step 1: Data acquisition and preprocessing.

[0115] Multi-source heterogeneous data of elderly C (female, 85 years old, with diabetes) over the past 14 days were obtained from a health profiling system. Features included: water intake, urine output, skin elastin score, daily steps, and daily activity level. The data were aggregated every 4 hours to form a feature vector F(t).

[0116] Step 2: Individualized dynamic health baseline modeling.

[0117] For water intake characteristics, the baseline center value was calculated to be μ=1200ml / day and MAD=150ml / day during the daytime activity period (06:00-22:00), and the individualized normal range was [750, 1650]ml / day.

[0118] The baseline update rate α was optimized through grid search, and the α with the highest AUC was selected as 0.05.

[0119] Step 3: Extract trend features from a multi-day sliding window.

[0120] The past 7 days were used as a sliding window. Trend characteristics of water intake were extracted: drift characteristics (mean difference from baseline was -180 ml / day), slope characteristics (slope = -25.7 ml / day², R² = 0.81), and fluctuation characteristics (CV = 12.3%). Simultaneously, trend characteristics of urine output were extracted: the mean difference from baseline was -90 ml / day, reflecting a decrease in urine output.

[0121] The extracted trend features are adjusted for confidence: Assuming the mean missing rate within the window is mean(M_i)=0.08 and the mean confidence level is mean(C_i)=0.95, then the adjusted water consumption slope is (-25.7)×(1-0.08)×0.95=-22.47.

[0122] Step 4: Personalized threshold adaptation and anomaly detection, and generation of anomaly confidence vector.

[0123] For water intake characteristics, setting k=1.0 (for elderly C who has no relevant chronic disease label, the baseline k value is used), then: - Lower threshold: θ_low = 1200 - 1.0×150 = 1050ml / day If a person's water intake is below the lower threshold for three consecutive days, it is considered abnormal (indicating possible dehydration).

[0124] Generate anomaly confidence vector: Calculate the confidence values ​​for abnormal heartbeat (0.18), abnormal respiration (0.22), abnormal fall (0.08), and abnormal rest (0.42). Combine them into anomaly confidence vector: Confidence_Vector(t) = [0.18, 0.22, 0.08, 0.42].

[0125] Step 5: Pre-prediction model construction and training, using the anomaly confidence vector from Step 4 as input features.

[0126] The Gradient Boosting Tree (GBDT) model was used, with the input being a complete feature vector (the trend feature vector extracted in step three (after confidence correction) plus the anomaly confidence vector generated in step four plus contextual features), and the output being the probability of dehydration risk occurring within the next 48 hours. The model training data included 300 historical dehydration events and 900 non-dehydrated samples. Cross-validation results showed AUC=0.89, F1=0.83, and a false positive rate of 9.7%.

[0127] Step 6: Generate interpretable output and map SOPs.

[0128] The model predicts that elderly person C has a 75% chance of developing dehydration within the next 48 hours. Using SHAP analysis, the top-3 contributing features are: slope of water intake trend (contribution 0.38), resting anomaly confidence (contribution 0.28), and degree of reduction in urine output (contribution 0.20).

[0129] The generated natural language explanation reads: "Increased risk of dehydration (75% probability). Main reasons: 1. Fluid intake has been decreasing over the past 3 days and has fallen below the individualized lower limit threshold. 2. The resting abnormality confidence score is 0.42, indicating a significant decrease in daytime activity, which may be related to fatigue. 3. Urine output has decreased synchronously, which may be related to dehydration." Simultaneously, by matching the SOP knowledge base, an intervention recommendation list is generated: 1. Assess the elderly person's skin elasticity and oral moisture; 2. Encourage increased fluid intake (in small, frequent sips); 3. Record 24-hour fluid intake and output; 4. Notify the nurse to assess whether intravenous fluid replacement is needed. Predicted outcome: The elderly patient developed signs of dehydration, including decreased skin elasticity and dry oral mucosa, after 36 hours. The nurse assessed the condition as mild dehydration. Due to early warning, timely fluid replacement measures were implemented, preventing the dehydration from worsening.

Claims

1. A method for predicting and interpreting physiological and pathological risks based on an individualized dynamic health baseline and a multi-day time-series trend window, characterized in that, include: (1) Acquire multi-source heterogeneous health monitoring data, form characteristic time series data, and simultaneously calculate the data confidence and data missing rate of the characteristic time series data; (2) Divide the day into multiple physiological periods according to physiological state, select steady-state samples from the characteristic time series data in each period, construct an individual dynamic health baseline based on the steady-state samples, and update it adaptively; (3) Using the individual dynamic health baseline as a reference, extract the multi-dimensional trend features of the feature time series data within a sliding window with configurable number of days, and correct the multi-dimensional trend features based on the data confidence and missing rate within the sliding window; (4) Based on the personalized threshold generated by the individual dynamic health baseline, anomaly determination is performed on the feature time series data, and in the case of anomalies, an anomaly confidence vector containing multi-dimensional anomaly values ​​is output. (5) Input the corrected trend features and the abnormal confidence vector into the prediction model and output the probability of the occurrence of pathological risk within the future time window; (6) Analyze the occurrence probability through SHAP value, combine it with pathological risk type, symptoms, SOP knowledge graph to match standardized intervention process, and output explanation and intervention suggestions.

2. The method according to claim 1, characterized in that, In step (1), the source of the multi-source heterogeneous health monitoring data is selected from at least one of the following: a structured database storing historical physiological signs and chronic disease tags, IoT sensors, and nursing records; the features include at least one of the following: physical signs features and sleep behavior features, and may also selectively include environmental features and nursing record features; wherein, physical signs features are composed of physiological parameter monitoring data, representing the basic physiological state of the elderly; sleep behavior features are composed of behavior monitoring data, representing the sleep and daily activity patterns of the elderly; environmental features are composed of environmental monitoring data, representing the environmental state of the elderly; and nursing record features are composed of nursing entry data, representing the subjective symptoms and objective nursing indicators of the elderly.

3. The method according to claim 1, characterized in that, In step (2), the individual dynamic health baseline is established and dynamically updated in the following manner: The day is divided into multiple physiological periods based on physiological state, including at least the nighttime sleep period, the daytime rest period, and the daytime activity period; For each type of physiological period, historical data within a preset period is retrieved, and data segments for the corresponding period are filtered out day by day. Data segments that meet the following conditions are included in the steady-state sample set: there are no records of acute health events in the data segment, the data missing rate is lower than the preset missing rate threshold, and the mean confidence level of the data is higher than the preset confidence level threshold. For each feature and each physiological time period, baseline parameters are calculated based on the corresponding steady-state sample set: the baseline center value is determined by the sample median, and the baseline fluctuation measure is determined by the median absolute deviation. The baseline employs an adaptive update mechanism: when the steady-state sample quality meets the preset standard for the first preset number of consecutive days, the baseline parameters are updated. The update rate is dynamically adjusted based on the mean confidence level of the steady-state samples within the most recent second preset number of consecutive days: when the mean confidence level is lower than the first preset threshold, the update rate is reduced to slow down the baseline iteration speed; When the average confidence level is higher than the second preset threshold, the update rate is increased by a preset amount; when the two thresholds are between, the current update rate remains unchanged.

4. The method according to claim 3, characterized in that, When performing baseline adaptive updates, if the pathological risk probability output by the prediction model is higher than the preset warning threshold, or if the system detects a continuous unidirectional drift trend, the baseline update mechanism is frozen to prevent pathological slow deterioration from contaminating the healthy baseline.

5. The method according to claim 3, characterized in that, During the cold start phase, if the monitored subject has no historical steady-state data, a general health baseline matching the age and chronic disease label of the population is used as the initial baseline. Furthermore, during the baseline adaptive update process, if the probability of pathological risk output by the prediction model exceeds the safety threshold, the update of the baseline center value is paused until the pathological risk returns to the safe range.

6. The method according to claim 1 or 3, characterized in that, The multi-dimensional trend features are at least one of the following: drift features, slope features, fluctuation features, persistent anomaly features, periodic disruption features, and multi-source synergy features, wherein: Drift-type features are generated based on the difference sequence between the feature time-series data within a sliding window and the individual's dynamic health baseline, representing the overall offset direction and cumulative offset of the feature relative to the baseline; Slope-type features are generated by linear regression of the numerical sequence within a sliding window based on the time-series data of the features, characterizing the rate of change and trend stability of the features; Fluctuation-type features are generated based on the statistical characteristics of the numerical sequence of the feature time series data within a sliding window, representing the degree of fluctuation and disorder of the feature. Persistent anomaly features are statistically generated based on the time-series data of the features, where the values ​​exceed the normal range within a sliding window, thus characterizing the persistence and severity of the abnormal state. Periodic disruption features are generated based on the difference in the degree of deviation of the time series data of the features from the individual's dynamic health baseline in different time periods, and characterize the degree of disorder of physiological rhythms. Multi-source collaborative features are generated based on the trend correlation analysis of multiple sets of associated monitoring features corresponding to preset disease risks within a sliding window, representing the collaborative pattern of abnormal changes in multiple features; The confidence level of the multi-dimensional trend features is corrected by weighting the data confidence and missing data rate within the corresponding sliding window, so that the corrected trend feature values ​​are positively correlated with the data quality within the window.

7. The method according to claim 1, characterized in that, In step (3), when the data missing rate in the sliding window is higher than the preset failure threshold or the data confidence is lower than the preset failure threshold, the following methods are adopted: (1) No numerical correction is made to the multi-dimensional trend features, no abnormal confidence vector is generated, the original multi-dimensional trend features are concatenated with the data confidence and data missing rate in the corresponding sliding window to form a joint feature vector, which is input into the prediction model and the subsequent interpretation output steps are executed; (2) Parallel data anomaly and equipment inspection warnings are triggered.

8. The method according to claim 1, characterized in that, In step (4), the personalized threshold is dynamically generated based on the baseline center value and baseline fluctuation measure of the individual's dynamic health baseline; Based on the baseline center value, the upper and lower thresholds are determined by combining the baseline fluctuation measurement and the preset offset coefficient to form an individualized normal range band, which is used to perform individualized anomaly judgment on each monitored feature in the feature time series data, and a multi-dimensional anomaly confidence vector is generated based on the anomaly judgment result as the input feature of the prediction model. The anomaly confidence vector is composed of anomaly confidence scores corresponding to multiple anomaly monitoring dimensions, and is used to quantify the degree of anomaly risk in multiple dimensions.

9. The method according to claim 1, characterized in that, The input feature vector of the prediction model also includes features characterizing the elderly person's baseline condition and the current monitoring period.

10. The method according to claim 1 or 9, characterized in that, The SHAP value is used to select the feature set with the highest contribution from all input features fed into the prediction model; combined with the temporal changes of features within the sliding window and personalized threshold exceeding information, the feature set is used as symptom evidence for the corresponding predicted pathological risk type; relying on the triple SOP knowledge graph containing pathological risk, symptom evidence, and nursing operations, the symptom evidence is mapped to standardized nursing intervention actions, generating an output result containing feature explanation statements and a list of intervention actions.

11. A device for predicting and interpreting pathological risks in elderly people in nursing homes based on individual baselines and multi-day trend windows, characterized in that, include: The data acquisition and preprocessing module is used to acquire multi-source heterogeneous health monitoring data, form feature vectors according to preset time windows, and simultaneously calculate the data confidence and data missing rate of each feature. The individual dynamic health baseline modeling module is used to filter steady-state samples by time period based on the feature vector, and to establish and adaptively update the individual dynamic health baseline. The configurable-day sliding window trend feature extraction module is used to extract multi-dimensional trend features within a configurable-day sliding window, with the individual dynamic health baseline as a reference, and to perform confidence correction based on the confidence level and missing rate of the data within the window. The personalized threshold adaptation and anomaly detection module is used to generate personalized thresholds based on the individual dynamic health baseline, perform anomaly detection based on the personalized thresholds, and output an anomaly confidence vector containing multi-dimensional anomaly values. The advance prediction model building and training module is used to input the confidence-corrected trend features and the abnormal confidence vector into the prediction model and output the probability of occurrence of a specific pathological risk within the future time window. The interpretability output generation and SOP mapping module is used to analyze feature contribution through SHAP values, and combine the risk type-symptom-SOP knowledge graph to match standardized intervention processes, outputting explanations and intervention recommendations.

12. A system for predicting and interpreting pathological risks in elderly people in nursing homes based on individual baselines and multi-day trend windows, characterized in that, Adopting an edge-cloud collaborative architecture, including: On the sensor side, deployed in the room of the monitored person and on their person, including wearable wristbands, smart mattresses and environmental sensors, are used to collect multi-source heterogeneous data; The edge computing gateway, deployed locally in the elderly care facility, communicates with the sensor terminal and is configured to perform data preprocessing, individual dynamic health baseline modeling, trend feature extraction and anomaly detection, and trigger local early warning using local cached data when the network is disconnected. The cloud platform, which communicates with the edge computing gateway, is configured to perform predictive model inference, SHAP value parsing, and SOP knowledge graph mapping, and maintain the SOP knowledge graph and audit evidence data. The terminal layer includes caregiver terminals, which are connected to the cloud platform to receive early warning information and a list of intervention suggestions, and to transmit the treatment results back.

13. A computer-readable storage medium storing a computer program, characterized in that, When the program is executed by the processor, it implements the method of any one of claims 1 to 10.

Images