Delirium risk assessment method

Abstract

The invention relates to a delirium risk assessment method. Aiming at delirium multiple induction factors including infection, drug use, metabolic disorder and sleep disorder, an integrated multi-dimensional data acquisition system is designed; data, including physiological parameters of heart rate variability, blood oxygen level and body temperature fluctuation, biochemical indexes including inflammatory factors, renal functions and electrolyte balance, and behavior data including sleep modes, activity amounts and cognitive test results, are collected in real time from physical sign monitoring equipment, an electronic medical record system and nursing records of a patient; a time sequence analysis technology is adopted, change modes of a patient on multiple physiological and behavioral parameters are mined, and delirium early symptoms are identified; a long-term short-term memory (LSTM) network deep learning algorithm is utilized to analyze parameter fluctuation trends of the patient in different time periods, and time sequence features of delirium precursor are extracted; and based on the time sequence analysis result of the delirium features, constructing a multi-factor fusion risk assessment model.

Description

Technical Field

[0001] The invention relates to a delirium risk assessment method. Background Art

[0002] The current methods for delirium risk assessment, such as a delirium risk assessment method and assessment system based on hospital information systems proposed in China's invention patent authorization CN2020104176573, have improved the ability to assess the delirium risk of special groups such as ICU patients to a certain extent, but their methods and systems still have many shortcomings and disadvantages. First, the existing method mainly relies on historical data and evaluation information in the hospital information system for secondary evaluation, which means that it is not flexible enough to respond to the dynamic changes in the patient's current health status. This evaluation method based on static data and historical records will miss the patient's current real-time changing physiological state and behavioral characteristics, especially in highly dynamic clinical environments such as ICU wards, where the patient's health status fluctuates dramatically in a short period of time. The dependence of the existing method makes its evaluation results more based on past medical record data rather than the patient's current physiological indicator changes, which obviously cannot meet the rapidly changing health needs of critically ill patients. Secondly, the "combined evaluation value" strategy adopted by the method, that is, combining certain characteristic values ​​of the patient for conditional judgment, is too rigid and dependent on pre-set evaluation conditions. Although this method can perform relatively rapid risk screening in specific situations, its core assessment mechanism is based on fixed logical rules and preset thresholds. This strategy ignores the multidimensionality and complexity of delirium, because there are many factors that induce delirium, and it is often manifested as a combination of different physiological and behavioral characteristics. The current method lacks in-depth mining and analysis of time series data, and only makes simple conditional judgments by combining several characteristic values, which can easily lead to underestimation or overestimation of delirium risk. In addition, once the preset assessment conditions are set, it is difficult to dynamically adjust them according to individual differences and changes in the patient's condition. The "static" nature of this assessment is difficult to adapt to the individualized treatment needs in the ICU ward.

[0003] Furthermore, the existing methods mainly focus on the evaluation of patients' language expression ability, physical activity ability and facial expressions, and these features are often limited by the patient's physical condition in the ICU ward. For example, the use of intubation and sedatives will significantly reduce the accuracy of the evaluation. Therefore, relying solely on these surface features will lead to misjudgment of delirium risk. For example, intubated patients cannot express language, and the use of sedatives leads to dull facial expressions, which in turn affects the evaluation results of the existing system. This method fails to fully consider the interference factors that patients are exposed to in the ICU environment, and such evaluation criteria have great limitations in practical applications. In addition, when the system detects that "the predetermined conditions are not met", historical information will be retrieved for secondary evaluation, but this process depends on the data integrity and update frequency of the hospital information system. If the patient's historical data is incomplete or not updated in time, it will further affect the accuracy of the evaluation. In addition, although the "dynamic delirium risk assessment strategy" is introduced in the technical solution of the comparative document, its dynamic nature is mainly reflected in the secondary evaluation combined with historical data, rather than dynamic adjustment based on the patient's current real-time data. In high-risk environments such as the ICU, the assessment of delirium risk requires higher real-time and sensitivity, but existing methods lag behind in this regard and lack the ability to capture and analyze the patient's current health status in real time. Time is of the essence for early intervention and prevention of delirium. If the assessment system cannot respond to changes in the patient's physiological signals in a timely manner, the best time for intervention will be missed.

[0004] In contrast, modern deep learning and time series analysis methods, such as technology based on LSTM (Long Short-Term Memory Network) models, can better capture the dynamic changes of patients' physiological signals, thereby providing more accurate real-time delirium risk assessment. Unlike the static data and rule-based judgment methods used in comparative documents, the LSTM model can extract complex dynamic features from patients' multidimensional time series data, and further integrate multiple inducing factors such as acute diseases, drug dependence, and long-term bed rest through integrated learning algorithms (such as XGBoost) to generate individualized delirium risk scores. This assessment method based on deep learning and data fusion not only improves the accuracy of delirium risk prediction, but also makes real-time adjustments based on the patient's latest data to ensure the timeliness of the assessment results. Summary of the invention

[0005] The purpose of the present invention is to provide a delirium risk assessment method to solve some of the drawbacks and shortcomings pointed out in the background technology.

[0006] The present invention solves the above-mentioned technical problems by adopting the following technical solution, which includes the following steps:

[0007] S1. Multidimensional data collection and dynamic feature capture of delirium-inducing factors:

[0008] S1.1. Design an integrated multidimensional data collection system for multiple delirium-causing factors including infection, drug use, metabolic disorders, and sleep disorders;

[0009] S1.2. Collect data in real time from patients’ vital sign monitoring devices, electronic medical record systems, and nursing records, including physiological parameters such as heart rate variability, blood oxygen levels, and temperature fluctuations, biochemical indicators including inflammatory factors, renal function, and electrolyte balance, and behavioral data including sleep patterns, activity levels, and cognitive test results;

[0010] S2. High-dimensional time series analysis of delirium characteristics:

[0011] S2.1. Use time series analysis techniques to mine the changing patterns of multiple physiological and behavioral parameters of patients and identify early signs of delirium;

[0012] S2.2. Use the long short-term memory network (LSTM) deep learning algorithm to analyze the parameter fluctuation trends of patients in different time periods and extract the time series characteristics of delirium precursors;

[0013] S3. Multi-factor integration for delirium risk assessment:

[0014] Based on the time series analysis results of delirium characteristics, a multi-factor fusion risk assessment model was constructed. An ensemble learning-based algorithm was used to fuse high-risk inducing factors including acute illness, long-term bed rest, and drug dependence with dynamic feature data to generate an individualized delirium risk score for patients.

[0015] S4. Real-time delirium risk dynamic warning and intervention strategy generation;

[0016] Through the risk assessment model, the patient's real-time data is continuously monitored, and a warning signal is generated when a high-risk state is detected. At the same time, based on the patient's individual characteristics and risk scores, including age, underlying diseases, and medication usage, targeted intervention strategy recommendations are generated, including adjusting medications, improving environmental stimulation, and optimizing nursing processes. The generation of intervention strategies uses decision trees and reinforcement learning algorithms.

[0017] Furthermore, the high-dimensional time series analysis method of delirium characteristics:

[0018] Collect multidimensional health data of patients from various data sources, including physiological parameters, behavioral data, cognitive function assessment results and drug use, and clean and pre-process the data; use a cleaning formula that combines derivative adjustment and exponential decay; detect abnormal data points and smooth them. The formula is as follows:

[0019]

[0020] Where: x represents the original input data point, including the patient's heart rate measurement; α adjusts the weight parameter to control the strength of outlier removal; represents the derivative operation on the data x, outliers(x) is an outlier detection function used to identify outliers in the data; Exponential decay factor, where λ controls the decay speed.

[0021] Furthermore, the high-dimensional time series analysis method of delirium characteristics:

[0022] The preprocessed multidimensional data is serialized and divided into multiple time windows of fixed length, including 1 hour and 3 hours; each time window contains multiple data points to describe the health status of the patient in that time period; a time smoothing method based on a modulation function is used to capture the periodic fluctuations in the patient's physiological signals; by introducing modulation terms, the periodic changes including heart rate and respiratory cycle are modeled, and the formula is as follows:

[0023]

[0024] Where V(t) is the time series input data, which represents the patient's physiological and behavioral data points collected at time t; the amplitude parameter of the γ modulation term is used to amplify or reduce the impact of periodic fluctuations; cos(ωt) is the periodic modulation term, where ω is the frequency parameter, which is used to capture the periodic fluctuations of the circadian rhythm in the data; e -βt The time decay factor controls the impact of old data on the current evaluation, where β determines the decay rate; the length of the T time window is used to define the integration interval.

[0025] Furthermore, the high-dimensional time series analysis method of delirium characteristics:

[0026] An LSTM neural network was introduced to analyze the patient's time series data and predict the risk of delirium. The risk prediction process of the LSTM network includes multiple layers: input layer, hidden layer, and output layer.

[0027] First, the smoothed time series data is input into the LSTM network; the model extracts the key features in the time series by recursively updating the hidden state; the risk score integration method is used to accumulate the risk score within the time window; the expression formula is used:

[0028]

[0029] Where R(t) represents the output risk score of the LSTM model at time t, indicating the patient's delirium risk at that time point; S(R) represents the Softmax function, which is used to convert the risk score R(t) into the probability of classification including "high risk", "medium risk" or "low risk", and controls the output risk value between 0 and 1; Δt time integration interval is used to define the cumulative risk assessment within time; f risk (R,Δt) The cumulative risk score is used to comprehensively analyze the patient's delirium risk in the future unit time; when the cumulative score exceeds the threshold θ, an early warning signal will be triggered; when the cumulative risk score exceeds the set threshold θ, an early warning will be generated and personalized intervention recommendations will be provided.

[0030] Furthermore, the multi-factor fusion method for delirium risk assessment includes:

[0031] First, data preprocessing is performed, including data cleaning, outlier removal, and time series smoothing. During the data cleaning process, an outlier cleaning method based on derivative adjustment and exponential decay is used to remove abnormal data points while retaining key time series features. The sensitivity of the calculated derivative to data fluctuations is adjusted. The cleaning formula is as follows:

[0032]

[0033] Where: x represents the original input data points, including the patient's heart rate measurement; α represents the adjustment parameter used to control the intensity of outlier cleaning; Represents the derivative of the input data, which is used to detect the rate of data change; detect(x) represents the outlier detection function, which identifies outliers in the data; represents an exponential decay factor, where λ controls the sensitivity to outliers, and a larger λ reduces the impact of outliers more quickly.

[0034] Furthermore, the multi-factor fusion method for delirium risk assessment includes:

[0035] Perform in-depth analysis on the preprocessed time series data; LSTM captures the time dependency of the data and extracts short-term and long-term health change trends related to delirium; after the input is processed by LSTM, the output hidden layer state H(t) is used to describe the patient's health status at time t;

[0036] A time series smoothing function is used; by introducing modulation terms and time decay factors, the periodic fluctuations of the patient's physiological data, including changes in heart rate and blood oxygen level, are captured:

[0037]

[0038] in:

[0039] H(τ) represents the hidden layer state of TM output, which represents the health characteristics of the patient at time τ; γ represents the modulation parameter, which is used to adjust the intensity of the periodic signal; sin(ωτ) represents the periodic modulation term, where ω is the frequency parameter, which is used to capture the periodic changes in the data, including the circadian rhythm; e -βτ Represents the time decay factor, where β controls the impact of old data on the current analysis, and β accelerates the decay of old data; t represents the length of the time window, which is used to define the integral range of the smoothing process.

[0040] Furthermore, the multi-factor fusion method for delirium risk assessment includes:

[0041] The XGBoost algorithm based on ensemble learning is used to integrate the state features including age, gender, and history of underlying diseases, as well as dynamic time series features. seq (H(t)) and inducing factor data were fused into a unified prediction model; an individualized delirium risk score was generated using the following multifactor fusion formula:

[0042]

[0043] Among them: F i represents the input features of the same triggering factors including acute illness, drug dependence, and activity level; w i is the weight, which is automatically learned and adjusted by the XGBoost model; φ(H(t)) represents the risk adjustment function of the time series feature H(t), which is used to enhance the sensitivity to dynamic features; Indicates that column features are integrated to accumulate the impact of historical data on current risks; T is the time range used to define the weight of historical data in risk assessment.

[0044] The delirium risk assessment method of the present invention has the following significant beneficial effects:

[0045] First of all, the present invention can accurately capture the changes in the health status of patients in intensive care environments such as the ICU by means of real-time multi-source data collection and dynamic time series analysis. Unlike traditional evaluation methods that rely on static data or historical medical records, the present invention can analyze and process the dynamic changes in the patient's physiology, behavior, and drug use in real time, and provide a more accurate and personalized delirium risk assessment. In particular, in the analysis of dynamic physiological data such as heart rate fluctuations, blood oxygen changes, and sleep interruptions, the precursors of delirium can be effectively identified, risks can be warned in advance, and timely intervention basis can be provided.

[0046] Secondly, the present invention adopts a LSTM (Long Short-Term Memory) neural network model based on deep learning, which can fully capture the long-term dependencies in time series data. For complex physiological and behavioral data, LSTM can retain historical information through its memory ability, thereby more accurately assessing the patient's dynamic health status. Compared with traditional rule-based methods, the LSTM model has obvious advantages in capturing the nonlinear relationships and time series characteristics of data, and can handle high-dimensional, multivariate data inputs, providing more accurate delirium risk prediction.

[0047] In addition, the present invention combines ensemble learning algorithms (such as XGBoost) to integrate static features, dynamic features and multiple predisposing factors (such as acute diseases, drug dependence, long-term bed rest, etc.) into the same model, making full use of the complementarity of various data sources to improve the accuracy and stability of the assessment. The ensemble learning method can effectively reduce the overfitting or bias problems of a single algorithm, and can provide more robust and reliable delirium risk assessment results in practical application scenarios with high data complexity. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 The present invention is a flow chart of a delirium risk assessment method.

[0049] Figure 2 The figure is a flow chart of the high-dimensional time series analysis method for delirium characteristics of the present invention.

[0050] Figure 3 The present invention is a flowchart of the multi-factor fusion method for delirium risk assessment. DETAILED DESCRIPTION

[0051] Combined with Figure 1 The delirium risk assessment method provided by the present invention includes multiple key steps, among which the S1 step is an important component, involving multi-dimensional data collection and dynamic feature capture of delirium inducing factors.

[0052] First, targeting multiple triggering factors of delirium, such as infection, drug use, metabolic disorders, and sleep disorders, this method designs an integrated multidimensional data acquisition system. The design goal of this system is to comprehensively capture various physiological and behavioral characteristic data related to the occurrence of delirium, and to monitor and analyze these data in real time, so as to timely identify the precursors of delirium in patients. In the specific implementation, the system integrates and collects information through multiple data sources such as the patient's vital sign monitoring equipment, electronic medical record system, and nursing records. Through the vital sign monitoring equipment, the patient's physiological parameters, such as heart rate variability, blood oxygen saturation, and body temperature fluctuations, can be obtained in real time. These physiological indicators are important bases for assessing the risk of delirium. Heart rate variability reflects the functional state of the autonomic nervous system and is usually closely related to the occurrence of delirium, while blood oxygen level and body temperature fluctuations can also reflect changes in the patient's physiological state. Abnormal fluctuations in these parameters indicate that the patient is at a higher risk of delirium. In addition to physiological parameters, this method also pays special attention to the collection of biochemical index data related to physiological changes, such as inflammatory factors, renal function, electrolyte balance, etc. The abnormal fluctuations of these indicators can help identify the risk of delirium caused by medical diseases or metabolic disorders. For example, renal failure or electrolyte imbalance is often closely related to the occurrence of delirium, so the real-time collection and analysis of these biochemical data is of great significance. At the same time, the system will also collect the patient's behavioral data, including sleep patterns, activity levels, and cognitive test results. These behavioral data can reflect changes in the patient's mental state and cognitive function. Sleep disorders are common causes of delirium, and activity levels and cognitive test results can reveal whether the patient has symptoms such as reduced physical activity or decreased cognitive function, which are important early warning signs of delirium. Therefore, by integrating various types of physiological, behavioral and biochemical data, and monitoring the dynamic changes of these data in real time, the present invention can provide strong support for the early identification and early warning of delirium risk. Through the collection and analysis of such multidimensional data, this method can effectively identify and capture the dynamic characteristics of the patient's physiological state, behavioral patterns, potential disease changes, etc., and provide accurate and comprehensive data support for subsequent risk assessment models, thereby improving the accuracy and timeliness of delirium risk assessment.

[0053] The core of step S2 is to use time series analysis technology to deeply explore the changing patterns of patients in multiple physiological and behavioral parameters and identify early signs of delirium. First of all, time series analysis technology is an effective tool that can help capture time dependence and trend changes in data. In the risk assessment of delirium, time series analysis can reveal the fluctuations of patients' physiological and behavioral data in different time periods, especially the abnormal changes in the short term. These abnormal fluctuations are often early signs of delirium, such as drastic fluctuations in physiological and behavioral parameters such as heart rate, blood oxygen saturation, and activity, which can reflect changes in the patient's health status. Therefore, by conducting a systematic time series analysis of the changes in these parameters, potential delirium risks can be identified earlier and intervention measures can be taken in advance.

[0054] In order to more accurately capture the changing patterns of these time dependencies, the present invention uses a deep learning algorithm called a long short-term memory network (LSTM). LSTM is a variant based on a recurrent neural network (RNN) and is specifically designed to process and predict long-term dependencies in time series data. Compared with traditional time series analysis methods, the LSTM algorithm has a stronger memory capacity and can effectively retain important historical information when processing long sequence data. This is particularly important for delirium risk assessment, because the occurrence of delirium is usually the long-term cumulative effect of multiple factors, which accumulate slowly over a period of time and show some minor changes. It is difficult to capture these long-term changes through traditional short-term data analysis. The LSTM algorithm can automatically adjust the memory weight of historical data through its "memory unit", thereby identifying the trend of parameter fluctuations in different time periods, revealing the long-term change trend of the patient's health status, and extracting precursor features related to delirium.

[0055] The input data of LSTM is preprocessed multidimensional time series data, including the patient's physiological parameters (such as heart rate variability, blood oxygen saturation, body temperature, etc.), behavioral data (such as activity level, sleep quality, etc.), and other factors that affect delirium (such as drug use records, etc.). The LSTM network extracts valuable time series features from these input data through multiple levels of nonlinear transformations and activation functions, and learns its fluctuation trends in different time periods. These fluctuation trends manifest themselves in the early stages as irregular fluctuations in the patient's heart rate, a slight drop in blood oxygen levels, or sudden changes in activity. Small abnormal changes in these data are signals of increased risk of delirium.

[0056] In this way, the present invention can use the LSTM algorithm to conduct in-depth analysis of the patient's health data and extract dynamic features that are closely related to the risk of delirium. For example, LSTM can analyze the patient's heart rate variability (HRV), which is generally considered to be an important indicator of the function of the autonomic nervous system, and its abnormal fluctuations indicate the occurrence of delirium. Similarly, LSTM can also analyze the patient's blood oxygen saturation changes. A decrease in blood oxygen saturation is often related to the occurrence of delirium, and the LSTM algorithm can capture this change trend in real time and provide early warning. Through in-depth analysis of these physiological and behavioral data, LSTM can effectively dig out the hidden change patterns in the patient's health status and extract potential precursors to delirium, thereby providing a scientific basis for subsequent risk assessment and intervention.

[0057] Step S4 first continuously monitors the patient's real-time data through a risk assessment model to ensure that the latest changes in the patient's health status can be obtained at any time. When the assessment model detects that the patient has entered a high-risk state, it can automatically generate an early warning signal to remind medical staff to pay attention to the patient's changes in a timely manner. This process relies on a variety of real-time data inputs, including the patient's physiological indicators, behavioral data, and related disease history. Through continuous monitoring of these data, the assessment model can sensitively capture abnormal fluctuations in the patient's heart rate, blood oxygen level, activity level, etc., so as to accurately identify the potential risks of delirium and provide warnings to medical staff in advance. Unlike traditional static risk assessment methods, the dynamic monitoring of the present invention can continuously update the risk assessment results based on real-time data changes, ensure the timeliness and accuracy of the assessment results, and avoid missing the best time for intervention.

[0058] After detecting a high-risk state and generating an early warning signal, the present invention further generates targeted intervention strategy recommendations based on the individual characteristics of the patient (such as age, underlying diseases, drug use, etc.) and the risk score obtained previously. These intervention strategies not only take into account the patient's current health status, but also combine their individual characteristics to ensure the personalization and precision of the intervention measures. The core of individualized intervention is to formulate appropriate intervention measures according to the patient's specific circumstances, such as adjusting medications, improving environmental stimulation, optimizing nursing processes, etc. For example, if the patient's medication history has a record of sedatives or opioids, targeted interventions include adjusting the dosage of these drugs or replacing other types of drugs to reduce the risk of delirium. For long-term bedridden patients, intervention strategies include improving environmental stimulation around the bed, increasing light or sound stimulation to maintain the patient's circadian rhythm and avoid cognitive dysfunction caused by environmental factors. Optimization of nursing processes includes regular cognitive function assessments, adjusting nursing schedules, or strengthening interactions with patients to ensure that patients receive adequate attention and care in hospital environments.

[0059] In order to ensure the effectiveness and flexibility of the intervention strategy, the present invention uses a decision tree and a reinforcement learning algorithm to generate an intervention strategy. A decision tree is a model that is based on patient historical data and risk scores for reasoning, and can quickly generate appropriate intervention measures based on different input conditions (such as the patient's medication use, disease history, etc.). By training a large amount of patient data, the decision tree can automatically identify the key factors that affect the occurrence of delirium and generate the best intervention recommendations for each situation. The reinforcement learning algorithm is an algorithm that learns the optimal decision-making strategy through interaction with the environment. In the present invention, reinforcement learning can continuously adjust the intervention strategy according to the patient's feedback information by simulating the effects of different intervention measures, so that the intervention measures can be optimized and adjusted as the patient's condition changes. The combination of these two algorithms makes the intervention strategy generation process more intelligent, personalized and adaptive, and can provide the most appropriate intervention plan for different patients and different situations.

[0060] Embodiment 1:

[0061] Combined with Figure 2 In order to better illustrate the application of the high-dimensional time series analysis method of delirium characteristics in the present invention, and to combine formula calculation to show how to use data cleaning and outlier detection technology to improve the accuracy of delirium risk assessment. The following example uses patient data as an example for detailed description.

[0062] Example background: Mr. Zhang, 68 years old, was hospitalized due to a sudden cerebral infarction and transferred to the intensive care unit (ICU). He has a history of hypertension and diabetes, and is receiving multiple medications, including antihypertensive drugs, blood sugar control drugs, and opioid analgesics. Due to long-term bed rest and the use of multiple drugs, the doctor suspected that he was at high risk of delirium, so it was decided to use the delirium risk assessment method of the present invention to dynamically monitor and evaluate his health status.

[0063] First, through the hospital information system and vital sign monitoring equipment, Mr. Zhang's multi-dimensional health data is continuously collected, including physiological parameters (such as heart rate, blood oxygen saturation, body temperature), behavioral data (such as sleep patterns, activity level) and cognitive function assessment results. The data points collected every day include: heart rate measurements (recorded every 5 minutes), blood oxygen saturation, body temperature, and hourly activity monitoring data. At the same time, drug use records, such as the dosage and frequency of opioid analgesics, are also obtained from the electronic medical record. After data collection, the system will pre-process the data, including data cleaning and outlier detection, to ensure the accuracy of the data.

[0064] During the data cleaning process, the system uses the following cleaning formula that combines derivative adjustment and exponential decay to detect and remove abnormal data points:

[0065]

[0066] explain:

[0067] x represents Mr. Zhang’s original input data point, such as heart rate measurement.

[0068] α is an adjustment weight parameter used to control the strength of outlier removal, and its value usually ranges from 0.1 to 0.5. In this example, α is set to 0.3.

[0069] Represents the derivative operation on the data x, which is used to detect the sensitivity of data fluctuations.

[0070] outliers(x) is an outlier detection function used to identify abnormal points in the heart rate data.

[0071] is an exponential decay factor, where λ controls the data decay speed and is usually in the range of 0.01 to 0.1. In this example, λ is set to 0.05.

[0072] Example data description and calculation:

[0073] The sequence of Mr. Zhang's heart rate measurements on a certain day is: [78, 82, 180, 79, 85], of which 180 is an obvious outlier. First, the system will use the outlier detection function outliers(x) to detect that 180 is an abnormal data point. Next, the system calculates the derivative adjustment term:

[0074]

[0075] Then, the smoothed heart rate value is calculated according to the cleaning formula:

[0076] f clean (180) = 180 - 0.3 × 0.0027 × 180 ≈ 179.85

[0077] The results show that although 180 is identified as an outlier, it is only adjusted to 179.85 after smoothing through the formula instead of being simply eliminated, thus retaining the integrity of the data while eliminating the interference of the anomaly on subsequent analysis.

[0078] After cleaning the data, the system inputs the processed data into the LSTM model for time series analysis. The LSTM network can identify the fluctuation trends of Mr. Zhang's heart rate, blood oxygen level, and body temperature, and extract potential precursor features of delirium from them. For example, the LSTM network will detect that Mr. Zhang's heart rate has continued to rise in the past 24 hours, while sleep time has decreased and blood oxygen level fluctuations have increased, which are early signs of delirium. After setting the LSTM model to analyze Mr. Zhang's data for the past three days, it was found that the heart rate variability (HRV) continued to decline, from 60ms fluctuations to 35ms, while blood oxygen saturation dropped from 98% to 94%, and nighttime activity increased. LSTM will summarize these time series features to calculate Mr. Zhang's delirium risk score.

[0079] Based on the LSTM time series analysis results and risk scores, the system of the present invention generates an individualized intervention strategy. In this example, the system detects that Mr. Zhang’s delirium risk score R delirium The HRV score of Mr. Zhang has improved significantly since the previous day, reaching 0.75, while the set warning threshold is 0.7. Therefore, the system automatically generates a warning signal and provides specific intervention suggestions to medical staff, such as reducing the dosage of opioids, increasing daytime activities, and improving the sleep environment at night. The reinforcement learning algorithm will further optimize the intervention strategy based on Mr. Zhang's real-time feedback. For example, if Mr. Zhang's HRV improves after reducing opioids, the system will continue to adjust the drug dosage until the risk score drops to a safe range.

[0080] Next, each step will be elaborated in detail, and combined with actual data and calculation results, the specific implementation and feasibility of the time series analysis solution in the present invention will be demonstrated.

[0081] During his hospitalization, Mr. Zhang collected a large amount of data about his health status through the hospital's information system, vital sign monitoring equipment and electronic medical records. These data not only include common physiological parameters such as heart rate, respiratory rate, blood oxygen saturation, etc., but also behavioral data such as sleep quality, activity level and cognitive test results, and even information such as drug use. Through the collection of these data, the system will perform data preprocessing, cleaning and time series, and ultimately use them for subsequent high-dimensional time series analysis.

[0082] During the data preprocessing process, Mr. Zhang's data was detected and cleaned for outliers to ensure the accuracy of the input data. Afterwards, the system divided these multidimensional data into multiple time windows, which were 1 hour and 3 hours respectively, aiming to capture the changing trends on different time scales. The data in each time window included multiple data points such as heart rate, respiratory rate, blood oxygen saturation, and activity level, reflecting Mr. Zhang's health status during that time period.

[0083] For example, the heart rate data of Mr. Zhang for a certain period of time (from 12:00 to 15:00) is set as: [78, 80, 82, 81, 75, 79, 83, 80, 78, 76]]. This period of data is divided into a 1-hour time window and used to analyze his physiological fluctuations within this hour.

[0084] Next, in order to further capture the potential periodic fluctuations in Mr. Zhang's health data, a time smoothing method based on modulation function was used. This method models the patient's physiological signals and analyzes their periodic fluctuation characteristics, especially the changes in heart rate and respiratory rate. This periodic fluctuation is closely related to factors such as circadian rhythm, medication use, and pain level, and is therefore crucial for early warning of delirium.

[0085] In the algorithm of the present invention, a modulation term is introduced to enhance or weaken the influence of periodic changes, and the following formula is specifically used:

[0086]

[0087] Formula explanation:

[0088] V(t) represents the time series input data, which here represents Mr. Zhang’s physiological data (such as heart rate, breathing rate, etc.) at time t.

[0089] γ is the amplitude parameter of the modulation term, which is used to control the intensity of the periodic fluctuation and can be adjusted according to the actual situation of the data. In this example, γ is set to 0.5, which means that the influence of the periodic fluctuation will be amplified or reduced with moderate intensity.

[0090] cos(ωt) is a periodic modulation term, where ω represents the periodic frequency in the data, which is usually related to the circadian rhythm cycle, respiratory cycle, etc. In this example, ω is set to 2π / 24 hours (i.e., 24-hour periodic fluctuation).

[0091] e -βt It is a time decay factor, which is used to reduce the influence of old data on the current evaluation. Setting β = 0.05 means that the data gradually decays in time and the influence will decrease over time.

[0092] T is the length of the time window. In this example, it is set to 1 hour, so the integration interval is from 0 to 1 hour.

[0093] Through this formula, the system will smooth Mr. Zhang's physiological data in each time window, so as to accurately capture the periodic fluctuations of his health status. For example, set Mr. Zhang's heart rate data to 78, 80, 82, 81, 75, 79, 83, 80, 78, 76, substitute these data into the formula for smoothing, and calculate the weighted periodic fluctuation trend. Through the influence of the modulation function and the time attenuation factor, the system will provide Mr. Zhang with a smoother heart rate change curve, which will help identify potential precursors to delirium.

[0094] Through the above method, the system periodically modulates and attenuates the data in each time window to obtain a smoother and more reliable health data sequence. Next, the system inputs these processed data into the deep learning model for further analysis. Based on these smoothed data, the LSTM network can better identify the changing trend of Mr. Zhang's health status and capture any potential precursors to delirium in a timely manner. For example, when the LSTM model is set to analyze Mr. Zhang's data over the past few hours, it is found that after an obvious periodic fluctuation, the fluctuation amplitude of the heart rate increases, and the sleep quality drops sharply. These changes are early warning signs of delirium. The LSTM analysis is set to conclude that Mr. Zhang's delirium risk score in the past 3 hours is 0.85, which exceeds the set threshold of 0.7. The system will immediately issue a risk warning, prompting medical staff to take necessary intervention measures, such as adjusting medications and optimizing nursing processes.

[0095] By introducing LSTM neural network to analyze and predict delirium risk, this paper will elaborate on how to use LSTM model for efficient time series analysis and explain the specific feasibility of each step. First, by inputting the preprocessed time series data into the LSTM network, the system can extract key features by recursively updating the hidden state, thereby capturing the dynamic changes of the patient's physiology and behavior. This process will be explained in detail by substituting actual data into specific calculation formulas.

[0096] In Mr. Zhang’s case, his heart rate, blood oxygen saturation, body temperature, and activity data for the past 24 hours were collected through the hospital information system. Set in a certain time period (12:00 to 15:00), Mr. Zhang’s heart rate data is: [78, 80, 82, 81, 75, 79, 83, 80, 78, 76]. These data are first smoothed and divided into multiple time windows. Each time window includes 1 hour of data, and these data are input into the LSTM network for the next step of risk assessment.

[0097] In an LSTM network, the input layer first receives this data and then passes it to the hidden layer through a recursive process, which gradually updates its state to capture patterns in the time series. The recursive nature of LSTM enables it to remember previous states and adjust its predictions based on new inputs.

[0098] The LSTM network is trained to generate a risk score R(t) at each time point, which represents the patient's delirium risk at time t. In order to transform these risk scores into more specific predictions, the Softmax function S(R) is introduced to classify the risk, mapping the risk score R(t) to three categories: "high risk", "medium risk" or "low risk".

[0099] The expression of the Softmax function is as follows:

[0100]

[0101] Among them, R(t) is the risk score at time point t, R(t) low ,R(t) mid The Softmax function converts the risk score into a probability value between 0 and 1 as a reference value for the “low risk” and “medium risk” classifications. For example, if Mr. Zhang’s risk score at 12:30 is 0.8, the Softmax function will map it to a high risk probability of 0.85, meaning that his risk of delirium at this point is 85%.

[0102] Next, the system integrates the risk scores in each time window to calculate the cumulative risk score f for the future period of time. risk (R,Δt) in order to more comprehensively assess the patient's risk changes. The integral formula is as follows:

[0103]

[0104] Where R(t) is the risk score at time point t, S(R(t)) is the probability value after Softmax conversion, and Δt is the time integration interval, which indicates the cumulative risk calculated within this time period.

[0105] For example, if Mr. Zhang is between 12:00 and 15:00, the risk score output by the LSTM model is: [0.6, 0.7, 0.75, 0.85, 0.9, 0.95, 0.9, 0.85, 0.8, 0.75], the cumulative risk score in this time period can be calculated. If Δt = 1 hour, then between 12:00 and 13:00, the integrated risk score is:

[0106]

[0107] Assume that R(t) changes continuously within the time period, and S(R(t)) also changes accordingly, and finally the cumulative risk score within this time period can be obtained.

[0108] If the cumulative risk score f calculated by integration risk If (R, Δt) exceeds the set threshold θ (the threshold is set to 0.7), the system will trigger a warning signal for delirium risk and generate a personalized intervention strategy based on this. For example, if the cumulative risk score from 12:00 to 15:00 is calculated to be 0.75, which exceeds the set threshold of 0.7, the system will issue a "high risk" warning signal and provide intervention suggestions, including adjusting medications, improving environmental stimulation, and increasing nursing intervention.

[0109] In Mr. Zhang’s case, because his cumulative risk score during that time period exceeded the threshold, the system recommended adjusting the dose of sedatives he was taking, optimizing his sleeping environment, reducing noise disturbances, and strengthening nighttime monitoring by nursing staff. All of these interventions are based on personalized risk assessment results to ensure that Mr. Zhang can receive timely and effective treatment and care.

[0110] Embodiment 2:

[0111] See also Figure 3 , Mr. Zhang was at high risk of delirium during his treatment in the intensive care unit (ICU). When providing a personalized risk assessment for Mr. Zhang, the system first applied a multi-factor fusion method and went through a series of data preprocessing steps to ensure the accuracy of the final analysis results. The following is a detailed analysis of Mr. Zhang's case, which includes key links such as data preprocessing, outlier cleaning, time series smoothing, and risk assessment.

[0112] First, the system collected a variety of physiological data from Mr. Zhang’s medical records, including heart rate, blood oxygen saturation, body temperature, activity level, etc. These data come from the patient’s vital sign monitoring equipment, electronic medical record system, and nursing records. For example, the raw data of Mr. Zhang’s heart rate over a period of time is as follows: [72, 75, 73, 79, 100, 82, 78, 74, 80, 105, 70]. Obviously, the “100” and “105” in the data are outliers, which are caused by problems with the monitoring equipment or measurement errors.

[0113] In order to clean these data, an outlier cleaning method based on derivative adjustment and exponential decay is used. First, the system calculates the derivative of the data fluctuation and uses the derivative to detect the rate of change of the data. The formula is as follows:

[0114]

[0115] Here, x represents the original data point (such as heart rate measurement), and α is a tuning parameter used to control the strength of outlier cleaning; It means to calculate the derivative of the input data x to help identify the rate of change of the data; detect(x) is an outlier detection function used to identify abnormal points in the data; is an exponential decay factor, where λ is a parameter that controls the sensitivity of the decay.

[0116] By setting α = 0.5 and λ = 0.05, the derivative of each data point can be calculated, and the abnormal values ​​in the heart rate data can be identified and smoothed. For example, for 100 and 105 in the heart rate data, their weights are reduced by this method, and these abnormal points are smoothed to values ​​close to the adjacent normal data, thereby ensuring the stability and accuracy of the data.

[0117] Through this data cleaning process, Mr. Zhang's heart rate data was corrected to: [72, 75, 73, 79, 78, 82, 78, 74, 80, 80, 70]. These values ​​are more in line with the normal physiological fluctuation range and eliminate abnormal fluctuations in the measurement.

[0118] After cleaning the data, the system performs time series smoothing. Through smoothing, short-term high-frequency fluctuations are eliminated, making long-term trends more obvious. In particular, a modulation function-based time smoothing method is used to capture the periodic fluctuations in Mr. Zhang's physiological signals. For example, Mr. Zhang's heart rate and respiratory cycle show circadian fluctuations, and this periodic fluctuation is crucial to the occurrence of delirium.

[0119] In order to model these periodic fluctuations, the system introduces a modulation term to capture the frequency changes of the fluctuations. Assuming that the system detects that the fluctuation frequency of Mr. Zhang's heart rate in a day is ω = 0.2 (that is, the heart rate will experience about 5 fluctuations per day), the system performs smoothing using the following formula:

[0120]

[0121] Here, V(t) represents time series data (for example, Mr. Zhang’s heart rate data), γ is the amplitude of the modulation term, which is used to amplify or reduce the impact of periodic fluctuations; ω is the frequency of the fluctuation, which is used to represent periodic changes; e -βt is the time decay factor, where β = 0.01 represents the decay speed of the data and controls the impact of old data on the current evaluation.

[0122] Through this smoothing process, Mr. Zhang's heart rate fluctuations are effectively smoothed in time, making the pattern of periodic fluctuations clearer. For example, the smoothed heart rate data becomes: [74,76,78,79,81,80,79,77,76,75,74], which reflects the normal fluctuations of heart rate and is not affected by short-term abnormal fluctuations.

[0123] After completing the data preprocessing, the system uses a multi-factor fusion method to combine Mr. Zhang's physiological parameters, behavioral data, drug usage and other multi-dimensional factors to conduct a comprehensive risk assessment. Specifically, the system calculates the risk weight of each factor and fuses the risk scores of different data sources. After time series analysis of Mr. Zhang's heart rate fluctuations, blood oxygen levels, body temperature and activity, the following risk scores were obtained: heart rate score of 0.75, blood oxygen score of 0.85, body temperature score of 0.7, and activity score of 0.65. Using the weighted average method, the system fuses these scores into an overall risk score:

[0124]

[0125] Among them, w1=0.3, w2=0.3, w3=0.2, w4=0.2 are the weights of each factor. and R Activity These are Mr. Zhang’s risk scores in these aspects. By taking these factors into consideration, the total risk score is obtained:

[0126] R total =0.3 0.75+0.3 0.85+0.2 0.7+0.2 0.65=0.76

[0127] This total risk score indicates that Mr. Zhang is at a medium risk level at the current stage. The system will continue to monitor his changes in subsequent time periods and issue a warning signal when the score exceeds the set threshold.

[0128] When Mr. Zhang's risk score R total When a certain threshold is reached (the threshold is set to 0.8), the system will trigger an early warning and generate a personalized intervention strategy. For example, when Mr. Zhang's risk score is 0.76, although it has not reached the early warning threshold, he is still within a certain risk range, so the system recommends adjusting his medication to target sedatives that induce delirium. At the same time, the system recommends strengthening monitoring of Mr. Zhang's sleep quality and activity level to prevent symptoms from worsening.

[0129] Through this series of data preprocessing, outlier cleaning, time series smoothing and multi-factor fusion methods, the system can accurately assess Mr. Zhang's delirium risk and provide effective intervention strategies with the support of actual data.

[0130] During Mr. Zhang's delirium risk assessment process, the system used a multi-factor fusion method, especially in deep analysis and time series smoothing processing. It captured the time dependency of the data through the LSTM network and further extracted short-term and long-term health change trends related to the occurrence of delirium.

[0131] First, the system performs time series preprocessing on Mr. Zhang’s physiological data, including data points such as heart rate, blood oxygen level, and body temperature. Mr. Zhang’s heart rate data is as follows: [72, 75, 73, 79, 78, 82, 80, 77, 79, 78, 81], and blood oxygen data is: [96, 97, 98, 96, 95, 94, 93, 95, 96, 97, 98]. These data represent Mr. Zhang’s heart rate and blood oxygen level measurement results for 11 consecutive hours.

[0132] Then, the preprocessed data was input into the LSTM model. LSTM (Long Short-Term Memory Network) can capture the time dependency in the data, especially in capturing the short-term and long-term trends of changes in the patient's health status. In the LSTM model, the state H(t) of the hidden layer is used to describe the patient's health status at time t. For Mr. Zhang's data, the LSTM network outputs the hidden layer state H(t) describing his health status based on the input time series data.

[0133] For example, after being processed by the LSTM network, the value of Mr. Zhang's health status H(6) at time point t = 6 hours is, for example, 0.78, indicating that the patient's health status at this time point is relatively good and there is a certain safety distance from the occurrence of delirium. At t = 8 hours, the status H(8) is 0.52, indicating that the health status is poor and the risk of delirium begins to increase.

[0134] Next, to better capture the periodic fluctuations in the patient's physiological data, the system uses a time series smoothing function. In particular, the system introduces modulation terms and time decay factors to capture changes in heart rate and blood oxygen levels. These fluctuations are affected by factors such as circadian rhythms, so the data needs to be smoothed and processed by integration within a time window.

[0135] The time series smoothing function used is as follows:

[0136]

[0137] In this formula, H(τ) represents the hidden layer state output by the LSTM model, that is, the health characteristics of the patient at different time points; γ is the modulation parameter used to adjust the strength of the periodic signal; sin(ωτ) is the periodic modulation term, where ω is the frequency parameter used to capture periodic changes in the data (for example, Mr. Zhang’s heart rate and blood oxygen have circadian rhythmic fluctuations); e-βτ is the time decay factor, where β controls the impact of old data on the current analysis. Usually, β ranges from 0.01 to 0.1, and a larger β accelerates the decay of old data and emphasizes the impact of recent data.

[0138] Set the modulation parameter γ=0.3, the frequency parameter ω=0.2 (indicating a periodic fluctuation of once every 5 hours), the attenuation factor β=0.05, and the time window length T=12 hours (i.e., smoothing the data within the past 12 hours). Then, Mr. Zhang's health data within the past 12 hours can be smoothed to capture the periodic fluctuations in his heart rate and blood oxygen.

[0139] Assume that Mr. Zhang’s hidden layer state H(6) at t=6 hours is 0.78, the healthy state at t=7 hours is 0.75, at t=8 hours is 0.52, and at t=9 hours is 0.56. The smoothed values ​​are:

[0140]

[0141] This integration process can be solved by numerical calculation methods, and the result is set to 0.60, which represents the health status value after time smoothing. By smoothing the time series of Mr. Zhang's health data, a more stable health status trend is obtained. These stable data are used to further extract periodic fluctuation trends and build a comprehensive risk assessment model by fusing them with other health data (such as blood oxygen level, body temperature, etc.). Based on the health status output by LSTM and the results of time series smoothing, the system can extract short-term and long-term health change trends, thereby predicting the probability of delirium.

[0142] Suppose that in the next few hours, the system continuously monitors Mr. Zhang's health status and fluctuates. Combined with other health parameters, the system can output a comprehensive risk score. For example, the system calculates that Mr. Zhang's delirium risk score in the next few hours is 0.82, indicating that the patient has entered a higher risk area. At this time, the system will trigger an early warning and provide clinicians with personalized intervention recommendations, such as drug adjustments, enhanced care, and improved environmental stimulation.

[0143] In the process of assessing Mr. Zhang's delirium risk, the system used the XGBoost algorithm based on ensemble learning to integrate multiple data sources into a unified prediction model. The model not only includes the patient's static characteristics, such as age, gender, and underlying diseases, but also incorporates the patient's dynamic time series characteristics and inducing factor information, and generates an individualized delirium risk score through calculation.

[0144] First, the system collects multi-dimensional data of Mr. Zhang, including static features and dynamic features. The static features of Mr. Zhang are set as follows:

[0145] Age: 65

[0146] Gender: Male

[0147] Underlying diseases: hypertension, diabetes

[0148] The dynamic time series features come from Mr. Zhang’s heart rate, blood oxygen, body temperature and other data. Assuming that after LSTM processing, Mr. Zhang’s hidden layer state sequence H(t) in the past 12 hours reflects his health status, and the time series features after smoothing are:

[0149] H(t)=[0.78,0.75,0.72,0.70,0.68,0.66,0.63,0.60,0.58,0.57,0.55,0.53]

[0150] Each value represents Mr. Zhang’s health status at a different time point.

[0151] In addition, Mr. Zhang’s predisposing factor data were also incorporated into the model, including acute illness history, drug dependence, and activity level.

[0152] History of acute illness: None

[0153] Drug dependence: Yes (using sedatives)

[0154] Activity level: Low (due to long-term bed rest, reduced activity)

[0155] To construct an individualized delirium risk score, the system uses the XGBoost (Extreme Gradient Boosting) algorithm, which is a powerful machine learning method based on ensemble learning. XGBoost can automatically learn and adjust feature weights based on training data to optimize the risk prediction model. By taking Mr. Zhang’s static features, dynamic features, and predisposing factors as input, the XGBoost algorithm outputs the weight w of each feature. i , these weights reflect the relative importance of each feature in delirium risk assessment.

[0156] The XGBoost model is set to learn and output the following weights for static features (such as age, gender, and underlying diseases) and dynamic features (such as time series health status):

[0157] w 年龄 =0.25

[0158] w 性别 =0.15

[0159] w 高血压 =0.2

[0160] w糖尿病 =0.1

[0161] w 药物依赖 =0.3

[0162] w 低活动水平 =0.25

[0163] These weights are automatically adjusted by XGBoost to represent the contribution of each feature to delirium risk.

[0164] Next, all input features (both static and dynamic) are combined using a fusion formula to generate an individualized delirium risk score. The formula is as follows:

[0165]

[0166] in:

[0167] F i Represents input features, such as Mr. Zhang’s static features (age, gender, underlying diseases, etc.) and predisposing factors (drug dependence, activity level, etc.).

[0168] w i are the feature weights that are automatically learned and adjusted by the XGBoost model.

[0169] φ(H(t)) is the risk adjustment function of the time series characteristic H(t), which is used to enhance the sensitivity to dynamic characteristics (such as physiological fluctuations). It means integrating dynamic features and calculating the impact of historical data on current risks. T is the defined time range, usually the past 12 hours or 24 hours.

[0170] The risk calculation for Mr. Zhang is set to proceed as follows:

[0171] 1. Static feature contribution:

[0172] F 年龄 =65,w 年龄 =0.25, so the contribution is 0.25×65=16.25

[0173] F 性别 =1 (1 for male and 0 for female), w 性别 =0.15, the contribution is 0.15×1=0.15

[0174] F 高血压 =1,w 高血压 =0.2, the contribution is 0.2×1=0.2

[0175] F 糖尿病 =1,w 糖尿病 =0.1, the contribution is 0.1×1=0.1

[0176] F 药物依赖 =1,w 药物依赖 =0.3, the contribution is 0.3×1=0.3

[0177] F 低活动水平 =1,w 低活动水平 =0.25, the contribution is 0.25×1=0.25

[0178] Total static feature contribution = 16.25 + 0.15 + 0.2 + 0.1 + 0.3 + 0.25 = 17.25

[0179] 2. Dynamic feature contribution:

[0180] For Mr. Zhang's dynamic time series characteristics H(t), the risk adjustment function φ(H(t)) is set to make the following adjustments based on Mr. Zhang's health status. According to the time series data, the change of the function φ(H(t)) will be affected by heart rate fluctuations, blood oxygen changes, etc. The integral result is set as:

[0181]

[0182] This is the result obtained by weighted integration of health status data in the past 12 hours, indicating the impact of historical data on current risk.

[0183] Final risk score calculation:

[0184] Add the static feature contribution to the dynamic feature contribution:

[0185] R delirium =17.25+0.45=17.7

[0186] According to the set risk score threshold, if R delirium >15, Mr. Zhang is assessed as a high-risk group, and the system will trigger a delirium risk warning and generate personalized intervention recommendations, such as adjusting medication use, strengthening care, and improving environmental stimulation.

[0187] Through the multi-factor fusion method based on XGBoost, combined with Mr. Zhang's static features, dynamic features and inducing factors, the system successfully calculated an individualized delirium risk score. This comprehensive evaluation method can not only accurately predict Mr. Zhang's risk of delirium, but also provide clinicians with detailed intervention recommendations, help optimize patient care plans, and improve the accuracy and effectiveness of clinical decision-making.

Claims

1. A delirium risk assessment method, characterized in that: The following steps are involved: S1. Multidimensional data collection and dynamic feature capture of delirium-inducing factors: S1.

1. Design an integrated multidimensional data collection system for multiple delirium-causing factors including infection, drug use, metabolic disorders, and sleep disorders; S1.

2. Collect data in real time from patients’ vital sign monitoring devices, electronic medical record systems, and nursing records, including physiological parameters such as heart rate variability, blood oxygen levels, and temperature fluctuations, biochemical indicators including inflammatory factors, renal function, and electrolyte balance, and behavioral data including sleep patterns, activity levels, and cognitive test results; S2. High-dimensional time series analysis of delirium characteristics: S2.

1. Use time series analysis techniques to mine the changing patterns of multiple physiological and behavioral parameters of patients and identify early signs of delirium; S2.

2. Use the long short-term memory network (LSTM) deep learning algorithm to analyze the parameter fluctuation trends of patients in different time periods and extract the time series characteristics of delirium precursors; S3. Multi-factor integration for delirium risk assessment: Based on the time series analysis results of delirium characteristics, a multi-factor fusion risk assessment model was constructed. An ensemble learning-based algorithm was used to fuse high-risk inducing factors including acute illness, long-term bed rest, and drug dependence with dynamic feature data to generate an individualized delirium risk score for patients. S4. Real-time delirium risk dynamic warning and intervention strategy generation; Through the risk assessment model, the patient's real-time data is continuously monitored, and a warning signal is generated when a high-risk state is detected. At the same time, based on the patient's individual characteristics and risk scores, including age, underlying diseases, and medication usage, targeted intervention strategy recommendations are generated, including adjusting medications, improving environmental stimulation, and optimizing nursing processes. The generation of intervention strategies uses decision trees and reinforcement learning algorithms.

2. A delirium risk assessment method according to claim 1, characterized in that High-dimensional time series analysis method for delirium characteristics: Collect multidimensional health data of patients from various data sources, including physiological parameters, behavioral data, cognitive function assessment results and drug use, and clean and pre-process the data; use a cleaning formula that combines derivative adjustment and exponential decay; detect abnormal data points and smooth them. The formula is as follows: Where: x represents the original input data point, including the patient's heart rate measurement; α adjusts the weight parameter to control the strength of outlier removal; represents the derivative operation on the data x, outliers(x) is an outlier detection function used to identify outliers in the data; Exponential decay factor, where λ controls the decay speed.

3. A delirium risk assessment method according to claim 2, characterized in that The high-dimensional time series analysis method for delirium characteristics: the preprocessed multidimensional data is serialized and divided into multiple time windows of fixed length including 1 hour and 3 hours; each time window contains multiple data points for describing the health status of the patient in the time period; a time smoothing method based on a modulation function is used to capture the periodic fluctuations in the patient's physiological signals; and the periodic changes including heart rate and respiratory cycle are modeled by introducing modulation terms.

4. A delirium risk assessment method according to claim 3, characterized in that High-dimensional time series analysis method for delirium characteristics: An LSTM neural network was introduced to analyze the patient's time series data and predict the risk of delirium. The risk prediction process of the LSTM network includes multiple layers: input layer, hidden layer, and output layer. First, the smoothed time series data is input into the LSTM network; The model extracts key features from the time series by recursively updating the hidden state; it uses the risk score integration method to accumulate the risk score within the time window; and uses the expression formula: f risk (R,Δt)=∫ t t+Δt S(R(t))dR Among them, R(t) represents the output risk score of the LSTM model at time t, which indicates the patient's delirium risk at that time point; S(R) represents the Softmax function, which is used to convert the risk score R(t) into the probability of classification including "high risk", "medium risk" or "low risk", and controls the output risk value between 0 and 1; Δt time integration interval is used to define the cumulative risk assessment within time; f risk (R,Δt) The cumulative risk score is used to comprehensively analyze the patient's delirium risk in the future unit time; when the cumulative score exceeds the threshold θ, an early warning signal will be triggered; when the cumulative risk score exceeds the set threshold θ, an early warning will be generated and personalized intervention recommendations will be provided.

5. A delirium risk assessment method according to claim 1, characterized in that The multi-factor fusion method for delirium risk assessment includes: First, data preprocessing is performed, including data cleaning, outlier removal, and time series smoothing. During the data cleaning process, an outlier cleaning method based on derivative adjustment and exponential decay is used to remove abnormal data points while retaining key time series features. The derivative is calculated to adjust the sensitivity of data fluctuations.

6. A delirium risk assessment method according to claim 5, characterized in that The multi-factor fusion method for delirium risk assessment includes: The preprocessed time series data is deeply analyzed; LSTM captures the time dependency of the data and extracts the short-term and long-term health change trends related to delirium; after the input is processed by LSTM, the hidden layer state H(t) is output to describe the patient's health status at time t; a time series smoothing function is used; by introducing modulation terms and time decay factors, the periodic fluctuations of the patient's physiological data, including changes in heart rate and blood oxygen level, are captured: g seq (H(t))=∫0 T H(τ)·(1+γsin(ωτ))e -βτ dt in: H(τ) represents the hidden layer state of TM output, which represents the health characteristics of the patient at time τ; γ represents the modulation parameter, which is used to adjust the intensity of the periodic signal; sin(ωτ) represents the periodic modulation term, where ω is the frequency parameter, which is used to capture the periodic changes in the data, including the circadian rhythm; e -βτ Represents the time decay factor, where β controls the impact of old data on the current analysis, and β accelerates the decay of old data; t represents the length of the time window, which is used to define the integral range of the smoothing process.

7. A delirium risk assessment method according to claim 6, characterized in that The multi-factor fusion method for delirium risk assessment includes: The XGBoost algorithm based on ensemble learning is used to integrate the state features including age, gender, and history of underlying diseases, as well as dynamic time series features. seq (H(t)) and inducing factor data were fused into a unified prediction model; an individualized delirium risk score was generated using the following multifactor fusion formula: Among them: F i represents the input features of the same triggering factors including acute illness, drug dependence, and activity level; w i is the weight, which is automatically learned and adjusted by the XGBoost model; φ(H(t)) represents the risk adjustment function of the time series feature H(t), which is used to enhance the sensitivity to dynamic features; Indicates that column features are integrated to accumulate the impact of historical data on current risks; T is the time range used to define the weight of historical data in risk assessment.

Images