A method and device for predicting chronic disease trends based on health record time-series data

Abstract

This application discloses a method and apparatus for predicting chronic disease trends based on time-series health record data. It acquires users' structured records and organizes them into a clean time-series sequence. Based on the clean time-series sequence, it calculates the average first-order rate of change and its direction, as well as the average second-order acceleration and its direction. It calculates a rate of change factor based on the absolute value of the average first-order rate of change, a current state factor based on the degree to which the current measurement deviates from the reference range, and an abnormal persistence factor based on the number of consecutive unidirectional changes or consecutive out-of-range occurrences. These factors are then weighted and summed to obtain a risk score, which is mapped to an interpretable risk level. Based on the risk level, key indicators, and user profile, a structured reference prevention plan is generated, along with a trend prediction report. This application achieves early trend warnings and individualized prevention recommendations earlier than a single-point threshold through dual quantification of time-series characteristics, three-factor weighted grading, and a historical response feedback loop.

Description

Technical Field

[0001] This application relates to the field of medical and health data processing technology, specifically to a method and device for predicting chronic disease trends based on time-series health record data. Background Technology

[0002] With the rapid development of the economy and society and the acceleration of population aging, chronic non-communicable diseases (hereinafter referred to as "chronic diseases") have become a major threat to the health of Chinese residents. Chronic diseases such as hypertension, diabetes, and dyslipidemia are characterized by long disease courses, insidious onset, and serious complications. Their effective prevention and control highly depend on long-term, continuous health monitoring and early intervention. Therefore, the importance of chronic disease management and health monitoring is increasingly prominent, as it not only relates to improving individual quality of life but also to the rational allocation of public health resources.

[0003] In chronic disease management and health monitoring, "early detection and early intervention" is a recognized core principle. However, when users face physical examination or laboratory reports, they usually only get a static comparison of a "current value" and a "reference range." This presentation method can only help users judge whether the indicator is "high" or "low," but it cannot answer three deeper and more practical questions:

[0004] (1) A single test is not enough to tell the trend. For example, if fasting blood glucose drops from 7.2 mmol / L to 6.5 mmol / L, it may seem like an improvement, but the user cannot tell whether it is a continuation of a slow decline or an occasional drop in a short period of fluctuation.

[0005] (2) The rate of deterioration is ignored. If blood pressure rises steadily from 130 / 85 to 150 / 95 within three months, each single-point measurement may only trigger an "out of range" alert, but cannot convey the trend signal of "continuous deterioration", thus missing the opportunity for intervention.

[0006] (3) Prevention advice is uniform. The general advice of health apps on the market for "high blood sugar" (such as "eat less sweets") is almost the same for men aged 30, women aged 65, patients diagnosed with diabetes or pregnant women. It lacks individualized information such as age, gender, medical history and medication history.

[0007] To solve the above problems, trend analysis is an indispensable foundation, and trend calculation must be based on continuous historical data. In fact, the frequency of routine physical examinations in my country is once a year, follow-up for chronic diseases is once every 1 to 3 months, and the interval for acute examinations is even shorter. This means that the vast majority of users have accumulated historical data from 3 to 50 time points, which provides the data foundation for trend mining.

[0008] Unfortunately, current health apps, smart bracelets, and insurance membership services generally have rudimentary "health record" or "health trend" modules. Their mainstream approach is to use "historical graphs" and "reaching-target reminders" as the primary visualization methods, with only a few applications introducing single-dimensional quantitative indicators such as "weekly / monthly change rates." While these functions provide some reference, they mostly stop at "data display" rather than "trend interpretation," failing to effectively answer the three deeper questions mentioned above, let alone translate them into personalized intervention strategies. Therefore, it is clear that there is still a significant gap in product and understanding between "having data" and "understanding data."

[0009] If we categorize and analyze these mainstream practices, we can find that their limitations are not accidental, but rather due to inherent shortcomings in their respective design logics:

[0010] (1) The "historical curve chart" commonly used in smart bracelets and health apps does not quantify time-series information: This type of function connects measurement points into a broken line along the time axis, leaving all the raw data to the user to interpret. Its shortcomings are: the speed of the trend requires subjective judgment by human eyes and lacks quantitative slope indicators; even if the user can see that it is "rising", they cannot know whether the speed of rise is accelerating, that is, they lack acceleration information; more importantly, each point on the broken line is still regarded as an independent measurement value, and the system does not give it the overall semantic meaning of "continuous deterioration". What the user sees is just a series of isolated time points, rather than a continuous trend.

[0011] (2) The "out-of-range alert" of the single-point threshold causes the warning to be delayed. That is, when a certain measurement value crosses the reference range, a push notification is triggered. This threshold mechanism has three limitations: First, the reference range is derived from population statistics and is not combined with the individual baseline. For example, the normal blood pressure range of a 65-year-old male is different from that of a young man, but he receives the same alarm threshold. Second, the alert is delayed. The indicator must "exceed the standard first and then be notified", and it is impossible to give an early warning before the trend crosses the boundary. Third, it cannot distinguish between "one-time abnormality" and "continuous deterioration". Blood pressure fluctuations after one exercise and steady increases over three consecutive months will be treated equally by the same push.

[0012] (3) General health knowledge base recommendation to prevent templates from becoming disconnected from history: When abnormal indicators are detected, the system directly pushes several suggested texts from the preset knowledge base. The problem with this type of method is that the recommended content is completely static and does not change with the user profile (age, gender, medical history, medication history); the system does not remember the user's historical responses, so even if the user ignored the "increase exercise" suggestion last time, it will still push it again without any changes; more importantly, the suggestions lack risk grading, and the guidance obtained for blood sugar that is 1 unit higher and 3 units higher is almost identical, which fails to reflect the urgency of differentiated intervention. Summary of the Invention

[0013] Therefore, this application provides a method and device for predicting chronic disease trends based on health record time-series data, in order to solve the problems of unquantified time-series information, delayed early warning, and disconnect between prevention templates and historical data in existing chronic disease management and health monitoring technologies.

[0014] To achieve the above objectives, this application provides the following technical solution:

[0015] Firstly, a method for predicting chronic disease trends based on time-series health record data includes:

[0016] Step 1: Obtain the user's structured records from the health record system, sort the multiple measurements of the same indicator in the structured records by timestamp, and then process them by handling missing values ​​and outlier marking to form a clean time series sequence;

[0017] Step 2: Based on the time interval between adjacent effective points in the cleaning time series, calculate the first-order rate of change between each adjacent effective point, and average all first-order rates of change to obtain the average first-order rate of change and its direction.

[0018] Step 3: Perform adjacent differences on the first-order rate of change sequence, calculate the second-order acceleration, and average all second-order accelerations to obtain the average second-order acceleration and its direction;

[0019] Step 4: Calculate the rate of change factor based on the absolute value of the average first-order rate of change, calculate the current state factor based on the degree to which the current measured value deviates from the reference range, and calculate the anomaly persistence factor based on the number of consecutive changes in the same direction or consecutive out-of-range occurrences; when the direction of the average first-order rate of change is opposite to the direction of the average second-order acceleration, increase the weight of the anomaly persistence factor and correspondingly decrease the weight of the rate of change factor.

[0020] Step 5: Perform a weighted summation of the rate of change factor, the current state factor, and the anomaly persistence factor to obtain a risk score, and map the risk score to an interpretable risk level;

[0021] Step 6: Based on the risk level and attention indicators, match the preset rule set and inject the user profile to generate a structured reference prevention plan; the strength of the reference prevention plan is dynamically adjusted according to the user's historical response behavior to the previous prevention plan.

[0022] Step 7: Based on the risk level, the focus indicators, and the reference prevention plan, generate a trend prediction report, record the initial status of the plan items in the trend prediction report, and write the user's adoption or ignoring behavior of the plan items into the feedback log for the next adjustment of the intensity of the reference prevention plan.

[0023] Optionally, in step 1, the missing value processing specifically involves: if the interval between adjacent time points is less than or equal to 30 days, linear interpolation is used to fill in the missing values, and the filled data points are marked as interpolation points; if the interval between adjacent time points is greater than 180 days, the missing values ​​are retained without interpolation; the interpolation points are not used in subsequent calculations of the first-order rate of change and the second-order acceleration.

[0024] Optionally, in step 1, the outlier marking process specifically involves: using the individual baseline ± 3 times the standard deviation or the upper and lower limits of the reference interval multiplied by 1.5, taking the union of the two criteria, and marking the measured values ​​that exceed the range as outliers; the outliers are not used in the subsequent calculation of the first-order rate of change and the second-order acceleration.

[0025] Optionally, in step 2, the method for determining the direction of the average first-order rate of change is as follows: the average first-order rate of change is compared with a preset direction threshold. If it is greater than the preset direction threshold, it is determined to be an upward direction; if it is less than the preset direction threshold, it is determined to be a downward direction; otherwise, it is determined to be a stable direction.

[0026] Optionally, in step 3, the method for determining the direction of the average second-order acceleration is as follows: if the average second-order acceleration is greater than 0, it is determined to be an acceleration deviation direction; if the average second-order acceleration is less than 0, it is determined to be a deceleration direction; if the absolute value of the average second-order acceleration is less than a preset threshold, it is determined to be an insignificant acceleration.

[0027] Optionally, in step 4, the calculation process of the rate of change factor is as follows: configure a severe threshold set according to the indicator type, compare the absolute value of the average first-order rate of change with the severe threshold set, and normalize it to the [0,1] interval by looking up a table.

[0028] Optionally, in step 4, the calculation process of the current state factor is as follows: calculate the difference between the current measured value and the lower limit of the reference interval, divide it by the difference between the upper and lower limits of the reference interval to obtain the deviation, and normalize the deviation to the [0,1] interval; if the current measured value is within the reference interval, calculate the normalized value based on the one closer to the upper or lower limit of the reference interval.

[0029] Optionally, in step 4, the calculation process of the abnormal persistence factor is as follows: count the number of consecutive changes in the same direction or consecutive out-of-range events, and map the number to 1.0 when the number is greater than or equal to 3, to 0.6 when it is equal to 2, to 0.3 when it is equal to 1, and to 0 when it is equal to 0.

[0030] Optionally, in step 5, when mapping the risk score to an interpretable risk level, the interval [0, 0.4) is mapped to low risk, the interval [0.4, 0.7) is mapped to medium risk, and the interval [0.7, 1.0] is mapped to high risk.

[0031] Secondly, a chronic disease trend prediction device based on health record time-series data includes:

[0032] The data acquisition module is used to acquire the user's structured records from the health record system, sort the multiple measurements of the same indicator in the structured records by timestamp, and then process them into a clean time series after missing value processing and outlier labeling.

[0033] The first-order rate of change calculation module is used to calculate the first-order rate of change between each adjacent effective point based on the time interval between adjacent effective points in the cleaning time series, and to calculate the average first-order rate of change and its direction by averaging all the first-order rates of change.

[0034] The second-order acceleration calculation module is used to perform adjacent differences on the first-order rate of change sequence, calculate the second-order acceleration, and calculate the average of all second-order accelerations to obtain the average second-order acceleration and its direction.

[0035] The three-factor weighting module is used to calculate the rate of change factor based on the absolute value of the average first-order rate of change, calculate the current state factor based on the degree to which the current measured value deviates from the reference range, and calculate the anomaly persistence factor based on the number of consecutive changes in the same direction or consecutive out-of-range events. When the direction of the average first-order rate of change is opposite to the direction of the average second-order acceleration, the weight of the anomaly persistence factor is increased, and the weight of the rate of change factor is decreased accordingly.

[0036] The risk grading module is used to perform a weighted summation of the rate of change factor, the current state factor, and the abnormal persistence factor to obtain a risk score, and then map the risk score to an interpretable risk level.

[0037] The prevention plan generation module is used to match a preset rule set and inject user profiles based on the risk level and attention indicators to generate a structured reference prevention plan; the strength of the reference prevention plan is dynamically adjusted according to the user's historical response behavior to the previous prevention plan.

[0038] The report generation and feedback module is used to generate a trend prediction report based on the risk level, the focus indicators, and the reference prevention plan, record the initial status of the plan items in the trend prediction report, and write the user's adoption or ignoring behavior of the plan items into the feedback log for the next adjustment of the intensity of the reference prevention plan.

[0039] Compared with the prior art, this application has at least the following beneficial effects:

[0040] This application provides a method for predicting chronic disease trends based on time-series health record data. It acquires users' structured records and organizes them into a clean time-series sequence. Based on the time interval between adjacent valid points in the clean time-series sequence, it calculates the average first-order rate of change and its direction. It performs adjacent differencing on the first-order rate of change sequence to calculate the second-order acceleration, and averages all second-order accelerations to obtain the average second-order acceleration and its direction. It calculates a rate of change factor based on the absolute value of the average first-order rate of change, a current state factor based on the degree to which the current measurement deviates from the reference interval, and an abnormal persistence factor based on the number of consecutive unidirectional changes or consecutive out-of-range occurrences. These factors are then weighted and summed to obtain a risk score, which is mapped to an interpretable risk level. Based on the risk level, key indicators, and user profile, a structured reference prevention plan is generated, along with a trend prediction report. This application achieves early trend warnings and individualized prevention suggestions earlier than single-point thresholds through dual quantification of time-series characteristics, three-factor weighted grading, and a historical response feedback loop. This solves the problems of unquantified time-series information, delayed warnings, and disconnect between prevention templates and historical data in existing technologies. Attached Figure Description

[0041] To more intuitively illustrate the prior art and this application, exemplary drawings are provided below. It should be understood that the specific shapes and structures shown in the drawings should not generally be regarded as limiting conditions for implementing this application; for example, based on the technical concept disclosed in this application and the exemplary drawings, those skilled in the art are able to easily make conventional adjustments or further optimizations to the addition / reduction / classification, specific shapes, positional relationships, connection methods, size ratios, etc. of certain units (components).

[0042] Figure 1 A flowchart of a chronic disease trend prediction method based on health record time-series data provided in Embodiment 1 of this application;

[0043] Figure 2 This is a schematic diagram of a chronic disease trend prediction method based on health record time series data provided in Embodiment 1 of this application. Detailed Implementation

[0044] The present application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0045] In the description of this application: unless otherwise stated, "a plurality of" means two or more. The terms "first," "second," "third," etc., in this application are intended to distinguish the objects referred to and do not have any special meaning in terms of technical connotation (e.g., they should not be construed as an emphasis on importance or order). Expressions such as "including," "comprising," and "having" also mean "not limited to" (certain units, components, materials, steps, etc.).

[0046] The terms used in this application, such as "upper," "lower," "left," "right," and "middle," are generally used to indicate the general relative positional relationship for the purpose of intuitive understanding by referring to the accompanying drawings, and are not absolute limitations on the positional relationship in the actual product.

[0047] Example 1

[0048] Please see Figure 1 and Figure 2 This embodiment provides a method for predicting chronic disease trends based on time-series health record data. This method runs on a WeChat mini-program or a standalone app client, and includes:

[0049] S1: Obtain the user's structured records from the health record system, sort the multiple measurements of the same indicator in the structured records by timestamp, and then process the missing values ​​and outlier labels to form a clean time series.

[0050] Specifically, the purpose of this step is to organize multiple measurements of the same indicator from the structured records of the health record system into a time series of equal length that can be calculated, providing clean input for downstream rate of change calculation.

[0051] The missing value handling process is as follows: if the interval between adjacent time points is less than or equal to 30 days, linear interpolation is used to fill in the missing values, and the filled data points are marked as interpolation points; if the interval between adjacent time points is greater than 180 days, the missing values ​​are retained and no interpolation is performed; the interpolation points are not used in subsequent calculations of the first-order rate of change and the second-order acceleration.

[0052] The outlier labeling process is as follows: take the individual baseline ± 3 times the standard deviation or the upper and lower limits of the reference interval multiplied by 1.5, take the union of the two criteria, and mark the measurements that are out of range as outliers; outliers are not used in subsequent calculations of first-order rate of change and second-order acceleration.

[0053] Example:

[0054] Input: A structured collection of records R = {r_i}, where each record contains the fields {patient_id, indicator_code, value, unit_std, ref_range, time, source_org}.

[0055] The processing logic for this step is as follows:

[0056] S101: Group by (patient_id, indicator_code) primary key;

[0057] S102: Sort the groups in ascending order by time (timestamp), generating the original sequence X=[ , , ..., ] and the corresponding timestamp T = [ , , ..., ];

[0058] S103: Missing value handling. If the interval between adjacent time points is >180 days, it is marked as "long interval" and the missing value is retained without imputation. If the missing value occurs within the interval between adjacent time points ≤30 days, linear interpolation is used to fill in the missing value. The imputation point is marked with is_imputed=True and the point is not included in the difference calculation in the downstream calculation.

[0059] S104: Outlier marking. Use one of the two criteria, "personal baseline ±3σ" or "reference interval upper and lower limits ×1.5" (take the union), to mark points whose measured values ​​are far outside the reasonable range with is_outlier=True. These points will not be included in the difference calculation in the downstream calculation.

[0060] S105: Sequence length check. If the number of valid points (is_imputed=False and is_outlier=False) is less than 3, it is marked as "insufficient sequence" and enters the "insufficient data" branch (see abnormal branch a).

[0061] Output: The “clean time series” for each metric includes: an array of valid values, an array of valid timestamps, a mask of valid values, the series length, and the individual baseline mean ± std.

[0062] Abnormal branch:

[0063] a. Insufficient sequence: Do not proceed to subsequent steps; mark this indicator as "unable to calculate trend".

[0064] b. Long interval ratio > 50%: This is considered insufficient sampling density, only the first-order rate of change is calculated, and the second-order acceleration is skipped;

[0065] c. If all points are outliers, remove that indicator from the current forecast.

[0066] It should be noted that in this step, ① missing value imputation can be replaced with multiple imputation (MI) or "Last Observation Carryover" (LOCF) based on the previous valid observation; ② outlier criterion can be replaced with "personal baseline ± 2.5σ" or "density-based LOF algorithm"; ③ sampling density determination can be replaced with "adaptive threshold based on index" (such as blood pressure for 14 days, weight for 30 days).

[0067] S2: Based on the time interval between adjacent effective points in the cleaning time series, calculate the first-order rate of change between each adjacent effective point, and average all first-order rates of change to obtain the average first-order rate of change and its direction.

[0068] Specifically, the purpose of this step is to convert the time series into two comparable quantization quantities: "direction" and "velocity".

[0069] The method for determining the direction of the average first-order rate of change is as follows: the average first-order rate of change is compared with a preset direction threshold. If it is greater than the preset direction threshold, it is determined to be an upward direction; if it is less than the preset direction threshold, it is determined to be a downward direction; otherwise, it is determined to be a stationary direction.

[0070] Example:

[0071] Input: Cleaning time sequence X, T, mask;

[0072] The processing logic for this step is as follows:

[0073] S201, First-order rate of change (i.e., difference slope or rate): for each pair of adjacent valid points ( , According to the time difference Δt = - (Day) Calculate the slope per unit time The unit is "indicator unit / day";

[0074] S202, First-order direction determination: For each index Find the mean of a sequence And compare it with the "indicator-specific directional threshold δ": > +δ is judged as "rising". < -δ is judged as "decreasing", otherwise it is judged as "stable".

[0075] Output: (for each metric) , ).

[0076] It should be noted that the first-order rate of change in this step can be replaced with the difference after exponentially weighted moving average (EWMA) to reduce the weight of long-term data; in addition, the direction determination can be replaced with the sign test based on the confidence interval to avoid misjudgment due to slight differences.

[0077] S3: Perform adjacent differences on the first-order rate of change sequence, calculate the second-order acceleration, and average all second-order accelerations to obtain the average second-order acceleration and its direction;

[0078] Specifically, the method for determining the direction of the average second-order acceleration is as follows: if the average second-order acceleration is greater than 0, it is determined to be an acceleration deviation direction; if the average second-order acceleration is less than 0, it is determined to be a deceleration direction; if the absolute value of the average second-order acceleration is less than a preset threshold, it is determined to be an insignificant acceleration.

[0079] Example:

[0080] S301, Second-order change (i.e., adjacent slope difference or acceleration): Perform another adjacent difference on the first-order slope sequence to obtain the acceleration sequence. Similarly, calculate the mean. ;

[0081] S302, Acceleration direction determination: >0 indicates that "the indicator is deviating at an accelerating rate" (in a worsening direction). < 0 indicates that "the indicator is slowing down", | | If the threshold is less than the threshold, it is judged as "insignificant acceleration".

[0082] Assuming a user's fasting blood glucose sequence is [6.2, 6.5, 6.8, 7.0, 7.2, 7.1] (mmol / L), with time intervals evenly spaced at 30 days, the calculation yields... = +0.018 mmol / L / day = -0.0012 mmol / L / day², the direction is determined as "rising but the acceleration is slowing down", which means that blood sugar is rising but the rise is slowing down.

[0083] After steps S2 and S3, the quadruple for each indicator can be obtained ( , , , (i.e., "speed + direction + acceleration + acceleration direction").

[0084] Abnormal branch:

[0085] a. When step S1 is marked as "insufficient sequence" or "first order only", only output... ;

[0086] b. direction and If there is a directional contradiction (one rising and one falling), a "contradictory signal" is output, which is then handled by the abnormal persistence factor in the subsequent step S4.

[0087] It should be noted that in this step, the acceleration can be replaced with "second-order coefficients of quadratic polynomial fitting", which is more robust to noise; in addition, the direction determination can also be replaced with "sign test based on confidence interval" to avoid misjudgment due to "slight difference".

[0088] S4: Calculate the rate of change factor based on the absolute value of the average first-order rate of change, calculate the current state factor based on the degree to which the current measured value deviates from the reference range, and calculate the anomaly persistence factor based on the number of consecutive changes in the same direction or consecutive out-of-range events; when the direction of the average first-order rate of change is opposite to the direction of the average second-order acceleration, increase the weight of the anomaly persistence factor and correspondingly decrease the weight of the rate of change factor.

[0089] Specifically, the purpose of this step is to construct a three-factor weighted risk-sharing model based on the current state and the quantified results of steps S2 and S3.

[0090] The calculation process of the rate of change factor R is as follows: configure a severe threshold set according to the indicator type, compare the absolute value of the average first-order rate of change with the severe threshold set, and normalize it to the [0,1] interval by looking up a table.

[0091] The current state factor S is calculated as follows: the difference between the current measured value and the lower limit of the reference interval is divided by the difference between the upper and lower limits of the reference interval to obtain the deviation, and the deviation is normalized to the [0,1] interval; if the current measured value is within the reference interval, the normalized value is calculated based on the one closer to the upper or lower limit of the reference interval.

[0092] The calculation process of the abnormal persistence factor P is as follows: count the number of consecutive changes in the same direction or consecutive out-of-range events. When the number is greater than or equal to 3, it is mapped to 1.0; when it is equal to 2, it is mapped to 0.6; when it is equal to 1, it is mapped to 0.3; and when it is equal to 0, it is mapped to 0.

[0093] Example:

[0094] Input: (for each metric) , , , ), current value Reference range [L, U], Abnormal persistence count Individual baseline μ.

[0095] Processing logic:

[0096] S401, Factor 1 R (rate of change): by | | Normalize to the [0,1] interval by looking up a table in the "Severity Threshold Set". The severity threshold set is configured by indicator (e.g., fasting blood glucose). ≥0.05 mmol / L / day is calculated as 1.0, 0.03-0.05 is calculated as 0.6, and the rest are calculated using linear interpolation).

[0097] S402, Factor 2 S (Current State): Calculate "Deviation" = ( - L) / (U - L) represents the degree to which the value exceeds the interval, and is normalized to [0,1]; if the value is within the interval, it is normalized to the value closer to the upper and lower limits of the distance.

[0098] S403, Factor Three P (Abnormal Persistence): Indicates continuous out-of-range or continuous The number of times in the same direction, ≥3 counts as 1.0, =2 counts as 0.6, =1 counts as 0.3, =0 counts as 0.

[0099] S5: The risk score is obtained by weighted summation of the rate of change factor, current state factor, and abnormal persistence factor, and the risk score is mapped to an interpretable risk level;

[0100] Specifically, the formula for calculating the risk score is as follows:

[0101] = 0.4×R + 0.35×S + 0.25×P.

[0102] It should be noted that the weights of the three factors can be configured differently according to the type of indicator (e.g., when the indicator is blood pressure, the weight of R is 0.5, and when the indicator is weight, the weight of P is 0.4).

[0103] When mapping risk scores to interpretable risk levels, the interval [0, 0.4) is mapped to low risk, the interval [0.4, 0.7) is mapped to medium risk, and the interval [0.7, 1.0] is mapped to high risk.

[0104] The final interpretable risk level includes: Risk Level + Three-Factor Decomposition Value + Top-N Key Indicators (based on...) Take the first 3-5 items in descending order.

[0105] Assuming the user's blood glucose level in the example above is: R = 0.6 (between 0.03 and 0.05), S = 0.85 (7.2, very close to the upper limit of 7.0), and P = 1.0 (five consecutive increases in the same direction), then... = 0.4×0.6+0.35×0.85+0.25×1.0 = 0.24+0.30+0.25 = 0.79, which is judged as "high risk"; the three-factor decomposition also shows "moderate rate, state close to the upper limit, and extremely strong persistence", which makes it easy for doctors to locate at a glance. Then, the risk score, risk level, (risk factor decomposition) of each indicator are output, as well as the list of Top-N indicators of concern for this prediction.

[0106] Abnormal branch:

[0107] a. and Contradictory – the P-factor weight was temporarily increased to 0.4, while the R-factor was decreased to 0.25;

[0108] b. Missing user profile (e.g., age not filled in) - Use the default profile ("General Adult") and mark "Incomplete profile, reduced reference value for age classification" in the report;

[0109] c. If the same indicator has already had a high-risk result within the last 7 days, this result will be "suppressed" and changed to "continued monitoring" to avoid repeatedly sending high-intensity reminders to users.

[0110] It should be noted that the risk level judgment can be changed from risk score to "decision tree model" or "lightweight logistic regression model" (which is still an interpretable model and does not introduce a black box); the classification threshold [0,0.4) / [0.4,0.7) / [0.7,1.0] can be changed to automatic optimization based on historical data ROC curve.

[0111] S6: Based on the risk level and key indicators, match the preset rule set and inject the user profile to generate a structured reference prevention plan; the strength of the reference prevention plan is dynamically adjusted according to the user's historical response behavior to the previous prevention plan.

[0112] Specifically, the purpose of this step is to generate an actionable, referential prevention plan based on risk levels and user profiles, with the output being structured entries rather than paragraphs.

[0113] Example:

[0114] Input: Risk level (low / medium / high), list of key indicators, user profile Q (age, gender, medical history, medication history, last response), dimensional rule set RULES, knowledge base KB.

[0115] The processing logic for this step is as follows:

[0116] S601: Dimensional rule set structure - each rule is a (trigger, conditions, outputs) triple; the trigger is in the form of (risk_level, indicator_code), the conditions are the conjunctive normal form of the profile field, and the outputs are 1-3 planning items {dimension, content, intensity};

[0117] S602: Rule matching – For each (risk level × key indicator) combination, iterate through RULES to find a match;

[0118] S603: User Profile Injection - Substitute field values ​​from Q into placeholders in outputs (e.g., "Age > 60 and indicator = blood sugar → Change the recommended ≥150 minutes of moderate-intensity exercise per week to ≥120 minutes of moderate-intensity exercise per week (elderly-friendly version)").

[0119] S604: Historical Response Feedback - If a user previously marked a plan item as "not adopted", the intensity of that type of plan will be automatically reduced this time, and the "not adopted" behavior will be written to the historical response field. The intensity will be reduced by one level again next time. If there are 3 consecutive "not adopted" entries, the intensity of that type of plan will be suspended for 1 month.

[0120] S605: Knowledge Base Retrieval – For each generated plan item, call KB to retrieve 1-2 bases (item summaries) as the "Source Description" field of the plan item;

[0121] S606: Deduplication and Sorting - Merge duplicate entries and sort them by "Dimension Priority = Review > Medication > Diet > Exercise > Rest and Work > Psychological > Risk Concern > Other".

[0122] Assuming the user in the example above with "high-risk blood sugar" is a 30-year-old male with no history of diabetes and who did not adopt the exercise suggestion last time, the generated 5 plan items are: ① [Follow-up] Follow up with fasting blood glucose and glycated hemoglobin within 3 months (high intensity); ② [Medication] It is recommended to consult an endocrinologist to assess whether medication intervention is needed (medium intensity); ③ [Diet] Halve the amount of staple food and replace it with whole grains (medium intensity); ④ [Exercise] Walk briskly for 90 minutes per week (already automatically reduced from 150 minutes, low intensity); ⑤ [Risk Monitoring] Seek medical attention immediately if symptoms of "three highs and one low" appear (medium intensity).

[0123] Output: A structured array of plan items, each containing (dimension, content, intensity, source description, historical response status).

[0124] Abnormal branch:

[0125] a.RULES has no matching rules - use the fallback template "It is recommended to pay attention to indicator X, maintain the current treatment plan, and have regular check-ups";

[0126] b. Missing profile – Only the trigger dimension is used; the conditions section is skipped.

[0127] c. KB search failed - fill in "internal knowledge base" as a placeholder in the source description field.

[0128] It should be noted that the reference prevention plan can also be generated through the following methods: ① The rule set RULES can be changed to be maintained by a visual editor and configured by non-technical personnel; ② The user profile can be expanded to fields such as "occupation", "work and rest time", and "exercise preference"; ③ The historical response factor can be changed to "time decay-based weighted response rate", with a higher weight for recent responses; ④ The knowledge base can be changed to a two-layer structure of "hierarchical caching + on-demand retrieval".

[0129] S7: Based on the risk level, key indicators, and reference prevention plan, generate a trend forecast report, record the initial status of the plan items in the trend forecast report, and write the user's adoption or neglect of the plan items into the feedback log for the next adjustment of the intensity of the reference prevention plan.

[0130] Specifically, the purpose of this step is to aggregate risk levels, key indicators, and planned items into a "trend prediction report" and write the user's subsequent behavior back into the profile.

[0131] Example:

[0132] Input: All outputs from steps S5-S6. Processing logic for this step:

[0133] S701: Report Structure – Rendered using a seven-part template: ① Risk Overview ② Top-N Key Indicators (including text summaries of change rates / accelerations) ③ Trend Data (line chart data + shaded reference intervals) ④ Individualized Plan Items ⑤ Historical Response Summary ⑥ Disclaimer ⑦ Report Generation Timestamp;

[0134] S702: Feedback Writing - When this report is generated, record the report_id, generation time, risk level snapshot, and the user's initial status for the planned item as "Pending Execution";

[0135] S703: Client Push - Push frequency according to risk level: "Low" once every 30 days, "Medium" once every 14 days, "High" once every 7 days;

[0136] S704: User Response Write-back - When a user clicks "Accept / Ignore / Completed" on each plan item in the client, the action is written to user_action_log and used as input for the "behavioral response factor" other than P in the next classification (the behavioral response factor does not participate in the risk_score calculation, but is only used to adjust the plan strength in step S6).

[0137] S705: Closed-loop example - The user marked the exercise suggestion as "not adopted" in the previous report. The exercise intensity was automatically reduced in the previous report. If the user does not adopt the suggestion for the third time in this report, this type of plan will be suspended for 1 month to avoid the accumulation of invalid information.

[0138] Output: Trend forecast report (structured JSON + rendered user view) + feedback log entries.

[0139] Abnormal branch:

[0140] a. Push notification failed - This will be downgraded to a pop-up notification the next time you enter the app's home screen;

[0141] b. If the report was generated less than 24 hours ago, it will be merged into the previous report and marked as "updated" instead of being created.

[0142] It should be noted that the output and feedback can be replaced by the following methods: ① The seven-zone report template can be changed to three versions: doctor version, user version, and family version, each focusing on different content; ② The push frequency can be changed to "dynamically adjusted according to risk level" (two consecutive medium-risk upgrades will result in high-frequency pushes); ③ The feedback writing can be changed to "event sourcing" mode for easier auditing and playback.

[0143] This embodiment provides a chronic disease trend prediction method based on health record time-series data. It simultaneously extracts the first-order rate of change and second-order acceleration from the health record time-series data, quantifying the "direction" and "velocity" into comparable values. A risk grading model is constructed, weighted by three factors: rate of change, current state, and abnormal persistence, outputting an interpretable risk level (low / medium / high). Based on the grading results and user profiles (including personal baseline, age group, underlying medical history, and historical responses), a personalized preventative plan is generated, and the user's actual adoption of the plan serves as input for the next grading, forming a closed loop. It should be noted that this method outputs trend indicators and preventative suggestions for reference only and does not constitute a medical diagnosis or treatment opinion. Users should interpret the results and develop treatment plans under the guidance of a professional physician.

[0144] The chronic disease trend prediction method based on health record time-series data provided in this embodiment has the following advantages compared with the prior art:

[0145] Compared to historical curves: historical curves require human estimation of slope and lack quantification. This embodiment outputs... and Two values, which can be directly written into a doctor's report, can be processed by a machine, and can be compared across users ("this user's blood sugar") "0.02 higher than the last visit" is a machine-readable statement. This embodiment also provides "acceleration" information—when the slope of the curve remains constant but the curvature changes (e.g., ...). flat but (Flipped), neither the user nor the historical graph can see it, but the method provided in this embodiment can identify it.

[0146] In contrast to single-point threshold alerts, which only alert when an indicator goes out of range, this implementation uses a combination of R+P to trigger "medium" or "high" risk even when the indicator is within the range but has deviated from the individual's baseline and continues to change—something single-point thresholds cannot do. The P factor equates "five consecutive instances of the same direction" to "one instance of severe out-of-range deviation"; while existing single-point threshold alerts treat both situations identically.

[0147] Compared to general recommendations: Existing general recommendations present the same message to everyone. This example adjusts the plan strength based on user profiles (age / gender / medication history / medication) and behavioral response factors, outputting 5 structured items with varying strengths instead of a single sentence. Each plan item includes "Source Explanation" and "Historical Response Status" fields, allowing users to trace the basis of each suggestion and their past adoption history.

[0148] In summary, this embodiment upgrades "single-point anomaly detection" to "trend deterioration detection," "general suggestions" to "personalized suggestions driven by profiling and response," and "one-time output" to "feedback loop"—these three upgrades together transform the existing "post-event reminder" tool into a predictive reference tool that supports early intervention. Through dual quantification of time-series features, three-factor weighted grading, and historical response feedback loop, it achieves pre-event trend warnings and personalized prevention suggestions earlier than the single-point threshold, solving the problems of unquantified time-series information, delayed warnings, and disconnect between prevention templates and history in the existing technology.

[0149] Example 2

[0150] This embodiment provides a chronic disease trend prediction device based on health record time-series data, including:

[0151] The data acquisition module is used to acquire the user's structured records from the health record system, sort the multiple measurements of the same indicator in the structured records by timestamp, and then process them into a clean time series after missing value processing and outlier labeling.

[0152] The first-order rate of change calculation module is used to calculate the first-order rate of change between each adjacent effective point based on the time interval between adjacent effective points in the cleaning time series, and to calculate the average first-order rate of change and its direction by averaging all the first-order rates of change.

[0153] The second-order acceleration calculation module is used to perform adjacent differences on the first-order rate of change sequence, calculate the second-order acceleration, and calculate the average of all second-order accelerations to obtain the average second-order acceleration and its direction.

[0154] The three-factor weighting module is used to calculate the rate of change factor based on the absolute value of the average first-order rate of change, calculate the current state factor based on the degree to which the current measured value deviates from the reference range, and calculate the anomaly persistence factor based on the number of consecutive changes in the same direction or consecutive out-of-range events. When the direction of the average first-order rate of change is opposite to the direction of the average second-order acceleration, the weight of the anomaly persistence factor is increased, and the weight of the rate of change factor is decreased accordingly.

[0155] The risk grading module is used to perform a weighted summation of the rate of change factor, the current state factor, and the abnormal persistence factor to obtain a risk score, and then map the risk score to an interpretable risk level.

[0156] The prevention plan generation module is used to match a preset rule set and inject user profiles based on the risk level and attention indicators to generate a structured reference prevention plan; the strength of the reference prevention plan is dynamically adjusted according to the user's historical response behavior to the previous prevention plan.

[0157] The report generation and feedback module is used to generate a trend prediction report based on the risk level, the focus indicators, and the reference prevention plan, record the initial status of the plan items in the trend prediction report, and write the user's adoption or ignoring behavior of the plan items into the feedback log for the next adjustment of the intensity of the reference prevention plan.

[0158] For details on the specific implementation of each module in a chronic disease trend prediction device based on health record time-series data, please refer to the above description of the limitations of a chronic disease trend prediction method based on health record time-series data, which will not be repeated here.

[0159] The technical features of the above embodiments can be combined in any way (as long as there is no contradiction in the combination of these technical features). For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described; these embodiments not explicitly written should also be considered to be within the scope of this specification.

Claims

1. A method for predicting chronic disease trends based on time-series health record data, characterized in that, include: Step 1: Obtain the user's structured records from the health record system, sort the multiple measurements of the same indicator in the structured records by timestamp, and then process them by handling missing values ​​and outlier marking to form a clean time series sequence; Step 2: Based on the time interval between adjacent effective points in the cleaning time series, calculate the first-order rate of change between each adjacent effective point, and average all first-order rates of change to obtain the average first-order rate of change and its direction. Step 3: Perform adjacent differences on the first-order rate of change sequence, calculate the second-order acceleration, and average all second-order accelerations to obtain the average second-order acceleration and its direction; Step 4: Calculate the rate of change factor based on the absolute value of the average first-order rate of change, calculate the current state factor based on the degree to which the current measured value deviates from the reference range, and calculate the anomaly persistence factor based on the number of consecutive changes in the same direction or consecutive out-of-range occurrences; when the direction of the average first-order rate of change is opposite to the direction of the average second-order acceleration, increase the weight of the anomaly persistence factor and correspondingly decrease the weight of the rate of change factor. Step 5: Perform a weighted summation of the rate of change factor, the current state factor, and the anomaly persistence factor to obtain a risk score, and map the risk score to an interpretable risk level; Step 6: Based on the risk level and the indicators of concern, match the preset rule set, inject the user profile, and generate a structured reference prevention plan; The strength of the reference prevention plan is dynamically adjusted based on the user's historical response behavior to the previous prevention plan; Step 7: Based on the risk level, the focus indicators, and the reference prevention plan, generate a trend prediction report, record the initial status of the plan items in the trend prediction report, and write the user's adoption or ignoring behavior of the plan items into the feedback log for the next adjustment of the intensity of the reference prevention plan.

2. The chronic disease trend prediction method based on health record time-series data according to claim 1, characterized in that, In step 1, the missing value processing is specifically as follows: if the interval between adjacent time points is less than or equal to 30 days, linear interpolation is used to fill in the missing values, and the filled data points are marked as interpolation points; if the interval between adjacent time points is greater than 180 days, the missing values ​​are retained and no interpolation is performed; the interpolation points are not used in the subsequent calculation of the first-order rate of change and the second-order acceleration.

3. The chronic disease trend prediction method based on health record time-series data according to claim 1, characterized in that, In step 1, the outlier labeling process is specifically as follows: using the individual baseline ± 3 times the standard deviation or the upper and lower limits of the reference interval multiplied by 1.5, the union of the two criteria is taken, and the measured values ​​that exceed the range are labeled as outliers; the outliers are not used in the subsequent calculation of the first-order rate of change and the second-order acceleration.

4. The method for predicting chronic disease trends based on time-series health record data according to claim 1, characterized in that, In step 2, the method for determining the direction of the average first-order rate of change is as follows: the average first-order rate of change is compared with a preset direction threshold. If it is greater than the preset direction threshold, it is determined to be an upward direction; if it is less than the preset direction threshold, it is determined to be a downward direction; otherwise, it is determined to be a stationary direction.

5. The method for predicting chronic disease trends based on time-series health record data according to claim 1, characterized in that, In step 3, the method for determining the direction of the average second-order acceleration is as follows: if the average second-order acceleration is greater than 0, it is determined to be an acceleration deviation direction; if the average second-order acceleration is less than 0, it is determined to be a deceleration direction; if the absolute value of the average second-order acceleration is less than a preset threshold, it is determined to be an insignificant acceleration.

6. The method for predicting chronic disease trends based on time-series health record data according to claim 1, characterized in that, In step 4, the calculation process of the rate of change factor is as follows: configure a severe threshold set according to the indicator type, compare the absolute value of the average first-order rate of change with the severe threshold set, and normalize it to the [0,1] interval by looking up a table.

7. The method for predicting chronic disease trends based on time-series health record data according to claim 1, characterized in that, In step 4, the calculation process of the current state factor is as follows: calculate the difference between the current measured value and the lower limit of the reference interval, divide it by the difference between the upper and lower limits of the reference interval to obtain the deviation, and normalize the deviation to the [0,1] interval; if the current measured value is within the reference interval, calculate the normalized value based on the one closer to the upper or lower limit of the reference interval.

8. The method for predicting chronic disease trends based on time-series health record data according to claim 1, characterized in that, In step 4, the calculation process of the abnormal persistence factor is as follows: count the number of consecutive changes in the same direction or consecutive out-of-range events. When the number is greater than or equal to 3, it is mapped to 1.0; when it is equal to 2, it is mapped to 0.6; when it is equal to 1, it is mapped to 0.3; and when it is equal to 0, it is mapped to 0.

9. The method for predicting chronic disease trends based on time-series health record data according to claim 1, characterized in that, In step 5, when mapping the risk score to an interpretable risk level, specifically: the interval [0, 0.4) is mapped to low risk, the interval [0.4, 0.7) is mapped to medium risk, and the interval [0.7, 1.0] is mapped to high risk.

10. A chronic disease trend prediction device based on health record time-series data, characterized in that, include: The data acquisition module is used to acquire the user's structured records from the health record system, sort the multiple measurements of the same indicator in the structured records by timestamp, and then process them into a clean time series after missing value processing and outlier labeling. The first-order rate of change calculation module is used to calculate the first-order rate of change between each adjacent effective point based on the time interval between adjacent effective points in the cleaning time series, and to calculate the average first-order rate of change and its direction by averaging all the first-order rates of change. The second-order acceleration calculation module is used to perform adjacent differences on the first-order rate of change sequence, calculate the second-order acceleration, and calculate the average of all second-order accelerations to obtain the average second-order acceleration and its direction. The three-factor weighting module is used to calculate the rate of change factor based on the absolute value of the average first-order rate of change, calculate the current state factor based on the degree to which the current measured value deviates from the reference range, and calculate the anomaly persistence factor based on the number of consecutive changes in the same direction or consecutive out-of-range events. When the direction of the average first-order rate of change is opposite to the direction of the average second-order acceleration, the weight of the anomaly persistence factor is increased, and the weight of the rate of change factor is decreased accordingly. The risk grading module is used to perform a weighted summation of the rate of change factor, the current state factor, and the abnormal persistence factor to obtain a risk score, and then map the risk score to an interpretable risk level. The prevention plan generation module is used to match a preset rule set and inject user profiles based on the risk level and attention indicators to generate a structured reference prevention plan. The strength of the reference prevention plan is dynamically adjusted based on the user's historical response behavior to the previous prevention plan; The report generation and feedback module is used to generate a trend prediction report based on the risk level, the focus indicators, and the reference prevention plan, record the initial status of the plan items in the trend prediction report, and write the user's adoption or ignoring behavior of the plan items into the feedback log for the next adjustment of the intensity of the reference prevention plan.

Images