A personalized medical service recommendation method based on user portrait

Abstract

The present application relates to the technical field of healthcare informatics, and discloses a personalized medical service recommendation method based on user portrait, comprising: dynamically constructing a chronic portrait layer and an acute context layer by calculating the occurrence frequency, survival duration and baseline sign statistical deviation of health data items, and getting rid of the dependence on static labels; checking the activation state of the acute context layer; when there are multiple activated acute items, making a clinical priority decision to select a dominant context; based on the dominant or unique acute context, preferentially recommending, and taking it as a temporary contraindication filter to review the silent conflict between chronic and secondary acute recommendations, the present application dynamically infers the clinical state and internally builds a priority decision, realizes the context awareness and dynamic suitability of the recommendation result, and avoids the clinical risks caused by state mismatch.

Description

TECHNICAL FIELD

[0001] The present application relates to a personalized medical service recommendation method based on user portrait, belonging to the technical field of medical health informatics. BACKGROUND

[0002] The current way of providing personalized medical service recommendation based on user's electronic health record or personal health record is a common technical method, which usually relies on the analysis of user health data, builds user portrait, and tends to process these data into an accumulated historical chronic portrait, such as marking the user as having a specific chronic disease or having a history of allergy, and the system matches and recommends corresponding services, such as chronic disease management plan, based on this relatively stable portrait. This method has its application value in dealing with stable chronic diseases; however, the key feature of medical health data is its instantaneity and context dependence. The above-mentioned method relying on static portrait matching does not fully consider this point in principle. When a user with a chronic portrait, such as a diabetic patient, encounters a short-term acute context, such as acute gastroenteritis, this method shows its inherent limitations. The system will continue to recommend a low-carb meal that may contain high-fiber foods based on the historical label of diabetes, which is clinically inappropriate for patients with acute gastroenteritis.

[0003] The limitation of the existing method is that it lacks a mechanism to distinguish between the user's long-term chronic state and his short-term acute context at the logical level of information processing. Moreover, even if existing technologies try to introduce dynamic monitoring, they often fail to fundamentally solve the problem due to the limitations of monitoring dimensions. For example, the Chinese invention patent with the authorization announcement number CN114283917B discloses a pre-warning analysis method and system based on chronic disease medication big data. This method builds user purchase portrait and conducts group classification to monitor the difference between the user and his group behavior to pre-warn the abnormality. The fundamental defect of this technical idea is that its pre-warning logic is completely based on the statistical rules of purchase behavior, not the actual clinical state of the user. It also cannot effectively distinguish between the user's chronic state and acute context, and may misjudge the compliant behavior of suspending purchase due to acute illness as an abnormal behavior that needs to be pre-warned, thus producing ineffective or even false guidance.

[0004] This leads to an inherent technical limitation that the recommended historical relevance and current appropriateness are difficult to balance, and when the historical portrait and the current context conflict with the clinical logic, the system has no mechanism to arbitrate or avoid risks, which is reflected in the following aspects: 1, the system lacks dynamic awareness of the timeliness of the user's health state, and tends to regard all health data as equally time-sensitive historical accumulation, and cannot identify the current acute health events that should be paid most attention to; 2, the information processing logic of the system is a static matching based on historical labels, rather than a dynamic coordination based on clinical safety, and when the clinical logic of the historical portrait and the current context is inconsistent, the system lacks an active safety review mechanism; 3, this leads to the loss of clinical appropriateness of the recommended results in a specific context, which may cause potential risks in medical safety, which is a technical problem to be solved in the field of medical information science. Therefore, how to provide a new recommendation method, which can dynamically distinguish and preferentially respond to the acute context of the user within the logical framework of information processing, while making it safe and consistent with the clinical logic of the chronic portrait, so as to ensure the clinical appropriateness and safety of the recommended results, has become a technical problem to be solved by the present application. SUMMARY

[0005] The present application provides a personalized medical service recommendation method based on user portrait, which mainly aims to solve the problem that the prior art lacks a mechanism to distinguish between long-term chronic state and short-term acute context of the user at the information processing logic level, resulting in loss of clinical appropriateness and safety of the recommended results in a specific scenario.

[0006] To achieve the above-mentioned purpose, the present application provides a personalized medical service recommendation method based on user portrait, comprising the following steps:

[0007] Step a, portrait dynamic layering step, obtaining health record data of the user, the portrait dynamic layering step is executed by the processor to perform the following operations: calculating the occurrence frequency and duration of the data items in the health record data in a preset long-term time window, when the occurrence frequency and duration both exceed the preset chronic baseline threshold, the data item is classified into the chronic portrait layer; detecting the statistical deviation of the recent low-frequency events newly recorded in the preset recent time window compared with the long-term personal mean of the user baseline sign data stream, when the recent low-frequency event or the statistical deviation exceeds the preset acute deviation threshold, the event or the abnormal sign is classified into the acute context layer;

[0008] Step b, context gating check step, checking whether there is an activated acute context item in the acute context layer;

[0009] Step c, the rule of the gate-driven coordination step, when the context gate check step checks that there are multiple activated acute context entries in the acute context layer, then triggers the clinical priority decision, based on the preset clinical priority label of the multiple acute context entries in the preset clinical knowledge base, selects the acute context entry with the highest priority as the dominant acute context; preferentially executes the preset recommendation rule matched with the dominant acute context or the unique acute context entry, and at the same time, takes the dominant acute context or the unique acute context entry as a temporary contraindication filter, reviews and silences the recommendations from the non-dominant acute context entry based on the recommendations generated by the chronic image layer; when the acute context layer is in a non-activated state, the preset recommendation rule matched with the chronic image layer is executed.

[0010] Preferably, in step a: the long-term time window is three hundred and sixty-five days; the chronic baseline threshold is: the occurrence frequency is greater than twelve times and the survival time is greater than one hundred and eighty days; the recent time window is fourteen days; the recent low-frequency event is a data item with a historical frequency lower than the chronic baseline threshold; the statistical deviation degree is the absolute value of the difference between the recent value of the baseline sign data stream and its long-term personal mean value; the acute deviation threshold is 2 times the standard deviation of the historical data corresponding to the long-term personal mean value.

[0011] Preferably, in step a, the acute context entry belonging to the acute context layer is automatically associated with an expected survival time set based on the clinical knowledge base; in step b, the activated state means that there is an acute context entry whose expected survival time has not been exceeded.

[0012] Preferably, the silencing processing in step c means using a preset acute-chronic clinical conflict knowledge base to identify and prevent recommendations from the chronic image layer that are clinically conflicting with the current dominant acute context or the unique acute context entry.

[0013] Preferably, before or during the image dynamic layering step a, it also includes: performing a term sniffing based on a preset acute keyword set on the user's preset unstructured data source; when the term sniffing matches the keywords in the acute keyword set, a to-be-checked acute context marker is automatically placed in the acute context layer, which is regarded as an acute context entry for checking by the context gate check step b.

[0014] Preferably, the rule coordination step c further comprises: for the non-conflicting chronic recommendations generated based on the chronic image layer and not silenced by the temporary contraindication filter, applying a priority dynamic attenuation process; the priority dynamic attenuation process comprises: temporarily de-prioritizing the pre-set baseline clinical priority of the non-conflicting chronic recommendations by one level, and placing the presentation priority of the non-conflicting chronic recommendations behind the recommendation rules matching the dominant acute context or the unique acute context entry.

[0015] Preferably, the rule coordination step c further comprises: for the recommendations generated based on the chronic image layer and reviewed by the temporary contraindication filter as non-conflicting, presenting the recommendations to the user in a de-prioritized or collapsed manner; monitoring the positive interaction behavior of the user on the de-prioritized or collapsed presented recommendations; when the positive interaction behavior is monitored, resolving the behavior as an implicit termination signal, and dynamically terminating the active state of the acute context entry based on the implicit termination signal, and the consequent silencing is released.

[0016] Preferably, the image dynamic stratification step a further comprises: logically binding the acute context entry with one or more pre-set baseline indicator data streams in the chronic image layer, the logical binding refers to establishing a corresponding relationship between the acute context entry and the baseline indicator data stream; the method further comprises a baseline regression monitoring step, the baseline regression monitoring step comprises: monitoring whether the bound baseline indicator data stream recovers from an acute phase disorder feature to a regular feature, the acute phase disorder feature refers to the data stopping updating or the value exceeding the pre-set personal historical mean threshold, the regular feature refers to the data recovering regular recording or the value returning to the personal historical mean threshold; when the regular feature regression is monitored, automatically terminating the active state of the acute context entry, and the countdown of the expected duration is terminated accordingly.

[0017] Preferably, the image dynamic stratification step a further comprises: automatically associating a clinical severity level label for the acute context entry in the acute context layer, the label is derived from a clinical knowledge base; and the silencing in the rule coordination step c is executed according to the clinical severity level label: when the clinical severity level label is high, performing complete silencing on all non-conflicting chronic image layer recommendations; when the clinical severity level label is low, performing silencing only on the chronic image layer recommendations determined as absolute conflict by the acute-chronic clinical conflict knowledge base and presenting the recommendations determined as non-conflicting after adding a pre-set context-aware label.

[0018] Preferably, in step c, the clinical priority label is divided into three levels of P1, P2 and P3; the clinical priority decision follows the following rules: when there is an entry of P1 level in multiple acute context entries, ; wherein, is the dominant acute context, is an acute context entry with P1 level; when there is no entry with P1 level but there is an entry with P2 level, then the entry with P2 level is selected as the .

[0019] Compared with the prior art, the beneficial effects of the present application are:

[0020] 1. In the profile dynamic stratification step, the user's health data is immediately separated into a chronic profile layer and an acute context layer, and the context gating and rule coordination mechanism makes the activated entry in the acute context layer have a dual function: it not only serves as a matching basis for the recommended rules that are executed in priority, but also is used as a temporary taboo filter that actively reviews all recommendations triggered by the chronic profile layer and performs silent processing on any recommendations that have clinical logical conflicts. This information processing method closely couples the priority of the recommendations with clinical safety, so that the system not only responds to the user's current needs, but also avoids potential risks that may be caused by state mismatching.

[0021] 2. In this method, specific clinical keywords in unstructured data sources are matched in conjunction with the profile dynamic stratification step and the context gating check. This design makes a matching signal derived from the user's immediate complaint text used as a trigger condition for activating the acute context layer. As a result, the temporary taboo review and silent processing logic in the rule coordination step driven by the gating can be executed before the generation of structured diagnoses, so that the clinical appropriateness of the recommendation service not only depends on the historical archives, but also responds to the user's immediate text state. In the profile dynamic stratification step, an activated acute context entry can be logically bound to one or more baseline indicator data streams in its chronic profile layer. The system continuously analyzes the behavior characteristics of the baseline indicator data stream, such as the frequency of data updates or the regularity of value fluctuations, through a baseline regression monitoring logic. When it is monitored that the data stream recovers from acute disorder characteristics to regular characteristics, the recovery state is used as an objective signal to automatically terminate the activation state of the acute context entry, which makes the duration of the context gating no longer dependent on the preset fixed time length, but dynamically responds to the user's individualized objective sign data stream.

[0022] 3. In the profile dynamic stratification, clinical priority attributes are set for different acute context entries. When the context gating check finds that there are multiple activated acute entries, a decision-making mechanism is executed to select the entry with the highest priority as the dominant acute context. The dominant acute context is then used as the only basis for the temporary taboo filter, which not only reviews and silences recommendations that conflict with the chronic profile layer, but also reviews and silences recommendations from other secondary acute contexts that conflict with the current dominant state, so as to maintain the consistency of the recommendation logic in multiple concurrent health states. BRIEF DESCRIPTION OF DRAWINGS

[0023] Fig. 1 Flow chart of the whole process of the dynamic stratification and gating collaborative recommendation method of the present application;

[0024] Fig. 2 Graph of the relationship between the time window and the identification rate of chronic and acute image layers of the present application;

[0025] Fig. 3 System activation state transition diagram driven by gating check of the present application. DETAILED DESCRIPTION

[0026] In order to make the purpose, technical solutions and advantages of the present application clearer and more apparent, the present application will be further described in detail below in conjunction with specific embodiments; it should be noted that the embodiments here are only used to explain the present application, and are not used to limit the protection scope of the present application.

[0027] The personalized medical service recommendation method based on user portrait provided by the present application runs in the technical environment of medical health information science, for example, is executed on the server side of a medical information system or a terminal device with corresponding data processing capability, and mainly includes the following steps in the information processing process: a portrait dynamic stratification step for distinguishing a chronic portrait layer and an acute context layer from original health record data; a context gating check step for judging whether the acute context is activated; and a rule coordination step driven by gating, which executes clinical priority decision when multiple acute contexts exist according to the result of gating check, and finally decides whether to execute acute recommendation preferentially, silent conflict chronic recommendation, or execute regular chronic recommendation.

[0028] In practical applications, the health profile data of a user acquired by the system can be sourced from a hospital's electronic health record (EHR) system, a user's personal health record (PHR) application, or a data interface of an Internet of Things wearable device, whose data structure is usually a database entry containing an ICD-10 or SNOMED CT diagnosis code, a drug generic name, a laboratory examination item, a vital sign reading, and a corresponding operation or record timestamp; the step a, the step of profiling dynamic stratification, of the present method is performed by a processor on these raw data, the task of which is to dynamically infer the clinical status of a user through the time series characteristics of data streams without relying on static classification labels; to identify the chronic profiling layer, the processor performs time series characteristic calculation on each type of data item in the database (for example, a diagnosis of type 2 diabetes or a drug of metformin), specifically, the processor performs an operation equivalent to a database query to calculate the occurrence frequency and duration of the data item in a preset long-term time window (for example, three hundred sixty-five days); the processor then compares the calculation result with a preset chronic baseline threshold, which is calibrated according to the general recognition of the field for chronic disease management, for example, a preferred threshold is calibrated to an occurrence frequency of more than twelve times and a duration of more than one hundred and eighty days, which respectively correspond to monthly follow-up visits, regular medication records or health problems lasting for half a year within a year; when a data item meets both conditions, the processor classifies it into a temporary, effective chronic profiling layer data structure in this recommendation session; at the same time, to identify the acute context layer, the processor performs two detections in parallel, the first one is to detect recent low-frequency events, that is, data items (for example, a first-time diagnosis of acute gastroenteritis) newly recorded in a preset recent time window (for example, the past fourteen days) and whose historical frequency and duration are lower than the above-mentioned chronic baseline threshold, the second one is to detect the statistical deviation of the user's baseline vital sign data stream (for example, daily blood glucose or body temperature readings), which is quantified as the absolute value of the difference between the recent value of the baseline vital sign data stream and its long-term personal mean value; the processor compares the statistical deviation with an acute deviation threshold, which can be calibrated to 2 times (i.e. 2-Sigma) of the standard deviation of the historical data corresponding to the long-term personal mean value, which is used in data analysis to identify statistically significant fluctuations representing potential acute physiological state changes; when a recent low-frequency event or a statistical deviation exceeds the acute deviation threshold, the event or abnormal vital sign (for example, acute gastroenteritis or body temperature 38.5°C) is classified into the acute context layer.

[0029] In step a, when an acute context entry is committed into the acute context layer, the clinical logic requires the state to be time-bounded, for this purpose, the system processor automatically queries a pre-set clinical knowledge base; the knowledge base is a decision support tool, for example, a key-value storage database, which maps ICD-10 codes or SNOMED CT identifiers to specific clinical parameters, including an expected time-to-live (TTL) set based on clinical practice, for example, {ICD:K29.7:{TTL_days:7}, ICD:J06.9:{TTL_days:5}}; the processor automatically associates the retrieved TTL value (e.g., 7 days) to the entry and stores it into the acute context layer; further, step b, the context gate check step is a logical check that must be invoked before any recommendation matching is performed, the processor iterates through the acute context layer and checks whether there exists an acute context entry that has not exceeded its expected time-to-live, the check logic is equivalent to As long as there exists at least one entry that satisfies this condition, the acute context layer is determined to be in an active state, thus triggering the prioritized coordination logic in step c.

[0030] Step c, the rule of the gated driving synchronization step is the main body of the method to achieve clinical safety; if the check result of step b is the non-activated state (i.e. the acute context layer is empty, or the TTL of all entries has expired), the gate is opened, the processor restores the standard process, and executes the preset recommended rule (for example, recommend routine reexamination of diabetes) matched with the chronic image layer; otherwise, if the check result is the activated state, the processor first checks the number of activated entries; when it is found that there are multiple activated acute context entries, for example, P2: knee arthroplasty and P1: acute upper respiratory tract infection with high fever exist at the same time, the system immediately triggers the clinical priority decision, which queries the aforementioned clinical knowledge base to obtain the preset clinical priority label for each entry. These labels are divided into P1 (life-threatening or systemic risk), P2 (local or functional problem), P3 (mild symptoms) levels according to the safety principle; the decision logic follows a deterministic selection rule, for example: when there are P1 level entries in multiple acute context entries, the entry with P1 level is selected as the dominant acute context; when there are no P1 level entries but P2 level entries, the P2 level entry is selected as the dominant acute context, and so on, so as to select a unique dominant acute context (in this example, P1: high fever); then, the processor preferentially executes the recommended rule (for example, recommends rest and water supplementation) matched with the dominant acute context (or the only acute context entry in the single acute scenario); at the same time, the processor takes the dominant acute context (i.e. high fever) as the key of the temporary contraindication filter to query a preset acute-chronic clinical conflict knowledge base, which is a safety rule set (for example, a conflict matrix) for identifying and preventing clinically conflicting recommendations; the processor uses this filter to review all recommendations generated based on the chronic image layer (for example, diabetes exercise program) and all recommendations from non-dominant acute context entries (for example, knee rehabilitation training); if the conflict knowledge base returns a conflict (for example, ), the conflicting recommendations are executed in silence, i.e. removed from the final recommendation list or prevented from being generated.

[0031] To further enhance the adaptability of the scenario, the method can also include several preferred embodiments; one way is for the processing of unstructured data, before or in step a, the system can be configured to perform a term sniffing step, which searches the user's recent unstructured data sources (such as health diary) for a preset acute keyword set (a list containing words such as diarrhea, fever, severe pain, etc.) through high-speed string matching or regular expressions; when the keywords are matched, the system does not make a diagnosis, but immediately automatically places a pending acute context label in the acute context layer, for example, a general symptom label with a short TTL (such as 24 hours) and a medium priority (P2), which is considered an acute context item, thus ensuring that the gating check of step b is activated, and the temporary contraindication filter of step c is started to respond to the immediate text complaint; another preferred embodiment is to grade the silent processing; in this way, in step a, while associating TTL, the system also automatically associates a clinical severity level label (for example, high severity level, low severity level) for the acute context item from the clinical knowledge base; accordingly, the silent processing in step c is configured to be executed according to the clinical severity level label: when the level is high severity level (such as acute myocardial infarction), the system performs complete silence, i.e. silences all non-conflicting chronic recommendations to avoid any information interference; when the level is low severity level (such as mild ankle sprain), the system only performs silent processing on recommendations that are determined by the acute-chronic clinical conflict knowledge base to be absolutely conflicting (such as running suggestion), while for non-conflicting recommendations (such as diabetic diet), the system recommends them after attaching a preset context-aware label, the content of which is, for example: reminder: please pay attention to your ankle condition when performing this item; another preferred embodiment is to address the recommendation fatigue problem, in step c, for those non-conflicting chronic recommendations that pass the temporary contraindication filter review, the system applies a priority dynamic attenuation process to them before output; this process queries the baseline clinical priority (for example, high, medium, low) preset for each rule in the recommendation rule library and temporarily degrades it by one level (for example, from medium to low); finally, when presented to the user, the presentation priority of these degraded non-conflicting chronic recommendations is placed after the acute recommendations of high priority, for example, they are displayed in a long-term health suggestion column.

[0032] In addition, in order to solve the problem of static rigidity of expected time to live (TTL), the method also provides a dynamic termination mechanism; the first is based on implicit feedback, which requires that the chronic recommendations that are dynamically attenuated in priority in step c must be presented in the form of weight reduction or folding; the system monitors the positive interaction behavior of the user to the weight-reduced or folded chronic recommendations in the background (for example, the user actively clicks to expand the folded menu of long-term health recommendations); when this behavior is monitored, the click event is parsed as an implicit termination signal, and the processor immediately dynamically terminates the active state of the acute context item (for example, sets its TTL to expired), and the silent processing is lifted; the second is based on objective signs, which further includes in step a logically binding the acute context item (such as influenza) to one or more baseline indicators (such as blood glucose data stream) in the chronic image layer, and the binding relationship comes from the clinical knowledge base; the system then starts the baseline regression monitoring step, which monitors whether the blood glucose data stream recovers from the acute period disorder feature (defined as data stops updating or numerical value exceeds the preset personal historical mean threshold) to the regular feature (defined as data resumes regular recording or numerical value returns to the personal historical mean threshold); when the regular feature is recovered, the recovery signal is considered as an objective clinical recovery signal, and the system automatically terminates the active state of the acute context item, at which time the counting of the expected survival time is also terminated; it should be noted that the health record data processed by the method are all considered as personal information; before performing step a to obtain data, the system has obtained data access permission through the standard user authorization protocol; during processing, all data are preferentially processed in the user-authorized secure server environment, or locally executed on the user terminal if possible (such as the term sniffing step), to reduce the transmission of raw data; the final result output by the method is the recommended service list or risk prompt (such as the context-aware label), rather than any form of medical diagnosis conclusion, and all recommendations are intended to assist users in health management and prompt users to seek professional medical advice.

[0033] In one (medical informatics) application scenario, the health record data of a user is acquired and processed by the system, the method is executed according to the steps; the system processor first executes step a, the portrait dynamic layering step, analyzes the data through the time sequence feature calculation; the processor traverses the health record data of the user, identifies that the occurrence frequency of the 2 type diabetes related diagnosis and medication record in the past 365 days is 15 times, and the duration is three years, which all exceed the preset chronic baseline threshold (occurrence frequency > 12 times, duration > 180 days), so the 2 type diabetes is classified into the chronic portrait layer; the processor also detects a newly recorded knee arthroscopy postoperative diagnosis 2 days ago in the preset recent time window (the past 14 days), the historical frequency of the diagnosis is lower than the chronic baseline threshold, and is identified as a recent low frequency event; and the processor also detects an acute upper respiratory tract infection diagnosis recorded 1 day ago, and synchronously acquires the baseline sign data stream of the user, displays that the recent body temperature value of the user is 38.8℃, and the long-term personal average value is 36.7℃, and the statistical deviation (2.1℃) exceeds the acute deviation threshold (2 times of the standard deviation of the long-term average value); therefore, the knee arthroscopy postoperative and the acute upper respiratory tract infection with high fever are all classified into the acute context layer; in step a, the processor automatically queries the preset clinical knowledge base, associates a clinical priority label P2 (local functional problem) and an expected duration of 14 days for the knee arthroscopy postoperative item, and associates a clinical priority label P1 (systemic risk) and an expected duration of 5 days for the acute upper respiratory tract infection with high fever item; then, the system executes step b, the context gating check step, the processor detects that there are two acute context items in the acute context layer which have not exceeded their expected duration, so it is determined that the acute context layer is in the active state; further, the system executes step c, the rule coordination step driven by the gate, since it is detected that there are multiple activated acute context items, the processor immediately triggers the clinical priority decision subprogram; the decision logic compares the clinical priority labels (P1 and P2) of the two items based on the preset rule (P1 level is prior to P2 level), and selects the acute upper respiratory tract infection with high fever (P1) item with the highest priority as the only dominant acute context.

[0034] After determining the dominant acute context, the processor proceeds to execute the collaborative logic of step c; first, the pre-set recommended rules matching the dominant acute context (P1 : high fever) are executed preferentially, for example, generating online respiratory department referral consultation and light diet and electrolyte water supplement recommendations; at the same time, the processor takes the dominant acute context (P1 : high fever) as a temporary contraindication filter to review all other possible generated recommendations; at this time, the system should have generated a recommendation for a diabetes routine exercise program based on the chronic profile layer (type 2 diabetes) and a recommendation for knee joint rehabilitation training based on a non-dominant acute context item (P2: postoperative); the temporary contraindication filter determines that both the diabetes routine exercise program and the knee joint rehabilitation training are in clinical conflict with the high fever state (clinical requirement for rest) by querying the acute-chronic clinical conflict knowledge base; therefore, the processor executes a silent treatment on the two conflicting recommendations; finally, the user only receives online respiratory department referral consultation and light diet and electrolyte water supplement recommendations matching the current dominant acute context (high fever) in this recommendation session, and all recommendations (from the chronic profile layer and other acute context layers) that are not suitable or conflicting in the current clinical situation are effectively prevented.

[0035] Example 2: To objectively verify the technical effect of the method of the present application in complex scenarios compared with the static image matching method, the present comparative test is set up; the test platform is based on a server-side simulation environment, which has a user database containing 1000 anonymized and synthesized health record data, which is constructed into typical scenarios of common chronic diseases and acute events in the field; at the same time, the test platform presets a preset clinical knowledge base and an acute-chronic clinical conflict knowledge base, which defines the clinical conflict relationship between a specific acute context (such as acute gastroenteritis) and a specific chronic disease recommendation (such as a high-fiber diet) according to clinical guidelines in advance; the test sets up two treatment groups: the control group, which uses the static image matching method of the prior art, which only processes the user's chronic image layer data and does not have the ability to identify the acute context layer, and does not have the context gating check step or the rule coordination step driven by the gate; the sample group of the present application, which fully deploys the method of the present application as in the detailed description, including step a, the image dynamic layering step (including time sequence feature calculation, term sniffing, hierarchical and priority label association), step b, the context gating check step (including TTL check), and step c, the rule coordination step driven by the gate (including clinical priority decision, temporary contraindication filter, silent processing, priority dynamic attenuation, and hierarchical execution logic); the evaluation index of the test is set as the inappropriate recommendation incidence, which is defined as: when there is an active acute context entry in the user's health record data, if the recommendation result generated by the recommendation system is clinically conflicted with the entry marked in the acute-chronic clinical conflict knowledge base, it is counted as an inappropriate recommendation; in the test, 1000 user record data are input into the recommendation method of the control group and the sample group of the present application respectively, the generated recommendation results of the two groups are recorded and compared, and the number of inappropriate recommendations is counted according to the above evaluation index; 8 representative typical scenarios are selected from 1000 scenarios, and the processing process and results are compared as shown in Table 1.

[0036] Table 1: Comparison of processing logic and results of the method of the present application and the control group method in typical scenarios.

[0037]

[0038] Analyzing the data shown in Table 1 (see Table 1) and the statistical results of all 1000 scenarios, the control group (static mode) lacks recognition and gating mechanism for acute context layer, and in typical conflict scenarios such as Case-001, Case-002, etc., it generates inappropriate recommendations that conflict with the user's current acute state based on the chronic image layer. In the total test of 1000 samples, the incidence of inappropriate recommendations is 15.2%; the sample group of the present application (gating mode) correctly identifies and activates the acute context layer in all the above conflict scenarios through the cooperation of steps a and b, and prevents the generation of inappropriate recommendations through the temporary taboo filter and silent processing mechanism of step c. In Case-002, the clinical priority decision mechanism further handles the acute-acute conflict, and in Case-003, the judgment of non-activated state or expected duration of expiration also enables the method to normally restore the chronic management recommendation when there is no acute context or the acute context has recovered. In the total test of 1000 samples, the incidence of inappropriate recommendations of the sample group of the present application is 0%.

[0039] Embodiment 3: This embodiment combines Figs. 1 to 3 a user image-based personalized medical service recommendation method, as shown in Fig. 1 , which acquires user health record data (EHR, PHR, wearable device data, etc.) from the beginning, and then performs step a: image dynamic layering, dynamically constructs a chronic image layer such as diabetes, hypertension and an acute context layer such as acute gastroenteritis, postoperative, through time series feature calculation such as occurrence frequency, duration, and statistical deviation, and then enters step b: context gating check to determine whether the acute context layer is activated. If it is in a non-activated state, it will execute the chronic image recommendation, that is, execute the regular chronic management rules, and generate a recommendation list and end, if it is in an activated state, it will execute the clinical priority decision when there are multiple acute entries, which provides the clinical priority and the expected duration of time from the clinical knowledge base to select a unique dominant acute context, and then execute step c: rule coordination driven by gating, which takes the dominant / unique acute context as a temporary taboo filter to review all recommendations. This step refers to the acute-chronic clinical conflict knowledge base to define acute-chronic clinical conflict rules and review recommendations from the chronic image layer (as shown by the dashed line). The final output of step c is integrated into the recommendation list, and the context-aware, dynamically appropriate and safe recommendation result (example: priority acute recommendation, silent conflict chronic recommendation) is output, and the process ends.

[0040] As Fig. 2As shown, the chart demonstrates the relationship between the time window length X-axis, from day 1 to day 365, and the recognition rate % Y-axis, where the chronic profile recognition rate (solid line) significantly increases with the extension of the time window, growing from about 10% on day 1 to over 90% on day 180, and approaching 100% on day 365, while the acute contextual recognition rate (dashed line) shows the opposite trend, with the highest recognition rate of 95% on day 1, gradually decreasing with the extension of the time window, and dropping to about 50% on day 365. As Fig. 3 As shown, the flow enters a loop state after receiving the data-triggered contextual gating check, and proceeds through the judgment of receiving new data / timing check. If the judgment result is acute context layer = inactive, the system enters the inactive state, executes the preset recommended rules matched with the chronic profile layer, and returns to the check loop. If the judgment result is acute context layer = active, the system enters the acute context active state, where it sequentially executes 1. performs clinical priority arbitration, 2. determines the dominant acute context, 3. prioritizes acute recommendations, 4. applies temporary contraindication filters, and 5. silences conflicting chronic recommendations. This active state will continue until the exit conditions are met, i.e., the occurrence of events: expected duration of stay expired OR baseline regression monitoring meets the standard OR receiving implicit termination signal, at which point the system will return to the inactive state.

[0041] In a specific (medical informatics) deployment scenario, the personalized medical service recommendation method of the present application needs to be applied to a hospital's diabetes chronic disease management information system, which has accumulated a large amount of (order of magnitude more than 500 patients, at least two years for each patient) anonymized historical electronic health record (EHR) data, including structured diagnosis codes (ICD-10), medication records, laboratory test results (such as glycosylated hemoglobin), vital signs (such as body temperature, blood glucose monitoring data) and corresponding time stamps; the engineering problem faced is that the chronic baseline threshold (a general setting of occurrence frequency > 12 times / year, duration > 180 days) and acute deviation threshold (a general setting of 2 times standard deviation) for step a, the dynamic stratification step mentioned in the specific implementation, and the determination threshold of acute period disorder features and regular features for baseline regression monitoring step, if directly using the general setting, may not achieve the expected discrimination effect on the health data pattern of this specific patient group, and a operable calibration procedure is needed to determine the parameter value; To solve this problem, an offline parameter calibration process based on historical data is performed before the system officially enables the recommendation function; The input of this process is the aforementioned anonymized EHR data set, the object of action is the time series features of each record in the data set, and the execution environment is configured as a server with database access and statistical analysis functions (capable of executing SQL queries and running Python statistical libraries); First, for the calibration of the chronic baseline threshold, the processor extracts all items labeled as chronic disease related (such as ICD-10 code E11 starting diagnosis, metformin medication record) and acute events (such as ICD-10 code J06.9 acute upper respiratory tract infection, K29.7 acute gastroenteritis) from the data set; For each type of item, calculate its actual occurrence frequency (annualized) and duration (time span from first record to last record) in different patient records; Using these labeled data, use statistical classification methods (such as constructing a receiver operating characteristic curve, ROC curve) to find a combination threshold of occurrence frequency and duration that can distinguish chronic items from acute items according to a specific statistical indicator (such as maximizing the Youden index); In this specific hospital data set analysis, the chronic baseline threshold obtained is determined to be greater than 10 times / year and greater than 150 days, and this set of thresholds is subsequently configured into the execution logic of step a.

[0042] Secondly, for the calibration of the acute deviation threshold and the baseline regression monitoring related threshold, the processor focuses on the baseline sign data stream, such as blood glucose monitoring data; For each patient, first identify and exclude the data segment in its historical record that overlaps in time with known acute events (such as infection, surgery), and only keep the blood glucose readings representing its chronic stable period; Based on these stable period data, calculate the long-term personal mean and historical data standard deviation of each patient's blood glucose value ; then, analyze the distribution of statistical deviation of blood glucose values from the patient's long-term personal mean during the period of known acute events (e.g. the time period with J06.9 diagnosis recorded and body temperature record elevated > 38℃); by comparing the deviation distribution between acute and stable periods (e.g. using Kolmogorov-Smirnov test), determine an acute deviation threshold that can distinguish the two states, so that in the dataset, blood glucose values deviating from the personal mean more than 2.5 times the historical data standard deviation are found to be more associated with acute events, so the acute deviation threshold is set to ; at the same time, analyze the time course of blood glucose data from disorder (e.g. 3 consecutive readings exceeding the range, or data record interruption for more than 48 hours) to normal (e.g. 5 consecutive readings returning to the range, and daily record recovery); the processor calculates the statistical distribution (e.g. median, quartile) of blood glucose recovery time corresponding to different types of acute events (e.g. infection, stress), which is used to set the specific numerical conditions for determining the return of normal characteristics in the baseline regression monitoring step (e.g. 5 consecutive readings within the ), on the other hand, the typical recovery time (e.g. for J06.9 infection, the median blood glucose recovery time is 6 days) is used to set or update the expected time to live (TTL) of the corresponding acute context item in the preset clinical knowledge base; by performing the above standardized calibration procedures based on historical data, the key thresholds used in the specific deployment of the method (chronic baseline threshold, acute deviation threshold, baseline regression monitoring threshold) and part of the parameters in the clinical knowledge base (e.g. TTL) are determined through an objective and reproducible process; in this way, the dynamic hierarchical logic of step a, the gating activation judgment of step b and the automatic termination mechanism of baseline regression monitoring can adapt to the actual mode of user data in specific application scenarios, thereby improving the adaptability of the method in information processing.

[0043] In the application of the method of the present application to the deployment of a (medical informatics) system, the core logic thereof relies on the accuracy and timeliness of preset clinical knowledge bases (containing expected duration of stay TTL, clinical priority label, clinical severity level label information) and acute-chronic clinical conflict knowledge bases; in order to clarify the content sources of the two knowledge bases, standardize the construction process and enable them to reflect the current clinical consensus, the system adopts a set of standardized knowledge base construction and maintenance procedures before the first deployment and in subsequent operation; the execution environment of the procedure is configured as a server with data processing capability, the input is standard resources obtained publicly, and the output is structured knowledge base data files; in the initial construction stage of the knowledge base, first extract the basic information from standard resources, these resources include international disease classification codes (such as ICD-10), systematic medical terminology sets (such as SNOMED CT), publicly released clinical practice guidelines, and drug instructions and drug interaction databases; for the preset clinical knowledge base, the processor associates specific disease codes or clinical states (such as acute gastroenteritis corresponding to K29.7) with their typical clinical characteristics according to these resources, extracts or determines their expected duration of stay (such as 7 days), clinical priority label (such as P2) and clinical severity level label (such as moderate) through clinical personnel evaluation; for the acute-chronic clinical conflict knowledge base, the processor analyzes the contraindications or caution recommendations for specific acute conditions in the clinical guidelines and drug databases, formalizes these information into conflict rules, for example, if the guidelines indicate that a high-fiber diet should be avoided during acute gastroenteritis (K29.7), a rule is generated: IF AcuteContext = K29.7 AND ChronicRecommendation.DietType = HighFiber THEN Conflict = TRUE; all extracted and formalized knowledge entries and conflict rules are cross-checked and confirmed by a review team composed of at least two clinical personnel with experience in the relevant field before being stored in the knowledge base.

[0044] To cope with the update of medical knowledge, the procedure further comprises a mechanism for time-effectiveness guarantee and reconstruction of the knowledge base; the system sets a fixed review period (e.g. every 6 months), or is triggered when a version update of the core clinical guidelines or drug database relied upon is detected; upon review or update trigger, the processor automatically or with human assistance re-executes the above process of extracting information from standard resources, identifies the newly added clinical states, updated treatment recommendations or new drug conflict information; the review team evaluates these changes and accordingly revises the existing entries in the knowledge base (e.g. adjusts the TTL or priority of a certain disease), supplements new entries, or discards obsolete rules; the updated version of the knowledge base is only officially deployed to replace the old version after regression testing (using historical case data to verify its logical consistency); this maintenance mechanism enables the core clinical knowledge and safety rules relied upon by the method of the present application to keep pace with the progress in the field.

[0045] Example 6: To ensure that the definition, matching and priority processing of the preset recommendation rules in the method of the present application have clear operating procedures, so as to guarantee the reproducibility and logical consistency of the recommendation results, the system adopts a standardized rule base construction and execution process when deployed; the input of the process is the user's chronic profile layer and acute context layer data, as well as the clinical knowledge base and acute-chronic clinical conflict knowledge base, the execution environment is a server processor configured with a rule engine, and the output is a sorted list of recommended services; in the construction stage of the rule base, each preset recommendation rule is structured as a data record containing the following fields: rule ID (unique identifier), target profile layer (specifying whether the rule applies to the chronic profile layer or the acute context layer), trigger condition set (one or more standard codes such as ICD-10 or SNOMED CT representing the profile or context entries required to trigger the rule), recommended action (describing the specific service or information content recommended), and baseline clinical priority (a numerical value, e.g. 1-100, or a classification label such as high, medium, low); the matching logic of the rule is defined as follows: after the system determines the target profile layer to be processed according to steps b and c (e.g. acute rules are prioritized when the acute context layer is activated, and chronic rules are processed when the acute context layer is not activated), the processor traverses the rules corresponding to the target profile layer in the rule base; for each rule, the processor checks whether all the codes in its trigger condition set exist in the user's current corresponding profile layer data (i.e. the dominant acute context / unique entry in the chronic profile layer or acute context layer); if the condition is met, the rule is activated, and its recommended action and baseline clinical priority are passed to the next step of processing.

[0046] The priority processing and sorting phase is performed according to the logic of step c; if the acute context layer is in the inactive state, all activated chronic rules are directly arranged in descending order of their baseline clinical priority values to generate the final recommendation list; if the acute context layer is in the active state, the temporary contraindication filter is first applied to review all activated rules (including rules from the chronic layer and non-dominant acute context, if applicable); rules determined to be in conflict are directly silenced (removed from the sorted list); for non-conflicting chronic recommendation rules from the chronic profile layer that pass the review, the processor performs priority dynamic attenuation processing on them, which can be quantified as multiplying their baseline clinical priority values by a preset attenuation factor (e.g. 0.5), or reducing their classification label by one level, e.g. from medium priority to low priority, corresponding to a value from 60 to 30; then, the processor aggregates all rules that have not been silenced (including acute rules that are given priority, whose priority remains unchanged; and chronic rules that have undergone priority attenuation processing), and arranges them in descending order of their current (or attenuated) priority values; if there are rules with the same priority value, their relative positions can be determined according to the rule ID or other preset order; the final sorting result is the recommendation list presented to the user, with acute recommendations first and non-conflicting but downgraded chronic recommendations second or displayed in a collapsed manner.

[0047] To further illustrate the essential differences and technical effects of the method of the present application relative to the prior art from the specific technical implementation logic level, the following comparative examples are set.

[0048] Comparative Example 1: This comparative example aims to verify the technical effect of a conventional recommendation method that only uses background technology to perform static matching based on historical labels when dealing with concurrent health states with clinical logic conflicts similar to Example 1 of the present application; in order to ensure fairness and objectivity of the comparison, this comparative example uses the same anonymized user health record data as input as Example 1, which clearly indicates that there is a long-term stable chronic profile (type 2 diabetes) and two newly occurring acute contexts (knee arthroscopy and acute upper respiratory tract infection with high fever) in the user profile, in addition, the basic conditions such as the recommendation rule base and clinical knowledge base used in this comparative example are also completely consistent with the configuration in Example 1; the processor of the conventional recommendation method used in this comparative example is set to: traverse all health labels in the user profile (including type 2 diabetes, knee arthroscopy, and acute upper respiratory tract infection with high fever), and independently and in parallel match all preset recommendation rules corresponding to each label. This method does not include the profile dynamic layering step, context gating check step, and rule coordination step driven by gating that are unique to the present application at the logical level of information processing, so it lacks the mechanism to identify acute contexts, prioritize, and review and silence other recommendations as temporary contraindications. By executing the above conventional recommendation method, the processor processes the same health record data input, and the processing process and output recommendation results are shown in Table 2 below.

[0049] Table 2: Results table of conventional static matching method processing concurrent conflict state.

[0050]

[0051] The test results show that the conventional static profile matching method, due to the fact that it cannot identify the time effectiveness difference and clinical priority of the user's health state in its core processing logic, will output the two recommendations of the diabetes regular exercise program and knee rehabilitation training that directly conflict with the acute upper respiratory tract infection with high fever (clinical requirement for rest) state that the user should prioritize, together with appropriate recommendations. If this result is adopted by the user, it may exacerbate the symptoms of the systemic infection and affect postoperative recovery, thereby causing potential clinical risks, and fails to achieve dynamic appropriate and safe personalized recommendations. The root cause of this technical defect lies in the lack of a mechanism that can dynamically distinguish between acute and chronic states and coordinate the processing of recommendations with built-in clinical safety logic.

[0052] It is obvious to those skilled in the art that the present application is not limited to the details of the above exemplary embodiments, and can be implemented in other specific forms without departing from the spirit or essential characteristics of the present application.

[0053] Finally, it should be noted that the above examples are merely intended to illustrate the technical solutions of the present application and not to limit the present application. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or equivalently replaced without departing from the spirit and scope of the present application.

Claims

1. A method for recommending personalized medical services based on user profiles, characterized in that, Includes the following steps: Step a, the dynamic stratification step of the profile, obtains the user's health record data, and performs the following operations through the processor: calculates the frequency of occurrence and duration of data items in the health record data within a preset long-term time window. When both the frequency of occurrence and duration of duration exceed the preset chronic baseline threshold, the health record data item is classified into the chronic profile layer; detects the statistical deviation of recently recorded low-frequency events within a preset recent time window from the user's baseline vital signs data stream compared to its long-term personal mean. When a recent low-frequency event is detected or the statistical deviation exceeds the preset acute deviation threshold, the low-frequency event or abnormal vital signs is classified into the acute context layer. Step b, context gating check step, checks whether there are active acute context entries in the acute context layer; Step c, the gate-driven rule coordination step: when the context gating check step detects that there are multiple active acute context entries in the acute context layer, the clinical priority decision is triggered. Based on the preset clinical priority labels of multiple acute context entries in the preset clinical knowledge base, the acute context entry with the highest priority is selected as the dominant acute context. Pre-defined recommendation rules that match the dominant acute context or the only acute context entry are prioritized for execution. At the same time, the dominant acute context or the only acute context entry is used as a temporary taboo filter to review and silence recommendations generated based on the chronic profile layer and recommendations from non-dominant acute context entries. When the acute context layer is inactive, the pre-defined recommendation rules that match the chronic profile layer are executed. And, in step a: recent low-frequency events are data items whose historical frequency is lower than the chronic baseline threshold; In step a, acute context entries included in the acute context layer are automatically associated with an expected duration set based on the clinical knowledge base; in step b, the active state refers to the existence of acute context entries that have not yet exceeded their expected duration; in step c, the inactive state refers to the acute context layer being empty, or the expected duration of all entries having expired.

2. The personalized medical service recommendation method based on user profile according to claim 1, characterized in that, In step a: the long-term time window is 365 days; the chronic baseline threshold is: the frequency of occurrence is greater than 12 times and the duration is greater than 180 days; the recent time window is 14 days; the statistical deviation is the absolute value of the difference between the recent value of the baseline vital signs data stream and its long-term individual mean; the acute deviation threshold is twice the standard deviation of the historical data corresponding to the long-term individual mean.

3. The personalized medical service recommendation method based on user profiles according to claim 1, characterized in that, The silencing process in step c refers to using a pre-defined acute-chronic clinical conflict knowledge base to identify and block recommendations from the chronic profile layer that clinically conflict with the current dominant acute context or the only acute context entry.

4. The personalized medical service recommendation method based on user profile according to claim 1, characterized in that, Before or during the dynamic layering step a of the profile, the method further includes: performing a term sniffing based on a preset acute keyword set on the user-preset unstructured data source; when the term sniffing matches a keyword in the acute keyword set, an acute context tag to be checked is automatically placed in the acute context layer, and the acute context tag to be checked is regarded as an acute context entry for the context gating inspection step b to be checked.

5. The personalized medical service recommendation method based on user profile according to claim 1, characterized in that, The gated rule collaboration step c further includes: applying a dynamic priority decay process to non-conflicting chronic recommendations generated based on the chronic profile layer and not silently processed by the temporary contraindication filter; the dynamic priority decay process includes: temporarily downgrading the preset baseline clinical priority of the non-conflicting chronic recommendations by one level, and placing the presentation priority of the non-conflicting chronic recommendations after the recommendation rule that matches the dominant acute context or the only acute context entry.

6. The personalized medical service recommendation method based on user profile according to claim 1, characterized in that, The rule coordination step c driven by gating further includes: presenting recommendations generated based on the chronic profile layer and deemed non-conflicting by the temporary taboo filter to the user in a downweighted or collapsed manner; monitoring the user's positive interaction behavior with the downweighted or collapsed recommendations; when positive interaction behavior is detected, resolving the positive interaction behavior as an implicit termination signal, and dynamically terminating the activation state of the acute context entry based on the implicit termination signal, thereby lifting the silent processing.

7. The personalized medical service recommendation method based on user profile according to claim 1, characterized in that, The dynamic stratification step a further includes: logically binding acute context entries with one or more preset baseline indicator data streams in the chronic profile layer. Logical binding refers to establishing a correspondence between acute context entries and baseline indicator data streams. The method also includes a baseline regression monitoring step, which includes: monitoring whether the bound baseline indicator data stream recovers from acute phase disordered features to normal features. Acute phase disordered features refer to data ceasing to update or values ​​exceeding a preset personal historical average threshold. Normal features refer to data recovering to normal records or values ​​returning to within the personal historical average threshold. When normal feature regression is detected, the activation state of the acute context entry is automatically terminated, and the timing of the expected duration is terminated accordingly.

8. The personalized medical service recommendation method based on user profile according to claim 1, characterized in that, The dynamic hierarchical modeling step a further includes: automatically associating an acute context entry in the acute context layer with a clinical severity level label derived from the clinical knowledge base; and the silent processing in the gated rule collaboration step c is performed hierarchically based on the clinical severity level label: when the clinical severity level label is high severity, all non-conflicting chronic modeling layer recommendations are completely silent; when the clinical severity level label is low severity, silent processing is only performed on chronic modeling layer recommendations that are determined to be absolutely conflicting by the acute-chronic clinical conflict knowledge base, and recommendations for chronic modeling layer recommendations that are determined to be non-conflicting are made after adding a preset context-aware label.

9. The personalized medical service recommendation method based on user profile according to claim 1, characterized in that, In step c, clinical priority labels are divided into three levels: P1, P2, and P3; clinical priority determination follows the following rule: when multiple acute context entries contain entries at the P1 level, ;in, For the dominant acute context, An acute context entry with a P1 level is selected; if no P1 level entry exists but a P2 level entry does, then the P2 level entry is chosen. .

Images