Artificial intelligence-based patient visit path planning system

Abstract

This invention relates to the field of medical intelligent technology, specifically to an artificial intelligence-based patient treatment path planning system, comprising: state inversion, etiology refinement, department matching, path planning, and conflict resolution modules. The state inversion module acquires the patient's initial state vector and, through multiple rounds of reverse traversal of the clinical decision tree, infers all prior pathological states and forms a set; the etiology refinement module removes duplicates from the set and merges them to generate a set of potential etiology nodes; the department matching module matches this set with a departmental capability knowledge graph to calculate the transfer cost; the path planning module invokes a reinforcement learning engine to generate a recommended path, using transfer cost as feedback and minimizing treatment time; and the conflict resolution module rearranges time overlap checks and generates a planning table. This system can comprehensively uncover etiologies, accurately match departments, optimize the treatment process, avoid overlapping examinations, and improve treatment efficiency.

Description

Technical Field

[0001] This invention relates to the field of medical intelligent technology, and in particular to a patient visit path planning system based on artificial intelligence. Background Technology

[0002] Current patient care pathway planning primarily relies on manual diagnosis or simple rule-based matching logic. Most systems directly assign the initial consultation department based on the patient's initial symptoms, which then guides the patient to subsequent relevant departments based on preliminary examination results. Some auxiliary planning systems introduce clinical decision tree technology, employing a forward traversal approach to deduce possible causes and related departments from the patient's initial symptoms. This assists in the initial planning of the care pathway, providing patients with basic guidance for their medical visits.

[0003] Current technologies cannot fully uncover the underlying pathological conditions that trigger a patient's current symptoms; they can only identify causes directly related to the current symptoms. This can easily lead to omissions or misjudgments of causes, resulting in errors in department matching. Furthermore, the pathway planning process does not consider the treatment capabilities of each department to calculate transfer costs, nor does it prioritize optimizing consultation time. It simply arranges departments in sequence, which can easily lead to overlapping times for cross-departmental examinations and redundant pathways, resulting in cumbersome patient procedures and excessively long waiting times. It is necessary to address the problems of incomplete cause identification, inaccurate department matching, and low consultation efficiency caused by unoptimized pathways in current technologies to achieve more accurate and efficient consultation pathway planning. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing an artificial intelligence-based patient visit path planning system.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: an artificial intelligence-based patient visit path planning system, comprising: The state inversion module obtains the patient's initial state vector and, based on the patient's initial state vector, initiates a multi-round state inversion process. By traversing the preset clinical decision tree in reverse, it deduces all the preceding pathological states that lead to the current symptoms, forming a set of preceding pathological states. The etiology refinement module removes duplicates and merges the set of prior pathological states to generate a set of potential etiology nodes. The department matching module matches the potential etiology node set with a pre-set department capability knowledge graph. The department capability knowledge graph defines the treatment authority and priority level of each clinical department for specific diseases or symptoms. Based on the matching results, the module calculates the transfer cost from the patient's initial state vector to each target department. The path planning module calls a path planning engine based on reinforcement learning. The path planning engine uses the transfer cost as environmental feedback and aims to minimize the total consultation time to perform positive reasoning calculations and generate a recommended path sequence from the initial consultation department to the final diagnosis department. The conflict resolution module performs conflict detection on the recommended path sequence. If time overlap is detected for cross-departmental examination items, the recommended path sequence is rearranged on the time axis, the rearranged recommended path sequence is fixed, and a medical treatment path planning table containing time nodes, department names and examination items is generated.

[0006] As a further aspect of the present invention, the step of obtaining the patient's initial state vector, and based on the patient's initial state vector, initiating a multi-round state inversion process, and by reverse traversing a pre-set clinical decision tree, deducing all the preceding pathological states that lead to the current condition, forming a set of preceding pathological states, including: The patient's initial state vector consists of the target patient's chief complaint information, present medical history data, and basic health records; The patient's initial state vector is mapped to the leaf nodes of the clinical decision tree, where each leaf node represents the current diagnostic conclusion. Activate the backtracking mechanism of the clinical decision tree, starting from the leaf node, and trace upwards along the reverse path of the parent-child relationship; In each backtracking operation, the necessary conditions that lead to the establishment of the current node are extracted, and the necessary conditions are defined as a new pre-pathological state. Repeat the backtracking operation until the root node of the clinical decision tree is reached or the preset backtracking depth threshold is reached; Collect all the preceding pathological states generated during the backtracking process, remove duplicate entries, and obtain the set of preceding pathological states.

[0007] As a further aspect of the present invention, the set of preceding pathological states is deduplicated and merged to generate a set of potential etiological nodes, including: A semantic similarity calculation model is established to measure the degree of overlap between two preceding pathological states in terms of medical concepts. Traverse each pair of preceding pathological states in the set of preceding pathological states and calculate the semantic similarity score between them. A similarity threshold is set, and preceding pathological states with semantic similarity scores higher than the similarity threshold are judged as equivalent states; The equivalent states are merged into a more general parent concept, which is the node in the set of potential pathogenesis nodes; For any preceding pathological state for which no equivalent state was found, it is directly included in the potential etiological node set.

[0008] As a further aspect of the present invention, matching the potential etiology node set with a pre-set departmental capability knowledge graph includes: Parse the medical terminology code for each node in the potential etiology node set; In the departmental capability knowledge graph, retrieve departmental nodes that contain the same medical term codes or semantically similar term codes; Record the department node corresponding to each potential cause node, as well as the treatment priority weight of the department node in the department's capability knowledge graph; If a potential cause node matches multiple department nodes, the department node with the highest treatment priority weight is selected as the department to which the potential cause node belongs. Construct a mapping table from the set of potential etiological nodes to the departmental nodes in the departmental capability knowledge graph.

[0009] As a further aspect of the present invention, the step of calculating the transfer cost from the patient's initial state vector to each target department based on the matching results includes: The transfer cost is determined by a combination of factors, including the difficulty of registration, waiting time, and the probability of interdisciplinary referral. Read the department corresponding to each potential cause node in the mapping table; Query the real-time scheduling data in the hospital information system to obtain the current appointment availability index and average waiting time for the department to which the appointment belongs; Assess the geographical distance from the current location of medical treatment to the department to which the patient belongs, and convert it into travel time; The complexity of the potential etiological node set is analyzed. If multi-system lesions are involved, the coordination cost of interdisciplinary joint diagnosis and treatment is calculated. The coordination cost is reflected in the additional communication and waiting time. The weighted sum of the appointment availability index, average waiting time, travel time, and coordination cost yields the transfer cost from the patient's initial state vector to their assigned department.

[0010] As a further aspect of the present invention, the invocation of a reinforcement learning-based path planning engine, wherein the path planning engine uses the transfer cost as environmental feedback and aims to minimize the total consultation time, performs positive reasoning calculations to generate a recommended path sequence from the initial consultation department to the final diagnosis department, including: Initialize an empty path sequence and set the patient's initial state vector to the current state; The path planning engine selects an action from all actions based on the current state, and the action represents going to a specific department for medical treatment; After the action is performed, the environment returns to the next state and the corresponding transfer cost, the next state being updated from the department's diagnosis result; The path planning engine updates its own policy network parameters using the transfer cost, and the policy network parameters determine the probability distribution of future action selections. Repeat execution until the current state meets the termination condition, i.e., the diagnosis is clear or the path length exceeds the limit; The selected actions throughout the loop are concatenated in sequence to form the recommended path sequence.

[0011] As a further aspect of the present invention, conflict detection is performed on the recommended path sequence. If time overlap is detected between cross-disciplinary examination items, the recommended path sequence is rearranged along the timeline, including: Analyze the list of examination items included in each step of the recommended medical treatment process; Extract the standard preparation time and estimated execution time for each inspection item; Based on the scheduling of appointment dates, a time window is allocated for each examination item; Check if two different inspection items are assigned to the same time period. If so, it is determined to be a time overlap conflict. For each time overlap conflict detected, adjust the execution order of one of the inspection items, moving it to an adjacent idle time period, until all conflicts are eliminated.

[0012] As a further aspect of the present invention, the step of allocating a time window for each examination item based on the scheduling of the consultation date includes: Obtain appointment calendar data for the hospital's radiology, laboratory, and functional testing departments; Identify the department type to which all examination items requiring appointments in the recommended path sequence belong; Based on the priority of department type, try to fill the examination items into the earliest available time slot in turn; If a certain inspection item cannot be inserted due to equipment being occupied, the search will continue to find the next available time slot. Record the final start and end times for each inspection item to create a preliminary schedule.

[0013] As a further aspect of the present invention, a semantic similarity calculation model is established, which is used to measure the degree of overlap between two preceding pathological states in terms of medical concepts, including: Obtain all standardized disease codes, symptom codes, and pathophysiological concepts from the medical knowledge base, and construct a unified medical concept ontology network. The medical concept ontology network consists of concept nodes and relation edges, and the relation edges include parent-child relationships, association relationships, and causal reasoning relationships. The text descriptions of all the preceding pathological states are extracted from the set of preceding pathological states, and natural language processing technology is used to perform word segmentation, stop word removal and medical terminology normalization on the text descriptions. The normalized medical terminology codes are mapped to the corresponding concept nodes in the medical concept ontology network; Calculate the shortest path distance between the concept nodes corresponding to two previous pathological states in the medical concept ontology network, and calculate the distance similarity score based on the shortest path distance; Calculate the depth of the common parent node of the concept nodes corresponding to two previous pathological states in the medical concept ontology network, and calculate the semantic hierarchical similarity score based on the depth of the common parent node. Based on the preset distance similarity weight and semantic similarity weight, the distance similarity score and semantic similarity score are weighted and summed to obtain the semantic similarity score, which measures the degree of overlap between two preceding pathological states in medical concepts.

[0014] As a further aspect of the present invention, the complexity of the potential causal node set is analyzed. If multi-system lesions are involved, the coordination cost of interdisciplinary joint diagnosis and treatment is calculated. The coordination cost is reflected in the additional communication and waiting time, including: For each node in the potential etiology node set, query its classification system in the medical terminology system according to its corresponding medical term code; The number of different medical systems involved in all nodes of the potential cause node set is counted, and the number is defined as the system complexity index. When the system complexity index is greater than one, it is determined to involve multiple systemic lesions; The average coordination time for cross-departmental diagnosis and treatment cases involving the same combination of medical systems recorded in the hospital information system is queried. The coordination time includes the average time from the issuance of the cross-departmental consultation request to the arrival of the consulting doctor, as well as the average delay time caused by resource scheduling when scheduling cross-departmental examination items. The coordination time for a unit is dynamically assessed based on the number of departments currently involved in diagnosis and treatment, the physical distance between departments, and the current workload of each department. The coordination cost of the interdisciplinary joint diagnosis and treatment is obtained by multiplying the system complexity index, the average coordination time of interdisciplinary diagnosis and treatment cases involving the same medical system combination, and the unit coordination time.

[0015] Compared with the prior art, the advantages and positive effects of the present invention are as follows: The process involves multiple rounds of state inversion, traversing a pre-set clinical decision tree in reverse to deduce all preceding pathological states that lead to the current symptom, forming a set of preceding pathological states. This set is then deduplicated and merged to generate a set of potential causal nodes. This approach comprehensively captures various preceding pathological information that may cause the current symptom, avoiding the limitations of conventional techniques that can only identify directly related causes. This makes causal investigation more comprehensive, providing a more complete and accurate basis for subsequent department matching, reducing department matching bias caused by missed causes, and making the starting point of the treatment path more targeted.

[0016] This approach utilizes a reinforcement learning-based path planning engine, employing the transfer cost calculated through department matching as environmental feedback. With the goal of minimizing total consultation time, it performs forward reasoning to generate a recommended path sequence from the initial consultation department to the final diagnosis department. This method dynamically combines the processing capacity and transfer cost of each department to achieve intelligent optimization of the consultation path. It breaks away from the simple sequential recommendation model of conventional technologies, rationally planning the consultation order of each department, shortening ineffective waiting time, and laying the foundation for time coordination of subsequent cross-departmental examinations, thus reducing process redundancy. Attached Figure Description

[0017] Figure 1 This is a sequence diagram of the patient visit path planning system based on artificial intelligence as described in this invention; Figure 2 A flowchart for generating a set of potential etiological nodes; Figure 3 A comparison chart of standardized total transfer costs for each department under different weights; Figure 4 A visual heatmap showing the matching degree between departmental capability knowledge graphs and diseases; Figure 5 This is a graph showing the relationship between the number of conflicts and the time required to resolve them. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0019] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0020] See Figure 1 After the system starts, the state inversion module first obtains the patient's initial state vector, which integrates key information such as the patient's chief complaint and medical history. Based on this vector, the state inversion module initiates a multi-round state inversion process. Its core is to reverse-traverse a pre-generated and stored clinical decision tree, starting from the node representing the current symptom, and inferring all possible antecedent pathological states that could lead to the symptom, collecting these states to form an antecedent pathological state set. The etiology refinement module processes this set, generating a more refined and non-redundant set of potential etiology nodes by deduplicating and merging semantically similar pathological states. The department matching module compares and matches this set of potential etiology nodes with a pre-defined departmental capability knowledge graph that details the diagnostic and treatment permissions and priorities of each department. Based on the matching results and combined with real-time hospital data, it calculates the transfer cost required to travel from the patient's current state to each possible target department. The path planning module is invoked, and its built-in reinforcement learning-based path planning engine optimizes the process by minimizing the total consultation time. It uses the calculated transfer cost as environmental feedback to perform positive reasoning and decision-making calculations, generating a recommended consultation path sequence from the suggested initial department to the final possible diagnosis department. The conflict resolution module analyzes and detects conflicts in this recommended path sequence, focusing on whether there is overlap in the timing of examinations in different departments. If such time overlaps are detected, the entire path sequence is reordered and rescheduled along the timeline. After eliminating all conflicts, the optimized path sequence is solidified into a structured consultation path planning table, which clearly lists the time nodes, department names, and examinations to be performed for each step.

[0021] In one embodiment of the present invention, the system integrates the target patient's chief complaint information, present medical history data, and basic health records to form a structured patient initial state vector. This module maps this patient initial state vector to a leaf node of a pre-defined clinical decision tree, where the leaf node represents the most likely diagnostic conclusion based on the current information. After mapping, the module activates the backtracking mechanism built into the clinical decision tree, starting from the leaf node and tracing upwards layer by layer along the reverse path of the parent-child relationships in the tree structure. In each backtracking operation, the system extracts the logically necessary conditions that lead to the current node's validity and defines these necessary conditions as a new pre-existing pathological state. This backtracking process is repeated until the root node of the clinical decision tree is reached, or a preset backtracking depth threshold is reached. All pre-existing pathological states generated during the backtracking process are collected, and the system initially removes completely duplicate entries, thus forming a set of pre-existing pathological states.

[0022] In its implementation, the state inversion module constructs an initial patient state vector by integrating multi-source medical information of the target patient. This initial state vector consists of three parts: the patient's stated complaints, current medical history data recording the progression of their illness, and a basic health record containing past health conditions. These three types of information are extracted from the hospital information system via a data interface and structured to form an initial vector representing the patient's current health status. The module maps this structured initial patient state vector to a leaf node of a pre-generated clinical decision tree stored in a knowledge base. Each leaf node represents a current diagnostic conclusion derived from the vector information. The mapping process involves matching the symptom and sign codes in the vector with the feature conditions on the decision tree nodes, finding the leaf node with the highest matching degree as the backtracking starting point. The module activates the built-in backtracking mechanism of the clinical decision tree, starting from the successfully matched leaf node and tracing upwards layer by layer along the reverse path of the parent-child relationship in the decision tree structure. Each backtracking corresponds to a logical deduction of the pathological cause.

[0023] In some embodiments, in each specific backtracking operation, the system extracts all the logically necessary conditions that lead to the current decision tree node state, defining each necessary condition as a new pre-pathological state. These necessary conditions may include pathophysiological changes, examination results, and symptom clusters. The backtracking operation is executed iteratively, with each iteration using the node reached in the previous backtracking as the new current node, and continuing to trace upwards to its parent node or higher-level etiological nodes. This process continues until the tracing path reaches the root node of the entire clinical decision tree, or the cumulative backtracking depth reaches a preset backtracking depth threshold. The backtracking depth threshold is used to prevent infinite tracing when the pathological chain is too long. Throughout the backtracking iteration, the system dynamically collects the new pre-pathological states generated by each backtracking operation, and after collection, performs deduplication on the collected pre-pathological state list, removing entries with completely duplicate codes, ultimately forming a set of pre-pathological states without duplicates for use by subsequent modules. Optionally, the structured fields of the patient's initial state vector may include, but are not limited to, symptom codes, quantitative values ​​of vital signs, vital sign parameters, and codes of key laboratory indicators. The construction of the clinical decision tree can be based on authoritative clinical guidelines and expert experience. The backtracking depth threshold can be preset and configured by the system administrator according to the characteristics of different disease specialties. For example, for diseases with relatively simple etiologies, the backtracking depth threshold can be set to 3, and for complex multi-system diseases, the backtracking depth threshold can be set to 5. The matching algorithm used in the mapping process can be to calculate the cosine similarity between the patient's initial state vector and the feature vectors of each leaf node of the decision tree, expressed by the formula: in: Represents the patient's initial state vector. The feature vectors representing the leaf nodes of the decision tree, along with the calculation of the vector dot product and modulus, are used to quantify the degree of matching. The similarity calculation result is used to select the leaf node most relevant to the patient's initial state vector from among many leaf nodes as the mapping target. In some embodiments, the deduplication process for the preceding pathological states is preliminary, removing only entries with identical codes. For semantically similar but differently coded state entries, deduplication and merging will be handled by a dedicated semantic similarity calculation model in the subsequent etiology refinement module.

[0024] In one embodiment of the present invention, see [reference] Figure 2This module establishes a semantic similarity calculation model to measure the overlap of medical concepts between any two preceding pathological states. To build this model, all standardized disease codes, symptom codes, and pathophysiological concepts from the medical knowledge base are acquired, and a unified medical concept ontology network is constructed based on this. This network consists of concept nodes and relation edges, including parent-child relationships, association relationships, and causal reasoning relationships. The module extracts text descriptions of all pathological states from the set of preceding pathological states and uses natural language processing techniques to segment these texts, remove stop words, and normalize medical terminology. The normalized medical terminology codes are mapped to the corresponding concept nodes in the medical concept ontology network. The shortest path distance between the corresponding concept nodes of two preceding pathological states in the ontology network is calculated, and a distance similarity score is calculated based on this distance. Simultaneously, the depth of the common parent node of these two concept nodes in the ontology network is calculated, and a semantic hierarchical similarity score is calculated based on this depth. The distance similarity score and the semantic hierarchical similarity score are weighted and summed according to preset weights to obtain the final semantic similarity score. After the model is ready, the module iterates through each pair of states in the set of preceding pathological states, calculating their semantic similarity scores. A similarity threshold is set, and pairs of states with scores higher than this threshold are considered semantically equivalent. Equivalent states are merged into a more general parent concept, which becomes a node in the potential cause node set. Preceding pathological states for which no equivalent state is found in the set are directly included in the potential cause node set. After generating the potential cause node set, the department matching module parses the standardized medical terminology code of each node in the set. The module retrieves department nodes containing the same medical terminology code or semantically highly similar terminology codes from a pre-defined departmental capability knowledge graph. The module records the department nodes matched to each potential cause node, as well as the predefined treatment priority weights of these department nodes in the knowledge graph. If a potential cause node matches multiple department nodes simultaneously, the department node with the highest treatment priority weight is selected as the final department to which the potential cause node belongs. After the traversal is complete, the module constructs a clear mapping table from the set of potential cause nodes to the department nodes in the department capability knowledge graph.

[0025] In its implementation, the etiology refinement module receives a set of preceding pathological states output by the state inversion module. This module establishes a semantic similarity calculation model, the core function of which is to measure the degree of overlap between any two preceding pathological states in terms of their medical conceptual essence. Establishing this model requires acquiring all standardized disease codes, symptom codes, and pathophysiological concepts from the medical knowledge base. Based on these standardized codes and concepts, a unified medical concept ontology network is constructed. This network consists of concept nodes and relational edges. Relational edges include parent-child relationships expressing hierarchical relationships, association relationships expressing relevance, and causal reasoning relationships expressing causal logic. The etiology refinement module extracts the original text descriptions of all preceding pathological states from the set of preceding pathological states. It then uses natural language processing techniques to segment, remove stop words, and normalize medical terminology in the original text descriptions. Finally, it maps the processed medical terminology codes to the corresponding concept nodes in the medical concept ontology network. In some embodiments, for any two state entries in the set of preceding pathological states, the semantic similarity calculation model begins to calculate their semantic similarity score. The calculation process involves two main parts. The first part is to calculate the shortest path distance between the concept nodes corresponding to the two preceding pathological states in the medical concept ontology network, and calculate the distance similarity score based on the shortest path distance; the shorter the distance, the higher the score. The second part is to calculate the depth of the common parent node of the two concept nodes in the medical concept ontology network, and calculate the semantic hierarchical similarity score based on the depth of the common parent node; the higher the level of the common parent node, the lower the score. The semantic similarity calculation model combines the scores from the two parts through a preset weighted formula, which is expressed as: Where: symbol The symbol represents the final semantic similarity score obtained from the calculation. and These represent the preset distance similarity weight and semantic similarity weight, respectively, and satisfy the following conditions: .symbol The function represents the shortest path distance between two concept nodes in the medical concept ontology network. It is the shortest path distance The function that converts to distance similarity scores is typically a monotonically decreasing function. (Symbol) The function represents the depth of the common parent node of two concept nodes in the medical concept ontology network. It is to deepen The function converts the data into semantic similarity scores, typically a monotonically decreasing function. The semantic similarity calculation model traverses the set of preceding pathological states, calculating a semantic similarity score for each pair of state entries. Optionally, medical terminology normalization can refer to the International Classification of Diseases (ICD) coding system or the Systematic Clinical Medical Terminology Set (SNOMEDCT), and the construction of the medical concept ontology network can be based on UMLS super ontology or domain-specific ontology. One specific implementation could be ,in It is a positive decay constant. Function One specific implementation could be ,in It is a decay factor between 0 and 1. The etiology refinement module sets a fixed similarity threshold, classifying any pair of preceding pathological states with a semantic similarity score higher than the threshold as semantically equivalent. For a set of preceding pathological states judged as equivalent, the etiology refinement module merges them into a more generalized parent concept, which becomes an independent node in the potential etiology node set. After traversing and processing all possible state pairs, for those preceding pathological state entries for which no equivalent state is found in the entire set of preceding pathological states, the etiology refinement module directly includes them in the final generated potential etiology node set.

[0026] In some embodiments, after generating a set of potential cause nodes, the department matching module begins operation. This module parses the standardized medical terminology code of each node in the set. The department matching module searches a pre-defined departmental capability knowledge graph, which defines the treatment authority and priority levels of each clinical department for specific diseases or symptoms. The search goal is to find department nodes in the knowledge graph that have the exact same medical terminology code as the potential cause node, or whose codes are different but are determined to be highly semantically similar by a semantic similarity calculation model. The department matching module records all department nodes successfully matched to each potential cause node and simultaneously records the predefined treatment priority weights of these department nodes in the departmental capability knowledge graph. Optionally, if a potential cause node matches multiple different department nodes in the departmental capability knowledge graph, the department matching module compares the treatment priority weights of these department nodes and selects the department node with the highest treatment priority weight as the final department to which this potential cause node belongs. After the department matching module completes the matching and attribution determination of all nodes in the potential cause node set, it constructs a clear mapping relationship table from the potential cause node set to the specific department node in the department capability knowledge graph. The mapping relationship table includes at least the potential cause node code, the matched attribution department code, and the treatment priority weight.

[0027] In one embodiment of the present invention, after completing the mapping relationship construction, the department matching module performs transfer cost calculation. The transfer cost is determined by a combination of factors, including registration difficulty, waiting time, and the probability of inter-departmental referral. The module reads the department information corresponding to each potential cause node in the mapping relationship table. It queries the real-time scheduling data in the hospital information system to obtain the current appointment availability index and average waiting time data of these departments. The system assesses the geographical distance from the patient's current location to the target department and converts it into an estimated travel time based on the speed of movement within the hospital. Simultaneously, the module analyzes the complexity of the potential cause node set to determine whether it involves multi-system lesions. In specific implementation, for each node in the potential cause node set, its system of belonging in the standard medical classification system is queried according to its medical terminology code. The number of different medical systems involved in all nodes is counted, and this number is defined as the system complexity index. When the system complexity index is greater than one, it is determined that it involves multi-system lesions. Subsequently, the hospital information system's historical records were consulted to extract the average coordination time for past cross-departmental consultation cases involving the same medical system combination. This time includes the average time from submitting a cross-departmental consultation request to the arrival of the consulting physician, as well as the average delay caused by resource scheduling when scheduling cross-departmental examinations. A unit coordination time was dynamically assessed based on the number of departments currently involved, the physical distance between departments, and the current workload of each department. The system complexity index, the historical average coordination time, and the dynamically assessed unit coordination time were multiplied to obtain the coordination cost of cross-departmental joint consultation. Finally, the obtained appointment availability index, average waiting time, travel time, and the calculated coordination cost were weighted and summed to obtain the transfer cost from the patient's initial state vector to each affiliated department.

[0028] In practical implementation, after constructing the mapping table from potential cause node sets to department nodes, the department matching module calculates the transfer cost from the patient's initial state vector to each target department. The transfer cost is determined by a combination of factors, including registration difficulty, waiting time, and the probability of inter-departmental referral. The module reads the department information corresponding to each potential cause node in the mapping table, queries real-time scheduling data in the hospital information system, and obtains the current appointment availability index and average waiting time for each department. The appointment availability index reflects the ratio of available appointments to the number of patients waiting for treatment, while the average waiting time is estimated based on historical and real-time queue data. The system assesses the geographical distance from the patient's current location to the target department. This geographical distance is obtained based on path planning on the hospital's electronic map and converted into travel time according to a preset average travel speed within the hospital. The department matching module standardizes these raw indicators, converting the appointment availability index, average waiting time, and travel time corresponding to registration difficulty into dimensionless standardized values. In some embodiments, the complexity of the potential causal node set is analyzed to determine whether multi-system lesions are involved. For each node in the potential causal node set, its system affiliation in the standard medical classification system is queried according to its corresponding standardized medical terminology code. The standard medical classification system is divided into respiratory, circulatory, digestive, etc. The number of different medical systems involved in all nodes in the potential causal node set is counted, and this number is defined as the system complexity index. When the system complexity index is greater than one, it is determined that the current patient's condition involves multi-system lesions. Past cross-departmental diagnosis and treatment cases involving the same combination of medical systems are queried from the hospital information system. The average coordination time of these cases is extracted and calculated. The average coordination time includes the average time from the formal issuance of the cross-departmental consultation request to the actual arrival of the relevant consulting doctor, as well as the average delay time caused by resource scheduling conflicts such as equipment and personnel when arranging cross-departmental examination items. Based on the number of departments currently involved in the diagnosis and treatment, the physical distance between departments, and the current workload of each department, a unit coordination time is dynamically evaluated. The unit coordination time is an adjustment coefficient that integrates the current real-time operational status. The system complexity index, the average coordination time for cross-disciplinary cases involving the same medical system combination, and the dynamically assessed unit coordination time are multiplied to obtain a raw coordination time. This raw coordination time is then standardized to convert it into a dimensionless standardized coordination cost. Optionally, the formula for calculating the total transfer cost can be expressed as a weighted sum of each standardized cost factor. The formula is as follows: Where: symbol This represents the total transfer cost calculated from the patient's initial state vector to a specific department. It is a dimensionless scalar. (Symbol) , , and These represent the preset weighting coefficients for four cost factors: registration difficulty, waiting time, travel time, and coordination costs. All weighting coefficients are dimensionless scalars and satisfy the following conditions: .symbol This represents the dimensionless cost of registration difficulty obtained after standardizing the original registration availability index. (Symbol) This represents the dimensionless waiting time cost obtained after standardizing the original average waiting time. (Symbol) This represents the dimensionless movement cost obtained after standardizing the original movement time. (Symbol) The standardized coordination cost represents the dimensionless coordination cost obtained after standardizing the original coordination time. If the system complexity exponent is no greater than one, then the standardized coordination cost is... Set to zero. The department matching module calculates the standardized total transfer cost for each department in the mapping table according to this process and formula. It is understandable that standardization can employ methods such as min-max normalization or Z-score standardization to map raw data with different dimensions to a unified, unitless standard scale, for example, mapping to the [0,1] interval. Weighting coefficients to The specific values ​​can be adjusted by the system administrator based on the hospital's operational strategy. The final calculated standardized total transfer cost for each department will serve as a key environmental feedback signal input to the path planning module, which will then be used by the reinforcement learning-based path planning engine when making decisions.

[0029] See Figure 3 This is a comparison chart of standardized total transfer costs for various departments under different weights. The core analysis is the changing patterns of standardized total transfer costs for each department under three weight configurations. Cardiology and neurology consistently rank high, representing the two departments with the highest consultation costs, primarily due to limited appointment slots, long waiting times, or coordination costs associated with multi-system diseases. Orthopedics and gastroenterology have the lowest costs; orthopedics has zero coordination costs due to the absence of multi-system diseases, while gastroenterology excels in both basic and coordination costs. Respiratory medicine and endocrinology fall in the middle range, showing relatively mild effects from weight adjustments. Weighted refactoring 2 places greater emphasis on waiting times and travel time, resulting in a slight increase in costs for cardiology and endocrinology, and a slight decrease in costs for neurology. Weighted refactoring 3 places greater emphasis on coordination costs, significantly decreasing costs for departments with multi-system diseases, while neurology experiences a slight increase due to its high proportion of coordination costs. Weighted refactoring 1 serves as a baseline, reflecting the cost distribution under the hospital's default operating strategy.

[0030] In one embodiment of the invention, the path planning module invokes a reinforcement learning-based path planning engine to generate a path. The engine first initializes an empty path sequence and sets the patient's initial state vector as the current state. At each decision step, the path planning engine, based on its internal policy network, selects an action from all possible actions according to the current state; here, an action represents going to a department for consultation. After executing the action, the system simulates the environment and returns the next state and the corresponding transition cost, where the next state is formed by updating the current state with the diagnosis result of the selected department. The path planning engine uses the transition cost obtained in this step as environmental feedback to update the parameters of its internal policy network. These parameters determine the probability distribution of selecting various actions in similar states in the future. The process of selecting an action, executing the action, obtaining feedback, and updating the network is repeated until the current state meets a preset termination condition, such as a definitive diagnosis or the path sequence length exceeding a limit. The engine sequentially concatenates the actions selected throughout the loop to form a recommended path sequence from start to finish.

[0031] In practice, the path planning module invokes a reinforcement learning-based path planning engine to generate recommended path sequences. The engine starts with the patient's initial state vector, aims to minimize the total time from initial consultation to obtaining a definitive diagnosis, and uses the transfer cost calculated by the department matching module as the environmental feedback signal. The path planning engine initializes an empty path sequence and sets the patient's initial state vector as the current state. This initial state vector contains structured information such as the patient's symptoms, signs, and basic medical history. At each decision step, the reinforcement learning-based path planning engine selects an action from all executable actions based on its internal policy network and the current state. Each action in the action space represents a visit to a department that exists in the department matching module's mapping table.

[0032] In some embodiments, after executing the selected action of proceeding to a specific department, the simulated treatment environment returns the next state and the corresponding transfer cost. The next state is formed by updating the current state in conjunction with the typical diagnostic results of the selected department, such as adding the preliminary diagnosis or examination results that the department might give to the state vector. The transfer cost is directly taken from the transfer cost pre-calculated for that department by the department matching module. The reinforcement learning-based path planning engine uses the obtained transfer cost as environmental feedback to update the parameters of its internal policy network. The parameters of the policy network determine the probability distribution of selecting each possible action when encountering similar or identical patient states in the future. The goal of the update is to make the policy more inclined to select action sequences that yield lower long-term cumulative transfer costs. The process of selecting an action, executing the action, obtaining environmental feedback, and updating the policy network parameters is repeated in a loop. Each loop updates the current state to the next state returned after the previous action execution. Optionally, the policy network can be a deep neural network, whose input is a feature representation of the patient state and whose output is the probability distribution of each selectable action. The update of the policy network parameters can use a policy gradient algorithm, the core update formula of which involves calculating the gradient of the expected reward. An exemplary state-action value function update formula can be expressed as: Where: symbol Represents the state Select action The state-action value function estimate. (Symbol) Represents the learning rate, controlling the step size for parameter updates. (Symbol) Represents the execution of actions The negative value of the immediate transfer cost following post-environmental feedback. (Symbol) This represents the discount factor, used to measure the current value of future rewards. (Symbol) Represents the next state The state-action value function estimate is the largest among all available actions. The update process gradually brings the value estimate closer to the actual optimal value. See Table 1, which shows the probability distribution of the policy network for the three available department actions in a given decision step.

[0033] Table 1: Probability distribution of the policy network for actions to the three available departments in a certain decision step. Understandably, the repetitive loop process will continue until a preset termination condition is met. The termination condition is typically set as follows: the diagnostic confidence in the current state vector exceeds a threshold indicating a definitive diagnosis, or the path sequence length reaches a preset maximum step limit. When the termination condition is met, the reinforcement learning-based path planning engine stops exploring and sequentially concatenates all actions selected throughout the entire loop from the initial state to the termination state, forming a complete recommended path sequence from the suggested initial consultation department to the final confirmed diagnosis department. This recommended path sequence is output as a department code sequence, such as "Emergency Department -> Cardiology Department -> Echocardiography Department," and is passed to the subsequent conflict resolution module for processing.

[0034] See Figure 4 This is a heatmap visualizing the matching degree between departmental capability knowledge graphs and diseases. The more yellow the color, the higher the matching score; the darker the purple, the lower the matching score. The Emergency Department consistently scores around 80 points for all systemic diseases, making it the only department with a high matching degree for multi-system diseases. This aligns with the Emergency Department's role as a "first-line care provider and emergency treatment provider for multiple systems," making it suitable as the initial entry point for patients with multi-system diseases. All departments show low matching degrees for endocrine system diseases, indicating that endocrine system diseases are not currently clearly assigned to specific departments or require multi-departmental collaboration. Single-system diseases should be prioritized for matching to corresponding specialties to maximize treatment efficiency; multi-system diseases or critical illnesses can be initially handled by the Emergency Department before being triaged to other specialties.

[0035] In one embodiment of the present invention, the conflict resolution module refines the scheduling of the recommended path sequence. The module first parses the list of examination items included in each step of the consultation process in the recommended path sequence and extracts the standard preparation time and estimated execution time for each examination item. The module obtains appointment calendar data from auxiliary departments such as the hospital's radiology department, laboratory department, and functional examination department, identifying all examination items requiring appointment in the path sequence and their respective department types. According to the preset priority of department types, the module sequentially attempts to fill each examination item into the earliest available time slot in its corresponding department. If an examination item cannot be inserted due to equipment occupancy or a full time slot, it automatically postpones to find the next available time slot. Through this process, a specific time window is assigned to each examination item, and its determined start and end times are recorded, forming a preliminary schedule. Based on this, the module checks whether two different examination items are assigned to the same time slot; if so, it is determined to be a time overlap conflict. For each detected time overlap conflict, the module adjusts the execution order of one of the examination items, moving it to an adjacent free time slot. This process is iterated until all detected time overlap conflicts are eliminated.

[0036] In practical implementation, the conflict resolution module receives the recommended path sequence generated by the path planning module. The module parses the list of examination items included in each step of the recommended path sequence, extracting the standard preparation time and estimated execution time for each examination item. Standard preparation time refers to the standardized preparation time required for a patient to undergo a specific examination, and estimated execution time refers to the average time taken for the examination itself. The conflict resolution module obtains appointment calendar data from the hospital's radiology, laboratory, and functional examination departments, identifying all examination items requiring appointment in the recommended path sequence and their respective department types. Department type priority can be preset by the system administrator; for example, setting the radiology department, which has more limited resources, as the department for MRI examinations, as a higher priority. In some embodiments, the conflict resolution module attempts to fill the examination items into the earliest available time slots according to the preset priority order of department types. A time slot is a period of time divided into fixed durations in the appointment calendar. For an examination item to be scheduled, the conflict resolution module searches the appointment calendar of its department for the first consecutive free time slot that can accommodate the estimated execution time of the examination item, starting from the start time of the currently planned appointment date. If a time slot is found to be occupied by another inspection item during the search, the conflict resolution module continues searching for the next available consecutive free time slot. After successfully allocating a time window for each inspection item, the conflict resolution module records the final planned start and end times for that inspection item, forming a preliminary schedule. Optionally, the constraints for allocating inspection item time windows can be expressed by a formula to ensure that items do not overlap and resources are sufficient. The formula can be expressed as: Where: symbol and Represents any two distinct check item indices. (Symbol) Representative inspection items The assigned start time of the plan, symbol Representative inspection items The assigned plan has a completion time and meets the following conditions: ,in It is an inspection item Standard preparation time It is an inspection item The estimated execution time. (Symbol) Indicates the items to be inspected The formula indicates that for any two different inspection items, the intersection of their respective allocated time window intervals must be empty, meaning that time overlap is not allowed. The conflict resolution module, based on the preliminary schedule, checks whether two different inspection items are assigned to the same time period. The criterion is whether there is a non-empty intersection between the time window intervals of the two inspection items. If a non-empty intersection exists, it is determined to be a time overlap conflict. In essence, after completing the preliminary schedule, the conflict resolution module executes conflict detection logic, traversing all inspection item pairs in the preliminary schedule and checking whether the time windows of each pair overlap. For each detected time overlap conflict, the conflict resolution module initiates an adjustment procedure. This procedure selects one inspection item from the conflict pair and moves its entire planned time window to the next available free time period adjacent to the appointment calendar of its department. After the move, it checks again whether new overlaps occur with the time windows of other items. This iterative process of adjustment, checking, and readjustment continues until the intersection of the time windows of all inspection item pairs in the preliminary schedule is empty, thus eliminating all detected time overlap conflicts. After eliminating all time overlap conflicts, the conflict resolution module combines the rearranged time window sequence with the department order in the original recommended path sequence to solidify a final conflict-free medical treatment path planning table, which contains clear time nodes, department names, and examination item information.

[0037] See Figure 5 This is a graph showing the relationship between the number of conflicts and resolution time, illustrating the positive correlation between the number of conflicts in cross-departmental examinations within the patient's treatment path and the time required to resolve them. It serves as a core performance visualization of the conflict resolution module. The more conflicts there are, the longer the resolution time increases non-linearly. With zero conflicts, the resolution time is zero, consistent with the logic that no conflict requires no adjustment. Each additional conflict increases the average resolution time by approximately 24 minutes, and the incremental time increases significantly when the number of conflicts is ≥4. In the low-conflict stage (1-3 conflicts), the incremental time stabilizes at 15 minutes per conflict, indicating that a small number of conflicts can be quickly resolved by simply extending the time slot. In the high-conflict stage (4-5 conflicts), the incremental time rises to 30 minutes per conflict, indicating that when multiple conflicts overlap, frequent resource coordination and adjustments to the overall timeline are required, significantly increasing complexity. During the reinforcement learning path planning stage, patient paths with ≥4 conflicts should be avoided as much as possible to control the total treatment time.

[0038] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A patient visit path planning system based on artificial intelligence, characterized in that, include: The state inversion module obtains the patient's initial state vector and, based on the patient's initial state vector, initiates a multi-round state inversion process. By traversing the preset clinical decision tree in reverse, it deduces all the preceding pathological states that lead to the current symptoms, forming a set of preceding pathological states. The etiology refinement module removes duplicates and merges the set of prior pathological states to generate a set of potential etiology nodes. The department matching module matches the potential etiology node set with a pre-set department capability knowledge graph. The department capability knowledge graph defines the treatment authority and priority level of each clinical department for specific diseases or symptoms. Based on the matching results, the module calculates the transfer cost from the patient's initial state vector to each target department. The path planning module calls a path planning engine based on reinforcement learning. The path planning engine uses the transfer cost as environmental feedback and aims to minimize the total consultation time to perform positive reasoning calculations and generate a recommended path sequence from the initial consultation department to the final diagnosis department. The conflict resolution module performs conflict detection on the recommended path sequence. If time overlap is detected for cross-departmental examination items, the recommended path sequence is rearranged on the time axis, the rearranged recommended path sequence is fixed, and a medical treatment path planning table containing time nodes, department names and examination items is generated.

2. The patient visit path planning system based on artificial intelligence as described in claim 1, characterized in that, The process involves obtaining the patient's initial state vector, initiating a multi-round state inversion process based on this vector, and reversing the pre-set clinical decision tree to deduce all preceding pathological states that lead to the current symptom, forming a set of preceding pathological states, including: The patient's initial state vector consists of the target patient's chief complaint information, present medical history data, and basic health records; The patient's initial state vector is mapped to the leaf nodes of the clinical decision tree, where each leaf node represents the current diagnostic conclusion. Activate the backtracking mechanism of the clinical decision tree, starting from the leaf node, and trace upwards along the reverse path of the parent-child relationship; In each backtracking operation, the necessary conditions that lead to the establishment of the current node are extracted, and the necessary conditions are defined as a new pre-pathological state. Repeat the backtracking operation until the root node of the clinical decision tree is reached or the preset backtracking depth threshold is reached; Collect all the preceding pathological states generated during the backtracking process, remove duplicate entries, and obtain the set of preceding pathological states.

3. The patient visit path planning system based on artificial intelligence as described in claim 2, characterized in that, The set of pre-existing pathological states is deduplicated and merged to generate a set of potential etiological nodes, including: A semantic similarity calculation model is established to measure the degree of overlap between two preceding pathological states in terms of medical concepts. Traverse each pair of preceding pathological states in the set of preceding pathological states and calculate the semantic similarity score between them. A similarity threshold is set, and preceding pathological states with semantic similarity scores higher than the similarity threshold are judged as equivalent states; The equivalent states are merged into a more general parent concept, which is the node in the set of potential pathogenesis nodes; For any preceding pathological state for which no equivalent state was found, it is directly included in the potential etiological node set.

4. The patient visit path planning system based on artificial intelligence as described in claim 3, characterized in that, Matching the set of potential etiological nodes with a pre-set departmental capability knowledge graph includes: Parse the medical terminology code for each node in the potential etiology node set; In the departmental capability knowledge graph, retrieve departmental nodes that contain the same medical term codes or semantically similar term codes; Record the department node corresponding to each potential cause node, as well as the treatment priority weight of the department node in the department's capability knowledge graph; If a potential cause node matches multiple department nodes, the department node with the highest treatment priority weight is selected as the department to which the potential cause node belongs. Construct a mapping table from the set of potential etiological nodes to the departmental nodes in the departmental capability knowledge graph.

5. The patient visit path planning system based on artificial intelligence as described in claim 4, characterized in that, The step of calculating the transfer cost from the patient's initial state vector to each target department based on the matching results includes: The transfer cost is determined by a combination of factors, including the difficulty of registration, waiting time, and the probability of interdisciplinary referral. Read the department corresponding to each potential cause node in the mapping table; Query the real-time scheduling data in the hospital information system to obtain the current appointment availability index and average waiting time for the department to which the appointment belongs; Assess the geographical distance from the current location of medical treatment to the department to which the patient belongs, and convert it into travel time; The complexity of the potential etiological node set is analyzed. If multi-system lesions are involved, the coordination cost of interdisciplinary joint diagnosis and treatment is calculated. The coordination cost is reflected in the additional communication and waiting time. The weighted sum of the appointment availability index, average waiting time, travel time, and coordination cost yields the transfer cost from the patient's initial state vector to their assigned department.

6. The patient visit path planning system based on artificial intelligence as described in claim 5, characterized in that, The call utilizes a reinforcement learning-based path planning engine. This engine uses the transfer cost as environmental feedback and aims to minimize the total consultation time. It performs positive reasoning calculations to generate a recommended path sequence from the initial consultation department to the final diagnosis department, including: Initialize an empty path sequence and set the patient's initial state vector to the current state; The path planning engine selects an action from all actions based on the current state, and the action represents going to a specific department for medical treatment; After the action is performed, the environment returns to the next state and the corresponding transfer cost, the next state being updated from the department's diagnosis result; The path planning engine updates its own policy network parameters using the transfer cost, and the policy network parameters determine the probability distribution of future action selections. Repeat execution until the current state meets the termination condition, i.e., the diagnosis is clear or the path length exceeds the limit; The selected actions throughout the loop are concatenated in sequence to form the recommended path sequence.

7. The patient visit path planning system based on artificial intelligence as described in claim 6, characterized in that, Conflict detection is performed on the recommended path sequence. If time overlap is detected for cross-disciplinary examination items, the recommended path sequence is rearranged along the timeline, including: Analyze the list of examination items included in each step of the recommended medical treatment process; Extract the standard preparation time and estimated execution time for each inspection item; Based on the scheduling of appointment dates, a time window is allocated for each examination item; Check if two different inspection items are assigned to the same time period. If so, it is determined to be a time overlap conflict. For each time overlap conflict detected, adjust the execution order of one of the inspection items, moving it to an adjacent idle time period, until all conflicts are eliminated.

8. The patient visit path planning system based on artificial intelligence as described in claim 7, characterized in that, The process involves allocating a time window for each examination item based on the scheduling of appointment dates, including: Obtain appointment calendar data for the hospital's radiology, laboratory, and functional testing departments; Identify the department type to which all examination items requiring appointments in the recommended path sequence belong; Based on the priority of department type, try to fill the examination items into the earliest available time slot in turn; If a certain inspection item cannot be inserted due to equipment being occupied, the search will continue to find the next available time slot. Record the final start and end times for each inspection item to create a preliminary schedule.

9. The patient visit path planning system based on artificial intelligence as described in claim 8, characterized in that, The step involves establishing a semantic similarity calculation model, which measures the degree of overlap between two preceding pathological states in terms of medical concepts, including: Obtain all standardized disease codes, symptom codes, and pathophysiological concepts from the medical knowledge base, and construct a unified medical concept ontology network. The medical concept ontology network consists of concept nodes and relation edges, and the relation edges include parent-child relationships, association relationships, and causal reasoning relationships. The text descriptions of all the preceding pathological states are extracted from the set of preceding pathological states, and natural language processing technology is used to perform word segmentation, stop word removal and medical terminology normalization on the text descriptions. The normalized medical terminology codes are mapped to the corresponding concept nodes in the medical concept ontology network; Calculate the shortest path distance between the concept nodes corresponding to two previous pathological states in the medical concept ontology network, and calculate the distance similarity score based on the shortest path distance; Calculate the depth of the common parent node of the concept nodes corresponding to two previous pathological states in the medical concept ontology network, and calculate the semantic hierarchical similarity score based on the depth of the common parent node. Based on the preset distance similarity weight and semantic similarity weight, the distance similarity score and semantic similarity score are weighted and summed to obtain the semantic similarity score, which measures the degree of overlap between two preceding pathological states in medical concepts.

10. The patient visit path planning system based on artificial intelligence as described in claim 9, characterized in that, The complexity of the potential etiological node set is analyzed. If multi-system diseases are involved, the coordination cost of interdisciplinary joint diagnosis and treatment is calculated. This coordination cost is reflected in the additional communication and waiting time, including: For each node in the potential etiology node set, query its classification system in the medical terminology system according to its corresponding medical term code; The number of different medical systems involved in all nodes of the potential cause node set is counted, and the number is defined as the system complexity index. When the system complexity index is greater than one, it is determined to involve multiple systemic lesions; The average coordination time for cross-departmental diagnosis and treatment cases involving the same combination of medical systems recorded in the hospital information system is queried. The coordination time includes the average time from the issuance of the cross-departmental consultation request to the arrival of the consulting doctor and the average delay time caused by resource scheduling when scheduling cross-departmental examination items. The coordination time for a unit is dynamically assessed based on the number of departments currently involved in diagnosis and treatment, the physical distance between departments, and the current workload of each department. The coordination cost of the interdisciplinary joint diagnosis and treatment is obtained by multiplying the system complexity index, the average coordination time of interdisciplinary diagnosis and treatment cases involving the same medical system combination, and the unit coordination time.

Images