Methods, systems, and media for mining the sequential diagnosis and treatment logic of TCM based on dynamic efficacy feedback.

Abstract

This invention discloses a method, system, and medium for mining the sequential diagnosis and treatment logic of Traditional Chinese Medicine (TCM) based on dynamic efficacy feedback. The method includes: preprocessing TCM time-series diagnosis and treatment data; constructing a sequence of consultation pairs containing multiple patient visits, each pair including consultation status and efficacy feedback; mapping symptoms and prescriptions into structured vectors based on TCM theory to generate state transition tuples; calculating the dynamic evolution pattern of the composite state of the state transition tuples using a state transition probability formula, combined with efficacy feedback information; extracting rules from the composite state vector mapping using a sequential rule mining algorithm, and generating rules consistent with TCM syndrome differentiation thinking using an interpretable decision tree model. This invention mines the hidden dynamic adjustment patterns of TCM syndrome differentiation and treatment in medical records, realizing the visualization and rule-based approach to diagnosis and treatment logic.

Description

Technical Field

[0001] This invention relates to the field of medical data analysis technology, and in particular to a method, system, and medium for mining the sequential diagnosis and treatment logic of traditional Chinese medicine based on dynamic efficacy feedback. Background Technology

[0002] The core of Traditional Chinese Medicine (TCM) diagnosis and treatment lies in dynamic syndrome differentiation and treatment. In clinical practice, treatment principles and methods must be continuously adjusted based on the patient's condition evolution and efficacy feedback, forming a sequential diagnostic and treatment logic. However, current technologies face the following bottlenecks when mining and analyzing TCM time-series medical record data:

[0003] Traditional electronic health record (EHR) analysis methods often focus on static associations within a single visit, making it difficult to capture the dynamic evolution of "symptom-prescription" states between visits. For example, the Apriori algorithm can discover co-occurrence patterns between symptoms and medications, but it cannot reveal dynamic pathogenesis transformation pathways such as "qi deficiency and blood stasis → phlegm-dampness obstruction → qi and yin deficiency." While sequence pattern mining techniques can process time-series data, their ability to model complex states is insufficient, and they cannot effectively integrate the driving role of efficacy feedback on state transitions.

[0004] Existing research often treats efficacy as a general label rather than a core conditional variable that triggers adjustments in syndrome differentiation. For example, time series models (such as LSTM and HMM) can predict disease progression, but their "black box" nature makes it impossible to explicitly link clinical decision-making logic such as "ineffective treatment → change prescription." This makes it difficult to quantify and model the dynamic thinking of traditional Chinese medicine, which emphasizes "no change of prescription if effective, change if ineffective."

[0005] While deep learning models can handle complex temporal relationships, the rules they extract lack consistency with traditional Chinese medicine (TCM) theories. For example, a model might output the association of "symptom A → drug B," but it cannot explain the intrinsic connection between this association and treatment principles such as "soothing the liver and relieving stagnation → regulating qi and promoting blood circulation," making it difficult for clinicians to trust and apply. Summary of the Invention

[0006] Based on the shortcomings of the existing technologies, this invention provides a method, system, and medium for mining the sequential diagnosis and treatment logic of traditional Chinese medicine with dynamic efficacy feedback, thereby realizing the visualization, standardization, and clinical application of the diagnosis and treatment logic.

[0007] To address the aforementioned technical problems, the first aspect of this invention discloses a method for mining the sequential diagnosis and treatment logic of Traditional Chinese Medicine based on dynamic efficacy feedback, comprising:

[0008] Preprocess TCM time-series diagnosis and treatment data to construct a sequence of consultation pairs containing multiple visits by patients. Each consultation pair includes the consultation status and efficacy feedback.

[0009] Based on traditional Chinese medicine theory, symptoms and prescriptions are mapped into structured vectors to generate state transition tuples;

[0010] By combining therapeutic feedback information, the dynamic evolution pattern of the composite state of the state transition tuple is calculated using the state transition probability formula; the rules of the composite state vector mapping are extracted using the sequential rule mining algorithm, and the rules that conform to the TCM syndrome differentiation thinking are generated by combining the interpretable decision tree model.

[0011] In some implementations, the dynamic evolution pattern of the composite state is calculated using a state transition probability formula, incorporating therapeutic feedback information, including:

[0012] Based on the sequence of visits, the state transition probability under different efficacy feedback is calculated, and a state transition probability matrix is ​​generated.

[0013] The frequent state sequence mining algorithm is used to extract the rules of previous state-treatment-next state, and the strong association rules that meet the support and confidence thresholds are selected.

[0014] An interpretable decision tree model is constructed to predict subsequent prescription strategies based on efficacy feedback and previous state, and rules that conform to the diagnostic thinking of traditional Chinese medicine are extracted.

[0015] In some implementations, the state transition probability under different therapeutic feedback is calculated using the following formula:

[0016] P(State {i+1}|State i, Efficacy {i+1}) = Count(State i, Efficacy {i+1}, State {i+1}) / Sum {all S'} Count(State i, Efficacy {i+1}, S')

[0017] Where P is the state transition probability, State i is the state at the i-th time, Efficacy {i+1} is the structured evaluation at the (i+1)-th time, and S' is the set of all possible next states under the current state State i and efficacy Efficacy {i+1}.

[0018] In some implementations, an interpretable decision tree model is constructed to predict subsequent prescription strategies based on efficacy feedback and previous states, and extracts data that conforms to the rules of TCM syndrome differentiation, including:

[0019] Extract defined features and target variables from the diagnosis pair transformation dataset T, and construct a feature matrix;

[0020] Using the C4.5 algorithm, information gain and split information are calculated. Based on the information gain rate as the splitting criterion, nodes with higher gain rates are selected as decision nodes to construct a decision tree.

[0021] Overfitting branches are removed by pessimistic pruning algorithm, while decision-making paths that conform to clinical experience are preserved;

[0022] Rules that conform to the diagnostic thinking of traditional Chinese medicine are extracted based on the decision tree model.

[0023] In some implementations, overfitting branches are pruned using a pessimistic pruning algorithm, preserving decision paths that align with clinical experience, including:

[0024] Calculate the pessimistic error estimate of the subtree, the adjusted error number of the subtree, and the error number of leaf nodes after pruning;

[0025] Compare the adjusted error count of the subtree with the adjusted error count of the leaf nodes after pruning. If the adjusted error count of the leaf nodes after pruning is not greater than the adjusted error count of the subtree, then perform the pruning operation.

[0026] In some implementations, rules that conform to the diagnostic thinking of traditional Chinese medicine are extracted, including:

[0027] Identify the efficacy split point in the decision tree and analyze the first and second branches corresponding to the efficacy split point; the first branch is the branch with good efficacy, and the second branch is the branch with poor efficacy.

[0028] Compare the differences between the first and second branches, and statistically analyze the proportion of the adherence rule in the first branch and the proportion of the substitution rule in the second branch to quantify the driving weight of the therapeutic effect on the diagnostic decision-making.

[0029] By pessimistically pruning branches with support below a preset threshold, the IF-THEN rule set that conforms to the diagnostic thinking of traditional Chinese medicine is extracted.

[0030] In some implementations, rules for previous state-treatment-next state are extracted through frequent state sequence mining, and strong association rules that meet support and confidence thresholds are selected, including:

[0031] Count the frequency of each state in all state sequences from the state sequence database D, calculate the support of a single state, filter out states with support greater than the minimum support, and generate a frequent 1-sequence set L1.

[0032] Iteratively generate a candidate k-sequence set Lk, remove infrequent subsequences through pruning operations, calculate the support of candidate sequences, and filter candidate sequences with support greater than the minimum support to form a frequent k-sequence set Lk;

[0033] Based on the frequent k-sequence set Lk and the state sequence database D, the support and confidence of the rules are calculated, and strong association rules are generated by combining the efficacy feedback.

[0034] In some implementations, it also includes:

[0035] Using composite states as nodes and efficacy feedback as edge weights, visualization rules are generated; the visualization rules include state transition diagrams or flowcharts; the composite states include symptom patterns and prescription strategies.

[0036] Secondly, a dynamic efficacy feedback-based TCM sequential diagnosis and treatment logic mining system was disclosed, including:

[0037] The preprocessing module preprocesses TCM time-series diagnosis and treatment data, constructs a sequence of consultation pairs containing multiple visits by the patient, and each consultation pair includes the consultation status and efficacy feedback;

[0038] The state transition tuple generation module maps symptoms and prescriptions into structured vectors based on traditional Chinese medicine theory, generating state transition tuples.

[0039] The rule extraction module, in conjunction with therapeutic feedback information, calculates the dynamic evolution pattern of the composite state of the state transition tuple using the state transition probability formula; it uses a sequential rule mining algorithm to extract the rules of the composite state vector mapping, and combines an interpretable decision tree model to generate rules that conform to the diagnostic thinking of traditional Chinese medicine.

[0040] Thirdly, a computer storage medium is disclosed, on which a computer program is stored, which, when executed by a processor, implements the method for mining the sequential diagnosis and treatment logic of TCM with dynamic therapeutic feedback as described in any of the above.

[0041] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0042] This invention deeply integrates dynamic efficacy feedback with the sequential diagnosis and treatment logic of Traditional Chinese Medicine (TCM). Based on structured vector mapping and state transition probability calculation, it transforms scattered diagnosis and treatment records into a dynamically evolving "syndrome-prescription" knowledge network. It accurately quantifies the driving weight of efficacy feedback on syndrome differentiation decisions, overcoming the limitations of static association in traditional methods. Through sequential rule mining and interpretable decision tree models, it extracts high-confidence rules consistent with TCM thinking, such as "do not change the prescription if effective, change it if ineffective," and constructs a multi-dimensional linkage decision path library of "efficacy-symptom-prescription," providing clinicians with evidence-based real-time decision support. Attached Figure Description

[0043] Figure 1 A schematic diagram of the TCM sequential diagnosis and treatment logic mining method with dynamic efficacy feedback provided by the present invention.

[0044] Figure 2 This is a flowchart illustrating step S3 of the TCM sequential diagnosis and treatment logic mining method with dynamic efficacy feedback provided by the present invention.

[0045] Figure 3 This is a flowchart illustrating step S32 of the TCM sequential diagnosis and treatment logic mining method with dynamic efficacy feedback provided by the present invention.

[0046] Figure 4 This is a flowchart illustrating step S33 of the TCM sequential diagnosis and treatment logic mining method with dynamic efficacy feedback provided by the present invention.

[0047] Figure 5 The flowchart of step S333 of the dynamic efficacy feedback TCM sequential diagnosis and treatment logic mining method provided by the present invention is shown.

[0048] Figure 6 This is a schematic diagram of the decision tree for the TCM sequential diagnosis and treatment logic mining method with dynamic efficacy feedback provided by the present invention. Detailed Implementation

[0049] To better understand and implement this invention, the technical solutions in the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0050] The terms “comprising” and “having” and any variations thereof in this invention are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device that includes a series of steps or modules is not necessarily limited to those steps or modules that are explicitly listed, but may include other steps or modules that are not explicitly listed or that are inherent to such processes, methods, products or devices.

[0051] The embodiments of the present invention disclose a method for mining the sequential diagnosis and treatment logic of traditional Chinese medicine based on dynamic efficacy feedback, which extracts dynamic adjustment rules that are easy to understand and conform to the thinking of traditional Chinese medicine.

[0052] like Figure 1 As shown, this method includes:

[0053] Step S1: Preprocess TCM time-series diagnosis and treatment data to construct a sequence of consultation pairs containing multiple visits by the patient. Each consultation pair includes the status of two adjacent visits, efficacy feedback, and the status of the subsequent visit.

[0054] Traditional Chinese medicine (TCM) time-series diagnosis and treatment data are retrieved from the database. This TCM time-series diagnosis and treatment data includes EMPI (Enterprise Master Patient Index, a unique value for patient ID), consultation date, prescription, symptoms, and time-series medical record data of the previous consultation's efficacy evaluation.

[0055] The TCM time-series diagnosis and treatment data are preprocessed, including serialization and standardization. Serialization involves grouping the data according to EMPI and sorting each group in ascending order based on the date of visit, generating an ordered sequence of visits Vp={Visit 1, Visit 2, ..., Visit k} for each patient, where p is the patient and k is the total number of visits for that patient.

[0056] Standardization involves standardizing the names of Chinese herbal medicines, symptom descriptions, and efficacy evaluations. A standard Chinese herbal medicine dictionary is used for mapping, unifying alternative names and variant spellings. Dosage information can be selectively extracted or standardized. A standard set of Chinese medicine symptom terms is used for mapping, handling synonyms and negative words. Efficacy evaluations are mapped to predefined ordered categorical variables, set as: {"Significantly Effective": 3, "Effective": 2, "Ineffective": 1, "Aggravated": 0,}.

[0057] Based on the preprocessed TCM time-series diagnosis and treatment data, a sequence of visit order pairs is constructed. For each patient p's visit order sequence Vp, if k>1, then k-1 visit order pairs Pair i = (Visit {i}, Visit {i+1}), where i ranges from 1 to k-1. Each visit order pair contains:

[0058] EMPI p

[0059] Symptoms Raw i, the original list of symptoms at the i-th visit;

[0060] Prescription Raw i, the original prescription list for the i-th visit;

[0061] Efficacy {i+1} is the structured evaluation of the efficacy of the i-th prescription recorded at the (i+1)-th visit.

[0062] Symptoms Raw {i+1}, the original list of symptoms at the (i+1)th visit;

[0063] Prescription Raw {i+1} represents the original prescription list for the (i+1)th visit.

[0064] The above-mentioned diagnosis sequence is stored in an intermediate database or data structure.

[0065] Step S2: Based on traditional Chinese medicine theory, map symptoms and prescriptions into structured vectors respectively, and generate state transition tuples.

[0066] Matching is performed using a predefined "symptom combination-syndrome" rule base based on rule mapping. For each consultation, the raw symptom list (Symptoms Raw) is converted into a structured symptom pattern representation (SymptomPattern). The output is the symptom pattern representation (SymptomPattern i) for the i-th consultation and the symptom pattern representation (SymptomPattern {i+1}) for the (i+1)-th consultation.

[0067] Based on the attribute mapping of Chinese herbal medicines in the Chinese Pharmacopoeia, which maps the four properties, five flavors, and meridian tropism of each herb in the prescription, the original prescription list Prescription Raw for each consultation is converted into a structured Chinese herbal medicine prescription representation PrescriptionStrategy. The i-th original prescription list PrescriptionStrategy i and the (i+1)-th original prescription list PrescriptionStrategy {i+1} are output.

[0068] Define the state of consultation i as State i = (SymptomPattern i, PrescriptionStrategy i).

[0069] For each consultation, generate a state transition tuple Transition i = (State i, Efficacy {i+1}, State {i+1}) for Pair i.

[0070] Step S3: Combining therapeutic feedback information, calculate the dynamic evolution pattern of the composite state using the state transition probability formula. For example... Figure 2 As shown, it includes the following steps:

[0071] Step S31: Based on the diagnosis-to-treatment sequence, calculate the composite state transition probability under different efficacy feedback and generate a state transition probability matrix;

[0072] The state transition probability under different efficacy evaluation conditions is calculated based on the state transition tuple Transition i. The composite state transition probability under different efficacy feedback is calculated using the following formula:

[0073] P(State {i+1} | State i, Efficacy {i+1}) = Count(State i, Efficacy {i+1}, State {i+1}) / Sum {all S'} Count(State i, Efficacy {i+1}, S')

[0074] Where P is the state transition probability, State i is the state at the i-th time, Efficacy {i+1} is the structured evaluation at the (i+1)-th time, and S' is the set of all possible next states under the current state State i and efficacy Efficacy {i+1}.

[0075] A state transition probability matrix is ​​constructed based on the state transition probabilities. The nodes of the state transition probability matrix are states, the edges are transitions, and the weights of the edges represent the state transition probabilities. The matrix is ​​displayed hierarchically according to efficacy.

[0076] Step S32: Extract rules from the previous state-treatment-next state using a frequent state sequence mining algorithm, and filter out strongly associated rules that meet the support and confidence thresholds. For example... Figure 3 As shown, it specifically includes:

[0077] Step S321: Count the frequency of each state in all state sequences from the state sequence database D, calculate the support of a single state, filter the states with support greater than the minimum support, and generate a set of frequent 1-sequences L1.

[0078] Step S322: Iteratively generate a candidate k-sequence set Lk, remove infrequent subsequences through pruning operations, calculate the support of candidate sequences, and filter candidate sequences with support greater than the minimum support to form a frequent k-sequence set Lk;

[0079] Step S323: Based on the frequent k-sequence set Lk and the state sequence database D, calculate the rule support and confidence, and generate strong association rules by combining efficacy feedback.

[0080] Specifically, each patient p corresponds to a state sequence Sp, which is composed of states in chronological order, Sp =<State {p,1}, State {p,2}, ..., State {p,k}> The state {p,i} is defined as (SymptomPattern {p,i}, PrescriptionStrategy {p,i}). A state sequence database D is constructed based on the state sequences Sp of multiple patients. A clinic-to-clinic transition dataset T is constructed, which consists of the set of state transition tuples Transition i from step S2, and is used to calculate the support and confidence of the rules in subsequent calculations.

[0081] A frequent state sequence mining algorithm is used to mine the state sequence database and the diagnosis-transformation dataset. Frequent state sequences are identified from the state sequences. The efficacy feedback is used as a condition variable to extract the rules of previous state-efficacy-subsequent state, and strong association rules that meet the support and confidence thresholds are selected.

[0082] First, the frequency of each single state X in all state sequences is counted from the state sequence database D. For example, patient A's state sequence might be "Qi deficiency and blood stasis → Phlegm-dampness obstruction → Qi and Yin deficiency". All patients' state sequences are scanned, and the independent frequency of each single state (such as "Qi deficiency and blood stasis") is counted. States with support exceeding a preset threshold are selected, forming a set of frequent 1-sequences. Their support is then calculated.

[0083] Support(State X)=Count(Sequences containing State X) / Total number of sequences in D

[0084] Here, Count(Sequences containing State X) refers to the number of patient state sequences that contain at least one State X.

[0085] Assign a set of frequent sequences L1 of length 1 to all states, State X, that satisfy Support(State X) >= min sup. L1 = {<State X> | Support(State X)>= min sup}

[0086] Next, candidate 2-sequences are generated through join operations. For example, "Qi deficiency and blood stasis" and "phlegm-dampness obstruction" from the frequent 1-sequences are combined into the candidate sequence "Qi deficiency and blood stasis → phlegm-dampness obstruction". A candidate k-sequence set Ck is generated based on L{k-1}. For generating C2, any two frequent states from the frequent sequence set L1 are selected.<State A> and<State B> Generate all possible 2-sequence candidates<State A, State B> For k > 2, joining State 1, State 2 ∈ L {k-1} requires that State 1, after removing the first element, equals State 2, after removing the last element; that is, removing duplicate elements before joining. For example, if...<A, B> and<B, C> In L2, candidates are generated.<A, B, C> .

[0087] These candidate sequences are pruned, removing those containing infrequent subsequences. For example, if "phlegm-dampness obstruction → qi and yin deficiency" is infrequent, its parent sequence is removed. For each generated candidate k-sequence c∈Ck, check if all its (k-1)-length subsequences are in L{k-1}. If any subsequence is not in L{k-1}, remove candidate c from Ck. For example, for candidate...<A, B, C> It needs to be checked.<A, B> and<B, C> Are they all in candidate sequence L2?

[0088] The state sequence database D is rescanned, and the state sequences Sp of all patients are traversed. The support of candidate sequences is calculated, and sequences that meet the threshold are retained as frequent k-sequences. The support is calculated using the following formula:

[0089] Support(c) =Count(Sequences containing c as a subsequence) / Totalnumber of sequences in D

[0090] Wherein, Sequences containing c as a subsequence are state sequences containing c as a subsequence, and Total number of sequences in D is the total number of sequences in the state sequence database D.

[0091] This process iterates until no higher-order frequent sequences can be generated. Its principle is based on the "prior property" of the Apriori algorithm—if a sequence is infrequent, its supersequence must also be infrequent, thus significantly reducing the generation of invalid candidates. Each element in L2 is of the form...<State A, State B> This indicates that State A followed by State B is a relatively frequent pattern. Hierarchical filtering significantly reduces computational complexity while preserving statistically significant high-frequency patterns, providing a reliable foundation for subsequent rule extraction. For example, if the support for "liver qi stagnation → qi stagnation and blood stasis" reaches 3%, it indicates that this pattern repeatedly occurs in clinical practice, potentially reflecting a common pathogenesis evolution pattern.

[0092] Since the rule format is State A -> [Efficacy] -> State B, the most important part is the frequent 2-sequence set L2. Each element in the sequence set L2 is of the form...<State A, State B> This indicates that State A being followed by State B is a relatively frequent pattern.

[0093] Based on the frequent k-sequence set Lk and the state sequence database D, rule support and confidence are calculated, and strong association rules are generated by combining efficacy feedback. The sequence set L2 is iterated, and for each frequent 2-sequence in sequence set L2...<State A, StateB> To obtain the efficacy value X for this state, where 3 = significantly effective, 2 = effective, 1 = ineffective, and 0 = worsened. In the clinic-to-clinic transformation dataset T, find the number of all tuples of the form (State A, Efficacy X, State B), representing the numerator of the rule State A -> [Efficacy=X] -> State B. Simultaneously, find the number of all tuples of the form (State A, Efficacy X, Any State) in the clinic-to-clinic transformation dataset T, representing the total number of times the premise State A appears and is immediately followed by an efficacy evaluation of X, used as the denominator for calculating the confidence score.

[0094] Rule support is calculated using the following formula:

[0095] Support(Rule:State A->[Efficacy=X]->State B) = Count(State A,Efficacy X, State B) / Total number of transitions in T

[0096] The total number of transitions in T is the total number of tuples in the transition dataset T.

[0097] Rule confidence is calculated using the following formula:

[0098] Confidence(Rule: StateA->[Efficacy=X]->State B)=Count(State A,Efficacy X, State B) / Count(State A, Efficacy X)

[0099] If Count(State A, Efficacy X) is 0, then the confidence level of the rule is undefined or is 0.

[0100] Retain all rules that simultaneously satisfy minimum rule support and minimum confidence to generate strongly correlated rules. Quantify the driving effect of therapeutic efficacy on state transitions by formulating conditional probabilities. For example:

[0101] The support for the rule "Liver Qi stagnation → [Efficacy = Ineffective] → Qi stagnation and blood stasis" was 1.2%, with a confidence level of 85%, indicating that this adjustment strategy has high credibility when the efficacy is ineffective.

[0102] The rule “Invigorating Qi and Blood → [Efficacy = Significant Effect] → Invigorating Qi and Blood (Standard Formula)” has a confidence level of 90%, directly verifying the clinical experience of “not changing the formula when it is effective” in traditional Chinese medicine.

[0103] Step S33: Construct an interpretable decision tree model to predict subsequent prescription strategies based on efficacy feedback and previous state, extracting strategies that conform to the rules of TCM syndrome differentiation. For example... Figure 4 As shown, it includes the following steps:

[0104] Step S331: Extract the defined features and target variables from the state transition tuple Transition i and construct the feature matrix; use the C4.5 algorithm, with information gain rate as the splitting criterion, and prioritize the node with the higher gain rate as the decision node;

[0105] From the set of state transition tuples Transition i mentioned above, extract the features and target variables defined above for each tuple to construct an MxN feature matrix X (M transition instances, N-1 features) and an Mx1 target vector Y. Randomly divide the dataset F into a training set Ftrain (70%) and a test set Ftest (30%), train on Ftrain, and evaluate on Ftest.

[0106] Configure and use the selected C4.5 algorithm library, and input the training set F train into the C4.5 algorithm, wherein the training set F includes feature X train and target Y train.

[0107] Calculate the entropy of the current node dataset S relative to the target variable PrescriptionStrategy {i+1}, namely Entropy(S).

[0108] Entropy(S) = - Σ [ P(cj) * log2(P(cj)) ]

[0109] Where P(cj) is the proportion of category cj (a certain PrescriptionStrategy {i+1}) in dataset S.

[0110] For each candidate feature A (e.g., SymptomPattern i, Efficacy {i+1}, etc.), calculate its information gain Gain(S, A) and split information SplitInfo(S, A).

[0111] Gain(S, A) = Entropy(S) - Σ [ (|Sv| / |S|) * Entropy(Sv) ]

[0112] Where Sv is a subset of feature A in S with value v.

[0113] SplitInfo(S, A) = - Σ[ (|Sv| / |S|) * log2(|Sv| / |S|)]

[0114] That is, for all possible values ​​v of feature A.

[0115] Calculate the gain ratio GainRatio(S, A) for each feature A:

[0116] GainRatio(S, A) = Gain(S, A) / SplitInfo(S, A)

[0117] If SplitInfo(S, A) is 0, then the gain rate is also 0, or special processing is required.

[0118] The feature A* with the highest gain ratio is selected as the splitting criterion for the current node. Based on different values ​​of A*, the dataset S is divided into subsets Sv. For each non-pure subset Sv (containing multiple target categories), the C4.5 algorithm is recursively called to construct a subtree. All samples in a node belong to the same category or there are no remaining features available for splitting.

[0119] Step S332: Remove overfitting branches using a pessimistic pruning algorithm, retaining decision paths that conform to clinical experience.

[0120] Step S3321: Calculate the pessimistic error estimate of the subtree, the adjusted error number of the subtree, and the error number of leaf nodes after pruning;

[0121] Step S3322: Compare the number of errors in the adjusted subtree with the number of errors in the leaf nodes after pruning. If the number of errors in the leaf nodes after pruning is not greater than the number of errors in the adjusted subtree, then perform the pruning operation.

[0122] Starting from the bottom of the tree, visit each non-leaf node (Internal Node) upwards. For each non-leaf node T, consider pruning it, that is, replacing the entire subtree rooted at T with a leaf node.

[0123] To calculate the pessimistic error estimate of a subtree, let Subtree(T) denote the entire subtree rooted at node T. Traverse all leaf nodes in Subtree(T). For each leaf node L in Subtree(T): N(L): the total number of training samples reaching that leaf node. E(L): the number of training samples misclassified at that leaf node.

[0124] The Pessimistic Error Count for Subtree is calculated by using Continuity Correction to adjust the error count, adding 0.5 error to each leaf node to compensate for the uncertainty in estimating the generalization error from the training data.

[0125] Adjusted Errors(Subtree(T)) = Σ[ E(L) + 0.5 ] , which sums the values ​​of all leaf nodes L in Subtree(T).

[0126] Calculate the pessimistic error estimate after pruning node T: If node T is replaced with a leaf node, the predicted class of that leaf node will be the most numerous class among the training samples arriving at node T. Let N(T) be the total number of training samples arriving at node T.

[0127] Let E'(T) be the number of training samples misclassified at node T if node T becomes a leaf node predicting its majority class. Apply continuity correction to calculate the pruned PessimisticError Count (Pruned) for node T:

[0128] Adjusted Errors(T as Leaf) = E'(T) + 0.5

[0129] Compare the pessimistic error count of the retained subtree (Adjusted Errors(Subtree(T))) with the pessimistic error count of pruning node T to a leaf node (Adjusted Errors(T as Leaf)). Compare the adjusted error count of the subtree with the pruned leaf node error count. If the pruned leaf node error count is not greater than the adjusted error count of the subtree, perform the pruning operation. Otherwise, do not prune, and retain node T and its subtree.

[0130] Step S333: Extraction rules based on the decision tree model, such as... Figure 5 As shown, it includes:

[0131] Step S3331: Determine the efficacy split point in the decision tree and analyze the first branch and second branch corresponding to the efficacy split point; the first branch is the branch with good efficacy, and the second branch is the branch with poor efficacy.

[0132] Step S3332: Compare the differences between the first branch and the second branch, and count the proportion of the prescription rule in the first branch and the proportion of the prescription change rule in the second branch to quantify the driving weight of the efficacy on the diagnosis and decision-making.

[0133] Step S3333: Remove branches with support below a preset threshold through pessimistic pruning, and extract the IF-THEN rule set that conforms to the TCM syndrome differentiation thinking.

[0134] Specifically, starting from the root node of the pruned decision tree, the path is traversed to each leaf node. Each path corresponds to a rule. Each internal node on the path represents an AND clause in the IF condition, including features and values / ranges. The leaf node represents the prediction result of the THEN part: PrescriptionStrategy {i+1}.

[0135] For example, IF PrescriptionStrategy i = "Invigorate Qi and Strengthen Spleen" AND Efficacy {i+1} ="Ineffective" AND SymptomPattern {i+1} = "Damp-Heat Accumulation" THEN PrescriptionStrategy {i+1} = "Clear Heat, Eliminate Dampness, and Strengthen Spleen" IF PrescriptionStrategy i = "Activate Blood Circulation and Remove Blood Stasis" AND Efficacy {i+1} ="Significantly Effective" THEN PrescriptionStrategy {i+1} = "Activate Blood Circulation and Remove Blood Stasis" (Preservative Principle)

[0136] When observing the decision tree, focus on features (SymptomPattern i, PrescriptionStrategy i, Efficacy {i+1}, SymptomPattern {i+1}) appearing near the root. Identify rule patterns representing common dialectical adjustment strategies, such as: maintaining the original prescription, adding ingredients, changing the prescription, and targeting concurrent syndromes.

[0137] For example, for a node that splits on Efficacy_{i+1}, it is necessary to clarify under what conditions efficacy is considered. This requires tracing back along the path from the root node of the tree to this efficacy split node. The splitting conditions of all other nodes along the path collectively define the "context" or "preconditions" of this node.

[0138] For example: Suppose the path is IF PrescriptionStrategy_i == "Soothe the liver and relieve depression" AND SymptomPattern_i == "Liver Qi stagnation", and it encounters a split in Efficacy_{i+1}. This means: For patients who were previously treated with the "Soothe the liver and relieve depression" strategy for "Liver Qi stagnation", how will subsequent decisions differ based on their feedback on the treatment effect? ​​The path determines the context of this efficacy.

[0139] Continue down the branch representing "good efficacy" (e.g., Efficacy{i+1}>1.5, Efficacy{i+1}== "significantly effective", or Efficacy{i+1}== "effective"). Observe whether this branch splits further, and if so, based on what feature (e.g., Symptom Pattern{i+1} or subdivision of Prescription Strategy i). What Prescription Strategy{i+1} is predicted by the final leaf node? Record which subsequent prescription strategies the model mainly recommends under the "good efficacy" path.

[0140] Similarly, continue down the branch representing "poor efficacy," such as Efficacy{i+1} <= 1.5, Efficacy{i+1} == "ineffective," or Efficacy{i+1} == "worsening." Observe the subsequent splits of this branch to see if the features used are different from those of the "good efficacy" branch. What kind of structured TCM prescription Prescription Strategy{i+1} is predicted by the final leaf node? Record which subsequent prescription strategies the model mainly recommends under this "poor efficacy" path.

[0141] like Figure 6 As shown, given that the previous prescription strategy was "tonifying Qi and promoting blood circulation," the model determines the next prescription strategy (PrescriptionStrategy+1) based on efficacy feedback (Efficacyi+1) and the current symptom pattern (SymptomPatterni+1). The purple-background boxes in the figure...<divclass="node"> This represents a decision node and includes a conditional statement. (Green background box) The leaf node represents the final prediction result, i.e., `PrescriptionStrategy_{i+1}`. Lines represent decision branches, and the accompanying text, such as "Yes (poor efficacy)" or "No (good efficacy)," explains the conditions for choosing that branch. The entire decision tree structure is read from top to bottom and from left to right.

[0142] Find the node in the graph that makes a judgment based on Efficacy_{i+1}. In this example, it is node B. Its condition is Efficacy_{i+1} <= 1.5 ?, used to distinguish between "poor efficacy" (ineffective / worsening) and "good efficacy" (effective / significant effect). This split occurs under the premise that the "invigorating qi and promoting blood circulation" strategy was used in the previous instance. Analyze the branches with good efficacy and poor efficacy, and compare the differences between the two branches.

[0143] Starting from node B in the graph, selecting "No," i.e., path BD, and after the "good therapeutic effect" condition (Efficacy{i+1}) is met, a leaf node D is immediately reached. The prediction of leaf node D is "invigorating Qi and promoting blood circulation." This indicates that when the previous "invigorating Qi and promoting blood circulation" strategy was effective or significant, the best practice learned by the model is to continue using this strategy (defensive approach). The model believes that in this case, there is no need to make further judgments based on the current symptom pattern (SymptomPattern{i+1}).

[0144] Starting from node B in the diagram, selecting "Yes" leads to paths BCE, FG, or H. After Efficacy{i+1} determines "poor efficacy," this path doesn't immediately reach a conclusion but continues splitting. The next split is based on the current symptom pattern SymptomPattern_{i+1}: node C indicates phlegm-dampness obstruction, and node F indicates worsening qi stagnation. Depending on the current symptoms, different prescription strategies are ultimately adopted: if it's phlegm-dampness obstruction, node E is predicted, focusing on strengthening the spleen, resolving phlegm, and promoting blood circulation; if it's worsening qi stagnation, node G is predicted, focusing on strengthening qi and promoting blood circulation. In other cases, node H is predicted, adjusting the qi-tonifying and blood-activating formula. This indicates that when the previous "qi-tonifying and blood-activating" strategy is ineffective or worsens the condition, the model no longer adheres to the original formula. The model believes a reassessment of the current specific condition is necessary; SymptomPattern_{i+1} becomes the key factor determining the next strategy, and poor efficacy triggers re-diagnosis. The final strategy is to adjust or change the formula based on the current diagnosis.

[0145] Comparing the two branches, the "good efficacy" path is short and direct, while the "poor efficacy" path is long and requires more judgment. When the efficacy is "good," the efficacy itself is the main determining factor; when the efficacy is "poor," the efficacy is a trigger, and the current symptoms become the main determining factor for the subsequent course of events. "Good efficacy" tends to maintain the original strategy, while "poor efficacy" tends to change / adjust the strategy.

[0146] Analyzing this decision tree clearly shows that the efficacy feedback (Efficacy_{i+1}) acts as a "decision logic converter" in the model. Based on the effectiveness of the previous treatment, it guides decisions to different branches: one is a simple path that tends towards "maintaining the status quo" (adhering to the original approach), and the other is a complex path that triggers "reassessment and strategy adjustment." In this complex path, the current symptom pattern (SymptomPattern_{i+1}) takes over from the efficacy feedback, becoming the key to determining the specific direction of adjustment. This intuitively reflects the TCM thinking mode of dynamically adjusting diagnosis and treatment based on efficacy feedback.

[0147] The performance of the obtained decision tree model is evaluated. The predictive ability and generalization of the pruned C4.5 decision tree model are verified using multi-dimensional metrics. The test set Ftest, containing features Xtest, is input into the trained model, and the output prediction result Ypred is compared with the true label Ytest to calculate the following metrics:

[0148] 1. Overall accuracy: The proportion of correctly predicted samples out of the total number of samples, reflecting the model's global predictive ability.

[0149] 2. Confusion Matrix: Displays the number of true positives (TP), false positives (FP), true negatives (TN), and false negatives (FN) in each category in matrix form, revealing the model's misclassification patterns in specific categories. For example, "Qi deficiency and blood stasis" is misclassified as "phlegm and dampness obstruction".

[0150] 3. Precision: {Precision} = TP / (TP + FP), which measures the proportion of results that the model predicts as belonging to a certain category that are actually correct. It avoids over-adjusting prescriptions, such as the risk of "promoting blood circulation and removing blood stasis" being incorrectly recommended.

[0151] 4. Recall: TP / (TP + FN) assesses the model's ability to capture real cases and reduce the risk of missed diagnoses, such as "qi stagnation and blood stasis" not being identified.

[0152] 5. F1 score: 2 (Precision Recall) / (Precision + Recall), which comprehensively balances precision and recall, and is especially suitable for scenarios where TCM syndrome categories are unbalanced, such as robustness assessment when there are few "significantly effective" cases.

[0153] Step S4: Using the composite state as the node and the efficacy feedback as the edge weight, generate a state transition diagram or flowchart.

[0154] By using composite states, such as combinations of symptom patterns and prescription strategies, as nodes and efficacy feedback as edge weights, a directed graph structure is constructed based on a probability matrix to generate a state transition graph that intuitively displays the evolution path of high-frequency states.

[0155] Sequential rules or decision tree logic are decomposed into multi-level judgment nodes through flowcharts, such as efficacy branches and syndrome conditions, transforming complex rules into interactive graphical paths, such as "ineffective treatment → syndrome differentiation of phlegm and dampness → modified heat-clearing formula". When selecting typical patient cases, the complete state trajectory is traced back, and the syndrome differentiation logic of each step is marked, forming a visual report of the diagnosis and treatment path. For example, initial diagnosis "liver qi stagnation" → follow-up diagnosis "qi stagnation and blood stasis", "ineffective treatment triggers enhanced qi-regulating and blood-activating formula"). The graphical interface lowers the understanding threshold of TCM dynamic syndrome differentiation logic, assisting doctors to quickly locate key adjustment nodes; quantifying probability and rule confidence improves decision credibility, and high-probability paths can be prioritized in practical applications; typical cases verify the consistency between the system and clinical practice, promoting the implementation of experience inheritance and intelligent assisted diagnosis and treatment.

[0156] This invention, by systematically integrating efficacy feedback information with TCM time-series diagnostic and treatment data, achieves for the first time the quantitative mining and visualization of dynamic syndrome differentiation logic. The technical solution transforms traditional static and fragmented diagnostic and treatment records into a dynamically interconnected knowledge network, revealing the patterns of the "syndrome differentiation-prescription" state evolving with efficacy between treatment sessions. It quantifies the impact of efficacy on subsequent decisions through probability calculations and rule extraction; for example, the probability of maintaining the prescription in "significantly effective" cases reaches 90%, while the probability of changing the prescription in "ineffective" cases exceeds 70%. At the clinical level, the system can provide doctors with references for syndrome differentiation adjustments in similar cases, reducing decision-making uncertainty. In the research field, it provides data support for theoretical research on syndrome evolution and prescription compatibility (such as mining the common transformation path of "damp-heat → phlegm-stasis"). Furthermore, the generated state transition diagrams and decision tree flowcharts, combined with the full-cycle state evolution trajectory analysis of typical medical cases, intuitively recreate the dynamic syndrome differentiation thinking of TCM, forming a reusable teaching case library to facilitate the inheritance of clinical experience and the innovation of TCM education models.

[0157] This application also provides a dynamic efficacy feedback-based TCM sequential diagnosis and treatment logic mining system, including:

[0158] The preprocessing module preprocesses TCM time-series diagnosis and treatment data, constructs a sequence of consultation pairs containing multiple visits by the patient, and each consultation pair includes the consultation status and efficacy feedback;

[0159] The state transition tuple generation module maps symptoms and prescriptions into structured vectors based on traditional Chinese medicine theory, generating state transition tuples.

[0160] The rule extraction module, in conjunction with therapeutic feedback information, calculates the dynamic evolution pattern of the composite state of the state transition tuple using the state transition probability formula; it uses a sequential rule mining algorithm to extract the rules of the composite state vector mapping, and combines an interpretable decision tree model to generate rules that conform to the diagnostic thinking of traditional Chinese medicine.

[0161] This invention deeply integrates dynamic efficacy feedback with the sequential diagnosis and treatment logic of Traditional Chinese Medicine (TCM). Based on structured vector mapping and state transition probability calculation, it transforms scattered diagnosis and treatment records into a dynamically evolving "syndrome-prescription" knowledge network. It accurately quantifies the driving weight of efficacy feedback on syndrome differentiation decisions, overcoming the limitations of static association in traditional methods. Through sequential rule mining and interpretable decision tree models, it extracts high-confidence rules consistent with TCM thinking, such as "do not change the prescription if effective, change it if ineffective," and constructs a multi-dimensional linkage decision path library of "efficacy-symptom-prescription," providing clinicians with evidence-based real-time decision support.

[0162] Based on the same inventive concept, the present invention also provides a computer device, comprising: a processor and a memory; wherein the memory stores a computer program adapted to be loaded by the processor and executed by the processor to perform the steps of the above-described method for mining the sequential diagnosis and treatment logic of TCM with dynamic therapeutic feedback.

[0163] The processing methods for computer devices can be referred to the description of the methods above, and will not be repeated here.

[0164] This application also provides a non-transitory machine-readable storage medium storing an executable program, which, when run by a microprocessor, causes the processor to execute the method provided in the above embodiments.

[0165] This invention discloses a computer-readable storage medium storing a computer program for electronic data interchange, wherein the computer program causes a computer to perform the described methods.

[0166] This invention discloses a computer program product including a non-transitory computer-readable storage medium storing a computer program operable to cause a computer to perform the described method.

[0167] The embodiments described above are merely illustrative. The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0168] Through the detailed description of the above embodiments, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, including read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), one-time programmable read-only memory (OTPROM), electrically-Erasable Programmable Read-Only Memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, disk storage, magnetic tape storage, or any other computer-readable medium that can be used to carry or store data.

[0169] Finally, it should be noted that the embodiments disclosed in this invention are merely preferred embodiments of this invention and are only used to illustrate the technical solutions of this invention, not to limit it. Although this invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this invention.

Claims

1. A method for mining the sequential diagnosis and treatment logic of Traditional Chinese Medicine based on dynamic efficacy feedback, characterized in that, include: Preprocess TCM time-series diagnosis and treatment data to construct a sequence of consultation pairs containing multiple visits by patients. Each consultation pair includes the consultation status and efficacy feedback. Based on traditional Chinese medicine theory, symptoms and prescriptions are mapped into structured vectors to generate state transition tuples; By combining therapeutic feedback information, the dynamic evolution pattern of the composite state of the state transition tuple is calculated using the state transition probability formula; the sequential rule mining algorithm is used to extract the rules of the composite state vector mapping, and the rules that conform to the TCM syndrome differentiation thinking are generated by combining the interpretable decision tree model. Combining therapeutic feedback information, the dynamic evolution pattern of composite states is calculated using the state transition probability formula, including: calculating the state transition probability under different therapeutic feedback based on the consultation pair sequence, and generating a state transition probability matrix; extracting rules of previous state-therapeutic effect-subsequent state through a frequent state sequence mining algorithm, and screening strong association rules that meet the support and confidence thresholds; constructing an interpretable decision tree model to predict subsequent prescription strategies with therapeutic feedback and previous state as features, and extracting rules that conform to the TCM syndrome differentiation thinking. The probability of state transition under different therapeutic responses is calculated using the following formula: P(State {i+1}|State i, Efficacy {i+1}) = Count(State i, Efficacy {i+1},State {i+1}) / Sum {all S'} Count(State i, Efficacy {i+1}, S') Where P is the state transition probability, State i is the state at the i-th time, Efficacy {i+1} is the structured evaluation at the (i+1)-th time, and S' is the set of all possible next states under the current state State i and efficacy Efficacy {i+1}.

2. The method for mining the sequential diagnosis and treatment logic of traditional Chinese medicine based on dynamic efficacy feedback according to claim 1, characterized in that, An interpretable decision tree model is constructed to predict subsequent prescription strategies using efficacy feedback and previous states as features, and extracts features that conform to the rules of TCM syndrome differentiation, including: Extract defined features and target variables from the diagnosis pair transformation dataset T, and construct a feature matrix; Using the C4.5 algorithm, information gain and split information are calculated. Based on the information gain rate as the splitting criterion, nodes with higher gain rates are selected as decision nodes to construct a decision tree. Overfitting branches are removed by pessimistic pruning algorithm, while decision-making paths that conform to clinical experience are preserved; Rules that conform to the diagnostic thinking of traditional Chinese medicine are extracted based on the decision tree model.

3. The method for mining the sequential diagnosis and treatment logic of traditional Chinese medicine based on dynamic efficacy feedback according to claim 2, characterized in that, Overfitting branches are removed using a pessimistic pruning algorithm, preserving decision paths that align with clinical experience, including: Calculate the pessimistic error estimate of the subtree, the adjusted error number of the subtree, and the error number of leaf nodes after pruning; Compare the adjusted error count of the subtree with the adjusted error count of the leaf nodes after pruning. If the adjusted error count of the leaf nodes after pruning is not greater than the adjusted error count of the subtree, then perform the pruning operation.

4. The method for mining the sequential diagnosis and treatment logic of traditional Chinese medicine based on dynamic efficacy feedback according to claim 1, characterized in that, Extract rules that conform to the diagnostic thinking of Traditional Chinese Medicine, including: Identify the efficacy split point in the decision tree and analyze the first and second branches corresponding to the efficacy split point; the first branch is the branch with good efficacy, and the second branch is the branch with poor efficacy. Compare the differences between the first and second branches, and statistically analyze the proportion of the adherence rule in the first branch and the proportion of the substitution rule in the second branch to quantify the driving weight of the therapeutic effect on the diagnostic decision-making. By pessimistically pruning branches with support below a preset threshold, the IF-THEN rule set that conforms to the diagnostic thinking of traditional Chinese medicine is extracted.

5. The method for mining the sequential diagnosis and treatment logic of traditional Chinese medicine based on dynamic efficacy feedback according to claim 4, characterized in that, Rules for previous state-treatment-next state are extracted through frequent state sequence mining, and strong association rules that meet support and confidence thresholds are selected, including: Count the frequency of each state in all state sequences from the state sequence database D, calculate the support of a single state, filter out states with support greater than the minimum support, and generate a frequent 1-sequence set L1. Iteratively generate a candidate k-sequence set Lk, remove infrequent subsequences through pruning operations, calculate the support of candidate sequences, and filter candidate sequences with support greater than the minimum support to form a frequent k-sequence set Lk; Based on the frequent k-sequence set Lk and the state sequence database D, the support and confidence of the rules are calculated, and strong association rules are generated by combining the efficacy feedback.

6. The method for mining the sequential diagnosis and treatment logic of traditional Chinese medicine based on dynamic efficacy feedback according to claim 5, characterized in that, Also includes: Using composite states as nodes and efficacy feedback as edge weights, visualization rules are generated; the visualization rules include state transition diagrams or flowcharts; the composite states include symptom patterns and prescription strategies.

7. A dynamic efficacy feedback-based TCM sequential diagnosis and treatment logic mining system, characterized in that, include: The preprocessing module preprocesses TCM time-series diagnosis and treatment data, constructs a sequence of consultation pairs containing multiple visits by the patient, and each consultation pair includes the consultation status and efficacy feedback; The state transition tuple generation module maps symptoms and prescriptions into structured vectors based on traditional Chinese medicine theory, generating state transition tuples. The rule extraction module, in conjunction with therapeutic feedback information, calculates the dynamic evolution pattern of the composite state of the state transition tuple using the state transition probability formula; it uses a sequential rule mining algorithm to extract the rules of the composite state vector mapping, and combines an interpretable decision tree model to generate rules that conform to the diagnostic thinking of traditional Chinese medicine. Combining therapeutic feedback information, the dynamic evolution pattern of composite states is calculated using the state transition probability formula, including: calculating the state transition probability under different therapeutic feedback based on the consultation pair sequence, and generating a state transition probability matrix; extracting rules of previous state-therapeutic effect-subsequent state through a frequent state sequence mining algorithm, and screening strong association rules that meet the support and confidence thresholds; constructing an interpretable decision tree model to predict subsequent prescription strategies with therapeutic feedback and previous state as features, and extracting rules that conform to the TCM syndrome differentiation thinking. The probability of state transition under different therapeutic responses is calculated using the following formula: P(State {i+1}|State i, Efficacy {i+1}) = Count(State i, Efficacy {i+1},State {i+1}) / Sum {all S'} Count(State i, Efficacy {i+1}, S') Where P is the state transition probability, State i is the state at the i-th time, Efficacy {i+1} is the structured evaluation at the (i+1)-th time, and S' is the set of all possible next states under the current state State i and efficacy Efficacy {i+1}.

8. A computer storage medium, characterized in that, It stores a computer program, which, when executed by a processor, implements the steps of the TCM sequential diagnosis and treatment logic mining method with dynamic efficacy feedback as described in any one of claims 1-6.

Images