Sports medicine intelligent agent diagnosis system

Abstract

The invention relates to a sports medicine intelligent agent diagnosis system, and the system comprises a data obtaining unit which is used for obtaining the multi-modal diagnosis data of a patient; the memory graph unit is used for constructing a memory graph by taking the multi-modal diagnosis data of a plurality of historical patients as nodes and taking comprehensive association weights among the data as connecting edges; the matching node screening unit is used for screening to obtain a matching node set through a synapse-like propagation mechanism based on the matching nodes obtained through preliminary screening and comprehensive association weights among the nodes in the memory map; and the diagnosis result output unit is used for performing membrane potential accumulation activation on each matching node through an LIF activation method based on a dynamic forgetting mechanism, and obtaining a diagnosis result of the to-be-diagnosed patient based on the diagnosis and treatment result corresponding to the activated matching node. The problem that in the prior art, a sports medicine diagnosis system does not consider the correlation of time sequences between cases, and consequently the accuracy of a diagnosis result is poor is solved.

Description

Technical Field

[0001] This invention relates to the field of clinical medical technology, and in particular to a sports medicine intelligent body diagnostic system. Background Technology

[0002] In recent years, sports injuries have become increasingly common, including anterior cruciate ligament (ACL) reconstruction, rotator cuff repair, and meniscus repair. Postoperative rehabilitation training is crucial for these patients. However, recovery speed and physical condition vary greatly among patients, and traditional, standardized rehabilitation training models often fail to accurately adapt to each individual's situation. Furthermore, existing sports medicine diagnostic systems typically only consider the semantic similarity between patients and cases in their memory database, neglecting the temporal correlation of symptom development between cases; or, during diagnosis and treatment, they rely heavily on the physician's experience, leading to limited diagnostic accuracy and potentially resulting in unsatisfactory rehabilitation outcomes or even secondary injuries. Summary of the Invention

[0003] Based on the above analysis, the present invention aims to provide a sports medicine intelligent body diagnostic system; it solves the problem that existing sports medicine diagnostic systems do not consider the correlation between cases over time, resulting in poor diagnostic accuracy.

[0004] The objective of this invention is mainly achieved through the following technical solutions:

[0005] This invention discloses a sports medicine intelligent body diagnostic system, comprising:

[0006] The data acquisition unit is used to acquire multimodal diagnostic data from patients.

[0007] The memory graph unit is used to construct a memory graph by using the high-dimensional feature representations corresponding to the multimodal diagnostic data of multiple historical patients as nodes and the comprehensive correlation weights between the high-dimensional feature representations as connecting edges; the multimodal diagnostic data of the historical patients are associated with the corresponding diagnosis and treatment results of each patient.

[0008] The matching node filtering unit is used to perform preliminary filtering of matching nodes based on the semantic similarity and temporal correlation strength between the high-dimensional feature representation of the patient to be diagnosed and the high-dimensional feature representation of each node in the memory graph; and based on the matching nodes obtained from the preliminary filtering and the comprehensive correlation weight between nodes in the memory graph, a set of matching nodes is obtained through a synaptic-like propagation mechanism.

[0009] The diagnostic result output unit is used to activate each matching node in the matching node set by accumulating membrane potential using the LIF activation method based on the dynamic forgetting mechanism, and to obtain the diagnostic result of the patient to be diagnosed based on the diagnosis and treatment results corresponding to the activated matching nodes.

[0010] Furthermore, the multimodal diagnostic data includes multiple key events related to the patient's disease and the time of occurrence of each key event; features are extracted from the multimodal diagnostic data, and each key event is sorted according to its occurrence time to obtain a high-dimensional feature representation of each multimodal diagnostic data, including the time series of key events.

[0011] Furthermore, the memory map is constructed using the following method:

[0012] Iterate through and obtain the semantic similarity between the high-dimensional feature representations of any two historical patients;

[0013] Elastic matching is performed on the high-dimensional feature representations corresponding to any two historical patients to obtain the temporal correlation strength between any two high-dimensional feature representations.

[0014] Based on the temporal association strength and semantic similarity, the comprehensive association weight of any two corresponding nodes is obtained;

[0015] Two nodes whose overall association weight is greater than a preset association threshold are connected by an edge, and the association weight is used as the edge weight to obtain the memory graph.

[0016] Furthermore, the matching node filtering unit obtains the matching node set by filtering using the following method:

[0017] Semantic matching is performed between the high-dimensional feature representation corresponding to the patient to be diagnosed and the high-dimensional feature representation of each node in the memory graph to obtain candidate associated nodes whose feature similarity is greater than a preset similarity threshold.

[0018] Based on the temporal correlation strength between the high-dimensional feature representation of the patient to be diagnosed and each candidate associated node, candidate associated nodes below the preset temporal correlation threshold are filtered to obtain the preliminary screening of matching nodes.

[0019] The initially selected matching nodes are used as target association nodes in the first layer and activated. Using the target association nodes in the first layer as starting nodes, a synaptic-like propagation mechanism is used to propagate the stimulus signal hierarchically along the connection edges of the memory map. Based on a preset activation threshold, multiple target association nodes corresponding to each layer are obtained. Based on the target association nodes corresponding to each layer, the matching node set is obtained.

[0020] Furthermore, the initially selected matching nodes are used as target association nodes in the first layer and activated. Using these target association nodes as starting nodes, a synaptic-like propagation mechanism is employed to propagate stimulus signals hierarchically along the connection edges of the memory map. Based on a preset activation threshold, multiple target association nodes corresponding to each layer are obtained. Based on the target association nodes corresponding to each layer, the matching node set is obtained, including:

[0021] S1: Set an initial stimulus intensity for each target-associated node in the first layer; and take each target-associated node in the first layer as a parent node, set the inactive nodes adjacent to each parent node as child nodes, and take the preset initial stimulus intensity of each target-associated node in the first layer as the current stimulus intensity.

[0022] S2: Based on the current stimulus intensity and the comprehensive association weight of the connection edge between the parent node and the child node, perform synaptic propagation on the activation intensity to obtain the attenuated stimulus intensity of each child node, and filter the child nodes whose stimulus intensity is greater than the activation threshold to obtain the target associated node of the next layer.

[0023] S3: Take each target associated node in the next layer as the parent node, take the inactive nodes adjacent to each parent node as child nodes, and take the stimulation intensity of each target associated node in the next layer as the current stimulation intensity. Repeat S2 until the stimulation intensity of each inactive node decays to below the activation threshold. A set of matching nodes is obtained based on the target associated nodes in each layer.

[0024] Furthermore, in synaptic-like propagation, if the stimulus propagation from multiple parent nodes points to the same child node, the optimal parent node is selected using the following formula:

[0025]

[0026] Where, p * For the optimal parent node, S p The stimulus intensity of parent node p, P score (p,c) represents the comprehensive association weight between node p and node c, where P i (c) is the set of all possible parent nodes corresponding to child node c.

[0027] Furthermore, the activation intensity is propagated in a synaptic-like manner to obtain the attenuated stimulation intensity of each child node, which is expressed as:

[0028]

[0029] in, Let be the stimulus intensity of the j-th node in the (i+1)-th layer. The comprehensive association weight between the k-th node in layer i and the j-th node in layer (i+1) Let be the stimulus intensity of the k-th node in the i-th layer.

[0030] Furthermore, the step of activating the matching nodes in the matching node set by accumulating membrane potential using the LIF activation method based on a dynamic forgetting mechanism includes:

[0031] Initialize the input current and membrane potential of each matched node;

[0032] The key event time series corresponding to the patient to be diagnosed and the key event time series corresponding to each matching node are traversed. If a matching key event and occurrence time are found, a stimulation pulse is generated and a continuous input current is formed.

[0033] The input current is converted into a membrane potential, and the membrane potential is accumulated; both the input current and the membrane potential decay based on a dynamic time constant; when the membrane potential of the matching node reaches the discharge threshold, the matching node is activated to discharge.

[0034] Furthermore, the generation of stimulation pulses and the formation of a continuous input current are expressed as follows:

[0035]

[0036] The membrane potential is accumulated and expressed as:

[0037]

[0038] Among them, I X (t) represents the input current of node X at time t; τ X (t) is the time constant of current decay at node X at time t; S X (t s ) represents node X at time t s The intensity of the stimulus received; δ(tt) s ) is the Dirac function, Γ X τ represents the set of timestamps for the discrete stimulus pulses received by node X; V (t) is the decay constant of the membrane potential at time t; V X (t) represents the membrane potential of node X at time t.

[0039] Furthermore, the dynamic time constant is expressed as:

[0040]

[0041] Where, τ X(V) (t) represents τ V (t) or τ X (t), τ X (t) is the time constant of current decay at node X at time t; τ V (t) is the decay constant of the membrane potential at time t; Δt is the time interval between two stimulation pulses.

[0042] The present invention can achieve at least the following beneficial effects:

[0043] 1. This invention incorporates the temporal pattern of disease occurrence and development into the calculation of inter-node association weights when constructing a memory graph. This weighting method, which combines time and semantics, ensures that each edge in the graph simultaneously reflects the medical semantic relevance and temporal similarity of cases. Compared to existing association methods that only consider semantic similarity, this invention can more accurately capture the temporal patterns and associations of disease evolution in sports medicine scenarios, thereby improving the accuracy and reliability of relevant case retrieval.

[0044] 2. In the matching node screening process, this invention adaptively limits the propagation depth and breadth through layer-by-layer attenuation, ensuring that only medically relevant paths can effectively transmit signals, thereby activating memory nodes truly related to the current case. This avoids noise introduced into the retrieval process due to excessively long paths or strained relationships, helping to reduce interference from irrelevant cases. Secondly, this mechanism allows valuable information to propagate indirectly through multiple intermediate case nodes, enabling the system to not only be limited to directly similar cases but also to discover related case chains spanning one or two intermediary links, thus expanding the retrieval scope. This propagation mechanism greatly improves the accuracy and biological rationality of the system in retrieving relevant cases in complex memory maps.

[0045] 3. This invention introduces a cumulative triggering model from the biological brain by adopting a LIF neural activation model based on dynamic time constants, thereby improving the reliability of diagnostic decisions. Nodes must accumulate sufficient multimodal evidence to cross the activation threshold, significantly reducing the risk of drawing conclusions based on a single anomaly and ensuring that diagnostic conclusions are well-supported. Furthermore, dynamic forgetting and sensitivity adjustment are achieved based on dynamic time constants, balancing new and old information. Traditional inference models often process evidence with a fixed forgetting rate or fixed time window, which may lead to information overload or forgetting of important clues. This invention, by dynamically adjusting the time constant, enables nodes to maintain high sensitivity to rapid, continuous information flows while retaining memory of information with long intervals. This adaptive mechanism improves the system's robustness in processing medical data at different time scales.

[0046] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings. Attached Figure Description

[0047] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.

[0048] Figure 1 This is a schematic diagram of a sports medicine intelligent body diagnostic system according to an embodiment of the present invention;

[0049] Figure 2 This is a schematic diagram of the matching node filtering process according to an embodiment of the present invention. Detailed Implementation

[0050] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0051] This invention provides a sports medicine intelligent body diagnostic system, such as... Figure 1 As shown, it includes:

[0052] The data acquisition unit is used to acquire multimodal diagnostic data from patients.

[0053] The memory graph unit is used to construct a memory graph by using the high-dimensional feature representations corresponding to the multimodal diagnostic data of multiple historical patients as nodes and the comprehensive correlation weights between the high-dimensional feature representations as connecting edges; the multimodal diagnostic data of the historical patients are associated with the corresponding diagnosis and treatment results of each patient.

[0054] The matching node filtering unit is used to perform preliminary filtering of matching nodes based on the semantic similarity and temporal correlation strength between the high-dimensional feature representation of the patient to be diagnosed and the high-dimensional feature representation of each node in the memory graph; and based on the matching nodes obtained from the preliminary filtering and the comprehensive correlation weight between nodes in the memory graph, a set of matching nodes is obtained through a synaptic-like propagation mechanism.

[0055] The diagnostic result output unit is used to activate each matching node in the matching node set by accumulating membrane potential using the LIF activation method based on the dynamic forgetting mechanism, and to obtain the diagnostic result of the patient to be diagnosed based on the diagnosis and treatment results corresponding to the activated matching nodes.

[0056] Specifically, the multimodal diagnostic data includes multiple key events related to the patient's disease and the time of occurrence of each key event; in particular, the key events in this embodiment include disease-related symptoms, locations of occurrence, and related events extracted from medical records (such as symptoms such as knee pain, swelling, and limited function, and events related to knee joint diseases such as collisions and falls), as well as examination data obtained through data acquisition devices (such as MRI, CT, ultrasound images, electromyography (EMG), and electrocardiogram (ECG) data).

[0057] The data acquisition unit in this embodiment includes motion sensors, a machine vision system, a heart rate monitor, medical imaging equipment (such as MRI, CT, and ultrasound), and bioelectrical signal acquisition equipment (such as electromyography (EMG) and electrocardiography (ECG). Among these, the patient can wear an IMU sensor to monitor joint range of motion and movement trajectory in real time. The machine vision system can be installed at the rehabilitation site for motion capture. In addition, the data acquisition unit is also used to acquire subjective assessment questionnaires filled out by the patient periodically and the patient's medical records to collect relevant data such as the patient's symptoms.

[0058] After acquiring the multimodal diagnostic data, the method further includes extracting features from the multimodal diagnostic data separately through a feature extraction unit, and sorting and fusing the extracted feature representations according to the occurrence time of the corresponding key events to obtain high-dimensional feature representations of each multimodal diagnostic data, including the time series of key events.

[0059] The diagnostic system in this embodiment acquires multimodal diagnostic data from patients through a standard medical data interface. Image data (such as MRI, CT, and ultrasound) can be retrieved via the DICOM protocol interface of PACS; text-based medical record data (such as doctor's diagnostic records and examination reports) is extracted from EMR through the HL7 / FHIR interface of the hospital information system; wearable device data (such as motion sensors and heart rate monitors) is uploaded via an IoT platform or a dedicated application. For different types of acquired data, the system performs corresponding preprocessing: for example, medical images undergo format conversion and normalization (DICOM files are decoded into pixel matrices, and resolution / grayscale is adjusted); medical record texts undergo word segmentation, entity recognition, and other cleaning; physiological signals undergo denoising and normalization. The preprocessed modal data are then input into a feature extraction unit (deep learning model) to extract feature representations. For example, image data can be processed by a trained convolutional neural network to extract visual features of lesion areas; text data can be encoded using a medical language model to extract symptoms and diagnostic points; and time-series signals can be processed by a recurrent neural network or Transformer model to extract trend features. Subsequently, the system fuses feature representations from different modalities at a high level to form a unified high-dimensional feature representation (embedding). Thus, the data for each patient case is transformed into a comprehensive feature vector in the same semantic space, ensuring that content from different information sources can be directly compared and correlated.

[0060] Furthermore, the memory map unit is constructed to obtain the memory map using the following method:

[0061] Iterate through and obtain the semantic similarity between the high-dimensional feature representations of any two historical patients;

[0062] Elastic matching is performed on the high-dimensional feature representations corresponding to any two historical patients to obtain the temporal correlation strength between any two high-dimensional feature representations.

[0063] Based on the temporal association strength and semantic similarity, the comprehensive association weight of any two corresponding nodes is obtained;

[0064] Two nodes whose overall association weight is greater than a preset association threshold are connected by an edge, and the association weight is used as the edge weight to obtain the memory graph.

[0065] It should be noted that the high-dimensional feature representation corresponding to each patient is stored as a case memory node in the system's memory graph unit. Each node represents a case memory unit, corresponding to a single diagnosis and treatment process of a historical patient case (including images, examination results, symptom descriptions, and the time of occurrence). The high-dimensional feature representation maintained within the node includes a series of key events and timestamps for the case (e.g., symptom onset time, injury date, and time points of each examination and treatment), representing the time series information of key events related to the disease. During the operation of the diagnostic system, new case nodes can be continuously added to the memory graph unit, allowing the knowledge base of the memory graph unit to be gradually expanded with clinical use. For common sports medicine injuries, corresponding knowledge nodes (e.g., typical injury cases and their standard treatment protocols) can also be pre-imported into the memory graph and associated with actual case nodes to provide prior expert knowledge support.

[0066] This embodiment considers both semantic similarity and temporal correlation strength between case memory nodes to determine the weights of edges between nodes. By introducing a temporal correlation triggering mechanism, Dynamic Time Warping (DTW) can be used to measure the correlation strength of the high-dimensional feature representations of two case memory nodes in the temporal dimension. Each case node contains a series of time-ordered "stimulus" events (i.e., key events, such as the time of symptom onset, the time of obtaining examination results, etc.), which can be regarded as a memory pulse sequence. The two memory sequences are elastically matched using the DTW algorithm to calculate their cumulative correlation score L(n,m). Subsequently, the score is normalized to the (0,1] interval using the Sigmoid function to obtain the temporal correlation strength, expressed as:

[0067]

[0068] Simultaneously, the semantic similarity score C between each node is calculated. score (A,B), for example, cosine similarity can be used to measure the similarity of the high-dimensional feature representations corresponding to the multimodal data of two cases.

[0069] Then, the temporal association strength and semantic similarity are multiplied to obtain the comprehensive association weight P. score (a, B) is represented as:

[0070]

[0071] In constructing the memory graph, this invention assigns a weighted edge to each pair of case nodes, with the weight being the comprehensive association weight P. score The value reflects the semantic relevance of two nodes in terms of disease and the temporal similarity of events. If two case nodes are not semantically related enough (e.g., their symptoms or imaging features are too different), no connection will be formed between them, or the connection weight will be very low and ignored. If two case nodes are semantically similar but completely mismatched in terms of time series (e.g., the progression and rhythm of disease development), the final weight will also be significantly reduced, and the connection may be ignored. For example, even if two cases of anterior cruciate ligament injury have similar symptoms, if one is an acute trauma and the other is a chronic degeneration, their time series patterns will be very different. The temporal association strength (Tscore) between the former and the latter will be very low, resulting in a low overall association weight (Pscore), and they will not be considered closely related nodes in the graph. Compared with traditional methods that rely purely on semantic similarity, the time-semantic fusion connection weights in this embodiment can better capture the temporal patterns and associations of disease occurrence and development in sports medicine, improving the accuracy of sports medicine diagnosis.

[0072] Furthermore, the matching node filtering unit obtains the matching node set by filtering using the following method:

[0073] Semantic matching is performed between the high-dimensional feature representation corresponding to the patient to be diagnosed and the high-dimensional feature representation of each node in the memory graph to obtain candidate associated nodes whose feature similarity is greater than a preset similarity threshold.

[0074] Based on the temporal correlation strength between the high-dimensional feature representation of the patient to be diagnosed and each candidate associated node, candidate associated nodes below the preset temporal correlation threshold are filtered to obtain the preliminary screening of matching nodes.

[0075] The initially selected matching nodes are used as target association nodes in the first layer and activated. Using the target association nodes in the first layer as starting nodes, a synaptic-like propagation mechanism is used to propagate the stimulus signal hierarchically along the connection edges of the memory map. Based on a preset activation threshold, multiple target association nodes corresponding to each layer are obtained. Based on the target association nodes corresponding to each layer, the matching node set is obtained.

[0076] The method of propagating stimulus signals hierarchically along the connection edges of the memory map using a synaptic-like propagation mechanism includes:

[0077] S1: Set an initial stimulus intensity for each target-associated node in the first layer; and take each target-associated node in the first layer as a parent node, set the inactive nodes adjacent to each parent node as child nodes, and take the preset initial stimulus intensity of each target-associated node in the first layer as the current stimulus intensity.

[0078] S2: Based on the current stimulus intensity and the comprehensive association weight of the connection edge between the parent node and the child node, perform synaptic propagation on the activation intensity to obtain the attenuated stimulus intensity of each child node, and filter the child nodes whose stimulus intensity is greater than the activation threshold to obtain the target associated node of the next layer.

[0079] S3: Take each target associated node in the next layer as the parent node, take the inactive nodes adjacent to each parent node as child nodes, and take the stimulation intensity of each target associated node in the next layer as the current stimulation intensity. Repeat S2 until the stimulation intensity of each inactive node decays to below the activation threshold. A set of matching nodes is obtained based on the target associated nodes in each layer.

[0080] More specifically, the synaptic propagation mechanism is a brain-inspired node activation method that employs a layer-by-layer propagation of stimulus signals to retrieve and activate the nodes most relevant to the input case within a constructed memory map. At each layer of propagation, the stimulus signal attenuates according to the weights between nodes (similar to the signal transmission and attenuation process between neurons).

[0081] First, the semantic similarity between the input case (the high-dimensional feature representation corresponding to the diagnostic data of the patient to be diagnosed) and the high-dimensional feature representations of each node in the memory graph is calculated, and the semantic similarity exceeding the preset threshold cosine similarity is selected. th The candidate node set is used for initial screening, ensuring that subsequent propagation occurs only on memory nodes semantically relevant to the current case. Building upon this, a temporal association triggering mechanism is introduced for further filtering: for each semantic candidate node, the temporal association strength (Tscore) between it and the current case in the time series pattern is obtained using the DTW method. If the association strength is lower than a set temporal association threshold, the candidate node is removed. Through dual semantic and temporal filtering, memory nodes that match the current case (the case to be diagnosed) in both medical context and temporal progression can be selected, avoiding the misselection of irrelevant cases based solely on superficial similarity.

[0082] The initial relevant nodes obtained through double filtering are considered as the first-layer target-related node set N1. Each target-related node is assigned an initial stimulus intensity, such as S{X1} = 1, simulating the unit stimulus generated by the current case input. Subsequently, the stimulus signal is propagated outward in the memory map in a hierarchical manner: the target-related node in the i-th layer (called the "parent node") will pass the stimulus to the adjacent child node in the next layer through its memory map connection. Parent node X i The propagation strength between a candidate child node and a given child node is determined by the combined association weight of the two nodes, expressed as:

[0083]

[0084] in, Let be the stimulus intensity of the j-th node in the (i+1)-th layer. The comprehensive association weight between the k-th node in layer i and the j-th node in layer (i+1) Let be the stimulus intensity of the k-th node in the i-th layer.

[0085] By iteratively applying the above formula, the stimulus intensity received by the nth layer node along the propagation path can be expressed as the result after layer-by-layer weight decay:

[0086]

[0087] This demonstrates that the longer the path or the smaller the correlation weight along the way, the more exponentially the stimulus will weaken when it reaches the terminal node. This attenuation characteristic of synaptic transmission ensures that only nodes that pass through highly correlated pathways (i.e., each transmission is medically closely related) can receive sufficient stimulation to be activated, reducing interference from irrelevant memories.

[0088] As mentioned earlier, this embodiment only allows nodes whose stimulus intensity exceeds the threshold but have not yet been activated to continue propagating outward as parent nodes, i.e., the parent node must meet the following conditions:

[0089]

[0090] Among them, P i Let p represent the set of qualified parent nodes at level i, p be the candidate parent nodes, Fire represent the set of activated nodes, and stim represent the set of nodes at level i. th N is the activation threshold. i This represents the set of target associated nodes at level i. Parent nodes that meet the specified conditions will be used to generate the candidate child node set for the next level.

[0091] For each parent node, select all nodes whose semantic similarity exceeds the similarity threshold cosine. th And the inactive nodes are selected as the candidate child node set C. i (p), represented as:

[0092]

[0093] That is, select nodes that have a sufficiently high cosine similarity to the parent node and have not yet appeared in the activated layer as candidate child nodes.

[0094] To ensure that each child node receives only a single-path stimulus from its most relevant parent node, preventing multiple paths from repeatedly activating or overactivating the same node, and thus maintaining the biological rationality of the propagation chain and the reliability of the results, if multiple parent nodes point to the same child node, this embodiment selects only the node that provides the most effective stimulus as the optimal parent node. Through optimal parent node selection, only the optimal parent node p... * It will pass its stimulus to child node c, represented as:

[0095]

[0096] Among them, P i (c) is the set of all possible parent nodes.

[0097] After performing the above steps, the set of child nodes in the next layer that actually receive stimulation can be determined. This set is the union of all child nodes activated after being selected by the optimal parent node. The child node sets of each parent node constructed by the method in this embodiment do not overlap, ensuring that each node in each layer has a unique origin.

[0098] Furthermore, the diagnostic result output unit activates the matching nodes in the matching node set by accumulating membrane potential using a LIF activation method based on a dynamic forgetting mechanism, including:

[0099] Initialize the input current and membrane potential of each matched node;

[0100] The key event time series corresponding to the patient to be diagnosed and the key event time series corresponding to each matching node are traversed. If a matching key event and occurrence time are found, a stimulation pulse is generated and a continuous input current is formed.

[0101] Specifically, in embodiments of the present invention, the occurrence time of critical events is represented using a relative time method. For example, the consultation time of the patient to be diagnosed is used as a baseline time point (denoted as 0), and the time difference between the occurrence time of the critical event and this baseline time point is used as the event's time label value: events occurring before the consultation are recorded as negative values ​​(e.g., 3 days before the consultation is recorded as -3 days), and events occurring after the consultation are recorded as positive values ​​(e.g., 5 days after the consultation is recorded as +5 days). Through this relative time labeling method, the difference in absolute dates between different patients can be eliminated, enabling the system to match events based on their relative temporal relationship in the course of disease development, thereby solving the problem of incomparable temporal sequences caused by inconsistent actual dates of critical events for different patients.

[0102] Furthermore, in this embodiment, key events include, but are not limited to: disease-related symptoms (such as knee pain, swelling, and limited function) obtained from medical records, patient reports, or monitoring devices; injury-related triggering events (such as falls or collisions); and examination results (such as positive MRI scans or abnormal synovial fluid analysis). After acquiring the above key events, the system sorts them according to relative time to form a key event time sequence based on the time of consultation. For example, if a patient falls and injures their knee 3 days before consultation, and develops knee swelling 2 days before consultation, then the patient's key event time sequence can be represented as: -3 days: fall, knee injury; -2 days: knee swelling; 0 days: consultation.

[0103] During the diagnostic reasoning process, the system iterates through the key event time series of the patient to be diagnosed and the key event time series corresponding to each matching node in the memory map. If a matching item exists with both the event type and the relative occurrence time being consistent, a stimulation pulse is generated at the corresponding matching node, forming a continuous input current. This input current is accumulated using the LIF activation method of the dynamic forgetting mechanism to form a membrane potential. When the membrane potential reaches a preset discharge threshold, the matching node is activated and discharged, and the diagnostic result corresponding to that node is output. By employing a relative time alignment and time pattern matching mechanism, this invention can accurately capture the common temporal characteristics of diseases such as sports injuries in different patients' disease courses, improving the accuracy and reliability of disease diagnosis.

[0104] The input current is converted into a membrane potential, and the membrane potential is accumulated; both the input current and the membrane potential decay based on a dynamic time constant; when the membrane potential of the matching node reaches the discharge threshold, the matching node is activated to discharge.

[0105] Specifically, during the membrane potential accumulation activation process, each node receives cumulative stimuli from key events in the patient's input case. To simulate the response characteristics of biological neurons to continuous stimulation, this invention introduces a dynamic leaky integral discharge model (dynamic LIF model) on the memory nodes, treating node activation as a neuron-like process of membrane potential accumulation and discharge. When the accumulated stimulation at a node reaches a threshold, a "discharge" is triggered, meaning the node is activated and can be used as a candidate diagnostic node.

[0106] In the dynamic LIF model, each memory node X maintains a time-varying input current I. X (t) and membrane potential V X (t). Discrete input stimuli (from each relevant key event) will be converted into continuous input currents, described in differential equation form as follows:

[0107]

[0108] Among them, I X (t) represents the input current of node X at time t; τ X S is the time constant for the decay of the current at node X; X (t s ) represents node X at time t s The intensity of the received stimulus (event intensity, such as X1 = 1.0); δ(tt) s ) represents the Dirac function; Γ represents the set of timestamps of discrete stimulus pulses received by the node.

[0109] Specifically, in the dynamic LIF model, stimulus intensity is used to characterize the contribution weight of a key event to the accumulation process of the membrane potential at the matching node. The stimulus intensity is directly related to the event type; different types of key events have different diagnostic values ​​in the course of disease development, and therefore, the corresponding stimulus intensity values ​​are also different. For example, in the diagnostic scenario of anterior cruciate ligament injury or postoperative rehabilitation, "significant joint swelling before consultation" can be assigned a high stimulus intensity value because it usually indicates an acute or subacute inflammatory response; "high-intensity running and jumping activity before consultation" indicates a history of joint stress and can be assigned a moderate intensity value; "mild discomfort without significant limitation of movement" has a lower diagnostic contribution and can be assigned a lower intensity value.

[0110] In practical applications, the values ​​can be obtained in the following ways: First, statistical analysis can be performed based on a large amount of historical clinical case data in sports medicine, combining the frequency of occurrence of the event in the process of diagnosing a specific disease, its correlation with the diagnostic results, and its medical importance, and then normalized to determine the values; Second, initial parameter values ​​can be directly set by clinical medical experts based on their past clinical experience during system initialization; In addition, the system can also dynamically optimize the stimulus intensity through machine learning methods during operation to adapt to the characteristic differences of different patient groups and disease stages.

[0111] During the diagnostic reasoning process, when a critical event of a patient to be diagnosed matches a critical event of a matching node in both event type and relative occurrence time, the system injects the stimulus intensity value corresponding to that event into the node's input current calculation formula as an input quantity, and marks the timestamp of the event's occurrence with a Dirac function, thereby generating discrete pulses at the corresponding moments and converting them into continuous input current. Different stimulus intensity values ​​corresponding to different event types will directly affect the accumulation rate of membrane potential and the timing of reaching the discharge threshold, thus enabling priority response to high diagnostic value events during the diagnostic process.

[0112] Through the dynamic LIF activation mechanism, discrete stimulus pulses are injected via a Dirac function to form a continuous input current, exhibiting exponential leakage. Each instantaneous stimulus pulse received by a node will increase the current I. XThe current increases suddenly, while it decays exponentially when there is no stimulus input (from -I). X (t) simulates the leakage effect of ion channels in biological neurons. This ensures that the influence of previous evidence gradually diminishes, thus simulating the natural forgetting of medical memories, while keeping nodes sensitive to new inputs.

[0113] Where the Dirac function δ(tt) s () is an idealized definition of a function, with the following characteristics: at t = t s At that time, δ(tt) s It is infinite; it is 0 at other times; the integral over the entire time domain is 1, which can be expressed as:

[0114]

[0115] During LIF activation, within a matching node, for each key event (e.g., joint swelling, positive test report, worsening symptoms, abnormal imaging findings, etc.) in the high-dimensional feature representation corresponding to the multimodal diagnostic data from the patient to be diagnosed, if a key event is related to a key event in the matching node, a specific stimulus pulse will be generated. Each stimulus pulse has a specific timestamp t. s (The moment the event occurs). After receiving multiple discrete stimulus pulses (i.e., each pulse occurs at a specific time t), the node... s (Generated), through integral conversion, a series of discrete stimulation pulses are converted into a continuously varying input current I over time. X (t).

[0116] Furthermore, the membrane potential V of the node X (t) reflects its intrinsic activation state, influenced by both the accumulation of input current and its own forgetting. The change in membrane potential follows the following differential equation:

[0117]

[0118] Where, τ V (t) is the decay constant of the membrane potential at time t; V X (t) represents the membrane potential of node X at time t; I X (t) represents the input current at node X at time t. Where -V X The (t) term ensures that the membrane potential does not accumulate indefinitely and become uncontrollable (corresponding to continuous natural decay), while I X The (t) term converts the continuous input current into an increase in membrane potential. Due to the introduction of current buffering, the rise in membrane potential is smooth relative to the stimulus pulse, thereby achieving selective activation of the matching node, that is, the membrane potential can only cross the threshold to trigger activation when continuous stimulation accumulates to a sufficient intensity.

[0119] In practical calculations, through physical modeling and parameter normalization, membrane potential and input current can be mapped to a unified dimensional system, such as being represented as dimensionless normalized values ​​(e.g., mapped to real values ​​in the range [0,1] or [-1,1]). Their relative magnitudes reflect the activation state and stimulus intensity of the node during the diagnostic reasoning process.

[0120] The membrane potential decay constant controls the natural decay rate of the membrane potential over time, corresponding to a normalized simulation of the membrane capacitance discharge process in biological neurons. The input current term has been dimensionally matched using a current-to-potential conversion coefficient, reflecting the immediate boost to the membrane potential from continuous stimulation triggered by key events. In the intelligent agent diagnostic system for sports medicine diseases of this invention, this conversion coefficient can be determined based on historical case data and the statistical characteristics of physiological signal amplitudes, ensuring that input stimuli from different sources and amplitudes are converted into normalized values ​​that can be directly added to the membrane potential before entering the membrane potential calculation formula.

[0121] However, with a fixed input current and membrane potential time constant τ X and τ V It is difficult to simultaneously achieve rapid response to new input and retention of long-term memory. Therefore, this invention proposes to... X and τ V Dynamic, adapting its value according to the temporal distribution of the stimulus, as expressed as:

[0122]

[0123] Where, τ X(V) (t) represents τ V (t) or τ X (t), τ X (t) is the time constant of current decay at node X at time t; τ V (t) is the decay constant of the membrane potential at time t; V X (t) represents the membrane potential of node X at time t; I X (t) represents the input current of node X at time t; Δt represents the time interval between two stimulus pulses.

[0124] This invention designs the current decay time constant and the membrane potential decay time constant as dynamically variable time constants, which are adaptively updated according to the time distribution of stimulus arrival. This simultaneously considers both sensitive response to rapid, continuous inputs and memory retention for inputs with long time intervals. When stimuli are very frequent (the time interval Δt between two stimuli is very small, meaning new stimuli arrive in succession), e -Δt When the value is close to 1, τ can be preserved through the mapping function. X and τ VIt remains unchanged near smaller values. Smaller τ X and τ V This indicates that the current and membrane potential decay rapidly, thus maintaining the node's high sensitivity to frequent inputs: each new stimulus causes the membrane potential to rise rapidly, but at the same time, old charges also leak out quickly, preventing excessive accumulation and premature triggering. When stimuli are very sparse (Δt is large, occurring only once every long period), e -Δt As the value approaches 0, the mapping function will gradually increase τ. X and τ V (Set the upper limit to 1, for example, initially τ starts from 0 and gradually approaches 1). Larger τ X and τ V This indicates that the current or membrane potential decays very slowly, effectively slowing down the rate of forgetting. Thus, even if there is a long interval between two stimulus pulses, the effect of the previous pulse can be retained longer, preventing it from completely disappearing before the next pulse arrives. Through this dynamic adjustment, the node can automatically adapt to the temporal pattern of the input stimulus: responding promptly to rapid stimuli and prolonging memory with slow stimuli. This approach of adjusting the forgetting rate based on Δt simulates the synaptic plasticity of biological neurons, giving the model of this invention both agility to frequent signals and persistence to isolated signals, significantly improving the intelligence of the reasoning process.

[0125] It should be noted that the input current reflects the sustained effect of the instantaneous stimulus intensity triggered by the critical event on the node, and its decay time constant τ X This invention is used to simulate the natural leakage process of postsynaptic currents in biological neurons. In this invention, the time constant is adaptively adjusted according to the time interval Δt between two stimulation pulses: when stimulation is very frequent (Δt is small), the time constant is kept at a small value, making the current highly sensitive to new inputs but decaying rapidly to prevent excessive accumulation; when the stimulation interval is long (Δt is large), the time constant is gradually increased to slow down the decay rate and retain the memory of isolated events.

[0126] Membrane potential represents the overall activation state of a node, formed by the accumulation of input currents from multiple events. Its decay time constant is used to simulate the discharge and fallback process of the cell membrane in the absence of continuous stimulation. Similar to the input current, the membrane potential decay time constant can also be designed to be dynamically variable to balance rapid response to continuous inputs and long-term maintenance of intermittent inputs. However, in some implementations, the membrane potential time constant can be set to a different adjustment range and update rate than the input current to more accurately simulate the information processing needs of "short-term symptom response" and "long-term disease memory" at different time scales in clinical sports medicine diagnosis.

[0127] In terms of mathematical formula structure, the input current and membrane potential time constants can employ similar adaptive update models (such as nonlinear mapping based on the interval Δt between two stimulations) to maintain consistency and maintainability in implementation. However, in parameter design, it is recommended to "set a smaller upper limit for the time constant and a higher decay rate for the input current" and "set a larger upper limit for the time constant and a lower decay rate for the membrane potential" to achieve a "fast response + slow forgetting dual-layer mechanism." For example, the time constant of the input current can be in the range [0.1, 0.5], while the membrane potential can be in the range [0.5, 1.0], and the slope parameter of the mapping function can also be different to distinguish the response sensitivity of the two.

[0128] In the sports medicine disease-related intelligent agent diagnostic system of the present invention, the dual-layer dynamic time constant mechanism can simultaneously achieve the following during the reasoning process: rapid response to acute symptom changes (such as sudden swelling within 48 hours after surgery); and long-term memory retention of sparse but important events in the course of the disease (such as abnormal MRI results 3 weeks after surgery).

[0129] Through this dual regulation, the system avoids noise accumulation due to frequent inputs and premature forgetting of key diagnostic clues due to excessively long event intervals when matching and activating case nodes, thereby improving the accuracy and robustness of diagnosis.

[0130] Furthermore, when the membrane potential V of a certain memory node X (t) When the level gradually increases with the accumulation of stimulation and exceeds a preset threshold, the node is considered activated and triggered, i.e., it enters the "discharge" state. Once the node discharges, its state will reset to the initial resting value (e.g., the membrane potential resets to V). rest The current is reset to I. rest A post-event reset mechanism prevents nodes from remaining in an active state for extended periods, thus avoiding the overwhelming of new evidence input by persistent activation of a particular case memory. For sports medicine diagnosis, this means that once a potential diagnosis is "triggered" with sufficient evidence, the system marks it as a confirmed diagnostic candidate and temporarily stops accumulating stimuli for it, focusing instead on accumulating evidence for other candidate conclusions. Only when a new key event is subsequently triggered can the node gradually reactivate. This design ensures the contextual relevance of the diagnostic process: each time, only the conclusion node that best fits the current comprehensive evidence is activated, allowing the system to dynamically adjust its confidence level in the diagnostic conclusion over time and with the arrival of new information.

[0131] After the LIF activation process described above, when one or more matching nodes are triggered, the system can generate corresponding diagnostic results and treatment recommendations. Typically, the memory graph pre-contains the treatment results of historical patients, i.e., the association between common sports medicine diseases and their recommended treatment plans. For example, a "ruptured anterior cruciate ligament" case node might be linked to a "surgical treatment" recommendation, and a "muscle strain" node might be linked to a "conservative treatment (rest and physical therapy)" recommendation. Once a memory node is activated, the system extracts the disease name or characteristics it represents and, based on the associations in the graph, finds the corresponding treatment recommendations as output to assist clinicians or patients in generating treatment plans.

[0132] In practical applications, when a node X enters the Fire state, the system retrieves the treatment plan information connected to that node and outputs it as a text-based treatment suggestion node for doctors' reference. If multiple nodes are activated simultaneously (e.g., multiple possible diseases coexist), the system can determine the priority based on the peak membrane potential or discharge order of each node, selecting the plan corresponding to the node with the higher peak potential or earlier activation as the primary suggestion, and listing the suggestions from other activated nodes as secondary options for reference, ensuring the objectivity and comprehensiveness of the treatment recommendations.

[0133] In summary, the memory graph unit of the sports medicine intelligent agent diagnostic system proposed in this invention incorporates the temporal pattern of disease occurrence and development into the correlation weights between nodes. This weight assignment method, combining time and semantics, ensures that each edge in the graph simultaneously reflects the medical semantic relevance and temporal similarity of cases. This allows for more accurate capture of the temporal patterns and correlations of disease evolution in sports medicine scenarios, thereby improving the accuracy and reliability of relevant case retrieval. Furthermore, during the node matching and screening process, the system adaptively limits the propagation depth and breadth through layer-by-layer attenuation, ensuring that only medically relevant paths can effectively transmit signals, thus activating memory nodes truly related to the current case. This avoids noise introduced during the retrieval process due to excessively long paths or strained relationships, helping to reduce interference from irrelevant cases. The system also incorporates a biological brain-like cumulative triggering model based on a LIF neural activation model with a dynamic time constant, improving the reliability of diagnostic decisions. Nodes must accumulate sufficient multimodal evidence to cross the activation threshold, significantly reducing the risk of drawing conclusions based on a single anomaly, ensuring that diagnostic conclusions are supported by sufficient evidence. Finally, dynamic forgetting and sensitivity adjustment based on the dynamic time constant achieve a balance between new and old information. Traditional inference models often process evidence with a fixed forgetting rate or a fixed time window, which may lead to information overload or forgetting important clues. This invention, however, dynamically adjusts the time constant, enabling nodes to maintain high sensitivity to rapid, continuous information flows while also retaining memory of information from long intervals. This adaptive mechanism improves the system's robustness in processing medical data at different time scales.

[0134] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0135] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A sports medicine intelligent body diagnostic system, characterized in that, include: The data acquisition unit is used to acquire multimodal diagnostic data from patients. The memory graph unit is used to construct a memory graph by using the high-dimensional feature representations corresponding to the multimodal diagnostic data of multiple historical patients as nodes and the comprehensive correlation weights between the high-dimensional feature representations as connecting edges; the multimodal diagnostic data of the historical patients are associated with the corresponding diagnosis and treatment results of each patient. The matching node filtering unit is used to perform preliminary filtering of matching nodes based on the semantic similarity and temporal correlation strength between the high-dimensional feature representation of the patient to be diagnosed and the high-dimensional feature representation of each node in the memory graph; and based on the matching nodes obtained from the preliminary filtering and the comprehensive correlation weight between nodes in the memory graph, a set of matching nodes is obtained through a synaptic-like propagation mechanism. The diagnostic result output unit is used to activate each matching node in the matching node set by accumulating membrane potential using the LIF activation method based on the dynamic forgetting mechanism, and to obtain the diagnostic result of the patient to be diagnosed based on the diagnosis and treatment results corresponding to the activated matching nodes.

2. The sports medicine intelligent body diagnostic system according to claim 1, characterized in that, The multimodal diagnostic data includes multiple key events related to the patient's disease and the time of occurrence of each key event; features are extracted from the multimodal diagnostic data, and each key event is sorted according to its occurrence time to obtain a high-dimensional feature representation of each multimodal diagnostic data, including the time series of key events.

3. The sports medicine intelligent body diagnostic system according to claim 2, characterized in that, The memory map was constructed using the following method: Iterate through and obtain the semantic similarity between the high-dimensional feature representations of any two historical patients; Elastic matching is performed on the high-dimensional feature representations corresponding to any two historical patients to obtain the temporal correlation strength between any two high-dimensional feature representations. Based on the temporal association strength and semantic similarity, the comprehensive association weight of any two corresponding nodes is obtained; Two nodes whose overall association weight is greater than a preset association threshold are connected by an edge, and the association weight is used as the edge weight to obtain the memory graph.

4. The sports medicine intelligent body diagnostic system according to claim 1, characterized in that, The matching node filtering unit obtains the set of matching nodes by filtering them using the following method: Semantic matching is performed between the high-dimensional feature representation corresponding to the patient to be diagnosed and the high-dimensional feature representation of each node in the memory graph to obtain candidate associated nodes whose feature similarity is greater than a preset similarity threshold. Based on the temporal correlation strength between the high-dimensional feature representation of the patient to be diagnosed and each candidate associated node, candidate associated nodes below the preset temporal correlation threshold are filtered to obtain the preliminary screening of matching nodes. The initially selected matching nodes are used as target association nodes in the first layer and activated. Using the target association nodes in the first layer as starting nodes, a synaptic-like propagation mechanism is used to propagate the stimulus signal hierarchically along the connection edges of the memory map. Based on a preset activation threshold, multiple target association nodes corresponding to each layer are obtained. Based on the target association nodes corresponding to each layer, the matching node set is obtained.

5. The sports medicine intelligent body diagnostic system according to claim 4, characterized in that, The process involves activating the initially selected matching nodes as target association nodes in the first layer, using the target association nodes in the first layer as starting nodes, and employing a synaptic-like propagation mechanism to propagate the stimulus signal hierarchically along the connection edges of the memory map, thereby obtaining multiple target association nodes corresponding to each layer based on a preset activation threshold. Based on the target associated nodes corresponding to each layer, the set of matching nodes is obtained, including: S1: Set an initial stimulus intensity for each target-associated node in the first layer; and take each target-associated node in the first layer as a parent node, set the inactive nodes adjacent to each parent node as child nodes, and take the preset initial stimulus intensity of each target-associated node in the first layer as the current stimulus intensity. S2: Based on the current stimulus intensity and the comprehensive association weight of the connection edge between the parent node and the child node, perform synaptic propagation on the activation intensity to obtain the attenuated stimulus intensity of each child node, and filter the child nodes whose stimulus intensity is greater than the activation threshold to obtain the target associated node of the next layer. S3: Take each target associated node in the next layer as the parent node, take the inactive nodes adjacent to each parent node as child nodes, and take the stimulation intensity of each target associated node in the next layer as the current stimulation intensity. Repeat S2 until the stimulation intensity of each inactive node decays to below the activation threshold. A set of matching nodes is obtained based on the target associated nodes in each layer.

6. The sports medicine intelligent body diagnostic system according to claim 5, characterized in that, In synaptic-like propagation, if the stimulus propagation from multiple parent nodes points to the same child node, the optimal parent node is selected using the following formula: Where, p * For the optimal parent node, S p The stimulus intensity of parent node p, P score (p,c) represents the comprehensive association weight between node p and node c, where P i (c) is the set of all possible parent nodes corresponding to child node c.

7. The sports medicine intelligent body diagnostic system according to claim 5, characterized in that, The activation intensity is propagated in a synaptic-like manner to obtain the attenuated stimulation intensity of each child node, which is expressed as: in, Let be the stimulus intensity of the j-th node in the (i+1)-th layer. The comprehensive association weight between the k-th node in layer i and the j-th node in layer (i+1) Let be the stimulus intensity of the k-th node in the i-th layer.

8. The sports medicine intelligent body diagnostic system according to claim 2, characterized in that, The step of activating the matching nodes in the matching node set by accumulating membrane potential using the LIF activation method based on a dynamic forgetting mechanism includes: Initialize the input current and membrane potential of each matched node; The key event time series corresponding to the patient to be diagnosed and the key event time series corresponding to each matching node are traversed. If a matching key event and occurrence time are found, a stimulation pulse is generated and a continuous input current is formed. The input current is converted into a membrane potential, and the membrane potential is accumulated; both the input current and the membrane potential decay based on a dynamic time constant; when the membrane potential of the matching node reaches the discharge threshold, the matching node is activated to discharge.

9. The sports medicine intelligent body diagnostic system according to claim 8, characterized in that, The generation of stimulation pulses and the formation of a continuous input current are expressed as follows: The membrane potential is accumulated and expressed as: Among them, I X (t) represents the input current of node X at time t; τ X (t) is the time constant of current decay at node X at time t; S X (t s ) represents node X at time t s The intensity of the stimulus received; δ(tt) s ) is the Dirac function, Γ X τ represents the set of timestamps for the discrete stimulus pulses received by node X; V (t) is the decay constant of the membrane potential at time t; V X (t) represents the membrane potential of node X at time t.

10. The sports medicine intelligent body diagnostic system according to claim 9, characterized in that, The dynamic time constant is expressed as: Where, τ X(V) (t) represents τ V (t) or τ X (t), τ X (t) is the time constant of current decay at node X at time t; τ V (t) is the decay constant of the membrane potential at time t; Δt is the time interval between two stimulation pulses.

Images