A neural rehabilitation clinical reasoning system and method based on a multi-layer bayesian causal network

Abstract

This invention discloses a neurorehabilitation clinical reasoning system and method based on a multi-layer Bayesian causal network. The system constructs a six-layer neurorehabilitation causal directed acyclic graph (NRI-DAG) containing 96 nodes and 155 directed edges. The six layers correspond to the lesion feature layer (L1), neurophysiological mechanism layer (L2), clinical syndrome layer (L3), compensatory pattern layer (L4), secondary injury layer (L5), and functional prognosis layer (L6), respectively. A hybrid conditional probability table construction method prioritizing literature anchors is employed. Unbiased quantification of intervention benefits is achieved through Pearl do-calculus surgery, and the leave-one-out method is used to rank the influence of evidence nodes. The system also integrates a geometric scoring module for nine compensatory movement patterns based on skeletal posture estimation. This invention provides an interpretable causal reasoning mechanism, quantifies the expected functional benefits of individualized intervention programs, solves the technical problem of insufficient interpretability in stroke rehabilitation clinical reasoning, and is applicable to neurorehabilitation clinical decision support.

Description

Technical Field

[0001] This invention belongs to the field of medical artificial intelligence technology, specifically involving a neurorehabilitation clinical reasoning system and method based on multi-layer Bayesian causal networks, which is applicable to interpretable reasoning assessment of post-stroke motor dysfunction, hierarchical identification of compensatory patterns, and decision support for individualized rehabilitation intervention. Background Technology

[0002] Stroke is one of the leading causes of adult disability worldwide, with more than 13 million new cases annually. Post-stroke motor dysfunction is the most common sequela, affecting approximately 80% of patients to varying degrees. Neurorehabilitation plays a crucial role in improving post-stroke motor function; however, the development of clinical treatment plans heavily relies on therapists' clinical experience, lacking systematic mechanistic analysis tools and quantitative decision-making criteria.

[0003] The neurophysiological mechanisms of post-stroke motor dysfunction involve the synergy and compensation of multiple motor control pathways: Corticospinal tract (CST) injury directly affects fine distal motor control; the reticulospinal tract (RST) plays a crucial ascending compensatory role in the recovery of compensatory motor function; the cortico-motor neuron system (CM) determines the voluntary movement ability of distal fingers; the basal ganglia-brainstem loop (BG-BS loop) regulates brainstem nuclei through basal ganglia output, influencing postural tone and the quality of anticipatory postural adjustment (APA); and the referent control framework explains the regulation mechanism of the tonic stretch reflex threshold (TSRT) and the occurrence of spasticity from the perspective of muscle free length regulation. Complex causal interactions exist among these pathways, and the degree of damage to each pathway exhibits a non-linear causal relationship with the patient's motor performance.

[0004] Existing stroke rehabilitation assessment tools suffer from the following technical deficiencies: First, tools based on traditional scales (such as the Fugl-Meyer Assessment Scale, Barthel Index, and Modified Ashworth Scale) only provide quantitative scores at a single time point, failing to reveal the causal relationship between neurophysiological mechanisms and clinical manifestations, and lacking mechanistic explanatory power. Second, predictive models based on machine learning (such as deep neural networks) are black-box characteristics, making it difficult for clinicians to understand their reasoning process and integrate AI recommendations into individualized treatment decisions. Third, compensatory movement patterns are a core clinical issue in stroke rehabilitation; excessive or persistent compensatory strategies can lead to secondary injuries such as learned disuse, shoulder pain, and joint contractures, but currently, there is a lack of technical means to systematically quantify the degree of activation of compensatory patterns and their causal chains. Fourth, existing tools cannot simulate counterfactual reasoning regarding "how functional prognosis will change after treatment intervention at a certain nerve or clinical node," failing to quantify the expected benefits of each candidate intervention at the individual patient level, thus hindering the realization of precision rehabilitation.

[0005] Bayesian Networks (BNs) are directed acyclic graph (DAG) models that combine probability theory and graph theory. They can represent the conditional independence structure between variables, support probabilistic inference under incomplete observations, and provide completely transparent and traceable causal chains. Pearl's do-calculus mathematically distinguishes between observational conditionation and causal intervention, overcoming the theoretical deficiency of traditional correlation models in accurately assessing intervention effects. However, current research on the application of Bayesian networks in neurorehabilitation has failed to simultaneously integrate multiple motor control theories to construct a unified six-layer causal model, achieve anchor point coverage of conditional probability tables based on evidence-based medicine literature, systematically detect compensatory patterns and vicious cycles, and quantify the individualized expected benefits of multiple intervention programs. [0005a] There are three unique technical obstacles to transferring Bayesian network technology to the field of clinical reasoning in neurorehabilitation, which prevent existing general-purpose Bayesian network technology from being directly applied to this field: Technical Obstacle 1: The Challenge of Heterogeneous Integration of Multiple Motor Control Theories. In the field of stroke rehabilitation, at least five independently developed motor control theories exist (CST / RST dual-pathway theory, Gracies spasticity pathophysiological model, Feldman reference control framework, BG-BS dual control system, and Bobath conceptual neurodevelopmental therapy). Each theory employs different conceptual frameworks and variable definitions, and there is no universal mapping relationship between them. Unifying these five theories into a single DAG with the assumption of compatibility and independence requires solving the semantic alignment problem across theoretical nodes—for example, the "stretch reflex threshold range (TSRT)" in the Gracies model and the "reference angle (μ)" in the Feldman reference control framework are mathematically equivalent but differ in operational measurement. A bridge needs to be established through the intermediate node L2_referent_control→L2_TSRT_range, which is a domain-specific engineering task that general BN construction tools cannot automatically complete. Technical Obstacle Two: There is no standard method for converting clinical literature to CPT parameters. Existing neurorehabilitation literature typically reports marginal probabilities (e.g., "the incidence of low CST function is 75-80% when there are large-volume lesions in the cortical motor area"), while Bayesian networks require conditional probabilities P(child node | all combinations of parent node states). For nodes with multiple parent nodes (e.g., L3APAintegrity has 6 parent nodes, generating 3...), ... 6 =729 state combinations), and the literature cannot cover all combinations; moreover, the research populations, measurement tools, and cutoff values ​​of different literatures are heterogeneous, and direct aggregation will introduce uncontrollable biases. The "maximum specificity literature anchor priority + synthetic prior completion" hybrid method proposed in this invention is a domain-specific CPT construction strategy designed to address the above challenges, and cannot be replaced by general Bayesian network learning algorithms. Technical Obstacle 3: Selection Bias in Intervention Benefit Estimation. In stroke rehabilitation assessment scenarios, the patient's evidence set (e.g., low spasticity) exhibits strong confounding associations with other favorable conditions (good CST integrity, high neuroplasticity). If traditional conditional probability P(prognosis|intervention node = target value, evidence) is used to estimate intervention benefits, confounding variables will systematically overestimate intervention benefits through common causal paths, leading to incorrect clinical decision prioritization. This invention employs Pearldo-calculus surgery (truncating all incoming edges of the intervention node) to achieve unbiased intervention benefit estimation at the sampling level, an operation not provided by existing general-purpose Bayesian inference tools and requiring specialized algorithm design. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a clinical reasoning system and method for neurorehabilitation based on multi-layer Bayesian causal networks, which simultaneously achieves: causal interpretable reasoning for neurorehabilitation assessment, hierarchical identification and quantification of compensation patterns, unbiased causal estimation of intervention benefits, and leave-one-out quantitative analysis of evidence information.

[0007] The technical solution of the present invention to solve the above-mentioned technical problems is as follows: A clinical reasoning method for neurorehabilitation based on multi-layer Bayesian causal networks includes the following steps: Step S1: Construct a six-layer neurorehabilitation causal directed acyclic graph (NRI-DAG) The reasoning process for neurorehabilitation is modeled as a directed acyclic graph G=(V, E), with the node set V divided into six layers according to the neurorehabilitation hierarchy: The first layer (L1, lesion feature layer) node set V1 contains nodes describing the anatomical features of stroke lesions, including lesion location nodes (status: cortical motor area / internal capsule / cerebellum / brainstem / subcortical), lesion volume nodes (status: low / medium / high), stroke course nodes (status: acute phase / subacute phase / chronic phase), affected hemisphere nodes (status: dominant side / non-dominant side), and bilateral affected nodes (status: yes / no). The second layer (L2, neurophysiological mechanism layer) node set V2 contains nodes describing the neurophysiological mechanisms of motor control. Key nodes include: corticospinal tract integrity node (reflecting CST functional integrity), stretch reflex threshold range node (corresponding to the μ parameter of the Feldman reference control frame), postural tension node (reflecting the type I and type II spasticity mechanisms of the Gracies spasticity model), motor cortex excitability node, cortical motor system node, reticulospinal tract node, basal ganglia-brainstem circuit node, proprioception node, neglect syndrome node, and aphasia node. The third layer (L3, clinical syndrome layer) node set V3 contains nodes that describe the clinical motor performance after stroke, including: spasticity level node (modified Ashworth scale / TSRT measurement), upper limb Fugl-Meyer score node, gait speed node, balance function node, hand function node, automatic posture adjustment integrity node, antigravity extension function node, and distal voluntary movement node. The fourth layer (L4, the compensatory pattern layer) node set V4 contains nodes that describe compensatory movement strategies, including: whole-body compensatory pattern nodes, associated response nodes, trunk compensatory nodes, and scapular compensatory nodes. The fifth layer (L5, secondary injury layer) node set V5 contains nodes that describe the risk of secondary injury, including: shoulder pain node, learned disuse node, shoulder subluxation node, and joint contracture node. The sixth layer (L6, functional prognosis layer) node set V6 contains nodes describing functional prognosis, including: Barthel index node, modified Rankin scale node, living independence node, return to work node, quality of life node, and community walking ability node. In the directed edge set E, the direction of each edge (u→v) follows the hierarchical causal flow (i.e., u∈Vᵢ, v∈Vⱼ, i≤j), and each edge has a causal probability weight p(u→v)∈(0,1), which reflects the influence strength of the upstream node on the downstream node; Step S2: Construct a Hybrid-CPT (Conditional Probability Table) For each node Xᵢ in the NRI-DAG, construct its conditional probability distribution P(Xᵢ | Pa(Xᵢ)) under all state combinations of its parent node set Pa(Xᵢ), using a hybrid method prioritizing literature anchors: (a) Document anchor matching: For each parent node state combination pa, traverse the node's preset document anchor set {a k}, each anchor point a k Includes the activation condition set when(a k (a set of parent node-target state pairs) and the corresponding document source probability distribution dist(a) k ); Select the anchor point a* that satisfies the conditions and has the largest number of constraints according to the principle of maximum specificity: a* = argmax_{a k : pa satisfies all when(a) k )condition} |when(a k )| If a* exists, then the conditional probability distribution under this state combination is the normalized dist(a*). (b) Synthetic prior completion: For state combinations without matching literature anchors, basic prior distributions are assigned to nodes according to level: nodes in the compensation layer (L4) and secondary damage layer (L5) adopt low state bias priors (bias coefficient -0.15), nodes in the functional prognosis layer (L6) adopt uniform priors, and nodes in other layers adopt high state bias priors (bias coefficient +0.10); based on the basic priors, the offset is affected by the weighted product of the normalized state values ​​of each parent node and the edge weights, and after normalization and truncation to the range of [0.01, 0.98], the final conditional probability distribution is obtained; Step S3: Bayesian posterior inference based on likelihood-weighted sampling Given a clinical evidence set E = {e1:s1, e2:s2, ..., e k :s k (Node-State Pairs), for the target node set T, perform N likelihood-weighted samplings (default N=400): Traverse all nodes in NRI-DAG topological order; for evidence node eᵢ: fix its state as the observation value sᵢ, and multiply the current sampling weight by P(X). e ᵢ=sᵢ | Pa(X e For latent variable nodes: forward sample the state according to the current CPT distribution; count the weighted counts of each state of each target node, and obtain the posterior probability distribution P(T | E) after normalization; Step S4: Quantification of intervention benefits of Pearl do-calculus-based surgery For the candidate intervention set {(Tⱼ, sⱼ*)}, the expected functional prognosis score after intervention is calculated using graph calculus (do-calculus): For the intervention node Tⱼ, during the likelihood-weighted sampling process, its state is fixed to the target value sⱼ*, but P(Tⱼ=sⱼ* |Pa(Tⱼ)) is not multiplied into the sampling weight, which is equivalent to truncating all edges entering Tⱼ in graph G and performing forward sampling on the modified graph G'; the difference between the comprehensive functional prognosis score of L6 layer after intervention and the baseline score is calculated as the average intervention benefit ATE(Tⱼ); all candidate interventions are sorted in descending order of ATE, and the Top-K recommended intervention schemes are output. Step S5: Compensation Pattern Recognition and Vicious Cycle Detection (a) Compensation mode scoring: Batch query the posterior probability distribution of all L4 layer compensation nodes and L5 layer secondary damage nodes; calculate the weighted compensation score for L4 nodes: score(L4ᵢ) = P(high) + 0.5×P(medium); filter nodes whose scores exceed the threshold θ (default 0.3) and output the list of active compensation modes in descending order of score; (b) Vicious cycle activation detection: For a pre-defined set of vicious cycles C = {c1, c2, ..., c m Each vicious cycle cᵢ contains a set of participating nodes N(cᵢ) and a preset intervention target. For each participating node v in cycle cᵢ, if v∈E and E[v]∈{high, severe, yes, moderate}, or the posterior probability of v satisfies P(v=adverse)>0.5, it is counted as an unfavorable activation. The activation ratio activationratio(cᵢ) = number of unfavorable activation nodes / |N(cᵢ)|. When activationratio≥ 0.4, the vicious cycle cᵢ is determined to be activated, and the activation cycle name, activation ratio and corresponding intervention target are output. Step S6: Leave-one-out (LOO) evidence sensitivity analysis For each evidence node eᵢ in the complete evidence set E, calculate its contribution to functional prognosis: Perform batch queries on the complete evidence set E and the set E\{eᵢ} after removing eᵢ, respectively, to calculate the change in the L6 functional prognosis composite score ΔL6(eᵢ) and the change in the L4 compensation mean score ΔL4(eᵢ); Composite impact score: combined_impact(eᵢ) = 0.6 × |ΔL6(eᵢ)| + 0.4 × |ΔL4(eᵢ)|; Output in descending order of composite impact score, high impact nodes (>0.05) are recommended for precise collection, and low impact nodes (<0.02) can be simplified for evaluation when resources are limited;

[0008] Furthermore, the present invention also includes step S7: compensatory motion scoring based on geometric analysis of skeletal key points. Three-dimensional world coordinates (in meters) of 33 skeletal key points of the subjects were obtained using skeletal posture estimation technology. Quantitative scores for nine compensatory movement patterns were calculated using geometric measurement methods: trunk tilt (absolute value of the difference in longitudinal coordinates between the two acromions), scapular elevation (vertical height deviation of the acromion between the affected and healthy sides), trunk flexion (sagittal plane coordinate difference between the midpoint of the hip joint and the midpoint of the shoulder joint), upper limb swing asymmetry (variance ratio of lateral displacement of the two wrist joints), insufficient weight-bearing on the affected side (longitudinal coordinate difference between the two ankle joints), pelvic rotation (angle between the line connecting the hip joints and the sagittal plane), knee flexion compensation (proportion of insufficient knee angle change during the gait cycle), elbow flexion fixation (proportion of continuous elbow flexion), and head and neck compensation (longitudinal deviation of the nasal tip key point relative to the midpoint of the shoulder joint). The scores for each pattern were normalized to the [0,1] interval. The scores were used as supplementary evidence for the L4 layer nodes to further constrain the posterior distribution of compensatory patterns in Bayesian inference.

[0009] The system technical solution adopted in this invention is as follows: A neurorehabilitation clinical reasoning system based on multi-layer Bayesian causal networks, comprising: The model management module (13) stores the topology of the six-layer neurorehabilitation causal directed acyclic graph, the node state definition, and the mixed conditional probability table of each node. The evidence collection module (12a) collects multi-source clinical evidence, including lesion imaging features, neurological function assessment results (corticospinal tract integrity rating, proprioception, aphasia, neglect), and clinical scale scores (Fugl-Meyer scale, gait test, Berg balance scale, modified Ashworth scale). The inference engine module (13a) performs likelihood-weighted sampling Bayesian inference based on the clinical evidence set, obtains the posterior probability distribution of each node, and supports optional invocation of the exact variable elimination algorithm; The intervention simulation module (13b) uses Pearl do-calculus surgery to perform counterfactual reasoning on candidate intervention options, calculates the average intervention benefit (ATE) of each intervention option, and outputs a priority ranking. The compensation analysis module (13c) identifies the activation compensation mode, detects the activation status of five preset vicious cycles, and assesses the risk of secondary damage. The sensitivity analysis module (13d) performs leave-one-out sensitivity analysis to quantify the impact of each evidence node on functional prognosis and guide the priority allocation of assessment resources. The posture analysis module (15) performs skeletal key point detection on the subject's motion video or image and calculates the geometric scores of nine compensatory motion patterns. The report generation module (16) integrates the reasoning results of the above modules to generate a structured clinical report that includes neurophysiological status assessment, compensation analysis, intervention suggestions and functional prognosis probability distribution;

[0010] The beneficial effects of this invention are as follows: (i) Explainable causal reasoning: Compared with black-box deep learning models, the reasoning mechanism of this invention based on causal directed acyclic graph is completely transparent. Each reasoning conclusion can be traced back to the specific causal chain and the conditional probability of the literature source, enabling clinicians to understand the reasoning process and make informed decisions. (ii) Unbiased intervention benefit estimation: Through Pearl do-calculus surgery, this invention can accurately calculate intervention benefits, eliminate selection bias caused by confounding variables through common cause paths, and provide intervention benefit predictions that are more in line with clinical reality than traditional conditional probability. (III) Systematic evaluation of compensation model: This invention realizes hierarchical quantitative scoring of L4 compensation model and automatic activation detection of five clinically important vicious cycles, and incorporates compensation analysis into a unified causal reasoning framework. (iv) Evidence-based individualized decision support: Leave-one-out sensitivity analysis helps therapists quantify the independent contribution of each assessment indicator to the patient's functional prognosis, enabling priority allocation of evidence-based assessment resources; (v) Multimodal evidence fusion: Integrating structured clinical assessment data with video motion analysis (skeletal posture estimation) into a unified Bayesian inference framework provides multi-dimensional quantitative assessment of compensatory patterns and enhances the comprehensive judgment ability of compensatory movements. (vi) Unexpected Technical Effects: Under sparse parameterization conditions using only 82 literature anchors (covering 9 key nodes out of 96 nodes), the literature backtesting of five key causal links of this invention passed the ±10% consistency standard (maximum error 5.8%), which is far superior to the average error (approximately 23%) of the pure synthetic prior version without anchors in the early stage of research and development. This result shows that the "maximum specificity anchor priority + synthetic prior completion" hybrid CPT construction method proposed in this invention can achieve higher-than-expected quantitative accuracy under sparse clinical literature conditions, an effect not reported in existing literature. In addition, compared with ordinary Bayesian networks (excluding do-calculus surgery), the ATE prediction value of this invention for the same intervention (spasticity management) is 19-31% lower (due to the elimination of systematic overestimation caused by selection bias), which is closer to the actual intervention effect reported in randomized controlled clinical trials. (VII) Legal Nature of the Invention: This invention provides an information processing technology for clinical decision support. All its outputs are in the form of data such as probability distributions, priority rankings, and statistical reports, and do not constitute diagnostic conclusions or treatment orders. The outputs of this system are provided only as supplementary reference information to licensed physicians or rehabilitation therapists. Final clinical decisions (including diagnostic judgments, treatment plan selection, and medication decisions) must be made by qualified medical professionals based on their professional judgment. The technical subject matter protected by this invention is information processing methods and systems, not methods for diagnosing or treating diseases. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings are briefly described below: Figure 1 This is the overall architecture diagram of the neurorehabilitation clinical reasoning system of the present invention, wherein: 11 is the front-end interactive interface, 12 is the API service layer, 13 is the NRI reasoning engine layer, 14 is the data persistence layer, 15 is the skeletal posture analysis module, 16 is the clinical report generation module, 13a is the reasoning engine submodule, 13b is the intervention simulation submodule, 13c is the compensation analysis submodule, and 13d is the sensitivity analysis submodule. Figure 2 This is a schematic diagram of the six-layer neurorehabilitation causal directed acyclic graph (NRI-DAG) structure of the present invention, showing representative nodes of the six layers from L1 to L6 and their causal connections; Figure 3 The flowchart of the Bayesian inference engine and intervention benefit quantification of this invention shows the calculation process of likelihood-weighted sampling inference (step S3) and do-calculus graph surgery (step S4); Figure 4 This is a six-step clinical evaluation workflow diagram of the present invention, which shows the complete clinical application process from evidence collection to structured report generation; Detailed Implementation

[0012] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The embodiments listed in this invention are merely preferred embodiments and are not intended to limit the scope of protection of this invention; any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

[0013] The following definitions apply to the entire specification and claims of this invention: "Directed Acyclic Graph (DAG): A graph structure in which nodes are connected by directed edges and do not contain directed loops, used to represent conditional independence and causal dependency between variables; "Conditional Probability Table (CPT)": A lookup table describing the conditional probability distribution of a discrete random variable under all combinations of states of its parent node; "Literature Anchor Point": Probability distribution parameters extracted from published evidence-based clinical studies under specific parent node state conditions, with traceable literature sources; "Graph Surgery (do-calculus)": An intervention simulation method in Pearl's causal reasoning framework, which achieves unbiased estimation of intervention effects by truncating all incoming edges of intervention nodes, and is fundamentally different from observation conditionation; "Leave One Out (LOO) method": systematically remove each piece of evidence one by one, and quantify the information contribution of each piece of evidence by comparing the changes in the reasoning results; "Vicious cycle": In the context of NRI-DAG, a positive feedback structure consisting of multiple mutually reinforcing nodes, where the adverse states of each node promote each other, forming a self-reinforcing pathological pattern. "Compensatory mode": The alternative motor strategies adopted by stroke patients to complete motor tasks. In the short term, it can maintain some function, but in the long term, it may lead to secondary musculoskeletal damage and inhibit the remodeling of target neural pathways. Example 1: NRI-DAG Model Structure

[0014] like Figure 2 As shown, this invention constructs a six-layer NRI-DAG containing 96 nodes and 155 directed edges. This directed acyclic graph integrates five major motor control theories: the corticospinal tract theory (Stinear PREP2 framework), the Gracies spasticity pathophysiological model, the reticular spinal tract compensation theory (Park 2018; Kubota 2024), the reference control framework (Feldman 1986; Jobin 2000), and the basal ganglia posture regulation theory.

[0015] The L1 lesion feature layer contains five root nodes (without parent nodes), covering the anatomical features of stroke lesions. The L1 lesionsite node has five states (cortical motor cortex, internal capsule, cerebellum, brainstem, and subcortical), reflecting the differentiated impact of different anatomical locations on the extent of CST and RST damage. The L1 lesion volume node uses a three-level classification (low / medium / high), corresponding to MRI lesion volume measurements.

[0016] The L2 neurophysiological mechanism layer contains approximately 15 nodes and is the core intermediate layer of the entire NRI-DAG. The corticospinal tract integrity node (L2CSTintegrity) is determined by the lesion characteristics of the L1 layer and directly affects distal voluntary motor ability; the stretch reflex threshold range node (L2TSRTrange) corresponds to the reference angle range of muscle activation in the Feldman reference control frame and is regulated by nodes such as L2posturaltone (postural tension); the basal ganglia-brainstem loop node (L2BGBSloop) receives influences from multiple upstream nodes through three parent sides (L2motorcortexexcitability, L1lesionsite, L1_bilateral) and regulates the quality of automatic posture adjustment (APA).

[0017] The L3 clinical syndrome layer comprises approximately 20 nodes, mapping neurophysiological mechanisms to observable clinical manifestations and scale scores. Among them, the L3 antigravity extension node is directly innervated by the reticulospinal tract node, while the L3 distal voluntary movement node is mainly determined by the integrity of the CST and the cortical motor system, reflecting the differentiated contributions of the RST-CST dual pathway to motor function.

[0018] The node design of the L4 compensation pattern layer and the L5 secondary injury layer reflects the hierarchical nature of compensatory movements: the L4 node reflects active compensation strategies, while the L5 node reflects structural secondary injuries that occur when the compensation strategies are not effectively controlled. The trunk compensation node (L4 compensation trunk) has three parent sides (L2 reticulospinal, L3 balance, L3APAintegrity), reflecting that trunk compensation is jointly influenced by the function of the reticular spinal cord and postural control ability.

[0019] The NRI-DAG defines five clinically important vicious cycles: (1) Spasticity-compensation cycle: Increased postural tension → reduced TSRT range → worsening spasticity → exacerbated combined reactions → solidification of compensatory strategies; (2) Disuse-pain cycle: Learned disuse → muscle atrophy of the affected limb → weakened shoulder support → worsening shoulder pain → movement avoidance → exacerbation of learned disuse; (3) Tension-contraction cycle: Persistent high postural tension → persistent overactivation of muscles → decreased joint range of motion → decreased passive stretching effect; (4) Compensation-functional degeneration cycle: Compensatory strategies dominate movement → reduced activation of target muscle groups → reorganization of neural plasticity in an unfavorable direction; (5) Neglect-disuse cycle: Spatial neglect syndrome → decreased frequency of use of the affected limb → progression of learned disuse → loss of motor experience → weakened remodeling drive. Example 2: Construction of Hybrid-CPT (Construction of Hybrid-CPT)

[0020] This invention employs a conditional probability table covering nine key clinical nodes (see Table 1) using 82 literature anchors. Taking the corticospinal tract integrity node (L2CSTintegrity) as an example, the literature anchor examples are as follows (Source: Stinear 2017 PREP2; Byblow 2015): Lesion location = motor cortex and lesion volume = height 0.75 0.15 0.10 Stinear 2017 Lesion location = internal capsule and lesion volume = height 0.70 0.20 0.10 Byblow 2015 Lesion location = motor cortex and lesion volume = low 0.25 0.45 0.30 Stinear 2012 Lesion location = subcortical and lesion volume = medium 0.45 0.35 0.20 Stinear 2017 For state combinations without literature anchor matching (e.g., lesion location = cerebellum and lesion volume = low), a synthesis method based on edge weights is adopted: First, based on the high state bias prior of L1 node [0.25, 0.35, 0.40], the comprehensive influence intensity is calculated by multiplying the normalized state value (stateindex / maxindex) of each parent node with the corresponding edge weight. Then, a proportional offset of influence × α (α = 0.6) is applied to the basic prior, normalized and truncated to the range of [0.01, 0.98]. Table 1: List of Document Anchor Point Coverage Nodes L2_CST_integrity 9 Stinear 2017 PREP2; Byblow 2015; Stinear 2012 L2_TSRT_range 8 Feldman 1986; Jobin 2000 L3_spasticity 10 Gracies 2005; Bhimani 2014; Lance 1980 L3_FMA_UE 8 Kim 2014; Byblow 2015 L4_compensation_global_pattern 9 Cirstea 2000; Levin 2000 L5_shoulder_pain 8 Lindgren 2007; Niessen 2008 L5_learned_nonuse 8 Taub 2002; Wittenberg 2003 L6_Barthel 11 Veerbeek 2011; Jørgensen 1999 L6_mRS 11 Quinn 2009; Banks 1988 Example 2 (continued): Example of complete conditional probability representation of L3_spasticity nodes [0020a] To illustrate the synergistic mechanism of maximum specificity matching of literature anchors and synthetic prior completion, the conditional probability table under its nine parent node state combinations is fully displayed using the spasticity node (L3_spasticity) as an example. This node has two parent nodes: L2posturaltone (postural tone, three-state: low / medium / high) and L2TSRTrange (TSRT threshold adjustment range, three-state: low / medium / high), forming a total of 3×3=9 parent node state combinations. The preset literature anchors of this node cover 6 of the highly clinically relevant combinations, and the remaining 3 are completed through synthetic priors. (a) Example of maximum specificity matching - conflict resolution of multiple candidate anchor points When the parent node's state is posturaltone=high and TSRTrange=low, the system retrieves three candidate anchor points that meet the activation conditions from the node's preset anchor point set: Anchor point A postural_tone=high 1 Gracies 2005 (single condition) Anchor point B TSRT_range=low 1 Feldman 1986 (single condition) Anchor point C postural_tone=high and TSRT_range=low 2 Gracies 2005 (two-condition combination) According to the principle of maximum specificity (a* = argmax | when(a) k The system selects anchor point C (N=2) with the most constraints, and uses its probability distributions P(low)=0.05, P(medium)=0.20, P(high)=0.75 as the conditional probability of this state combination, discarding anchor points A and B that only satisfy a single condition. This principle ensures that when multiple literature anchor points meet the activation conditions, literature evidence with more specific and concrete clinical contexts is given priority. (b) Complete CPT for L3_spasticity node (9 combinations of parent node states) 1 high low 0.05 0.20 0.75 Literature anchor C (Gracies 2005, two conditions, N=2) 2 high medium 0.10 0.30 0.60 Literature anchor point D (Gracies 2005, two conditions, N=2) 3 high high 0.24 0.35 0.41 **Synthetic Priors** (No literature anchor coverage, see calculation process) 4 medium low 0.15 0.35 0.50 Literature anchor point E (Feldman 1986; Jobin 2000, two conditions, N=2) 5 medium medium 0.34 0.41 0.25 **Synthetic Priors** (No Literature Anchor Coverage) 6 medium high 0.52 0.33 0.15 Literature anchor F (Jobin 2000, two conditions, N=2) 7 low low 0.28 0.44 0.28 Reference anchor G (Feldman 1986, single-condition TSRT_range=low, N=1) 8 low medium 0.51 0.34 0.15 **Synthetic Priors** (No Literature Anchor Coverage) 9 low high 0.72 0.20 0.08 Literature anchor H (Gracies 2005, two conditions, N=2)† †All row probabilities are normalized, and the probabilities of each state are truncated to the interval [0.01, 0.98]. (c) Example of the synthesis prior calculation process (line 3: postural_tone=high, TSRT_range=high) This combination lacks a direct joint probability report in the literature (high postural tension theoretically increases the risk of spasticity, but a wide TSRT range means that the reference control mechanism still retains a relatively wide regulatory space, which has a partial offsetting effect on spasticity), thus triggering synthetic prior completion: Step 1: Assign basic priors. The L3 layer (clinical syndrome layer) nodes adopt high-state bias priors (bias coefficient +0.10), with a basic distribution of [0.27, 0.35, 0.38] (low / medium / high); Step 2: Calculate the overall influence strength. Use the formula: influence = Σⱼ(Normalized state value of parent node j - 0.5) × edge_weight(j → current node) / Number of parent nodes. • L2posturaltone=high: Normalized state value = 2 / 2 = 1.0, (1.0-0.5) × edge weight 0.75 = +0.375; • L2TSRTrange=high: High TSRT_range is a protective factor against spasticity (wide threshold range → low spasticity), and the edge weight is set to −0.70, (1.0-0.5)×(−0.70) = −0.350; • influence = (0.375 + (−0.350)) / 2 = +0.0125 (Net effect: high tension slightly dominant); Step 3: Stack the offsets and normalize. Adjustment amount = influence × α (default α = 0.6) = +0.0075; Stack the offsets onto the basic prior high direction and subtract them equally from the low direction, resulting in [0.263, 0.35, 0.388]. After normalization, truncate to [0.01, 0.98], and finally P(low) = 0.24, P(medium) = 0.35, P(high) = 0.41. Step 4: Clinical Consistency Validation. The results were consistent with clinical expectations: with high tension + wide TSRT, the tendency for spasticity was moderate to high but not severe (P(high)=0.41), significantly lower than in the first row of high tension + narrow TSRT (P(high)=0.75), demonstrating the important protective and regulatory role of TSRT range in the occurrence of spasticity, consistent with the theoretical predictions of the Feldman reference control framework. Example 3: Quantification of Intervention Benefits (do-calculus surgery)

[0021] like Figure 3 As shown, this invention implements Pearl do-calculus through the following steps, taking candidate intervention (intervention node T, target state s*) as an example: Step (a): Baseline assessment. Perform batch likelihood-weighted sampling (N=200 times) on the L6 node set to obtain the baseline functional prognostic distribution P(L6 | E), and calculate the baseline L6 comprehensive score: baselinescore = Σ{v∈L6} Σ_k P(v=k|E) × value(v,k), where the value function for positive prognostic nodes (Barthel, QoL, etc.) is high state = 1.0, medium state = 0.5, and low state = 0.0; negative prognostic nodes (mRS, nursing burden, etc.) are mapped inversely. Step (b): Intervention simulation (graph surgery). Perform N samplings, traversing each particle in topological order: for intervention node T, force sample[T]=s*, but do not multiply P(T=s*|Pa(T)) into the weights; process the remaining nodes according to the standard likelihood weighted sampling rules; count the weighted counts of L6 nodes and calculate the L6 comprehensive score sim_score after intervention; Step (c): Benefit calculation. ATE(T, s*) = simscore - baselinescore;

[0022] This embodiment uses spasticity management intervention (do(L3_spasticity="low"), corresponding to botulinum toxin injection) as an example to illustrate the essential difference between graphical surgery and conditionalization: Conditioned P(L6|L3_spasticity="low", E): The system uses low spasticity as evidence to infer that its parent nodes (postural tension, TSRT range, etc.) are also in a favorable state, systematically overestimating the overall functional prognosis after intervention, while actual spasticity management only directly improves spasticity and does not change the degree of upstream nerve damage. The graph operation P(L6|do(L3_spasticity="low"), E) fixes the belief of the spasticity node as low but does not update its parent node, and only propagates the effect of spasticity improvement through the downstream causal chain. It correctly simulates the clinical scenario of "giving botulinum toxin injection only changes the spasticity itself" and outputs an unbiased estimate of the expected functional benefit of the intervention. Example 4: Compensation Pattern Recognition and Vicious Cycle Detection

[0023] In this embodiment, a patient with severe hemiplegia (FMA-UE=18 / 66, modified Ashworth grade 2) with a large volume of lesions in the cortical motor area during the subacute phase of the disease is used as an example: L4 layer compensation score (example): • L4compensationtrunk: P(high)=0.58, P(medium)=0.32 → score= 0.74 (activated, high priority) • L4synergy pattern (combined reaction): P(high)=0.41, P(medium)=0.43 → score =0.62 (activation) • L4 scapular compensation: P(high)=0.28, P(medium)=0.31 → score = 0.44 (activation) Vicious cycle detection results: • Spasticity-compensatory cycle (6 involved nodes, 4 of which are in an adverse state): activationratio=0.67 → Activated; Recommended intervention target: L2TSRT_range (stretch reflex threshold regulation, corresponding to kinesiology patch / oral baclofen). • Disuse-pain cycle (5 involved nodes, 2 of which are in an adverse state): activationratio=0.40 → Activated; Recommended intervention target: L5learned_nonuse (Constraint-induced movement therapy CIMT); Example 5: Sensitivity Analysis Using Leave-One-Out Method

[0024] Using the patient evidence set from Example 4 as input, a leave-one-out sensitivity analysis was performed, and the results are as follows (example): L2_CST_integrity L2 0.148 0.163 0.124 Precise data acquisition (MEP inspection) L3_spasticity L3 0.097 0.102 0.089 Precise harvesting (improved Ashworth) L2_postural_tone L2 0.083 0.089 0.075 Precise collection L3_FMA_UE L3 0.071 0.078 0.061 Precise collection L1_bilateral L1 0.018 0.021 0.013 Low impact, can be simplified Analysis conclusion: For this patient, the integrity of the corticospinal tract (L2CSTintegrity) is the most important single piece of evidence affecting the prediction of functional prognosis, and motor evoked potential (MEP) testing is recommended as a priority; while the bilateral impairment status (L1_bilateral) has the least impact on the inference conclusion and can be temporarily excluded when assessment resources are limited. Example 6: Six-Step Clinical Assessment Workflow

[0025] like Figure 4 As shown, the complete six-step application process of the system of the present invention in a clinical setting is as follows: Step 1 (Patient Basic Information): Enter the patient's basic information (name, age, gender), stroke type (ischemic / hemorrhagic), onset time, and affected side; Step 2 (Imaging and Lesion Information): Enter the lesion location, volume estimation, damaged hemisphere, whether it is bilateral, and the corresponding L1 layer evidence node; Step 3 (Neurological Function Assessment): Input the corticospinal tract integrity rating (based on MEP / TCMS / DTI), proprioception, aphasia degree, neglect syndrome assessment, corresponding L2 layer evidence nodes; Step 4 (Clinical Scale Entry): Enter the Fugl-Meyer Upper Limb Scale (0-66 points), 10-meter gait test (m / s), Berg Balance Scale, and Modified Ashworth Scale ratings, corresponding to the L3 level evidence nodes. Step 5 (Posture Video Acquisition): Acquire patient movement videos, and the skeletal posture analysis module (15) automatically calculates scores for nine compensatory movement patterns to generate a comprehensive compensatory index as supplementary evidence for L4 layer; Step 6 (NRI Inference and Report Generation): Trigger the NRI inference engine to sequentially execute posterior inference (S3), intervention benefit quantification (S4), compensation pattern identification (S5), and leave-one-out sensitivity analysis (S6). The report generation module (16) then generates a structured clinical report, which includes: neurophysiological state atlas (posterior probability of each node in L2 layer), list of compensation pattern activation and vicious cycle warning, functional prognostic probability distribution (six prognostic indicators in L6 layer), intervention priority ranking and expected ATE value, and heatmap of key evidence contribution.

[0026] The accuracy of the following key inference links has been verified in literature backtesting (errors are all within ±10%): PLIC lesions → Low CST function 79.7% 75-80% (Stinear 2017) <5% (passed) CM+CST complete → distal voluntary movement 78.3% >75% (Kubota 2024) <3% (passed) RST complete → Anti-gravity stretching 77.9% >70% (Park 2018) <8% (passed) Reference control → TSRT normal 66.6% >60% (Feldman 1986) <7% (passed) BG-BS + Basal ganglia → APA complete 69.2% ~75% (Related Literature) 5.8% (passed) The above verification results show that the NRI-DAG model and hybrid CPT construction method of the present invention can accurately reflect the core causal relationship of post-stroke neurological rehabilitation, and provide evidence-based and reliable quantitative basis for clinical reasoning. Example 7: Complete NRI-DAG Topology Disclosed

[0027] To enable those skilled in the art to fully reproduce this invention, the node list, edge set grouping, and vicious cycle structure of the NRI-DAG are fully disclosed below in the form of an appendix table. The complete NRI causal model (nricausalmodel.json) used in the implementation of this invention contains 96 nodes, 155 directed edges, 5 vicious cycles, and 31 intervention nodes, with the specific structure as follows. Table A: Complete list of NRI-DAG nodes (96 nodes) 1 L1_IC_subsite Internal capsule injury zones (PLIC / ALIC / Knee) L1 posterior_limb / anterior_limb / genu / NA none 2 L1_age Age stratification L1 low / medium / high none 3 L1_comorbidity Comorbidity Index L1 low / medium / high none 4 L1_lesion_side Damage side L1 left / right / bilateral none 5 L1_lesion_site Damaged area L1 cortex_motor / internal_capsule / corona_radiata / basal_ganglia / thalamus / brainstem / cerebellum / multiple none 6 L1_lesion_volume Damage volume (normalized) L1 low / medium / high none 7 L1_premorbid_mobility Pre-illness activity level L1 low / medium / high none 8 L1_time_since_stroke Disease course L1 hyperacute / acute / subacute / chronic none 9 L2_BG_BS_loop Basal ganglia-brainstem circuit function L2 low / medium / high L1_lesion_site, L1_lesion_volume 10 L2_BG_inhibition Basal ganglia motor inhibition function L2 low / medium / high L1_lesion_site 11 L2_BG_initiation Basal ganglion kinetic initiation function L2 low / medium / high L1_lesion_site 12 L2_CM_system Cortical motor neuron system L2 low / medium / high L1_lesion_site, L2_CST_integrity, L1_IC_subsite 13 L2_CPG_integrity Central pattern generator integrity L2 low / medium / high L2_BG_BS_loop 14 L2_CST_integrity Corticospinal tract integrity L2 low / medium / high L1_lesion_site, L1_lesion_volume, L1_IC_subsite 15 L2_TSRT_range TSRT threshold adjustment range L2 low / medium / high L2_referent_control 16 L2_attention Attention / Execution Function L2 low / medium / high none 17 L2_body_schema Body schema integrity L2 low / medium / high none 18 L2_brainstem_automatic brainstem automotor regulation L2 low / medium / high L2_BG_BS_loop 19 L2_corticoreticulospinal Corticoreticular Spinal System L2 low / medium / high L1_lesion_site, L1_IC_subsite, L1_lesion_volume 20 L2_cuneate_gating wedge-beam core gating efficiency L2 low / medium / high L1_lesion_site, L2_proprioception 21 L2_cutaneous Skin sensation L2 low / medium / high none 22 L2_motor_abundance Motion richness (good variation ratio) L2 low / medium / high L1_lesion_volume, L2_CST_integrity 23 L2_motor_learning_rate Learning rate of motion L2 low / medium / high L1_time_since_stroke, L1_age, L2_sensory_gating 24 L2_postural_tone Postural muscle tension L2 low / medium / high L2_BG_inhibition 25 L2_proprioception Proprioception L2 low / medium / high L1_lesion_site 26 L2_referent_control Reference control accuracy L2 low / medium / high L1_lesion_site, L1_lesion_volume 27 L2_reticulospinal Reticulospinal tract function L2 low / medium / high L1_lesion_site, L1_lesion_volume 28 L2_rubrospinal Red nucleus spinal tract function L2 low / medium / high none 29 L2_sensory_gating Feeling the gating efficiency L2 low / medium / high L2_vestibular, L2_visual, L2_cuneate_gating 30 L2_vestibular Vestibular function L2 low / medium / high L2_proprioception 31 L2_vestibulospinal vestibular spinal tract function L2 low / medium / high L1_lesion_site 32 L2_visual visual function L2 low / medium / high none 33 L3_APA_integrity Pre-positioning integrity L3 low / medium / high L2_reticulospinal, L2_corticoreticulospinal, L2_BG_BS_loop, L2_body_schema, L2_attention, L2_BG_initiation 34 L3_APA_latency APA incubation period L3 low / medium / high L3_APA_integrity 35 L3_ARAT ARAT rating L3 low / medium / high L3_grasp_quality, L3_dexterity 36 L3_FMA_UE Fugl-Meyer Upper Extremity Score (Normalized) L3 low / medium / high L2_CST_integrity, L2_motor_learning_rate, L3_multi_joint_coordination 37 L3_WMFT_time WMFT Function Task Time L3 low / medium / high L3_reaching_quality 38 L3_ankle_coactivation Ankle joint co-activation L3 low / medium / high L3_coactivation 39 L3_ankle_dorsiflexion Ankle dorsiflexion angle (swing phase) L3 low / medium / high L3_ankle_coactivation 40 L3_antigravity_extension Anti-gravity stretching ability L3 low / medium / high L2_reticulospinal, L3_APA_integrity, L2_postural_tone 41 L3_circumduction Circular gait L3 low / medium / high none 42 L3_coactivation Antagonistic muscle co-activation L3 low / medium / high L2_TSRT_range, L3_spasticity 43 L3_dexterity Hand dexterity L3 low / medium / high L3_fractionated_movement, L3_distal_voluntary_movement 44 L3_distal_voluntary_movement Distal voluntary movement ability L3 low / medium / high L2_CM_system, L2_CST_integrity 45 L3_foot_drop Foot drop L3 absent / mild / moderate / severe L2_CST_integrity, L2_BG_BS_loop, L3_ankle_dorsiflexion, L3_ankle_coactivation 46 L3_fractionated_movement Separation of motion capabilities L3 low / medium / high L2_CM_system, L3_distal_voluntary_movement 47 L3_gait_initiation_time Gait initiation time L3 low / medium / high L2_BG_initiation, L2_attention, L3_APA_integrity, L3_APA_latency 48 L3_gait_speed pace L3 low / medium / high L2_BG_BS_loop, L3_APA_integrity, L3_step_length, L3_gait_initiation_time, L3_ankle_coactivation 49 L3_gait_symmetry Gait symmetry L3 low / medium / high none 50 L3_grasp_quality Grasp quality L3 low / medium / high L3_proximal_distal_dissociation, L3_dexterity 51 L3_hip_hiking Hip lift compensation L3 low / medium / high none 52 L3_movement_efficiency Exercise efficiency (metabolic cost) L3 low / medium / high L3_coactivation, L3_hip_hiking, L3_circumduction 53 L3_movement_initiation Motor start-up ability L3 low / medium / high none 54 L3_movement_smoothness Motion smoothness L3 low / medium / high L3_coactivation 55 L3_multi_joint_coordination Multi-joint coordination L3 low / medium / high L2_motor_abundance 56 L3_muscle_tone_asymmetry Muscle tone asymmetry L3 low / medium / high none 57 L3_postural_sway Posture sway amplitude L3 low / medium / high L2_vestibulospinal, L2_proprioception, L2_vestibular, L2_sensory_gating, L3_slr_integrity 58 L3_proximal_distal_dissociation Proximal-distal separation motion L3 low / medium / high L3_shoulder_girdle_stability 59 L3_reaching_quality Reaching motion quality L3 low / medium / high L3_proximal_distal_dissociation, L3_shoulder_girdle_stability 60 L3_rigidity Muscle rigidity L3 low / medium / high L2_postural_tone 61 L3_selective_movement Selective motor skills L3 low / medium / high L2_motor_abundance 62 L3_shoulder_girdle_stability shoulder strap stability L3 low / medium / high L2_reticulospinal 63 L3_slr_integrity Short latency postural response integrity L3 low / medium / high L2_reticulospinal, L2_vestibulospinal 64 L3_spasticity Degree of spasm L3 low / medium / high L2_postural_tone, L2_TSRT_range 65 L3_step_length Step length L3 low / medium / high L3_hip_hiking、L3_circumduction 66 L4_compensation_foot Foot compensation (foot drop / inversion) L4 low / medium / high L3_foot_drop 67 L4_compensation_global_pattern Holistic movement pattern (non-selective) L4 low / medium / high L2_CST_integrity, L2_motor_abundance 68 L4_compensation_head_neck Head and neck compensation L4 low / medium / high none 69 L4_compensation_hip Hip compensation (elevation / external rotation) L4 low / medium / high L3_foot_drop 70 L4_compensation_shoulder Shoulder strap compensation (lifting / extending) L4 low / medium / high L2_CM_system, L3_selective_movement, L3_dexterity 71 L4_compensation_trunk Thoracolumbar spine compensation (lateral flexion / rotation) L4 low / medium / high L3_antigravity_extension, L3_APA_integrity, L3_shoulder_girdle_stability 72 L4_compensation_vision_dependence Visual Dependence Compensation L4 low / medium / high L2_proprioception, L2_vestibular 73 L4_trunk_rotation_asymmetry Torso rotation asymmetry L4 low / medium / high L4_compensation_shoulder 74 L4_weight_shift_asymmetry Asymmetric shift of center of gravity L4 low / medium / high none 75 L5_anxiety anxiety L5 low / medium / high L5_depression 76 L5_cardiovascular_deconditioning Cardiovascular adaptation L5 low / medium / high none 77 L5_contracture Joint contracture L5 low / medium / high L1_time_since_stroke, L3_spasticity, L5_muscle_atrophy 78 L5_depression Depression L5 low / medium / high L5_shoulder_pain 79 L5_fall_history History of Falls L5 absent / mild / moderate / severe L3_postural_sway, L5_osteoporosis 80 L5_fatigue fatigue L5 low / medium / high L3_movement_efficiency, L4_compensation_trunk, L5_depression 81 L5_fear_of_falling Fear of falling L5 low / medium / high L3_postural_sway, L4_compensation_vision_dependence, L5_osteoporosis, L5_fall_history, L5_anxiety 82 L5_impingement_syndrome Subacromial impingement syndrome L5 no / yes none 83 L5_learned_nonuse Learned disuse L5 low / medium / high L1_time_since_stroke、L5_shoulder_pain 84 L5_maladaptive_plasticity Poor neuroplasticity L5 low / medium / high L5_learned_nonuse 85 L5_muscle_atrophy Muscle atrophy L5 low / medium / high L3_spasticity, L3_coactivation, L5_cardiovascular_deconditioning 86 L5_osteoporosis Osteoporosis risk L5 low / medium / high L3_circumduction, L5_cardiovascular_deconditioning 87 L5_shoulder_pain shoulder pain L5 low / medium / high L5_shoulder_subluxation, L5_impingement_syndrome 88 L5_shoulder_subluxation shoulder subluxation L5 absent / mild / moderate / severe none 89 L5_social_isolation social distancing L5 low / medium / high L3_gait_speed, L5_fear_of_falling 90 L6_Barthel Barthel index (normalized) L6 low / medium / high L1_premorbid_mobility, L3_gait_speed, L3_FMA_UE, L3_ARAT 91 L6_QOL Quality of life L6 low / medium / high L5_depression, L6_Barthel, L6_participation 92 L6_caregiver_burden Caregiver burden L6 low / medium / high none 93 L6_healthcare_utilization Medical resource utilization L6 low / medium / high L1_comorbidity, L5_shoulder_pain, L5_contracture, L5_fall_history, L5_depression 94 L6_mRS Modified Rankin Scale L6 absent / mild / moderate / severe L6_Barthel 95 L6_mortality_risk Risk of death L6 low / medium / high L1_age, L1_comorbidity, L5_cardiovascular_deconditioning, L6_healthcare_utilization 96 L6_participation social participation L6 low / medium / high L3_gait_speed

[0028] Table B: Statistics of NRI-DAG directed edge sets (155 edges) L1→L2 22 L1_lesion_site→L2_CST_integrity (0.8); L1_lesion_site→L2_CM_system (0.75); L1_lesion_site→L2_reticulospinal (0.75); (Total 22 items) L1→L5 2 L1_time_since_stroke→L5_contracture (0.8); L1_time_since_stroke→L5_learned_nonuse (0.75) L1→L6 4 L1_age→L6_mortality_risk (0.7); L1_comorbidity→L6_mortality_risk (0.75); L1_comorbidity→L6_healthcare_utilization (0.7); (Total 4 items) L2→L2 12 L2_CST_integrity→L2_CM_system (0.8); L2_CST_integrity→L2_motor_abundance (0.7); L2_BG_BS_loop→L2_brainstem_automatic (0.75); (Total 12 items) L2→L3 31 L2_CST_integrity→L3_FMA_UE (0.7); L2_CST_integrity→L3_foot_drop (0.65); L2_CM_system→L3_fractionated_movement (0.75); (Total 31 items) L2→L4 5 L2_CST_integrity→L4_compensation_global_pattern(0.75); L2_CM_system→L4_compensation_shoulder(0.8); L2_proprioception→L4_compensation_vision_dependence (0.75); (Total 5 items) L3→L3 32 L3_APA_integrity→L3_gait_initiation_time (0.85); L3_APA_integrity→L3_gait_speed (0.75); L3_distal_voluntary_movement→L3_dexterity (0.85); (Total 32 items) L3→L4 7 L3_selective_movement→L4_compensation_shoulder(0.75); L3_foot_drop→L4_compensation_foot (0.9); L3_foot_drop→L4_compensation_hip (0.85); (7 items in total) L3→L5 8 L3_postural_sway→L5_fall_history (0.7); L3_movement_efficiency→L5_fatigue (0.75); L3_spasticity→L5_muscle_atrophy (0.7); (8 items in total) L3→L6 4 L3_gait_speed→L6_Barthel (0.8); L3_gait_speed→L6_participation (0.8); L3_FMA_UE→L6_Barthel (0.85); (4 items in total) L4→L4 1 L4_compensation_shoulder→L4_trunk_rotation_asymmetry (0.7) L4→L5 2 L4_compensation_trunk→L5_fatigue (0.7); L4_compensation_vision_dependence→L5_fear_of_falling(0.65) L5→L5 15 L5_shoulder_pain→L5_learned_nonuse (0.85); L5_shoulder_pain→L5_depression (0.8); L5_shoulder_subluxation→L5_shoulder_pain (0.8); (15 items in total) L5→L6 6 L5_shoulder_pain→L6_healthcare_utilization(0.75); L5_contracture→L6_healthcare_utilization(0.7); L5_fall_history→L6_healthcare_utilization(0.8); (6 items in total) L6→L6 4 L6_Barthel→L6_mRS (0.85); L6_Barthel→L6_QOL (0.8); L6_participation→L6_QOL (0.85); (4 items in total)

[0029] Table C: The Complete Structure of Five Vicious Cycles The vicious cycle is defined as a positive feedback structure in the NRI-DAG consisting of multiple mutually reinforcing pathological state nodes, where adverse states of each participating node can mutually promote each other through causal chains, forming a self-sustaining pathological pattern. This invention presupposes five clinical vicious cycles as follows: VC-A Upper limb disuse closed loop Damage to the CM system → Decreased fine motor skills of the hand → Shoulder girdle compensation → Acromial impingement → Shoulder pain → Learned disuse → Poor plasticity → Further functional decline L2_CM_system, L3_dexterity, L4_compensation_shoulder, L5_impingement_syndrome, L5_shoulder_pain, L5_learned_nonuse, L5_maladaptive_plasticity, L3_reaching_quality (8) Shoulder girdle stabilization training, scapular control, shoulder pain management, CIMT, motor imagery, TMS A / B VC-B Fear of falling to adapt to a closed loop BG-BS disorder → APA deficiency → Postural wobble → Falls → Fall phobia → Reduced activity → Cardiovascular deadaptation / muscle atrophy / osteoporosis → Repeated falls L2_BG_BS_loop, L3_APA_integrity, L3_postural_sway, L5_fall_history, L5_fear_of_falling, L5_social_isolation, L5_cardiovascular_deconditioning, L5_muscle_atrophy, L5_osteoporosis (9) Postural control training, balance training, cognitive behavioral therapy, progressive activity planning, resistance training A / B VC-C Feeling like a closed loop of depression in learning Impaired proprioception → decreased efficiency of cuneate nucleus gating → decreased motor learning efficiency → slow rehabilitation progress → depression / anxiety → decreased adherence → even slower progress L2_proprioception, L2_cuneate_gating, L2_sensory_gating, L2_motor_learning_rate, L3_FMA_UE, L5_depression, L5_anxiety, L5_social_isolation (8) Proprioceptive training, robot-assisted training, external focus attention, augmented feedback, motivational interviews B VC-D Spasm efficiency fatigue closed loop Limited range of motion in TSRT → worsening spasticity → increased coactivation → decreased exercise efficiency → increased fatigue → reduced training volume → decreased muscle strength → worsening spasticity L2_TSRT_range, L3_spasticity, L3_coactivation, L3_movement_efficiency, L5_fatigue, L3_FMA_UE (6) Botulinum toxin injections, oral antispasmodics, stretching, selective exercise training, FES (fiber optic stimuli), and low-intensity endurance training. A / B VC-E Gait abnormality joint load closed loop Foot drop → circumduction gait / hip elevation → abnormal joint load → osteoarthritis → pain → further gait abnormalities → muscle atrophy L3_ankle_dorsiflexion, L3_foot_drop, L4_compensation_hip, L3_circumduction, L5_osteoporosis, L3_gait_speed, L5_muscle_atrophy (7) Ankle-foot orthosis (AFO), functional electrical stimulation (FES), gait retraining, weight-bearing gait training, joint protection education B / C

Claims

1. A neural rehabilitation clinical reasoning method based on a multi-layer Bayesian causal network, characterized in that, Includes the following steps: S1: Construct a six-layer causal directed acyclic graph (NRI-DAG) for neurorehabilitation. Model the neurorehabilitation reasoning problem as a directed acyclic graph G=(V, E). The node set V is divided into six layers according to the neurorehabilitation level: lesion feature layer (L1), neurophysiological mechanism layer (L2), clinical syndrome layer (L3), compensation mode layer (L4), secondary injury layer (L5), and functional prognosis layer (L6). Each edge in the directed edge set E has a causal probability weight p(u→v)∈(0,1), and the edge direction follows the hierarchical causal flow direction. S2: A hybrid method prioritizing literature anchors is used to construct the conditional probability table (CPT) for each node. The hybrid method includes: for each parent node state combination, traversing the preset set of literature anchors for that node, filtering candidate anchors that satisfy all activation conditions of the current parent node state combination, selecting the candidate anchor with the most activation conditions from the candidate anchors that satisfy the conditions as the optimal matching anchor (maximum specificity principle), and normalizing its literature source probability distribution as the conditional probability of the state combination; if there are no literature anchors that satisfy the conditions, a synthetic prior distribution based on the weighted offset of the parent node's influence intensity is used. The synthetic prior distribution is assigned different basic biases according to the node's level (the compensation mode layer L4 and the secondary damage layer L5 nodes use low state biases, and the other layer nodes use high state normal biases). S3: Receives clinical evidence set E as input, performs Bayesian posterior inference through likelihood-weighted sampling, traverses each node in NRI-DAG topological order, fixes the observation state of evidence nodes and multiplies the conditional probability likelihood to the sampling weight, samples latent variable nodes forward according to the current conditional probability distribution, and uses weighted counting to statistically analyze the posterior probability distribution of query nodes. S4: For the candidate intervention set, the average intervention benefit (ATE) of each intervention scheme is quantified by Pearl do-calculus graph surgery. The graph surgery is implemented as follows: during the likelihood weighted sampling process, the target state of the intervention node is fixed but the conditional probability likelihood is not multiplied into the sampling weight, which is equivalent to truncating all incoming edges of the intervention node. The difference between the comprehensive functional prognosis score of the sixth layer after intervention and the baseline score is calculated as ATE. The intervention priority is output in descending order of ATE. S5: Batch query the posterior probability distribution of the fourth-layer compensation mode nodes and the fifth-layer secondary damage nodes, identify the activated compensation mode according to the weighted compensation scoring formula, and detect the activation status of each vicious cycle in the preset vicious cycle set; wherein the vicious cycle is a positive feedback structure composed of multiple mutually reinforcing pathological state nodes, and the adverse states of each participating node promote each other through causal edges to form a self-sustaining pathological mode. S6: Perform leave-one-out (LOO) evidence sensitivity analysis. For each evidence node, calculate the weighted comprehensive impact score of the change in the sixth-level functional prognosis score and the change in the fourth-level compensation score caused by its removal. Output the evidence nodes in descending order of their comprehensive impact scores.

2. A neural rehabilitation clinical reasoning system based on multi-layer Bayesian causal networks, characterized in that, include: The model management module (13) stores the topology, node state definition and mixed conditional probability table of the six-layer neurorehabilitation causal directed acyclic graph. The six layers include the lesion feature layer (L1), the neurophysiological mechanism layer (L2), the clinical syndrome layer (L3), the compensation mode layer (L4), the secondary injury layer (L5) and the functional prognosis layer (L6). The evidence collection module (12a) collects multi-source clinical evidence, including lesion imaging features, neurological function assessment results, and clinical scale scores. The inference engine module (13a) performs likelihood-weighted sampling Bayesian inference based on the clinical evidence set to obtain the posterior probability distribution of each node; The intervention simulation module (13b) uses Pearl do-calculus surgery to perform counterfactual reasoning on candidate intervention programs, calculate the average intervention benefit of each intervention and output a priority ranking; The compensation analysis module (13c) identifies activated compensation patterns, detects the activation state of vicious cycles, and assesses the risk of secondary damage. The sensitivity analysis module (13d) performs leave-one-out sensitivity analysis to quantify the impact of each evidence node on functional prognosis. The report generation module (16) integrates the reasoning results of each module to generate a structured clinical report that includes neurophysiological status assessment, compensation analysis, intervention recommendations and functional prognosis prediction.

3. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the neurorehabilitation clinical reasoning method based on multilayer Bayesian causal networks as described in claim 1.

4. The method as described in claim 1, characterized in that, In step S2, the maximum specificity matching algorithm for the document anchor points includes: traversing all preset document anchor points of the node, filtering anchor points that satisfy all activation conditions of the current parent node state combination, selecting the anchor point with the most activation conditions from the anchor points that satisfy the conditions as the optimal matching anchor point, normalizing its document source probability distribution and truncating it to the range of [0.01, 0.98], and using it as the conditional probability distribution under the parent node state combination.

5. The method as described in claim 1, characterized in that, In step S2, the basic priors of the synthetic prior distribution are allocated hierarchically as follows: the nodes of the compensation mode layer (L4) and the secondary damage layer (L5) adopt low-state bias priors with a bias coefficient of -0.15; the nodes of the functional prognosis layer (L6) adopt uniform priors. The remaining layer nodes adopt a high-state bias prior with a bias coefficient of +0.

10. Based on the basic prior, the comprehensive influence intensity is calculated according to the formula influence = Σⱼ(normalized state value of parent node j - 0.5)×edge_weight(parent node j → current node) / number of parent nodes. The final distribution is obtained by normalizing the influence offset after superimposing it with a superposition ratio of α (default 0.6).

6. The method as described in claim 1, characterized in that, In step S4, the essential difference between Pearl do-calculus graph surgery and observation conditionation is that the conditionation operation treats the intervention node as an evidence node, and its conditional probability likelihood is multiplied into the sampling weight, causing confounding variables to affect the inference result through the common cause path, resulting in the intervention benefit being systematically overestimated; the graph surgery operation fixes the state of the intervention node but does not multiply the conditional probability likelihood, truncates the statistical dependency between the intervention node and its causal variables, achieves an unbiased estimate of the intervention effect, and correctly simulates the clinical scenario of "medical intervention forcibly fixing the node to the target state".

7. The method as described in claim 1, characterized in that, In step S5, the weighted compensation score formula is: score(L4i) = P(L4i = high) + 0.5 × P(L4i = medium); in the vicious cycle activation detection, the activation ratio is calculated as: activationratio = |{participating cycle node v: v is in an unfavorable state (through direct observation or Bayesian inference)}| / |total number of participating cycle nodes|; when the activation ratio is not less than 0.4, the vicious cycle is determined to be in an activated state, and the corresponding clinically recommended intervention target is output.

8. The method as described in claim 1, characterized in that, In step S6, the formula for calculating the combined impact score is: combined_impact(eᵢ) = 0.6 × |ΔL6(eᵢ)| + 0.4 × |ΔL4(eᵢ)|; where ΔL6(eᵢ) is the absolute change in the sixth-level functional prognosis comprehensive score after removing the evidence node eᵢ, and ΔL4(eᵢ) is the absolute change in the fourth-level compensation mean score after removing the evidence node eᵢ; nodes with a combined impact score greater than 0.05 are marked as high-impact nodes, and it is recommended to collect them accurately; nodes with a combined impact score less than 0.02 are marked as low-impact nodes, and the calculation can be simplified when assessment resources are limited.

9. The method as described in claim 1, characterized in that, The process also includes step S7: obtaining the three-dimensional coordinates of 33 key skeletal points of the subject based on skeletal posture estimation technology, and scoring nine compensatory movement patterns using geometric calculation methods. The nine compensatory movement patterns include: trunk tilt, scapular elevation, trunk flexion, asymmetrical upper limb swing, insufficient weight-bearing on the affected side, pelvic rotation, knee flexion compensation, elbow flexion fixation, and head and neck compensation. The scores of each compensation pattern are normalized to the [0,1] interval, and the scoring results are used as supplementary evidence for the fourth layer node and input into the NRI inference engine.

10. The method as described in claim 9, characterized in that, The geometric scoring methods for the nine compensatory movement patterns are as follows: trunk tilt score is obtained by calculating the absolute value of the longitudinal coordinate difference between the bilateral acromion key points; scapular elevation score is obtained by the longitudinal height deviation between the shoulder key points on the affected and healthy sides; trunk flexion score is obtained by the coordinate difference between the midpoint of the hip joint and the midpoint of the shoulder joint in the sagittal plane; upper limb swing asymmetry score is obtained by the variance ratio of the lateral displacement of the bilateral wrist joints; insufficient weight-bearing on the affected side score is obtained by the longitudinal coordinate difference between the bilateral ankle joints; and pelvic rotation score is obtained by calculating the angle between the line connecting the hip joints and the forward direction.

11. The method as described in claim 1, characterized in that, Step S3 also supports batch multi-target queries: during a single topology traversal sampling process, the weighted counts of multiple target nodes are counted simultaneously, with a default sampling count of 400 times; Step S4’s intervention benefit quantification supports batch parallel computation, performing independent graph surgery simulations on all candidate intervention schemes under the same evidence set.

12. The method as described in claim 1, characterized in that, The NRI-DAG pre-sets at least five clinically important vicious cycles, namely: spasticity-compensation cycle, disuse-pain cycle, tension-contraction cycle, compensation-functional decline cycle, and neglect-disuse cycle. Each vicious cycle pre-sets a set of evidence-based recommended clinical intervention target nodes.

13. The method as described in claim 1, characterized in that, In step S2, the preset literature anchor points include at least 82 anchor point entries, covering nine key clinical nodes: corticospinal tract integrity node, stretch reflex threshold range node, spasticity degree node, upper limb Fugl-Meyer score node, whole body compensation pattern node, shoulder pain node, learned disuse node, Barthel index node, and modified Rankin scale node. The probability distribution of the anchor points comes from evidence-based clinical studies published by Stinear, Byblow, Gracies, Feldman, Veerbeek, etc.

14. The system as described in claim 2, characterized in that, The inference engine module (13a) also includes a precise variable elimination submodule, which uses a variable elimination algorithm to perform precise Bayesian inference; when the precise inference exceeds a preset time threshold, it automatically degrades to likelihood-weighted sampling approximate inference, thereby achieving an adaptive balance between inference accuracy and real-time response.

15. The system as described in claim 2, characterized in that, It also includes an AI report generation submodule, which calls the large language model interface to compress the NRI inference results into structured input and generate a natural language clinical report containing prognostic description, intervention rationale and rehabilitation goals; the compression process removes probability terms with confidence levels below a preset threshold and truncates the intervention recommendation list to the first K terms in order to control the input scale of the large language model and improve the clinical relevance of the report.

Images