A system and method for auxiliary assessment of lumbar disc herniation based on plantar pressure gait data
The lumbar disc herniation assessment system, which combines hardware and software, solves the problems of complex equipment, unstable data, and unclear result interpretation in existing technologies. It enables auxiliary screening and assessment of the degree of functional impairment of lumbar disc herniation, and provides efficient data management and result traceability.
Patent Information
- Application Number
- CN202611171476.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-04
- Publication Date
- 2026-11-17
AI Technical Summary
Existing lumbar disc herniation assessment systems suffer from problems in gait data processing, including high equipment costs, complex operation, difficulty in reflecting the force distribution in different areas of the foot, unstable signal acquisition, unclear interpretation levels of results, and inconsistent data management. Furthermore, they are difficult to use to assess the degree of functional impairment.
The system employs a hardware acquisition system and a host computer software system to work together. Through bipedal pressure data acquisition, standardized gait sample construction, dual-task temporal analysis, and session-level result aggregation, it generates a structured assessment report and links the results to cloud archiving, thereby enabling auxiliary screening and functional impairment assessment for lumbar disc herniation.
It improves the consistency and traceability of gait assessment data, avoids misjudgment of phase difference caused by alternating left and right foot landing, retains the relative load differences before channel-by-channel standardization, and provides clear data correspondence and data basis for multiple assessments.
Smart Images

Figure CN123120169A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of smart healthcare, wearable biomechanical signal acquisition and time-series data processing, and specifically to an auxiliary assessment system and method for lumbar disc herniation based on plantar pressure gait data, comprising a hardware acquisition system, a host computer software system, a report generation module and a cloud data management module. Background Technology
[0002] Lumbar disc herniation is a common degenerative spinal disease, usually related to changes such as lumbar disc degeneration, annulus fibrosus damage, and nucleus pulposus protrusion. The herniated tissue can irritate or compress nerve roots, causing low back pain, radiating pain in the lower extremities, numbness, paresthesia, decreased muscle strength, and limited walking function. Affected by factors such as pain stimulation, nerve root compression, and abnormal neuromuscular control, patients may exhibit abnormal manifestations such as decreased gait speed and cadence, shortened stride length, prolonged double support phase, and asymmetry in gait.
[0003] The aforementioned gait changes can also be accompanied by alterations in lower limb load transfer patterns, manifesting as abnormalities in pressure amplitude, contact time, pressure change sequence, and the relationship between the left and right feet at different biomechanical locations on the sole. Studies have shown that patients with lumbar disc herniation may differ from healthy individuals in gait spatiotemporal parameters, gait symmetry, and plantar pressure distribution. Therefore, plantar pressure gait data can provide objective technical data for characterizing walking dysfunction associated with lumbar disc herniation from the perspective of lower limb movement and force application.
[0004] Clinically, lumbar disc herniation usually requires a comprehensive assessment combining medical history, physical examination, and imaging studies. Scales such as the Oswestry Disability Index (ODI) are primarily used to evaluate the impact of low back and leg pain on daily activities and functional status. Existing gait assessment methods include clinical observation, optical motion capture, label-free visual analysis, force tables, pressure tracks, inertial sensors, and plantar pressure sensors. Among these, clinical observation is simple to operate but heavily relies on the experience of the assessor; optical motion capture, force tables, and pressure tracks can obtain comprehensive kinematic or kinetic parameters, but they typically suffer from high equipment costs, complex operating procedures, limited testing space, and are not suitable for long-term continuous monitoring; label-free visual analysis is easily affected by occlusion and joint positioning accuracy; inertial sensors cannot directly reflect the force distribution in different areas of the foot and may be affected by factors such as installation location and signal drift.
[0005] With the development of wearable sensor technology, smart shoes and smart insoles can continuously collect plantar pressure data in natural walking scenarios, showing potential applications such as portability, dynamic monitoring, and long-term follow-up. However, existing related systems are mostly geared towards general gait monitoring, typically focusing on displaying pressure curves, plantar pressure cloud maps, or calculating routine gait parameters. They lack specialized modeling for lumbar disc herniation-related left-right foot force asymmetry, local pressure waveforms and their bimodal variations, as well as long-term temporal dependencies between multiple gait cycles. In some systems, signal acquisition, sample construction, intelligent analysis, and result management are separated, and a complete technical chain for stably converting plantar pressure signals collected by wearable devices into disease-related auxiliary assessment information has not yet been formed.
[0006] For bipedal pressure data, directly inputting the left and right foot channels into a general time-series model in a fixed order presents three easily overlooked processing issues: First, independent normalization of the left and right corresponding regions may compress the true relative load differences; second, during normal walking, the left and right feet alternately strike the ground, and directly comparing pressure values at the same sampling time will introduce gait phase misalignment; third, the affected side of the patient is not fixed, and the model may mistakenly interpret the fixed left and right channel positions as disease features. Therefore, it is necessary to explicitly complete the pairing of corresponding regions, phase alignment of the left and right support phases, and left-right mirror consistency processing before or at the beginning of the model, so that asymmetric features originate from comparable gait stages and are modeled together with local waveforms and cross-cycle changes.
[0007] Existing intelligent analysis solutions based on gait data often focus on classifying and identifying patient samples and healthy controls, resulting in relatively singular output dimensions. This makes it difficult to simultaneously assess the degree of functional impairment based on a unified scale mapping rule while simultaneously assisting in the screening of lumbar disc herniation. Some solutions also fail to clearly distinguish between gait feature analysis results, disease-assisted screening information, functional impairment assessment results, and clinical diagnostic conclusions, leading to unclear interpretation levels and application boundaries for the output results. Furthermore, raw plantar pressure data, visualization results, assessment reports, and subject information typically lack a unified mechanism for coordinating and managing these elements, hindering multiple assessments, efficacy comparisons, and continuous follow-up. Summary of the Invention
[0008] Purpose of the invention: This invention aims to provide an auxiliary assessment system and method for lumbar disc herniation based on plantar pressure gait data. Through the collaboration of a hardware acquisition system and a host computer software system, it realizes the acquisition of bipedal pressure data, construction of standardized gait samples, dual-task temporal analysis, and session-level result summarization. The analysis results are linked with report generation and cloud archiving, providing objective and traceable information for auxiliary screening, functional impairment assessment, and follow-up management of lumbar disc herniation.
[0009] Technical Solution: To achieve the above objectives, this invention provides an auxiliary assessment system for lumbar disc herniation based on plantar pressure gait data, comprising: a hardware acquisition system, host computer software, a report generation module, and a cloud data management module; wherein,
[0010] The hardware acquisition system collects pressure signals from preset biomechanical regions of both feet, generating and transmitting multi-channel pressure time-series data with timestamps. The host computer software communicates with the hardware acquisition system, performing validity screening, time window segmentation, channel-by-channel normalization, and channel combination on the multi-channel pressure time-series data to generate plantar pressure gait samples. It extracts local and long-term temporal features, and outputs in parallel the auxiliary screening results for lumbar disc herniation and the functional impairment assessment results based on the Oswestry Disability Index. It also performs quality-weighted summarization of valid window results within the same assessment session. The report generation module links session-level auxiliary assessment results, subject information, quality records, and visualization charts to form a structured assessment report. The cloud-based data management module archives subject information files, raw plantar pressure data files, quality records, and assessment report files according to the assessment session.
[0011] This invention provides an auxiliary assessment method for lumbar disc herniation based on plantar pressure gait data, comprising the following steps:
[0012] Step 1: Collect pressure signals from preset areas of both feet using a hardware acquisition system, generate multi-channel pressure time-series data with timestamps, and perform local backup and wireless transmission;
[0013] Step 2: The received data is time-aligned, validity-filtered, time-window-segmented, and channel-by-channel-normalized using the host computer software. The bipedal pressure channel and the combined force channel of the left and right feet are combined to form a plantar pressure gait sample.
[0014] Step 3: Extract local temporal features and long-term temporal features from the plantar pressure gait samples respectively to obtain the window-level lumbar disc herniation auxiliary screening category probability and the functional impairment degree category probability based on the Oswestry Disability Index four-level label;
[0015] Step 4: Summarize the probabilities of the two window-level categories within the same evaluation session according to the proportion of valid data in the window, and generate session-level auxiliary evaluation results or insufficient data indicators;
[0016] Step 5: Use the report generation module to link the conversation-level auxiliary assessment results, subject information, quality records, and visualization charts to form a structured assessment report;
[0017] Step 6: Archive the subject information file, raw plantar pressure data file, quality record and assessment report file by assessment session through the cloud data management module.
[0018] Specific implementation: Further, in one embodiment, pressure sensors are respectively installed in the corresponding areas of the heel, first metatarsal, third metatarsal, and fifth metatarsal on the left and right shoe bodies. When a full-foot pressure sensor array is used, the array data can be aggregated according to the above-mentioned biomechanical regions.
[0019] Furthermore, the shoe body may be made of flexible thermoplastic polyurethane material, and the sole may have mounting slots corresponding to the pressure sensor. The power management unit may include a rechargeable battery, a charging interface, a voltage regulator circuit, and a battery protection circuit.
[0020] Furthermore, each analysis window contains 500 time points. The data preprocessing module uses valid data from the same evaluation session to determine the basic normalized parameters for the eight regional pressure channels of both feet, and calculates the resultant force channels for the left and right feet based on these regional pressure channels, thus constructing a 500×10 basic gait sample consisting of eight regional pressure channels and two single-foot resultant force channels. To avoid the basic channel-by-channel normalization weakening the relative amplitude differences between the two feet, the system simultaneously retains the calibrated original pressure view, or performs pairwise normalization for each pair of left and right corresponding regions using a common minimum, maximum, or robust scale for subsequent symmetry feature calculations.
[0021] The window quality checks include channel missing checks, abnormal constant value checks, timestamp continuity checks, effective time point ratio checks, bipedal contact state checks, and complete left and right support pairing checks. Windows that do not meet the preset quality requirements are not included in the session-level result aggregation, or their weights are reduced according to their effective data ratio and complete gait pairing ratio; windows that only lack complete paired support phases but whose basic pressure timing is still valid can retain the basic model output without generating symmetrical branch outputs.
[0022] Furthermore, the local temporal model can employ a one-dimensional convolutional network. For the four corresponding left and right regions of the heel, first metatarsal, third metatarsal, and fifth metatarsal, the pressure sequence and first-order temporal difference are respectively fed into the left and right convolutional branches with shared parameters. The shared convolutional kernel ensures that the same local waveform detector operates on the left and right feet with the same parameters, reducing bias caused by fixed foot sides and individual sensor differences. The outputs of each group of left and right branches are further combined into feature sums, absolute differences, normalized differences, and / or element-wise products, and fused with the left and right force imbalance features through 1×1 convolutions or fully connected layers for cross-regional fusion. The convolutional kernel length can be set to 15 time points, corresponding to a local time range of approximately 150 ms at 100 Hz sampling, used to extract pressure changes such as the rising edge of landing, the falling edge of takeoff, local peaks, or bimodal morphology. After the second-level convolution expands the local receptive field, the class probability is output after random deactivation, temporal convergence, and classification mapping.
[0023] Furthermore, the long-range time series model can employ the Mamba time series model or an equivalent selective state-space network. Unlike the approach that simply projects the 10 pressure channels directly, this embodiment incorporates the first-order time difference of pressure, left / right contact state, the normalized phase position of the current support phase, and the asymmetric features of the region that can be mapped to that phase into the input markers at each time step. The input-related state parameters are generated jointly by these markers, enabling the model to retain more state information during periods of rapid pressure change, load transfer, and increased bipedal differences, while suppressing continuous non-contact periods or small acquisition disturbances. Multiple complete gait cycles are processed in the order of acquisition to extract consistency, volatility, and cumulative changes between cycles. In the offline evaluation implementation, forward and reverse state encoding can also be performed separately and their convergence features fused.
[0024] The local and long-range temporal models share the basic sample construction and bipedal symmetry feature construction rules, but the model structure and output head are trained independently. The default configuration uses a one-dimensional convolutional model for the binary classification auxiliary screening task and a Mamba temporal model for the ODI four-class classification task. An optional cross-configuration adds an ODI four-class auxiliary output head to the one-dimensional convolutional model and a binary classification auxiliary output head to the Mamba temporal model. Cross-execution targets different structural outputs for the same label task; it does not derive four-class results from binary classification results, nor does it infer disease diagnosis labels from ODI categories.
[0025] Furthermore, considering that the affected side may be located on the left or right side, during model training, the corresponding region channels of the left and right feet and the left and right combined force channels are interchanged with a preset probability, and the sign of the signed difference feature is changed simultaneously, while the disease binary classification label and ODI level label remain unchanged. During inference, the probabilities of the original sample and the left and right mirror images can be calculated separately and averaged. If the clinical information records the affected side, the input can be uniformly rearranged into the affected side / non-affected side order, while retaining the original foot side label for report display.
[0026] Furthermore, the host computer software system retains the main output probability, optional auxiliary output probability, effective data ratio, complete left and right support pairing ratio, differences between the two models for the same task, and quality records for each analysis window. It then weights and summarizes the category probabilities of the effective windows according to the effective data ratio, pairing completeness, and model consistency. When the number of effective windows, the amount of effective data, the number of complete paired gaits, model consistency, or the summary confidence level are lower than the corresponding thresholds, it does not directly output a deterministic evaluation result. Instead, it outputs a flag indicating insufficient data or that the result needs verification, prompting for re-collection, supplementary data collection, or verification by professionals.
[0027] Furthermore, the host computer software system can be configured with interfaces for session establishment, real-time monitoring and visualization, intelligent analysis, report generation, and data management and cloud synchronization. These interfaces are used to bind subjects and devices, display stress data, call analysis models, generate evaluation reports, and trigger session data packet uploads, respectively.
[0028] Furthermore, the report generation module and cloud data management module can be deployed as software functional modules on the host computer or in the cloud, and maintain association with hardware-collected data and host computer analysis results through a unified evaluation session identifier.
[0029] Before uploading the session data package, the cloud-based data management module performs document existence and consistency checks on the participant information file, raw data file, quality records, analysis results, and evaluation report. In the event of an upload failure, the complete session data package is retained locally and re-upload is permitted to maintain the correlation and traceability between various data types.
[0030] Furthermore, the Oswestry Disability Index (ODI) label is used to characterize the degree of functional limitation associated with low back pain, and is not used to characterize the radiographic or pathological severity of lumbar disc herniation. The auxiliary screening results and functional impairment assessment results are intended to provide supplementary reference for professionals and do not replace clinical diagnosis.
[0031] Beneficial effects: Compared with the prior art, the present invention has at least the following beneficial effects:
[0032] (1) Connect bipedal pressure collection, standardized sample construction, intelligent analysis, result summarization, report generation and cloud archiving into a unified assessment process to improve the consistency and traceability of gait assessment data.
[0033] (2) By adding left and right support pairing, same region pairing normalization and explicit symmetry features in addition to the basic pressure sample, the phase difference caused by the alternating landing of the left and right feet is avoided as load asymmetry, and the relative load difference before channel-by-channel standardization is retained.
[0034] (3) The one-dimensional convolution model uses local convolution with shared weights for the left and right regions with the same name, and explicitly integrates the sum and difference of the left and right features, so that the local pressure waveform changes and the asymmetry of the bipedal load have a clear data correspondence.
[0035] (4) The Mamba time series model generates input-related state update parameters by combining pressure changes, contact state, gait phase and symmetry information. It selectively retains load transfer and asymmetric changes within multiple gait cycles, thereby forming an input representation and state update method that is tailored to the long-term characteristics of plantar pressure.
[0036] (5) By setting optional binary and quadruple cross-output heads for one-dimensional convolutional models and Mamba time series models, heterogeneous model consistency verification can be performed on the same task; when the output difference is large, a verification mark is generated, thereby avoiding mechanically binding the two tasks to a single model structure.
[0037] (6) By checking window quality, checking complete gait pairing, weighted quality summarization and data shortage prompts, the impact of missing channels, data interruption, insufficient effective gait or inconsistent output of dual models on session-level results can be reduced.
[0038] (7) By using a unified assessment session identifier to link subject information, raw data, symmetry indicators, model consistency records, analysis results and assessment reports, a data foundation is provided for multiple assessment comparisons and follow-up management. Attached Figure Description
[0039] Figure 1 This is a hierarchical and data flow diagram of the hardware acquisition system, host computer software system, report generation module, and cloud data management module of the present invention;
[0040] Figure 2 These are actual images and exploded views of the smart shoe of this invention;
[0041] Figure 3 This is the PCB schematic diagram of the present invention;
[0042] Figure 4 This is a diagram of the one-dimensional convolutional neural network structure of the auxiliary screening unit of this invention;
[0043] Figure 5 This is a schematic diagram of the Mamba block of the four-category assessment unit for the degree of functional impairment of the present invention;
[0044] Figure 6 This is the host computer interface of the present invention. Detailed Implementation
[0045] The technical solution of the present invention will be further described below with reference to the accompanying drawings.
[0046] like Figure 1 As shown, the auxiliary evaluation system of this invention includes a hardware acquisition system, host computer software, a report generation module, and a cloud data management module. These four modules are linked through a unified evaluation session identifier; the latter two modules can be deployed on a host computer or in the cloud, but respectively handle report generation and session data archiving.
[0047] Hardware data acquisition system. The hardware data acquisition system includes a pair of left and right shoe components, pressure sensors, a signal acquisition and processing unit, a wireless communication module, and a power management unit. The system synchronously acquires pressure data from both feet, writes device identifiers, channel identifiers, and timestamps to each channel, and performs local backup and wireless transmission.
[0048] After the host computer software establishes an evaluation session, it verifies the device identifiers, data packet sequence numbers, and timestamps of the left and right shoes respectively, and aligns the data of both feet according to the timestamps; in the event of a short-term communication interruption, the data source can be supplemented by local backup.
[0049] The acquisition circuit converts the resistance changes of four flexible piezoresistive sensors inside each shoe into voltage signals. In one embodiment, the sensor pull-up voltage is 3.3 V, and a 74HC4051 or equivalent multiplexer is used to poll each channel. A voltage follower is set at the output to increase the input impedance and isolate the subsequent circuitry. The analog-to-digital converter converts the 0-3.3 V analog voltage into a digital signal, and the microcontroller completes the packaging and storage.
[0050] Four pressure sensors are installed on the left and right shoe bodies at the heel, first metatarsal, third metatarsal, and fifth metatarsal positions, respectively. When using a full-foot pressure sensor array, the array channels are aggregated according to the above-mentioned areas. The shoe body can be made of flexible thermoplastic polyurethane material with pre-set mounting slots; the power management unit includes a rechargeable battery, a charging interface, and voltage regulation and protection circuitry.
[0051] The host computer software system includes a data preprocessing module, a data visualization module, and an intelligent analysis module, which respectively complete bipedal alignment and sample construction, stress curve and distribution display, dual-task analysis, and window quality recording.
[0052] The first analysis branch undertakes the task of auxiliary screening for lumbar disc herniation, employing a one-dimensional convolutional model. The model takes 500×10 baseline gait samples and temporal difference and symmetry features constructed from their corresponding original / paired normalized pressures as input. It uses a one-dimensional convolutional branch with shared parameters and a kernel length of 15 for four groups of left and right corresponding regions. Then, it performs cross-channel fusion of the sum and absolute difference of left and right local features, as well as the left-right force imbalance features, to directly characterize local pressure changes and bipedal load asymmetry. Subsequently, it performs random inactivation, temporal convergence, and fully connected mapping, outputting window-level probabilities of lumbar disc herniation samples and healthy control samples via Softmax. This model can also be configured with an ODI four-class auxiliary output head.
[0053] The second analysis branch undertakes the task of assessing the degree of ODI (Obstructive Disorder Injury) using a Mamba time-series model. In one embodiment, 500×10 baseline pressure, corresponding first-order time difference, bipedal contact state, normalized support phase position, and symmetry features are mapped to a 64-dimensional latent space via a linear layer or 1×1 convolution. At least two concatenated selective state space blocks generate state updates and gating parameters using the current input labels, selectively accumulating load transfer and asymmetric changes over multiple gait cycles covered by 500 time steps. The model outputs four probabilities of functional impairment after time-averaged pooling, multilayer perceptron, and Softmax. The four-level labels are generated by dividing the ODI percentage using three incremental thresholds, which are saved with the model configuration and remain unchanged during the inference phase. This model can also be configured with a binary classification auxiliary output head for lumbar disc herniation.
[0054] One example dataset includes 118 participants clinically diagnosed with lumbar disc herniation by a professional physician and 130 healthy controls, totaling 248 participants. Clinical diagnosis labels are used for the first analysis branch; samples that completed the ODI scale assessment are labeled for the second analysis branch according to pre-training fixed rules. Foot pressure data at 100 Hz are collected from each participant for 30 to 200 seconds; continuous recordings are windowed in 500-point (5-second) segments, with segments shorter than 500 points excluded from analysis. The dataset is divided by participant to avoid the same participant's window being included in both the training and test sets simultaneously.
[0055] The host computer processing flow in this embodiment is summarized in the following six steps; the thresholds, number of network layers, and training parameters for each step can be adjusted according to the data scale and equipment conditions:
[0056] (1) Session establishment and data reception: Establish an evaluation session and bind the subject information and bipedal device identifier, verify the data packet sequence number, timestamp and channel identifier, and write the continuous pressure time sequence into the session cache and local raw data file.
[0057] (2) Window quality control: The continuous record is divided according to the preset window length; the original numerical domain is checked for missing channels, abnormal constant values, discontinuous timestamps, effective time point ratios, and whether the combined force of the two feet is continuously lower than the threshold. Windows that do not meet the requirements are removed and the reasons for removal are recorded.
[0058] (3) Channel construction, phase pairing and standardization: The pressure of the four regions of the left foot and the right foot is summarized to form two resultant force channels, and combined with the pressure channels of the eight regions of both feet in a fixed order to form 500×10 basic gait samples; In addition to the basic channel-by-channel normalization, the original calibration pressure is retained or the joint statistics of the left and right corresponding regions are used to perform shared scale normalization; The complete support phase is extracted according to the rising and falling edges of the left and right resultant forces exceeding the contact threshold, and a set of temporally adjacent left and right support phases are resampled to the same phase length to form paired gait data that can be compared phase by phase, and the regional time difference and bipedal symmetry characteristics are calculated.
[0059] (4) Dual-structure, dual-task analysis: The basic gait samples and bipedal symmetry features are input into the local temporal model and the long-range temporal model, respectively; by default, the binary classification head of the one-dimensional convolutional model and the four-class classification head of the Mamba temporal model are used as the main output. Optionally, the four-class classification head of the one-dimensional convolutional model and the binary classification head of the Mamba temporal model are enabled simultaneously to form auxiliary probabilities with different structures but consistent labels for the same task; the main and auxiliary probabilities are weighted and fused based on validation set performance or calculated for consistency differences, and the labels are not derived between the binary classification and the four-class classification.
[0060] (5) Session-level summary: Based on the proportion of effective data in each window, the proportion of complete left and right support pairings, and the weight allocation of the consistency of the two models in the same task, summarize the binary and four-class probabilities in the same evaluation session respectively; when the number of effective windows, the total amount of effective data, the number of complete pairing gaits, the consistency of the two models, or the summary confidence is insufficient, no definite category will be output, and the generated data will be insufficient or the result needs to be reviewed and the quality reasons will be indicated.
[0061] (6) Results output and archiving: Save window-level main output probability, optional auxiliary output probability, model difference, symmetry index, quality score, reasons for elimination, session-level results and visualization charts. The report generation module generates a structured evaluation report, which is then archived by the cloud data management module according to the evaluation session.
[0062] Based on the above written flowchart, the following is an algorithmic expression to illustrate its feasibility. Let the calibration pressure matrix of the effective window be... The duration is The original number of channels in each region is 8. The base gait samples can be further normalized by calculating the minimum and maximum values for each channel within the same evaluation session, using the following formula; when the denominator is 0, the corresponding channel is set to 0. The bipedal symmetry branch is calculated by calibration pressure before this independent normalization, or using the shared scale normalization result of the same region described later, to preserve the left-right relative amplitude relationship:
[0063]
[0064] After normalization, eight regions of the two feet are retained according to a fixed channel order, and the resultant force representation channels of the left and right feet are calculated separately. If channels 1-4 correspond to the four regions of the left foot and channels 5-8 correspond to the four regions of the right foot, then:
[0065]
[0066]
[0067] For the The left and right regions with the same name in the group. Within the valid data of the same evaluation session, the scale is jointly determined by both the left and right sides. and And calculate the pairwise normalized sequences L̄ᵢ and R̄ᵢ:
[0068]
[0069]
[0070]
[0071]
[0072] The complete support phases are obtained from the start and end points where the combined left and right forces exceed the corresponding contact thresholds. A set of temporally adjacent left and right support phases are then interpolated to... Normalized phase points This process compares the position of the left and right feet at the same percentage point in the support phase, rather than directly comparing the same absolute sampling time during alternating walking. For the first... Groups of the same name and the resultant force of both feet can construct a sign difference. Normalized asymmetry and resultant force imbalance :
[0073]
[0074]
[0075]
[0076]
[0077]
[0078] In addition to the phase-by-phase sequence mentioned above, the contact duration difference, peak value difference, pressure-time integral difference, and peak phase difference of the left and right support phases can also be calculated and used as window-level symmetry statistics. If multiple complete left and right support phases exist within the window, the characteristics of each group are retained according to the acquisition order, enabling the long-range time series model to analyze the asymmetric changes between multiple gait cycles.
[0079] Therefore, we obtain Gait sample matrix And obtain the first-order difference of pressure, contact state, gait phase position and , , Symmetry enhancement representation of the composition . and Input channels can be concatenated, or they can be encoded separately and then fused at the temporal convergence feature layer. The local temporal branch generates auxiliary screening main probabilities and optional ODI auxiliary probabilities through shared-weight convolution and left-right pairwise fusion; the long-range temporal branch first maps the fused input to a sequence representation e. t Then, the input-related state recursion is performed. Its core processing can be summarized as follows:
[0080]
[0081]
[0082] Among them, Āt, B̄t, and Ct are derived from the current input e, which includes information on pressure changes, contact state, gait phase, and bipedal symmetry. t Generate, h t The hidden state is represented by the temporal convergence feature, which connects the ODI quad-classifier head and the optional disease binary classifier head, respectively. Similarly, the convergence feature of the one-dimensional convolutional branch connects the disease binary classifier head and the optional ODI quad-classifier head, respectively. This representation is used to summarize the relationship between plantar pressure enhancement input, selective state update, temporal convergence, and classification mapping. The specific state dimension, number of layers, and gating form can be adjusted according to the training data scale.
[0083] Suppose that the same task is output by a one-dimensional convolutional model and a Mamba temporal model, respectively, resulting in probabilities pCNN and pMamba. Then the differences in fusion probability and structural consistency can be expressed as follows:
[0084]
[0085]
[0086] Among them, the binary classification task can be increased. With a bias towards one-dimensional convolution as the main output, the four-class classification task can be reduced. The bias is towards the main Mamba output, with specific weights determined by the performance of the validation set or the quality of the window based on the participants. When the preset threshold is exceeded, the system retains the original probabilities of the two models and generates a result verification mark; this consistency judgment only compares the outputs within the same label space.
[0087] For N valid windows within the same assessment session, let pn be the category probability of the nth window and qn be the proportion of valid data. Then, the session-level probability P can be obtained using the following formula; the auxiliary screening branch and the functional impairment degree branch are summarized separately:
[0088]
[0089] When N is lower than the minimum number of valid windows, Σqn is lower than the valid data volume threshold, or max(P) is lower than the summary confidence threshold, the system outputs a data insufficiency flag and the corresponding quality reason, but does not output a specific category. Window-level results, quality scores, reasons for rejection, and session-level results are saved simultaneously for verification of their source.
[0090] Model training employs the AdamW optimizer and cross-entropy loss, and the subjects are divided into training, validation, and test sets. In one embodiment, the initial learning rate of the one-dimensional convolutional model is 1×10⁻³, the weight decay coefficient is 1×10⁻⁴, and the Dropout inactivation probability is 0.3–0.5; the initial learning rate of the Mamba temporal model is 1×10⁻⁴, the weight decay coefficient is 1×10⁻², and the Dropout inactivation probability is 0.1–0.2. Linear warm-up and cosine annealing are used during training, the gradient norm pruning threshold can be set to 1.0, the batch size can be set to 32–64, and early stopping is implemented by monitoring validation set metrics. The above parameters are used to illustrate possible implementations and do not constitute a limitation on the scope of protection.
[0091] To address the issue of inconsistent left and right foot sides, paired left-right mirror enhancement is performed during the training phase: four sets of corresponding regions, left / right resultant force, and contact state channels are swapped, and the swapped data exhibits signed characteristics. and Inverse, absolute asymmetry The binary and ODI four-class labels remain unchanged. For models configured with both binary and four-class heads, the training loss can be weighted by the cross-entropy loss of the four task heads and the consistency loss between the two models on the same task. Samples that have not completed the ODI scale assessment will only have their corresponding four-class loss masked, without affecting their participation in binary training. The training, validation, and test sets are still divided according to the subjects, and mirrored samples must not cross the original subject's data subset.
[0092] In this embodiment, with the existing data partitioning, the training set accuracy of the first analysis branch is 95%, and the test set accuracy is 85%. This result is only used to illustrate that the process is feasible; actual performance still needs to be verified on independent, consistent data.
[0093] The graphical user interface of the host computer software system includes session establishment, real-time monitoring and visualization, intelligent analysis, report generation, and cloud synchronization. Each interface is linked to data through a unified session identifier and displays window quality records, left and right support pairing status, regional and resultant force symmetry charts, consistency between main output and optional auxiliary output, session-level results, and prompts for insufficient data or data requiring review.
[0094] Report generation module. The report generation module writes subject information, assessment time, data collection quality, stress chart, left and right support pairing information, regional and resultant force symmetry indicators, two types of session-level results, optional dual-model consistency records and quality tips into a structured assessment report, and saves the session number, model version, feature construction version and generation time.
[0095] Cloud-based data management module. In response to the upload operation, the cloud-based data management module assembles the subject information, raw stress data, quality records, analysis results, and assessment report corresponding to the current assessment session into a session data package, and creates a directory according to the subject identifier, assessment date, and session number.
[0096] Before uploading, perform file existence and checksum verification; record progress and status during upload; retain local files for retrying in case of failure. Cloud archiving maintains the correlation between original data, analysis results, and reports.
[0097] This invention provides an auxiliary assessment system for lumbar disc herniation using plantar pressure gait data as technical input. The system outputs auxiliary screening results for lumbar disc herniation through a binary classification model, and outputs assessment results of functional impairment degree through an independent four-class classification model based on a preset ODI four-level mapping rule. Simultaneously, it manages the raw plantar pressure gait data, visualization charts, auxiliary assessment results, and subject information in a correlated manner, thus forming a complete processing flow covering signal acquisition, data preprocessing, sample construction, model analysis, result display, and data management. This provides technical support for objective auxiliary assessment and continuous follow-up in outpatient, community, and home settings.
Claims
1. A system for assisting in the assessment of lumbar disc herniation based on plantar pressure gait data, characterized in that, include: Hardware acquisition system, host computer software, report generation module, and cloud data management module; among them, The hardware acquisition system collects pressure signals from preset biomechanical regions of both feet, generating and transmitting multi-channel pressure time-series data with timestamps. The host computer software communicates with the hardware acquisition system, performing validity screening, time window segmentation, channel-by-channel normalization, and channel combination on the multi-channel pressure time-series data to generate plantar pressure gait samples. It extracts local and long-term temporal features, and outputs in parallel the auxiliary screening results for lumbar disc herniation and the functional impairment assessment results based on the Oswestry Disability Index. It also performs quality-weighted summarization of valid window results within the same assessment session. The report generation module links session-level auxiliary assessment results, subject information, quality records, and visualization charts to form a structured assessment report. The cloud-based data management module archives subject information files, raw plantar pressure data files, quality records, and assessment report files according to the assessment session.
2. The lumbar disc herniation auxiliary assessment system based on plantar pressure gait data according to claim 1, characterized in that, The hardware acquisition system includes a pair of left and right shoe bodies, with pressure sensors installed on the left and right shoe bodies in the corresponding areas of the heel, first metatarsal bone, third metatarsal bone, and fifth metatarsal bone, respectively. When using a pressure sensor array covering the entire sole, the array channels are aggregated according to the above four areas.
3. The lumbar disc herniation auxiliary assessment system based on plantar pressure gait data according to claim 1, characterized in that, The hardware acquisition system also includes a signal acquisition and processing unit, a wireless communication module, and a power management unit. The signal acquisition and processing unit is used to complete analog-to-digital conversion, timestamp addition, local backup, and communication packetization.
4. The lumbar disc herniation auxiliary assessment system based on plantar pressure gait data according to claim 1, characterized in that, The host computer software divides the continuous recording into analysis windows of preset length. For each window, it performs channel missing detection, abnormal constant value detection, effective time point ratio check, and double foot contact state check. For valid windows that pass the check, it determines the normalization parameters of the pressure channels in each region based on the valid data in the same evaluation session and performs channel-by-channel normalization. It calculates the left and right foot combined force channels respectively and combines them with the pressure channels of each region of the double feet in a fixed order to form the plantar pressure gait sample.
5. The lumbar disc herniation auxiliary assessment system based on plantar pressure gait data according to claim 1, characterized in that, The host computer software includes a first analysis branch and a second analysis branch. The two analysis branches use the same plantar pressure gait sample construction rules but are trained independently. The first analysis branch uses a one-dimensional convolution model to extract local temporal features and outputs window-level probabilities of lumbar disc herniation sample categories and healthy control sample categories. The second analysis branch uses a selective state-space model to extract long-term temporal features and outputs the probability of functional impairment category generated according to the preset Oswestry Disability Index four-level mapping rules.
6. The lumbar disc herniation auxiliary assessment system based on plantar pressure gait data according to claim 5, characterized in that, The one-dimensional convolutional model performs two-level convolution, nonlinear activation, and random deactivation along the time dimension, and outputs the auxiliary screening category probability after time dimension convergence and classification mapping; the selective state space model sequentially performs input projection, input-related state update, gating modulation, time dimension convergence, and classification mapping, and outputs the category probability of the degree of functional impairment.
7. The lumbar disc herniation auxiliary assessment system based on plantar pressure gait data according to claim 5, characterized in that, The host computer software assigns weights to each valid window within the same evaluation session according to the proportion of valid data, and performs weighted summation of the window-level category probabilities output by the first analysis branch and the second analysis branch to generate session-level auxiliary screening probability and session-level functional impairment probability. When the number of valid windows, the amount of valid data, or the summation confidence level is lower than the preset threshold, it outputs a data insufficiency indicator and the corresponding quality reason.
8. The lumbar disc herniation auxiliary assessment system based on plantar pressure gait data according to claim 1, characterized in that, The report generation module writes subject information, assessment time, window quality records, stress visualization charts, session-level auxiliary screening results, and session-level functional impairment assessment results into a structured assessment report, and records the session number, model version, and report generation time. The cloud data management module organizes session data packets according to subject identifier, assessment date, and session number, performs file existence and consistency checks before uploading, and retains the complete local session data packet and allows re-uploading if the upload fails.
9. A method for auxiliary assessment of lumbar disc herniation based on plantar pressure gait data, characterized in that, Includes the following steps: Step 1: Collect pressure signals from preset areas of both feet using a hardware acquisition system, generate multi-channel pressure time-series data with timestamps, and perform local backup and wireless transmission; Step 2: The received data is time-aligned, validity-filtered, time-window-segmented, and channel-by-channel-normalized using the host computer software. The bipedal pressure channel and the combined force channel of the left and right feet are combined to form a plantar pressure gait sample. Step 3: Extract local temporal features and long-term temporal features from the plantar pressure gait samples respectively to obtain the window-level lumbar disc herniation auxiliary screening category probability and the functional impairment degree category probability based on the Oswestry Disability Index four-level label; Step 4: Summarize the probabilities of the two window-level categories within the same evaluation session according to the proportion of valid data in the window, and generate session-level auxiliary evaluation results or insufficient data indicators; Step 5: Use the report generation module to link the conversation-level auxiliary assessment results, subject information, quality records, and visualization charts to form a structured assessment report; Step 6: Archive the subject information file, raw plantar pressure data file, quality record and assessment report file by assessment session through the cloud data management module.
