Off-bed activity multi-mode monitoring risk early warning method and system

Abstract

The invention relates to the technical field of medical monitoring, in particular to an off-bed activity multi-mode monitoring risk early warning method and system. According to the method, data such as bed surface pressure, non-contact posture / displacement and vital signs are collected, time alignment and modal confidence evaluation are carried out, cross-modal consistency is calculated and gated fusion is carried out, and a risk score is obtained by combining off-bed urgency degree prediction, postural intolerance physiological response and pipeline traction risk. And outputting a prompt / early warning / alarm and adaptively updating a threshold value and parameters under nursing feedback, thereby reducing false alarm and missing alarm and realizing advanced intervention.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical monitoring, and in particular to a multi-modal monitoring risk early warning method and system for off-bed activities. BACKGROUND

[0002] With the aggravation of aging, the promotion of accelerated recovery surgery (ERAS) and standardized management of intensive care unit (ICU), the risk of off-bed related adverse events at the bedside of hospitalized patients (especially postoperative, sedation / delirium, limited neurological function or patients with multiple catheter indwelling) significantly increases. The so-called off-bed related adverse events include not only the common falls and bed falls, but also dizziness / syncope induced by body position change, transient disturbance of consciousness caused by postural hypotension, and bleeding, catheter dislocation or drainage obstruction caused by catheter, drainage tube and infusion tube being pulled. Clinically, bedside accompanying, restraint or call bell reminder are commonly used to reduce the risk, but these methods have the disadvantages of high labor cost, lagging response and strong subjective dependence. Therefore, using sensor and information processing technology to automatically monitor the off-bed activities of patients and realize risk early warning has become an important direction of intelligent ward and nursing informatization.

[0003] In the prior art, off-bed monitoring usually focuses on whether to get off the bed / abnormal off-bed duration, and uses pressure sensors, infrared sensors, cameras, millimeter wave radars or wearable inertial devices for detection and alarm. For example, Chinese patent document CN111568437A discloses a non-contact off-bed real-time monitoring method, which uses millimeter wave radar to collect electromagnetic echo and identify motion and direction, uses depth camera to obtain image and perform visual analysis, further performs data fusion to improve recognition accuracy, and alarms for the abnormal situation of off-bed for too long. The advantages of this scheme lie in the non-contact and fusion detection idea which can improve the usability of off-bed recognition, and take into account the drift, environmental interference or false alarm problems that may be caused by traditional pressure, infrared and wearable solutions. However, from the perspective of off-bed risk management, its core output is still biased towards the monitoring and alarm of off-bed events / off-bed time: on the one hand, off-bed activities often have continuous and periodic characteristics of lying position, sitting up, sitting on the bedside, standing and off-bed, and it may be difficult to depict the imminent state of off-bed in time only according to the off-bed duration; on the other hand, the off-bed risk of postoperative patients is not only determined by the action itself, but also highly coupled with the hemodynamic response under body position change and the state of the pipeline, and simple off-bed duration alarm cannot cover high-risk situations such as standing syncope and pipeline being pulled; in addition, in the real clinical environment of shielding, light changes or bedside personnel intervention, multiple sources of data may conflict, and if there is no quantitative processing of the quality / reliability of each mode, false alarms and missed alarms may still occur, affecting the clinical usability.

[0004] In addition, in order to achieve more comprehensive ward monitoring, there are schemes that coordinate vision with bed equipment, wearable equipment and vital sign collection. Chinese patent document CN116013548A discloses a smart ward monitoring method and device based on computer vision, which describes that the first detection equipment cooperates with the second detection equipment installed on the bed to obtain action posture information and vital sign information, and ensures time alignment through the same sampling frequency or communication method, thereby improving data availability and reliability. The advantage of this type of scheme is that it can obtain posture and vital signs from multiple devices, and emphasizes synchronization alignment and data reliability, which is beneficial to form continuous patient state evaluation. However, from the perspective of off-bed activity risk warning, which is a more focused clinical demand, there is still room for further improvement in the existing disclosure: first, smart ward monitoring usually faces general action posture recognition and monitoring control, and the special logic required for off-bed risk warning, such as grading trigger strategy, warning urgency, clinical prescription constraints, etc. It is not necessarily available; second, although it involves multi-device alignment, it is still difficult to avoid false positives if there is no dynamic evaluation and conflict handling mechanism for different modalities quality in the presence of common problems such as signal loss, drift, occlusion, noise mutation, etc. in the clinical field; third, the body position change-vital sign abnormal response and pipeline pulling risk of postoperative / sedation patients is often simplified as an ordinary vital sign threshold or a general event record under the general monitoring framework, making it difficult to form a targeted characterization of off-bed risk.

[0005] For example, for the fall prevention alarm system in the nursing scene, there are also many disclosures. Chinese patent document CN117423210B discloses a patient fall prevention intelligent sensing alarm system for nursing, which monitors the patient's position change through position detection related sensors and algorithms, and triggers an alarm when the patient leaves the bed surface or approaches the dangerous area near the bed, while combining remote monitoring and other means to reduce the risk of falling. This type of scheme embodies the practical idea of position / area trigger + alarm linkage, which is suitable for rapid deployment in nursing workflows. However, its limitations are: on the one hand, position / area triggers are often sensitive to clinical interference factors, such as turning over, bedside nursing operations, and accompanying personnel contact, which can cause short-term abnormalities in position characteristics. If there is no quantitative check on the consistency of multi-modal evidence, false positives are likely to occur; on the other hand, for postoperative or critically ill patients, the risk comes not only from whether they are close to the bed, but also from whether they have the ability to safely stand up / stand, whether vital signs allow body position changes, and whether pipelines are at risk of being pulled off / coming off. The traditional bed-side dangerous area trigger cannot cover these complex risks. In addition, the nursing scene generally emphasizes low false alarm rate and explainability, and if the alarm trigger logic cannot reflect patient individual differences, alarm fatigue may occur, reducing the trust and response efficiency of medical staff to the system.

[0006] In summary, although the prior art provides beneficial exploration in bed exit event detection, non-contact monitoring, multi-device alignment, and nursing alarm linkage, etc., in the application of bed exit activity risk warning for high-risk scenarios such as postoperative / sedation / ICU, etc., the following problems still exist: (1) Insufficient risk characterization: focusing on bed exit events or bed exit duration, it is difficult to capture the urgent state and phased changes of imminent bed exit, and the physiological risks such as body position intolerance are insufficient; (2) Insufficient processing of multi-modal conflicts and quality fluctuations: the complex clinical environment leads to frequent occurrence of occlusion, loss, drift and noise mutation, and when there is a lack of dynamic assessment of modal reliability and consistency conflict disposal, false positives / false negatives are easily caused; (3) Insufficient fusion of postoperative specific risk factors: multi-line indwelling, sedation / delirium, pain and medication status significantly affect bed exit safety, and existing public solutions often lack linkage mechanisms with these prior information and nursing prescriptions, making it difficult to form a deployable graded warning strategy; (4) Insufficient consideration of clinical usability and privacy: while meeting the requirements of high sensitivity warning, low false positives, explainability, easy deployment and data compliance, the general monitoring framework may not be able to achieve a balance in bed exit risk warning.

[0007] Therefore, there is an urgent need for a bed exit activity multi-modal monitoring risk warning method and system that can maintain stability in the case of multi-source data quality fluctuations and evidence conflicts, and can combine postoperative high-risk factors and nursing prescriptions for graded intervention, in order to improve the timeliness and accuracy of warning, reduce false positives and improve the adaptability of clinical workflow. SUMMARY

[0008] The technical purpose of the present application is to provide a bed exit activity multi-modal monitoring risk warning method and system for high-risk hospitalized patients such as postoperative / sedation / ICU, etc., by time alignment, modal confidence assessment and cross-modal consistency gating fusion of multi-source data such as bed surface load / pressure, non-contact posture displacement, vital signs and optional wearable motion and pipeline state, combined with bed exit urgency prediction, body position intolerance physiological response identification and pipeline pull risk assessment, output graded warning and individualized threshold adaptive update combined with nursing feedback, thereby reducing false positives / negatives, identifying high-risk bed exit behavior in advance and reducing adverse events such as falls, syncope and pipeline dislodgement in complex ward environments.

[0009] In the first aspect, in order to achieve the above-mentioned purpose, the present application adopts the following technical scheme:

[0010] A bed exit activity multi-modal monitoring risk warning method, the method comprising the following steps:

[0011] S1, multi-modal acquisition: taking the sampling time as the index, obtaining a multi-modal time series data set of the same monitored object.

[0012] S2, alignment and confidence evaluation: aligning and calculating the confidence of each modality is the set of modalities; when , is the confidence threshold, determining that the modality is a low-confidence modality at the moment and reducing its fusion weight;

[0013] S3, stage recognition: inputting the preset timing recognition model, outputting the bed exit activity stage probability vector , and defining the set of bed exit activity stages as ;

[0014] S4, gated fusion and postoperative risk scoring: weighting and fusing the modalities according to the modality confidence to obtain the fused stage probability , and calculating the cross-modality consistency index ; when , is the consistency threshold, calculating the bed exit urgency , the physiological response consistency index , and the pipeline risk index , and obtaining the risk score accordingly;

[0015] S5, graded warning and closed-loop update: comparing the risk score with the warning threshold to obtain the risk level and trigger the corresponding warning action.

[0016] As a preferred, in step S1, the at least includes at least two of: bed surface load / pressure distribution data , non-contact posture or displacement data , vital sign data , and optionally includes wearable motion data and / or pipeline state data ;

[0017] wherein is the timestamp, is the bed surface pressure matrix or load vector, is the posture key point, point cloud or displacement feature obtained by depth / infrared / millimeter wave, at least includes systolic blood pressure , heart rate ​​​​​​​oxygen saturation one of, is an acceleration / angular velocity or equivalent motion feature, is a catheter, urinary catheter or infusion line in-situ status, displacement or tension characterization.

[0018] As a preference, in step S2, the fusion weight and the fusion result satisfy:

[0019] ,

[0020] ;

[0021] wherein, is a normalized fusion weight of the modality , is a modality confidence.

[0022] As a preference, the cross-modality consistency index satisfies:

[0023] wherein, is a number of modalities, is a Kullback-Leibler divergence.

[0024] As a preference, the risk score satisfies:

[0025] ;

[0026] wherein, is a Sigmoid function, is a coefficient vector, is a weight coefficient, is a bias term, is a postoperative risk prior vector, at least containing one of a sedation / delirium assessment, a pain assessment, a medication status, a history of falls or a nursing risk classification;

[0027] is a bed exit urgency and is calculated from a predicted bed exit time :

[0028] ,

[0029] in seconds; is a line risk index; is a physiological response consistency indicator.

[0030] As a further preference, the at least includes two of the following variables: a sedation / agitation score , a pain score an indicator of vasoactive drug dose the number of postoperative indwelling lines or a fall risk scale score ; wherein is a quantifiable sedation / agitation scale, is a quantifiable pain scale, is a binary or dose classification variable of vasoactive drug use, is a count of lines in place, is a care scale score.

[0031] As a preference, the physiological response consistency indicator is constituted by a vital sign excursion amount associated with body position change, and at least comprises:

[0032] and ,

[0033] wherein, and is the mean value within a stable window;

[0034] and when and , the value of is increased to represent the risk of body position intolerance; wherein is a systolic blood pressure drop threshold, is a heart rate increase threshold.

[0035] As a preference, the line risk index is determined by at least one line in place status and a line pull characterization quantity, the line pull characterization quantity comprising at least one of: line end point displacement , line path bending change or line tension ; wherein is a displacement amount from a baseline position, is a curvature change amount, is a tension or equivalent tension indicator.

[0036] As a preference, in step S2, the vital sign baseline is calculated simultaneously based on the initial postoperative lying position stable window .

[0037] As a preference, in step S5, the early warning threshold is associated with a postoperative rehabilitation activity prescription level , and satisfies that when indicates that autonomous out-of-bed is prohibited, the and is reduced to early warning, when​ Instructions allow for improvement during bedside sitting / standing training. and To reduce false alarms; among which It is a discrete rank variable.

[0038] Preferably, in step S5, in the feedback confirmation window... Get the real results tags inside and based on right and / or , , At least one of them is updated individually; where .

[0039] Secondly, the present invention also provides a multimodal monitoring and risk warning system for out-of-bed activities, the system being used to implement the method, the system comprising:

[0040] Multimodal acquisition unit, used for acquisition and optional and form ;

[0041] Alignment and confidence evaluation unit, used to generate And calculate fusion weight and baseline vital signs ;

[0042] Stage identification unit, used for output With the probability of the fusion stage ;

[0043] Consistency gating and postoperative risk scoring unit, used for calculation And in Time calculation With risk score ;

[0044] The tiered early warning and closed-loop update unit is used to... Output risk level And trigger an early warning action, and in Internal receiving tag To update the threshold or model parameters.

[0045] Preferably, the multimodal acquisition unit includes at least one of a mattress pressure sensor array and a non-contact sensing sensor, wherein the non-contact sensing sensor is one of a depth camera, an infrared camera, or a millimeter-wave radar; and vital sign data. Data is collected via the bedside monitor interface.

[0046] Preferably, the consistency gating and postoperative risk scoring unit are configured to be based on the cross-modal consistency index. Greater than When entering conflict resolution mode, the conflict resolution mode includes at least one of the following: extending the decision holding time window. Request additional modal data collection, or adjust the risk level. Downgrade; among them The unit is seconds.

[0047] Preferably, the system is deployed as an edge computing architecture, outputting only at the bedside. And event summaries that do not contain recognizable images, and localize the original image frames without uploading them.

[0048] Thirdly, the present invention also provides a computer-readable storage medium having a computer program or instructions stored thereon, which, when executed by a processor, implement the steps of the method.

[0049] Fourthly, the present invention also provides a computer program product, including a computer program or instructions that, when executed by a processor, implement the steps of the method.

[0050] This invention employs a fusion mechanism of multimodal time alignment, modal confidence assessment, and cross-modal consistency gating to ensure stable output from bed exit monitoring even under complex conditions such as ward obstruction, changes in lighting, nursing interventions, and single sensor drift / missing. This significantly reduces false alarms and missed alarms caused by single-modal mistriggers in traditional bed exit alarms. Furthermore, this invention identifies bed exit activities in stages: lying down—sitting up—sitting at the bedside—standing—leaving the bed, and introduces bed exit urgency (predicting the time to leave the bed) to provide early warning before bed exit, shifting nursing intervention from post-event alarms to pre-event prevention. Moreover, this invention incorporates postoperative specific risks into a unified risk scoring framework, using postural changes to trigger... Consistent vital sign responses (such as decreased systolic blood pressure and compensatory changes in heart rate) identify the risk of postural intolerance / syncope. Combined with the in-place status of tubing and traction characteristics, the system assesses the risk of catheter, drainage tube, and infusion tubing dislodgement or traction bleeding. This not only prompts patients to get out of bed but also indicates whether getting out of bed is safe and where the risk originates. In addition, the system performs individualized adaptive updates of thresholds / weights based on nursing prescription levels and feedback tags, continuously suppressing alarm fatigue and improving adaptability and interpretability for different patients and different postoperative stages. Ultimately, this achieves a comprehensive technical effect of reducing falls, syncope, and tubing-related adverse events, improving nurses' response efficiency, reducing care costs, and enhancing ward safety management. Attached Figure Description

[0051] Figure 1 This is a schematic diagram of the structure of a multimodal monitoring and risk warning system for off-bed activity according to the present invention.

[0052] Figure 2 This is a flowchart illustrating a method for monitoring and warning risks of out-of-bed activity according to the present invention.

[0053] Figure 3 This is a schematic diagram of the multimodal data time alignment and modal confidence assessment process of the present invention.

[0054] Figure 4 This is a schematic diagram of the structure of the off-bed activity phase identification model of the present invention.

[0055] Figure 5 This is a schematic diagram of the cross-modal consistency gating and postoperative risk score calculation process of the present invention.

[0056] Figure 6 This is a schematic diagram of the hierarchical early warning and closed-loop update mechanism of the present invention.

[0057] Figure 7 This is a schematic diagram illustrating the calculation of the consistency index of postural intolerance physiological response and the pipeline risk index in the postoperative / ICU setting according to the present invention.

[0058] Figure 8 This is a schematic diagram comparing the detection sensitivity of different methods for real bed exit events in Example 1 (postoperative general ward scenario).

[0059] Figure 9 This is a schematic diagram comparing the specificity of different methods for non-bed leave events in Example 1 (postoperative general ward scenario).

[0060] Figure 10 This is a schematic diagram comparing the false alarm frequency of different methods in Example 1 (postoperative general ward scenario).

[0061] Figure 11 This is a schematic diagram comparing the advance warning time of different methods in Example 1 (postoperative general ward scenario).

[0062] Figure 12 This is a schematic diagram comparing the detection performance of pipeline traction near-loss events in Example 2 (ICU multi-pipeline in-situ semi-physical scenario).

[0063] Figure 13 This is a schematic diagram comparing the performance of orthostatic intolerance (high physiological risk) event identification in Example 3 (PACU / physiological risk scenario during sedation recovery period).

[0064] Figure 14 This is a schematic diagram of the risk score curve for a typical out-of-bed event.

[0065] Figure 15 This is a schematic diagram of the multimodal confidence curves for a typical out-of-bed event data segment.

[0066] Figure 16 The time-series heatmap of bed surface pressure is collected for a typical bed exit event.

[0067] Figure 17 Modal confidence time-series heatmaps of typical bed exit events.

[0068] The module includes: 110—Multimodal acquisition unit; 111—Bed surface pressure acquisition module; 112—Non-contact posture / displacement acquisition module; 113—Vital signs acquisition module; 114—Wearable motion acquisition module; 115—Pipeline status acquisition module; 120—Alignment and confidence assessment unit; 121—Synchronization and resampling module; 122—Missing information handling module; 123—Quality index calculation module; 124—Confidence calculation and gating module; 130—Stage identification unit; 131—Feature construction module; 132—Modal encoder; 133—Temporal backbone network; 134—Output head; 140—Conformity gating and postoperative risk scoring unit; 141—Weighted fusion module; 142—Conformity gating module; 143—Escape urgency prediction module; 144— and and Calculation module; 145—Risk scoring output module; 150—Graded early warning and linkage unit; 151—Grade determination module; 152—Linkage output module; 153—Prescription linkage module; 160—Data storage and feedback closed-loop unit; 161—Baseline establishment module; 170—Nurse station / mobile terminal. Detailed Implementation

[0069] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the protection scope of the present invention.

[0070] I. Terminology Explanation

[0071] Out-of-bed activity phase: refers to the continuous set of states of the patient from lying down to getting out of bed, including at least "lying down", "sitting up", "sitting by the bedside", "standing", and "getting out of bed".

[0072] Modality: refers to a data source or sensing type, such as bed pressure, non-contact posture / displacement, vital signs, wearable IMU, pipeline status, etc.

[0073] Modal confidence : indicates mode At any moment Quantitative indicators of data quality and reliability are used to integrate weighted and gating.

[0074] Cross-modal consistency index A quantitative indicator that measures whether the distribution of judgments on the bed exit stage at the same time is consistent across different modalities, used for conflict resolution and false alarm suppression.

[0075] Urgency to leave bed From prediction to bed exit time The calculated urgency index is used for early warning.

[0076] Postoperative risk prior vector : A set of static or slowly changing information related to individual postoperative / sedation / ICU risks, such as sedation / delirium assessment, pain assessment, medication status, history of falls, number of indwelling catheters, nursing prescription level, etc.

[0077] Pipeline Risk Index Indicators reflecting the risk of traction, twisting, dislodgement, or bleeding of urinary catheters, drainage tubes, infusion tubes, etc.

[0078] Physiological response consistency index An indicator that reflects abnormal vital sign responses (such as orthostatic hypotension / intolerance) triggered by changes in body position (sitting up / standing).

[0079] Warning threshold Risk scoring thresholds are used to output prompts / warnings / alarms.

[0080] Feedback confirmation window : Labels used to collect actual results after an early warning And update the time window for individualized parameters.

[0081] II. System Structure of the Invention

[0082] like Figure 1 As shown, the multimodal monitoring and risk warning system for out-of-bed activities of the present invention includes at least the following units (which can be deployed in postoperative wards, ICUs, PACUs, neurosurgery departments, etc.):

[0083] 1. Multimodal acquisition unit

[0084] Used to collect and generate multimodal time series datasets. It should contain at least two of the following modules, preferably three or more:

[0085] Bed surface bearing / pressure acquisition module: such as a mattress pressure sensor array (e.g., 32×16 or 64×32 dot matrix), outputting a bed surface pressure matrix. Or the carrying vector.

[0086] Non-contact attitude / displacement acquisition module: such as one or a combination of depth camera / infrared camera / millimeter-wave radar, outputting attitude key points, point cloud or displacement features. In privacy-focused implementations, the contactless module only outputs key skeleton points or low-dimensional displacement features, without outputting the original, recognizable image.

[0087] Vital signs acquisition module: acquires data via bedside monitor interface. , , Composed of respiratory rate, etc. .

[0088] Wearable motion acquisition module (optional): such as wristband / ankleband IMU, outputting acceleration and angular velocity characteristics. .

[0089] Pipeline status acquisition module (optional): used to obtain The following methods can be used: (1) miniature strain / tension sensors at the pipeline clamping point; (2) low-power IMUs at the pipeline endpoints; (3) vision / depth tracking of key points along the pipeline path; and (4) RFID / reed switch detection for presence / dislodgement. In practical engineering, one or a combination of these methods can be selected to form a stable pipeline tension characterization quantity.

[0090] 2. Alignment and Confidence Assessment Unit

[0091] This is used to unify timestamps across modalities, resample, fill in missing data, and output modal confidence scores. And its gating results. This includes: a time synchronization module (hardware synchronization or network clock / NTP synchronization); a resampling module (unifying different sampling rates to the target frequency). ); Quality assessment module (evaluation of missing, occlusion, noise, and drift); Confidence calculation and threshold gating module (reducing weight or removing low-confidence modes).

[0092] 3. Stage Identification Unit (Model Reasoning Unit)

[0093] Used according to Output the probability distribution of the bed exit stage. This unit may contain a multi-branch encoder, a temporal modeling network, and a classification head (see the model structure and training disclosure below for details).

[0094] 4. Consistency Gating and Postoperative Risk Scoring Unit

[0095] Used to calculate the crossmodal consistency index Risk scores are calculated when consistency conditions are met. Integrating the urgency of leaving the bed Consistency of physiological response Pipeline risks Postoperative pre-existing conditions .

[0096] 5. Tiered early warning and linkage unit

[0097] It is used to output three levels of results: prompts, warnings, and alarms, triggering actions such as bedside audio and visual prompts, voice prompts, push notifications from nurses' stations / mobile devices, and nursing calls, and to implement escalation / degradation rules and suppression mechanisms (such as window keeping, de-shaking, and conflict handling).

[0098] 6. Data storage and feedback closed-loop unit

[0099] Used to record event summaries, risk curves, and early warning logs, and to display them in the feedback confirmation window. Internal collection tags (Actual bed exit / false alarm / cancellation / nursing confirmation) Update thresholds and individualized parameters to reduce false alarms and alarm fatigue.

[0100] III. Specific Technical Route for Implementing the Method of the Invention

[0101] like Figure 2 As shown, the method of the present invention is executed according to S1–S5, of which S2, S4, and S5 contribute more to reducing false alarms, achieving early warning and adapting to complex postoperative risks, and are therefore disclosed in detail.

[0102] 3.1 Step S1, Multimodal sampling

[0103] In one implementation, the system sets a target uniform sampling frequency. (Can be adjusted according to computing power and scenario). Examples of raw sampling rates for each modality: Pressure Array Depth camera Millimeter-wave radar Vital signs IMU Pipeline tension The system uses timestamps. Unified data management leads to:

[0104] : Bed surface pressure matrix or bearing vector;

[0105] : Coordinates of key posture points (such as head / shoulder / hip / knee / ankle), centroid trajectory of point cloud, chest wall micro-movement, bedside approach distance, etc.;

[0106] : wait;

[0107] Statistical characteristics of acceleration / angular velocity;

[0108] : Pipeline end displacement, tension, or changes in path curvature, etc.

[0109] To ensure the feasibility of the project, We recommend using low-dimensional feature output methods: for example, only outputting the displacement / velocity of the skeleton key points in the bed coordinate system, without retaining the original video frames; or millimeter-wave radar only outputting the centroid and velocity vector after clustering the target point cloud, in order to balance privacy compliance and bandwidth.

[0110] 4.2 Step S2, Alignment and Confidence Assessment

[0111] like Figure 3 As shown, the goal of step S2 is to transform multi-source data into fusionable and reliable data. And output the confidence level for each mode. This ensures that subsequent fusion is not misled by noise, occlusion, or drift.

[0112] S2-1 Time Alignment and Resampling

[0113] For each mode The original sequence ,implement:

[0114] 1) Clock unification: Hardware-triggered synchronization is used, or NTP alignment is used to reach a unified time base;

[0115] 2) Resampling: Resample all modes to Low-frequency signals such as vital signs are analyzed using zero-order hold or linear interpolation; high-frequency IMUs can perform sliding statistical convergence. ;

[0116] 3) Missing Entry: Short Missing (e.g.) Neighbor interpolation can be used; long missing segments are marked as missing segments and used for confidence penalty.

[0117] The resampled aligned data is denoted as .

[0118] S2-2 Quality Index Calculation

[0119] In the sliding quality assessment window Within (unit: seconds), calculate at least the following three types of metrics (expandable):

[0120] 1) Missing rate :window Percentage of samples with missing information;

[0121] 2) Signal-to-noise ratio It can be estimated by the ratio of target frequency band energy to noise energy; for vision / radar types, it can be replaced by target detection confidence / point cloud clustering stability.

[0122] 3) Drift amount It reflects slow-changing offsets, such as zero-point drift of the pressure array, changes in radar static background, and changes in camera extrinsic parameters. It can be measured by the statistical difference between the baseline segment and the current segment.

[0123] S2-3 Modal Confidence Calculation and Gating

[0124] In a preferred embodiment, interpretable multiplicative combinations are used to obtain... :

[0125] ;

[0126] in:

[0127] Modality exist Confidence level at any given moment;

[0128] Missing rate; : Missing penalty coefficient;

[0129] Signal-to-noise ratio or equivalent quality score; Reference quality baseline; Scale parameters;

[0130] : Drift amount; Drift penalty coefficient;

[0131] : Sigmoid function.

[0132] Confidence threshold Used for gating: when This mode is marked as a low-confidence mode. Low-confidence modes are not directly hard-rejected (to prevent extreme cases where no modes are available), but instead undergo weighted fusion: their weights are automatically reduced in step S4. If the system strategy requires a more conservative approach, then the duration of continuous low reliability will exceed [a certain value]. The mode is briefly removed at (seconds) and a sensor anomaly alert is triggered.

[0133] S2-4 Postoperative vital signs baseline establishment

[0134] Postoperative / ICU patients exhibit significant variations in vital signs, making fixed thresholds prone to false alarms. Therefore, the system uses an initial positional stability window. Establish baseline An optimal implementation is as follows: when the probability of the reclining position output by the stage recognition unit is consistently higher than... And duration At that time, the average vital signs were taken: , , etc. constitute For subsequent calculations With individualized threshold updates.

[0135] 4.3 Step S3, Stage Identification

[0136] like Figure 4 As shown, step S3 outputs the stage probability vector for each mode. and define the stage set as To ensure feasibility, two equivalent implementation paths are given below; each project can choose one:

[0137] Option A: Independent recognition of each modality + subsequent fusion

[0138] For each mode Establish an encoder, a temporal network, and a classification head respectively, and output... ;

[0139] Advantages: Each modality can be trained independently and is easy to maintain; Disadvantages: There are many models.

[0140] Option B: Shared timing backbone + modal branch (Recommended)

[0141] A dedicated encoder is established for each modality, and the multimodal features are aligned to a unified dimension and then fed into the shared temporal backbone.

[0142] Shared trunk output hidden state Then output from the modal splitter. It also supports mask training for missing modalities.

[0143] The following is a structured disclosure using Option B as an example:

[0144] (1) Input feature construction

[0145] At a unified frequency Next, construct the modal feature vector for each time step:

[0146] Pressure characteristics It can be obtained by dimensionality reduction from the pressure matrix, including total load, left and right / front and back pressure centers, bedside pressure gradient, upper body / lower body load ratio, etc.

[0147] Non-contact characteristics Key skeletal features include height of key points on the skeleton, trunk angle, distance from the bedside, vertical velocity, and angular velocity of the upper body when rising.

[0148] vital signs : , , and its short-term rate of change;

[0149] IMU characteristics Statistical analysis of acceleration modulus, angular velocity modulus, and attitude change rate;

[0150] Pipeline characteristics : Pipeline end displacement, tension, curvature changes, etc.

[0151] The above splicing forms and in the window A sequence is formed within (seconds) As input to the model.

[0152] (2) Encoder Design

[0153] Pressure encoder: two layers Convolutional or lightweight CNN + fully connected, output dimension ;

[0154] Non-contact encoder: MLP (two-layer fully connected + ReLU) or Convolution, output ;

[0155] Vital signs encoder: MLP output ;

[0156] IMU encoder: Convolution + Pooling Output ;

[0157] Pipeline encoder: MLP output .

[0158] After encoding, a unified feature vector is obtained. ,in (Missing modes are set to zero by the mask and added to the mask vector).

[0159] (3) Temporal backbone network

[0160] Options include BiGRU, TCN, or TransformerEncoder. For ease of deployment, a two-layer GRU (which can be replaced with LSTM) is preferred:

[0161] Number of hidden GRU units in Layer 1 ;

[0162] Number of hidden GRU units in layer 2 ;

[0163] Output hidden state .

[0164] (4) Classification header output

[0165] For each mode Set a classification head (fully connected + Softmax) and output the probability of 5 classes. Each component corresponds to In the middle stage, and the sum is 1.

[0166] 4.4 Step S4, Gated Fusion and Postoperative Risk Scoring

[0167] like Figure 5 As shown, step S4 is the key step in the present invention to form an early warning + anti-false alarm + postoperative composite risk effect. The core includes: fusion weight calculation, cross-modal consistency gating, prediction of bed urgency, identification of physiological response consistency, pipeline risk assessment, and final risk score output.

[0168] S4-1 Confidence-Weighted Fusion

[0169] First, the confidence levels of each modality are obtained from S2. Calculate the fusion weights And obtain the probability of the fusion stage. :

[0170] ;

[0171] in: The set of modalities participating in the fusion; For modality Probability distribution for the 5 stages; To normalize the weights, ensure ; The probability of the merged stage can be directly used for stage determination (taking the stage with the highest probability) or for subsequent scoring.

[0172] S4-2 Crossmodal Consistency Index and Gating

[0173] To suppress false alarms caused by modal conflicts, a consistency index is calculated. :

[0174] ;

[0175] in:

[0176] The number of modes;

[0177] The divergence is Kullback-Leibler.

[0178] Gating strategy: When When this occurs, enter conflict resolution mode (without immediately escalating the alarm) and execute at least one of the following:

[0179] 1) Keep window : Extend the observation period and require continuous fulfillment of conditions before upgrading;

[0180] 2) Additional evidence: Request the activation of backup modes (such as turning on the radar, increasing the sampling rate) or prompt nursing staff for confirmation;

[0181] 3) Degraded output: Limit the output to prompts or warnings to avoid accidental call triggering.

[0182] S4-3 Bed Exit Urgency Prediction

[0183] Bed exit urgency is used to provide early warning before bed exit. It predicts the time to bed exit. The regression head output (sharing the main branch with step S3) can be used: the regression input is a sequence. ;Regression Output (Unit: seconds)

[0184] Bed urgency is defined as:

[0185] ,

[0186] in, The smaller, The larger the number, the more urgent it is.

[0187] S4-4 Physiological Response Consistency Index (Identification of postural intolerance)

[0188] like Figure 7 As shown, postoperative patients commonly experience a drop in blood pressure, cardiac compensation, and dizziness / syncope when sitting up / standing. This invention uses a baseline... The offset is calculated for reference:

[0189] ;

[0190] ;

[0191] And define the threshold , (Units are mmHg and bpm respectively). One optimal logic is: when... and When a tendency for postural intolerance is identified, the risk is increased. In engineering Desirable Continuous values, such as combinations of offsets mapped by the Sigmoid function, or implemented using piecewise functions (facilitating real-time calculations for edge devices).

[0192] S4-5 Pipeline Risk Index (Identification of risks of traction / detachment)

[0193] like Figure 7 As shown, for the indwelling pipeline, the system uses... The tensile characterization quantity is obtained as follows:

[0194] Pipeline end displacement (Relative baseline displacement);

[0195] Pipeline path bends and changes ;

[0196] Pipeline tension Or equivalent tension index.

[0197] When the patient enters the sitting-up / bedside sitting / standing stage, if or Exceeding the threshold ,but Increase; and can be stacked with the number of in-situ An amplification of overall risk to reflect the higher risk in patients with multiple access routes.

[0198] S4-6 Risk Score With parameter definition

[0199] Based on the above information, calculate the risk score:

[0200] ;

[0201] Parameter definition:

[0202] Risk score;

[0203] :Sigmoid;

[0204] : Probability vector of the fusion stage; : Phase contribution coefficient vector (e.g., assigning higher weights to standing / leaving bed);

[0205] : Urgency to leave the bed; Urgency weight;

[0206] Pipeline risk index; Pipeline weights;

[0207] Physiological response consistency index; Physiological weight;

[0208] Postoperative prior vector; Prior weight vector;

[0209] : Bias term.

[0210] Step 4.5, S5: Graded Early Warning and Closed-Loop Improvement.

[0211] like Figure 6 As shown, step S5 addresses the implementation of nursing prescription constraints related to large individual differences in alarm fatigue.

[0212] S5-1 Tiered Early Warning Rules

[0213] Set threshold Output:

[0214] Hints (optional, log only or provide a brief hint);

[0215] Warning (bedside audio and visual alerts + push notifications from nurses' devices);

[0216] Alarm (strong sound and light + nursing call / priority push).

[0217] To suppress transient noise, a duration window is set. : Must meet continued It was then upgraded to an alarm; and a debouncing window was set. Avoid repeated bombing.

[0218] S5-2 Combining Nursing Prescription Levels Adaptive threshold

[0219] Postoperative patients often have activity restrictions (e.g., prohibited from getting out of bed independently, allowed to sit at the bedside, allowed to practice standing). This invention adjusts the threshold accordingly. Adjustment: When To prevent people from getting out of bed on their own and reduce [the risk of infection] (More sensitive, intervene earlier); when To allow for training, the threshold should be increased moderately to reduce false alarms. It can be configured by doctors / nurses in the system, or automatically issued according to the disease progression stage.

[0220] S5-3 Feedback Confirmation and Tag Collection

[0221] In the feedback confirmation window Internally collected real results labels :

[0222] Real high-risk events (real bed-breaking incidents, fall risks, fainting warning signs, confirmation of tubing traction, etc.);

[0223] False alarms or negligible events (nursing operation interference, accidental triggering during turning over, accidental triggering by caregivers, etc.).

[0224] Tag sources may include: nurse confirmation / cancellation, nursing call records, manual review of bedside video (only in authorized scenarios), and bed exit event counters, etc.

[0225] S5-4 Personalized Update Strategy

[0226] To ensure feasibility, two types of update strategies are provided (choose one or a combination):

[0227] Strategy 1: Threshold Adaptation

[0228] Maintaining false alarm rate estimates for each patient dimension Risk of underreporting When it appears consecutively In case of false alarm, increase or improve When it appears And without prior warning, reduce or increase The upper limit of the weight of the corresponding item.

[0229] Strategy 2: Small online updates

[0230] With the main model parameters fixed, only the coefficients of the last layer are considered. The learning rate is used for either small-step gradient updates or recursive least squares updates. Pick And set safety constraints (coefficient range, update frequency upper limit) to ensure that drift is not caused by a small amount of noise labels.

[0231] V. Model Structure, Training Methods, Parameter Selection, and Dataset Usage

[0232] This section contains key disclosures on implementation, ensuring that those skilled in the art can reproduce the work without inventive effort.

[0233] 5.1 Dataset Construction and Labeling

[0234] (1) Data collection objects and scenarios

[0235] Subjects: Postoperative general ward, ICU, PACU patients and control group;

[0236] Scene coverage: turning over in a supine position, sitting up, sitting on the edge of the bed, preparing to stand, getting out of bed, nursing intervention, caregiver intervention, equipment movement, shielding, etc.

[0237] Acquisition modality and synchronization: Consistent with step S1, ensure a unified timestamp.

[0238] (2) Labeling system

[0239] It must contain at least two types of tags:

[0240] Stage Labels : Obtained by manual or semi-automatic annotation (e.g., by caregivers clicking on the stage switching time on the annotation tool under privacy compliance conditions);

[0241] Event Tags This indicates whether a real high-risk bed-related event has occurred within the warning window.

[0242] For postural intolerance, additional labels can be added: for example, a drop in SBP exceeding the threshold after standing and accompanied by symptoms is positive; for tubing risks, traction / dislodgement / bleeding risks can be added as positive.

[0243] (3) Data partitioning

[0244] Divide the training / validation / test sets according to the patient dimension (to avoid the same patient being leaked into different sets), with a recommended ratio of 70% / 15% / 15%; and perform stratified sampling for real high-risk events to alleviate class imbalance.

[0245] 5.2 Model Training Objective and Loss Function

[0246] Multi-task training (recommended):

[0247] Stage recognition loss: cross-entropy ;

[0248] Time-out regression loss: Huber or MSE, denoted as ;

[0249] Confidence / Quality Auxiliary Loss (Optional): For example, supervision with low confidence for missing segments is denoted as... ;

[0250] Total loss: in For weight hyperparameters.

[0251] 5.3 Key Hyperparameter Recommendations

[0252] Unified sampling frequency ;

[0253] sequence window Used for stage identification;

[0254] Regression Window For ;

[0255] Quality assessment window Baseline window ;

[0256] Training optimizer: Adam; learning rate Start with cosine annealing or step descent; batch size is 16 / 32 based on computing power.

[0257] Category imbalance handling: weight the bed-out / standing stage, or use focal loss as an alternative. ;

[0258] Example of initial threshold value: , (can be done) Adjustment);

[0259] Conflict threshold : Selected by minimizing false positives on the validation set;

[0260] Postural threshold: , (Can be individualized).

[0261] 5.4 Inference Deployment and Resource Constraints

[0262] The stage identification and regression model is preferably deployed on bedside edge devices (such as ARM or small GPU boxes), and quantization / pruning is used to reduce latency; the output only contains Instead of uploading the original image frames, the event summary should be displayed. For vital signs and pipeline modules, asynchronous thread updates can be used to avoid blocking the main pipeline.

[0263] VI. Typical Application Examples

[0264] Three typical application examples (laboratory / semi-physical verification) are given below, all based on the system structure shown in the attached diagram of the manual. Figure 1 The system includes a multimodal acquisition unit, an alignment and confidence assessment unit, a stage identification unit, a consistency gating and postoperative risk scoring unit, a graded early warning and closed-loop update unit, etc., as well as the methodology (…). Figure 2 (S1–S5). All experiments were conducted using a nursing simulation laboratory prototype and standardized script scenarios. Comparison schemes and quantitative data are provided to demonstrate the technical effectiveness of this invention in early warning, false alarm suppression, identification of postural intolerance, identification of pipeline traction risk, and closed-loop adaptive testing.

[0265] 1. General test platform and evaluation indicators

[0266] A. Test Platform

[0267] Bed surface pressure / load-bearing module: 64×32 dot matrix pressure pad, output pressure matrix (10Hz).

[0268] Non-contact attitude / displacement module: One of a depth / infrared camera or millimeter-wave radar (15Hz, output skeleton key points / displacement features) (No original video frames are stored).

[0269] Vital signs module: Monitor interface or vital signs simulator output , including (1Hz) (1Hz) (1Hz).

[0270] Wearable IMU (optional): Wristband IMU output (50Hz, converged to 10Hz).

[0271] Pipeline status (optional): Tension output from miniature tension / strain sensor at pipeline clamp. (10Hz) or endpoint displacement .

[0272] Unified sampling frequency ,according to Figure 2 Alignment is completed in step S2. .

[0273] B. Comparison Method (Baseline Approach)

[0274] Comparison Option B0 (Single-modal mattress leave-bed alarm): using only An alarm will sound when the total load decreases or the pressure center approaches the edge of the bed for ≥2 seconds (a common bed-off alarm logic in engineering).

[0275] Comparison with Solution B1 (multimodal fusion but without confidence / consistency gating): using (optional) The probability of equal-weighted fusion stage is not calculated. and This does not include conflict resolution.

[0276] Comparison with option B2 (which includes stage identification but does not introduce postoperative physiological responses) Pipeline risks ): Multimodal stage identification and bed urgency are used, but the risk score does not include Items, and prescription linkage / closed-loop updates are not performed.

[0277] The proposed solution is executed according to steps S1–S5 of this application, including: Confidence weighted Consistency gating, bed urgency Consistency of physiological response Pipeline risks And closed-loop individualized threshold updates.

[0278] C. Evaluation Indicators (Event Level)

[0279] Event-level statistics are performed on bed exit events / high-risk events, defined as follows:

[0280] 1) The number of high-risk events that were correctly predicted; The number of high-risk events was underreported; : Number of false alarms; : Number of times the alarm was not triggered (statistics based on non-event windows).

[0281] 2) Recall / Sensitivity: in Sensitivity.

[0282] 3) Specificity: in Specificity.

[0283] 4) False alarm frequency (per patient day): in False alarm count / patient day To accumulate the number of days of patient monitoring.

[0284] 5) Lead time (for each actual bed exit event): in This is the actual time of getting out of bed (script / annotation). The moment when the warning is first triggered; This indicates an early warning.

[0285] 2. Application Example 1: Validation of Early Warning and False Alarm Suppression in Postoperative General Wards

[0286] 2.1 Experimental Objective

[0287] The study aimed to verify whether the invention could significantly reduce false alarms while maintaining high sensitivity and provide early warning before patients leave the bed, under conditions of obstruction and interference caused by high-frequency nursing operations (turning over, changing dressings, measuring vital signs, and touching by caregivers) in the postoperative ward.

[0288] 2.2 Experimental Design

[0289] Venue: Nursing simulation laboratory (simulated single-patient ward).

[0290] Subjects: 20 volunteers (weighing 45–92 kg) performed actions according to a standardized script; 2 nursing staff simulated routine interventions.

[0291] Data collection time: 40 minutes per person, totaling approximately 13.3 hours; calculated as a monitoring day for each patient. (Based on 24 hours / day).

[0292] Scene script (per person):

[0293] Normal rolling / movement: 6 times;

[0294] Sitting up but not getting out of bed: 4 times;

[0295] Sit on the edge of the bed: 3 times;

[0296] Standing ready but being persuaded to return: 2 times;

[0297] Actual getting out of bed: 3 times (including both fast and slow getting up);

[0298] Nursing interventions (covering / contacting the bed): 8 times.

[0299] Event definition:

[0300] Real-life bed-leaning event: feet off the edge of the bed and standing / taking a step (marked) );

[0301] High-risk intention to get out of bed: A continuous phase of sitting up → sitting at the bedside → standing up is observed within 10–30 seconds before getting out of bed, and the prescription is set to prohibit independent getting out of bed (simulating postoperative restrictions).

[0302] 2.3 Results Data

[0303] Table 1 Overall performance comparison of Example 1 (13.3h, actual bed exit events) )

[0304]

[0305] Key points for explanation and proof of effect: such as Figures 8-11 As shown, compared with B0, the present invention improves sensitivity (0.83→0.96) while reducing false alarm frequency. A significant decrease (30.9 → 5.5 times / patient day) indicates a confidence-weighted... With consistency gating It has a significant inhibitory effect on nursing obstruction and bedside contact interference. Compared with B1 / B2, the present invention... The further increase (8.9 / 10.1→13.2s) indicates the introduction of Once the time of bed alighting is predicted, the alarm can be moved from the time of bed alighting to the preparation stage, thus achieving a true pre-intervention window. The value was increased to 0.93, which reflects the effect of false alarm suppression and helps reduce alarm fatigue.

[0306] Table 2 Example 1: Analysis of False Alarm Sources in Typical Interference Scenarios (Statistical False Alarm FP=3)

[0307]

[0308] As can be seen, the false alarms of this invention are mostly limited to prompts / warnings and can be quickly suppressed by gating strategies, without causing frequent alarms.

[0309] 3. Application Example 2: Validation of Risk Identification for In-situ Traction / Dislodgement of Multiple Tubes in the ICU

[0310] 3.1 Experimental Objective

[0311] Verification of the invention under conditions where multiple lines are in place (drainage tube, urinary catheter, infusion tube) Can we identify traction risks in advance and reduce potential hazards in pipelines that are already under stress even if bed separation has not yet occurred? Simultaneously, can we verify the effectiveness of bed separation during the bed separation phase? The coupling effect.

[0312] 4.2 Experimental Design

[0313] Venue: Semi-physical ICU bedside experimental table (simulated hospital bed + pipeline fixing points + controllable tension source), which meets the requirements. Figure 1 structure.

[0314] Pipeline configuration: Category 3 pipelines, number of pipelines Three ranges; tension sensor range 0–20N, sampling 10Hz.

[0315] Scenario script: 120 trial rounds in total, each round lasting 30–90 seconds, including:

[0316] Sit up / sit on the edge of the bed (without leaving the bed);

[0317] Stand ready;

[0318] The patient's feet were not off the bed, but the tubing was pulled, nearly causing a near-loss event (human-set traction threshold).

[0319] Actual getting out of bed action (partial rounds).

[0320] Definition of near loss event: when Duration ≥1.0s or A pipeline event lasting ≥1.0s is classified as a high-risk event (labeled). ).

[0321] 2.3 Results Data

[0322] Table 3 Example 2: Pipeline Risk Event Detection Performance ( )

[0323]

[0324] in , This is the moment when the pipeline risk threshold is first met.

[0325] Key points for explanation and proof of effect: such asFigure 12 As shown, the detection rate of B0 / B2 is low (0.41 / 0.58) in scenarios where the pipeline has been pulled but not yet removed from the bed, because its triggering mainly depends on the characteristics of the danger zone away from the bed / bedside; this invention will Explicitly incorporate risk scoring and phase Coupling allows the system to provide early warnings during the standing / sitting-at-the-bed stage, when pipelines are under stress. The accuracy rate was 0.91, with an average lead time of 7.8 seconds, providing nursing staff with a suitable time window for securing / releasing tubing; the number of false alarms related to tubing risk remained low (2), indicating that... Gating and Weight reduction can suppress false triggering caused by occasional spikes in the tension sensor.

[0326] 4. Application Example 3: Early Warning Verification of Postural Intolerance (Consistency of Physiological Response) During PACU / Sedation Recovery Period

[0327] 4.1 Experimental Objective

[0328] This invention verifies that during the recovery period from sedation, patients may experience orthostatic hypotension / intolerance when changing from a supine to a sitting / standing position. (based on Can physiological risks be identified before getting out of bed, thus avoiding high-risk situations where a person has stood up but experiences dizziness and falls?

[0329] 4.2 Experimental Design

[0330] Venue: Post-anesthesia care unit (PACU) working condition simulation experimental area, using a vital signs simulator to output controllable... The curves were overlaid with real motion data; the motions were performed by 12 volunteers (sitting up / sitting on the edge of the bed / standing preparation / getting off the bed).

[0331] Event definition: When the postural intolerance criterion occurs Decrease within 30 seconds ≥ and Rise ≥ This is recorded as a high-risk physiological event and marked accordingly. This experiment is designed... , .

[0332] Total rounds: 96 rounds, including high-risk physiological events (The rest are normal changes in body position).

[0333] 4.3 Results Data

[0334] Table 4 Example 3: Identification effect of high-risk physiological events

[0335]

[0336] The accuracy rate of the linked alarm refers to the proportion of cases where an alarm is triggered during the standing preparation / getting out of bed stage and there is indeed a high physiological risk.

[0337] Key points for explanation and proof of effect: such as Figure 13 As shown, traditional bed-leaning / posture detection (B1 / B2) is not sensitive to physiological risks and is prone to two types of biases: actions that appear to indicate an impending bed-leaning motion but are not physiologically dangerous, or actions that indicate standing up but are physiologically abnormal. This invention utilizes a baseline window... Establish individual vital sign baselines, and use structure Included It significantly improved the detection rate of physiological risks (0.50→0.88) and controlled the false alarms to 2, proving that it can achieve risk warnings that are closer to clinical mechanisms during the sedation recovery period.

[0338] Figure 14 This is a schematic diagram of the risk score curve for a typical bed exit event segment, illustrating the risk score. The rising process of changes with the stage of bed exit activity is analyzed, and an early warning threshold is given. With alarm threshold The triggering time is used to demonstrate that the present invention can achieve graded early warning before bed exit occurs. Figure 15 This is a schematic diagram of multimodal confidence curves for typical bed exit event acquisition segments, showing the confidence levels of bed surface pressure mode, non-contact posture / displacement mode, and vital signs mode. Changes over time, and the occlusion area is marked. The threshold is used to demonstrate that the present invention can reduce the weight of low-confidence modes and suppress false alarms under occlusion / interference conditions. Figure 16 The image shows a time-series heatmap of bed surface pressure collected from a typical bed-leaving event. The horizontal axis represents time, the vertical axis represents the index along the length of the bed, and the color represents the pressure intensity. This is used to demonstrate that the present invention can capture the spatiotemporal evolution characteristics of bed surface load distribution during the process of lying down, sitting up, sitting at the bedside, preparing to stand, and leaving the bed. Figure 17 The image shows a time-series heatmap of modal confidence scores for typical bed-off event acquisition segments. The horizontal axis represents time, the vertical axis represents different modal rows, and the color represents the confidence score. This is used to demonstrate that the present invention can identify confidence drops and maintain overall judgment stability through gating fusion when non-contact modalities are occluded.

[0339] 5. Conclusion

[0340] In ward conditions with frequent nursing obstruction / intervention, this invention, while maintaining high sensitivity, significantly reduced the false alarm frequency from 30.9 times / patient-day (common mattress alarm) to 5.5 times / patient-day, and increased the average advance warning time to 13.2 seconds (Example 1). In semi-physical ICU conditions with multiple in-place tubing, this invention achieved a detection rate of 0.91 for tubing traction near-loss events, with an average advance warning time of 7.8 seconds, significantly better than that without... The comparative scheme (Example 2). In the scenario of positional intolerance risk during the sedation recovery period, the present invention achieved a detection rate of 0.88 for high-risk physiological events, with a lower false alarm rate, proving that it can not only prompt for getting out of bed, but also indicate whether getting out of bed is safe (Example 3).

[0341] The foregoing description of embodiments of the present invention, through which those skilled in the art are able to implement or use the present invention, will be readily apparent to those skilled in the art. Various modifications to these embodiments will be readily apparent to those skilled in the art. The general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novelty disclosed herein.

[0342] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0343] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0344] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0345] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0346] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0347] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, like read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0348] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

Claims

1. A method for multimodal monitoring and risk early warning of out-of-bed activities, characterized in that, The method includes the following steps: S1, multimodal acquisition: based on the sampling time Use this as an index to retrieve multimodal time series datasets of the same monitored object. S2, Alignment and Confidence Assessment: [The text abruptly ends here, likely due to an incomplete sentence Time alignment is obtained And for each mode Calculate modal confidence , For modal sets; when hour, The confidence threshold is used to determine whether the mode is in The time interval is the low-confidence mode and its fusion weight is reduced; S3 Stage Identification: Based on Input a pre-defined time-series identification model, output a probability vector of the out-of-bed activity stage. And define the set of out-of-bed activity stages as S4, Gated fusion and postoperative risk scoring: based on modal confidence level Weighted fusion is performed to obtain the fusion stage probability. And calculate the cross-modal consistency index. ;when hour, Calculate bed urgency using the consistency threshold. Physiological response consistency index Pipeline Risk Index And obtain a risk score based on this. ; S5, Tiered Early Warning and Closed-Loop Update: [This will...] With warning threshold Comparison to obtain risk level And trigger the corresponding warning action.

2. The method according to claim 1, characterized in that, In step S1, the At least include: bed surface load / pressure distribution data Non-contact attitude or displacement data Vital signs data At least two of them, and optionally including wearable motion data. and / or pipeline status data ; in, For timestamps, This refers to the bed surface pressure matrix or bearing vector. For attitude key points, point clouds, or displacement features obtained from depth / infrared / millimeter wave, At least systolic blood pressure Heart rate Blood oxygen saturation One of them, For acceleration / angular velocity or equivalent motion characteristics, It is a measure of the in-situ status, displacement, or tension of drainage tubes, urinary catheters, or infusion lines.

3. The method according to claim 1, characterized in that, In step S2, the fusion weights and fusion results satisfy: ， ； in, For modality Normalized fusion weights, Modal confidence; And / or, the cross-modal consistency index satisfies: in, For the number of modes, The Kullback-Leibler divergence; And / or, the risk score meets the following requirements: ； in, For the Sigmoid function, For the coefficient vector, These are the weighting coefficients. For bias terms, The postoperative risk prior vector must include at least one of the following: sedation / delirium assessment, pain assessment, medication status, history of falls, or nursing risk grading. The urgency of getting out of bed and the predicted time to get out of bed The calculation yielded: ， The unit is seconds; Pipeline risk index; It serves as an indicator of consistency in physiological responses.

4. The method according to claim 3, characterized in that, The Include at least two of the following variables: sedation / agitation score Pain score Indication dose of vasoactive drugs Number of indwelling catheters after surgery or fall risk scale score ;in To quantify the level of calm / agitation, To quantify pain levels, For whether or not vasoactive drugs are used, a binary or dose-gradient variable. For the in-situ pipeline count, The nursing scale score; And / or, the pipeline risk index It is determined by at least one pipeline in-situ state and pipeline tension characterization quantity, wherein the pipeline tension characterization quantity includes: pipeline end displacement. Pipeline path bends and changes or pipeline tension At least one of them; This represents the displacement relative to the baseline position. The change in curvature It is a tension or equivalent tension index.

5. The method according to claim 3, characterized in that, In step S2, the initial postoperative bed rest stability window is also considered. Calculate baseline vital signs The physiological response consistency index Composed of vital sign shifts related to body position changes, and including at least: and , in, and Don't be a stable window The mean within; And when and At that time, improve The value of is used to characterize the risk of orthostatic intolerance; among which The threshold for systolic blood pressure drop. This is the threshold for increased heart rate.

6. The method according to claim 1, characterized in that, In step S5, the warning threshold Prescription levels for postoperative rehabilitation activities Related, and satisfy when Lower when instructing "Do not get out of bed independently" and With early warning, when Increase when the instruction "Allow bedside sitting / standing training" is given. and To reduce false alarms; among which It is a discrete rank variable; And / or, in step S5, in the feedback confirmation window Get the real results tags inside and based on right and / or , , At least one of them is updated individually; where .

7. A multimodal monitoring and early warning system for out-of-bed activities, characterized in that, The system is used to implement the method according to any one of claims 1-6, the system comprising: a multimodal acquisition unit for acquiring... and optional and form Alignment and confidence evaluation unit, used to generate And calculate fusion weight and baseline vital signs Stage identification unit, used for output With the probability of the fusion stage Consistency gating and postoperative risk scoring unit, used to calculate And in Time calculation With risk score A tiered early warning and closed-loop update unit is used to... Output risk level And trigger an early warning action, and in Internal receiving tag To update the threshold or model parameters.

8. The system as described in claim 7, characterized in that, The multimodal acquisition unit includes at least one of a mattress pressure sensor array and a non-contact sensing sensor, wherein the non-contact sensing sensor is one of a depth camera, an infrared camera, or a millimeter-wave radar; and vital sign data. Data collected via the bedside monitor interface; And / or, the consistency gating and postoperative risk scoring unit are configured to support cross-modal consistency indices. Greater than When entering conflict resolution mode, the conflict resolution mode includes at least one of the following: extending the decision holding time window. Request additional modal data collection, or adjust the risk level. Downgrade; among them The unit is seconds; And / or, the system is deployed as an edge computing architecture, outputting only at the bedside. And event summaries that do not contain recognizable images, and localize the original image frames without uploading them.

9. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method according to any one of claims 1-6.

10. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method according to any one of claims 1-6.

Images