Critical patient rehabilitation evaluation system based on artificial intelligence

Abstract

The invention discloses a critical patient rehabilitation evaluation system based on artificial intelligence, particularly relates to the technical field of medical information and intelligent rehabilitation evaluation, and is used for solving the problems that in an intensive care unit, multi-source data are not synchronous, nursing and treatment events are often caused, sedation and ventilation fluctuation pollution is caused, and the problem of serious illness is solved. And thus, the problem of distortion in judgment of which level cannot move nowadays and can move to can be solved. A unified time base is established through three types of anchor points of nursing, body position and cardiopulmonary rhythm, passivity and autonomy are judged in an influence time period, and the passivity is deducted and reserved; and then sedation, circulation, respiratory support and medication are compiled into safety budget with delay, linkage evidence quantity and uncertainty adjustment release rhythm, and degradation and rollback are triggered by an anti-fact track of margin segmentation, so that an executable prescription of'activity, reachable grade and reason 'is output beside a bed, event pollution and physiological fluctuation are inhibited, and the aim of preventing and treating the patient is fulfilled. And maloperation and missed operation are reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical information and intelligent rehabilitation evaluation, in particular to a critical patient rehabilitation evaluation system based on artificial intelligence. BACKGROUND

[0002] The existing critical rehabilitation evaluation mainly depends on artificial scale and grading management. The bedside data sources are scattered, including nursing records, body position changes, monitoring and ventilator parameters, etc. The time labels are not uniform, and there are many omissions and delays. Nursing behaviors often cause severe body movements or ventilation changes in a short period of time, which are easily misjudged as patient self-activity, thereby interfering with the judgment of "whether to carry out and to which level". The existing artificial intelligence solutions mainly focus on modeling the long-term outcomes such as mortality and hospitalization length, lack of bedside feasibility evaluation on a minute scale, and are difficult to directly guide specific actions such as sitting, side-bed sitting and standing.

[0003] Existing data intelligent attempts are usually based on time stamp alignment and black box probability output, without embedding clinical hard conditions such as sedation level, hemodynamic stability, ventilator support intensity and recent period medication into the calculation process as a gating mechanism. The influence of nursing events is often removed or simply weighted, which not only loses evidence but also fails to eliminate the pollution source. There is a lack of systematic traceable processing for data missing, delay and cross-device switching, resulting in insufficient stability in cross-department deployment. At the same time, the existing system mainly releases suggestions in a fixed rhythm, lacks a release strategy linked to evidence sufficiency, and rarely sees counterfactual reasoning and risk characterization based on safety margin for "how will the key physiological quantities change if the activity level is improved". It is difficult to provide a "reason chain" for execution and audit.

[0004] Under this background, in the critical environment with multiple source strong noise and sedation fluctuation, how to establish the influence period of nursing and body position events under the unified time base, distinguish passive and autonomous and suppress event pollution; how to solidify sedation, circulation, respiratory support and medication as calculable and stable safety constraints, continuously acting on evaluation and suggestion generation; how to run through evidence sufficiency in data processing and release rhythm, output action suggestions and their safety basis corresponding to rehabilitation grading on a minute scale, and form a replayable evidence chain for bedside execution and post-audit. SUMMARY

[0005] In view of the deficiencies in the prior art, the present application provides a critical patient rehabilitation evaluation system based on artificial intelligence to solve the problems mentioned in the background.

[0006] To achieve the above purpose, the present application provides the following technical scheme: a critical patient rehabilitation evaluation system based on artificial intelligence, comprising: Timebase anchor module: for establishing care, position, cardiorespiratory anchor uniform timebase and phase lock, connecting acquisition components, interacting with time sequence and event interface to carry time, phase, credibility in anchor structure; Event modeling module: for generating influence period around anchor, separating passive and autonomous and deducting interference, connecting timebase anchor module to carry identification, range, weight in influence window structure; Constraint budget module: for analyzing sedation, circulation, respiratory support and medication, generating safety budget and gate, connecting information system and ventilator to carry threshold, penalty, gate state in budget structure; Evidence encoding module: for encoding posture, body motion spectrum, oxygenation and laboratory backfill and generating uncertainty, connecting event modeling module and constraint budget module to carry evidence amount, uncertainty in evidence structure; Grade prescription module: for calculating rehabilitation feasibility score under constraint and mapping activity grade and prescription, connecting evidence encoding module to carry grade, care equipment, monitoring frequency, key points in prescription structure; Simulation publishing module: for generating counterfactual trajectory and risk curve under safety budget, setting publishing rhythm and degradation, connecting grade prescription module and constraint budget module, outputting suggestion through subscription publishing protocol to carry time sequence, factor, degradation reason in evidence chain structure, and landing.

[0007] In a preferred embodiment, the timebase anchor module comprises: Anchor classification unit, receiving writing according to care anchor, position anchor, cardiorespiratory rhythm anchor, and forming unique key with name, time, source, credibility; Phase locking unit, performing two-phase binding with ECG peak and respiratory phase zero-crossing in search window, performing timestamp alignment when phase is missing and recording reason; Conflict resolution unit, determining main event according to care priority over position, position priority over rhythm and writing apportioned weight; Interpolation governance unit, performing degradation interpolation when missing and marking non-interpolable; Interface verification unit, verifying signature and timestamp, writing rollback for late arrival, and writing coverage with correction mark; Database landing unit, outputting normalized time sequence and anchor table and writing into timebase database.

[0008] In a preferred embodiment, the event modeling module comprises: Event modeling module establishes front window and back window with care anchor, position anchor, cardiorespiratory rhythm anchor as zero point under uniform timebase; Retrieving care markers and bed angle and posture transition, ventilator trigger ratio change, cardiorespiratory phase continuity in sequence, determining passive activity and autonomous activity according to given priority order; The passive strength is output in parallel by calculating and subtracting weights from the body motion features within the passive window based on anchor point type and anchor point confidence. When adjacent windows overlap, the windows are merged while retaining the source list and coverage ratio. Conflicts are re-adjudicated based on nursing priority, body position, and rhythm correction. When both passive and autonomous evidence exist, output mixed annotations and prevent the generation of upgraded prescriptions in this window; A list of decontamination characteristics and impact windows is generated for downstream use.

[0009] In a preferred embodiment, the event modeling module input interface includes time period, anchor point identifier, range, weight fields, and carries a signature and timestamp; The window number is composed of the patient identifier, anchor time, and an increasing sequence and is unique. The idempotent key is composed of the window number and the timestamp to ensure deduplication and order. The output provides a list of decontamination features and impact windows, and records the window number, source list, coverage ratio, judgment type, deduction ratio, passive intensity, boundary time, processing result, operator, and timestamp in the evidence chain; When the motion channel is missing, only passive annotations and window boundaries are published to maintain downstream recognition. Recording and publishing are both completed under a unified time base and permission verification.

[0010] In a preferred embodiment, the constrained budget module includes: The constrained budget module receives sedation level, circulatory parameters, ventilator parameters, and medication records through a parameter subscription protocol. The messages include name, value, unit, and timestamp. Based on the calibration table corresponding to the device number and version number, the threshold is verified, the restriction is converted into a single amount and the minimum value is taken to generate a comprehensive budget. When the hysteresis time window is stably crossed, the gating status is updated. The gating status is allowed, restricted, or prohibited. The upstream credibility is reduced in stages. In case of conflict, the order of priority is: circulation, then respiration, then sedation, and finally medication. The gating release includes the triggering basis, the direction of the delay, the evidence chain number, and the timestamp.

[0011] In a preferred embodiment, the evidence encoding module includes: The evidence encoding module receives decontamination characteristics, safety budget and gating status, oxygenation and laboratory backfilling, aligns and unifies the time base and completes integrity and timing verification, executes interpolation and elimination rules based on upstream credibility stripes and records them into the evidence chain; The representation and evidence are generated by encoding attitude, body motion spectrum, and oxygenation together. The representation vector is output in the order of fields: attitude stability, body motion energy, oxygenation fluctuation, and gated state summary. At the same time point, the amount of evidence and uncertainty are propagated and synthesized step by step. When the channel conclusions are mutually exclusive and not within the nursing event window, the decision is made sequentially based on posture, oxygenation, and body movement. When the uncertainty reaches the threshold, the conservation strategy is adopted. The characteristic bus outputs the characterization, evidence quantity, uncertainty, and timestamp.

[0012] In a preferred embodiment, the graded prescription module includes: The graded prescription module performs integrity and timing checks on characterization, amount of evidence, uncertainty, safety budget, and gating status. If the gate is not ready, it enters the waiting queue; if the gate is full, the conservative prescription will be output. Scores are generated based on physiological baselines and weighted by the amount of evidence and uncertainty. Safety budget and gating status are applied to the scores as upper limit constraints and continuous penalties, respectively. The activity level is mapped according to the version-locked level mapping table, and a prescription is generated from the version-locked prescription template library. The prescription includes protective gear, monitoring frequency, execution points and review points. When the level is higher than the security budget, the pre-set degradation path will be used. The output includes a chain of evidence, which contains a timeline, key factors, gating entries, and version identifiers.

[0013] In a preferred embodiment, the simulated release module includes: Consistency verification unit, counterfactual solution unit, risk mapping unit, and downgrade rollback unit; The consistency verification unit aligns the multi-channel representations under a unified time base, and generates the total amount of evidence based on the channel confidence and sensor health, and writes it into the evidence chain. The counterfactual calculation unit uses the steady-state segment as a baseline, and adds the activity load to heart rate, mean arterial pressure, blood oxygen saturation and respiratory trigger rate one by one, and calculates the remaining margin of the relative safety budget. The risk mapping unit forms a risk curve based on the margin segmentation; The downgrade and rollback unit generates recommendations and determines downgrade and rollback based on the risk curve, termination conditions, and safety budget assessment results.

[0014] In a preferred embodiment, the simulated release module further includes a release control unit and a trace auditing unit; The release control unit switches channels between the hospital message bus and the edge direct link through the publish-subscribe protocol. The send buffer is resent in first-in-first-out order, and deduplication is performed by forming an idempotent key with patient ID, suggestion number and timestamp. At the same time, the release rhythm is adjusted according to the total amount of evidence and the verified safe prescription is maintained in the case of insufficient evidence. The audit unit writes recommendations, activity levels, protective gear and staffing configurations, risk curve points, termination condition lists, evidence chain numbers and unified time base timestamps to both the prescription database and the evidence database. It uses patient identification, time period and version as index keys, and generates a reconciliation list and completes the rewriting and cancellation when an anomaly occurs.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. By establishing a unified time base using three anchor points—nursing care, body position, and cardiopulmonary rhythm—passive and voluntary actions are determined within the influencing period, and passive contributions are "deducted and retained." Sedation, circulatory support, respiratory support, and medication are then compiled into a safety budget with hysteresis, and the amount of evidence and uncertainty are linked to the release rhythm. Combined with counterfactual trajectory triggering degradation and rollback based on margin segmentation, an executable prescription that "can be moved, the level of attainment, and the reason" is output at the bedside is achieved, which suppresses interference caused by event contamination and physiological fluctuations, and reduces the risk of mis-movement and missed movement.

[0016] 2. By ensuring time consistency through unique keys and conflict resolution, recording the entire chain of evidence and encrypting signatures to prevent replay, combining publish / subscribe with backup channels and idempotent reissue, and reconciling prescription and evidence databases for consistency, a rehabilitation assessment system that is auditable, traceable, and stable in real wards is achieved, facilitating cross-department replication and supporting continuous improvement. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the structure of the artificial intelligence-based rehabilitation assessment system for critically ill patients according to the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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 skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Example: Figure 1 A schematic diagram of the structure of the AI-based rehabilitation assessment system for critically ill patients of this invention is provided. The AI-based rehabilitation assessment system for critically ill patients includes: Time base anchor module: used to establish a unified time base and phase lock for nursing, body position and cardiopulmonary anchor points, connected to the acquisition component, and interacted with the timing and event interface, with the anchor structure carrying the time, phase and reliability; Event modeling module: used to generate impact time periods around anchor points, separate passive and autonomous events and deduct interference, connect to time-based anchor point module, and influence window structure bearing identifier, range, and weight; Constraint Budget Module: Used to analyze sedation, circulation, respiratory support and medication, generate safety budgets and gating, connect information systems and ventilators, and use budget structures to carry thresholds, penalties and gating states; Evidence coding module: used to encode attitude, body motion spectrum, oxygenation and laboratory backfill and generate uncertainty, connecting the event modeling module and the constraint budget module, so that the evidence structure carries the evidence quantity and uncertainty; Graded prescription module: used to calculate rehabilitation feasibility score under constraints and map activity level to prescription, connected to evidence coding module, with prescription structure to carry level, protective gear, monitoring frequency, key points; Simulated release module: Used to generate counterfactual trajectories and risk curves under a safety budget, set release rhythm and degradation, connect the grade prescription module and the constraint budget module, output suggestions through the subscription release agreement, and use the evidence chain structure to carry the time sequence, factors, degradation reasons, and put them on disk.

[0020] The technical connections and implementation logic of the six modules are as follows:

[0021] The time-base anchor module first aggregates nursing, body position, and cardiopulmonary rhythm data to achieve a unified time base and phase lock. It then outputs standardized time series and anchor table data downstream via time series and event interfaces. The event modeling module generates impact time periods on both sides of each anchor point, determines passive and autonomous events, produces decontamination features and impact window structures, and writes back the associated anchor point identifiers. The constraint budget module subscribes to sedation, circulation, respiratory support, and medication data in parallel, generating safety budgets and gating states under a unified time base, forming hard constraint signals. The evidence encoding module integrates decontamination features and gating signals, combining oxygenation and laboratory backfill data for evidence processing. The system encodes data and outputs a representation stream carrying evidence quantity and uncertainty. The graded prescription module receives the representation stream and gating through the inference interface, calculates minute-level rehabilitation feasibility scores, maps them to activity levels and prescriptions, and provides information on protective gear configuration, monitoring frequency, and key execution points. The simulation and release module performs counterfactual deduction on prescriptions under safety budget constraints, generates counterfactual trajectories and risk curves, and determines maintenance, downgrade, and rollback based on margin segmentation and termination conditions. Recommendations are delivered to bedside terminals via a publish-subscribe protocol, and the time series, factor contributions, and downgrade reasons are written into the evidence chain and synchronously stored in the prescription and evidence databases. The overall data flows unidirectionally: "anchor point → window → budget and evidence → score and prescription → simulation and release." Anomalies are triggered by conservative branches based on evidence quantity and gating, balancing real-time performance and traceability.

[0022] Time-Based Anchor Module: Used to establish a unified time base and phase lock for nursing, positioning, and cardiopulmonary anchor points, connected to the acquisition unit, and interacting through the timing and event interface. The anchor structure carries the time, phase, and reliability. Specific implementation is as follows: This module addresses the issue of chaotic multi-source timing in intensive care units, aiming to clarify and trace the events of "who did what, when" and "what physiological phase they were in at the time." On the input side, anchor points are categorized into three types with standardized terminology: nursing anchor points include turning over, suctioning, nebulization, passive joint movement, and extubation assessment; positional anchor points include supine, semi-sitting, sitting up, left lateral decubitus, and right lateral decubitus; rhythmic segments include ECG peaks and respiratory phase zero crossings. Terminology is standardized here: standardized timing refers to a unified timeline that has been aligned and phase-locked; the chain of evidence refers to a replayable trajectory recording the source, criteria, and result; the day boundary is defined by the local zero hour of the ward; degraded interpolation refers to restricted interpolation used when confidence decreases; and rollback refers to the reversal of invalid or duplicate writes.

[0023] The system establishes a unified time base based on the local time of the ward, uses the hospital's network for time synchronization, and retains the original time zone field of the collected data. When a clock rollback occurs, the system segments the preceding and following segments using a "time synchronization event" as a cutter and marks them in the evidence chain. Each anchor entry must simultaneously provide four items: name, time, source, and credibility; any missing item will result in rejection. If the handshake fails within five seconds, the system will automatically retry three times. The credibility is not an arbitrary percentage but is generated according to rules and can be recalculated: manual annotation has a higher weight than device events; millisecond-level timestamps have a higher weight than manual entries; consecutive segments with the same source and equal intervals receive bonus points; device self-test anomalies result in a deduction of points. After merging the four items, the weight is truncated to zero or one, and each contribution is written into the evidence chain.

[0024] Phase locking opens a search window centered on the event time, with a default of 20 seconds before and after. If the patient's rhythm is slow, the search window should be no less than one respiratory cycle and no less than two cardiac cycles. First, the most recent ECG peak is searched within the window and paired with the most recent respiratory phase zero crossing. If both are present, it is bound as "biphase"; if only one is present, it is marked as "single-phase"; if both are missing, the timestamps are reverted for alignment, and the reason is explained. Drift and missing data are not arbitrarily imputed: imputation is triggered when the interval between adjacent homologous records exceeds three times the median interval of that source in the past hour; forward hold is prioritized, with a single segment not exceeding 30 seconds; if the upper limit is exceeded, a linear transition between the preceding and following points is used, with the total duration not exceeding 90 seconds; if this is still exceeded, it is marked as "not suitable for imputation," and subsequent steps are weighted accordingly. Conflicts cannot be resolved by simply stating "subject to so-and-so": When two events occur simultaneously, they are adjudicated according to a matrix of "nursing care takes precedence over body position, and body position takes precedence over rhythm." The primary event is fully involved in subsequent characteristics, while the secondary event is only recorded. When apportionment is required, nursing care and body position are weighted at a ratio of 7:3, and body position and rhythm are weighted at a ratio of 9:1. The apportionment ratio is then incorporated into the chain of evidence.

[0025] Late arrivals and out-of-order requests are managed using a unique key, consisting of the name, time, and source. Late packages with versions earlier than existing ones are rolled back as invalid. Corrections must include a "correction flag," and the system overwrites the old value with the new one, recording the source and time of the correction. Interfaces are kept as simple and clear as possible: write requests have four fixed fields, and responses return the record number and version. Common errors are uniformly enumerated as missing fields, time out of bounds, time reversal, duplicate writes, signature verification failure, and requests exceeding the window limit, with corresponding handling of rejection and rollback, request resending, or recording a conflict list. Request bodies must be signed and timestamped, with an acceptance window of ±30 seconds; requests exceeding the window are rejected. After processing, two types of results are output: one is a normalized time series after alignment and phase locking, and the other is a structured anchor table. Both are simultaneously written to the "time base library," with the main table storing the time series and the secondary table storing the anchors. The main index is the patient number plus the date boundary. Support is provided for playback in ascending order of time and batch export by patient number plus date boundary. Data is stored in layers, with hot layers retaining data for 30 days and cold layers retaining data for at least six months.

[0026] The project makes the following engineering commitments regarding concurrency and timeliness: no fewer than fifty concurrent connections per ward, end-to-end latency no more than ten seconds, and a one-hour buffer; in case of network jitter, data is first written locally to disk, and then resent in chronological order after the link is restored to ensure the timeline is not reversed. Permissions are divided into three levels—read-only, write, and audit—based on the principle of minimum necessity, and keys are rotated monthly; all processing rules are locked with version numbers, and when changes are made, the version number, effective time, and change summary are written into the evidence chain, and a rollback point is retained.

[0027] Routine quality control employs a fixed procedure: At least twenty events are randomly selected from patients daily, and the event-phase fit distribution and interpolation error distribution are calculated. The fit rate is no less than 95%, and the median interpolation time error is no greater than two seconds. Substandard segments are automatically re-run; if still substandard, they are sent for manual review, and all actions are rewritten into the evidence chain. To illustrate real-world behavior, two scenarios are presented: A positive example is a nurse recording suctioning, raising the bed angle from supine to 45 degrees, and a short-term increase triggered by the ventilator. The system first aligns suctioning and bed angle to a unified time base, then finds the nearest ECG peak and respiratory phase zero-crossing near the event to complete "dual-phase" binding. An influence window is established three minutes before and after suctioning, marking this significant change in body movement as a passive contribution. Subsequent feature calculations deduct its weight to avoid misclassifying "nursing movement" as "patient movement." A negative example is a late nursing record that conflicts with the bed angle record, along with a short missing ECG segment and network jitter. The system identifies late writes with a unique key, and rollbacks are invalid if no correction flag is included. Due to ECG segment gaps, "single-phase" locking is completed using only the respiratory phase, and the reason is explained. When the network is unstable, local disk writes are performed first, and then resends in sequence after recovery. The normalized timing of the output remains monotonically increasing.

[0028] Overall, the input items, rules, thresholds, interfaces, storage, quality control, and security closed loop have all been implemented. Based on this, those skilled in the art can reproduce the same timeline and evidence chain without creative effort. The phase locking strategy, message channel, and deployment location allow equivalent replacements without changing the field and timing commitments, ensuring both feasibility and retaining flexibility for implementation at the departmental level.

[0029] Event modeling module: Used to generate impact time periods around anchor points, separate passive and autonomous events and subtract interference, connect to the time-based anchor point module, and determine the impact window structure's load-bearing identifier, range, and weight. Specific implementation is as follows: The event modeling module begins operation after the unified time base is established, receiving normalized time series, anchor point tables, and body movement and posture features as input. The normalized time series is a sequence of multi-source data aligned under the same time base, including second-level timestamps, sampling start and end dates, and integrity markers. All timestamps are timed in the hospital's standard time zone, with Beijing time as the default. A deviation of no more than one second between the device clock and the unified time base is allowed; exceeding this deviation is recorded as "clock deviation" and triggers correction. The anchor point table records the occurrence time, source, and reliability of nursing anchor points, position anchor points, and cardiopulmonary rhythm anchor points. Body movement and posture features are derived from wearable and bedside components, reflecting triaxial acceleration power distribution, posture category, bed angle, and position switching markers.

[0030] The module generates an impact period around each anchor point, setting a front and back window with the anchor point occurrence time as zero: nursing events have a default front window of three minutes and a back window of seven minutes; body position changes have a default front window of two minutes and a back window of five minutes; rhythm anchor points are only used for boundary correction, with each correction not exceeding thirty seconds; when the anchor point reliability is not less than 80% and the body movement signal integrity rate is not less than 90%, the front and back window lengths are allowed to be finely adjusted by one-third within a predetermined range; otherwise, a fixed setting is used. The body movement signal integrity rate is defined as the ratio of the number of effective sampling points within the window to the theoretical number of sampling points. The sampling jitter tolerance is not more than one second, and when determining whether an adaptive window is allowed, a score of not less than 90% is used as the criterion; effective sampling points refer to samples that are not missing, not saturated, and not marked as outliers. The saturation criterion is that the sensor's full scale lasts for more than one second, and outliers are samples that exceed the calibration range or whose difference between adjacent samples exceeds one-tenth of the full scale.

[0031] When adjacent windows overlap by more than half, they are automatically merged. After merging, the source list and their respective coverage ratios are retained, and their contributions are allocated according to their respective coverage durations within the merged window. Those with a coverage ratio of less than 10% are merged into "Other" and the reason is recorded. The source list only includes five categories: nursing records, body position sensors, bed control systems, ventilator logs, and manual entries. Conflicts at the same time are adjudicated in the following order: nursing care first, body position second, and rhythm correction third.

[0032] Within each period of influence, three pieces of evidence are examined sequentially: First, whether nursing or positioning markers exist, and whether transitions in bed angle and posture are verified. A sudden change in bed angle is defined as an angle change of no less than ten degrees within ten seconds. A change in posture refers to a change in posture from supine or supine to lateral, sitting up, or sitting on the side of the bed. Second, whether the ventilator trigger rate has increased by no less than 10 percent relative to the baseline of the previous minute and the absolute value exceeds one per minute. The baseline is the trigger count of the previous minute divided by one minute. If the previous minute includes nursing or positioning events, the most recent one-minute period without events is used as the baseline. Third, whether the cardiopulmonary phase is continuous. Phase continuity means that the change in phase difference between two adjacent respiratory cycles and the ECG cycle is no more than one-tenth of the cycle length and there is no jump within three consecutive cycles.

[0033] The judgment order is as follows: first, check nursing or positional markings; second, check the ventilator trigger rate; and finally, check phase continuity. Passive activity is defined as any one of these criteria being met without any contrary evidence. Spontaneous activity is defined as the median change in body motion power (calculated in 10-second increments) not exceeding 20% ​​for at least one minute, and the body motion power consistently exceeding 1.5 times the patient's resting median body motion power. When both types of evidence coexist, a mixed label is output, prompting downstream users to adopt a conservative strategy. No escalation prescriptions can be generated within this window; only maintenance or downgrade recommendations are allowed. Body motion characteristics within the passive window are weighted according to anchor point type and confidence level. Nursing care weights are 60% to 80%, positional weights are 40% to 60%, and rhythm correction weights are 10% to 30%. Each weight is linearly reduced proportionally with anchor point confidence, and the cumulative deduction ratio for a single window does not exceed 80%. Deduction is numerical, not deleted. Passive contributions are fully recorded in the parallel labeling channel for subsequent safety budget reference. Passive intensity is also calculated and output.

[0034] Passive intensity is calculated as the average of the squares of the three gravitational acceleration components within the sampling window, with the unit being the square of gravitational acceleration. The "patient's median resting body dynamics power" serves as the baseline, calculated by taking the median of a five-minute window within the most recent ten minutes that excludes nursing care and positional anchor points and whose body dynamics power is lower than the patient's median value for that shift. This median is updated every five minutes. If the current shift cannot form this window, it reverts to the baseline of the previous shift using the same caliber. Shifts are typically eight hours; if no shift field is available, an eight-hour window is used instead. Inputs and outputs are published in one-minute increments, all carrying second-level timestamps, with an alignment tolerance of no more than one second. Data spanning different days or exceeding the sampling range is truncated at the boundary, and the reason for "out-of-bounds truncation" is recorded.

[0035] The interface follows the principle of minimum necessary data. Required fields in the request body include time period, anchor point identifier, range, and weight. All requests must be signed and timestamped, with a allowed deviation of no more than five minutes. Requests exceeding the window limit are rejected with a reason code. Missing fields return 101, reverse time returns 102, signature verification failure returns 103, and after three failed retries, 104 is returned and the request is written to the local cache. Once the channel is restored, requests are resent in chronological order. The window number is generated by concatenating the patient number, anchor point time, and an increasing sequence, ensuring uniqueness and idempotency. The idempotency key consists of the window number and timestamp, ensuring no replays and no missed requests. Internal module states progress in the order of window creation, identification, deduction, and publication, with window priority determined by the order of occurrence. End-to-end processing latency is no more than ten seconds, and at least fifty windows can be processed concurrently in the same ward. The cache retention period is one hour. When the motion channel is temporarily unavailable, only passive annotations and window boundaries are published to ensure downstream recognition of nursing impacts.

[0036] All outputs are provided in two categories: downstream users subscribe to decontamination features to perform evidence coding and scoring, and simultaneously subscribe to impact windows to grasp passive traces and boundaries; the allocation list includes window number, allocation ratio, and deduction start and end. The minimum traceable fields of the evidence chain are fixed as window number, source list, coverage ratio, judgment type, deduction ratio, passive intensity, boundary time, processing result, operator, and timestamp. All of these are written into the evidence database to support audit playback and dispute review. Read and write operations are subject to hierarchical access control and signature verification, and anti-replay verification is performed within the timestamp window.

[0037] Acceptance testing employs bedside record playback, selecting no fewer than one hundred nursing events as the gold standard, covering common scenarios such as turning over, suctioning, nebulization, and position elevation. Passive identification accuracy is no less than 90%, window alignment error is no more than one second, and the consistency rate between mixed labeling and manual review is no less than 90%. Taking turning over as an example, the module sets a front window before turning over and a back window after turning over. In the back window, the duration of short body movement pulses is less than 30 seconds, and the peak body movement power is more than three times the patient's median resting body movement power. At the same time, a position change marker appears, accompanied by a ventilator trigger rate that meets the criteria. The phase is continuous, so it is labeled as passive and deducted according to the nursing weight, while retaining the passive intensity and boundary. If the nursing record is missing but the position change and short body movement exist, the passive tendency is output with a downgrade marker, which is used by downstream systems to reduce the amount of evidence. To adapt to the rhythms and equipment differences in different wards, the module provides two strategies: fixed window and adaptive window. Both expose the same output structure and record the strategy and level used in the evidence chain. The deployment mode can be centralized at the bedside edge or nurse station. Communication can be carried out through message bus or local call. Any implementation that does not change the above timing and field caliber is considered equivalent.

[0038] In the overall process, a unified time base provides a reliable anchor point, windowing and merging limit the impact of nursing care to an interpretable time frame, judgment rules clarify the distinction between passive and autonomous actions and leave complete traces, parallel outputs of deduction and retention ensure clean input for subsequent scoring and provide traceable evidence for safety budgets, and interface contracts, fault tolerance and auditing mechanisms guarantee stable operation and verifiability in real ward environments, thereby meeting the requirements of openness, sufficiency and feasibility.

[0039] The constraint budget module is used to analyze sedation, circulation, respiratory support, and medication administration, generate safety budgets and gating mechanisms, and connect the information system and ventilator. The budget structure carries thresholds, penalties, and gating states. Specifically, the implementation is as follows: The constrained budget module is designed for real-world bedside processes, transforming hard conditions such as sedation, circulation, respiratory support, and proximal medication administration into calculable and traceable safety budgets, and issuing gating statuses for downstream use. The module continuously subscribes to four types of data from the hospital information system, monitoring equipment, and ventilator: sedation level records including rating category, assessor, and timestamp; circulatory parameters including heart rate, mean arterial pressure, pulse pressure, lactate, and their proximal changes; ventilator parameters including inhaled oxygen concentration, positive end-expiratory pressure, tidal volume, trigger ratio, and minute ventilation; and proximal medication administration including drug name, route of administration, dosage, start time, and end time.

[0040] The transmission uses a parameter subscription protocol. Messages must carry four items: name, value, unit, and timestamp. Messages lacking units or with reversed timestamps are rejected, and the reason is recorded. Simultaneously, a retransmission is requested from upstream to prevent inconsistent data from entering the budget link. After data enters, it is first aggregated. For the same type of data from different channels at the same time, the one with the highest credibility is retained, and the rest are attached as backup evidence to the same time point to ensure that the data retrieval is verifiable.

[0041] Thresholds and ranges are uniformly locked based on device number and version number: when the message carries device number and version number, the corresponding calibration table is executed; when the device number is missing, the department-level default table is used, and the use of the default table and the reason are written into the evidence chain; all calibration tables and mapping rules are managed by version number, and the effective time and modifier are entered into the change log, and during playback, each record is strictly restored according to the version. Subsequently, threshold verification is performed, and all quantities within the allowable range are reviewed. Values ​​exceeding the limit are capped or capped according to the nearest rule, and an alarm event is issued and the evidence chain and trigger source are recorded. Budget calculation adopts discrete tiering and conservative synthesis: each type of physiological restriction is first converted into a single quota, and a dimensionless value between zero and one is taken, where zero indicates prohibition and one indicates that the item does not constitute a restriction.

[0042] Sedation reduction is divided into four levels: mild corresponds to medium level, moderate to low level, severe to very low level, and deep depression to zero level. Circulatory reduction is based on the trend of mean arterial pressure and lactate; a high level is given if the mean arterial pressure is in the safe zone and has recently stabilized or rebounded; a medium to low level is given if the mean arterial pressure is close to the lower limit and decreasing; and a very low level is given if the mean arterial pressure decreases and is accompanied by an increase in lactate. Respiratory support reduction is mainly based on inhaled oxygen concentration and positive end-expiratory pressure; a low or very low level is given if the mean arterial pressure is in a relatively high level. Medication reduction for increases in strong sedatives or vasopressors is reduced by one or two levels based on the dosage intensity and duration of action. When all individual levels are combined into a comprehensive budget, the minimum value is taken to prioritize safety. To suppress tremors, a hysteresis band is superimposed on the comprehensive budget; the gating state only changes if the hysteresis threshold is stably crossed within a continuous three to five-minute time window. The preferred hysteresis time window is three minutes, configured by the department within permissible limits and managed by version number.

[0043] The budget is updated in a rolling one-minute increment based on a unified time base, with the increment aligned with the event sequence. The system time base alignment uses the hospital's standard clock, and timestamps are based on the hospital's standard time, with an alignment error preferably not exceeding three seconds. Gating is divided into three levels: allowed, restricted, and prohibited, corresponding to the highest, middle, and lowest segments of the comprehensive budget, respectively. In the event of conflicting evidence near the boundary, the order of priority is: cyclical administration first, followed by respiration, then sedation, and finally medication. When a gating control is issued, the triggering criteria, hysteresis direction, and effective duration are included to facilitate easy reference by the nursing station and downstream personnel.

[0044] Considering latency and missing data in the data acquisition process, an upstream credibility reduction is introduced before budget calculation. Credibility is divided into three levels: if below the minimum threshold, it is uniformly downgraded to the lowest level and a ban is issued; if between the minimum and intermediate thresholds, it is downgraded by one level and a restriction is issued; if above the intermediate threshold, no further reduction is applied. Preferably, the minimum threshold is set at 30%, and the intermediate threshold at 60%. The reduction is lifted after credibility recovers and remains for several minutes, while the hysteresis principle remains effective. If sedation records and medication prescriptions are inconsistent, or if loop parameters contradict alarm records, conflict entries are immediately generated and pushed to the department for review. The review time limit is in minutes, preferably ten minutes. If review is not completed within the time limit, the system maintains the restriction or ban according to a conservative strategy until the data on either side is corrected. All conflict entries include the trigger source, comparative evidence, and handling results, and are uniformly written into the evidence chain and audit log.

[0045] To ensure timeliness, the module employs a lightweight queue with rate limiting and depth control, keeping end-to-end processing latency below five seconds. Budget updates are generated on a minute-by-minute basis, with sudden high-frequency reports smoothed by rate limiting. When the queue overflows, the earliest backup evidence is discarded in chronological order, retaining the primary evidence. In case of network jitter or upstream timeout, data is first cached and then replenished in chronological order upon recovery. All writes use the patient identifier plus time period as the idempotent key, and duplicate reports are not counted twice. The interface is read-only, with only the minimum necessary fields disclosed. All write actions require signature verification and a timestamp window to prevent replay, meeting hospital tiering and auditability requirements. The retention period for the evidence chain and audit logs is preferably six months, but at least three months.

[0046] The output side primarily focuses on safety budgets and gating status, retrieved by downstream users through a read-only subscription channel. Output fields include the current budget value, gating level, triggering basis, effective duration, evidence chain number, and timestamp, stored in the pre-prescription database and indexed by patient identifier, time period, and version. Error code ranges are allocated by function segment: anchor-related errors are segment 1, window-related errors are segment 2, parameter subscription-related errors are segment 3, and message publishing-related errors are segment 6. Unlisted entries are processed as catch-all codes and incorporated into the evidence chain for easy fault location and review. The module commits to two core acceptance criteria: false release rate and coverage rate, serving as the main indicators of safety and effectiveness, respectively. The preferred targets are set at no more than 1% for false release rate and no less than 95% for coverage rate. Statistics are conducted weekly, by disease area and bed, with the statistical period automatically extended until the target is met if the sample size is insufficient. Verification is based on cross-referencing of bedside execution records and adverse event records.

[0047] To accommodate the differences in equipment and processes across different hospital areas, the module allows for equivalent substitutions without altering the calculation principles and output objects: budget synthesis can be achieved using a segmented or weighted approach, subscription channels can be replaced with an intra-hospital message bus or direct bedside connection, and deployment locations can be placed at the bedside edge or nurse station server. However, the gating level, evidence chain field, hysteresis caliber, and credibility level remain consistent, allowing upstream and downstream systems to connect without modification.

[0048] A typical scenario is as follows: the inhaled oxygen concentration is at a high level and the positive end-expiratory pressure is at a high level; the overall circulatory parameters are stable, but the mean arterial pressure is close to the lower limit and has slightly decreased recently; sedation is slightly restricted and has not been increased recently. After the rule is implemented, the single-item allowance for respiratory support falls to the low level, the single-item allowance for circulation falls to the low-to-medium level, and the single-item allowance for sedation is at the medium level. Medication is not reduced, and the minimum value of the comprehensive budget is taken to reach the restricted level. The gating is restricted, allowing only sitting up and sitting on the side of the bed, and the evidence chain records respiratory support as the primary cause and circulatory trend as the secondary cause. If the mean arterial pressure rises and the inhaled oxygen concentration drops to the safe zone within the next three minutes, and the budget crosses the hysteresis threshold and remains stable, the gating switches from restricted to allowed. If the credibility drops below the intermediate threshold during this period, the restriction is maintained according to the reduction rule, and the effective duration is extended until the credibility recovers.

[0049] The overall approach has clearly defined what data to collect, what rules to use for judgment, how to synthesize the budget, when to change the gating, how to resolve conflicts, and how to record and audit the results at the executable level. It also clarifies clock alignment, retention period, and error code segmentation, which facilitates engineering implementation and multi-center reproduction. It also makes it easy to directly reference key mechanisms such as graded reduction, minimum value synthesis, three-segment gating, hysteresis time window, three-level confidence reduction, evidence chain field, and idempotent key in the claims, ensuring that the disclosure is sufficient, the expression is clear, and the verification is possible.

[0050] Evidence coding module: Used to encode attitude, body motion spectrum, oxygenation, and laboratory backfill and generate uncertainty. Connects the event modeling module and constraint budgeting module, using evidence structure to carry evidence quantity and uncertainty. Specific implementation is as follows: This stage receives decontamination characteristics, safety budgets, and gating status from upstream outputs, as well as oxygenation and laboratory backfilling. All inputs are first standardized to the same timeframe, and their sources and timestamps are clearly marked to ensure traceability. Before processing, completeness and temporal continuity are checked. If missing timestamps, abrupt changes in sampling intervals, or batch delays are found, interpolation or rejection is selected based on the previously established credibility bands, and the results are written into the evidence chain. The time and window dimensions are fixed as follows: body movement and posture use a short window (default 30 seconds, sliding step 10 seconds, allowable tolerance ±5 seconds); oxygenation and cross-channel consistency use a long window (default 3 minutes, sliding step 30 seconds, allowable tolerance ±15 seconds). All timestamps are based on the ward clock, recorded at the second level after timebase correction, and segmented at midnight across days, with a segmentation marker written into the evidence chain.

[0051] Then, parallel coding is performed according to the channel: the attitude focuses on the body position category and stability. The body position category is limited to five levels: supine, semi-reclining, sitting up, sitting on the side of the bed, and standing. The attitude comes from the fusion result of bed corner sensing and inertial measurement. The bed corner is measured in degrees, and the inertial measurement is measured in multiples of gravitational acceleration. After fusion, a single attitude conclusion and stability label are produced. Stability is assessed by combining the number of posture changes and the amplitude of bed angle changes within adjacent short windows. Zero changes and bed angle changes of no more than five degrees are considered stable. One change or a bed angle change of five to fifteen degrees is considered mild. Multiple changes or a bed angle change of more than fifteen degrees are considered moderate or severe. The body motion spectrum focuses on the distribution and abrupt changes of low-frequency and mid-frequency energy within the short window. The low-frequency band is taken as 0.1 to 0.5 Hz, and the mid-frequency band is taken as 0.5 to 3 Hz. A low-to-mid frequency ratio between 0.5 and 2 is considered normal activity. A ratio below 0.5 is considered very weak activity. A ratio above 2 and within the event influence window is considered passive disturbance dominant. Oxygenation focuses on the rate of decrease and recovery time of blood oxygen saturation. A decrease of no more than two percentage points within thirty seconds and a recovery within three minutes are considered stable. A decrease of more than three percentage points or failure to recover within three minutes is considered abnormal fluctuation.

[0052] Each channel generates a representation and an evidence measure for the current time period. The representation condenses key morphological points and information density, while the evidence measure quantifies the data quality and reliability of the conclusions for that channel, taking into account factors such as whether it falls within the event impact window, whether gating restrictions are triggered, and whether cross-device switching occurs. The field order of the representation vector is fixed as attitude stability level, body energy level, oxygenation fluctuation level, and gating state summary. Each field only contains the level number and textual annotation of the level basis to avoid implicit meaning drift. After the results from multiple channels are generated in parallel at the same time, the evidence measure and uncertainty are propagated sequentially in a fixed order, followed by synthesis and publication. This allows uncertainties to accumulate gradually from the bottom up, avoiding the risk being discovered only at the last step and requiring a complete rollback.

[0053] Evidence quantity and uncertainty are categorized using a hierarchical approach without formulas: Evidence quantity is divided into five levels, with an initial level of four. Levels below four are applied when evidence is at the end of the influence window and incomplete data is deducted. Levels below three are applied if there are sampling jumps or cross-device switching issues (corrected but not corrected, down two levels). Levels below three are applied if the data is restricted at the gating point, and down one level if manual data is late for more than one long window. Evidence quantity of level three or higher is considered sufficient, and below three is considered insufficient. Uncertainty is also divided into five levels, inversely linked to evidence quantity. Uncertainty increases by one level for each decrease in evidence quantity. If cross-channel data is mutually exclusive and no consensus is reached, it increases by an additional level, up to a maximum of five levels. Modules only release data at a regular rhythm when uncertainty does not exceed level three; otherwise, a conservation strategy is adopted.

[0054] The condition for entering the conservation phase is that the amount of evidence is below level three or the uncertainty is above level three within any short window. After entering, the release frequency is reduced from once per minute to once every three minutes. Exiting the conservation phase requires that the amount of evidence recovers to at least level three and the uncertainty is not higher than level three within two consecutive long windows. The start and end times and reasons for entering and exiting are recorded. When the channel conclusions are mutually exclusive, first check if it is within the nursing event window; if not, make a decision in the order of posture, oxygenation, and body movement, and score them in parallel with confidence weighting, with posture weighted at 40%, oxygenation weighted at 40%, and body movement weighted at 20%. If the difference in scores after parallel scoring is less than 10%, mark it as undetermined and raise the uncertainty by one level, while issuing a supplementary sampling prompt; if it is still undetermined after two consecutive long windows, maintain the conservation phase and record the decision number.

[0055] To ensure stability during peak periods, the module has a concurrency threshold of 50 channels. If this threshold is exceeded, flow control is implemented, prioritizing attitude and oxygenation processing. Body motion is then placed in a waiting queue with a depth of 20 and a single waiting time limit of 20 seconds. If this limit is exceeded, the current short window body motion spectrum is discarded, a passive degradation flag is output, and conservation is initiated. Backoff employs exponential backoff with a maximum of 30 seconds, and the number of backoffs and their reasons are recorded in the evidence chain. Under normal circumstances, the end-to-end latency from complete input to valid output is no more than 10 seconds. Intermediate results from the most recent hour are cached locally. Upon cache expiration, the cache is automatically pushed to disk, and the index is written to the audit database. Once the cache is full, the results are processed using a first-in-first-out (FIFO) disk strategy.

[0056] When input validation fails, rollback is performed according to priority: if posture is missing but body movement and oxygenation are complete, output continues as usual, a missing marker is filled in at the posture position, and the evidence quantity is reduced by one level; if oxygenation is missing, the conservation branch is entered directly and a supplementary sampling prompt is issued; if all three channels are missing simultaneously, no output is issued, only the time period gap is recorded and an upstream link self-check is triggered; for late arrivals within a long window, history is not rolled back, only the evidence chain is rewritten and the current uncertainty is updated; for those arriving beyond a long window, the current output is not updated, only an expired storage mark is placed on disk. External output provides four types of fields in read-only mode via the feature bus: characterization, evidence quantity, uncertainty, and timestamp; invalid values ​​are uniformly marked as null values ​​and the reason is written in the evidence chain; all units are uniformly fixed: bed angle is measured in degrees, blood oxygen is measured as a percentage, body movement energy is measured as an energy ratio of multiples of gravitational acceleration, and time is measured in seconds, and the interface layer does not change with the device; any missing evidence quantity will trigger a conservative branch and issue a supplementary sampling prompt.

[0057] Data is uniformly anonymized and stored in the database. The minimum set of fields in the evidence chain is fixed as follows: timestamp, data source, window parameters, classification basis, downgrade reason, weight decision, version number, and the identifier of the writer or writing program. All fields are mandatory, and all read and write operations are traceable. The representation stream is the sole input to downstream grade calculations and an important component of the interpretation chain. Any prescription or publication downstream can be traced back to the sampling, decision, and gating status at the time of encoding along this chain. In scenarios with limited computing power or access points, the representation generation can be switched to an equivalent scheme of rule-based representation: attitude stability classification, body energy classification, and oxygenation fluctuation classification are directly given by engineering rules, and output according to the same evidence quantity and uncertainty classification caliber, with consistent interfaces, timing, auditing, and security boundaries. Regardless of whether a learning-based or rule-based representation is used, the input, processing, and output boundaries of this module remain unchanged. If it is deployed in conjunction with adjacent links, it will not affect the above data caliber and evidence chain records.

[0058] The expected outcome for a typical bedside case is as follows: within three minutes of the turning event, the motion spectrum shows an increase in mid-frequency energy and is within the influence window, the postural stability is slightly altered, the oxygenation decreases by no more than two percentage points and recovers within three long windows, the evidence level should be no lower than level three and the uncertainty no higher than level three, and the module should release data at a normal rhythm. If the motion amplitude remains low, the posture is stable, the oxygenation fluctuates only slightly and is at the end of less than one long window after the nursing event has ended, the evidence level drops to level two and the uncertainty rises to level four. The module then enters conservation mode and slows down the release rhythm, while simultaneously sending a supplementary sampling prompt to the access side. It automatically exits conservation mode after the evidence level recovers to at least level three for two consecutive long windows and the uncertainty decreases.

[0059] Through the aforementioned mechanisms of fixed standards, clear grading, explicit thresholds, evidence-based decisions, traceable flow control and retreat mechanisms, and verifiable sample expectations, this step clearly explains, remembers, and connects the three key issues: "whether the data is credible, where the risks come from, and how to tighten the gating boundaries." This provides a stable representational flow and an auditable evidentiary basis for subsequent grade calculations, prescription generation, and publication.

[0060] The graded prescription module is used to calculate rehabilitation feasibility scores under constraints and map activity grades to prescriptions. It is linked to the evidence coding module, using a prescription structure to carry grade, protective gear, monitoring frequency, and key points. Specifically, it is implemented as follows: After receiving the characterization, evidence quantity and uncertainty, as well as the safety budget and gating status, this module first performs integrity and timing verification: all inputs are uniformly calibrated by the hospital's clock service, with a time stamp error not exceeding three seconds; timestamps must not be earlier than the previous version; if gating is not ready, it enters a waiting queue, retrying at 30-second intervals, with a cumulative waiting time not exceeding five minutes. If the time limit is exceeded, a conservative prescription marked "gating not ready conservative output" is output and a gating mismatch event is reported. No new recommendations are issued during the waiting period. After verification, the module enters the scoring process: a physiological baseline is established using the moving median and interquartile range of the past 15 minutes; weights are assigned to posture, body motion spectrum, oxygenation, and laboratory backfill according to the evidence quantity, with elements with low evidence quantity or high uncertainty automatically reduced in weight; then, the safety budget and gating status are applied to the scoring, with items with insufficient budget forming upper limit constraints, and items approaching the threshold being continuously penalized and reduced, resulting in minute-level scores, with "elevation factors and suppressed factors" clearly indicated in the scoring description.

[0061] After the score is generated, a version-locked level lookup table is mapped to the executable level. This lookup table aligns item by item with commonly used scales in the department, including the minimum score, necessary preconditions, prohibition conditions, recommended monitoring frequency range, default protective equipment set, default execution points, and effective version information. Taking "sitting on the side of the bed" as the smallest example: the minimum score is 0.6, the preconditions are sedation level not lower than -2, mean arterial pressure not lower than 60 mmHg, inhaled oxygen concentration not higher than 60%, and positive end-expiratory pressure not higher than 10 cmH2O; the prohibition condition is a continuous increase in lactate from baseline of not less than 0.5 mmol / L for not less than 30 minutes; the recommended monitoring frequency is once to twice per hour (default is once per hour); the default protective equipment is a gait belt and bedside rails; the default execution points are two people assisting, first raising the head of the bed to 45 degrees, observing for two minutes, and then slowly sitting on the side.

[0062] Prescriptions are generated from a version-locked template library. Fields include a list of protective gear, upper and lower limits for monitoring frequency, execution points, and review points. When a patient is at high risk of falling or has unstable circulation, the template automatically adds support straps or increases the monitoring frequency. If the "level is too high and the budget is insufficient," the module generates a more conservative version according to the built-in downgrade path in the comparison table, and records the budget items and percentages hit in the prescription. Before execution, each review point is checked. If the heart rate increases by more than 20 beats per minute from the baseline, or the mean arterial pressure decreases by more than 10 mmHg from the baseline, or the blood oxygen saturation decreases by at least 3 percentage points from the baseline for at least 30 seconds, a downgrade is triggered. If any of the above indicators fluctuates within half of the threshold and lasts for less than 30 seconds, only "close monitoring" is indicated. If there is a rapid change in consciousness, or the patient complains of difficulty breathing and the ventilator trigger rate increases by at least 20% for at least 60 seconds, or similar equipment alarms accumulate at least three times within five minutes, the prescription is immediately stopped and reverted to the last verified safe prescription, and the reason and time are recorded.

[0063] All outputs are provided through a unified inference interface. Fields include score, grade, protective gear, monitoring frequency, execution points, and timestamp, all of which are mandatory. Missing any field returns "Required field missing," and reverse time order returns "Reverse time order" and a resend request is made. For the same patient, time slice, and grade, duplicates are removed using an idempotent key: patient identifier, time segment, and grade. Older versions are automatically invalidated. To ensure timeliness, both scoring and publishing have an end-to-end latency limit of five seconds. In high-concurrency scenarios, queuing and rate limiting strategies are employed. If a timeout occurs, a conservative branch is initiated, publishing only low-intensity prescriptions and review prompts along with the timeout reason. If the publishing channel is abnormal, a backup channel is switched first. If both channels are unavailable, the data is cached locally and resent according to a fixed strategy. If the resend limit is exceeded, an offline notification is generated for manual review.

[0064] Each prescription is accompanied by a chain of evidence. The minimum set of fields includes a timeline of scoring, release, and execution feedback times; a list of key factors sorted by contribution; hit gating items; budget occupancy percentage; downgrade or termination clauses; a comparison table and template version; generation node identifier; and signature and verification results. Prescriptions and evidence chains are cross-indexable, supporting playback of the "version used during generation." During the online and operational phases, weekly statistics are compiled on prescription executability pass rate, safety event rate, downgrade occurrence rate and cause distribution, and gating incompatibility trigger rate, with alternating weekly retests for comparison. When executability falls below 80% or the safety event rate exceeds 1%, the lowest grade score, monitoring frequency, and preconditions are adjusted first, followed by an evaluation of upstream evidence. All changes to comparison tables and templates are clearly stated with reasons and impact assessments. Old versions are retained in the evidence library for at least six months, and any playback displays the version at that time.

[0065] The entire process, from input verification to scoring and mapping, from formulating to release and record keeping, first establishes solid evidence, then strengthens constraints, and finally clearly describes the actions and releases them after authorization verification. If any key input is missing, threshold is not met, or link is abnormal, it will enter the explicit conservative branch and be downgraded or terminated according to the rules. At the same time, the reason, time and version are written into the evidence chain, which is convenient for bedside execution and also convenient for reproducibility and review under the same standard for post-event auditing and continuous improvement.

[0066] The simulated release module is used to generate counterfactual trajectories and risk curves under a safety budget, set release rhythms and degradation, connect the graded prescription module and the constraint budget module, output suggestions through a subscription release protocol, and use an evidence chain structure to carry the timing, factors, and degradation reasons before finalizing the data. The specific implementation is as follows: After receiving prescription candidates, security budgets and gating status, as well as the latest characterization, this module first performs a consistency check: channels with high evidence volume are used as the primary judgment, while channels with low evidence volume are only used as references, and the steady-state segments of the current time period and the previous time period are aligned to a unified time base; the calculation method of evidence volume is simultaneously determined as "the total evidence volume weighted by the channel credibility", with the weight derived from the credibility of the upstream anchor point and the health of the sensor, and the value is between zero and one. It is version-locked in the evidence library and written into the evidence chain with each batch.

[0067] Subsequently, without exceeding the budget limit, a short-term counterfactual analysis was conducted on the "Level 1 Activity Enhancement": using the most recent five-minute stable segment as a baseline, the activity load was progressively added to heart rate, mean arterial pressure, blood oxygen saturation, and respiratory trigger rate; if a validated sample of the same level of activity existed for the patient, the peak offset of that sample was prioritized; if no sample existed, the departmental baseline offset from the prescription database was used to generate three to five sampling points in one-minute increments, and the remaining margin to the budget boundary was calculated at each sampling point, marking the moment of minimum margin as the key monitoring point. The risk curve was formed according to the "margin mapping" and the segmentation criteria were clearly defined: a margin of not less than 30% of the budget limit was considered low risk, between 10% and 30% was considered medium risk, and less than 10% was considered high risk; the segment threshold number was written into the evidence chain with each version. The termination conditions are determined independently and include at least four categories: mean arterial pressure below the threshold, blood oxygen saturation dropping by more than the limit for two consecutive points, heart rate rising by more than the limit for two consecutive points, and respiratory trigger rate suddenly increasing. Once triggered, the current improvement is deemed infeasible, and the boundary variables, trigger time, and recovery suggestions are given.

[0068] Evidence volume and release rhythm are linked: when the evidence volume is not less than 60% and is consistent across multiple channels, it is released once per minute; when it is between 30% and 60%, it is released once every three minutes with a "pending review" label; when it is below 30%, the previously verified safe prescription is maintained, without proactive upgrades, and a prompt for supplementary collection is given; the above thresholds are locked in the evidence library by version for each batch, and the effective time and number are written in the evidence chain. The trigger thresholds for downgrading and rollback are clear: when the evidence volume falls into the low range for three consecutive minutes, or any point on the risk curve reaches a high value, or the termination condition is triggered, or the budget library hard threshold is updated, causing the current recommendation to no longer be approved, it is automatically downgraded to the previously verified safe prescription; if the new recommendation conflicts with the prescription in execution, it is rolled back to the safe version, and the source of the conflict, the rollback time, and subsequent actions are recorded in the evidence chain.

[0069] External announcements are made via a publish-subscribe approach. The main channel is the hospital's internal message bus, while the backup channel is a direct edge link. The message body includes the suggested text, activity level, protective gear and personnel configuration, risk curve points, a list of termination conditions, evidence chain number, and a second-level timestamp under a unified time base. The backup channel is switched upon the first timeout of the main channel. The sent buffer is resent in a first-in-first-out order while maintaining the timing. The main channel is automatically switched back upon recovery. Idempotency is used throughout the process to remove duplicates. The idempotency key consists of the patient number, suggestion number, and timestamp. If the receiving end receives duplicates within a five-minute window, only the message is updated and read, without repeated execution. The resend window is thirty minutes, with three resends at incremental intervals. Messages exceeding the window are marked as expired and reported as alarms.

[0070] To ensure hospital security and compliance, symmetric encryption and digest signing are performed before message transmission. The receiving end verifies the signature and checks whether the timestamp falls within the allowed window. Key rotation and time windows are uniformly managed by the security policy version, and the version number is written into the evidence chain along with the message. Collection and disk storage follow the minimum necessity principle, with identifiable information being anonymized in a hierarchical manner, and access permissions being hierarchically assigned to the department level. All recommendations, along with their corresponding risk curves, termination conditions, and explanations, are synchronously written into the prescription and evidence databases, with the index keys uniformly set as patient number, time period, and version. If writing to either database fails, a consistent reconciliation list is generated locally and rewritten to prevent prescriptions from becoming disconnected from evidence.

[0071] End-to-end latency shall not exceed 15 seconds; when the amount of evidence decreases, the release rhythm shall be automatically slowed down to 3 minutes; network jitter and channel timeout shall be addressed by local alarms, sequential resending, and backup channels to ensure order and reachability. Before going live, offline replay data shall be used to calculate the achievement rate, false release rate, and timeliness on a rolling weekly basis. After going live, weekly retesting shall be conducted and alternating weekly comparisons shall be adopted. Revised versions shall be solidified only if the false release rate is not higher than the previous version and the timeliness is not lower than the previous version; otherwise, they shall be automatically rolled back.

[0072] Taking a typical time period of standing intention as an example: Under the premise of high evidence and budget allowance, the module superimposes standing load on a five-minute baseline to obtain the following results: the mean arterial pressure at the second sampling point has the smallest margin and approaches the segment boundary, blood oxygen slightly decreases, and heart rate rises but remains within the allowable range within the next three minutes. The risk curve reaches medium risk at this point but does not cross the boundary. The module does not directly allow standing but downgrades the recommendation to sitting at the side of the bed, and clarifies the key monitoring points and termination conditions. At the same time, the event timeline, key factor contribution, budget occupancy ratio, segment threshold number, and downgrade reason are fully written into the evidence chain. Under the premise of not changing the above causal constraints and timing requirements, counterfactual inference can be achieved by short-term state model or empirical inference. The release channel can use the hospital message bus or edge local broadcast. Deployment can be completed at the bedside edge or nurse station centralized node, as long as the time delay, sequence, and traceability requirements are met. Overall, the input boundaries, processing procedures, threshold caliber, and traceability index are all clearly defined, and those skilled in the art can directly implement and audit playback based on this.

[0073] All calculations involved in the embodiments are dimensionless numerical calculations, and the preset parameters and thresholds in the calculations are set by those skilled in the art according to the actual situation.

[0074] It should be noted that this invention can be deployed on the device itself to realize embedded applications, or it can run on a PC or other terminal with a user interface, thereby meeting various hardware environments and usage requirements.

[0075] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wireless or wired transmission; wired transmission methods include optical fiber, twisted pair, coaxial cable, etc.; wireless transmission includes infrared, microwave, etc. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center containing one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0076] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and modules described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0077] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0078] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0079] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0080] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0081] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0082] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An artificial intelligence-based rehabilitation assessment system for critically ill patients, characterized in that, include: Time base anchor module: used to establish a unified time base and phase lock for nursing, body position and cardiopulmonary anchor points, connected to the acquisition component, and interacted with the timing and event interface, with the anchor structure carrying the time, phase and reliability; Event modeling module: used to generate impact time periods around anchor points, separate passive and autonomous events and deduct interference, connect to time-based anchor point module, and influence window structure bearing identifier, range, and weight; Constraint Budget Module: Used to analyze sedation, circulation, respiratory support and medication, generate safety budgets and gating, connect information systems and ventilators, and use budget structures to carry thresholds, penalties and gating states; Evidence coding module: used to encode attitude, body motion spectrum, oxygenation and laboratory backfill and generate uncertainty, connecting the event modeling module and the constraint budget module, so that the evidence structure carries the evidence quantity and uncertainty; Graded prescription module: used to calculate rehabilitation feasibility score under constraints and map activity level to prescription, connected to evidence coding module, with prescription structure to carry level, protective gear, monitoring frequency, key points; Simulated release module: Used to generate counterfactual trajectories and risk curves under a safety budget, set release rhythm and degradation, connect the grade prescription module and the constraint budget module, output suggestions through the subscription release agreement, and use the evidence chain structure to carry the time sequence, factors, degradation reasons, and put them on disk.

2. The artificial intelligence-based rehabilitation assessment system for critically ill patients according to claim 1, characterized in that, The time base anchor module includes: Anchor point classification units receive and write data according to nursing anchor points, body position anchor points, and cardiopulmonary rhythm anchor points, and each anchor point is assigned a unique key consisting of name, time, source, and credibility. The phase-locking unit performs biphase binding based on the ECG peak and the zero crossing of the respiratory phase within the search window. When a phase is missing, it performs timestamp alignment and records the reason. The conflict resolution unit determines the primary event based on the priority of nursing care over body position and body position over rhythm, and writes the allocation weight accordingly. The interpolation management unit performs downgraded interpolation and marks the data as non-interpolable when a measurement is missing. The interface verification unit verifies the signature and timestamp, rolls back late writes, and overwrites writes with correction flags. The time base unit outputs a normalized timing sequence and anchor point table and writes it into the time base library.

3. The artificial intelligence-based rehabilitation assessment system for critically ill patients according to claim 1, characterized in that, The event modeling module includes: Under a unified time base, the event modeling module establishes front and back windows with nursing anchors, body position anchors, and cardiopulmonary rhythm anchors as zero points; Nursing markers, bed corner and posture transitions, ventilator trigger ratio changes, and cardiopulmonary phase continuity were retrieved sequentially, and passive and spontaneous activities were determined according to a predetermined priority order. The passive strength is output in parallel by calculating and subtracting weights from the body motion features within the passive window based on anchor point type and anchor point confidence. When adjacent windows overlap, the windows are merged while retaining the source list and coverage ratio. Conflicts are re-adjudicated based on nursing priority, body position, and rhythm correction. When both passive and autonomous evidence exist, output mixed annotations and prevent the generation of upgraded prescriptions in this window; A list of decontamination characteristics and impact windows is generated for downstream use.

4. The artificial intelligence-based rehabilitation assessment system for critically ill patients according to claim 3, characterized in that: The input interface for the event modeling module includes time period, anchor point identifier, range, and weight fields, and carries a signature and timestamp. The window number is composed of the patient identifier, anchor time, and an increasing sequence and is unique. The idempotent key is composed of the window number and the timestamp to ensure deduplication and order. The output provides a list of decontamination features and impact windows, and records the window number, source list, coverage ratio, judgment type, deduction ratio, passive intensity, boundary time, processing result, operator, and timestamp in the evidence chain; When the motion channel is missing, only passive annotations and window boundaries are published to maintain downstream recognition. Recording and publishing are both completed under a unified time base and permission verification.

5. The artificial intelligence-based rehabilitation assessment system for critically ill patients according to claim 1, characterized in that, The budget constraint module includes: The constrained budget module receives sedation level, circulatory parameters, ventilator parameters, and medication records through a parameter subscription protocol. The messages include name, value, unit, and timestamp. Based on the calibration table corresponding to the device number and version number, the threshold is verified, the restriction is converted into a single amount and the minimum value is taken to generate a comprehensive budget. When the hysteresis time window is stably crossed, the gating status is updated. The gating status is allowed, restricted, or prohibited. The upstream credibility is reduced in stages. In case of conflict, the order of priority is: circulation, then respiration, then sedation, and finally medication. The gating release includes the triggering basis, the direction of the delay, the evidence chain number, and the timestamp.

6. The artificial intelligence-based rehabilitation assessment system for critically ill patients according to claim 1, characterized in that, The evidence encoding module includes: The evidence encoding module receives decontamination characteristics, safety budget and gating status, oxygenation and laboratory backfilling, aligns and unifies the time base and completes integrity and timing verification, executes interpolation and elimination rules based on upstream credibility stripes and records them into the evidence chain; The representation and evidence are generated by encoding attitude, body motion spectrum, and oxygenation together. The representation vector is output in the order of fields: attitude stability, body motion energy, oxygenation fluctuation, and gated state summary. At the same time point, the amount of evidence and uncertainty are propagated and synthesized step by step. When the channel conclusions are mutually exclusive and not within the nursing event window, the decision is made sequentially based on posture, oxygenation, and body movement. When the uncertainty reaches the threshold, the conservation strategy is adopted. The characteristic bus outputs the characterization, evidence quantity, uncertainty, and timestamp.

7. The artificial intelligence-based rehabilitation assessment system for critically ill patients according to claim 1, characterized in that, The graded prescription module includes: The graded prescription module performs integrity and timing checks on characterization, amount of evidence, uncertainty, safety budget, and gating status. If the gate is not ready, it enters the waiting queue; if the gate is full, the conservative prescription will be output. Scores are generated based on physiological baselines and weighted by the amount of evidence and uncertainty. Safety budget and gating status are applied to the scores as upper limit constraints and continuous penalties, respectively. The activity level is mapped according to the version-locked level mapping table, and a prescription is generated from the version-locked prescription template library. The prescription includes protective gear, monitoring frequency, execution points and review points. When the level is higher than the security budget, the pre-set degradation path will be used. The output includes a chain of evidence, which contains a timeline, key factors, gating entries, and version identifiers.

8. The artificial intelligence-based rehabilitation assessment system for critically ill patients according to claim 1, characterized in that, The simulated release module includes: Consistency verification unit, counterfactual solution unit, risk mapping unit, and downgrade rollback unit; The consistency verification unit aligns the multi-channel representations under a unified time base, and generates the total amount of evidence based on the channel confidence and sensor health, and writes it into the evidence chain. The counterfactual calculation unit uses the steady-state segment as a baseline, and adds the activity load to heart rate, mean arterial pressure, blood oxygen saturation and respiratory trigger rate one by one, and calculates the remaining margin of the relative safety budget. The risk mapping unit forms a risk curve based on the margin segmentation; The downgrade and rollback unit generates recommendations and determines downgrade and rollback based on the risk curve, termination conditions, and safety budget assessment results.

9. The artificial intelligence-based rehabilitation assessment system for critically ill patients according to claim 1, characterized in that: The simulated release module also includes a release control unit and a log auditing unit; The release control unit switches channels between the hospital message bus and the edge direct link through the publish-subscribe protocol. The send buffer is resent in first-in-first-out order, and deduplication is performed by forming an idempotent key with patient ID, suggestion number and timestamp. At the same time, the release rhythm is adjusted according to the total amount of evidence and the verified safe prescription is maintained in the case of insufficient evidence. The audit unit writes recommendations, activity levels, protective gear and staffing configurations, risk curve points, termination condition lists, evidence chain numbers and unified time base timestamps to both the prescription database and the evidence database. It uses patient identification, time period and version as index keys, and generates a reconciliation list and completes the rewriting and cancellation when an anomaly occurs.

Images