An artificial intelligence-based remote control analgesic pump system

Abstract

The present application relates to the technical field of artificial intelligence medical treatment and remote pain management, in particular to a remote control type analgesia pump system based on artificial intelligence, which comprises a multi-modal perception unit, a medical order model analysis unit, a subjective-physiological inconsistency evaluation unit, a medical order compliance monitoring unit, a dynamic alignment decision unit, a hierarchical safety constraint unit and an analgesic effect evaluation unit; the system synchronously collects physiological index data and subjective feedback data, and the core thereof is to calculate a subjective-physiological inconsistency index and a medical order compliance drift rate; the dynamic alignment decision unit weighs the above indexes and corrects them under hierarchical safety constraints to generate a final drug administration instruction; the present application realizes a change from masking or averaging data conflicts to actively quantifying subjective-physiological disconnection, can identify and solve data conflict problems, and provides a core basis for dynamic decision-making.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence medical and remote pain management technology, specifically to a remotely controlled analgesia pump system based on artificial intelligence. Background Technology

[0002] In the current field of remote analgesia, the operation of analgesia pumps is highly dependent on rigid medical orders set by physicians, while simultaneously responding to sparse and subjective analgesia requests from patients. Existing technologies often employ weighted averages or simple logic gates when integrating multimodal data. This approach masks inconsistencies between objective physiological indicators and patients' subjective experiences, leading to biased decision-making. This mechanism creates a core contradiction between the rigidity of medical orders and the fluctuations in individualized patient experiences. Furthermore, AI decision-making is often separated from safety constraints, lacking integrated safety guarantees. In addition, systems generally lack closed-loop evaluation of the true effectiveness of medication decisions, failing to form a learning loop. Therefore, how to construct a dynamic decision-making system that operates within a pre-set safety framework, strictly adhering to long-term medical orders while flexibly aligning with real-time patient experiences, and proactively quantifying and resolving inconsistencies between subjective and physiological data, is a pressing technical problem that needs to be solved in this field. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention provides a remotely controlled analgesia pump system based on artificial intelligence. Specifically, the technical solution of this invention includes: The multimodal sensing unit is used to simultaneously collect physiological indicator data, subjective feedback data, and real-time safety status data to generate physiological state vectors, subjective demand signals, and real-time safety status vectors. The medical order model parsing unit is used to receive the treatment parameters set by the physician and parse them into long-term treatment target benchmark signals; A subjective-physiological inconsistency assessment unit is used to receive the physiological state vector and the subjective demand signal, and generate a subjective-physiological inconsistency index. The medication adherence monitoring unit is used to receive long-term treatment target baseline signals and actual cumulative dosing records from the stratified safety constraint unit, and to calculate the medication adherence drift rate. A dynamic alignment decision unit is used to generate suggested dosing decisions based on a trade-off between the subjective-physiological inconsistency index and the medical order compliance drift rate. A hierarchical safety constraint unit is used to modify the suggested dosing decision based on preset safety constraints, and generate the final dosing instruction and safety boundary activation signal to be executed. An analgesic effect assessment unit is used to monitor the analgesic effect after drug administration and generate alignment effect feedback signals.

[0004] Preferably, the physiological index data are continuous high-frequency physiological index data, including heart rate variability (HRV) and skin conductance activity (EDA); the subjective feedback data are sparse low-frequency subjective feedback data, including patient-pressed PCA button requests and electronic pain score (VAS).

[0005] Preferably, the process by which the subjective-physiological inconsistency assessment unit generates the subjective-physiological inconsistency index is as follows: Based on the physiological state vector, the objective level of analgesia requirement is inferred; Based on the aforementioned subjective demand signals, the level of subjective analgesia demand is quantified; The subjective-physiological inconsistency index is generated by calculating the difference or distance between the objective analgesia demand level and the subjective analgesia demand level.

[0006] Preferably, the inference process of the objective analgesia requirement level is as follows: by using a built-in pre-trained deep learning model, the time series data of the physiological state vector is received, and the continuous objective analgesia requirement level is output in real time. The quantification process of the subjective analgesia demand level is as follows: within a preset time window, the sparse subjective demand signals are accumulated or weighted averaged and mapped to the subjective analgesia demand level. The process of generating the subjective-physiological inconsistency index is as follows: calculate the difference between two demand levels over time, which is obtained by weighted difference, dynamic time warping (DTW) distance, or Wasserstein distance.

[0007] Preferably, the dynamic alignment decision unit execution experience-medical order bidirectional coupling control mechanism operates as follows: When the subjective-physiological inconsistency index is high and the medical order compliance drift rate is lower than the drift budget threshold, the subjective-physiological inconsistency index is minimized first to generate an active intervention medication decision. When the subjective-physiological inconsistency index is low, or the medication adherence drift rate is high, the medication adherence drift rate is minimized first to generate a conservative regression dosing decision.

[0008] Preferably, the process by which the analgesic effect evaluation unit generates the alignment effect feedback signal is as follows: When the patient presses the PCA button, the system administers the medication. Start a timing window based on the onset time of the analgesic used; If, within this window, the patient's subjective-physiological inconsistency index or self-reported pain score fails to fall below the pain relief threshold; This request is then marked as an invalid PCA request, and the alignment effect feedback signal is generated based on the invalid PCA request rate.

[0009] Preferably, the dynamic alignment decision unit is implemented as a reinforcement learning agent, whose state input is the subjective-physiological inconsistency index and the medical order compliance drift rate, and whose action is the suggested dosing decision.

[0010] Preferably, the reward function of the reinforcement learning agent is configured as follows: A positive reward is given when the subjective-physiological inconsistency index decreases; When the medication adherence drift rate approaches the drift budget threshold, a negative penalty is applied. When the security boundary activation signal occurs, a strong negative penalty is imposed; Furthermore, the alignment effect feedback signal from the analgesic effect evaluation unit is used to automatically adjust the reward function of the reinforcement learning agent.

[0011] Compared with the prior art, the present invention has the following beneficial effects: 1. Unlike existing technologies that mask or average the conflict between objective physiological data and patients' subjective feelings, this system actively and quantitatively reveals the degree of this disconnect by creating a subjective-physiological inconsistency index. This enables the system to identify and resolve data conflict issues, providing a core basis for subsequent dynamic decision-making. 2. This system transforms rigid long-term medical orders set by physicians into flexible budgets by introducing the medical order adherence drift rate and drift budget threshold. The system uses this budget to intelligently and flexibly respond to the patient's real-time individualized pain fluctuations without violating the long-term treatment goals, thus resolving the core contradiction between the rigidity of medical orders and the fluctuation of patient experience. 3. This system uses a hierarchical safety constraint framework as the basic environment for system operation, rather than an external module independent of artificial intelligence decision-making; all dosing decisions recommended by the artificial intelligence unit must be modified by the safety unit with the highest veto power, ensuring that the final executed instructions are always within the absolute clinical safety boundary, thus solving the hidden danger of the separation of AI decision-making and safety constraints; 4. This system has innovatively constructed a slow learning loop based on the rate of invalid PCA requests through an analgesic effect assessment unit. This loop assesses not whether the decision was executed, but whether the decision truly and effectively relieved the pain. This feedback based on real effects enables the system to evolve from an executor into a learner capable of self-iteration and deep adaptation to individual patient differences. Attached Figure Description

[0012] The present invention will be further explained below with reference to the accompanying drawings and embodiments: Figure 1 This is a structural diagram of the system of the present invention. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0014] Example 1 Please see Figure 1 A remotely controlled analgesia pump system based on artificial intelligence, configured to operate within a pre-defined hierarchical safety constraint framework, includes: The multimodal sensing unit is used to simultaneously collect physiological indicator data, subjective feedback data, and real-time safety status data to generate physiological state vectors, subjective demand signals, and real-time safety status vectors. The medical order model parsing unit is used to receive the treatment parameters set by the physician and parse them into long-term treatment target benchmark signals; The subjective-physiological inconsistency assessment unit is used to receive physiological state vectors and subjective demand signals, and generate a subjective-physiological inconsistency index. The medication adherence monitoring unit is used to receive long-term treatment target baseline signals and actual cumulative dosing records from the stratified safety constraint unit, and to calculate the medication adherence drift rate. Dynamic alignment decision units are used to generate recommended dosing decisions based on a trade-off between the subjective-physiological inconsistency index and the medication adherence drift rate. The hierarchical safety constraint unit is used to modify the suggested dosing decision based on preset safety constraints and generate the final dosing instruction and safety boundary activation signal. An analgesic effect assessment unit is used to monitor the analgesic effect after drug administration and generate alignment effect feedback signals.

[0015] Physiological data are continuous, high-frequency data, including heart rate variability (HRV) and electrical activity of the skin (EDA); subjective feedback data are sparse, low-frequency data, including patient-pressed PCA button requests and electronic pain score (VAS).

[0016] This invention provides a specific implementation of an artificial intelligence-based remotely controlled analgesia pump system, which operates within a preset hierarchical safety constraint framework. The significance of this framework lies in the fact that it provides an insurmountable, highest-priority boundary for subsequent artificial intelligence decision-making, ensuring the safety of the patient's life. Unlike the existing technology where AI decision-making and safety boundaries are separated, this invention regards safety constraints as the basic environment for system operation, ensuring that the dynamic alignment of AI always occurs within clinically permissible limits. In this embodiment, the multimodal sensing unit is designed to comprehensively and in real-time capture the patient's individualized analgesia experience to address the biased decision-making caused by a single data source. This unit, implemented through wearable sensors and a patient interface, is configured to simultaneously collect two distinct types of data: continuous high-frequency physiological indicators, specifically heart rate variability (HRV) and electrical skin conductance (EDA) in this embodiment, which objectively reflect the patient's autonomic nervous system activity and stress levels; and sparse low-frequency subjective feedback data, specifically the patient's PCA button requests and electronic pain score (VAS), which directly reflect the patient's subjective feelings. After collection, the unit does not directly output the raw data but performs necessary preprocessing: for example, extracting time-domain and frequency-domain features from the HRV signal, and extracting skin conductance level (SCL) and skin conductance response (SCR) features from the EDA signal. These extracted, structured features are combined into a multidimensional physiological state vector, while the PCA request and VAS score are encoded as subjective demand signals. Both are then synchronously output to the subsequent subjective-physiological inconsistency assessment unit. The medical order model parsing unit aims to transform rigid clinical instructions set by physicians into long-term control goals that AI can understand and execute. This unit receives treatment parameters set by remote physicians, such as 24-hour total dose limits, basal flow rate, PCA single dose, and minimum lockout time. Its core function is to use these discrete parameters to generate a continuous long-term analgesia treatment model through a preset interpolation algorithm. This model can be represented by an ideal cumulative dosing trend curve that is smooth and represents the physician's expectations over a 24-hour period. The unit ultimately outputs a long-term treatment goal benchmark signal, which is transmitted to the medical order compliance monitoring unit as the gold standard for subsequent judgment on whether AI decisions deviate from the medical order. The Subjective-Physiological Inconsistency Assessment Unit is designed to address one of the core technical problems of this invention: quantifying the degree of discrepancy between objective physiological inferences and the patient's subjective experience. This is a key innovation that distinguishes this invention from existing technologies. Existing technologies often use weighted averages or OR gate logic when fusing multimodal data, which masks conflicts between data. This invention takes the opposite approach, actively calculating and quantifying this conflict and using it as a core input for subsequent decision-making. This unit receives physiological state vectors and subjective demand signals, and generates a quantitative index, namely the Subjective-Physiological Inconsistency Index (SPDI), through internal processing. The SPDI is a quantitative indicator used to characterize the degree of discrepancy between AI inferences from objective data and the patient's subjective experience. A high SPDI value indicates a serious deviation between AI inferences and patient experience, suggesting that the system needs intervention. This index is output to the dynamic alignment decision unit and the analgesic effect assessment unit. The purpose of the medication adherence monitoring unit is to quantify the degree to which the AI's immediate decisions deviate from the doctor's core intentions in the long term. This unit receives a long-term treatment target baseline signal and actual cumulative dosing records. The unit continuously compares the actual cumulative dosage with the target dosage expected by the long-term treatment target baseline signal at the same point in time. It calculates this cumulative deviation as a percentage and defines it as the medication adherence drift rate. This medication adherence drift rate is a quantitative indicator used to characterize the deviation of the long-term medication budget spent by the AI ​​decision in pursuing real-time comfort. Its value is obtained by calculating the difference between the actual cumulative dosage and the expected dosage of the baseline signal, and then dividing it by the expected dosage of the baseline signal to obtain a relative percentage. This drift rate is output as a real-time status signal to the dynamic alignment decision unit. The dynamic alignment decision unit is the core control unit of this invention for achieving dynamic alignment of medical orders and experiences. The innovative information flow of this invention converges here: the input received by this unit is not a single button signal from a traditional PCA pump, but two highly condensed state variables unique to this invention: the Subjective-Physiological Inconsistency Index (SPDI) and the Medical Order Compliance Drift Rate. Its core function is to dynamically weigh these two optimization objectives: that is, to find the optimal balance between minimizing SPDI and minimizing the drift rate. This unit ultimately outputs a suggested dosing decision, such as increasing the flow rate by 20%, maintaining, or decreasing the flow rate by 10%, and transmits it to the stratified safety constraint unit. The hierarchical safety constraint unit receives real-time safety status vectors and corrects suggested dosing decisions based on preset safety constraints. When the real-time safety status vectors reach preset physiological hazard thresholds, this unit immediately rejects all dosing decisions to generate a final dosing instruction and a safety boundary activation signal. It is the specific executor of the hierarchical safety constraint framework preset in this invention, aiming to ensure that all AI decisions are executed within the clinically absolutely safe boundary and has the highest veto power. This unit receives suggested dosing decisions and internally stores hard constraints set by physicians based on clinical safety standards. Its judgment logic is: if the suggested dosing decision does not reach the safety red line, it is directly approved; if the decision reaches the red line, the unit rejects the suggestion and corrects it to the maximum permissible decision within the safety boundary. This unit ultimately outputs two signals: the final dosing instruction, which is sent to the analgesia pump hardware for execution and its dosing record is synchronously sent to the medical order compliance monitoring unit to update the drift rate; and the safety boundary activation signal, which is generated only when the red line is reached and sent to the dynamic alignment decision unit. An analgesic effect assessment unit receives the subjective-physiological inconsistency index from the subjective-physiological inconsistency assessment unit and subjective feedback data from the multimodal perception unit, monitors the analgesic effect after drug administration, and generates an alignment effect feedback signal. Its purpose is to form a longer-term, slow-learning closed loop; this is another major innovation of this invention. Existing feedback loops typically only go as far as whether drug administration is required, while the closed loop of this invention extends to whether drug administration is effective. This unit assesses whether the decision of the dynamic alignment decision unit truly resolves the patient's pain. The unit monitors the analgesic effect after drug administration and ultimately generates an alignment effect feedback signal. The generation logic of this signal is based on the calculation of the invalid PCA request rate, used to measure the actual effect of the AI ​​decision. This feedback signal is transmitted back to the dynamic alignment decision unit to adjust its internal decision model. Through the collaborative work of the aforementioned units, this embodiment constructs a complete and self-consistent perception-evaluation-decision-execution-feedback closed-loop system. Compared with existing technologies, this invention has essential differences and non-obvious advancements: Creative information flow: This invention abandons the traditional linear logic of button-administration and creatively constructs a bivariate decision-making model based on the SPDI inconsistency index and the order drift rate; Dual-loop adaptive design: This invention constructs two closed loops: one is a fast decision-making closed loop driven by SPDI and the drift rate, used to weigh the experience and the order in real time; the other is a slow learning closed loop driven by the invalid PCA request rate, used to evaluate the effectiveness of the fast closed loop itself and adjust it; this nested dual-loop structure allows the system to evolve from an executor to a learner; Resolves the core contradiction: By creating two core quantitative indicators, SPDI and the order compliance drift rate, this invention enables AI to flexibly and safely dynamically align the patient's individualized real-time experience within a rigid order framework, resolving the core contradiction between the rigidity of orders and the fluctuation of experience, as well as the problem of the disconnect between subjective feelings and objective data.

[0017] Example 2: The process by which the Subjective-Physiological Inconsistency Assessment Unit generates the Subjective-Physiological Inconsistency Index is as follows: Based on the physiological state vector, the objective level of analgesia demand is inferred; Quantify the level of subjective analgesia demand based on subjective demand signals; The subjective-physiological inconsistency index is generated by calculating the difference or distance between the objective analgesia demand level and the subjective analgesia demand level.

[0018] The process of inferring the objective analgesia requirement level is as follows: the built-in pre-trained deep learning model receives time-series data of physiological state vectors and outputs continuous objective analgesia requirement levels in real time. The process of quantifying the subjective analgesia demand level is as follows: within a preset time window, sparse subjective demand signals are accumulated or weighted averaged and mapped to the subjective analgesia demand level. The process of generating the subjective-physiological inconsistency index is as follows: calculate the difference between two demand levels over time, which is obtained by weighted difference, dynamic time warping (DTW) distance, or Wasserstein distance.

[0019] To further clarify how the Subjective-Physiological Inconsistency Assessment Unit generates the Subjective-Physiological Inconsistency Index (SPDI), this embodiment provides a specific method that combines deep learning and time window processing to ensure that those skilled in the art can implement it. The inference process regarding the objective analgesia requirement level: The unit's built-in pre-trained deep learning model, such as a model based on a Temporal Convolutional Network (TCN) or a Long Short-Term Memory (LSTM) network, is chosen because it can effectively capture the complex long-term temporal dependencies in physiological signals. This pre-trained deep learning model refers to a model that has been trained on a large, diverse clinical dataset. This training dataset contains a massive amount of pain level labels synchronized with physiological state vectors and annotated by clinicians. The model is configured to receive time-series data of physiological state vectors and output continuous objective analgesia requirement levels in real time, such as a continuous value mapped to the range of 0-10, which represents the degree of patient pain inferred by the AI ​​from objective physiological data. Regarding the quantification process of subjective analgesia demand level: Since subjective demand signals are sparse, low-frequency data, which differ in scale from high-frequency physiological vectors, this invention employs a processing method within a preset time window to achieve scale alignment. The duration of this preset time window is a key configurable parameter. Its specific value is determined based on a comprehensive consideration of the clinical pain outbreak cycle and drug onset time, aiming to achieve a technical trade-off: the time window must be short enough to quickly respond to the patient's acute pain changes; at the same time, it must be long enough to smooth out meaningless, noisy PCA button presses caused by accidental touches or emotional fluctuations. Within this window, the system accumulates or performs weighted averaging on the sparse subjective demand signals. One specific processing method is: assigning a base score to each PCA button request within the window, directly using the value of each VAS score, and performing a time-decayed weighted average on all scores, ultimately normalizing the result to the same 0-10 range as the objective demand level, thereby quantifying it as the subjective analgesia demand level. Regarding the generation process of the Subjective-Physiological Inconsistency Index (SPDI): This unit receives two demand levels, both within the range of 0-10, and calculates the difference between the two demand levels over time. This difference is a distance metric used to characterize the degree of disconnect between inference and perception. This embodiment provides several calculation methods, which can be selected by those skilled in the art based on computing resources and required accuracy: Weighted difference: Calculates the absolute value of the difference between the two at the same time, which has the advantage of high computational efficiency; Dynamic Time Warping (DTW) distance: When considering the possible time delay of several minutes between objective physiological response and subjective perception, DTW is a superior choice because it can calculate the morphological similarity of the two time series without being affected by slight time axis misalignment; Wasserstein distance: When it is necessary to compare the distribution of two demand levels within a time window rather than just their numerical values, Wasserstein distance can be used; The calculated difference value is the final output Subjective-Physiological Inconsistency Index (SPDI). Through the above implementation methods, this invention concretizes the abstract inconsistency assessment in Example 1, ensuring feasibility; it uses a deep learning model to solve the problem of extracting objective pain states from high-frequency, nonlinear physiological data; it solves the problem of fusion of high-frequency objective data and low-frequency subjective data in terms of time scale and data sparsity by using time window weighting processing; it makes the calculation of SPDI more robust by introducing an advanced distance metric, which can accurately quantify the degree of disconnect between inference and perception, providing a reliable and quantifiable input basis for subsequent dynamic alignment decisions.

[0020] Example 3: The dynamic alignment decision-making unit execution experience-medical order bidirectional coupling control mechanism works as follows: When the subjective-physiological inconsistency index is high and the medication adherence drift rate is lower than the drift budget threshold, the subjective-physiological inconsistency index is minimized first to generate proactive intervention medication decisions. When the subjective-physiological inconsistency index is low or the medication adherence drift rate is high, the medication adherence drift rate should be minimized first to generate a conservative regression dosing decision.

[0021] This embodiment details the core algorithm executed within the dynamic alignment decision unit, namely the experience-medical order bidirectional coupling control mechanism. The purpose of this mechanism is to solve the trade-off problem in Embodiment 1, namely, to make intelligent decisions between the two conflicting goals of aligning patient experience and adhering to doctor's orders. This unit continuously receives two status inputs: the Subjective-Physiological Inconsistency Index (SPDI) and the Doctor's Order Compliance Drift Rate; its internal control logic is a condition-based dynamic priority switching mechanism. The operation of this mechanism depends on two key, configurable thresholds: SPDI High Threshold: This threshold is used to determine whether the patient's experience-inference level has reached a clinically unacceptable level; this value is not set arbitrarily, but is determined by statistical analysis of historical data to find the SPDI threshold that is strongly correlated with high rates of ineffective PCA requests or patient-reported dissatisfaction with analgesia. Drift budget threshold: This drift budget threshold is a key control parameter that represents the flexibility or budget that the physician allows the AI ​​to deviate from long-term medical orders for real-time comfort. This threshold is not a fixed single value; in a preferred implementation, it is a dynamic function that can be preset by the physician according to the patient's treatment stage and tolerance. Based on the aforementioned threshold, the coupling logic of this unit is triggered: In one scenario, when the subjective-physiological inconsistency index is high and the medication adherence drift rate is below the drift budget threshold, the system determines that the budget is sufficient and the patient's needs are urgent. The decision unit is configured to prioritize minimizing the subjective-physiological inconsistency index. At this point, it generates an active intervention medication decision, which temporarily increases the drift rate in order to quickly respond to the patient's real-time needs. Another scenario is when the subjective-physiological inconsistency index is low, or the medication adherence drift rate is high. Under these conditions, the system determines that the patient's needs are not urgent or that medication adherence is more critical, and the decision unit is configured to prioritize minimizing the medication adherence drift rate. In this case, it generates a conservative regression dosing decision, which aims to bring the system's cumulative dosing behavior back to the long-term treatment target benchmark, thereby saving budget for possible future SPDI fluctuations. This unit will output the generated proactive intervention dosing decisions or conservative regression dosing decisions as a recommended dosing decision to the hierarchical safety constraint unit. The experience-medical order bidirectional coupling control mechanism disclosed in this embodiment is the core innovation of this invention in achieving dynamic balance between medical orders and experience. By introducing the concept of drift budget threshold, it softens rigid medical orders into a manageable elastic constraint. This makes the system no longer a rigid executor of medical orders, but a dynamic interpreter of medical orders. It can intelligently allocate budget to cope with the patient's immediate pain fluctuations while ensuring long-term medical order compliance, thereby maximizing the patient's real-time comfort without deviating from the doctor's long-term treatment intentions.

[0022] Example 4: The process by which the analgesic efficacy assessment unit generates the alignment effect feedback signal is as follows: When the patient presses the PCA button, the system administers the medication. Start a timing window based on the onset time of the analgesic used; If, within this window, the patient's subjective-physiological inconsistency index or self-reported pain score fails to fall below the pain relief threshold; This request will then be marked as an invalid PCA request, and an alignment effect feedback signal will be generated based on the invalid PCA request rate.

[0023] This embodiment illustrates how the analgesic effect assessment unit generates alignment effect feedback signals; the purpose of this unit is to establish a slow learning loop, and its core innovation lies in assessing the effectiveness of decisions, rather than the execution of decisions. The processing logic of this unit is as follows: The unit detected an event: when the patient pressed the PCA button, the system administered medication; The system immediately initiates a timing window; the duration of this window is not fixed but is set based on the known pharmacological characteristic of the onset time of the analgesic drug used; this is crucial to ensuring the effectiveness of the assessment; for example, for intravenous morphine, the peak analgesic onset time is approximately 15–20 minutes, so the window can be set to 20 minutes; for fentanyl, the window may only need 5–10 minutes; the system is configured to automatically query and set this window duration from a pharmacological database based on the type of drug loaded in the pump; During the duration of this timing window, the unit continuously monitors changes in the patient's condition; At the end of the window, the system makes a determination: if the patient's Subjective-Physiological Inconsistency Index (SPDI) or self-reported pain score fails to drop below a preset pain relief threshold within the window; this pain relief threshold is a configurable standard that characterizes clinically effective relief; it can be defined as an absolute value or a relative value; those skilled in the art can flexibly set this standard according to clinical guidelines. If the standard is not met, the unit will mark this request as an invalid PCA request; This unit calculates the invalid PCA request rate cumulatively over a longer timescale and generates an alignment effect feedback signal based on this invalid PCA request rate. This alignment effect feedback signal is a meta-feedback signal characterizing the long-term effectiveness of AI decision-making. If the signal shows that the invalid PCA request rate is too high, it indicates that the internal decision-making logic of the current dynamic alignment decision unit is failing for that particular patient. This signal is fed back to the dynamic alignment decision unit to trigger adjustments to its internal model. Through the invalid PCA request evaluation mechanism disclosed in this embodiment, the present invention achieves a quantitative evaluation of the true analgesic effect; it constructs a slow learning closed loop based on effect rather than execution; this enables the system to no longer blindly pursue the instantaneous balance of SPDI and drift rate, but to calibrate its balance strategy based on the long-term actual effect of the decision; for example, when the invalid request rate is found to be high, the system will realize that it needs to adjust its response weight to SPDI, thereby achieving deep adaptive learning for individual differences among different patients.

[0024] Example 5: The dynamic alignment decision unit is implemented as a reinforcement learning (RL) agent, whose state inputs are the subjective-physiological inconsistency index and the medication adherence drift rate, and whose action is the suggested dosing decision.

[0025] The reward function for the reinforcement learning agent is configured as follows: When the subjective-physiological inconsistency index decreases, a positive reward is given. When the medication adherence drift rate approaches the drift budget threshold, a negative penalty is imposed. When a safety boundary activation signal appears, a strong negative penalty is imposed; Furthermore, the alignment effect feedback signal from the analgesic effect assessment unit is used to automatically adjust the reward function of the reinforcement learning agent.

[0026] This embodiment provides an advanced implementation of the dynamic alignment decision unit, which is a specific technical means of the experience-medical order bidirectional coupling control mechanism, ensuring the adaptability of the mechanism; In this embodiment, the dynamic alignment decision unit is implemented as a reinforcement learning agent; this reinforcement learning agent is an artificial intelligence model that learns how to choose actions to maximize long-term cumulative rewards through interaction with the environment and trial and error. The State input of the RL agent is precisely defined as: the Subjective-Physiological Inconsistency Index (SPDI) from the Subjective-Physiological Inconsistency Assessment Unit and the Medical Order Compliance Drift Rate from the Medical Order Compliance Monitoring Unit; The action space of the RL agent is defined as the proposed dosing decisions, such as a discrete set of actions including increasing the flow rate by 20%, maintaining, decreasing the flow rate by 10%, approving PCA, and rejecting PCA. The core of this RL agent, namely its reward function RewardFunction, is carefully designed to embody the multi-objective optimization concept of this invention; the reward function is configured as follows: Alignment Experience: When the Subjective-Physiological Inconsistency Index (SPDI) decreases, i.e. when the agent's actions successfully realign the AI's inference with the patient's feelings, the system provides a positive reward. Adherence to medical orders: When the medical order adherence drift rate approaches the drift budget threshold, i.e., when the agent's actions are exhausting the medical order budget, the system imposes a negative penalty; Ensuring safety: When a safety boundary activation signal appears, i.e. when the agent's actions attempt to cross the clinical safety red line, the system imposes a strong negative penalty; the weight of this penalty is set much higher than the previous two to ensure that the RL agent quickly learns during exploration that safety is an inviolable red line; The above reward / penalty values ​​are merely examples, and their relative magnitudes are key to the design of this invention: safety penalty >> medical order penalty > experience reward; Furthermore, this embodiment implements a meta-learning or slow learning closed loop, which is a non-obvious innovation that distinguishes this invention from traditional RL applications; the alignment effect feedback signal from the analgesic effect evaluation unit is used to automatically adjust the reward function of the reinforcement learning agent; This alignment effect feedback signal serves as a meta-feedback signal. When it shows a persistently high rate of invalid PCA requests, it indicates that the current reward function configuration is ineffective for the patient. At this point, the system is configured to automatically adjust the internal weights of the reward function. One feasible adjustment method is for the system to execute a preset meta-policy adjustment rule. For example, when the invalid PCA request rate is >30%, the system determines that the current strategy is insufficient in responding to the patient's pain, and automatically increases the positive reward weight of SPDI reduction by 10%, while / or decreases the negative penalty weight of drift rate by 10%. This forces the RL agent to relearn a new strategy that is more aggressive and more inclined to respond to SPDI. Conversely, if the drift rate is consistently too high, the opposite adjustment can be made. This adjustment is executed automatically, enabling the AI ​​to learn how to learn to adapt to individual differences in pharmacodynamics among different patients. By implementing the dynamically aligned decision unit as a reinforcement learning agent and supplementing it with an innovative self-adjusting reward function mechanism, this invention provides a powerful and adaptive implementation path for solving the dual-objective optimization problem of medical orders and patient experience. The design of the reward function transforms the abstract clinical objective into a computable and optimizable multi-objective function. The strong penalty of the safety boundary ensures that the RL agent can learn to actively operate within the safety margin, rather than being passively rejected by the safety boundary, thus achieving proactive safety margin management. The automatic adjustment mechanism of the reward function is a major innovation of this invention, which enables the system to evolve from a policy executor to a policy learner. The system can reflect on and correct its underlying value judgments based on the actual effects of the decisions, thereby achieving deep, dynamic, and truly individualized adaptation to the differences among individual patients.

[0027] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A remotely controlled analgesia pump system based on artificial intelligence, characterized in that, It is configured to operate within a predefined hierarchical safety constraint framework and includes: The multimodal sensing unit is used to simultaneously collect physiological indicator data, subjective feedback data, and real-time safety status data to generate physiological state vectors, subjective demand signals, and real-time safety status vectors. The medical order model parsing unit is used to receive the treatment parameters set by the physician and parse them into long-term treatment target benchmark signals; A subjective-physiological inconsistency assessment unit is used to receive the physiological state vector and the subjective demand signal, and generate a subjective-physiological inconsistency index. The medication adherence monitoring unit is used to receive long-term treatment target baseline signals and actual cumulative dosing records from the stratified safety constraint unit, and to calculate the medication adherence drift rate. A dynamic alignment decision unit is used to generate suggested dosing decisions based on a trade-off between the subjective-physiological inconsistency index and the medical order compliance drift rate. A hierarchical safety constraint unit is used to modify the suggested dosing decision based on preset safety constraints, and generate the final dosing instruction and safety boundary activation signal to be executed. An analgesic effect assessment unit is used to monitor the analgesic effect after drug administration and generate alignment effect feedback signals; The dynamic alignment decision-making unit execution experience-medical order bidirectional coupling control mechanism operates as follows: When the subjective-physiological inconsistency index is high and the medical order compliance drift rate is lower than the drift budget threshold, the subjective-physiological inconsistency index is minimized first to generate an active intervention medication decision. When the subjective-physiological inconsistency index is low, or the medication adherence drift rate is high, the medication adherence drift rate is minimized first to generate a conservative regression dosing decision. The process by which the analgesic effect evaluation unit generates the alignment effect feedback signal is as follows: When the patient presses the PCA button, the system administers the medication. Start a timing window based on the onset time of the analgesic used; If, within this window, the patient's subjective-physiological inconsistency index or self-reported pain score fails to fall below the pain relief threshold; This request is then marked as an invalid PCA request, and the alignment effect feedback signal is generated based on the invalid PCA request rate.

2. The remotely controlled analgesia pump system based on artificial intelligence according to claim 1, characterized in that, The physiological data are continuous, high-frequency physiological data, including heart rate variability (HRV) and electrical activity of the skin (EDA); the subjective feedback data are sparse, low-frequency subjective feedback data, including patient-pressed PCA button requests and electronic pain score (VAS).

3. The remotely controlled analgesia pump system based on artificial intelligence according to claim 1, characterized in that, The process by which the subjective-physiological inconsistency assessment unit generates the subjective-physiological inconsistency index is as follows: Based on the physiological state vector, the objective level of analgesia requirement is inferred; Based on the aforementioned subjective demand signals, the level of subjective analgesia demand is quantified; The subjective-physiological inconsistency index is generated by calculating the difference or distance between the objective analgesia demand level and the subjective analgesia demand level.

4. The remotely controlled analgesia pump system based on artificial intelligence according to claim 3, characterized in that: The process of inferring the objective analgesia requirement level is as follows: by using a built-in pre-trained deep learning model, the time series data of the physiological state vector is received, and the continuous objective analgesia requirement level is output in real time. The quantification process of the subjective analgesia demand level is as follows: within a preset time window, sparse subjective demand signals are accumulated or weighted averaged and mapped to the subjective analgesia demand level. The process of generating the subjective-physiological inconsistency index is as follows: calculate the difference between two demand levels over time, which is obtained by weighted difference, dynamic time warping (DTW) distance, or Wasserstein distance.

5. The remotely controlled analgesia pump system based on artificial intelligence according to claim 1, characterized in that, The dynamic alignment decision unit is implemented as a reinforcement learning agent, whose state input is the subjective-physiological inconsistency index and the medical order compliance drift rate, and whose action is the suggested dosing decision.

6. The remotely controlled analgesia pump system based on artificial intelligence according to claim 5, characterized in that, The reward function of the reinforcement learning agent is configured as follows: A positive reward is given when the subjective-physiological inconsistency index decreases; When the medication adherence drift rate approaches the drift budget threshold, a negative penalty is applied. When the security boundary activation signal occurs, a strong negative penalty is imposed; Furthermore, the alignment effect feedback signal from the analgesic effect evaluation unit is used to automatically adjust the reward function of the reinforcement learning agent.

Images