Nursing treatment vehicle medical article expiration date management system based on machine learning

Abstract

The invention relates to a nursing treatment vehicle medical article expiration date management system based on machine learning, and belongs to the technical field of medical informatization and intelligent management. According to the system, vehicle-mounted medical article expiration date and environment data are collected in real time through an article information sensing layer, and a hospital-wide dynamic expiration date map is constructed in combination with hospital-wide multi-source information; using a multi-task time sequence prediction model to output a consumption rate, a near-effect period risk and a deployment priority; and the collaborative scheduling module generates a transfer instruction and a use suggestion, and online optimization of the model is realized by executing a feedback closed loop. According to the scheme, hospital-wide effective period collaborative management is realized, and the resource utilization efficiency and the medication safety are improved.

Description

Technical Field

[0001] This invention belongs to the field of medical information and intelligent management technology, specifically relating to a machine learning-based management system for the expiration date of medical supplies in a nursing treatment vehicle. Background Technology

[0002] In modern medical institutions, nursing carts serve as temporary storage and distribution centers for clinical medicines and consumables, and their management efficiency directly impacts medical safety and resource utilization. With the expansion of hospital scale and the increasing demand for refined management, the control of expiration dates for medical supplies on carts has expanded from the level of individual devices to a hospital-wide collaborative scheduling system. Machine learning-based intelligent management systems are considered a crucial technological approach to improve the accuracy of expiration date warnings and the scientific nature of allocation decisions.

[0003] The machine learning-based nursing cart medical supplies expiration date management system aims to dynamically optimize drug allocation and rotation strategies by collecting real-time data on the types, quantities, and expiration dates of items in each cart, combined with historical usage patterns and departmental demand predictions. The goal of this type of system is to prioritize the use of near-expiration items and facilitate their rational interdepartmental circulation, thereby reducing resource waste caused by expiration and ensuring clinical medication safety.

[0004] While existing technologies have introduced barcode scanning or RFID identification to record item information in localized scenarios, they generally suffer from a lack of global collaborative capabilities: the expiration date management of individual treatment carts operates in isolation, failing to perceive the hospital-wide inventory distribution and departmental demand differences, leading to some departments experiencing drug stockpiles and expiration while others face shortages; the lack of effective fusion and learning mechanisms for multi-source heterogeneous data makes it difficult to generate forward-looking and adaptive allocation instructions. Achieving closed-loop management of medical item expiration dates based on intelligent prediction and collaborative scheduling across the entire hospital has become a technical challenge that needs to be addressed. Summary of the Invention

[0005] The purpose of this invention is to provide a machine learning-based management system for the expiration dates of medical supplies in nursing treatment vehicles, in order to solve the problems of drug stockpiling and clinical shortages caused by the isolated operation of expiration date management of treatment vehicles, lack of hospital-wide collaborative scheduling capabilities, and insufficient integration of multi-source data in the existing technology.

[0006] This invention provides a machine learning-based medical supplies expiration date management system for nursing treatment carts, comprising: The item information sensing layer is used to collect in real time the unique identifier, current quantity, production batch number, expiration date, and stored temperature and humidity data of the medical items on the vehicle through non-contact identification devices and environmental sensors deployed on each nursing and treatment vehicle; The hospital's inventory status aggregation platform is used to receive and integrate inventory information perception layer data from all nursing and treatment vehicles, and synchronously access the department schedule, patient admission and discharge records, historical prescription data and real-time inventory list of the inventory center in the hospital information system to build a dynamic inventory expiration map of the entire hospital. The machine learning prediction engine is used to train and run a multi-task time series prediction model based on the hospital-wide dynamic expiration date map of medical supplies, and output the predicted consumption rate of various medical supplies, near-expiration risk level score, and cross-department allocation priority sequence for each department in the next 72 hours. The collaborative scheduling decision module is used to generate a scheduling strategy package containing item transfer instructions, usage order suggestions, and early warning threshold adjustment parameters based on the consumption rate prediction, near-expiration risk level score, and allocation priority sequence. The execution feedback closed-loop unit is used to send the scheduling strategy package to the vehicle terminal of the target nursing treatment vehicle and the corresponding department nurse station, and to transmit the execution status, actual usage records and item removal confirmation information in real time, so that the machine learning prediction engine can perform online model fine-tuning and strategy iteration optimization.

[0007] Preferably, the non-contact identification device is a high-frequency RFID reader array, with the antenna layout covering each storage compartment of the treatment cart, ensuring that non-contact identification is completed when the item has not been removed. The environmental sensors include digital temperature and humidity sensors and light intensity detectors, with a sampling frequency of no less than once every 5 minutes, and are timestamped with the unique identifier of the item.

[0008] Preferably, the hospital-wide dynamic expiration date map is organized in a graph database structure, with nodes representing specific medical item instances and attribute fields including the item's generic name, specifications, supplier, expiration start and end dates, current treatment vehicle number, and department. Edge connections represent potential flow paths of items between different treatment vehicles or storage centers, with weights determined by historical allocation frequency and departmental proximity.

[0009] Preferably, the multi-task temporal prediction model adopts an architecture with a shared underlying feature extraction network and three independent task heads; The shared underlying layer is a bidirectional long short-term memory network. The input feature vector consists of the historical consumption sequence of items, the patient flow index during the same period, seasonal factors, and the remaining days of the current expiration date. The first task head outputs the predicted consumption rate for the next 72 hours in different time periods; The second task head outputs a near-term risk level score, with a value ranging from 0 to 100; The third task leader outputs a priority sequence for cross-departmental allocation.

[0010] Preferably, the calculation logic for the near-expiration risk level score is as follows: When the remaining validity period of an item is less than or equal to 3 days, the base score is 80. For every day the validity period is reduced, the score increases by 5; If the department where the medical supplies in the nursing treatment vehicle are located has a consumption rate of 0 in the past 24 hours and the inventory is greater than or equal to twice the average daily usage, an additional 10 points will be awarded. If the temperature and humidity of the storage environment for the item exceed the range specified in the pharmacopoeia for more than 30 minutes, an additional 15 points will be awarded.

[0011] Preferably, the specific rules for generating item transfer instructions by the collaborative scheduling decision module are as follows: items with a high-risk level score ≥90 and a low consumption rate prediction value less than 50% of the average daily usage are preferentially allocated to departments with a consumption rate prediction value greater than 120% of the average daily usage for the same type of items and a current inventory less than the safety threshold; the transfer path selection is based on the principle of the shortest physical distance and the fewest handover links.

[0012] Preferably, the usage order suggestion is pushed to the nurse station terminal in a visual form, and the similar items in the treatment cart are sorted and prompted according to the principle of first-to-last-expiry. When scanning for use, the system automatically verifies whether it conforms to the recommended order. If the nurse chooses to skip items with near expiry dates and use items with far expiry dates, the system will trigger a secondary confirmation pop-up and record the abnormal operation log.

[0013] Preferably, the early warning threshold adjustment parameters include the number of days in advance for expiration warning, the automatic freezing threshold for high-risk items, and the cross-departmental allocation trigger threshold; The parameters are dynamically adjusted based on the historical prediction accuracy of the machine learning prediction engine and the actual occurrence rate of expiration events; When the prediction accuracy is greater than 95% for 7 consecutive days and there are 0 expiration events, the number of days in advance for expiration warning is reduced by 1 day; if there are 2 expiration events that are not warned within a single week, the number of days in advance for warning is automatically increased by 1 day and the freezing threshold for high-risk items is increased by 5%.

[0014] Preferably, the execution feedback closed-loop unit uses the touch screen interface of the vehicle-mounted terminal of the treatment vehicle and a barcode secondary verification mechanism; All item removal operations must undergo two-factor authentication, meaning the operation will only be effective after the nurse's identity verification and the item's barcode scan are successfully matched.

[0015] Preferably, the system synchronously records the operation time, operator, batch of items, and destination, forming an unalterable operation audit chain.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention constructs an integrated closed-loop management system for the perception, prediction, scheduling, and feedback of the expiration dates of items covering the entire hospital's nursing and treatment vehicles, breaking through the limitations of traditional single-point expiration date monitoring; 2. By integrating multi-source heterogeneous data and adopting a multi-task machine learning architecture, we have achieved accurate quantitative assessment of the consumption trend and near-expiration risk of goods, enabling allocation decisions to shift from experience-driven to data-driven. 3. The system introduces a dynamic threshold adjustment mechanism and execution feedback loop to ensure that the prediction model continuously adapts to actual clinical changes, thereby improving the adaptability and robustness of expiration date management. 4. By enforcing the "first-come, first-served" principle and embedding an abnormal operation interception mechanism, the waste of near-expiry items caused by human negligence is eliminated from the process, ensuring the safety of clinical medication and the efficient use of medical resources. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention; Figure 2 This is a schematic diagram of the core principle framework of the multi-task time series prediction model in this invention; Figure 3 This is a data fusion logic flowchart of the item information perception layer and the in-hospital item status aggregation platform in this invention; Figure 4 This is a schematic diagram of the multi-level interaction relationship and data flow between the collaborative scheduling decision module and the execution feedback closed-loop unit in this invention. Detailed Implementation

[0018] This invention provides a machine learning-based management system for the expiration dates of medical supplies in a nursing treatment vehicle, the overall technical architecture of which is shown in the attached figure. Figure 1 As shown. This system constructs a five-layer closed-loop system covering the entire hospital—sensing, aggregation, prediction, decision-making, and feedback—to achieve real-time monitoring, risk quantification, intelligent scheduling, and dynamic optimization of the expiration dates of various medical supplies on nursing treatment vehicles. The following will combine this with the attached... Figure 1 To be continued Figure 4 It provides a detailed explanation of the functional implementation, data flow logic, algorithm principles, and engineering details of each component of the system.

[0019] At the item information sensing layer, each nursing cart deployed in clinical departments is equipped with a high-frequency RFID reader array and a multimodal environmental sensor group. The antennas of the high-frequency RFID reader array are precisely positioned inside or at the edge of each independent storage compartment on the cart, ensuring contactless identification of all medical items with passive UHF RFID tags within the compartment without the items being physically removed. Each identification cycle is 30 seconds. The system activates the antennas of each compartment sequentially through a polling mechanism to avoid signal crosstalk and uses a phase difference algorithm to eliminate misreading of tags between adjacent compartments. The collected data includes the item's unique identifier, current compartment number, item name, specifications, production batch number, expiration date, and tag read strength.

[0020] Digital temperature and humidity sensors and light intensity detectors are integrated inside the treatment vehicle, with a sampling frequency set at no less than once every 5 minutes. Each collected environmental parameter and timestamp are precisely bound to the unique identifier of the most recently successfully identified item, forming a three-dimensional data tuple of "item-environment-time". All sensing data is locally cached and initially verified by the vehicle's embedded processor before being encrypted and uploaded to the hospital's item status aggregation platform via the hospital's internal Wi-Fi 6 network.

[0021] The hospital's inventory status aggregation platform receives and integrates sensor data streams from all nursing and treatment vehicles across the hospital, and simultaneously connects to multiple heterogeneous data sources within the hospital's information system. Please refer to the appendix. Figure 3 The data fusion logic flowchart shown first illustrates how the platform establishes a unified time synchronization benchmark to address timestamp discrepancies caused by network latency or clock drift across different treatment vehicles. The maximum permissible deviation threshold is set at 2 seconds. Subsequently, the platform performs multi-dimensional correlation and matching between the received item information and the hospital information system's departmental schedules, patient admission and discharge records, historical prescription database, and the central pharmacy's real-time inventory list.

[0022] For example, when a treatment vehicle is located in the third ward of the internal medicine department, the system automatically assigns it to the node of that clinical department. If the clinical department has 8 new inpatients that day, and historical data shows that each new inpatient consumes an average of 3 vials of ceftriaxone sodium for injection, the platform can estimate the potential increase in demand for that clinical department that day. All the associated data is structured and stored in a distributed graph database to construct a hospital-wide dynamic expiration date map of supplies.

[0023] In this graph, each node represents a specific medical item instance. Its attribute fields include at least 15 dimensions of information, such as the item's generic name, specifications (e.g., 0.5 grams / piece), supplier name, expiration start date, expiration end date, current treatment cart number, department code, current quantity, and historical temperature and humidity sequence of the storage environment. The edges between nodes represent the physical flow paths that items may take between different treatment carts, departmental medicine cabinets, or central inventory. The weight of each edge is determined by two factors: first, the actual frequency of similar items being allocated along this physical flow path in the past 30 days; and second, the proximity coefficient calculated from the physical distance (in meters) and floor difference (in floors) between the source and target departments on the building floor plan. This hospital-wide dynamic item expiration map is fully updated every 10 minutes and supports millisecond-level subgraph queries and path recommendations.

[0024] The model architecture of the machine learning prediction engine is shown in the attached figure. Figure 2 As shown, this machine learning prediction engine employs a multi-task temporal prediction model, which consists of a shared underlying feature extraction network and three independent task output heads.

[0025] The shared underlying layer is a bidirectional long short-term memory network. The input feature vector is composed of four types of elements: the first type is the hourly consumption sequence of the item in the past 7 days, which is standardized by sliding window; the second type is the average daily patient flow index of the department in the same period, which is calculated by weighting the number of inpatients, the number of surgeries, and the proportion of outpatients to inpatients on that day; the third type is the seasonal factor, including preset Boolean flags such as whether it is the peak season for influenza, before or after holidays; and the fourth type is the remaining validity days of the current item, represented in integer form.

[0026] The bidirectional Long Short-Term Memory (LSTM) network consists of two layers of hidden units, each with 128 units. It employs the tanh activation function and introduces a dropout rate of 0.3 for regularization to prevent overfitting. During training, the model simultaneously optimizes the joint loss function for the three tasks through backpropagation.

[0027] The first task head is used to output the predicted consumption rate for different time periods within the next 72 hours. Its output layer is a fully connected neural network with ReLU activation function and loss function [missing information]. The mean square error is used, i.e.: Indicates the prediction time step The total number, This represents the actual consumption rate. These are model predictions. The goal of this task is to provide accurate demand forecasts for subsequent scheduling.

[0028] The second task head is responsible for generating near-expiration risk level scores, with values ​​ranging from 0 to 100 as continuous real numbers. Since high-risk samples (scores ≥ 90) account for less than 5% of the training set, resulting in a severe class imbalance, this task uses Focal Loss as the optimization objective to enhance the model's attention to hard-to-classify samples. Furthermore, the system incorporates a rule-based auxiliary scoring logic as a post-processing calibration mechanism for the model output.

[0029] Specifically, when an item's remaining shelf life is less than or equal to 3 days, the base score is set at 80. For each day the shelf life decreases, the score increases by 5. If the department where the item is located has a consumption rate of 0 in the past 24 hours, and the current inventory is greater than or equal to twice the department's average daily usage, an additional 10 points are awarded. If the temperature or humidity of the item's storage environment exceeds the storage conditions specified in the Pharmacopoeia of the People's Republic of China (e.g., for injections, 2℃ to 8℃ and relative humidity ≤60%) for more than 30 minutes, an additional 15 points are awarded. The final risk score is a weighted average of the model's predicted value and the rule-calibrated value, with weights of 0.7 and 0.3, respectively.

[0030] The third task head is used to output the priority sequence for cross-departmental allocation, which is essentially a ranking learning problem. This task head treats each item to be allocated as a candidate and outputs its relative priority score compared to other similar items. During training, the ListNet ranking loss function is used to learn the optimal ranking strategy in an end-to-end manner. In addition to shared features, the model input also incorporates contextual information such as the current inventory level, safety stock threshold, and historical out-of-stock frequency of the target department to improve the practicality of the allocation recommendations.

[0031] The collaborative scheduling decision module receives three outputs from the machine learning prediction engine and generates an executable scheduling strategy package based on preset business rules. Please refer to the appendix. Figure 4The multi-level interaction diagram shown illustrates how the collaborative scheduling decision-making module first sorts all item instances across the hospital in descending order of risk level score, filtering out "high-risk, low-demand" items with a score ≥90 and a predicted consumption rate less than 50% of the item's daily average usage over the next 72 hours. The system then searches the entire hospital for "high-demand, low-inventory" departments that meet the following criteria: their predicted consumption rate for this type of item over the next 72 hours is greater than 120% of the daily average usage, and their current inventory is less than a preset safety stock threshold (typically 1.5 times the daily average usage). Once a match is found, the system generates an item transfer instruction, including the source treatment vehicle number, target department code, item unique identifier, transfer quantity, and expected completion time window (typically within 2 hours). The transfer path planning follows a dual optimization principle: first, minimizing the physical distance (i.e., minimizing the walking or logistics robot path length from the source to the target); and second, minimizing handover steps (prioritizing direct paths that do not require transit through the central pharmacy). Path calculation is based on the 3D topology map in the hospital's Building Information Model (BIM) and implemented using the Dijkstra algorithm.

[0032] In addition to transfer instructions, the scheduling strategy package also includes usage order suggestions and early warning threshold adjustment parameters. Usage order suggestions are pushed to the terminal interface of the target department's nursing station in the form of a visual list. All items in the treatment cart with the same generic name and specifications are arranged in ascending order of their expiration date, and each item is labeled "Recommended for Priority Use" or "Use Deferred". When a nurse scans the item barcode with a handheld terminal to prepare for use, the system compares the selected items in real time to see if they conform to the recommended order. If it detects an operation that skips near-expiration items and selects far-expiration items, the system immediately pops up a secondary confirmation dialog box on the terminal, requiring the nurse to enter a reason for skipping, and automatically records this abnormal operation to the audit log.

[0033] The set of parameters for adjusting the early warning threshold includes: the number of days in advance for expiration warnings (initial value: 5 days); the automatic freezing threshold for high-risk items (initial value: 90 points); and the trigger threshold for cross-departmental allocation (initial value: 85 points). These parameters are not fixed but dynamically adjusted based on the historical performance of the machine learning prediction engine. The specific rules are as follows: If the system's accuracy in predicting the consumption rate of all items is greater than 95% for 7 consecutive days, and no unannounced expiration events occur throughout the hospital, the number of days in advance for expiration warnings will automatically decrease by 1 day to reduce unnecessary warning interference. Conversely, if more than two unannounced expiration events occur within a single week (i.e., the risk score is consistently less than 80 within 24 hours before the item expires), the system will automatically increase the number of days in advance for expiration warnings by 1 day and raise the freezing threshold for high-risk items by 5 percentage points to enhance risk interception capabilities. All parameter adjustments must be confirmed by the system administrator in the backend approval interface before taking effect to ensure a controllable change process.

[0034] The execution feedback closed-loop unit is responsible for sending the scheduling strategy package to the execution end and collecting the execution results, forming a complete control loop. This execution feedback closed-loop unit encrypts and pushes the strategy package to the vehicle-mounted terminal of the target nursing treatment vehicle and the management terminal of the corresponding department's nurse station via the hospital intranet. The vehicle-mounted terminal is equipped with a 7-inch capacitive touchscreen, runs a customized Android operating system, and allows nurses to view transfer instructions, confirm item removal, and report abnormal situations.

[0035] All item removal operations must pass a two-factor authentication mechanism: the first factor is the nurse's biometric authentication, and the second factor is optical scanning verification of the item's barcode. The removal operation is only permitted when the system confirms the operator's identity is legitimate and the scanned item completely matches the unique identifier specified in the dispatch instruction.

[0036] After the operation is completed, the system automatically records complete information such as operation time, operator ID, item batch number, quantity removed, and destination. This record is then hashed, encrypted, and written to a blockchain-based immutable operation audit chain. This immutable operation audit chain adopts a consortium blockchain architecture, with nodes distributed across servers in the Information Technology Department, Pharmacy Department, and Nursing Department, ensuring data transparency and tamper-proof protection. All execution statuses and actual usage records are transmitted back to the hospital's item status aggregation platform in real time, triggering the online fine-tuning mechanism of the machine learning prediction engine. Fine-tuning employs an incremental learning strategy, using only the latest 24-hour feedback data to perform lightweight updates on the last layer of the model, avoiding the computational overhead of full retraining. Fine-tuning occurs every 6 hours, or is triggered immediately when the deviation rate of a single scheduling execution exceeds 15%.

[0037] In summary, this embodiment constructs a medical supply expiration date management system for nursing treatment vehicles, possessing self-sensing, self-predicting, self-decision-making, and self-optimizing capabilities, through a highly integrated hardware sensing network, a multi-source fusion data aggregation platform, a multi-task collaborative prediction model, a rule-driven scheduling engine, and a closed-loop feedback execution mechanism. This system not only achieves accurate early warning and efficient circulation of near-expiration items, but also resolves the structural contradiction of drug stockpiling and clinical shortages, improving the efficiency of medical resource utilization and patient medication safety.

[0038] Building upon the aforementioned embodiments, this embodiment further introduces a cross-hospital model collaborative training mechanism based on federated learning to address the data silo problem among multiple branches of a large medical group, while simultaneously protecting patient privacy and data sovereignty across each hospital. Please refer to the appendix. Figure 1 In this embodiment, each branch hospital deploys a localized in-hospital item status aggregation platform and machine learning prediction engine, but no longer uploads the original patient data or prescription details to the group headquarters.

[0039] Each branch only encrypts and uploads the model's gradient update parameters or intermediate feature representations to the group-level federated aggregation server. The aggregation server uses a secure multi-party computation protocol to perform a weighted average of model updates from N branches, generating global model parameters, which are then distributed to each branch for local model replacement. The weight allocation is dynamically adjusted based on each branch's historical data quality score (calculated by comprehensively considering data integrity, timeliness, and annotation accuracy) to ensure that high-quality data sources contribute more to the global model.

[0040] In the item information sensing layer, this embodiment optimizes the high-frequency RFID reader array for resistance to metal interference. Since some treatment carts are made of stainless steel, traditional RFID signals are susceptible to reflection and attenuation, leading to a decrease in recognition rate. This embodiment attaches a 0.2 mm thick layer of ferrite absorbing material to the inner wall of each storage cell and fine-tunes the antenna operating frequency to the 865 MHz to 868 MHz band to avoid metal resonant frequencies.

[0041] In the data fusion stage of the hospital's inventory status aggregation platform, this embodiment adds a real-time monitoring function for changes in the medical insurance catalog. The system automatically obtains notifications of adjustments to drug reimbursement categories by connecting to the API interface of the National Medical Insurance Service Platform. Once a downgrade in the medical insurance category of an item is detected, the system immediately assesses the potential decline in clinical usage and adds a "demand shrinkage" risk label to the hospital-wide dynamic inventory expiration map for that item, triggering early warning and accelerated allocation strategies.

[0042] In this embodiment, the machine learning prediction engine incorporates an adversarial example detection module. Considering the potential for label corruption, sensor malfunctions, or human error in a clinical environment, the system performs anomaly detection using a lightweight one-dimensional convolutional autoencoder before input features enter the bidirectional long short-term memory network. If the reconstruction error exceeds a preset threshold (e.g., mean squared error > 0.15), the adversarial example is determined to be abnormal, automatically removed, or a manual review process is triggered to prevent noisy data from contaminating model training.

[0043] In this embodiment, the collaborative scheduling decision module extends its deep integration with the hospital's logistics robot. After the scheduling strategy package is generated, the system not only issues instructions to the nurses' station but also directly calls the logistics robot scheduling API, specifying the robot number, pickup coordinates, delivery coordinates, and priority level. The logistics robot navigates using the UWB indoor positioning system. Upon arriving at the source treatment vehicle, it automatically unlocks the electronic lock of the designated compartment, allowing the nurse to place the items into the robot's cargo compartment. After the compartment door closes, it automatically locks and departs. The entire process requires no manual intervention in path planning, significantly improving delivery efficiency.

[0044] In this embodiment, the execution feedback closed-loop unit incorporates voice interaction functionality. Nurses can confirm operations using natural language commands via the microphone of the vehicle-mounted terminal. The system parses the commands using a locally deployed speech recognition model and compares them with dispatch instructions. If they match, the verification process is completed automatically, reducing the burden of manual input. The speech model is periodically fine-tuned by collecting anonymized speech samples from various branches to accommodate different dialect accents.

[0045] Through the above enhancements, this embodiment achieves collaborative evolution of expiration date management capabilities across hospital areas while ensuring data privacy. It also improves the robustness and ease of use of the system in complex physical environments and high-noise data scenarios, making it suitable for large-scale deployment in large medical consortia.

Claims

1. A machine learning based nursing cart medical item expiration management system, comprising: The application comprises: An article information perception layer for collecting the unique identification, current quantity, production batch number, expiration date, and storage temperature and humidity data of the medical articles on the nursing treatment vehicle in real time through the non-contact identification device and environmental sensor deployed on each nursing treatment vehicle; An in-hospital article state aggregation platform for receiving and integrating the article information perception layer data from all nursing treatment vehicles, and synchronously accessing the department scheduling table, patient admission and discharge records, historical prescription data, and real-time inventory list of the inventory center in the hospital information system to construct a hospital-level dynamic article expiration date map; A machine learning prediction engine for training and running a multi-task time series prediction model based on the hospital-level dynamic article expiration date map to output the consumption rate prediction value, near-expiration risk level score, and cross-department allocation priority sequence of each type of medical article in each department within the next 72 hours; A collaborative scheduling decision module for generating a scheduling strategy package containing article transfer instructions, use sequence suggestions, and warning threshold adjustment parameters based on the consumption rate prediction value, near-expiration risk level score, and allocation priority sequence; An execution feedback loop unit for issuing the scheduling strategy package to the vehicle-mounted terminal of the target nursing treatment vehicle and the corresponding department nurse station, and real-time returning the execution status, actual use record, and article removal confirmation information for online model fine-tuning and strategy iteration optimization by the machine learning prediction engine.

2. The machine learning-based nursing care cart medical supply expiration management system according to claim 1, characterized by, The non-contact identification device is a high-frequency RFID reader array with an antenna layout covering each storage compartment of the nursing treatment vehicle to ensure inductive recognition without removing the article; The environmental sensor includes a digital temperature and humidity sensor and an illumination intensity detector with a sampling frequency of not less than 1 time per 5 minutes and time-stamped binding with the unique identification of the article.

3. The machine learning-based nursing care cart medical supply expiration management system of claim 2, wherein, The hospital-level dynamic article expiration date map is organized in a graph database structure, with nodes representing specific medical article instances and attribute fields containing the article generic name, specification, supplier, effective period start and end date, current treatment vehicle number, and belonging department; Edge connections represent the potential transfer path of the article between different treatment vehicles or inventory centers, and the weight is determined by the historical allocation frequency and department proximity.

4. The machine learning-based nursing care cart medical supply expiration management system of claim 3, wherein, The multi-task time series prediction model adopts a shared bottom feature extraction network and three independent task heads architecture; The shared bottom is a bidirectional long short-term memory network, and the input feature vector is composed of the article historical consumption sequence, contemporaneous patient flow index, seasonal factor, and current expiration date remaining days; The first task head outputs the future 72-hour time period consumption rate prediction value; The second task head outputs the near-expiration risk level score with a value range of 0 to 100; The third task head outputs the cross-department allocation priority sequence.

5. The machine learning-based nursing care cart medical supply expiration management system according to claim 4, characterized by, The calculation logic of the near-expiration risk level score is as follows: When the article remaining effective period is less than or equal to 3 days, the base score is 80; Each decrease of 1 day of effective period increases the score by 5; If the department where the nursing treatment vehicle medical article is located has a consumption rate of 0 within the past 24 hours and the inventory quantity is greater than or equal to 2 times the daily average consumption, an additional 10 points are added; If the article storage environment temperature and humidity exceeds the pharmacopoeia specified range for more than 30 minutes, an additional 15 points are added.

6. The machine learning-based nursing care cart medical supply expiration management system of claim 5, wherein, The specific rule of the collaborative scheduling decision module generating the article transfer instruction is: preferentially transferring the article with a high-risk level score greater than or equal to 90 and a low consumption rate prediction value less than 50% of the daily average consumption to a department with a same type article consumption rate prediction value greater than 120% of the daily average consumption and a current inventory less than a safety threshold; The transfer path selection is based on the principle of shortest physical distance and least transfer link.

7. The machine learning-based nursing care cart medical supply expiration management system of claim 6, wherein, The use sequence suggestion is pushed to the nurse station terminal in a visual form, the same type articles in the treatment cart are sorted and prompted according to the principle of first expiration first use, and whether the recommended sequence is met is automatically checked when scanning and using; If the nurse chooses to skip the near expiration date article and use the far expiration date article, the system will trigger a secondary confirmation pop-up window and record the abnormal operation log.

8. The machine learning-based nursing care cart medical supply expiration management system of claim 7, wherein, The early warning threshold adjustment parameters include expiration date early warning advance days, high-risk article automatic freezing threshold, and cross-department transfer triggering threshold; The parameters are dynamically adjusted according to the historical prediction accuracy of the machine learning prediction engine and the actual expiration event occurrence rate; When the prediction accuracy is greater than 95% for 7 consecutive days and there is no expiration event, the expiration date early warning advance days are reduced by 1 day; If there are 2 un-warned expiration events in a single week, the early warning advance days are automatically increased by 1 day and the high-risk article freezing threshold is increased by 5%.

9. The machine learning-based nursing care cart medical supply expiration management system of claim 1, wherein, The execution feedback closed loop unit is through the touch screen interaction interface of the treatment cart terminal and the barcode secondary verification mechanism; All article removal operations must be verified by two factors, i.e., the nurse identity authentication and the successful matching of the article barcode scanning.

10. The machine learning-based nursing care cart medical supply expiration management system of claim 9, wherein, The system synchronously records the operation time, operator, article batch and destination, forming an unalterable operation audit chain.

Images