Medical internship student information management method and related equipment

Abstract

The invention discloses a medical internship student information management method and related equipment. The method comprises the steps that a medical college and a practice hospital carry out model training locally to obtain local model parameter updating; under the condition of not exchanging local original training data of each terminal, generating a global optimization practice management model shared by each terminal through a federated learning algorithm based on local model parameter updating so as to distribute the shared global optimization practice management model to medical colleges and the practice hospitals; students to be internalized in a medical college are evaluated based on the global optimization internship management model, a matched internship cultivation plan is recommended, multi-dimensional ability evaluation is performed on the students based on the model, constraint conditions are comprehensively considered, and the matched internship cultivation plan is recommended for the students. The problem that medical internship management between hospitals and colleges mainly has data islands and collaborative barriers between hospitals and colleges is solved, and the scientificity and efficiency of internship management are improved on the premise that privacy is protected.

Description

Technical Field

[0001] The embodiments of this application relate to the field of smart healthcare, and more specifically, the present invention relates to a method and related equipment for managing information of medical interns. Background Technology

[0002] Currently, the main problems in medical internship management between hospitals and universities are data silos and collaboration barriers. Hospital internship management systems and university teaching systems are independent of each other, with inconsistent data formats and standards, creating information silos. This results in the inability to synchronize internship plans, student information, and assessment results in real time, leading to low management efficiency and high collaboration costs. Summary of the Invention

[0003] The summary section introduces a series of simplified concepts, which will be further explained in detail in the detailed description section. The summary section of this invention is not intended to limit the key features and essential technical features of the claimed technical solution, nor is it intended to determine the scope of protection of the claimed technical solution.

[0004] To address the issues of data silos and collaboration barriers in medical internship management between hospitals and universities, this invention proposes a method for managing medical intern information. This method includes: Medical colleges and affiliated hospitals each train the model locally to obtain local model parameter updates. Without exchanging their respective local original training data, a globally optimized internship management model is generated based on the local model parameter updates using a federated learning algorithm, and then the shared globally optimized internship management model is distributed to the medical school and the internship hospital. Based on the aforementioned global optimization internship management model, medical students awaiting internships are evaluated, and matching internship training plans are recommended.

[0005] Optional, also includes: Based on the aforementioned global optimization internship management model, the internship candidates and hospitals in medical colleges are evaluated, and matching internship hospitals are recommended for the interns and / or matching interns are recommended for the internship hospitals.

[0006] Optional, also includes: The interns' competency level is assessed based on the global optimization internship management model and their internship behavior information at the internship hospital to generate an internship assessment score.

[0007] Optional, also includes: After the aggregation is completed, the version identifier and effective time of the globally optimized internship management model are distributed along with the model so that the medical colleges and the internship hospitals can perform evaluations based on a consistent model version.

[0008] Optionally, the assessment of medical students awaiting internships based on the global optimized internship management model, and the recommendation of matching internship training plans, includes: Based on the global optimization internship management model, features are extracted from the past learning performance, standardized assessment results and interest preferences of the interns, and the matching degree and risk assessment results are output for the interns. Based on the preset thresholds and strategies, the internship training plan recommendation results are generated.

[0009] Secondly, the present invention also proposes a medical intern information management device, comprising: Local training units are used by medical schools and affiliated hospitals to train models locally and update local model parameters. The fusion optimization unit is used to generate a globally optimized internship management model shared by all parties through a federated learning algorithm based on the local model parameter updates without exchanging their respective local original training data, so as to distribute the shared globally optimized internship management model to the medical college and the internship hospital. The evaluation unit is used to evaluate medical students awaiting internships based on the global optimized internship management model and recommend matching internship training plans.

[0010] Thirdly, an electronic device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program stored in the memory to implement the steps of the medical intern information management method as described in any of the first aspects above.

[0011] Fourthly, the present invention also proposes a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the medical intern information management method of any of the above claims in the first aspect.

[0012] In summary, the medical intern information management method proposed in this application involves medical schools and internship hospitals training their models locally to obtain local model parameter updates. Without exchanging their original local training data, a globally optimized internship management model is generated based on these local model parameter updates using a federated learning algorithm. This shared globally optimized internship management model is then distributed to the medical schools and internship hospitals. Based on this globally optimized internship management model, interns awaiting internships at the medical schools are evaluated, and matching internship training plans are recommended. From a learning theory perspective, the empirical risk of the global model approximates the risk under a multi-institutional joint distribution, while the single-point model approximates the risk under a marginal distribution within the institution. In multi-domain tasks, the former is more robust to extrapolation for unknown interns and unfamiliar departments, thus achieving higher evaluation accuracy and lower misjudgment rate under the same privacy boundaries. The combination of a personalized layer and robust aggregation reduces the impact of abnormal patterns in a particular institution on the global update. Therefore, when the school's scoring is stricter, the hospital's operational standards are more detailed, or the case structure is different, the model will not be forced to a local minimum at any one end. Explicit alignment of competency vectors with job requirement vectors translates the reasons for recommending a particular major and extending a rotation into verifiable, quantifiable factors. This facilitates review by the teaching team and the academic affairs office, and also allows for tracing the basis for judgment in case of disputes. Since only parameter updates are exchanged, bandwidth pressure varies with model size rather than sample size. Introducing distillation can further control downlink distribution costs, making it easier to scale across multiple institutions and bases.

[0013] The medical intern information management method of the present invention, other advantages, objectives and features of the present invention will be apparent in part from the following description, and in part will be understood by those skilled in the art through study and practice of the present invention. Attached Figure Description

[0014] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit this specification. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 A schematic diagram of a method for managing medical intern information provided in this application embodiment; Figure 2 A schematic diagram of a medical intern information management device provided in this application embodiment; Figure 3 This is a schematic diagram of the structure of an electronic device for managing information of medical interns, provided as an embodiment of this application. Detailed Implementation

[0015] The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus. The technical solutions of the embodiments of this application will now be clearly and completely described in conjunction with the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them.

[0016] To address the issues of data silos and collaboration barriers between hospitals and universities in medical internship management, please refer to [link / reference needed]. Figure 1 This is a flowchart illustrating a method for managing information of medical interns provided in an embodiment of this application, which may specifically include steps S110 to S130.

[0017] S110, medical colleges and affiliated hospitals each train the model locally to obtain local model parameter updates.

[0018] S120, without exchanging their respective local original training data, based on the local model parameter updates, a globally optimized internship management model shared by all parties is generated through a federated learning algorithm, so as to distribute the shared globally optimized internship management model to the medical school and the internship hospital.

[0019] S130, Based on the global optimization internship management model, evaluate the medical students awaiting internships and recommend matching internship training plans.

[0020] Understandably, medical internship information suffers from significant data silos and domain differences across universities and hospitals. Universities tend to prioritize course grades, objective structured clinical exam scores, mannequin training records, knowledge gaps, and learning duration. Hospitals, on the other hand, prioritize real-world clinical scenarios such as departmental rotations, teaching records, participation in anonymized cases within the electronic medical record's visible range, compliance rates with standardized procedures, and adverse event reviews. Traditional methods for collaborative modeling typically require centralized data, which violates privacy laws and leads to accumulated biases due to inconsistent data sets. The basic idea of ​​federated learning is to train models locally at each participant, exchanging only securely processed parameter updates or gradient information, with the coordinator aggregating the data to obtain a shared global model. This achieves higher generalization capabilities than single-point training without exposing the original data. The reason this approach improves generalization is that the updates to each local parameter, statistically speaking, are equivalent to approximating different directions of the target risk function. The aggregated direction more closely approximates the risk reduction direction of the true overall distribution. In cases where there are differences in the distribution between universities and hospitals, introducing a personalization layer, weighted aggregation, and robust aggregation can reduce the bias of a single domain on the global model. Simultaneously, differential privacy and secure aggregation ensure that updated values ​​cannot be used to infer the original samples. Finally, a global optimization model is used to assess the abilities of interns, and the assessment vector, along with hospital job requirements, departmental workload, teaching resources, and risk constraints, are integrated into a matching and recommendation process to form an executable internship training plan.

[0021] For example, the university and hospital sides each complete feature governance and standard alignment locally. First, a semantic feature dictionary is established, mapping synonymous fields from different sources to the same semantic entry. For instance, the counting of non-standard hand hygiene and the counting of contamination in sterile areas are semantically distinguished as the frequency of sterile operation violations and the frequency of missed hand hygiene items, clearly specifying the counting period, sampling time, and unit of measurement. Second, time-related features are statistically analyzed using sliding time windows, with weekly, monthly, and rotational windows commonly used in internship assessments. Third, missing values ​​are imputed using a distribution-preserving strategy, and a missing value indicator feature is added to each imputed field to retain uncertainty information. Finally, labels required for training are generated locally. Labels can be binary (whether the department's independent operation access standard has been met), multi-level (expected level after rotation), or real-valued operation standardization scores. This transforms data from heterogeneous systems into a learnable, unified representation, reducing systematic errors in subsequent models caused by inconsistent standards. Each participant builds a model with isomorphic or equivalent expressive power locally. In practice, a deep network architecture consisting of a general representation layer, a domain adaptation layer, and a task output layer can be used. The general layer learns skill representations shared across institutions, the domain adaptation layer absorbs institutional distribution differences through conditional normalization or small-scale gating networks, and the task layer provides qualification probabilities, risk stratification, and capability vectors. The loss function uses cross-entropy in classification tasks, dual ranking loss in matching priority ranking scenarios, and uncertainty-weighted summation to automatically allocate task weights in multi-task scenarios. Local training employs mini-batch stochastic gradient descent or AdamW, with early stopping and gradient pruning to ensure the model converges towards stability before each round of federated communication. This allows each institution to produce stable, aggregateable parameter updates while retaining adaptability to local distributions. After several epochs of local training, the client performs safe processing on parameter updates. First, pruning is performed to limit the norm of single-client updates, controlling their contribution. Then, calibrated noise is superimposed to achieve a preset differential privacy budget, making the impact of any single sample on the update value statistically difficult to identify. Finally, the data is uploaded via a secure aggregation protocol. This protocol allows the coordinator to see only the encrypted composite updates from each party, not the plaintext updates from any single party. This reduces the risk of inversion attacks through stabilization and randomization, and ensures that the coordinator and other participants cannot reconstruct the original sample or sensitive tags from the transmitted content. In each round of communication, the coordinator receives the encrypted summation result, decrypts it, and then performs the aggregation. A basic strategy can be a sample-weighted average. When there is strong data imbalance or uneven data quality, robust aggregation is introduced, such as median coordinates, truncated mean, or outlier suppression methods like Krum, to reduce the interference of anomalous updates on the global direction. When certain institutions cover specific departments, cases, or equipment, dynamic weights based on information gain are used, giving higher weights to updates that contribute more to global generalization.The aggregated global optimized internship management model is used to monitor drift between rounds on the evaluation set. If non-monotonic improvement is detected, adaptive adjustments to the learning rate and weighting strategy are triggered. Multi-source local improvements are synthesized into global improvements that can be stably improved across a wider distribution, significantly reducing the degradation caused by single-domain shifts. The coordinating end distributes the latest global model back to universities and hospitals. Each end freezes the general representation layer locally, only making rapid fine-tuning to small-scale personalized and task layers, ensuring the model fits the institution's context without compromising sharing capabilities, such as the details of a hospital's puncture procedure or a university's scoring scale and semantics. For scenarios with limited end resources, knowledge distillation can be used to compress the global teacher model into a lightweight student model. This achieves the advantages of both sharing and personalization: the shared layer provides robust representations across institutions, and the personalized layer digests local specificities, thus avoiding the generalization loss of a one-size-fits-all approach. Before students enter internship matching, the university uses the global model to generate capability vectors for students. The capability vector dimensions can cover theoretical mastery, operational standardization rate, communication and record quality, night shift stress tolerance, and risk prevention awareness, among others. The hospital-side maintenance staff has a demand vector and capacity constraints, including department capacity, teaching qualifications, case structure, shift allocation limits, and compliance risk thresholds. The recommendation process includes feasibility screening to eliminate combinations that violate constraints, and optimality solving using a multi-objective function that considers student growth benefits, hospital teaching burden, patient safety risks, and school-hospital balance, with adjustable weights. Allocation schemes are obtained through heuristic stable matching or maximum weight matching algorithms. Penalties are applied to hard constraints such as night shifts, scarce specialties, and operational access requirements when necessary to ensure solutions meet management red lines. After system operation, data distribution will subtly change with semester changes, equipment updates, and minor adjustments to scoring criteria. Each end periodically triggers small-batch retraining, incorporating the latest round of teaching scores, practical quality control, and complication reviews into the sample. The coordination end monitors indicator trends and increases communication frequency or adjusts aggregation robustness when performance declines. For cold starts of newly opened departments or new evaluation dimensions, parameter initialization transfer and small-sample metric learning can be used to provide stable evaluations even with limited labeling. This step ensures the model maintains performance in real-world scenarios and is not dragged down by slow drift. Access control and the principle of least privilege limit the range of data accessible during training on each endpoint; training logs, model versions, and aggregate summaries are recorded locally, and hash values ​​of sensitive operations can be stored on the blockchain for auditing purposes. Differential privacy budgets are allocated centrally at the semester level to ensure that cumulative noise does not exceed a preset threshold.

[0022] In summary, the medical intern information management method provided in this application involves medical schools and internship hospitals training their models locally to update local model parameters. Without exchanging their original local training data, a globally optimized internship management model is generated based on the updated local model parameters using a federated learning algorithm. This globally optimized internship management model is then distributed to the medical schools and internship hospitals. Based on this globally optimized internship management model, the interns at the medical schools are evaluated, and matching internship training plans are recommended. From a learning theory perspective, the empirical risk of the global model approximates the risk under a multi-institutional joint distribution, while the single-point model approximates the risk under a marginal distribution within the institution. In multi-domain tasks, the former is more robust to extrapolation for unknown interns and unfamiliar departments, thus achieving higher evaluation accuracy and lower misjudgment rate under the same privacy boundaries. The combination of a personalized layer and robust aggregation reduces the impact of abnormal patterns in a particular institution on the global update. Therefore, when the school's scoring is stricter, the hospital's operational standards are more detailed, or the case structure is different, the model will not be forced to a local minimum at one end. Explicit alignment of competency vectors with job requirement vectors translates the reasons for recommending a particular major and extending a rotation into verifiable, quantifiable factors. This facilitates review by the teaching team and the academic affairs office, and also allows for tracing the basis for judgment in case of disputes. Since only parameter updates are exchanged, bandwidth pressure varies with model size rather than sample size. Introducing distillation can further control downlink distribution costs, making it easier to scale across multiple institutions and bases.

[0023] In some examples, it also includes: Based on the aforementioned global optimization internship management model, the internship candidates and hospitals in medical colleges are evaluated, and matching internship hospitals are recommended for the interns and / or matching interns are recommended for the internship hospitals.

[0024] Understandably, the global optimization internship management model can generate not only capability vectors for trainees but also job demand vectors and mentoring capacity vectors for internship hospitals and their respective departments. By incorporating trainee capabilities, job demands, capacity constraints, safety risks, and training benefits into a single evaluation and matching framework, the traditional one-way allocation driven by human experience can be transformed into a computable recommendation process oriented towards the preferences of both ends. Fit is measured in the same representation space, and a multi-objective function characterizes the mentoring benefits on the hospital side and the growth benefits on the trainee side. With capacity, qualifications, shifts, and safety red lines as hard constraints, the recommended list is obtained through interpretable optimization or stable matching.

[0025] For example, the global model outputs a competency vector for each intern at the institution level, covering dimensions such as theoretical mastery, standardized operation, communication records, shift workload, risk control, and self-motivated learning. At the hospital level, it periodically generates a job requirement vector for each department, including case structure, frequency of key operations, teaching strengths, record-keeping requirements, and psychological communication requirements. Simultaneously, it generates a capacity vector, specifying the nurse-to-bed ratio, number of mentors, night shift speaking privileges, and essential qualification thresholds. A unified semantic feature dictionary ensures that the two vectors are comparable within the same metric space. For any combination of intern and department, a fit score is first calculated, given by the weighted similarity between the competency vector and the requirement vector, incorporating risk penalties and growth potential bonuses. Then, a feasibility domain is selected based on the capacity vector; those who violate entry requirements, mentor quotas, nurse-to-bed ratio, night shift restrictions, or compliance red lines are directly eliminated. The system outputs a set of candidate hospitals for each intern and a pool of candidate interns for each hospital. The system generates a preference sequence for interns, sorted according to fit, growth potential, commuting radius, and personal development direction. A preference sequence is generated for the hospital side, ranked according to factors such as teaching benefits, safety risks, team structure balance, and the degree of filling departmental gaps. Key factors and weights for preference generation are visible to both the teaching group and the academic affairs office, facilitating human-machine collaborative decision-making. A set of matches is solved within the feasible region to optimize the overall objective function, which integrates four categories of indicators: student growth benefits, hospital teaching benefits, system fairness, and rotation balance. The solution can employ stable matching to ensure no mutually exclusive pairs, or maximum-weight matching with penalties for unfair solutions. If necessary, a hierarchical solution using hard constraints followed by soft objectives can be used. For arrangements with a time dimension, a phased variable is introduced, first determining the current period's objective, then continuously optimizing the next period's arrangement. For matched pairs, the hospital's schedule, university teaching calendar, and student exam schedules are automatically integrated to generate a training plan including rotation order, stage objectives, assessment nodes, mentor assignments, a list of required cases, and a safety training list. Key assessment points are backfilled as annotation sources for subsequent federated retraining to maintain model freshness. After deployment, monitor the adaptation indicators, such as the improvement of the standardization rate, the change of the error rate, the change of the complaint rate, and the fluctuation of the tutor's workload. If there is a continuous deterioration or abnormal peak, the system will automatically reduce the weight of the matching or temporarily freeze the relevant combination and trigger the person in charge to review.

[0026] In some examples, it also includes: The interns' competency level is assessed based on the global optimization internship management model and their internship behavior information at the internship hospital to generate an internship assessment score.

[0027] Understandably, the global optimization internship management model provides a shared capability representation and assessment boundary across institutions. However, trainees' real-time behavior in specific hospitals and departments is influenced by case structure, teaching style, shift intensity, and equipment differences. Relying solely on pre-admission or on-campus data will result in assessment lag and bias. By integrating internship behavioral information such as process indicators of key operations, teaching evaluation statements, electronic record integrity, patient interaction feedback, and time-based workload with the global model, static capability profiles can be transformed into contextualized capability assessments. Therefore, using the capability vector output by the global model as a priori representation, and then introducing temporal behavioral characteristics representing trainees' performance within their own hospital and department, the model can weigh the latest behavior through gating updates or attention weighting, thereby obtaining a capability level estimate that balances priori and on-site performance, and generating scores tailored to assessment scenarios.

[0028] For example, a behavioral feature set for internship scenarios is established on the hospital side. Structured features include aseptic technique violations, first-time puncture success rate, medication verification errors, missing handover records, standardized markings of doctor-patient communication language, night shift attendance stability, mentor intervention frequency, and patient satisfaction tags. Unstructured features include teaching comments, medical record writing content, and ward round Q&A records. A semantic feature dictionary maps synonymous indicators to unified entries, clearly defining the calculation granularity and time window; for example, a weekly window is used to calculate the stability of repetitive operations, and a rotation window is used to calculate the stage achievement rate. This ensures that behavioral data from different departments and information sources are comparable in the same semantic space. Textual comments are anonymized, removing identifying markers such as names and bed numbers. Numerical behavioral features are standardized in units and outlier truncation is applied. For each feature, the data source, the evaluator's identity category, and a timestamp are recorded, and a confidence weight is generated to suppress noise sources and low-confidence scores during subsequent evaluations. This ensures privacy and data quality. Before trainees enter the hospital or new department, the institution outputs a set of competency vectors based on a global model, covering dimensions such as theoretical mastery, standardized operation, record quality, communication skills, stress resistance, and risk prevention awareness, along with uncertainty estimates. These vectors are transmitted to the hospital via a secure channel for assessment purposes only and do not contain any original internal data. An assessment sub-model is deployed at the hospital, with input consisting of two parts: a prior competency vector and a chronologically ordered sequence of behavioral features. The sub-model can employ gated recursion or attention mechanisms to maximize the impact of recent high-confidence behaviors on competency estimation. For differences in mentors or case structures, conditional normalization is used as contextual conditions to avoid misjudging scenario difficulty as trainee inadequacy. The output is an updated competency level estimate and the contribution of each competency dimension. Based on the competency level estimate, and combined with the department's assessment syllabus and weighting strategy, a comprehensive score is calculated, along with dimensional sub-scores and suggested items. The assessment score can include both quantitative components, such as a percentage-based composite score and whether the entry threshold has been met, and qualitative conclusions, such as recommendations to increase the training time for aseptic draping and to reduce high-risk independent procedures during night shifts. To prevent a single, isolated event from excessively influencing the results, a time decay and robust aggregation strategy is introduced, ensuring that continuous improvement is reflected in the score trend. For the generated scores, the most influential behavioral evidence and model attention weights are shown to the teaching team. When the difference between the score and the instructor's subjective judgment exceeds a set threshold, the system prompts for review and allows the instructor to supplement with structured explanations. This information, provided it complies with regulations, can be included in the data source for subsequent federated retraining. The assessment sub-model runs locally in the hospital, and the behavioral data remains within the hospital. Parameter updates, after being pruned and noise-reduced, are only uploaded during the federated training cycle. The content displayed to trainees and institutions follows the principle of minimum usability; for example, institutions can only see the sub-scores and recommendations, not sensitive case details.

[0029] In some examples, the internship behavior information includes structured process indicators for the internship process, anonymized teaching comments, and case writing content. This internship behavior information is standardized and credible according to a preset time window and a semantic feature dictionary. The method also includes: When trainees enter the department, the prior capability vector output by the global optimization internship management model on the medical school side is obtained and transmitted to the internship hospital side through a secure channel for evaluation and initialization. A contextualized assessment sub-model is constructed on the hospital side. The prior capability vector and the behavioral feature sequence arranged in chronological order are used as inputs. Time location coding and load intensity are used as modulation quantities to distinguish steady-state performance from short-term performance affected by load. The output includes capability level results that simultaneously include steady-state capability estimation and temporary capability estimation. Based on the assessment outline and weighting strategy of the trainee's department, the ability level results are mapped into a comprehensive assessment score and sub-item scores, and improvement suggestions and evidence points for training and admission decisions are generated. The assessment sub-model runs locally at the internship hospital.

[0030] Understandably, in some cases, capacity does not increase monotonically; consecutive night shifts or peak outpatient visits may cause a periodic decline. Therefore, by modeling shift rhythms as an exogenously driven non-stationary process, a rhythm modulation unit is introduced to allow recent load to produce interpretable modulation of capacity estimates.

[0031] For example, a behavioral feature set oriented towards the internship process can be defined on the hospital side. Structured data may include, for instance, counts of aseptic technique violations, first-time puncture success rate, medication verification errors, missing handover records, alarm response time, mentor intervention frequency, night shift attendance stability, and patient satisfaction tags. Text data may include instructor comments and case notes. To avoid homonyms and statistical discrepancies, a semantic feature dictionary is established, clearly defining the data source, statistical granularity, and unit for each feature. Three types of time windows—weekly, monthly, and rotating—are used to generate statistics, and each record is assigned a collection timestamp and a credibility marker. For example, the weight of comments from novice raters is reduced, while the weight of counts that are consistent after sampling is increased. This maps multi-source information to a comparable, unified semantic space, reducing systematic errors caused by inconsistent standards in subsequent evaluations. The text is anonymized locally at the hospital, removing identifiable information such as names, bed numbers, and hospital admission numbers. It is then segmented into clauses and a glossary, representing semantic components (e.g., communication coherence, reassuring wording, clarity of medical advice explanation, and teamwork performance) and source components (e.g., evaluator's identity category, department, or time period). Identity embedding is introduced for source components in subsequent modeling. During training, a source discernibility penalty is added to the loss function, making it difficult for the model to identify source features from the ability representation. Simultaneously, consistency constraints ensure that semantic components align with the direction of relevant structured indicators; for example, negative communication statements should be positively correlated with the probability of co-occurrence with complaint tags. This reduces the penetration of subjective bias and improves the comparability of soft skills dimensions. Before trainees enter the department, the institution uses a globally optimized internship management model to output a prior ability vector covering dimensions such as theoretical mastery, standardized operation, communication records, stress tolerance, risk prevention, and self-motivated learning, along with uncertainty estimation. This vector is transmitted to the hospital through an authorized secure channel, used only as an initialization vector for evaluation, and does not carry original internal data or identifiable information. This approach utilizes minimal bandwidth and low leakage surface to introduce robust representations trained across institutions into the hospital setting, reducing the drift risk associated with purely local small-sample assessments. An assessment sub-model is deployed at the hospital, with input consisting of two parts: a prior capability vector and a sequence of behavioral features arranged chronologically. The model encodes the sequence of occurrence using temporal location encoding, modulates workload intensity (e.g., time-based workload or bed-to-nurse ratio peak outpatient / emergency index), and incorporates mentor identity, equipment model, and sampling strategy as contextual conditions into the conditional normalization unit, resulting in different weights for the same behavior in different scenarios. The model outputs a set of capability level results, including steady-state capability estimates and temporary capability estimates, along with dimensional contribution values ​​for subsequent interpretable presentation. This setup separates stable competence from short-term fluctuations, preventing the misjudgment of temporary declines caused by high workloads as long-term capability deficiencies. Based on the department's assessment syllabus and weighting strategy, the capability level results are mapped to a comprehensive assessment score and sub-item scores, while simultaneously generating improvement suggestions and key evidence points.To adapt to different departmental scoring scales, a set of threshold parameters is learned and slowly updated with quarterly data and case structure, giving the admission and qualification lines learnable flexibility. To suppress the excessive influence of incidental events, robust aggregation and time decay strategies are adopted, allowing continuous improvement to be gradually reflected in the scoring trend. The output includes not only a set of scores but also a list of minimum improvement actions corresponding to low-scoring dimensions. For example, in the communication dimension, it provides clear requirements for conducting three scenario drills in high-pressure pediatric outpatient and emergency departments and passing video quality control. Thus, firstly, the behavioral sequence is updated in a rolling time window, allowing the model to respond quickly to changes in trainees' status and avoid the lag caused by relying solely on pre-admission profiles. Secondly, the combined effects of source discernibility penalty, equipment context conditions, and load modulation enable the scoring to distinguish between performance decline due to high task difficulty and genuine inadequacy, reducing unfair penalties. Thirdly, the prior ability vector provides long-term representation, while the behavioral sequence provides proximal correction; the functional division and consistency constraints of these two types of information reduce the impact of incidental noise. Finally, the scoring mapping directly produces sub-item scores and a list of minimum improvement actions, facilitating teaching arrangements and admission decisions.

[0032] In some examples, it also includes: On the hospital side, a teaching evaluation representation containing semantic component representation and source component representation is constructed. The source component representation generates a source identity embedding based on the evaluator's identity category, department, and time period information. During local training at the internship hospital, the semantic component representation and the source identity embedding are jointly input into the evaluation sub-model. A penalty term for source discernibility is added to the loss function to suppress the penetration of source preference into the ability representation. Consistency constraints are used to ensure that the semantic components of the teaching evaluation and the structured process indicators are statistically consistent. During the assessment phase, the semantic components of the mentoring comments, which have undergone source ablation, are jointly input into the assessment sub-model along with the structured process indicators of the internship process to output the results of the intern's competency level. The original text of the mentoring comments is used in the assessment sub-model, which runs locally at the internship hospital.

[0033] Understandably, in some cases, such as when attitude is average or communication needs to be strengthened, such textual labels are subjective and context-dependent, and it is easy to mistake the tutor's personal style for the student's ability. Textual evidence can be broken down into semantic components and source components, and expression patterns and source preferences can be modeled separately. Source dissolution can be used to remove personal biases.

[0034] For example, on the hospital side, the teaching evaluation comments are de-identified by removing identifiable fields such as name, bed number, hospital number, or mobile phone number. Sentence and phrase segmentation is completed using a medical terminology list and a stop word list, forming stable units for phrases such as insufficient explanation in doctor-patient communication, omissions in shift handover, and appropriate preoperative reassurance. Simultaneously, meta-information such as the generation timestamp, the department to which the evaluation belongs, the shift time, and the evaluator's identity category are retained as context for subsequent source modeling. This transforms the original text into a computable corpus with controllable privacy risks and clarifies contextual boundaries, laying a traceable foundation for subsequent semantic and source decoupling. Based on a predefined semantic feature dictionary, the segmented text is aligned to semantic components of several capability dimensions, such as communication coherence, reassurance strategies, clarity of terminology explanation, teamwork, record-keeping standards, or risk awareness. Each component is given its frequency of occurrence, strength polarity, and contextual clues. Simultaneously, source identity embeddings are generated based on the evaluator's identity category, department, and time-period stress level, serving as a separate input channel. Orthogonally modeling semantic and source information at the representation level provides a clear interface for subsequent source suppression and consistency calibration, avoiding uncontrollable biases at the feature layer. When training the evaluation sub-model locally in the hospital, the model simultaneously achieves two objectives: first, predicting ability dimension scores based on textual semantic components and structured process indicators; second, minimizing source identity in the ability representation. To this end, a source discernibility penalty is added to the loss function, suppressing any tendency to infer which tutor or department wrote the evaluation based on the ability representation. Subspaces in the representation space that are insensitive to source but sensitive to ability discrimination can be found, resulting in scores closer to reality and reduced subjective bias, especially in environments with multiple tutors and shifts, significantly reducing the drift caused by changing tutors and scores. To mitigate the risk of drawing conclusions solely based on linguistic bias, cross-modal consistency constraints are introduced during training. This involves explicitly modeling the statistical relationships between semantic components and structured process indicators. For example, the probability of a joint occurrence of negative expressions related to communication coherence and a decrease in family satisfaction should be higher than random levels, and the correlation between positive expressions related to record-keeping standards and a decrease in the number of handover omissions should reach a preset threshold. Long-term deviations will prompt for data quality review or weight adjustment. The technical effect of this step is to anchor the impact of textual evidence on scoring to objective metrics, reducing the amplification of biases caused by a single modality. During the evaluation phase, the semantic components of the instructor's comments, after source ablation, are used together with structured process indicators from the same time window as input. Contextual factors such as shift workload, case complexity, equipment model, and instructor changes are used as conditional variables. The model first outputs a capability estimate decomposed by dimension, and then aggregates it into a comprehensive score and sub-item scores based on the departmental assessment syllabus. The original teaching evaluation comments and their de-identified copies are kept only locally at the hospital. The parameter updates of the evaluation sub-model are uploaded after norm pruning and noise injection during the federated training cycle. The coordinator uses robust aggregation to generate a global model for redistribution.Universities and other hospitals can only see the individual scores and recommendations, but cannot access any original text or reversible text traces.

[0035] In some cases, the rating scales of institutions and hospitals change slowly with semesters and mentor changes, showing no abnormalities in the short term but significantly reducing cross-institutional comparability after six months. Considering that ability representations and rating scales can be learned simultaneously during federated aggregation, the implicit rating thresholds for each department and each mentor are treated as learnable parameters and a stable prior is applied to improve this. In some examples, this also includes: Medical colleges and affiliated hospitals each train the model locally to obtain local model parameter updates and scoring scale parameter updates. The scoring scale parameters are associated with the scorer's identity category, department, and quarter and are used to represent the implicit scoring threshold. Without exchanging their respective local original training data, after performing norm pruning and noise injection on the local model parameter update and the scoring scale parameter update, robust aggregation is performed at the coordinating end through federated learning algorithm. Based on sample size weight and outlier suppression strategy, a shared global optimized internship management model and a global scoring scale are generated simultaneously. A stable prior is applied to the global scoring scale to suppress slow drift caused by semester change and tutor change. The shared global optimized internship management model and the global scoring scale are distributed to the medical colleges and the internship hospitals. The general representation layer is frozen on the end side, and the personalized layer and scoring scale are fine-tuned step by step based on the case structure, equipment differences and scoring style of the institution, so as to match the evaluation criteria with the local context while maintaining cross-institutional comparability. Based on the distributed and fine-tuned global optimized internship management model and global scoring scale, the interns in the medical colleges are evaluated, and a capability vector containing steady-state capability estimation and temporary capability estimation is generated. Based on this, a matching internship training plan is recommended. The recommendation is based on the constraints of department capacity, teaching qualifications, class restrictions and patient safety thresholds, and outputs the rotation order, stage goals and evaluation nodes with the goals of trainee growth benefits and hospital teaching benefits. During the operation phase, the scoring distribution, correlation of structured process indicators, and mentor change events are monitored. The scoring scale parameters and aggregation weights are automatically adjusted for detected scoring caliber drift, and the adjusted parameters are updated for subsequent federated training to maintain the stability and fairness of cross-comparison and matching decisions.

[0036] In some cases, the sampling frequency and alarm strategies of monitoring and infusion equipment differ across hospital campuses, leading to incomparability of process indicators with the same name. Considering the possibility of introducing device domain adaptation at the feature layer, mapping device model, firmware version, and sampling period to contextual conditions, and learning device-invariant representations can improve this. In some examples, the internship behavior information includes structured process indicators and anonymized teaching comments and case writing content. This internship behavior information is standardized and credited according to a preset time window and semantic feature dictionary. The method also includes: On the hospital side, a device metadata dictionary is maintained to record the model, firmware version, and sampling period of monitoring and infusion equipment. Based on the device metadata, the process curve is resampled and phase aligned to calibrate the time base of the same index. The device model, firmware version, and sampling period are used as contextual conditions to input the conditional normalization unit of the evaluation sub-model to eliminate device domain differences. When trainees enter the department, the prior capability vector output by the college based on the global optimization internship management model is obtained and transmitted to the hospital through an authorized security channel as an evaluation initialization vector. A contextualized assessment sub-model is constructed on the hospital side. The prior capability vector, the structured process index after equipment domain adaptive processing, and the semantic components of the teaching evaluation after source ablation are used as input in chronological order. Combined with the context of shift load and case complexity, a capability level result that includes both steady-state capability estimation and temporary capability estimation is generated. Based on the department's assessment outline and weighting strategy, the ability level results are mapped to a comprehensive assessment score and sub-item scores. Evidence points and improvement suggestions corresponding to low-scoring dimensions are output. The assessment sub-model runs locally in the hospital, and the original behavioral data and equipment metadata do not leave the hospital. Only during the federated training cycle are parameter updates processed by norm pruning and noise injection uploaded for aggregation, so as to maintain the consistency and comparability of scoring standards across hospital areas while ensuring privacy and traceability.

[0037] Please see Figure 2 One embodiment of the medical intern information management device in this application may include: Local training unit 21 is used by medical colleges and internship hospitals to train models locally and obtain local model parameter updates. The fusion optimization unit 22 is used to generate a globally optimized internship management model shared by each end through a federated learning algorithm based on the local model parameter update without exchanging their respective local original training data, so as to distribute the shared globally optimized internship management model to the medical school and the internship hospital. Evaluation unit 23 is used to evaluate medical students awaiting internships based on the global optimized internship management model and recommend matching internship training plans.

[0038] In summary, the medical intern information management device provided in this application allows medical colleges and internship hospitals to train their models locally and update local model parameters. Without exchanging their original local training data, a globally optimized internship management model is generated based on the updated local model parameters using a federated learning algorithm. This globally optimized internship management model is then distributed to the medical colleges and internship hospitals. Based on this globally optimized internship management model, the interns in the medical colleges are evaluated, and matching internship training plans are recommended. From a learning theory perspective, the empirical risk of the global model approximates the risk under a multi-institutional joint distribution, while the single-point model approximates the risk under a marginal distribution within the institution. In multi-domain tasks, the former is more robust to extrapolation for unknown interns and unfamiliar departments, thus achieving higher evaluation accuracy and lower misjudgment rate under the same privacy boundaries. The combination of a personalized layer and robust aggregation reduces the impact of abnormal patterns in a particular institution on the global update. Therefore, when the school's scoring is stricter, the hospital's operational standards are more detailed, or the case structure is different, the model will not be forced to a local minimum at one end. Explicit alignment of competency vectors with job requirement vectors translates the reasons for recommending a particular major and extending a rotation into verifiable, quantifiable factors. This facilitates review by the teaching team and the academic affairs office, and also allows for tracing the basis for judgment in case of disputes. Since only parameter updates are exchanged, bandwidth pressure varies with model size rather than sample size. Introducing distillation can further control downlink distribution costs, making it easier to scale across multiple institutions and bases.

[0039] like Figure 3 As shown, this application embodiment also provides an electronic device 300, including a memory 310, a processor 320, and a computer program 311 stored in the memory 310 and executable on the processor. When the processor 320 executes the computer program 311, it implements the steps of any of the above-mentioned methods for managing medical intern information.

[0040] Since the electronic device described in this embodiment is the device used to implement the medical intern information management device in the embodiments of this application, those skilled in the art can understand the specific implementation method and various variations of the electronic device in this embodiment based on the method described in the embodiments of this application. Therefore, how the electronic device implements the method in the embodiments of this application will not be described in detail here. Any device used by those skilled in the art to implement the method in the embodiments of this application is within the scope of protection of this application.

[0041] In practical implementation, when the computer program 311 is executed by the processor, it can achieve the following: Figure 1 Any of the corresponding implementation methods in the embodiments.

[0042] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0043] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0044] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0045] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0046] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0047] This application also provides a computer program product, which includes computer software instructions that, when executed on a processing device, cause the processing device to perform actions such as... Figure 1The corresponding embodiment describes the process for managing medical intern information.

[0048] A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of this application is generated. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions may be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available medium that a computer can store or a data storage device such as a server or data center that integrates one or more available media. The available medium may be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state disk (SSD)).

[0049] 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 units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0050] 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 units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units 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, or indirect coupling or communication connection between apparatuses or units, and may be electrical, mechanical, or other forms.

[0051] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0052] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0053] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it 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 all or part 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 of 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.

[0054] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for managing information of medical interns, characterized in that, include: Medical colleges and affiliated hospitals each train the model locally to obtain local model parameter updates. Without exchanging their respective local original training data, a globally optimized internship management model is generated based on the local model parameter updates using a federated learning algorithm, and then the shared globally optimized internship management model is distributed to the medical school and the internship hospital. Based on the aforementioned global optimization internship management model, medical students awaiting internships are evaluated, and matching internship training plans are recommended.

2. The method as described in claim 1, characterized in that, Also includes: Based on the aforementioned global optimization internship management model, the internship candidates and hospitals in medical colleges are evaluated, and matching internship hospitals are recommended for the interns and / or matching interns are recommended for the internship hospitals.

3. The method as described in claim 1, characterized in that, Also includes: The interns' competency level is assessed based on the global optimization internship management model and their internship behavior information at the internship hospital to generate an internship assessment score.

4. The method as described in claim 1, characterized in that, Also includes: After the aggregation is completed, the version identifier and effective time of the globally optimized internship management model are distributed along with the model so that the medical colleges and the internship hospitals can perform evaluations based on a consistent model version.

5. The method according to any one of claims 1 to 4, characterized in that, The assessment of medical students awaiting internships based on the global optimization internship management model, and the recommendation of matching internship training plans, includes: Based on the global optimization internship management model, features are extracted from the past learning performance, standardized assessment results, and interests of the trainees to be interns. Based on the extracted features, the trainees to be interns in medical schools are evaluated and a matching internship training plan is recommended.

6. The method as described in claim 3, characterized in that, The internship behavior information includes structured process indicators for the internship process, anonymized teaching comments, and case writing content. This internship behavior information is standardized and tagged with credibility according to a preset time window and semantic feature dictionary. The method also includes: When trainees enter the department, the prior capability vector output by the global optimization internship management model on the medical school side is obtained and transmitted to the internship hospital side through a secure channel for evaluation and initialization. A contextualized assessment sub-model is constructed on the hospital side. The prior capability vector and the behavioral feature sequence arranged in chronological order are used as inputs. Time location coding and load intensity are used as modulation quantities to distinguish steady-state performance from short-term performance affected by load. The output includes capability level results that simultaneously include steady-state capability estimation and temporary capability estimation. Based on the assessment outline and weighting strategy of the trainee's department, the ability level results are mapped into a comprehensive assessment score and sub-item scores, and improvement suggestions and evidence points for training and admission decisions are generated. The assessment sub-model runs locally at the internship hospital.

7. The method as described in claim 6, characterized in that, Also includes: On the hospital side, a teaching evaluation representation containing semantic component representation and source component representation is constructed. The source component representation generates a source identity embedding based on the evaluator's identity category, department, and time period information. During local training at the internship hospital, the semantic component representation and the source identity embedding are jointly input into the evaluation sub-model. A penalty term for source discernibility is added to the loss function to suppress the penetration of source preference into the ability representation. Consistency constraints are used to ensure that the semantic components of the teaching evaluation and the structured process indicators are statistically consistent. During the assessment phase, the semantic components of the mentoring comments, which have undergone source ablation, are jointly input into the assessment sub-model along with the structured process indicators of the internship process to output the results of the intern's competency level. The original text of the mentoring comments is used in the assessment sub-model, which runs locally at the internship hospital.

8. A medical intern information management device, characterized in that, include: Local training units are used by medical schools and affiliated hospitals to train models locally and update local model parameters. The fusion optimization unit is used to generate a globally optimized internship management model shared by all parties through a federated learning algorithm based on the local model parameter updates without exchanging their respective local original training data, so as to distribute the shared globally optimized internship management model to the medical college and the internship hospital. The evaluation unit is used to evaluate medical students awaiting internships based on the global optimized internship management model and recommend matching internship training plans.

9. An electronic device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor, when executing the computer program stored in the memory, implements the steps of the medical intern information management method as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the medical intern information management method as described in any one of claims 1-7.

Images