Cancer patient palliative care suitability evaluation method based on integrated time series forest strategy optimization

Abstract

The application discloses a kind of based on integrated time sequence forest strategy optimization cancer patient alleviation medical applicability evaluation method, through system integration patient time sequence physiological and psychological symptom score, static clinical features and historical intervention record, construct integrated time sequence forest model, realize the precise, explainable baseline prediction of the current state corresponding to patient current state alleviation medical level, realize the precise baseline prediction of current alleviation medical applicability grade.Meanwhile, use multi-output integrated tree regression model to construct state transfer predictor, dynamically simulate the symptom evolution trajectory of patient under different intervention strategies.On this basis, introduce reinforcement learning mechanism, learn optimal dynamic level adjustment strategy, and weighted fusion with baseline classifier, supplemented by clinical safety constraints, form a set of data-driven, dynamic evaluation, closed-loop optimization intelligent decision support method, improve the precision of alleviation medical resource allocation, timeliness of intervention measures and long-term benefit of patient quality of life.

Description

Technical Field

[0001] This invention relates to a medical data analysis and decision-making method, specifically to a method for evaluating the applicability of palliative care for cancer patients based on integrated temporal forest strategy optimization, belonging to the medical field. Background Technology

[0002] With advancements in cancer treatment concepts, palliative care has become increasingly important in improving the quality of life and prolonging survival for patients with advanced cancer, especially with early palliative care intervention proving to bring significant benefits. Accurately assessing a patient's suitability for palliative care and developing individualized, tiered intervention plans accordingly has become a crucial step in optimizing the allocation of medical resources and improving overall treatment outcomes.

[0003] Currently, the evaluation of the applicability of palliative care to cancer patients mainly relies on the clinical judgment of physicians or single-dimensional screening tools, such as thresholding based on symptom scales (e.g., ESAS-r) or coarse stratification based on functional status (e.g., KPS score). While these methods are simple, they have significant limitations:

[0004] First, the evaluation process is mostly based on static, single-point-in-time information, which makes it difficult to capture the dynamic evolution of patients' symptoms, psychological and functional status, resulting in the assessment results lagging behind the actual changes in the condition. Secondly, there is a lack of systematic integration and quantitative analysis of multi-source information (such as demographic characteristics, comorbidities, and temporal symptom trajectories), resulting in a single evaluation dimension and insufficient personalization. Third, existing methods mostly focus on current status assessment and fail to form a closed loop optimization with subsequent intervention level recommendations and long-term treatment effects (such as changes in quality of life scores), thus failing to achieve a continuous improvement cycle of "assessment-intervention-reassessment".

[0005] Therefore, there is an urgent clinical need for a palliative care applicability evaluation method that can integrate multi-dimensional temporal data, achieve dynamic and accurate assessment, and intelligently connect with stratified intervention recommendations and efficacy prediction. Summary of the Invention

[0006] This invention aims to address the problems of existing methods for evaluating the applicability of palliative care for cancer patients, which rely on static, single-dimensional assessments, are unable to capture the dynamic evolution of the disease, and are disconnected from stratified intervention decisions.

[0007] To address the aforementioned technical problems, the present invention discloses a method for evaluating the suitability of palliative care for cancer patients based on integrated temporal forest strategy optimization, characterized by comprising the following steps: Step 1: Collect patient information data and corresponding palliative care applicability evaluation levels to obtain patient sample data; Step 2: Perform feature engineering on each patient sample data to extract features and construct a feature set. For any patient... The patient sample data extracted features include: Static features ; Timing-based physiological and psychological symptom score , , Indicates the patient The time of diagnosis This indicates the update cycle for the palliative care suitability assessment level. Indicates the number of data collection cycles; Palliative care applicability rating ; Temporal dynamic characteristics ; Time characteristics This indicates the time since diagnosis. The number of cycles; Step 3: Using the feature set constructed in Step 2, train an ensemble temporal forest classifier. It is used for baseline classification of palliative care applicability assessment levels at the current time, and integrates a temporal forest classifier. The inputs are static features, temporal physiological and psychological symptom scores, temporal dynamic features, and time features. The output is a predicted value for the palliative care applicability rating. Always for the patient The integrated temporal forest classifier The output predicted value is represented as ; Step 4: Based on the feature set constructed in Step S2, build a state transition dataset, train a multi-output ensemble tree regression model, and predict the patient's state change at the next time point given the current state and the palliative care applicability rating. The inputs to the multi-output ensemble tree regression model are static features, the current period's temporal physiological and psychological symptom scores, symptom trend features, and the current period's data passed through the ensemble temporal forest classifier. The predicted values ​​obtained The changes in the palliative care applicability assessment level are also analyzed, and the output is used as a predictive value for the time-series physiological and psychological symptom scores for the next cycle. Among them, symptom trend characteristics , The direction of change in the palliative care applicability rating is the change in the current period's palliative care applicability rating compared to the previous week's palliative care applicability rating, based on the previous period's temporal physiological and psychological symptom scores. Step 5: An ensemble temporal forest classifier based on step 3. In conjunction with the multi-output ensemble tree regression model from step 4, a discrete state space and action space are defined. A reward function that integrates symptom improvement, intervention appropriateness, and resource optimization is designed. A simulated transfer dataset is generated, and a policy optimizer is trained using the Q-learning algorithm to obtain the optimal level adjustment policy. This is then integrated with an ensemble temporal forest classifier. Weighted fusion is used to achieve a recommendation for the applicability evaluation level of palliative care under clinical safety constraints.

[0008] Preferably, in step 1, the patient information data includes demographic information, diagnostic information, and medical history information.

[0009] Preferably, in step 2, the static feature ,in, Indicates the patient age, Indicates the patient gender, Indicates the patient Types of cancer diagnosis Indicates the patient Types of comorbidities.

[0010] Preferably, in step 2, the time-series dynamic features ,in: Indicates the patient In time The scores of physiological and psychological symptoms compared to the previous time point The difference in physiological and psychological symptom scores; Indicates the patient recent Mean scores of physiological and psychological symptoms at each time point; Indicates the patient In time The number of symptoms improved by comparing the physiological and psychological symptom scores at the previous time point; Indicates the patient At the point of time The number of symptoms that worsened was compared with the previous time point in terms of the scores of physiological and psychological symptoms. Indicates the patient In time The highest score on the physiological and psychological symptom assessment; Indicates the patient In time The number of symptoms with a score greater than 5 in the physiological and psychological symptom scores.

[0011] Preferably, in step 3, the integrated temporal forest classifier is... During training, the objective is to minimize the weighted cross-entropy loss function:

[0012] in: Indicate category The weights are different, and different categories represent different levels of palliative care applicability evaluation. Indicates an indicator function; This represents the probability distribution predicted by the ensemble tree model; This represents the true value of the palliative care suitability rating.

[0013] Preferably, in step 3, during the inference phase, the integrated temporal forest classifier is... The output prediction results are subject to time smoothing constraints: if , The integrated temporal forest classifier is then obtained after post-processing adjustments. Output .

[0014] Preferably, in step 4, when training the multi-output ensemble tree regression model, the objective is to minimize the multi-output mean square error.

[0015] In the formula, This represents the true value of the time-series physiological and psychological symptom scores.

[0016] Preferably, in step 4, during the inference phase, the predicted values ​​output by the multi-output ensemble tree regression model are... Nearest neighbor mapping, where This represents the number of dimensions of patient symptoms and signs, and the Pearson correlation coefficient matrix between each symptom is pre-calculated from the training data. Based on the Pearson correlation coefficient matrix For strongly correlated symptom pairs, if the direction of increase or decrease in the prediction results significantly deviates from the historical correlation, then fine-tuning is performed.

[0017] Preferably, step 5 includes the following steps: Step 501: Define the reinforcement learning environment and the state and action space: The patient's current state is represented as a discrete state. ,in: Indicates the severity of symptoms; This indicates the applicability ratings for the three most recent palliative care assessments, with each component having a specific value. ; Define the action space, including: Action set Where: 0 means "maintain the current rating", 1 means "upgrade", and 2 means "downgrade"; Evaluation level at the next time point after the action is performed. From the current level and actions Joint decision:

[0018] Step 502: Immediate Reward Defined as a weighted sum of three components:

[0019] in: , , These are adjustable weighting coefficients; Rewards for symptom improvement Used to evaluate intervention outcomes:

[0020] In the formula, Symptom levels after intervention Symptom levels before intervention; Appropriate rewards for intervention, used to evaluate decision quality: like , ; like , ; like , ; Resource optimization rewards are used to evaluate resource utilization. If the action is a downgrade and the symptom level does not worsen before and after the intervention. (Saving medical resources); If the action is a downgrade and the symptom severity worsens after the intervention compared to before and after the intervention, ; Other situations ; Step 503: Integrated Temporal Forest Classifier Based on Step 3 Together with the multi-output ensemble tree regression model from step 4, training data is generated, and for each sample in the training data, the actual state is... Starting from this point, try all the actions in sequence. The state transition predictor is used to predict the symptom score at the next time point, and then the next symptom level is calculated. And calculate instant rewards To form a quadruple All generated quadruplets Constructing a reinforcement learning training set ; Step 504, Q-learning training: Initialize action value function ; For each training epoch, from the reinforcement learning training set Randomly sampled mini-batch quadruples And update :

[0021] in, For learning rate, As a discount factor, Indicates the reward value. Indicates the value of an action; During training, using probability Randomly select actions, with probability Select Current The action with the highest value The value gradually decreases from the initial value to the minimum value as the number of training rounds increases; training stops when the strategy no longer changes significantly in several consecutive rounds, or when the preset maximum number of training rounds is reached. After training, the optimal policy is obtained: ; Step 505, for new patients At the point of time First, the ensemble temporal forest classifier from step 3 of step S3 is used. Get the initial recommendation level At the same time, based on the current state Query Optimal Strategy Get reinforcement learning recommended actions Then get the recommendation level. ; Final decision level The decision was made by weighted voting.

[0022] In the formula, For balance coefficient, This represents the rounding function.

[0023] This invention integrates patients' temporal physiological and psychological symptom scores, static clinical characteristics, and historical intervention records to construct an ensemble temporal forest model. This model enables accurate and interpretable baseline prediction of the palliative care level corresponding to the patient's current state, achieving precise baseline prediction of the current palliative care suitability level. Simultaneously, a state transition predictor is constructed using a multi-output ensemble tree regression model to dynamically simulate the evolution trajectory of patient symptoms under different intervention strategies. Building upon this, a reinforcement learning mechanism is introduced, with long-term quality of life improvement as the optimization goal. Combining multi-dimensional reward functions such as symptom improvement, intervention suitability, and resource conservation, the recommended intervention strategy is simulated and optimized to learn the optimal dynamic level adjustment strategy. This strategy is then weighted and fused with the baseline classifier, supplemented by clinical safety constraints, forming a data-driven, dynamically evaluated, and closed-loop optimized intelligent decision support method. This improves the accuracy of palliative care resource allocation, the timeliness of intervention measures, and the long-term benefits to patients' quality of life.

[0024] The technical solution provided by this invention effectively overcomes the limitations of existing methods, such as reliance on static single-point assessment, lack of longitudinal trend analysis, and disconnection from intervention decisions. It significantly improves the accuracy of allocating allocative medical resources and the timeliness of intervention measures. At the same time, it has the advantages of strong model interpretability, moderate training data requirements, and ease of deployment and promotion in clinical practice. Attached Figure Description

[0025] Figure 1 This is a flowchart illustrating the method disclosed in an embodiment of the present invention. Detailed Implementation

[0026] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined by the appended claims.

[0027] like Figure 1 As shown, this invention discloses a method for evaluating the applicability of palliative care for cancer patients based on integrated temporal forest strategy optimization, specifically including the following steps: S1. Use cancer patient assessment scales to screen for symptoms and signs caused by the disease. Connect to the hospital information system to collect information from self-reported questionnaires by cancer patients, as well as relevant demographic information, diagnostic information, medical history information, and historical palliative care applicability evaluations.

[0028] In one embodiment of the present invention, an optional implementation is that step S1 further includes the following steps: S1.1 Use the Cancer Patient Assessment Scale to screen for symptoms and signs caused by the disease. The scale includes... This is a multiple-choice question; you must select a score from 0 to 10. Example (ESAS-r scale, (See Table 1.) Connect to the hospital information system to collect information from self-administered questionnaires by cancer patients.

[0029] Table 1: Examples of self-reported assessment scales by cancer patients (ESAS-r scale)

[0030] S1.2 connects to the hospital information system, linking relevant patient information through patient ID and visit number, including: Demographic information, including age, gender, etc.; Cancer diagnosis information, including diagnosis name, diagnosis code, department name, etc.; Medical history information, including information on comorbidities, etc.; Historical assessment of palliative care suitability (judged by experienced palliative care team physicians). Palliative care suitability is assessed for patients with advanced cancer based on a comprehensive evaluation of multiple dimensions, including symptom burden, functional status, psychosocial needs, comorbidities, and disease prognosis, to determine the need for palliative care intervention and the urgency and intensity of such intervention. Its core principle is to accurately identify patient needs throughout the entire disease lifecycle (especially in the early stages) to ensure timely and appropriate allocation of palliative care resources.

[0031] The palliative care suitability assessment level (Table 2) is updated on the following schedule: The evaluation period is generally one month, and the evaluation level is 0, 1, 2, or 3.

[0032] Table 2: Grading Framework for Palliative Care Applicability Assessment

[0033] S2. Preprocess the data collected in step S1, including data cleaning, data aggregation and filtering, feature engineering, etc.

[0034] In one embodiment of the present invention, an optional implementation is that step S2 further includes the following steps: S2.1 Preprocesses the data collected in step S1, including: Data cleaning: reviewing, verifying, deleting, or filling in duplicate, missing, invalid, and erroneous values; Data aggregation and filtering: Various information is associated with patient ID and consultation number. A set of data, identified by patient ID and consultation number and featuring relevant patient information as characteristic variables, is treated as a sample. Data aggregation and filtering are then performed to obtain the following results. Patient sample data.

[0035] S2.2 is the data collected in step S2.1. Feature engineering was performed on patient sample data, including: (1) Constructing static features ,in: Indicates the patient Age can be binned and coded, for example, age can be binned into... And label encoding is performed. ; Indicates the patient gender, 0 represents female, and 1 represents male; Indicates the patient The cancer diagnosis type can be uniquely encoded. Indicates the patient The types of comorbidities can be uniquely coded.

[0036] In one embodiment of the present invention, an optional implementation is: cancer diagnosis type. This can include: head and neck tumors (such as nasopharyngeal carcinoma), respiratory system tumors (such as lung adenocarcinoma), digestive system tumors (such as colorectal cancer), breast tumors (such as breast cancer), genitourinary system tumors (such as prostate cancer), gynecological tumors (such as endometrioid carcinoma), endocrine and neuroendocrine tumors (such as papillary thyroid carcinoma), bone and soft tissue sarcomas (such as liposarcoma), skin and melanoma (such as basal cell carcinoma), central nervous system tumors (such as glioblastoma), and hematopoietic and lymphatic system malignancies (such as multiple myeloma).

[0037] In one embodiment of the present invention, an optional implementation is that the type of comorbidity... These can include: cardiovascular complications (such as hypertension), metabolic-endocrine complications (such as diabetes), respiratory-thoracic complications (such as asthma), renal-urinary complications (such as benign prostatic hyperplasia), and neuropsychiatric complications (such as Parkinson's disease). (2) Collect patients' temporal physiological and psychological symptom scores That is, the results of the self-administered scale in step S1.1, including the patient's diagnosis time. and subsequent The results of the self-administered questionnaire, including: This indicates the update cycle for the self-administered questionnaire (the same as the update cycle for the palliative care applicability assessment level mentioned above), which is generally one month. Indicates the number of data collection cycles; This indicates the number of dimensions of patient symptoms and signs, for example, the ESAS-r cancer patient assessment scale. Therefore, for Data from a sample of patients, including time-series physiological and psychological symptom scores. Each score; (3) Collect patient history of palliative care applicability assessment level , , ; (4) Constructing temporal dynamic features ,in: Indicates the patient In time The scores of physiological and psychological symptoms compared to the previous time point The difference in physiological and psychological symptom scores; if Then use filling; Indicates the patient recent The average scores of physiological and psychological symptoms at each time point; if insufficient If the current time point is specified, then the current score will be used. (Count of getting better) indicates the patient In time The scores of physiological and psychological symptoms were compared with the previous time point. The number of symptoms that improved with the physiological and psychological symptom scores (i.e., the number of symptoms whose scores decreased); if Then use filling; (Count of getting Worse) indicates the patient At the point of time The scores of physiological and psychological symptoms were compared with the previous time point. The number of symptoms that worsened due to the physiological and psychological symptom scores (i.e., the number of symptoms whose scores increased); if Then use filling; Indicates the patient In time The highest score on the physiological and psychological symptom assessment; Indicates the patient In time The number of symptoms with a score greater than 5 in the physiological and psychological symptom scores.

[0038] (5) Time characteristics Indicates the time since diagnosis The number of cycles.

[0039] S3. Based on the feature set constructed in step S2, train an integrated temporal forest classifier to perform baseline classification of the applicability evaluation level of palliative care at the current time.

[0040] In one embodiment of the present invention, an optional implementation is that step S3 further includes the following steps: S3.1 Construct the patient according to step S2 At the point of time training samples Each training sample include: Input variables: static features Physiological and psychological symptom scores Temporal dynamic characteristics Time characteristics ; And, output variable: Palliative care suitability rating. ; A patient-grouping strategy was adopted to divide the patients into groups. All training samples corresponding to the patient sample data were divided into training patient groups. Validation patient group Test patient group Each group includes all time points. The data.

[0041] In one embodiment of the present invention, an optional implementation is to train the patient group. Validation patient group Test patient group The division ratio can be , , .

[0042] S3.2 Employs the ensemble tree model algorithm, training the model based on the training data from step S3.1 to construct a baseline-level classifier. The objective is to minimize the weighted cross-entropy loss function:

[0043] in, Indicate category The weights, and the training set Medium category The proportion is inversely proportional. Indicates an indicator function, This represents the probability distribution predicted by the ensemble tree model.

[0044] In one embodiment of the present invention, an optional implementation method is that the ensemble tree model algorithm can be one of Random Forest, Gradient Boosting Decision Tree (GBDT), XGBoost, LightGBM, CatBoost, AdaBoost, etc.

[0045] S3.3 performs grouped K-fold cross-validation, calculating the evaluation metric for each fold on the validation set. During cross-validation, a hyperparameter grid search is performed, and based on the cross-validation results, the hyperparameter combination with the highest mean evaluation metric on the validation set is selected. The selected hyperparameter combination is then used for the training data. Retrain the final baseline classification model and on the validation set. The final model performance is then evaluated.

[0046] In one embodiment of the present invention, an optional implementation method is to include at least one of the following evaluation metrics: accuracy, Macro-F1, Micro-F1, weighted recall.

[0047] In one embodiment of the present invention, an optional implementation method is to perform a hyperparameter grid search based on the number of decision trees, the maximum depth (max_depth), etc.

[0048] In one embodiment of the present invention, an optional implementation is to select the hyperparameter combination with the highest average Macro-F1 on the validation set based on the cross-validation results.

[0049] In the S3.4 prediction phase, a time smoothing constraint is applied to avoid abrupt changes in the applicability evaluation level between adjacent time points. If , Then, the model output is obtained after post-processing adjustments. , .

[0050] S4. Based on the feature set constructed in step S2, construct a state transition dataset, train a multi-output ensemble tree regression model, and predict the patient's state change at the next time point given the current state and the palliative care applicability rating.

[0051] In one embodiment of the present invention, an optional implementation is that step S4 further includes the following steps: S4.1 Construct a state transition dataset to learn how a patient's state changes at the next time point, given the current state and the palliative care applicability rating.

[0052] In one embodiment of the present invention, an optional implementation is that each sample data in the state transition dataset contains the following information: Input variables include: static features Timing-based physiological and psychological symptom score Symptom trend characteristics Palliative care applicability assessment prediction level (Obtained through step S3), direction of change in palliative care suitability evaluation level (upgrade=1, downgrade=-1, maintain=0). And output variables, including: the patient's physiological and psychological symptom scores at the next time point. .

[0053] In one embodiment of the present invention, an optional implementation is that the value of the change direction of the palliative care suitability evaluation level is taken as 1, -1, or 0, where 1 represents "upgrade", -1 represents "downgrade", and 0 represents "maintain".

[0054] A patient-grouping strategy was adopted to divide the patients into groups. Patient sample data were divided into training patient group. Validation patient group Test patient group Each group includes all time points. The data.

[0055] In one embodiment of the present invention, an optional implementation is to train the patient group. Validation patient group Test patient group The division ratio can be , , .

[0056] S4.2 employs a multi-output ensemble tree regression model, training the regression model based on the training data from step S4.1 to construct a state transition predictor. The objective is to minimize the mean square error (MSE) of the multiple outputs: .

[0057] In one embodiment of the present invention, an optional implementation is to use minimizing the mean square error of multiple outputs (MSE) as the basic objective function, while employing mean absolute error (MAE) or Huber loss to enhance robustness.

[0058] In one embodiment of the present invention, an optional implementation method is that the multi-output ensemble tree regression model is one of the following: MultiOutput Random Forest Regressor, MultiOutput Gradient Boosting Regressor, LightGBM, CatBoost (which supports multiple outputs and can be set to objective='regression', multi_output=True).

[0059] S4.3 performs grouped K-fold cross-validation, calculating the evaluation metric for each fold on the validation set. During cross-validation, a hyperparameter grid search is performed, and based on the cross-validation results, the hyperparameter combination with the minimum mean evaluation metric on the validation set is selected. The selected hyperparameter combination is then used for the training data. Retrain the final baseline classification model and on the validation set. The final model performance is then evaluated.

[0060] In one embodiment of the present invention, an optional implementation method is that the evaluation indicators include at least one of the following: mean squared error (MSE), mean absolute error (MAE), symptom level prediction accuracy (the proportion of predicted continuous scores that differ from the actual scores by ≤1 after rounding to the nearest integer, i.e., tolerance for 1-point error), and directional consistency (the proportion of predicted change direction that is consistent with the actual change direction).

[0061] In one embodiment of the present invention, an optional implementation method is to perform hyperparameter grid search based on the number of decision trees, the maximum depth (max_depth), etc.

[0062] In one embodiment of the present invention, an optional implementation is to select the hyperparameter combination that minimizes the average MSE on the validation set based on the cross-validation results. In the S4.4 prediction phase, given the current input, the predicted values ​​output by the model in step S4.3 are... Nearest neighbor mapping is used. It can also perform correlation corrections between physiological and psychological symptom scores to improve prediction accuracy. The Pearson correlation coefficient matrix between each symptom is pre-calculated from the training data. For strongly correlated symptom pairs (such as fatigue and drowsiness, anxiety and depression), if the direction of increase or decrease in the prediction results significantly deviates from historical correlation, fine-tuning can be performed. For example, for each symptom dimension, the predicted average of its strongly correlated symptom groups can be used as the final output.

[0063] S5. Based on the baseline classifier in step S3 and the state transition predictor in step S4, define the discrete state space and action space (maintain, upgrade, downgrade), design a reward function that integrates symptom improvement, intervention suitability, and resource optimization, generate a simulated transition dataset, train the policy optimizer using the Q-learning algorithm, obtain the optimal level adjustment policy, and weight and fuse it with the baseline classifier to achieve a recommendation of palliative care suitability evaluation level under clinical safety constraints.

[0064] In one embodiment of the present invention, an optional implementation is that step S5 further includes the following steps: S5.1 defines the reinforcement learning environment and the state and action space, further including: S5.1.1 State-space discretization: Representing the patient's current state as a discrete state. ,in: The severity level can be indicated by the patient's total score on the scale at the current time point. Obtained through mapping. Assume... One mapping relationship is as follows: grade (Mild):

[0065] grade (Moderate):

[0066] grade (Severe):

[0067] grade (Extremely severe):

[0068] This indicates the grades of the three most recent palliative care suitability assessments (arranged from oldest to most recent), with each component having a value. For the initial assessment at the time of diagnosis, missing historical grades are filled with 0.

[0069] S5.1.2 Action space definition, including: (1) Action set ,in: 0 indicates "Maintain the current rating level". 1 indicates "upgrade" (increase the level by one level, with a maximum of 3); 2 indicates "downgrade" (lowering the level by one level, with a minimum of 0); (2) Evaluation level at the next time point after the action is performed. From the current level and actions Joint decision:

[0070] The S5.2 reward function design further includes: S5.2.1 Instant Rewards Defined as a weighted sum of three components:

[0071] in: , , These are adjustable weighting coefficients; Rewards for symptom improvement Used to evaluate intervention outcomes (whether symptoms have improved):

[0072] That is, a decrease (improvement) of 1 level in symptom severity is rewarded with 10 points, and an increase (worsening) of 1 level is penalized with 10 points; The appropriateness reward for intervention is used to evaluate the quality of decision-making (whether the intervention is suitable for the current state): Rewards will be given based on the degree of match between the intervention level and the symptom level: if (Complete match) ; if (The intervention level differs from the symptom level by 1 level). ; if (The intervention level and the symptom level are severely mismatched.) ; Resource optimization rewards are used to evaluate resource utilization. If the action is a downgrade ( And the condition has not worsened. ), (Saving medical resources); If the action is a downgrade ( And the condition is deteriorating. ), (Improper downgrading led to worsening of symptoms); Other situations .

[0073] S5.3 Baseline-level classifier trained in step S3 and the state transition predictor obtained from step S4 training Generate training data. For the training patient group... Each patient In actual condition Starting from this point, try all the actions in sequence. The state transition predictor is used to predict the symptom score at the next time point, and then the next symptom level is calculated. And calculate the instant reward according to step S5.2. To form a quadruple All generated quadruples Constructing a reinforcement learning training set .

[0074] In one embodiment of the present invention, an optional implementation is to enhance the completeness of the training data by performing state-action pairs that have not appeared in historical data. The state transition predictor actively generates simulated samples and fills the reinforcement learning training set. middle.

[0075] S5.4 Q-learning training further includes: Initialize action value function : Use optimistic initialization, assigning high initial values ​​to all possible state-action pairs (e.g., To encourage exploration; For each training epoch, from the reinforcement learning training set Randomly sampled mini-batch quadruples And update :

[0076] in, For the learning rate (e.g.) ), Discount factor (e.g.) This indicates the importance of future rewards. Indicates the reward value. Indicates the value of an action; During training, using probability Randomly select actions, with probability Select Current The action with the highest value As the number of training epochs increases, from the initial value (e.g.) ) gradually decays to a minimum value (e.g. When the strategy no longer changes significantly over several consecutive rounds (e.g., 50 rounds) (e.g.) Training will stop when the preset maximum number of training rounds (e.g., 2000 rounds) is reached.

[0077] After training, the optimal policy is obtained:

[0078] That is, for each discrete state Recommend the optimal action (maintain, upgrade, or downgrade).

[0079] S5.5 for new patients At the point of time First, the initial recommendation level is obtained using the baseline classifier from step S3. At the same time, based on the current state Query Optimal Strategy Get reinforcement learning recommended actions Then get the recommendation level. (Calculated through action conversion rules).

[0080] Final decision level The decision was made by weighted voting.

[0081] in, This is a balance factor and can be adjusted based on clinical trust levels. This represents the rounding function.

[0082] In one embodiment of the present invention, an optional implementation involves imposing clinical safety constraints, including: If a patient's symptom score deteriorates sharply within adjacent time points (e.g.) If the recommendation level rises by more than 2 levels, the recommendation level will be forcibly raised to the highest level 3. To avoid frequent adjustments, if the recommended level is the same as the previous time point and the symptoms are stable, the current level should be maintained first.

[0083] Those skilled in the art will understand that the features described in the various embodiments and / or claims of this invention can be combined or combined in various ways, even if such combinations or combinations are not explicitly described in this disclosure. In particular, the features described in the various embodiments and / or claims of this invention can be combined or combined in various ways without departing from the spirit and teachings of this disclosure. All such combinations and / or combinations fall within the scope of this disclosure.

[0084] The embodiments of the present invention have been described above. However, these embodiments are for illustrative purposes only and are not intended to limit the scope of this disclosure. Although various embodiments have been described above, this does not mean that the measures in the various embodiments cannot be used advantageously in combination. The scope of the present invention is defined by the appended claims and their equivalents. Various substitutions and modifications can be made by those skilled in the art without departing from the scope of the present invention, and all such substitutions and modifications should fall within the scope of this disclosure.

Claims

1. A method for evaluating the applicability of palliative care for cancer patients based on ensemble temporal forest strategy optimization, characterized in that, Includes the following steps: Step 1: Collect patient information data and corresponding palliative care applicability evaluation levels to obtain patient sample data; Step 2: Perform feature engineering on each patient sample data to extract features and construct a feature set. For any patient... The patient sample data extracted features include: Static features ; Timing-based physiological and psychological symptom score , , Indicates the patient The time of diagnosis This indicates the update cycle for the palliative care suitability assessment level. Indicates the number of data collection cycles; Palliative care applicability rating ; Temporal dynamic characteristics ; Time characteristics This indicates the time since diagnosis. The number of cycles; Step 3: Using the feature set constructed in Step 2, train an ensemble temporal forest classifier. It is used for baseline classification of palliative care applicability assessment levels at the current time, and integrates a temporal forest classifier. The inputs are static features, temporal physiological and psychological symptom scores, temporal dynamic features, and time features. The output is a predicted value for the palliative care applicability rating. Always for the patient The integrated temporal forest classifier The output predicted value is represented as ; Step 4: Based on the feature set constructed in Step S2, build a state transition dataset, train a multi-output ensemble tree regression model, and predict the patient's state change at the next time point given the current state and the palliative care applicability rating. The inputs to the multi-output ensemble tree regression model are static features, the current period's temporal physiological and psychological symptom scores, symptom trend features, and the current period's data passed through the ensemble temporal forest classifier. The predicted values ​​obtained The changes in the palliative care applicability assessment level are also analyzed, and the output is used as a predictive value for the time-series physiological and psychological symptom scores for the next cycle. Among them, symptom trend characteristics , The direction of change in the palliative care applicability rating is the change in the current period's palliative care applicability rating compared to the previous week's palliative care applicability rating, based on the previous period's temporal physiological and psychological symptom scores. Step 5: An ensemble temporal forest classifier based on step 3. In conjunction with the multi-output ensemble tree regression model from step 4, a discrete state space and action space are defined. A reward function that integrates symptom improvement, intervention appropriateness, and resource optimization is designed. A simulated transfer dataset is generated, and a policy optimizer is trained using the Q-learning algorithm to obtain the optimal level adjustment policy. This is then integrated with an ensemble temporal forest classifier. Weighted fusion is used to achieve a recommendation for the applicability evaluation level of palliative care under clinical safety constraints.

2. The method for evaluating the applicability of palliative care for cancer patients based on ensemble temporal forest strategy optimization as described in claim 1, characterized in that, In step 1, the patient information data includes demographic information, diagnostic information, and medical history information.

3. The method for evaluating the applicability of palliative care for cancer patients based on ensemble temporal forest strategy optimization as described in claim 1, characterized in that, In step 2, the static features ,in, Indicates the patient age, Indicates the patient gender, Indicates the patient Types of cancer diagnosis Indicates the patient Types of comorbidities.

4. The method for evaluating the applicability of palliative care for cancer patients based on ensemble temporal forest strategy optimization as described in claim 1, characterized in that, In step 2, the time-series dynamic features ,in: Indicates the patient In time The scores of physiological and psychological symptoms compared to the previous time point The difference in physiological and psychological symptom scores; Indicates the patient recent Mean scores of physiological and psychological symptoms at each time point; Indicates the patient In time The number of symptoms improved by comparing the physiological and psychological symptom scores at the previous time point; Indicates the patient At the point of time The number of symptoms that worsened was compared with the previous time point in terms of the scores of physiological and psychological symptoms. Indicates the patient In time The highest score on the physiological and psychological symptom assessment; Indicates the patient In time The number of symptoms with a score greater than 5 in the physiological and psychological symptom scores.

5. The method for evaluating the applicability of palliative care for cancer patients based on ensemble temporal forest strategy optimization as described in claim 1, characterized in that, In step 3, the integrated temporal forest classifier is... During training, the objective is to minimize the weighted cross-entropy loss function: in: Indicate category The weights are different, and different categories represent different levels of palliative care applicability evaluation. Indicates an indicator function; This represents the probability distribution predicted by the ensemble tree model; This represents the true value of the palliative care suitability rating.

6. The method for evaluating the applicability of palliative care for cancer patients based on ensemble temporal forest strategy optimization as described in claim 1, characterized in that, In step 3, during the inference phase, the integrated temporal forest classifier is... The output prediction results are subject to time smoothing constraints: if , The integrated temporal forest classifier is then obtained after post-processing adjustments. Output .

7. The method for evaluating the applicability of palliative care for cancer patients based on ensemble temporal forest strategy optimization as described in claim 1, characterized in that, In step 4, when training the multi-output ensemble tree regression model, the objective is to minimize the multi-output mean square error: In the formula, This represents the true value of the time-series physiological and psychological symptom scores.

8. The method for evaluating the suitability of palliative care for cancer patients based on ensemble temporal forest strategy optimization as described in claim 1, characterized in that, In step 4, during the inference phase, the predicted values ​​output by the multi-output ensemble tree regression model are processed. Nearest neighbor mapping, where This represents the number of dimensions of patient symptoms and signs, and the Pearson correlation coefficient matrix between each symptom is pre-calculated from the training data. Based on the Pearson correlation coefficient matrix For strongly correlated symptom pairs, if the direction of increase or decrease in the prediction results significantly deviates from the historical correlation, then fine-tuning is performed.

9. The method for evaluating the applicability of palliative care for cancer patients based on ensemble temporal forest strategy optimization as described in claim 1, characterized in that, Step 5 includes the following steps: Step 501: Define the reinforcement learning environment and the state and action space: The patient's current state is represented as a discrete state. ,in: Indicates the severity of symptoms; This indicates the applicability ratings for the three most recent palliative care assessments, with each component having a specific value. ; Define the action space, including: Action set Where: 0 means "maintain the current rating", 1 means "upgrade", and 2 means "downgrade"; Evaluation level at the next time point after the action is performed. From the current level and actions Joint decision: Step 502: Immediate Reward Defined as a weighted sum of three components: in: , , These are adjustable weighting coefficients; Rewards for symptom improvement Used to evaluate intervention outcomes: In the formula, Symptom levels after intervention Symptom levels before intervention; Appropriate rewards for intervention, used to evaluate decision quality: like , ; like , ; like , ; Resource optimization rewards are used to evaluate resource utilization. If the action is a downgrade and the symptom level does not worsen before and after the intervention. (Saving medical resources); If the action is a downgrade and the symptom severity worsens after the intervention compared to before and after the intervention, ; Other situations ; Step 503: Integrated Temporal Forest Classifier Based on Step 3 Together with the multi-output ensemble tree regression model from step 4, training data is generated, and for each sample in the training data, the actual state is... Starting from this point, try all the actions in sequence. The state transition predictor is used to predict the symptom score at the next time point, and then the next symptom level is calculated. And calculate instant rewards To form a quadruple All generated quadruplets Constructing a reinforcement learning training set ; Step 504, Q-learning training: Initialize action value function ; For each training epoch, from the reinforcement learning training set Randomly sampled mini-batch quadruples And update : in, For learning rate, As a discount factor, Indicates the reward value. Indicates the value of an action; During training, using probability Randomly select actions, with probability Select Current The action with the highest value The value gradually decreases from the initial value to the minimum value as the number of training rounds increases; training stops when the strategy no longer changes significantly in several consecutive rounds, or when the preset maximum number of training rounds is reached. After training, the optimal policy is obtained: ; Step 505, for new patients At the point of time First, the ensemble temporal forest classifier from step 3 of step S3 is used. Get the initial recommendation level At the same time, based on the current state Query Optimal Strategy Get reinforcement learning recommended actions Then get the recommendation level. ; Final decision level The decision was made by weighted voting. In the formula, For balance coefficient, This represents the rounding function.

Images