Individualized drug recommendation system for tumor patients and method thereof

Abstract

The application relates to the field of medical information technology, in particular to a tumor patient individualized drug recommendation system and a method thereof, the system comprising: a doctor module, which acquires patient case text description, extracts case features by using a bag-of-words model, and extracts features by using a convolutional neural network to obtain a case feature vector; a keyword semantic feature vector is constructed based on the case feature vector; cosine similarity calculation is performed on the keyword semantic feature vector and each to-be-matched item in clinical trial data; a drug recommendation module: similarity calculation is performed according to a case matching list and drug efficacy data, and an attention mechanism is applied to screen out the to-be-matched item with the highest similarity in the case matching list, acquire recommended drugs and efficacy evaluation of the to-be-matched item, in the matching process, the system skillfully uses cosine similarity calculation and the attention mechanism, not only can the case with the highest similarity be found out, but also intelligent weighting can be performed according to the importance of different features.

Description

Technical Field

[0001] This invention relates to the field of medical information technology, specifically to a personalized medication recommendation system and method for cancer patients. Background Technology

[0002] With the rapid development of medical technology, the field of cancer treatment is undergoing an unprecedented transformation. The concept of personalized medicine is gaining increasing popularity, emphasizing the development of the best treatment plan based on each patient's unique circumstances. However, achieving truly personalized medicine still faces many challenges.

[0003] Currently, medication regimens for cancer patients primarily rely on clinical guidelines and physicians' personal experience. While this approach is simple and direct, it often fails to adequately consider individual patient differences. For example, two patients with the same non-small cell lung cancer may respond vastly to the same drug due to differences in age, physical condition, and gene mutation type. Traditional methods struggle to accurately predict these differences, leading to unsatisfactory treatment outcomes and potentially even serious adverse reactions.

[0004] On the other hand, with the rise of precision medicine, personalized treatment plans based on gene testing are beginning to emerge. While this approach takes into account the patient's genetic background, it still has limitations. It primarily focuses on single gene mutations, neglecting the patient's overall clinical characteristics and past treatment history. Furthermore, the high cost of gene testing also limits its widespread application.

[0005] In recent years, the application of artificial intelligence (AI) technology in the medical field has attracted widespread attention. Some studies have attempted to use machine learning algorithms to analyze large amounts of medical data to provide doctors with decision-making support. However, most of these systems remain at the data analysis level and struggle to truly achieve personalized medication recommendations. They often only provide general statistical results and cannot offer precise advice tailored to the complex conditions of specific patients.

[0006] Furthermore, most existing intelligent assistance systems operate on a black-box model, lacking interpretability for their recommendations. This not only undermines doctors' trust in the system but also increases the risk of medical decisions. In the medical field, which involves life and health, every decision should be based on sufficient evidence and clear explanation.

[0007] To address these challenges, there is an urgent need for intelligent systems capable of truly personalized medication recommendations. This system should be able to comprehensively consider a patient's various clinical characteristics, genetic information, past treatment history, and other multi-dimensional data, combined with the latest clinical trial results, to provide each patient with the optimal medication plan. Simultaneously, the system should also be highly interpretable, enabling physicians to understand and trust its recommendations. Summary of the Invention

[0008] This invention addresses the aforementioned technical problems. It aims to develop an artificial intelligence-based personalized medication recommendation system for cancer patients, capable of comprehensively analyzing patient information, accurately matching clinical trial data, and generating personalized medication suggestions. This system not only considers the patient's gene mutation status but also fully utilizes various clinical features in the case description, achieving more accurate patient-trial matching through advanced natural language processing and machine learning algorithms.

[0009] This invention proposes a personalized medication recommendation system for cancer patients, characterized in that the system comprises:

[0010] Doctor Module: Obtain patient case text descriptions, extract case features using a bag-of-words model, and perform feature extraction using a convolutional neural network to obtain case feature vectors; construct keyword semantic feature vectors based on case feature vectors; calculate cosine similarity between keyword semantic feature vectors and each item to be matched in clinical trial data, select the top 3-5 items with the highest similarity and their feature vectors, and fuse them to complete the matching of cases with clinical trial data, obtaining a case matching list;

[0011] Medication recommendation module: Calculates similarity based on the case matching list and drug efficacy data, and uses an attention mechanism to filter out the matching items with the highest similarity in the case matching list, and obtains the recommended drug and its efficacy assessment for the matching item.

[0012] Preferably, the doctor module inputs the case text and case feature vector into the medication recommendation module. The medication recommendation module uses a semantically aware vector matching algorithm to automatically match cases in the clinical trial database with the case feature vector and provides a ranking table of case similarity based on the case feature vector. Based on the matching list and combined with drug efficacy data, the module generates recommended drugs and their efficacy assessments.

[0013] Preferably, the doctor module is composed of a doctor module, which provides a text input interface and receives medical records filled in by the doctor. The doctor module is connected to the medication recommendation module. The doctor module inputs text or structured data into the medication recommendation module.

[0014] Preferably, the doctor module includes a case collection and analysis module, a tumor type recommendation module, and a matching list generation module. The case collection and analysis module and the tumor type recommendation module are both connected to the matching list generation module, which is connected to the medication recommendation module. The doctor module also includes a storage unit for storing intermediate data during the case collection, tumor type recommendation, and matching list generation processes. After the case text is collected, it enters the case collection and analysis module. After relevant feature extraction and processing, feature extraction data is obtained. The tumor type recommendation module calculates and processes the feature extraction data to obtain a tumor recommendation type list. The matching list is generated by the matching list generation module and sent to the medication recommendation module.

[0015] Preferably, the case collection and analysis module inputs the case text into the tumor type recommendation module. The tumor type recommendation module uses a bag-of-words model to extract case features and uses a convolutional neural network to extract features to obtain a case feature vector. Based on the case feature vector, a keyword semantic feature vector is constructed. The keyword semantic feature vector is then compared with each item to be matched in the clinical trial data pool using cosine similarity calculation to obtain the matching list. The matching list is then input into the matching list generation module, which in turn inputs the matching list into the medication recommendation module.

[0016] Preferably, the medication recommendation module includes a matching degree calculation unit, a matching degree filtering unit, and an efficacy evaluation unit. The matching degree calculation unit is connected to the matching list generation module, and generates a matching feature vector list from the matching list. The matching degree calculation unit is connected to the matching degree filtering unit, and the matching degree filtering unit is connected to the efficacy evaluation unit. The matching degree calculation unit sends the matching feature vector list to the matching degree filtering unit, and the matching degree filtering unit calculates the similarity based on the matching feature vector list. The matching degree ranking unit calculates the recommendation priority based on the similarity. The efficacy evaluation unit evaluates the efficacy based on the efficacy data of the highest priority drug and generates a recommendation plan. The recommended efficacy data and recommended drug plans are displayed on the treatment recommendation page for doctors to choose from.

[0017] The method for a personalized medication recommendation system for cancer patients includes the following steps:

[0018] Step S1: The doctor module obtains the medical record text and inputs it into the medication recommendation module to obtain the matching list;

[0019] Step S2: Combine the matching list with drug efficacy data to calculate similarity, and apply an attention mechanism to filter out the matching item with the highest similarity in the case matching list, and obtain the recommended drug efficacy assessment for the matching item.

[0020] Preferably, step S1 further includes the following steps: the case text is collected and analyzed by the case collection and analysis module to obtain a case feature vector; a keyword semantic feature vector is constructed based on the case feature vector; the keyword semantic feature vector is used to calculate the similarity with each item to be matched in the clinical trial data, and the top 3-5 items with the highest similarity and their feature vectors are selected and fused to complete the matching of the case and the clinical trial data to obtain the matching list; the case collection and analysis module inputs the case text into the tumor type recommendation module, the tumor type recommendation module uses a bag-of-words model to extract case features and uses a convolutional neural network to extract features to obtain a case feature vector; a keyword semantic feature vector is constructed based on the case feature vector; the keyword semantic feature vector is used to calculate the cosine similarity with each item to be matched in the clinical trial data pool to obtain the matching list.

[0021] Preferably, step S2 further includes the following steps:

[0022] Step S21: Input the matching list into the matching degree calculation unit. The matching degree calculation unit sorts each matching item in the matching list according to its similarity value, sets a threshold, and deletes matching items with a similarity value lower than the threshold, generating the matching feature vector list.

[0023] Step S22: The matching degree filtering unit calculates the similarity based on the matching feature vector list; based on the matching feature vector list, it performs a dot product of the matching feature vectors and keyword semantic feature vectors of all matching items to calculate the similarity value, and sorts them according to the similarity value.

[0024] Step S23: The matching degree ranking unit calculates the recommendation priority based on similarity, obtains the highest priority matching item based on the recommendation priority, acquires the drug efficacy data from the efficacy evaluation unit based on the highest priority matching item, and generates recommended drugs based on the efficacy data, which are then displayed through the treatment recommendation interface; the matching item with the maximum similarity is the matching item with the highest recommendation priority; when the maximum similarity is the same and the maximum similarity is greater than a set threshold, the matching degree filtering unit adds the matching results in each matching feature vector to obtain the total matching degree value, and re-sorts the similarity based on the total matching degree value, with the maximum similarity being the highest priority matching item. The matching item with the highest similarity value is the matching item with the highest recommendation priority. When the maximum matching degree is different, the matching degree filtering unit extracts the vector corresponding to the keyword semantics in the matching feature vector of each matching item and adds them together to obtain the total matching degree value. The similarity is re-sorted according to the total matching degree value, and the matching item with the highest similarity value is the matching item with the highest recommendation priority. The drug efficacy data in the efficacy evaluation unit is obtained according to the highest matching item, and a recommended drug plan is generated according to the efficacy data. The case text with the highest similarity and the recommended drug efficacy evaluation are displayed through the diagnosis and treatment recommendation interface for doctors to select.

[0025] Preferably, the highest priority match is the match with the largest weighted average of similarity and matching score; if the similarity is less than the maximum similarity, the recommended drug for the highest priority match is the clinical trial drug corresponding to the match to be matched for the highest priority match, and the efficacy data in the clinical trial data corresponding to the highest priority match is used to generate the recommended drug regimen.

[0026] The beneficial technical effects of the technical solution of this invention are as follows:

[0027] The core of this invention lies in its unique algorithmic mechanism. First, it employs a bag-of-words model and convolutional neural networks to extract features from case texts, capturing subtle information that traditional methods might overlook. Then, by constructing keyword semantic feature vectors, the system can better understand the semantic content of the cases, improving matching accuracy. During the matching process, the system cleverly utilizes cosine similarity calculation and attention mechanisms, not only identifying the cases with the highest similarity but also intelligently weighting features based on their importance.

[0028] This approach has yielded significant benefits. First, it dramatically improves treatment effectiveness. Through precise matching and personalized recommendations, patients are more likely to receive the most suitable treatment plan, thus increasing the success rate. Second, it effectively reduces the incidence of adverse reactions. The system can predict potential side effects based on massive amounts of data, providing doctors with safer medication options. Third, it significantly improves patients' quality of life. By balancing efficacy and side effects, the system can find the optimal balance for patients, ensuring treatment effectiveness while minimizing the impact on their daily lives.

[0029] More importantly, the system of this invention has the ability to continuously learn and optimize. As the amount of data increases and the algorithm iterates, the accuracy of the system's recommendations will continuously improve. This means that it can not only help current patients but also benefit more patients in the future.

[0030] Furthermore, the system is designed with practical clinical needs in mind, offering excellent interpretability and operability. Doctors can clearly understand the basis for the system's recommendations, which not only enhances their trust in the system but also provides strong support for doctor-patient communication.

[0031] In summary, this invention provides a revolutionary solution for the field of cancer treatment. It perfectly combines artificial intelligence technology with clinical medical knowledge, taking a significant step towards achieving truly personalized precision medicine. Given the current high incidence of cancer, this innovative technology, which can improve treatment outcomes, reduce side effects, and enhance quality of life, undoubtedly has significant social value and application prospects. Attached Figure Description

[0032] Figure 1 This is the overall system logic block diagram of the present invention.

[0033] Figure 2 This is a block diagram of the internal structure logic of the doctor module of the present invention.

[0034] Figure 3 This is a block diagram of the internal structure logic of the medication recommendation module of the present invention.

[0035] Figure 4 This is a flowchart of the method of the present invention.

[0036] Figure 5 This is a block diagram of the matching degree calculation logic of the present invention.

[0037] Figure 6 This is a block diagram of the matching degree filtering logic of the present invention.

[0038] Figure 7 This is a logic diagram for efficacy evaluation. Detailed Implementation

[0039] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the specific implementation methods, structures, features, and effects are described in detail below with reference to the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0040] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0041] See Figure 1-7 This invention proposes a personalized medication recommendation system for cancer patients, which includes a doctor module 1 and a medication recommendation module 2.

[0042] The main function of Doctor Module 1 is to acquire patient case descriptions, extract case features using a bag-of-words model, and then use a convolutional neural network to extract features, resulting in a case feature vector. For example, when a lung cancer patient visits the clinic, the doctor might record in the case file: 65-year-old male, heavy smoking history, chest CT scan showed a 3cm nodule in the upper lobe of the right lung, biopsy confirmed non-small cell lung cancer. Doctor Module 1 will convert this text into a feature vector, such as [0.8, 0.6, 0.9, 0.7, 0.5], where each number represents the weight of key information such as age, gender, smoking history, tumor size, and tumor type.

[0043] Then, based on this case feature vector, the system constructs keyword semantic feature vectors. In one embodiment of the present invention, this process may use natural language processing techniques such as Word2Vec or GloVe to convert keywords such as non-small cell lung cancer and heavy smoking into high-dimensional vectors, for example [0.9, 0.8, 0.7, 0.6, 0.5, ...].

[0044] Next, the system calculates the cosine similarity between this keyword semantic feature vector and each item to be matched in the clinical trial data. The formula for calculating cosine similarity is as follows:

[0045]

[0046] Where A represents the keyword semantic feature vector, B represents the feature vector of the item to be matched, and ||A|| and ||B|| represent the Euclidean norms of vectors A and B, respectively. i B i Let n be the i-th component of vectors A and B, and n be the dimension of the vectors.

[0047] This invention preferably selects the top 3-5 similarity items and their feature vectors for fusion. This selection is based on experience, ensuring sufficient matching accuracy without excessive computational burden. Through this process, the system matches case data with clinical trial data, obtaining a case matching list.

[0048] Finally, the medication recommendation module 2 calculates the similarity based on the case matching list and drug efficacy data, and applies an attention mechanism to filter out the candidate items with the highest similarity in the case matching list, obtaining the recommended drug and its efficacy assessment for that candidate item. The calculation formula for the attention mechanism is as follows:

[0049]

[0050] Where Q is the query matrix, representing case features; K is the key matrix, representing the features of the items to be matched; V is the value matrix, representing the drug information corresponding to the items to be matched; d k The dimension of the key vector; the softmax function is defined as:

[0051]

[0052] Where: x i Let n be the i-th element of the input vector; n is the dimension of the vector.

[0053] The advantage of this method is that it can accurately match similar cases and extract the most relevant drug information, thereby providing doctors with personalized medication advice.

[0054] In one embodiment of the present invention, the doctor module 1 inputs the case text and case feature vector to the medication recommendation module 2. The medication recommendation module 2 uses a semantically aware vector matching algorithm to automatically match cases in the clinical trial database with the case feature vector and provides a ranking table of case similarity based on the case feature vector.

[0055] For example, suppose we have a lung cancer patient's case feature vector [0.8, 0.6, 0.9, 0.7, 0.5]. The system will compare this vector with all cases in the clinical trial database. For instance, the database might have three cases with feature vectors of [0.79, 0.61, 0.88, 0.72, 0.49], [0.5, 0.4, 0.6, 0.3, 0.2], and [0.82, 0.59, 0.91, 0.68, 0.51]. By calculating cosine similarity, the system will rank the similarities: third case > first case > second case.

[0056] Then, based on this matching list and combined with drug efficacy data, the system generates recommended drugs and their efficacy assessments. For example, if a third case uses a certain targeted drug and achieves good results, the system may recommend that drug and provide an expected efficacy assessment.

[0057] The advantage of this approach is that it not only considers the similarity of cases but also combines actual treatment effects, thereby providing more reliable and personalized medication recommendations.

[0058] In one embodiment of the present invention, the doctor module 1 is composed of a terminal that provides a text input interface and receives medical records entered by the doctor. The doctor module is connected to the medication recommendation module 2, and inputs text or structured data into the medication recommendation module 2.

[0059] In practical applications, this might be a software interface on a tablet or desktop computer. Doctors can input case information on this interface, such as patient age: 65 years old, gender: male, chief complaint: cough for 3 months, chest CT scan showing a 3cm nodule in the upper lobe of the right lung, etc. The system will automatically convert this information into structured data and then transmit it to the medication recommendation module 2 for processing.

[0060] The advantage of this design is that it retains the doctor's usual way of working (i.e., free text input) while converting information into a machine-processable format, thus achieving a seamless connection between human and machine collaboration.

[0061] In one embodiment of the present invention, the doctor module 1 includes a case collection and analysis module 11, a tumor type recommendation module 12, and a matching list generation module 13. Both the case collection and analysis module 11 and the tumor type recommendation module 12 are connected to the matching list generation module 13, which in turn is connected to the medication recommendation module 2.

[0062] In addition, the doctor module 1 also includes a storage unit 14 for storing intermediate data during the process of case collection, tumor type recommendation, and matching list generation. The advantage of this design is that it improves the system's processing speed because intermediate results can be cached, eliminating the need for repeated calculations.

[0063] In practice, the case text first enters the case collection and analysis module 11, where feature extraction and processing yield feature extraction data. Then, the tumor type recommendation module 12 calculates and processes this feature extraction data to obtain a list of recommended tumor types. Finally, the matching list generation module 13 generates a matching list and sends it to the medication recommendation module 2.

[0064] For example, suppose a case text describes a patient's symptoms: persistent dry cough, weight loss, and a chest X-ray showing an opacity in the upper lobe of the right lung. The case collection and analysis module 11 will extract keywords such as dry cough, weight loss, and right upper lobe opacity, and convert them into feature vectors. Then, the tumor type recommendation module 12 may recommend lung cancer as a possible tumor type based on these features. Finally, the matching list generation module 13 will search for similar lung cancer cases in the clinical trial database and generate a matching list.

[0065] This modular design enables the system to process complex medical information in a step-by-step and organized manner, thereby improving the system's efficiency and accuracy.

[0066] In one embodiment of the present invention, the entire processing flow from case text to matching list is as follows: First, the case collection and analysis module 11 inputs the case text into the tumor type recommendation module 12. The tumor type recommendation module 12 uses a bag-of-words model to extract case features and uses a convolutional neural network to perform feature extraction to obtain a case feature vector.

[0067] In this process, the bag-of-words model transforms the text into word frequency vectors. For example, for the text "patient has a persistent dry cough with chest pain", the bag-of-words model might generate a vector like this: [1,1,1,1,1,1], where each number represents the number of times the words "patient", "occur", "persistent", "dry cough", "with", and "chest pain" appear.

[0068] The convolutional neural network then processes this vector further to extract higher-level features. For example, it might recognize that persistent dry cough and chest pain are two symptoms that often co-occur in lung cancer cases, and therefore assign higher weights to these two features.

[0069] Next, the system constructs keyword semantic feature vectors based on the case feature vectors. This step may use word embedding techniques, such as Word2Vec or GloVe, to transform each keyword into a high-dimensional vector that can capture the semantic relationships between words.

[0070] Finally, the system calculates the cosine similarity between the keyword semantic feature vector and each item to be matched in the clinical trial data pool to obtain a matching list. This matching list is then input into the matching list generation module 13, which in turn inputs it into the medication recommendation module 2.

[0071] The advantage of this process is that it combines traditional text processing techniques (such as the bag-of-words model) with modern deep learning techniques (such as convolutional neural networks and word embeddings), enabling it to comprehensively and accurately capture key information in case texts, thereby improving matching accuracy.

[0072] In one embodiment of the present invention, the medication recommendation module 2 includes a matching degree calculation unit 21, a matching degree screening unit 22, and an efficacy evaluation unit 23.

[0073] The matching degree calculation unit 21 is connected to the matching list generation module 13, and generates a list of matching feature vectors from the matching list. In practical applications, this unit may use a deep neural network to process the matching list and extract more abstract and high-level features. For example, for a lung cancer case, it may focus on key information such as tumor size, location, and stage, and encode this information into feature vectors.

[0074] Then, the matching degree calculation unit 21 sends the list of matching feature vectors to the matching degree filtering unit 22. The matching degree filtering unit 22 calculates the similarity based on the list of matching feature vectors. The similarity calculation here may use methods such as cosine similarity or Euclidean distance. For example, if the feature vectors of two cases are [0.8, 0.6, 0.9] and [0.79, 0.61, 0.88], respectively, their cosine similarity will be very high, indicating that the two cases are very similar.

[0075] Next, the matching filtering unit 22 calculates the recommendation priority based on similarity. The priority calculation may consider multiple factors, such as similarity, treatment effect, and side effects. For example, even if two cases are very similar, if one case has a better treatment plan and fewer side effects, it will receive a higher priority.

[0076] Finally, the efficacy assessment unit 23 evaluates the efficacy of the highest-priority drug based on its efficacy data and generates a recommended treatment plan. This unit may use machine learning models, such as random forests or gradient boosting trees, to predict the potential efficacy of a particular drug for the current patient.

[0077] Recommended treatment data and medication regimens are displayed on the treatment recommendation page for doctors to choose from. This page may present expected treatment outcomes in chart form, such as the expected tumor shrinkage and extended survival time, to help doctors make more informed decisions.

[0078] This modular design allows the system to progressively refine recommendations, selecting the most suitable treatment plan for the current patient from a large number of possible matches, embodying the core concept of personalized medicine.

[0079] In one embodiment of the present invention, a method for recommending personalized medication for cancer patients is also disclosed. This method includes two main steps:

[0080] Step S1: Doctor module 1 obtains the medical record text and inputs it into medication recommendation module 2 to obtain a matching list.

[0081] Step S2: Combine the matching list with drug efficacy data to calculate similarity, and apply an attention mechanism to filter out the matching items with the highest similarity in the case matching list, and obtain the recommended drug efficacy assessment for the matching item.

[0082] The advantage of this method lies in its combination of the doctor's professional judgment (through text input of medical records) and the powerful analytical capabilities of machine learning algorithms. For example, when a doctor inputs a complex lung cancer case, the system can not only quickly find similar historical cases, but also filter out the most relevant treatment options.

[0083] In one embodiment of the invention, the application of an attention mechanism is particularly important. It can automatically identify the most critical features in a case and make matching decisions accordingly. For example, for a patient with advanced non-small cell lung cancer, the system might pay particular attention to the tumor's gene mutations, as these are often key factors determining the effectiveness of targeted therapy.

[0084] Specifically, the case texts are first collected and analyzed by the case collection and analysis module 11 to obtain case feature vectors. Then, keyword semantic feature vectors are constructed based on these case feature vectors. Finally, the similarity between the keyword semantic feature vectors and each item to be matched in the clinical trial data is calculated. The top 3-5 items with the highest similarity are selected and their feature vectors are fused to complete the matching between the cases and the clinical trial data, resulting in a matching list.

[0085] A key aspect of this process is selecting the top 3-5 similarity matches. This number is chosen based on a trade-off between experience and practical needs. Selecting too few matches may miss valuable information, while selecting too many may introduce noise and increase computational burden. In a preferred embodiment of the invention, the number 3-5 strikes a good balance between ensuring matching quality and computational efficiency.

[0086] For example, suppose there is a case of a patient with advanced pancreatic cancer. The system might sift through thousands of clinical trial data to find 3-5 most similar cases. These cases may include the same cancer type, similar age group, similar comorbidities, etc. By fusing the feature vectors of these 3-5 cases, the system can obtain a more comprehensive and robust matching result.

[0087] In one embodiment of the present invention, the role of the tumor type recommendation module 12 is particularly emphasized. First, the case collection and analysis module 11 inputs the case text into the tumor type recommendation module 12. Then, the tumor type recommendation module 12 uses a bag-of-words model to extract case features and uses a convolutional neural network to perform feature extraction to obtain a case feature vector.

[0088] In this process, the bag-of-words model can effectively capture keywords and their frequencies in the text. For example, for a case described as a patient experiencing persistent abdominal pain accompanied by jaundice, and a CT scan showing a mass in the head of the pancreas, the bag-of-words model might generate a high-dimensional vector, where the corresponding dimensional values ​​for words such as abdominal pain, jaundice, and pancreatic head are relatively high.

[0089] Next, the application of convolutional neural networks can further extract deeper features. It can identify relationships between words and contextual information; for example, the simultaneous occurrence of persistent abdominal pain and a mass in the head of the pancreas may suggest a higher likelihood of pancreatic cancer.

[0090] Finally, the system calculates the cosine similarity between the keyword semantic feature vectors and each item to be matched in the clinical trial data pool, resulting in a matching list. The formula for calculating cosine similarity is as follows:

[0091]

[0092] Where A represents the keyword semantic feature vector, B represents the feature vector of the item to be matched, and ||A|| and ||B|| represent the Euclidean norms of vectors A and B, respectively. i B i Let n be the i-th component of vectors A and B, and n be the dimension of the vector. The advantage of this method is that it focuses on the direction of the vectors rather than their magnitude, and can effectively compare texts of different lengths.

[0093] In one embodiment of the present invention, step S2 is divided into three sub-steps:

[0094] Step S21: Input the matching list into the matching degree calculation unit 21. The matching degree calculation unit 21 sorts each matching item in the matching list according to the similarity value, sets a threshold, deletes matching items with similarity values ​​lower than the threshold, and generates a matching feature vector list.

[0095] In one embodiment of the invention, this threshold can be set to 0.7. This means that only matches with a similarity greater than 0.7 will be retained. This threshold selection is based on a large amount of experimental data and can effectively filter out less relevant matches while retaining enough valuable information.

[0096] Step S22: Matching degree filtering unit 22 calculates the similarity based on the matching feature vector list.

[0097] Step S23: The matching degree ranking unit calculates the recommendation priority based on the similarity, obtains the highest priority matching item based on the recommendation priority, obtains the drug efficacy data in the efficacy evaluation unit 23 based on the highest priority matching item, and generates recommended drugs based on the efficacy data, which are then displayed on the diagnosis and treatment recommendation interface.

[0098] The design of these three steps reflects the system's layered screening and optimization process, ensuring that the final recommendation results are both relevant and effective.

[0099] Next, in step S21, the matching degree calculation unit 21 sorts each matching item in the matching list according to the similarity value, sets a threshold, deletes matching items with a similarity value lower than 1, and generates a matching feature vector list.

[0100] It's worth noting that the threshold is set to 1. This is a relatively strict standard, meaning only exact matches will be retained. In practice, this threshold may need to be adjusted based on the specific circumstances. For example, for rare cases, the threshold might need to be lowered to 0.8 or lower to ensure that enough relevant cases can be found for reference.

[0101] In one embodiment of the present invention, in step S22, based on the matching feature vector list, the matching feature vectors of all matching items and the keyword semantic feature vectors are multiplied by a dot product to calculate the similarity, and then sorted according to the similarity values.

[0102] The dot product is a simple and effective method for calculating similarity. Suppose we have two vectors A = [a1, a2, ..., an] and B = [b1, b2, ..., bn], their dot product is a1*b1 + a2*b2 + ... + an*bn. The larger this result, the more similar the two vectors are.

[0103] In a preferred embodiment of the invention, the system may also assign different weights to different features. For example, in lung cancer cases, features such as smoking history and gene mutation type may be given higher weights, because these factors have a significant impact on the selection of treatment options.

[0104] In one embodiment of the present invention, step S23 describes how to determine the highest priority match. First, the match with the highest similarity is the match with the highest recommendation priority. However, when the maximum similarity is the same and greater than a set threshold, the matching degree filtering unit 22 adds the matching results in each matching feature vector to obtain a total matching degree value, and re-sorts the similarity based on the total matching degree value. The match with the highest similarity value is the match with the highest recommendation priority.

[0105] In one embodiment of the invention, this threshold can be 0.95. That is, when the similarity of multiple matches is greater than 0.95, the system will further calculate the total matching score to distinguish them. This method can provide more detailed differentiation when the similarity is very close.

[0106] In one embodiment of the present invention, the processing method when the maximum matching degree is different is as follows: In this case, the matching degree filtering unit 22 extracts the vector corresponding to the keyword semantics in the matching feature vector of each matching item and adds them together to obtain the total matching degree value. The similarity is re-sorted according to the total matching degree value, and the matching item with the maximum similarity value is the matching item with the highest recommendation priority.

[0107] The advantage of this method lies in its ability to better capture the semantic information of keywords. For example, in a case described as advanced non-small cell lung cancer with EGFR mutation positivity, the system will pay particular attention to the keyword "EGFR mutation" because it has a decisive impact on the choice of treatment.

[0108] In one embodiment of the present invention, drug efficacy data in efficacy evaluation unit 23 is obtained based on the highest matching item, and a recommended drug regimen is generated based on the efficacy data. The case text with the highest similarity and the recommended drug efficacy evaluation are displayed through the treatment recommendation interface for doctors to select.

[0109] In practical applications, this interface may display expected therapeutic effects in the form of charts or text, such as the expected degree of tumor shrinkage, extended survival time, and possible side effects. This information helps doctors to more comprehensively evaluate treatment options and make the most suitable choice for the patient.

[0110] In one embodiment of the invention, the highest priority match is the match with the largest weighted average of similarity and total matching score, and the similarity is less than the maximum similarity. This design considers both similarity and total matching score, providing a more comprehensive evaluation.

[0111] In one embodiment of the invention, the weighted average can be set to 60% for similarity and 40% for total matching. This ratio is chosen based on extensive clinical data analysis, ensuring the importance of similarity while also fully considering the comprehensiveness of the matching.

[0112] Finally, the recommended drug for the highest priority match is the clinical trial drug corresponding to the highest priority match to be matched, and the efficacy data in the clinical trial data corresponding to the highest priority match is used to generate the recommended drug regimen.

[0113] This means the system doesn't simply recommend drugs for the most similar cases, but also considers the drug's actual efficacy in clinical trials. For example, if a historical case highly similar to the current case has used a novel targeted drug in a clinical trial and achieved significant efficacy, the system will prioritize recommending that drug.

[0114] In summary, the personalized medication recommendation system and method for cancer patients proposed in this invention, by combining natural language processing, machine learning, and clinical medical knowledge, can provide doctors with more accurate and personalized medication advice. It not only improves treatment outcomes but also reduces unnecessary side effects, which is of great significance for improving the quality of life and prolonging the survival of cancer patients. Furthermore, this system has learning and adaptive capabilities; as the amount of data increases and the algorithm is optimized, the accuracy of its recommendations will continuously improve, making a significant contribution to the development of precision medicine.

[0115] To fully demonstrate the superiority of this invention, we designed a series of experiments to compare the embodiments of this invention with conventional methods. Here, we selected 100 newly diagnosed non-small cell lung cancer patients as research subjects, and used the personalized medication recommendation system of this invention (Example 1) and the conventional clinical guideline recommendation method (Comparative Example 1) to formulate treatment plans.

[0116] Example 1 utilizes the personalized medication recommendation system for cancer patients of this invention. The system first obtains detailed patient medical information through the doctor module 1, including age, gender, smoking history, tumor size, stage, and gene mutation status. Then, the system uses a bag-of-words model and convolutional neural network to extract features and construct keyword semantic feature vectors. Next, the system matches these features with cases in the clinical trial database, selecting the 3-5 cases with the highest similarity for analysis. Finally, the system, through the medication recommendation module 2 and in conjunction with drug efficacy data, generates a personalized medication plan for each patient.

[0117] Comparative Example 1 uses the traditional clinical guideline recommendation approach. In this approach, physicians primarily rely on the National Comprehensive Cancer Network (NCCN) guidelines, selecting standardized treatment plans based on the patient's cancer stage and certain clinical characteristics, such as EGFR mutation status.

[0118] Both methods were followed up for 6 months, primarily assessing the following indicators: treatment effectiveness, improvement in quality of life, incidence of adverse reactions, and patient satisfaction. The detection criteria and methods for these indicators are as follows:

[0119] 1. Treatment response rate: The treatment response criteria for solid tumors (RECIST 1.1) were used, and changes in tumor size were assessed by CT or MRI scans. Complete response (CR) and partial response (PR) were considered effective.

[0120] 2. Improvement in quality of life: The European Organisation for Research and Treatment of Cancer (EO RTC) Quality of Life Questionnaire (EO RTC QLQ-C30) was used to assess the improvement before treatment and 6 months later.

[0121] 3. Incidence of adverse reactions: Adverse reactions were assessed according to the National Cancer Institute Common Toxicity Criteria (NCI-CTC AE v5.0), and the incidence of grade 3 and above adverse reactions was recorded.

[0122] 4. Patient satisfaction: A questionnaire with a 5-point Likert scale was designed to assess patients' satisfaction with the treatment plan.

[0123] The test results are shown in the table below:

[0124] Evaluation indicators Example 1 Comparative Example 1 Treatment effectiveness 68％ 52％ Degree of improvement in quality of life 35％ 20％ Adverse reaction incidence 15％ 28％ Patient satisfaction 4.2 / 5 3.5 / 5

[0125] From these results, we can clearly see the superiority of the present invention. Example 1 significantly outperforms Comparative Example 1 in all evaluation metrics. Let's analyze these results in more detail:

[0126] First, the most significant improvement is in treatment effectiveness. Example 1 achieved an effectiveness rate of 68%, 16 percentage points higher than Comparative Example 1. This result fully demonstrates that the personalized medication recommendation system of this invention can more accurately match patients with suitable treatment plans. By analyzing a large amount of clinical trial data, the system can capture subtle factors that might be overlooked by conventional methods, thereby improving the precision of treatment.

[0127] Secondly, the differences in the degree of improvement in quality of life are also considerable. Patients in Example 1 experienced a 3-5% improvement in quality of life, significantly higher than the 20% improvement in Comparative Example 1. This indicates that the present invention not only focuses on therapeutic effects but also fully considers the patient's quality of life. When recommending medication, the system weighs the efficacy and potential side effects of the drug to select the optimal balance for the patient.

[0128] Furthermore, the significant reduction in the incidence of adverse reactions is another highlight of this invention. The incidence of adverse reactions in Example 1 was only 15%, almost half that of Comparative Example 1. This result demonstrates that the system of this invention can better predict and avoid potential adverse reactions, allowing for the selection of the most suitable medication and dosage for patients. This not only improves patients' quality of life but also reduces treatment risks and medical costs.

[0129] Finally, the improved patient satisfaction also reflects the overall advantages of the invention. The patient satisfaction score in Example 1 reached 4.2 out of 5, significantly higher than the 3.5 in Comparative Example 1. This indicates that patients can truly experience the benefits of individualized treatment, including better efficacy, fewer side effects, and a higher quality of life.

[0130] It is worth mentioning that Example 1 also demonstrated good adaptability and learning ability. During the 6-month follow-up period, we observed a month-on-month improvement in the system's recommendation accuracy. This indicates that with the increase in data volume and continuous optimization of the algorithm, the system of the present invention can continuously improve its performance, laying a solid foundation for future long-term applications.

[0131] In summary, the personalized medication recommendation system for cancer patients of this invention demonstrates significant advantages in improving treatment effectiveness, enhancing patients' quality of life, reducing the incidence of adverse reactions, and increasing patient satisfaction. These results not only validate the innovation and practicality of this invention but also provide strong support for the application of precision medicine in the field of cancer treatment. We have reason to believe that, with further technological improvements and the accumulation of clinical experience, this invention will play an increasingly important role in future cancer treatment, bringing patients more personalized, precise, and effective treatment options.

[0132] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A personalized medication recommendation system for cancer patients, characterized in that, The system includes: Doctor Module: Obtain patient case text descriptions, extract case features using a bag-of-words model, and perform feature extraction using a convolutional neural network to obtain case feature vectors; construct keyword semantic feature vectors based on case feature vectors; calculate cosine similarity between keyword semantic feature vectors and each item to be matched in clinical trial data, select the top 3-5 items with the highest similarity and their feature vectors, and fuse them to complete the matching of cases with clinical trial data, obtaining a case matching list; Medication recommendation module: Calculates similarity based on the case matching list and drug efficacy data, and uses an attention mechanism to filter out the matching items with the highest similarity in the case matching list, and obtains the recommended drug and its efficacy assessment for the matching item.

2. The personalized medication recommendation system for cancer patients according to claim 1, characterized in that, The doctor module inputs the case text and case feature vector into the medication recommendation module. The medication recommendation module uses a semantically aware vector matching algorithm to automatically match cases in the clinical trial database with the case feature vector and provides a ranking table of case similarity based on the case feature vector. Based on the matching list and combined with drug efficacy data, the module generates recommended drugs and their efficacy assessments.

3. The personalized medication recommendation system for cancer patients according to claim 1, characterized in that, The doctor module is composed of a doctor module, which provides a text input interface and receives medical records filled in by doctors. The doctor module is connected to the medication recommendation module. The doctor module inputs text or structured data into the medication recommendation module.

4. The personalized medication recommendation system for cancer patients according to any one of claims 1-3, characterized in that, The doctor module includes a case collection and analysis module, a tumor type recommendation module, and a matching list generation module. Both the case collection and analysis module and the tumor type recommendation module are connected to the matching list generation module, which is also connected to the medication recommendation module. The doctor module further includes a storage unit for storing intermediate data during the case collection, tumor type recommendation, and matching list generation processes. After case text is collected, it enters the case collection and analysis module, where relevant features are extracted and processed to obtain feature extraction data. The tumor type recommendation module calculates and processes the feature extraction data to obtain a tumor recommendation type list. The matching list is generated by the matching list generation module and sent to the medication recommendation module.

5. The personalized medication recommendation system for cancer patients according to claim 4, characterized in that, The case collection and analysis module inputs the case text into the tumor type recommendation module. The tumor type recommendation module uses a bag-of-words model to extract case features and uses a convolutional neural network to extract features, obtaining a case feature vector. Based on the case feature vector, a keyword semantic feature vector is constructed. The keyword semantic feature vector is then compared with each item to be matched in the clinical trial data pool using cosine similarity calculation to obtain the matching list. The matching list is then input into the matching list generation module, which in turn inputs the matching list into the medication recommendation module.

6. The personalized medication recommendation system for cancer patients according to claim 5, characterized in that, The medication recommendation module includes a matching degree calculation unit, a matching degree screening unit, and a efficacy evaluation unit. The matching degree calculation unit is connected to the matching list generation module, and the matching degree calculation unit generates a matching feature vector list through the matching list. The matching degree calculation unit is connected to the matching degree filtering unit, and the matching degree filtering unit is connected to the efficacy evaluation unit. The matching degree calculation unit sends the matching feature vector list to the matching degree filtering unit, and the matching degree filtering unit calculates the similarity based on the matching feature vector list. The matching degree ranking unit calculates the recommendation priority based on the similarity, and the efficacy evaluation unit performs efficacy evaluation based on the drug efficacy data of the highest priority drug and generates a recommendation scheme. Recommended treatment data and recommended medication regimens are displayed on the treatment recommendation page for doctors to choose from.

7. The method of the personalized medication recommendation system for cancer patients according to claim 6, characterized in that, Includes the following steps: Step S1: The doctor module obtains the medical record text and inputs it into the medication recommendation module to obtain the matching list; Step S2: Combine the matching list with drug efficacy data to calculate similarity, and apply an attention mechanism to filter out the matching item with the highest similarity in the case matching list, and obtain the recommended drug efficacy assessment for the matching item.

8. The method for recommending individualized medication for cancer patients according to claim 7, characterized in that, Step S1 further includes the following steps: the case text is collected and analyzed by the case collection and analysis module to obtain a case feature vector; a keyword semantic feature vector is constructed based on the case feature vector; the keyword semantic feature vector is compared with each item to be matched in the clinical trial data to calculate similarity, and the top 3-5 items with the highest similarity and their feature vectors are selected and fused to complete the matching between the case and the clinical trial data to obtain the matching list; the case collection and analysis module inputs the case text into the tumor type recommendation module, which uses a bag-of-words model to extract case features and uses a convolutional neural network to extract features to obtain a case feature vector; a keyword semantic feature vector is constructed based on the case feature vector; the keyword semantic feature vector is compared with each item to be matched in the clinical trial data pool to calculate cosine similarity to obtain the matching list.

9. The method for recommending personalized medication for cancer patients according to claim 8, characterized in that, Step S2 further includes the following steps: Step S21: Input the matching list into the matching degree calculation unit. The matching degree calculation unit sorts each matching item in the matching list according to its similarity value, sets a threshold, and deletes matching items with a similarity value lower than the threshold, generating the matching feature vector list. Step S22: The matching degree filtering unit calculates the similarity based on the matching feature vector list; based on the matching feature vector list, it performs a dot product of the matching feature vectors and keyword semantic feature vectors of all matching items to calculate the similarity value, and sorts them according to the similarity value. Step S23: The matching degree ranking unit calculates the recommendation priority based on similarity, obtains the highest priority matching item based on the recommendation priority, acquires the drug efficacy data from the efficacy evaluation unit based on the highest priority matching item, and generates recommended drugs based on the efficacy data, which are then displayed through the treatment recommendation interface; the matching item with the maximum similarity is the matching item with the highest recommendation priority; when the maximum similarity is the same and the maximum similarity is greater than a set threshold, the matching degree filtering unit adds the matching results in each matching feature vector to obtain the total matching degree value, and re-sorts the similarity based on the total matching degree value, with the maximum similarity being the highest priority matching item. The matching item with the highest similarity value is the matching item with the highest recommendation priority. When the maximum matching degree is different, the matching degree filtering unit extracts the vector corresponding to the keyword semantics in the matching feature vector of each matching item and adds them together to obtain the total matching degree value. The similarity is re-sorted according to the total matching degree value, and the matching item with the highest similarity value is the matching item with the highest recommendation priority. The drug efficacy data in the efficacy evaluation unit is obtained according to the highest matching item, and a recommended drug plan is generated according to the efficacy data. The case text with the highest similarity and the recommended drug efficacy evaluation are displayed through the diagnosis and treatment recommendation interface for doctors to select.

10. The method for recommending personalized medication for cancer patients according to claim 9, characterized in that, The highest priority match is the match with the largest weighted average of similarity and matching score; the similarity is less than the maximum similarity; the recommended drug for the highest priority match is the clinical trial drug corresponding to the match to be matched for the highest priority match; the efficacy data in the clinical trial data corresponding to the highest priority match is used to generate the recommended drug regimen.

Images