ARTIFICIAL INTELLIGENCE-BASED CLINICAL DECISION SUPPORT SYSTEM AND PATIENT-SPECIFIC CONTRAST AGENCY DOSE DETERMINATION METHOD

Abstract

The invention encompasses a clinical decision support system based on a large language model retrained with domain-specific data and supported by a Retrieval-Augmented Generation architecture, which simultaneously processes numerous patient-specific clinical parameters and current pharmacological guidelines to optimize patient safety and image quality in medical imaging methods such as computed tomography and magnetic resonance imaging, and the dose determination method applied by this system.

Description

1 TARIFF AI-BASED CLINICAL DECISION SUPPORT SYSTEM AND PATIENT-SPECIFIC APPROACH. METHOD FOR DETERMINING CONTRAST AGENT DOSE TECHNICAL AREA 5 The invention is relevant to medical imaging techniques such as computed tomography and magnetic resonance imaging. Contrast agents used in these methods optimize patient safety and image quality. In order to do this, numerous clinical parameters of the patient and current pharmacological guidelines are combined. Information Access-Assisted Production (Retrieval-Augmented Generation—RAG) that operates in real-time Clinical 10 based on a large language model supported by its architecture and retrained with domain-specific data. It includes the decision support system and the dosage determination method applied by this system. PREVIOUS TECHNIQUE In current technology, computed tomography, magnetic resonance imaging and similar cross-sectional medical imaging techniques are used. The process of determining the dosage of contrast agent in imaging methods is generally used in clinical practice. empirical approaches that rely on fixed protocols or a limited number of clinical parameters This is carried out based on the principle that the contrast agent dose is determined solely on the patient's needs. The calculation is based on total body weight, and the injection is administered according to this calculation. This is how it is carried out. In some cases, the patient's gender is also used as an additional parameter, 20 However, a multi-parameter assessment that goes beyond this is not being carried out. Currently... approaches, glomerular filtration rate, body composition, lean body mass, blood values, factors affecting contrast agent pharmacokinetics, such as the patient's age, comorbidity status, and medical history. It omits critical clinical parameters that directly affect patient safety from being evaluated. This situation presents a significant technical problem in clinical practice: The patient's kidney 25 When functions, body composition, and other individual characteristics are not taken into account, some acute contrast-induced complications due to excessively high doses administered to patients The risk of kidney damage increases; in some patients, insufficient dose leads to decreased image quality. and the need arises to repeat the imaging procedure as its diagnostic value decreases. Each Both situations threaten patient safety, leading to unnecessary radiation exposure and additional costs. 30 This leads to... In addition, international clinical guidelines and current practices are followed in the application of contrast agents. Dynamic integration with pharmacological data is not possible. Clinical decision-making process Because it relies largely on the physician's personal experience or limited reference tables, the patient It is not possible to evaluate fundamental differences in a standardized and systematic way, This also leads to high variation and unpredictable results in clinical applications. 35 2 Artificial intelligence has been developed for determining the contrast agent dosage in the current technique, and When machine learning-based approaches are examined, it is observed that these systems are mostly single-model and It appears that these systems consist of single-layer architectures. These systems are used solely for image analysis or It can generate predictions based on a limited number of patient data; clinical guidelines and pharmacological information... It does not offer a multi-layered decision-making mechanism that processes patient parameters together. This 5 A significant portion of these types of models are able to explain the clinical rationale behind the results they produce. It lacks explainability in artificial intelligence systems used in healthcare. Although it is a critical requirement, models in current technology generally have a black box character. It displays and determines which clinical parameters or sources the recommended dose is based on. It is unable to present the generated data to the user. This situation affects physicians' 10% access to system outputs. This reduces confidence and limits the integration of such systems into clinical routine. The current Another significant shortcoming in the technology is the inability to integrate semantic information retrieval with generative modeling processes. It should not be used. Clinical guidelines, drug package inserts, current publications, and patient scenarios are examples of such sources. The failure to directly integrate rich information resources into the decision-making model, and the model outputs... This makes it difficult to rely on current and reliable information. Especially for large language models. 15 where used, the model's training data does not include or is not adequately represented The risk of producing inaccurate, unfounded, or clinically inconsistent results for clinical situations. This is known as the problem of hallucinations, and this risk is not adequately addressed in current techniques. Given these shortcomings, the current technique aims to optimize contrast agent dosage on a patient-by-patient basis. enabling dynamically updated clinical guidelines and explainable decisions. 20 It is far from offering a system that can produce and strengthen the problem of hallucinations. This Therefore, dose optimization improves patient safety in medical imaging procedures. A new technical solution is needed that personalizes and standardizes clinical decision-making processes. It is heard. The present invention aims to eliminate all these technical problems, Retrieval- A large 25-bit system, retrained with domain-specific data and powered by Augmented Generation architecture. It presents a language model-based clinical decision support system and its corresponding methodology. Patent document number KR102254971B1 describes the use of artificial intelligence to enhance contrast in images. a method and apparatus for converting contrast-free images into one another It is mentioned. The invention involves a non-contrast CT or MRI scan of a specific area of ​​a patient. 30 obtaining the image and inputting this image into an artificial neural network-based learning model. By administering it in this way, the patient receives contrast material as if it had been administered without actually having it injected. It enables the production of an output image such as this. Furthermore, the contrast-enhanced image is also produced using the same model. This makes it possible to convert the image to a non-contrast image. Thus, contrast The aim is to avoid adverse effects associated with substance injection. In the present invention, however, 35 Instead of producing images without administering contrast medium, the contrast medium dosage is determined on a patient-by-patient basis. For this purpose, multiparameter clinical data processing, supported by RAG architecture, is required. 3 These include elements such as the use of language models and semantic access from domain-specific knowledge bases, While this equivalent application focuses solely on image transformation, the current invention provides clinical decision support. It encompasses the entire process. Patent document number KR20250034586A describes how 5 can be achieved using non-contrast computed tomography. The text describes a method and system for detecting brain infarction using artificial intelligence. The invention uses all the slices that make up the complete CT image taken from the patient as input. by using and analyzing these images through a patient model processor, The system allows for the prediction of whether a patient has a brain infarction. For each patient... It performs classification at the image level by working separately for each. In the present invention, 10 However, disease detection is not done through imaging alone; instead, the patient's demographic, clinical, and integrated with laboratory parameters, international clinical guidelines and pharmacological information. It is processed in this way, and also involves calculations such as lean body mass and glomerular filtration rate. Contrast agent dosage is determined, and this decision can be explained using a large language model. is being introduced. While this equivalent application only offers an image-based diagnostic tool, the present invention 15 It is a clinical decision support system that shapes the decision-making process for administering contrast medium. Patent document number KR102939650B1 describes the use of non-contrast computed tomography. The text describes an artificial intelligence-based method and system for detecting brain hemorrhages. The invention involves acquiring a CT image of a patient and processing the entire image using an artificial neural network. By providing input to a network model, especially an LSTM-based architecture, the brain on the image is activated. Automatic detection of the presence, localization, and extent of bleeding. This ensures that the system operates on a personalized model for each patient. It estimates the presence of bleeding from non-contrast images. In the present invention, however, Rather than image analysis, determining the safe and effective dose of contrast agent to be administered to the patient is crucial. Determining this is essential. The present invention applies not only to the image but also to glomerular filtration rate and body mass index. index, lean body mass, blood values ​​and medical history, among numerous clinical parameters It is based on, and also utilizes, Retrieval-Augmented Generation architecture in line with current clinical guidelines. It generates decisions by dynamically bringing in information. This equivalent application is a diagnostic classification. While the system is 30, the current invention focuses on pre-treatment dose optimization and risk management. Patent document number US11100621B2 describes how a machine can produce contrast agents without the use of exogenous contrast agents. learning and simulated post-contrast T1-weighted magnetic resonance imaging using artificial intelligence techniques. The text describes a system and method for generating resonance images. The invention involves converting non-contrast MRI images of a patient into T1-weighted, T2-weighted, FLAIR, and diffusion-35 images. a retrospectively collected dataset consisting of different sequence types, such as weighted imaging By providing this as input to a machine learning model trained on it, the real contrast... 4 producing an output that resembles a contrast-enhanced image without requiring substance injection. This provides. In the present invention, however, it completely eliminates the use of contrast medium. Instead, the dosage of contrast agent to be used should be optimized on a patient-by-patient basis. This is the aim. The current invention not only performs image processing and synthesis, also based on glomerular filtration rate threshold values ​​for patient safety 5 contraindication control, imaging technique-specific pharmacokinetic calculations, and large tongue size The model offers a multi-layered architecture, such as that used to explain clinical decision-making. Patent document number CN113077439A describes microinvasion in hepatocellular carcinoma. The invention refers to an ultrasound contrast system for prediction. The invention is the Sonazoid 10. Based on the withdrawal pattern of the ultrasound contrast agent in the Kupffer phase, the patient's data... Segmentation of the hepatocellular carcinoma lesion area, figure, histogram, second from this region. Extraction of grade-specific histograms and tissue features, and statistical analysis to determine the most significant Attribute identification and patient clinical information, ultrasound image characteristics, and artificial intelligence. By combining radiographic features based on a model, the risk of microinvasion during surgery is reduced. 15 It enables prediction in the pre-infancy period. In the present invention, however, it is aimed at a specific pathology. It is not a diagnostic prediction system, but rather a comprehensive computed tomography and magnetic resonance imaging (MRI) scan. A universal contrast agent dose determination system applicable to these methods is presented. The present invention utilizes a large language model and, instead of radiomic feature extraction and statistical analysis, Semantic information retrieval from clinical guidelines via Retrieval-Augmented Generation architecture 20 It performs and uses generative artificial intelligence with deterministic calculations based on patient parameters. It adopts a hybrid approach that brings them together. Patent document number CN120656350A describes three dynamic devices targeting the uterus and fallopian tubes. A method and system for 3D ultrasound contrast training is described. The invention is a 25 Based on the patient's clinical data and medical history, artificial intelligence algorithms are used to identify specific pathological conditions. screening patient groups with specific characteristics and a corresponding training case study creation of the library, followed by virtual reality-based three-dimensional modeling techniques dynamic simulation of the anatomical structures of the uterus and fallopian tubes and in depth Using learning algorithms, the person being trained can adapt their ultrasound probe movements and image processing. It enables the generation of feedback by analyzing its quality in real time. In the present invention... However, it is not a simulation system for training clinical staff, but rather a clinical experience directly at the patient's bedside. A system is presented that supports the decision-making process. The present invention is a case study for educational purposes. instead of creating a library and virtual reality simulation, it works on real patient data. Calculation of contrast agent dosage, performance of safety checks and explainable dose 35 It includes functions directly related to clinical application, such as generating recommendations. Patent document number CN121354924A describes a reinforcement learning-based approach to gout treatment. The text discusses a method for predicting the prognosis. The invention involves data from multiple sources related to the patient. Creating a state space by collecting data, a Markov scale of the disease progression process modeling the decision-making process and using reinforcement learning algorithms to determine a specific Predicting future disease states with or without intervention 5 The system also provides contrasting phenomena and uncertainty measurement through It provides interpretable and verifiable visualization results. In the present invention, however, Reinforcement learning and Markov decision process are not used; instead, Retrieval-Augmented A large language model supported by Generation architecture is preferred. The current invention, Instead of predicting disease prognosis, contrast agent injection is a specific and urgent clinical action. 10 The focus is on determining the dosage beforehand, and this decision is based on clinical guidelines and pharmacological principles. It makes it explainable by substantiating it with documents. Patent document number CN104677974A describes multiple proteins for acute ischemic stroke. A combination diagnostic model based on markers is described. The invention involves a patient with 15 The serum contains eight proteins: SHIP1, IGF1, SDRP, MBL, CAM1, CRP, LPa, and MRC1. Measuring the expression levels of different proteins and a combination model of these eight proteins. By considering these factors together, early diagnosis of acute ischemic stroke is possible. This invention also enables the processing of SHIP1, SDRP, CAM1 and MRC1 proteins. Its association with disease is being demonstrated for the first time. In the present invention, proteomics or biomarker 20 There is no diagnostic system based on age, instead of the patient's serum protein profile. The current invention uses age, gender, height, weight, body mass index, glomerular filtration rate, blood values, medical history, and using parameters that are routinely measured in clinical practice, such as imaging techniques, and By combining these parameters with a large language model and semantic information retrieval, contrast agent dosage is determined. It determines. 25 Patent document number CA3086557A1 describes one or more substances contaminated with biological agents. a way to organize the decontamination or disinfection of excess objects The method is being discussed. The invention describes how each area labeled with Radio Frequency Identification (RFID) can be used to monitor the patient. 30 cleaning or disinfection history of the room or reusable medical equipment monitoring, for example, if a patient is moved from one room to another, this information updating and thus ensuring infection control in the hospital environment It is based on the present invention, which involves monitoring the disinfection of physical objects or areas. They have no relation whatsoever. The present invention is in an entirely different technical field, medical imaging. an artificial intelligence system for determining contrast agent dosage on a patient-by-patient basis beforehand 35 The system offers services such as patient room monitoring and equipment disinfection. It excludes. 6 Studies have shown that determining the dosage of contrast agents in medical imaging procedures... Existing technical solutions for this purpose are mostly limited to image conversion or a specific function. It is limited to narrow applications such as diagnostic detection of pathology and contrast agent It is far from offering a direct solution to the problem of optimizing dosage on a patient-by-patient basis. 5 It is observed that a significant portion of current systems completely eliminate the use of contrast agents. It adopts image synthesis approaches aimed at removal, and these approaches rather than managing the clinical risks associated with contrast agent injection, managing non-contrast methods It aims to convert between visual and contrast-enhanced images. Some other techniques... The solutions, on the other hand, are aimed at diagnosing a specific disease or pathological condition and involve 10 steps on the image. It offers functions such as lesion detection, segmentation or classification, but contrast medium There is no need to determine the dosage based on patient-specific safety and efficacy parameters. It does not include a mechanism. Furthermore, the artificial intelligence-based technology used in the current technique... The vast majority of approaches rely on a wealth of information, such as clinical guidelines and pharmacological documentation. It is unable to dynamically integrate its resources into the decision-making process, therefore especially 15 Any protection against the problem of hallucinations that can be seen in large language models It does not offer a mechanism. Another significant shortcoming of existing systems is glomerular filtration. pharmacokinetics specific to rate, lean body mass, body composition, and imaging technique. Deterministic methods that simultaneously evaluate critical clinical parameters such as requirements a hybrid architecture that combines computational mechanisms with generative artificial intelligence models 20 They lack these qualities. The present invention, however, aims to eliminate all these technical shortcomings. Supported by a Retrieval-Augmented Generation architecture, it targets numerous site-specific patients. It presents a clinical decision support system based on a large language model retrained with a specific scenario. This system uses patient data such as age, gender, height, weight, body mass index, glomerular filtration rate, and blood type. values, medical history, and imaging techniques, among numerous clinical variables, are evaluated simultaneously. 25 It starts with a data input module that functions as, then, James or Boer formulas. using lean body mass calculation and specific imaging techniques. Pharmacokinetic calculations based on total body weight or lean body mass It does so. However, the system adheres to international clinical guidelines and pharmacological principles. 30 A knowledge base created from documents and patient scenarios through vector representations It indexes and uses semantic similarity algorithms to identify the most relevant contextual content for patient queries. It dynamically retrieves this contextual information. This retrieved contextual information is then used in deterministic calculations and Glomerular filtration rate is combined with safety guidelines based on threshold values; site-specific A large language model retrained with data is used as a production component, creating a multi-layered and Structured requests can be explained through output templates, and contrast agent 35 is tailored to the patient. Dosage recommendations are being formulated. Consequently, the current invention hallucinates using large language models. a clinically sound, explainable, and structured approach that minimizes the problem and is supported by clinical guidelines. 7 By generating contrast agent dosage recommendations specific to each patient, none of the current systems in technology can achieve this. a multi-parameter, dynamic and safety-focused contrast agent dose optimization that it could not solve It solves the problem. Ultimately, the problems mentioned above, which cannot be solved with the current technology, are within the scope of 5 relevant technical fields. This has made it necessary to make an innovation. A BRIEF DESCRIPTION OF THE INVENTION The present invention is an artificial construct developed to eliminate the technical shortcomings mentioned above. Intelligence-based clinical decision support system and patient-specific contrast agent dose determination method 10 It is related to. The main purpose of the invention is to determine the dosage of contrast agents in medical imaging processes. A fundamental solution to the multi-parameter clinical data processing and patient safety optimization problem encountered. The aim is to offer a solution that considers only the patient's total body weight and gender, among other limited factors. 15 caused by traditional dose calculation methods based on parameters Risk of acute kidney injury due to contrast, poor image quality and examination due to low dose. repetition problems, glomerular filtration rate, body composition, lean body mass, blood values ​​are considered simultaneously by numerous clinical variables such as age, medical history, and imaging technique. It is eliminated by being considered as such. Another aim of the invention is to use deterministic pharmacokinetic calculations to develop generative artificial intelligence. By offering a hybrid architecture that brings together these models, it enables clinical applications of numerical dose calculations. The aim is to ensure that it is based on guidelines and pharmacological information. In this way, it is not just about numbers. not only to generate a dosage recommendation, but also to determine which clinical parameters and international standards this recommendation is based on. guidelines and the pharmacological data on which they are based are presented to the user in a clear and understandable way. This is presented in order to increase physicians' confidence in the system outputs and improve clinical decision-making. 25 The process is being made transparent. Another aim of the invention is to create clinical guidelines through a Retrieval-Augmented Generation architecture. a domain-specific knowledge base consisting of drug leaflets and patient scenarios in a large language by integrating it into the model, the model outputs are up-to-date, reliable, and clinically grounded. The aim is to ensure this. In this way, the problem frequently seen in large language models and the model's training 30 Incorrect, unfounded, or misleading information regarding clinical conditions not included in or adequately represented in the data. The problem of hallucinations, known for the risk of producing results that contradict clinical guidelines, is serious. By reducing dosages, clinically reliable and consistent dosage recommendations can be generated. 8 Another aim of the invention is to provide data based on the patient's body composition and imaging technique. The goal is to integrate deterministic algorithms into the system that will perform optimal dose calculations. In this way, lean body mass is calculated using the James or Boer formulas, Dose calculation based on lean body mass in computed tomography and X-ray applications. During this process, magnetic resonance imaging is based on total body weight and glomerular 5 Depending on the filtration rate threshold values, the use of contrast medium is completely prevented. or safety mechanisms such as dose reduction are activated to maximize patient safety. is being removed. Another objective of the invention is to create a large language model that can be retrained with numerous patient scenarios. using different age groups, kidney function levels, comorbidity profiles and imaging 10 The aim is to increase the model's sensitivity to these scenarios. This will allow the system's outputs to be more realistic. It is ensured that it reflects clinical scenarios and is adaptable to different patient populations, The problem of clinical variation caused by uniform dosing protocols is eliminated. Another purpose of the invention is structured requests engineering and rule-based routing. By controlling the responses produced by the large language model through its mechanisms, it prevents errors or 15 The aim is to minimize the risk of unfounded production. In this way, model outputs include dosage range, clinical rationale, in structured sections such as risk analysis, type of contrast agent, and general assessment. They are being developed, both explainability and clinical usability are being improved, and each proposal is also being evaluated. Compliance with international clinical guidelines is ensured. Another aim of the invention is to be easily integrated into clinical practice, providing existing healthcare organizations with knowledge. a system architecture compatible with transaction infrastructures and offering real-time decision support The aim is to provide this. In this way, physicians and a tool that provides quick and reliable dosage recommendations without interrupting technicians' workflow This also avoids unnecessary repeat tests and associated additional radiation. By preventing exposure, operational efficiency and cost effectiveness are achieved. 25 Another aim of the invention is to improve clinical decision-making processes in contrast agent administration. by standardizing, eliminating differences in practice between different centers and different physicians. The goal is to minimize variations in clinical outcomes resulting from the same clinical condition. patients, regardless of the imaging center where the imaging was performed or the evaluating physician This ensures that similar dosage recommendations are received, thus promoting equality, consistency, and quality in healthcare. 30 Standards are being raised. 9 All the purposes mentioned above and those that will emerge from the detailed explanation below. The present invention relates to an artificial intelligence-based clinical decision support system. the subject system; - patient's age, gender, height, weight, glomerular filtration rate, blood values, disease Data 5 is a multiparameter data structure that receives history and imaging technique information. Introduction Module, - calculates body mass index using data obtained from the aforementioned Data Entry Module, and A body composition that determines lean body mass using the James or Boer formula. Calculation Module, - Lean 10 as determined by the aforementioned Body Composition Calculation Module. body mass and imaging provided by the aforementioned Data Input Module. A Dose system uses technical knowledge to calculate contrast agent dose and volume. Calculation Engine, - glomerular filtration rate and disease provided by the aforementioned Data Entry Module Based on medical history information and clinical guidelines, safety warnings and contraindications 15 A Security Control Module that performs an assessment, - Text parsing and vector processing by processing clinical guidelines and pharmacological documents. A Document Processing and Vectorization Module that creates the representation, - vector generated by the aforementioned Document Processing and Vectorization Module Vectorizing patient queries onto representations, we identified the 20 most relevant to semantic similarity. An Information Retrieval Module that retrieves clinical content, - an open-source Big Language Model retrained with numerous patient scenarios and - the aforementioned Data Entry Module, the aforementioned Body Composition Calculation Module, The aforementioned Dose Calculation Engine, the aforementioned Safety Control Module and 25 using the aforementioned Great Language Model as a production component and within its own structure Dosage can be explained through defined structured prompt and output templates. A Decision Production Module that generates recommendations It includes. The invention also allows for patient-specific contrast agent doses of 30 using the system described above. It also includes the determination method, which consists of the following steps: - patient's age, gender, height, weight, glomerular filtration rate, blood values, disease It receives history and imaging technique information through the Data Entry Module and much more. It functions as a parameterized data structure. - calculates body mass index and lean body mass index using the James or Boer formula. It determines its mass through the Body Composition Calculation Module. 5 - based on total body weight or lean body mass according to imaging technique. Contrast agent dose and volume calculation is performed using the Dose Calculation Engine. It accomplishes. - Based on glomerular filtration rate and medical history information, as well as clinical guidelines. Safety warnings and contraindications assessment in Safety Check 10. It does this through the module. - Text parsing and vector processing by processing clinical guidelines and pharmacological documents. The representation creation process is carried out through the Document Processing and Vectorization Module. It accomplishes. - Vectorizing patient queries to retrieve the most relevant clinical content based on semantic similarity. 15 The retrieval process is carried out through the Information Retrieval Module. - Open-source Big Language retrained with numerous patient scenarios Through its model, it ensures the production of clinically relevant, consistent, and reliable outputs. - Data Entry Module, Body Composition Calculation Module, Dose Calculation Engine, All data provided by the Security Control Module and the Information Retrieval Module is 20 It combines; open-source Big Language retrained with numerous patient scenarios. By using its model as a production component, it defines itself internally. Dose recommendations can be explained through structured prompt and output templates. It creates and presents the product to the user via the Production Module. The structure of the current invention and the best understanding of its advantages, including additional elements, are outlined in section 25. This should be evaluated together with the figures explained below. BRIEF DESCRIPTION OF THE FIGURES Figure 1: Schematic view of an artificial intelligence-based clinical decision support system. REFERENCE NUMBERS 11 1 - Data Entry Module 2 - Body Composition Calculation Module 3 - Dose Calculation Engine 4 - Security Control Module - Document Processing and Vectorization Module 5 6 - Information Retrieval Module 7 - The Large Language Model 8 - Decision Making Module DETAILED DESCRIPTION OF THE INVENTION 10 This detailed explanation focuses solely on the innovation in the invention to provide a better understanding of the subject matter. This is conveyed without being limited to examples. Accordingly, in the following explanation and figures, artificial Intelligence-based clinical decision support system and patient-specific contrast agent dose determination method. It is explained. Figure 1 is a schematic view of an AI-based clinical decision support system. According to this view, the system: - patient's age, gender, height, weight, glomerular filtration rate, blood values, disease A data structure that receives history and imaging technique information as a multiparameter data structure. Introduction Module (1), - 20 which calculates body mass index with data received from the mentioned Data Entry Module (1). and a Body Mass Index that determines lean body mass using the James or Boer formula. Composition Calculation Module (2), - Lean as determined by the Body Composition Calculation Module (2) mentioned above body mass and imaging provided by the aforementioned Data Input Module (1). Dose 25 is a device that calculates contrast agent dose and volume using its technical knowledge. Calculation Engine (3), - glomerular filtration rate provided by the mentioned Data Input Module (1) and Safety warnings are based on medical history information and clinical guidelines. A Safety Control Module (4) that performs contraindication assessment, - Text fragmentation and vector processing by processing clinical guidelines and pharmacological documents 30 A Document Processing and Vectorization Module (5) that creates the representation, 12 - generated by the aforementioned Document Processing and Vectorization Module (5) Vectorizing patient queries onto vector representations and achieving semantic similarity An Information Retrieval Module (6) that retrieves relevant clinical content. - an open-source Big Language Model retrained with numerous patient scenarios (7) and, 5 - mentioned Data Entry Module (1), mentioned Body Composition Calculation Module (2), the mentioned Dose Calculation Engine (3), the mentioned Safety Control Module (4) and combines all data provided by the aforementioned Information Retrieval Module (6), using the aforementioned Great Language Model (7) as a production component and its own through the structured request and output templates defined within it 10 A Decision Production Module (8) that generates explainable dose recommendations It includes. The invention concerns an Artificial Intelligence-Based Clinical Decision Support System and Patient-Specific Contrast Agent. Dosage Determination Method refers to the patient-based dosage of contrast agents in medical imaging procedures. It is a hybrid system developed to optimize performance, and its working principle is deterministic. 15 Pharmacokinetic calculations and large languages ​​supported by Retrieval-Augmented Generation architecture. It is based on the integration of generative artificial intelligence based on the model. The invention is the subject of this. The preferred structure of the method is as follows: - To start the system, the patient's age, gender, height, weight, etc. must be entered into the Data Entry Module (1). glomerular filtration rate, blood values, medical history, and imaging technique 20 The information is received as a multi-parameter data structure. - This data is directly entered into Body Composition by the aforementioned Data Entry Module (1). Calculation Module (2) with Dose Calculation Engine (3) and Safety Control Module (4) is transmitted. - Body Composition Calculation Module (2), on the data it receives, the patient's body 25 It calculates the body mass index and lean body mass using the James or Boer formula. It determines this value and transmits it to the mentioned Dose Calculation Engine (3). - The Dose Calculation Engine (3) receives the image from the mentioned Data Input Module (1). with the technical knowledge he received from the Body Composition Calculation Module (2) mentioned above. Using lean body mass index, computed tomography or X-ray 30 in applications based on lean body mass, and in magnetic resonance imaging. It performs contrast agent dose and volume calculations based on total body weight. - Simultaneously, the Security Control Module (4) is connected to the aforementioned Data Entry Module (1) Based on the glomerular filtration rate and medical history information obtained, and according to clinical guidelines. 13 It uses these rules to generate contraindications, dose reductions, and safety warnings. It transmits the outputs to the Decision Production Module (8). - In a parallel process, the Document Processing and Vectorization Module (5) is already in the system. by processing the defined international clinical guidelines and pharmacological documents It creates text fragmentation and vector representation. 5 - Created by the aforementioned Document Processing and Vectorization Module (5) Vector representations are indexed and made searchable by the Information Retrieval Module (6). is brought. - The Information Retrieval Module (6) can be accessed directly from the aforementioned Data Entry Module (1) or The patient query received through the aforementioned Decision Production Module (8) 10 by vectorizing, by the aforementioned Document Processing and Vectorization Module (5) The most relevant clinical case is identified using a semantic similarity algorithm on the generated vector representations. It dynamically retrieves content and uses this contextual content in the aforementioned Decision Generation process. It transmits to module (8). - Decision Production Module (8), mentioned Data Entry Module (1), mentioned Body 15 Composition Calculation Module (2), mentioned Dose Calculation Engine (3), the mentioned Security Control Module (4) and the mentioned Information Retrieval Module (6) provides all data, calculated parameters, security rules and by combining the retrieved contextual clinical content, a multi-layered and structured approach is created. It creates a request. 20 - Decision Generation Module (8) is used to generate the final clinical decision output. It uses the Big Language Model (7) as a production component. - Big Language Model (7), retrained with numerous patient scenarios, open source It is a model and is made available for use in clinical practice by making the aforementioned Decision Production Module (8) available. It provides context-sensitive, consistent, and reliable output generation capability. 25 - The Decision Production Module (8) incorporates the aforementioned Big Language Model (7) within itself. Dosage interval using defined structured prompt and output templates, clinical rationale, risk analysis, type of contrast agent, and general assessment, etc. structured sections, explainable and based on clinical guidelines. It creates and presents a contrast agent dosage recommendation specific to the patient. This 30 In this way, the system solves the hallucination problem frequently seen in large language models. a clinical decision support process that is minimized, safe, and optimized on a patient-by-patient basis. It provides. 35

Claims

14 REQUESTS 1. It is an artificial intelligence-based clinical decision support system whose features include: • Patient's age, gender, height, weight, glomerular filtration rate, blood values, disease A data structure that receives history and imaging technique information as a multiparameter data structure. Introduction Module (1), 5 • Calculates body mass index with data received from the mentioned Data Entry Module (1) and A body composition that determines lean body mass using the James or Boer formula. Calculation Module (2), • Lean determined by the Body Composition Calculation Module (2) mentioned above body mass and imaging provided by the aforementioned Data Input Module (1) 10 A Dose system uses technical knowledge to calculate contrast agent dose and volume. Calculation Engine (3), • glomerular filtration rate and provided by the mentioned Data Input Module (1) Safety warnings are based on medical history information and clinical guidelines. A Safety Control Module (4) that performs contraindication assessment, 15 • Text parsing and vector representation by processing clinical guidelines and pharmacological documents. a Document Processing and Vectorization Module (5), • vector generated by the aforementioned Document Processing and Vectorization Module (5) Vectorizing patient inquiry onto representations to identify the most relevant clinical settings with semantic similarity. A Knowledge Retrieval Module (6) that retrieves content, 20 • an open-source Big Language Model (7) retrained with numerous patient scenarios and • mentioned Data Entry Module (1), mentioned Body Composition Calculation Module (2), the mentioned Dose Calculation Engine (3), the mentioned Safety Control Module (4) and combining all data provided by the aforementioned Information Retrieval Module (6), using the aforementioned Great Language Model (7) as a production component and having 25 within itself Dosage recommendations can be explained through defined structured prompt and output templates. a Decision Production Module (8) It includes.

2. This is a method for determining patient-specific contrast agent doses; its characteristic feature is: - patient's age, gender, height, weight, glomerular filtration rate, blood values, disease History and imaging technique information is received through the Data Entry Module (1). and processed as a multi-parameter data structure, - Calculation of body mass index and lean body mass using the James or Boer formula Determining the mass through the Body Composition Calculation Module (2), 5 - based on total body weight or lean body mass according to imaging technique. Dose Calculation Engine for contrast agent dose and volume calculation. (3) is carried out through, - Based on glomerular filtration rate and medical history information, as well as clinical guidelines. Safety warnings and contraindication assessment in Safety Control Module 10 (4) to be done through - Text parsing and vector processing of clinical guidelines and pharmacological documents. The process of creating the representation is done through the Document Processing and Vectorization Module (5). to be carried out, - Vectorizing patient queries and returning the most relevant clinical content based on semantic similarity. 15 The process of fetching information is done through the Information Fetch Module (6), - Open-source Big Language Model retrained with numerous patient scenarios (7) through which clinically relevant, consistent and reliable output production is ensured and - Data Entry Module (1), Body Composition Calculation Module (2), Dose Calculation The engine (3), Safety Control Module (4) and Information Retrieval Module (6) are powered by 20 Combining all provided data, Decision Production Module (8) of the Big Language Model (7) used as a production component and output with structured request. Dosage recommendations that can be explained through templates and presented to the user. It includes the steps. 30

Images