System and method for dose prediction

Abstract

The described embodiments relate to methods for predicting the toxicity of a drug to be administered to a patient. The method includes accessing at least one computed tomography (CT) torso slice associated with a patient; performing a tagging process on the at least one CT scan using a trained artificial intelligence (AI) segmentation model; determining at least one body composition parameter based on the at least one tagged CT slice; receiving at least one demographic parameter associated with the patient; receiving at least one dose parameter associated with a medicament to be administered to the patient; and generating an output based on the at least one body composition parameter, the at least one demographic parameter, and the at least one dose parameter using the trained AI prediction model, the output corresponding to the predicted toxicity of the drug to the patient.

Description

Technical Field

[0001] The described embodiments generally relate to systems and methods for dose prediction. Specifically, the described embodiments relate to systems and methods for performing drug dose prediction to minimize adverse drug reactions. Background Technology

[0002] The unpredictability of toxicity development in patients following drug treatments (such as chemotherapy) is significant. For example, in a cohort of colorectal cancer patients treated with oxaliplatin, more than 50% experienced serious or life-threatening drug toxicity.

[0003] The dosage of chemotherapy used to treat cancer is often roughly calculated based on parameters such as the patient's body surface area (BSA). This leads to a very high incidence of chemotherapy toxicity. These toxicities cannot be accurately predicted, thus causing serious complications in patients, including sepsis and neuropathy, requiring hospitalization for patients with severe problems. This overdose occurs because patients with the same body surface area (BSA) may have very different body compositions, resulting in very different absorption rates of chemotherapy doses.

[0004] Better methods are needed to select the dosage of a specific drug for a particular patient in order to minimize the patient’s chance of adverse reactions to the drug or reduce the severity of adverse reactions.

[0005] The aim is to address or improve upon one or more drawbacks or defects associated with prior systems and methods used for dose prediction, or at least provide a useful alternative. Documents, laws, materials, devices, articles, etc., included in this specification should not be construed as an admission that any or all of these constitutes part of the prior art or is common general knowledge in the field related to this disclosure, although they existed prior to the priority date of each appended claim. Summary of the Invention

[0006] A method is provided for predicting the toxicity of a drug to be administered to a patient. The method includes:

[0007] Receive at least one computed tomography (CT) slice of the torso associated with the patient;

[0008] A trained artificial intelligence (AI) segmentation model is used to perform labeling processing on at least one CT scan;

[0009] At least one body composition parameter is determined based on at least one labeled CT slice;

[0010] Receive at least one demographic parameter associated with the patient;

[0011] Receive at least one dosage parameter associated with the drug to be administered to the patient; and

[0012] An output is generated using a trained AI prediction model based on at least one body composition parameter, at least one demographic parameter, and at least one dosage parameter, the output corresponding to the predicted toxicity of the drug given to the patient.

[0013] In some embodiments, the method further includes training the AI ​​prediction model.

[0014] In some embodiments, the labeled CT slice includes one or more or all of the labeled regions of muscle, visceral adipose tissue (VAT), subcutaneous adipose tissue (SAT), and intramuscular / muscular adipose tissue (IMAT).

[0015] In some embodiments, the body composition parameter is at least one or all of the following: amount of muscle, amount of VAT, amount of SAT, amount of IMAT, radiometric density of muscle, radiometric density of VAT, radiometric density of SAT, or radiometric density of IMAT.

[0016] A method is provided for training an artificial intelligence model to assess drug dosage in order to minimize toxic reactions to a drug. The method includes:

[0017] Receive samples of labeled CT trunk slices associated with patient samples;

[0018] At least one body composition parameter is determined based on each labeled CT slice;

[0019] A set of predictive parameters is identified for each patient, including at least one body composition parameter, at least one patient demographic parameter, and at least one dose parameter;

[0020] For each combination of prediction parameters, a model fitting process is performed to enable the model to generate a predicted toxicity rating for each patient;

[0021] For each combination of predicted parameters, the predicted toxicity rating is compared with the measured toxicity rating to determine the accuracy of that combination of predicted parameters; and

[0022] Choose the optimal combination of prediction parameters, where the optimal combination of prediction parameters is the combination of prediction parameters that leads to the highest accuracy.

[0023] In some embodiments, the labeled CT slices include one or more or all of the labeled areas of muscle, VAT, SAT, and IMAT.

[0024] In some embodiments, the body composition parameter is at least one or all of the following: amount of muscle, amount of VAT, amount of SAT, amount of IMAT, radiometric density of muscle, radiometric density of VAT, radiometric density of SAT, and radiometric density of IMAT.

[0025] In some embodiments, the method further includes using a trained AI model to perform labeling of CT slice samples to generate a sample of labeled CT slices.

[0026] In some embodiments, the CT slice is taken from the abdomen.

[0027] In some embodiments, between two and one thousand CT slices are received.

[0028] A method for minimizing the toxic effects of a drug on patients is provided. The method includes:

[0029] i) Determine the amount of muscle, visceral adipose tissue (VAT), subcutaneous adipose tissue (SAT), and intramuscular / intramuscular adipose tissue (IMAT) in at least a portion of the patient's trunk.

[0030] ii) Based at least on i) determining the patient's risk of toxic reactions to the drug, and

[0031] iii) If a patient is determined to be at risk of toxicity, then reduce the standard recommended dose of the drug.

[0032] In some embodiments, the method further includes administering the drug.

[0033] In some embodiments, step i) is performed using any of the methods described above.

[0034] Throughout this specification, the word “comprising” or variations thereof, such as “comprises” or “comprising”, shall be understood to imply the inclusion of the stated element, whole or step, or group of elements, whole or steps, but does not exclude any other element, whole or step, or group of elements, whole or steps. Attached Figure Description

[0035] Figure 1 An example flowchart of a method for dose prediction according to some embodiments is shown;

[0036] Figure 2 The illustration shows what can be used to perform in some embodiments. Figure 1 Example architecture of the method;

[0037] Figure 3A The illustration shows a first example of a labeled CT image taken from the first patient;

[0038] Figure 3B The illustration shows a second example of a labeled CT image taken from a second patient;

[0039] Figure 4 The illustration shows an example pipeline for assessing toxicity in a specific patient according to some embodiments;

[0040] Figure 5 A graph showing the accuracy of dose prediction for different combinations of illustrated features; and

[0041] Figure 6 The following are examples of what can be used to perform [the task] according to some embodiments. Figure 1 and Figure 4 Example system of steps. Detailed Implementation

[0042] The described embodiments generally relate to systems and methods for dose prediction. Specifically, the described embodiments relate to systems and methods for performing drug dose prediction to minimize adverse drug reactions.

[0043] Some embodiments relate to tools for predicting the dose-limiting toxicity (DLT) of a drug. Examples include drugs for treating cancer, which may include breast cancer, lung cancer, liver cancer, colon cancer, bowel cancer, prostate cancer, colorectal cancer, and / or pancreatic cancer. Further examples include drugs for treating diabetes, hypertension, pain, autoimmune diseases, and / or neurodegenerative diseases. The described embodiments can help balance achieving a positive response to a drug (such as a reduction in the condition of the patient receiving the drug) with reducing at least one side effect (such as toxicity, neuropathy, or another adverse reaction). A drug can be any type of substance administered to a patient to treat or prevent a condition. Examples of drug types may include, but are not limited to, small molecules, antibodies, vaccines, and viruses.

[0044] A patient's absorption of a drug is based on many factors, including the patient's body composition. This can be estimated using measurements such as the patient's height, weight, BMI, and / or body surface area (BSA). However, none of these measurements can accurately predict how a patient's body will absorb a particular drug.

[0045] More accurate measurements of body composition can be obtained from computed tomography (CT) scans. Trained practitioners can analyze CT scans and determine which areas are associated with specific tissue types, such as muscle, visceral adipose tissue (VAT), and subcutaneous adipose tissue (SAT). However, this manual process is time-consuming and cumbersome, and is therefore generally used only for research purposes. Furthermore, measurements of specific tissues within a single CT scan image may not provide an accurate picture of a patient's overall body composition.

[0046] By automating the processing of analytical CT scans, the described embodiments allow for efficient analysis of multiple CT slices from a single patient via segmentation, thereby enabling more accurate calculation of body composition and more accurate dose prediction. The disclosed method further improves the accuracy of predictions by taking into account demographic and biostatistical parameters.

[0047] While some examples in this paper specifically relate to chemotherapy for colorectal cancer, it should be noted that the methods disclosed are equally applicable to other cancer types and diseases, including breast cancer, lung cancer, liver cancer, colon cancer, bowel cancer, prostate cancer, pancreatic cancer, diabetes, hypertension, pain, autoimmune diseases, and / or neurodegenerative diseases. More specifically, the segmentation method for CT scans used to calculate body composition remains the same for all drugs. The final model that maps those body composition parameters to toxicity or dosage can be trained individually for each drug, as disclosed herein.

[0048] As used herein, the term "patient" refers to mammals, including humans, livestock such as horses, cattle, sheep, and goats, and companion animals such as dogs and cats. In some embodiments, the patient is a human.

[0049] Computed tomography (computational computed tomography; formerly known as computed axial tomography or CAT scan) is a medical imaging technique used to acquire detailed images of the body's interior. Some CT scanners use rotating X-ray tubes and a row of detectors placed in a gantry to measure X-ray attenuation in different tissues within the body. The multiple X-ray measurements obtained from different angles are then processed on a computer using tomographic reconstruction algorithms to produce a cross-sectional image (virtual "slice") of the body. Therefore, when this article refers to a "CT scan," it refers to the entire dataset generated by a computer for a patient and may contain one or more "CT slices." A CT slice is a single image representing a cross-section of the patient's body.

[0050] This disclosure provides a computer-based method for evaluating drug dosage using multiple different artificial intelligence (AI) models. These models can also be described as machine learning (ML) models because there is a training process that tunes the model's internal parameters so that the model output best corresponds to the labels of the training data.

[0051] The first model is referred to as the segmentation model. The segmentation model is trained to recognize tissue regions shown on CT slices, which may include muscle tissue, visceral adipose tissue (VAT), subcutaneous adipose tissue (SAT), intramuscular / intramuscular adipose tissue (IMAT), bone, and organ tissue. The segmentation model can be trained to perform this process by providing it with training data comprising a series of manually labeled CT slices. Once trained, new CT slices are provided to the segmentation model and automatic labeling is performed. CT slices from a series of locations from a single patient can be provided to the model and labeled. In some embodiments, at least some slices are from the patient's torso. In some embodiments, at least some slices are from the patient's head, arm, or leg. In some embodiments, slices are taken between the patient's neck and coccyx. In some embodiments, slices are taken between the T1 and T12 vertebrae. In some embodiments, slices are taken between the L1 and L5 vertebrae. In some embodiments, at least some slices are taken from the patient's abdomen. In some embodiments, 2 to 1000 slices are analyzed. In some embodiments, 2 to 500 slices are analyzed. In some embodiments, 2 to 100 slices are analyzed. In some embodiments, 2 to 50 slices are analyzed. In some embodiments, 2 to 25 slices are analyzed. In some embodiments, 2 to 10 slices are analyzed.

[0052] Multiple body composition parameters can then be calculated based on the labeled CT scan data. These parameters may include: surface area of ​​muscle, VAT, SAT, IMAT, bone and / or organ tissues; radioactivity of muscle, VAT, SAT, IMAT, bone and / or organ tissues; volume of muscle, VAT, SAT, IMAT, bone and / or organ tissues; skeletal muscle index; and / or lean body mass, one or more of these. The surface area of ​​the L3 mid-muscle (cm²) can then be used as a metric. 2 The skeletal muscle index (SMI, cm) is calculated by dividing the height (m) by the number of skeletal muscles. 2 / m 2 Women with a SMI value <38.5 cm 2 / m 2 Male SMI value <52.4 cm 2 / m 2 This can be categorized as muscle loss. Lean body mass (LBM) can be calculated as [(L3 mid-muscle surface area (cm²)]].2 () × 0.3) + 6.06).

[0053] The second model is referred to as the predictive model. The predictive model is trained to provide toxicity predictions based on multiple input parameters. This model can be trained on data collected from patients and can receive the aforementioned body composition parameters, as well as other patient and treatment parameters, such as patient age at diagnosis, sex, smoking status, diabetes status, height, weight, BSA, body mass index, tumor location, cancer stage, genetic risk factors, and one or more of the absolute dose of each drug to be administered. The predictive model can be trained to identify an optimal set of parameters for predicting DLT, and then use that set of parameters for prediction. In some embodiments, the predictive model can be trained to output a toxicity rating of the input dose, which may be in the form of a binary toxicity / non-toxicity output. In some embodiments, the predictive model can be trained to output an optimal dose based on patient parameters. In some embodiments, the predictive model can be trained to output a toxicity score or rating, which may correspond to the predicted likelihood or severity of toxicity or side effects. In some embodiments, the predictive model can be trained to output predicted side effects. In some embodiments, the predictive model can be trained to output the predicted severity of side effects. In some embodiments, the predictive model can be trained to output the predicted duration of side effects. In some embodiments, the predictive model can be trained to output a score or rating corresponding to the predicted quality of life of the patient if the input drug dose is administered to the patient. In some embodiments, for example, the predictive model can be configured to output data corresponding to the parameters most important in forming the prediction, such as the parameters of specific body components that most significantly affect the prediction.

[0054] Figure 5 Figure 500 illustrates the accuracy of predictions generated by the predictive model for different combinations of input parameters. The x-axis 510 corresponds to the different combinations of input parameters tested, while the y-axis 520 shows the average accuracy of the predictions when using each combination. Line 530 shows the experimental results, and line 540 shows the maximum accuracy achieved at 0.773. This maximum accuracy was achieved for the tested model using a set of 10 parameters. The optimal set of parameters for identification included: LBM, oxaliplatin, absolute dose of oxaliplatin, absolute dose of 5FU, L3 muscle volume, stage_T, height, allmg, L3 SAT radioactivity, and age at diagnosis. The following table provides further accuracy values ​​for experimental measurements:

[0055]

[0056]

[0057] The table below provides performance measurements for a specific embodiment obtained through experiments:

[0058]

[0059] Some of the described embodiments of the predictive tool utilize one or more of patient demographics, disease, treatment, dosage, toxicity outcome, and CT images (3D) to predict treatment side effects. For example, for a patient with colorectal cancer, the predictive tool may utilize patient demographics, details of the patient's colorectal cancer, chemotherapy treatment, chemotherapy dosage, chemotherapy toxicity outcome, and radiomics details from CT images (3D) to predict chemotherapy-related toxicities. In some embodiments, CT image data may be received in accordance with the Medical Digital Imaging and Communications (DICOM) standard.

[0060] In one example, data from colorectal cancer patients treated over the past decade have been used to validate the AI ​​software and to develop the described predictive tool. In some embodiments, the described method is able to accurately predict 80% of toxic and non-toxic patients.

[0061] In the example above, patient CT scans were accessed, CT images were downloaded, and relevant patient data were identified. Results from patients who received colorectal cancer treatment over a 10-year period (2012–2021) were analyzed. Of the 203 colorectal cancer patients who received chemotherapy (such as oxaliplatin), 120 experienced dose-limiting toxicities. Of these 203 patients, 80% were accurately predicted by AI algorithms in some embodiments disclosed herein to be likely to develop these complications.

[0062] In some embodiments, patient CT scan images can be accessed and processed according to segmentation models of some described embodiments, and 3D body composition reports can be generated. Predictive models of some described embodiments can utilize patient data to generate predictions to identify patients at risk of developing toxicity. The described embodiments can be used in clinical studies to predict the toxicity of new drugs. The described embodiments can be embedded in a health IT cloud, where CT images are automatically available for use in the segmentation model. Clinical patient data can be automatically available for use in the predictive model, such as by sending data from an EMR system to cloud software that can generate predictive reports.

[0063] The described embodiments can determine the distribution and quality of various body tissues, including muscles, fat, bone, and internal organs in 3D segments of the body. Each patient has a different fat distribution, which may be located around one or more specific organs. Some patients may have more fatty marbling in certain areas of the body. The quality of tissues such as muscle, fat, bone, or organ tissues can be indicated by the brightness / darkness (which can be measured in Hounsfield units) of the areas of muscle, fat, bone, and organs captured in a CT scan, and this can vary in different parts of the muscle, fat, bone, or organ. The described embodiments can calculate patterns of fat and muscle from CT scans. According to some embodiments, scores can be given to the patterns of fat and muscle, and in some embodiments, these scores can be used as input parameters for a predictive model. In some embodiments, patterns of muscle and fat can be assigned to groups or categories, and in some embodiments, these groups or categories can be used as input parameters for a predictive model.

[0064] Figure 1 A flowchart of an embodiment of a method 100 for providing drug dosage prediction according to some embodiments is shown. In the illustrated embodiment, the method begins at 101 by accessing one or more CT images. For example, accessing CT images may include downloading them, retrieving them from a memory storage location, or receiving them from an external computing device. In some embodiments, the accessed CT images may be in DICOM format.

[0065] During the training phase, the images received at point 101 constitute the training dataset 102 and are manually labeled as regions representing different tissue types, such as regions representing muscle tissue and regions representing fat. Further annotations can indicate regions of organs and bones.

[0066] At point 103, the annotation algorithm generates an annotated copy of each CT image received at point 101, including annotations for image regions of fat, muscle, bone, organs, or other tissues. In some embodiments, the annotation algorithm may be an AI algorithm. The generated result 105 is sent to an evaluation model.

[0067] At 105, the evaluation module evaluates the accuracy of output 104. This is achieved by evaluating how closely the output annotations generated by annotation algorithm 103 are similar to the hand-annotated images of the training images. The evaluation model 105 can access the hand-annotated images from the training dataset to allow evaluation of the accuracy of annotation algorithm 103. The accuracy estimate 118 is passed back to adjustment module 106, which adjusts the internal parameters of annotation algorithm 103, such as through backpropagation and gradient descent, to improve the accuracy of result 104. By iteratively feeding training data 102 in the form of CT images 101 to annotation algorithm 103 and evaluating its performance, annotation algorithm 103 can be trained to accurately label CT images.

[0068] In some embodiments, the annotation algorithm 103 includes a U-net architecture consisting of a shrinking path and an expanding path, giving it a U-shape. The shrinking path is a convolutional network consisting of repeated applications of convolutions, each followed by a Rectified Linear Unit (ReLU) and a max-pooling operation. During shrinking, spatial information is reduced while feature information is increased. The expanding path combines the features and spatial information with high-resolution features from the shrinking path through a series of up-convolutions and connections. Further details are described below. Figure 2 , Figure 2 The example architecture 200 is illustrated. Note that other neural networks, convolutional neural networks, and various types of AI algorithms can be used in the same way.

[0069] To generate training dataset 102, a set of CT training scans can first be manually “segmented” or labeled by trained clinicians. These labeled images can be stored and used to train the annotation algorithm 103 as described above. The annotation algorithm 103 can be tested using a separate set of CT images. In some embodiments, manual labeling can be performed using software to “color” different body component types on the CT images, enabling the AI ​​algorithm to learn to “color” or segment the images itself.

[0070] In some embodiments, Hounsfield units can be used during training. Hounsfield units are radiometric measurements of how "bright" or "dark" a pixel is on a CT image. The annotation algorithm 103 can be trained to identify patterns and body composition based on the location of regions of specific tissue on the scan and in conjunction with how bright or dark the tissue appears, as indicated by Hounsfield units.

[0071] Details of the example U-Net architecture that can be used by annotation algorithm 103 are described in Olaf Ronneberger, Philipp Fischer, Thomas Brox, U-Net: Convolutional Networks for Biomedical Image Segmentation, Medical Image Computing and Computer-Assisted Intervention (MICCAI), Springer, LNCS, Vol.9351: 234–241, 2015, which is available at arXiv:1505.04597 and is incorporated herein by reference. According to that paper, the input images and their corresponding segmentation maps are used to train the network using a stochastic gradient descent implementation in Caffe. Due to padding-free convolutions, the output images have a constant boundary width smaller than the input images. To minimize overhead and maximize GPU memory utilization, annotation algorithm 103 can be configured to preferentially use large input tiles rather than large batch sizes, thus reducing a batch of input images to a single tile image comprising each image in that batch. Therefore, the annotation algorithm 103 can be configured to use high momentum (e.g., such as 0.99) so that a large number of previously seen training samples determine the update in the current optimization step.

[0072] The energy function can be calculated by combining pixel-wise softmax with the cross-entropy loss function on the final feature map. Softmax is defined as... ,in Marker at pixel position (in ) is the activation of feature channel k at position k. K is the number of categories. It is an approximately maximum function. That is, for a given function with maximum activation... k, For all other k, Then, the cross-entropy at each position is penalized for deviations from 1 using the following formula.

[0073]

[0074] in It is the actual label of each pixel. It is a weight map, which we introduce to give some pixels more importance during training.

[0075] The annotation algorithm 103 can be configured to pre-compute a weight map for each ground truth segment to compensate for the different frequencies of pixels of certain classes in the training dataset that may occur, and to force the network to learn small separation boundaries introduced between contact units.

[0076] Morphological operations can be used to compute the separation boundary. The weight map is then computed as follows:

[0077]

[0078] in It is a weighted graph that balances category frequencies. The distance from the marker to the boundary of the nearest cell. It is the distance to the boundary of the second nearest unit. In our experiment, we set... and Pixel.

[0079] In deep networks with many convolutional layers and different paths through the network, good initialization of weights is useful. Otherwise, some parts of the network may be overactivated while others never contribute. Ideally, the initial weights should be adjusted so that each feature map in the network has approximately unit variance. For networks with a U-net architecture (alternating convolutional layers with ReLU layers), this can be achieved by adjusting the weights from a set of features with a standard deviation of approximately 1 / 2. The initial weights are obtained from a Gaussian distribution, where N represents the number of input nodes to a neuron. For example, for a 3×3 convolution with 64 feature channels in the previous layer, .

[0080] According to some embodiments, this method allows adjustment of hyperparameters such as the learning rate, batch size, and number of epochs. Similar or different architectures can be chosen, such as those available for Google's TensorFlow library.

[0081] Note that annotation algorithm 103 can classify regions in a two-dimensional slice of a CT scan into one or more of muscle, fat, bone, or organs. In some examples, the CT slice may be taken from the third lumbar vertebra (L3). However, it can be difficult to clinically determine whether a particular slice is more clinically relevant than others. Therefore, in some embodiments, the method constructs a three-dimensional body model containing volumetric representations of one or more of muscle, fat, bone, organs, or other tissues. In one example, this involves processing all slices containing the L3 vertebra, processing slices from multiple vertebrae, or processing slices from all vertebrae. As a result, differences in clinical relevance between different CT slices are less correlated, and the final output is more accurate.

[0082] Once the annotation algorithm 103 is trained (i.e., its parameters are adjusted to reduce the difference between the result 104 and the manually labeled images to below a predetermined threshold), the annotation algorithm 103 can be applied to the test dataset 107, which includes other manually labeled CT slices. According to some embodiments, the annotation algorithm 103 is not further adjusted at this stage, but is applied only once to each slice of the test dataset 107 to generate a result 104 for evaluation by the evaluation model 105. If the evaluation model 105 determines that the annotation algorithm 103 has been sufficiently trained, such as by determining that the results of processing the test dataset 107 are within a predetermined level of accuracy, the parameters of the trained annotation algorithm 103 can be output as the best-fit model 108. The best-fit model 108 can be used as the final segmentation model 109. In some embodiments, the segmentation model 109 can continue to be trained and updated with further CT images, which may be CT images from different parts of the body (such as different vertebrae). As mentioned above, if more CT slices from the same or different vertebrae are used to bring "height" into the model, the segmentation model 109 can become three-dimensional. In some embodiments, this additional training is performed by the labeling algorithm 103 before the best-fit model 108 is output to the segmentation model 109.

[0083] When segmentation model 109 is available, method 100 calculates body composition measurements 111 based on input data 110. Body composition measurements 111 are measurements indicating the body composition of the patient in the captured image. These measurements 111 are quantitative measurements based on segmentation model 109. Spatial information determined by segmentation model 109 (which may include labeled regions or volumes) is mapped to one or more numerical measurements. For example, these may include one or more of the following: muscle mass, VAT mass, SAT mass, IMAT mass, bone mass, organ mass, radiometric density of muscle, VAT mass, SAT mass, bone mass, and / or organ mass. The quantities of muscles, VAT, SAT, IMAT, bones, or organs can be calculated as their respective surface areas in a 2D model (i.e., the surface area of ​​muscles, VAT, SAT, IMAT, bones, and organs) or their respective body volumes in a 3D model (i.e., the volume of muscles, VAT, SAT, IMAT, bones, and organs).

[0084] During the training phase, body composition measurements 111 generated based on training dataset 102 and test dataset 107 (and possibly other data from other patients) are used as input parameters to the second classification algorithm 116, as described below. Further inputs include patient and / or treatment parameters 112, which are parameters collected during treatment. These may include further measurements (such as body surface area and lean body mass), data from additional tests such as blood tests, and / or treatment options or parameters (such as the absolute dose or type of medication).

[0085] In this phase, method 100 may include an optional feature selection step to select the optimal combination of available input features (selected from available body composition measurements 111 and patient / treatment parameters 112). For this purpose, method 100 generates multiple feature combinations 113 and corresponding feature subsets 114. In one example, as the first step of feature selection, 13 parameters (including 6 3D body composition measurements and 7 other patient and clinical data) may be included to select the optimal combination as feature 113. In other words, in this example, k=13. Therefore, the number of combinations with 1 feature = 13; the number of combinations with 2 features = 78; the number of combinations with 3 features = 286; the number of combinations with 4 features = 715; ... the number of combinations with 13 features = 1, making a total of 8191 different combinations possible in this example.

[0086] These subsets 114 are then provided to the feature selection module 115, which includes a parameter selection algorithm 116 with adjustment parameters 117, which creates a final subset 120 of parameters determined by the best-fit model. The feature selection module 115 can train the parameter selection algorithm 116 using 10-fold cross-validation via iterative model fitting processing.

[0087] The final subset 120 can be expanded by adding further patient parameters 121 (such as age, sex, tumor location, etc.) that were not measured but captured from patient data. In some embodiments, these parameters can be added one by one to the final subset 120 to test whether their addition improves the accuracy of the final results obtained.

[0088] The prediction algorithm 122 is tuned using an expanded set of parameters and adjusted parameters. The prediction algorithm 122 is trained on data from training dataset 102 and test dataset 107, as well as data from other patients. According to some embodiments, approximately 1000 patients can be used to generate the trained prediction algorithm 122.

[0089] Once it is determined that the prediction algorithm 122 has been sufficiently trained, such as by determining that the results of processing the test database 107 are within a predetermined level of accuracy, the parameters of the trained prediction algorithm 122 can be output as the best-fit model 124. The best-fit model 124 can be used as the final prediction model 123. Once both the segmentation model 109 and the prediction model 123 have been trained, these models can be used to generate results 125 based on the input data 110. In some embodiments, result 125 may correspond to the predicted toxicity of a drug dose to be administered to a patient. In some embodiments, result 125 may correspond to the predicted safe dose of a drug to be administered to a patient. In some embodiments, result 125 may correspond to the optimal dose of a drug to be administered to a patient. In some embodiments, result 125 may correspond to a toxicity rating of the input drug dose, which may be in the form of a binary toxicity / non-toxicity output. In some embodiments, result 125 may correspond to a toxicity score or rating, which may correspond to the predicted likelihood or severity of toxicity or side effects. In some embodiments, result 125 may correspond to predicted side effects. In some embodiments, result 125 may correspond to the predicted severity of side effects. In some embodiments, result 125 may correspond to the predicted duration of side effects. In some embodiments, result 125 may correspond to a score or rating that matches the patient's predicted quality of life if the input drug dose is administered to the patient. In some embodiments, for example, result 125 may include parameters that are most important in forming the prediction, such as specific body composition parameters that most significantly affect the prediction.

[0090] When method 100 generates result 125 based on input data 110, the input data is first fed into segmentation model 109 to obtain body composition measurements 111. Patient / treatment parameters can also be retrieved or obtained from input data 110. The final subset 120 of features 113 (determined during training processing) is fed into prediction model 123 to generate result 125.

[0091] In some embodiments, there may be approximately 30 parameters available for analysis, if not more. Examining each parameter combination may be computationally too expensive. For example, 25 parameters result in 33,554,431 unique combinations. Method 100 can run different combination tasks in parallel; however, computational cost may still be a limitation. As previously mentioned, many patient factors are input into different models. Using repeatability analysis, it may be possible to determine which “mixtures” of patient factors and to what extent they should be incorporated into the “optimal” algorithm, and these parameters can be stored as a final subset 120 for all further processing of the input data 110 after model training has been completed. Another parameter can be added at 121, and it can be determined whether this improves the algorithm / prediction on an ad-hoc basis.

[0092] Different machine learning algorithms were compared, using all available input parameters to build predictive models to distinguish between the DLT group and the non-DLT group, in order to determine the optimal machine learning algorithm. Results showed that regularized generalized methods (i.e., Lasso and Elastic Network Regularized Generalized Linear Models) achieved the highest performance across various performance metrics, although other methods may be more efficient in certain situations. The package "Caret" supports automatic model tuning throughout the training process.

[0093] As the second step in the feature selection process, various feature combinations (but not all) are considered. Method 100 can incrementally add residual parameters (number = k) to the model and save the model with improved performance. Then, Method 100 adds residual parameters (number = k-1)... until the model performance no longer improves.

[0094] Figure 2 An example U-net architecture 200 is shown that can be used to train annotation algorithm 103 and / or segmentation model 109. Architecture 200 is merely an example of a U-net architecture, and alternative architectures may be used in some embodiments. Architecture 200 shows an input image 210 being encoded through a series of convolutional and pooling layers 220. For each layer, the processed image becomes smaller (indicated by the height of each block), but the number of channels processed increases (indicated by the thickness of each block). The image is then decoded via a series of upsampling and convolutional layers 230 to obtain an output segmentation map 240. During upsampling, the image is concatenated with the corresponding image from the pooling step, as shown in skip connection 250.

[0095] Figure 3A The illustration shows an example CT image 300 taken from the first patient, which can be generated by the segmentation module 109. Figure 3A Shaded areas indicate various types of identified tissues. Tissue 310 corresponds to muscle; tissue 320 corresponds to subcutaneous fat; tissue 330 corresponds to visceral fat; and region 340 corresponds to other tissues such as bones and organs.

[0096] Figure 3B The illustration shows another example CT image 350 taken from a second patient, which can be generated by the segmentation module 109. Image 350 has been labeled with the same regions as image 300. However, as can be seen from the illustration, the second patient has a smaller area of ​​muscle tissue 310 and a larger area of ​​subcutaneous fat 320 and visceral fat 330. Figure 4 The illustration shows a method 400 for processing patient data according to some embodiments.

[0097] At point 405, a CT scan was performed on the patient, and at least one CT image was captured.

[0098] At 410, at least one CT image is stored in a memory location, which in some embodiments may be a local storage location on a local computer or on an external computing device or in the cloud.

[0099] At 415, access at least one CT image. This can be done, for example, by downloading at least one CT image from a memory location or retrieving it from an external device.

[0100] At point 109, see above reference. Figure 1 The trained segmentation model 109 is applied to at least one CT image to generate a 2D or 3D segmentation image, wherein the model labels tissues based on its training. Based on this, body composition parameters are determined at 425. Optionally, these measurements are output or stored at 430 to allow clinicians to access them.

[0101] At 420, patient and / or treatment parameters are accessed and added as input data, and these parameters are passed to prediction model 123 for processing together with the determined body composition measurements.

[0102] Finally, the trained prediction model 123 was applied to body composition measurements and patient and / or treatment parameters. The prediction model 123 can be used in the above references. Figure 1 The parameters selected in the parameter selection step. In some embodiments, the segmentation model 109 includes a convolutional neural network because such networks are well-suited for image classification. When the input is not an image, the prediction model 123 can use an alternative network. Therefore, other classification models may perform better.

[0103] At position 435, the output is the result generated by prediction model 123. The output of the prediction model can be a classification of the input drug dose regarding the effect of that dose on the patient. For example, in some embodiments, the output can be one of red, yellow, or green, where...

[0104] Red – This means do not continue. There is a high risk of chemical toxicity. Consider a lower dose.

[0105] Yellow – This indicates that caution is advised. Toxicity may be mild but should be tolerable for the patient. Please monitor the patient carefully.

[0106] Green – means that it is safe to continue with the selected dosage.

[0107] In some embodiments, method 400 can calculate an accurate prescription for an appropriate dose, which may be a dose that has the highest efficacy while reducing the severity, chance, or duration of side effects. In other words, the output at 435 may be an accurate chemotherapy dose.

[0108] In some embodiments, the prediction model 123 is implemented by a package called "Caret," which enables automatic model tuning throughout the training process. Caret is an abbreviation for Classification and Regression Training and is available from the R Statistical Toolkit on the CRAN website https: / / cran.r-project.org / .

[0109] According to some embodiments, the model parameters selected for predictive model 123 can be those parameters with the highest accuracy. The accuracy of the predictive model can be determined by comparing AI classification of the two groups (toxic and non-toxic) with clinical judgment. Accuracy can be the metric for selecting the best model.

[0110] Other parameters that may be included in the model, such as those at 420, may include cancer characteristics, cancer aggressiveness (i.e., STAGE), patient demographics (e.g., age, weight, ethnicity), and / or patient blood parameters. Demographic, biostatistical, and cancer characteristic parameters are available from ACCORD. One or more of age at diagnosis, sex, smoking status, and diabetes status may be included as demographic parameters. Biostatistical parameters may include one or more of height (in cm), weight (in kg), body surface area (BSA), and body mass index (BMI). Cancer characteristic parameters may include one or more of tumor location, cancer staging based on the Australian Clinical Pathology Staging System (ACPS) (i.e., staging_acps), and the TNM staging system. Treatment parameters may include example presentations of various regimens (X represents the amount of dose), such as the FOLFOX regimen: FOLFOX (5FU: 1 mg over 46 hours, 2 mg over 3-5 minutes / oxaliplatin (Oxa) 3 mg / FA 4 mg).

[0111] The model can also incorporate information on genomics, particularly gene mutations that determine the body’s ability to metabolize chemotherapeutic agents. The gene in question could be a deficiency of dihydropyrimidine dehydrogenase (DPD), which is involved in the metabolism of 5-fluorouracil (5-FU) chemotherapy.

[0112] According to some embodiments, blood test data may be included. This may include data relating to the presence of dihydropyrimidine dehydrogenase (DPD) in a patient's blood sample.

[0113] Figure 6 It shows the method for execution Figure 1 and Figure 4An example system 600 of the method is provided. System 600 includes an arrangement of system components, including hardware and software that can be used to perform the methods disclosed herein. Those skilled in the art will readily understand that... Figure 6 The system described is merely one of many potential embodiments suitable for performing this method.

[0114] System 600 includes a clinician computing device 610, which can be controlled by a clinician who wishes to generate drug dosage predictions for a patient. In the illustrated embodiment, system 600 also includes a server system 620. User computing device 610 can communicate with server system 620 via network 630. However, in some embodiments, user computing device 610 can be configured to perform the described methods independently without accessing network 630 or server system 620.

[0115] The clinician computing device 610 may be, for example, a computing device such as a personal computer, laptop computer, desktop computer, tablet computer, or smartphone. The clinician computing device 610 includes a processor 611 configured to read and execute program code. The processor 611 may include one or more data processors for executing instructions, and may include one or more of a microprocessor, a microcontroller-based platform, a suitable integrated circuit, and one or more application-specific integrated circuits (ASICs).

[0116] The clinician computing device 610 also includes at least one memory 612. Memory 612 may include one or more memory storage locations, which may include volatile and non-volatile memory, and may be in the form of ROM, RAM, flash memory, or other memory types. Memory 612 may also include system memory, such as BIOS.

[0117] Memory 612 is arranged to be accessible by processor 611 and to store data that can be read and written by processor 611. Memory 612 may also contain program code 614 executable by processor 611 to enable processor 611 to perform various functions. For example, program code 614 may include a drug dosage prediction application 615. For example, processor 621 executing drug dosage prediction application 615 may be able to execute aspects of a drug dosage prediction method, such as those referenced above. Figure 1 and Figure 4 Some steps are described.

[0118] According to some embodiments, the drug dosage prediction application 615 may be a web browser application (such as Chrome, Safari, Internet Explorer, Opera, or any other alternative web browser application) that can be configured to access the web page providing the functionality via an appropriate Uniform Resource Locator (URL).

[0119] Program code 614 may include Figure 6 Additional applications not shown in the diagram, such as operating system applications, may be mobile operating systems if the clinician computing device 610 is a mobile device, or desktop operating systems, or alternative operating systems, if the clinician computing device 610 is a desktop device.

[0120] The clinician computing device 610 may also include user input and output peripherals 616. For example, these devices may include one or more of a display screen, touchscreen display, mouse, keyboard, speaker, microphone, and camera. User I / O 616 can be used to receive data and instructions from the user and to transmit information to the user.

[0121] The clinician computing device 610 may also include a communication interface 617 to facilitate communication between the clinician computing device 610 and other remote or external devices. The communication module 617 may allow wired or wireless communication between the clinician computing device 610 and the external device, and may use Wi-Fi, USB, Bluetooth, or other communication protocols. According to some embodiments, for example, the communication module 617 may facilitate communication between the clinician computing device 610 and the server system 620 via a network 630.

[0122] Network 630 may include one or more local area networks (LANs) or wide area networks (WANs) that facilitate communication between the elements of system 600. For example, according to some embodiments, network 630 may be the Internet. However, network 630 may include at least a portion of any one or more networks having one or more nodes that transmit, receive, forward, generate, buffer, store, route, switch, process, or a combination thereof, or one or more messages, packets, signals, or some combination thereof. Network 630 may include one or more of, for example, wireless networks, wired networks, the Internet, intranets, public networks, packet-switched networks, circuit-switched networks, ad hoc networks, infrastructure networks, public switched telephone networks (PSTN), cable networks, cellular networks, satellite networks, fiber optic networks, or some combination thereof.

[0123] Server system 620 may include one or more computing devices and / or server devices (not shown), such as one or more servers, databases, and / or processing devices communicating via a network, wherein the computing and / or server devices host one or more applications, libraries, APIs, or other software elements. Components of server system 620 may provide server-side functionality to one or more client applications, such as dose prediction application 615. Server-side functionality may include features such as user account management, login operations, and other features such as data sharing capabilities. According to some embodiments, server system 620 may include a cloud-based server system. Although a single server system 620 is shown, server system 620 may include multiple systems of servers, databases, and / or processing devices. According to some described embodiments, server system 620 may host one or more components of a platform for performing drug dose prediction.

[0124] Server system 620 may include at least one processor 621 and memory 622. Processor 621 may include one or more data processors for executing instructions, and may include one or more of a microprocessor, a microcontroller-based platform, a suitable integrated circuit, and one or more application-specific integrated circuits (ASICs). Memory 622 may include one or more memory storage locations, and may be in the form of ROM, RAM, flash memory, or other memory types.

[0125] Memory 622 is arranged to be accessible by processor 621 and contains data 623 configured to be read and written by processor 621. Data 623 may store data such as user account data, CT image data, and data relating to machine learning models trained to perform segmentation and prediction functions.

[0126] In the illustrated embodiment, data 623 includes training data 625, body composition measurements 111, and patient / treatment parameters 112. While this data is illustrated as residing in memory 622 of server system 620, in some embodiments, some or all of this data may alternatively or additionally reside in memory 612 of clinician computing device 610, or in alternative local or remote memory locations.

[0127] Training data 625 may include CT images 101 containing training dataset 102 and test dataset 107, as referenced above. Figure 1 As described above, the body composition measurement 111 can be stored in data 623 after being generated by the segmentation model 109, as referenced above. Figure 1 In some embodiments, patient and / or treatment parameters 112 may be stored in data 623 after being received from the clinician's computing device 610.

[0128] The memory 622 also includes program code 624 executable by the processor 621 to cause the processor 621 to perform a workflow. For example, program code 624 may include a server application executable by the processor 621 to cause the server system 620 to perform server-side functions. According to some embodiments, such as when the dose prediction application 615 is a web browser, the server application may include a web server, such as Apache, IIS, NGINX, GWS, or an alternative web server. In some embodiments, the server application may include an application server specifically configured to interact with the dose prediction application 615. The server system 620 may be provided with both a web server module and an application server module.

[0129] Program code 624 may also include one or more code modules, such as one or more of segmentation module 626 and prediction module 628. Segmentation module 626 and prediction module 628 may be configured to cause processor 621 to perform the functions of segmentation model 109 and prediction model 123, as referenced above. Figure 1 and Figure 4 As stated above.

[0130] The segmentation module 626 and prediction module 628 may be software modules, such as additional modules or plug-ins that operate in conjunction with the dose prediction application 615 to extend its functionality. In alternative embodiments, modules 626 and / or 628 may be built into the dose prediction application 615. In alternative embodiments, modules 626 and / or 628 may be standalone applications (running on user computing device 610, server system 620, or alternative server system (not shown)) that communicate with the dose prediction application 615, such as via network 630.

[0131] Modules 626 and 628 have been described and illustrated as part of / installed on server system 620, and can be configured as additional modules or extensions to server applications, separate standalone server applications communicating with server applications, or built-in parts of server applications. Inputs such as user interactions, patient data, and / or CT images can be provided at and / or received by clinician computing device 610, and then transmitted to server system 620, enabling dose prediction methods to be executed by components of server system 620.

[0132] In some alternative embodiments (not shown), the functionality provided by one or more of modules 626 and / or 628 may alternatively be provided by the clinician computing device 610 based on locally or remotely stored data. One or more of modules 626 and / or 628 may reside as an additional module or extension to the dose prediction application 615, a separate standalone application communicating with the dose prediction application 615, or a built-in part of the dose prediction application 615.

[0133] In alternative embodiments (not shown), all functions may be performed by server system 620. Alternatively, in some embodiments, an application programming interface (API) may be used to interface with server system 620 to perform the techniques disclosed herein.

[0134] Server system 620 may also include a communication interface 627 to facilitate communication between server system 620 and other remote or external devices. Communication module 627 may allow wired or wireless communication between server system 620 and external devices, and may use Wi-Fi, USB, Bluetooth, or other communication protocols. According to some embodiments, for example, communication module 627 may facilitate communication between server system 620 and clinician computing device 610.

[0135] Server system 620 may include additional functional components illustrated and described, such as one or more firewalls (and / or other network security components), load balancers (for managing access to server applications), and other components.

[0136] A retrospective study was conducted on stage 3 colon cancer patients who received oxaliplatin treatment after surgery at a single tertiary healthcare facility. A validated AI algorithm was used to calculate body composition from staging CT scans. The oxaliplatin dose / lean body mass toxicity cut-point was determined from participant operating characteristic analysis.

[0137] Between 2012 and 2021, 129 patients were identified (53% male, mean age 58 years (range 25–80 years)). Surgical procedures mainly included right hemicolectomy / extended right hemicolectomy (n=55), anterior hemicolectomy (n=44), Hartmann procedure (n=12), and other procedures (n=18).

[0138] Eighty-four (65.1%) patients experienced dose-limiting toxicities, and these toxicities were higher in female patients (p=0.025). Women had significantly lower muscle mass index and higher rates of obesity (p<0.001). Lean body mass (BSA) was similar in both the DLT and non-DLT groups. Lean body mass was weakly correlated with BSA (R0.025). 2=0.514, p=0.001). The optimal oxaliplatin cutoff point was identified as 3.41 mg / kg lean body mass. During the first four cycles of chemotherapy, 37 patients experienced DLT, of whom 29 (78%) received a dose equal to or greater than the predicted cutoff point (p=0.05).

[0139] This demonstrates that using AI-automated body composition measurements to predict the risk of drug-related toxicities can successfully predict early treatment toxicities.

[0140] Those skilled in the art will understand that various changes and / or modifications can be made to the above embodiments without departing from the broad scope of this disclosure. Therefore, these embodiments are to be considered illustrative rather than restrictive in all respects.

Claims

1. A method for predicting the toxicity of a drug to be administered to a patient, the method comprising: Access at least one computed tomography (CT) slice of the torso associated with the patient; The labeling process for the at least one CT scan is performed using a trained artificial intelligence (AI) segmentation model. At least one body composition parameter is determined based on at least one labeled CT slice; Receive at least one parameter associated with the patient; Receive at least one dosage parameter associated with the drug to be administered to the patient; as well as An output is generated using a trained AI prediction model based on at least one body composition parameter, at least one demographic parameter, and at least one dosage parameter, the output corresponding to the predicted toxicity of the drug administered to a patient.

2. The method of claim 1, wherein the method further comprises training the AI ​​prediction model.

3. The method of claim 1 or 2, wherein the labeled CT slices comprise one or more or all of the labeled regions of muscle, visceral adipose tissue (VAT), subcutaneous adipose tissue (SAT), intramuscular / intramuscular adipose tissue (IMAT), bone, and organs.

4. The method as described in any of the preceding claims, wherein the body composition parameters are at least one or more of the following: amount of muscle, amount of VAT, amount of SAT, amount of IMAT, amount of bone, amount of organ, radiometric density of muscle, radiometric density of VAT, radiometric density of SAT, radiometric density of IMAT, radiometric density of bone, or radiometric density of organ.

5. The method according to any one of claims 1 to 4, wherein the drug is a chemotherapy drug for treating one of breast cancer, lung cancer, liver cancer, colon cancer, intestinal cancer, prostate cancer, colorectal cancer, and pancreatic cancer.

6. The method of any one of claims 1 to 5, wherein the patient-associated parameters include parameters associated with the treatment to be administered to the patient.

7. The method of any one of claims 1 to 6, wherein the patient-associated parameters include parameters obtained from blood tests.

8. A method for training an artificial intelligence model to assess drug dosage in order to minimize toxic reactions to a drug, the method comprising: Receive samples of labeled CT trunk slices associated with patient samples; At least one body composition parameter is determined based on each labeled CT slice; A set of predictive parameters is identified for each patient, the predictive parameters including at least one body composition parameter, at least one patient demographic parameter, and at least one dose parameter; For each combination of prediction parameters, a model fitting process is performed to enable the model to generate a predicted toxicity rating for each patient; For each combination of predicted parameters, the predicted toxicity rating is compared with the measured toxicity rating to determine the accuracy of the combination of predicted parameters; and Select the optimal combination of prediction parameters, wherein the optimal combination of prediction parameters is the combination of prediction parameters that leads to the highest accuracy.

9. The method of claim 8, wherein the labeled CT slices comprise one or more or all of the labeled regions of muscle, VAT, SAT, IMAT, bone, and organs.

10. The method of claim 8 or 9, wherein the body composition parameters are at least one or all of the following: amount of muscle, amount of VAT, amount of SAT, amount of IMAT, amount of bone, amount of organ, radiometric density of muscle, radiometric density of VAT, radiometric density of SAT, radiometric density of IMAT, radiometric density of bone, or radiometric density of organ.

11. The method of any one of claims 8 to 10, wherein the method further comprises using a trained AI model to perform labeling of CT slice samples to generate a sample of labeled CT slices.

12. The method as described in any of the preceding claims, wherein the CT slices are taken from the abdomen.

13. The method of any of the preceding claims, wherein between two and one thousand CT slices are received.

14. A method for minimizing toxic reactions to a drug in a patient, the method comprising: i) Determine the amount of muscle, visceral adipose tissue (VAT), subcutaneous adipose tissue (SAT), and intramuscular / intramuscular adipose tissue (IMAT) in at least a portion of the patient's trunk. ii) Based at least on i) determining the patient's risk of toxic reactions to the drug, and iii) If a patient is determined to be at risk of toxicity, then reduce the standard recommended dose of the drug.

15. The method of claim 14, further comprising administering the drug.

16. The method of claim 14 or 15, wherein step i is performed using the method of any one of claims 1 to 7, 12 or 13.

17. The method of any one of claims 14 to 16, wherein the drug is a chemotherapy drug for treating one of breast cancer, lung cancer, liver cancer, colon cancer, intestinal cancer, prostate cancer, colorectal cancer, and pancreatic cancer.

18. A non-transitory computer-readable storage medium storing instructions that, when executed by a processing device, cause the processing device to perform the method as claimed in any one of claims 1 to 17.

19. A computer system configured to generate text, the computer system comprising: processor; as well as The storage medium as described in claim 18.

Images