Machine learning systems and methods for diagnosing rare diseases

Abstract

The present specification relates to the use of machine learning models for diagnosing patients with rare diseases based on medical data / records, and methods for training such models. According to a first aspect of the present specification, a computer-implemented method for generating a training dataset for training a machine learning model to identify individuals with rare diseases is described, the method including: receiving an initial dataset including medical data related to a plurality of individuals with a rare disease, wherein the medical data for each individual includes data indicating a plurality of symptoms of the rare disease experienced by the individual; identifying a plurality of clusters of individuals in the initial dataset using an unsupervised clustering method; identifying one or more of the clusters as least representative of the rare disease based on the medical data of the individuals in the clusters; removing one or more of the individuals from the one or more clusters identified as least representative based on the medical data of the one or more individuals to generate a pruned dataset; and combining the pruned dataset with a control dataset including a plurality of individuals without the rare disease to generate a training dataset.

Description

[Technical Field]

[0001] This specification relates to the use of machine learning models to diagnose patients with rare diseases based on medical data / records and methods for training such models. [Background technology]

[0002] It is estimated that approximately 300 million people worldwide suffer from rare diseases. Rare diseases typically require longer diagnosis times than more common diseases, with the average diagnosis time for a rare disease exceeding four years. This can lead to significant delays in treating rare diseases and potentially lower the chances of successful treatment. This delay is due to many factors, including physician unfamiliarity with rare diseases, the variable symptoms of a given rare disease, and the disease being masked by symptoms of more common diseases. While traditional diagnostic algorithms can sometimes be effective, they rely on HCP awareness and require validation of numerous clinical features, including differential diagnoses. These conditions are rarely met in the real world. Summary of the Invention [Means for solving the problem]

[0003] According to a first aspect of the present specification, a computer-implemented method for generating a training dataset for training a machine learning model to identify individuals with a rare disease is described. The method includes: receiving an initial dataset including medical data related to a plurality of individuals with a rare disease, wherein the medical data for each individual includes data indicative of a plurality of symptoms of the rare disease experienced by the individual; identifying a plurality of clusters of individuals in the initial dataset using an unsupervised clustering method; identifying one or more of the clusters as least representative of the rare disease based on the medical data of the individuals in the clusters; removing one or more of the individuals from the one or more clusters identified as least representative based on the medical data of the one or more individuals to generate a pruned dataset; and combining the pruned dataset with a control dataset including a plurality of individuals without the rare disease to generate a training dataset.

[0004] Identifying one or more of the clusters as least representative of the rare disease may include: identifying a representative symptom of the rare disease for each of the plurality of clusters; comparing the representative symptoms for each cluster to a predetermined set of known symptoms of the rare disease, the predetermined set of known symptoms including a set of more representative symptoms and a set of least representative symptoms; and identifying the cluster as least representative of the rare disease if the representative symptom of the cluster is included in the set of least representative symptoms.

[0005] Removing one or more individuals from one or more clusters identified as least representative can include: for each individual in the cluster identified as least representative: determining whether medical data associated with the individual meets a threshold condition, the threshold condition being based on symptoms of a rare disease; and removing the individual from the identified cluster if the threshold condition is not met. The threshold condition can include a threshold number of symptoms of the rare disease. The threshold condition can alternatively or additionally include having at least one symptom of the rare disease in a distinctive subset of symptoms of the rare disease.

[0006] Identifying a plurality of clusters of individuals in the initial dataset may include clustering the dataset into a predetermined number of clusters using hierarchical agglomerative clustering.

[0007] The method may further include: comparing medical data associated with individuals in the pruned dataset to medical data associated with individuals in the control dataset to identify one or more potential symptoms of the rare disease; and augmenting the training dataset with one or more potential symptoms of the rare disease.

[0008] The method may further include: using natural language processing to extract one or more potential symptoms of the rare disease from literature related to the rare disease; and augmenting the training dataset with the one or more potential symptoms of the rare disease.

[0009] The control dataset can include individuals without the rare disease who have at least a threshold number of symptoms of the rare disease.

[0010] Combining the pruned dataset with the control dataset can include matching a predetermined proportion of individuals in the control dataset to each individual in the pruned dataset, and the matching can be based on one or more demographic characteristics of the individuals.

[0011] According to a further aspect of the present specification, a computer-implemented method is described for training a machine learning model to classify an individual as having a rare disease based on the individual's medical records, the method including using supervised learning techniques to train the machine learning model using a labeled training dataset, the training dataset having been generated using any of the methods described herein for generating a training dataset.

[0012] The machine learning model can be trained on a subset of the data in the labeled dataset; the subset includes, for each individual with a rare disease in the dataset, medical data collected before the individual was diagnosed with the rare disease.

[0013] According to a further aspect of the present specification, a computer-implemented method for diagnosing a disease is described, the method including: inputting medical data associated with an individual into a machine learning model; processing the input medical data using the machine learning model to generate data indicating whether the individual associated with the medical data has the disease; and outputting data from the machine learning model indicating whether the individual associated with the medical data has the disease, wherein the machine learning model has been trained using any of the training methods described herein.

[0014] According to a further aspect of the present specification, a system is described that includes one or more processors and a memory, the memory including computer readable code that, when executed by the one or more processors, causes the system to perform any one or more of the methods described herein.

[0015] According to a further aspect of the present specification, a computer program product is described that includes computer readable code that, when executed by a computing device, causes the computing device to perform any one or more of the methods described herein.

[0016] As used herein, the term "rare disease" is preferably used to mean a disease that affects less than 1 in 2000 people in the general population. Currently, there are more than 6,000 known rare diseases, and new rare diseases are constantly being discovered. In the following written description, an example method related to Gaucher disease is described, but it will be understood that the method is equally applicable to other rare diseases.

[0017] Embodiments will now be described, by way of non-limiting example, with reference to the accompanying drawings, in which: [Brief explanation of the drawings]

[0018] [Figure 1] FIG. 1 illustrates an overview of an example pipeline for training a machine learning model to identify individuals with a rare disease. [Figure 2] FIG. 1 illustrates an overview of an example method for generating a training dataset used to train a machine learning model to identify individuals with a rare disease. [Figure 3] FIG. 1 shows an example of the identity of a cluster of patients with Gaucher disease identified by a clustering algorithm. [Figure 4] FIG. 1 illustrates an example of a method for narrowing down a dataset of patients with a rare disease. [Figure 5] FIG. 1 illustrates an example of a method for combining a rare disease dataset with a control dataset. [Figure 6] FIG. 1 illustrates an example of a method for extending a training dataset with additional features. [Figure 7]1 is a flowchart of an example method for generating a training dataset for training a machine learning model to identify individuals with a rare disease. [Figure 8] FIG. 1 illustrates an example of a computing system / device for performing the methods described herein. DETAILED DESCRIPTION OF THE INVENTION

[0019] This document describes a method for training a machine learning model to identify patients (also referred to herein as individuals) at risk of having a rare disease. The machine learning system uses a symptomatic approach to identify individuals at risk of a rare disease. The system is trained using electronic medical records of a substantial number of patients. The trained model can be used to determine whether an individual has or is at risk of having a particular rare disease based on the individual's symptoms / electronic medical records. This allows patients with rare diseases to be identified and caught early in the disease progression, thereby enabling them to receive effective treatment and avoid undue morbidity and burden.

[0020] Due to the low proportion of patients with rare diseases in the general population, datasets that can be used to train machine learning models to diagnose rare diseases are often noisy and / or unbalanced. This specification describes methods for generating reduced noise and / or balanced datasets that, when used to train machine learning models, result in more accurate models (i.e., models with fewer false positives and / or false negatives).

[0021] 1 illustrates an example pipeline 100 for training a machine learning model to identify individuals with a rare disease. Method 100 can be performed by a computer. Each of operations 102-110 of pipeline 100 can be performed independently or in combination with any one or more of the other operations.

[0022] The pipeline 100 includes a dataset generation operation 102 , a model training operation 104 , a model testing operation 106 , a model validation operation 108 , and a model output operation 110 .

[0023] During the dataset generation operation 102, a training dataset 112 is generated by combining a target dataset 114 containing medical data of a plurality of individuals with a rare disease with a control dataset 116 containing medical data of a plurality of individuals without the disease. Each individual in the training dataset 112 is associated with a binary label indicating whether the individual has a rare disease or not. The method of generating the training dataset is described in more detail below with reference to FIG. 2.

[0024] The medical data for each individual includes one or more symptoms experienced by the individual. The symptoms include known symptoms of rare diseases and may further include other symptoms (or combinations of symptoms) that may not be specifically associated with rare diseases. For a given individual, the medical data may include all recorded symptoms experienced by the individual. The medical data may include medical procedures and conditions (e.g., indications, treatments, comorbidities) associated with the individual. The medical data may further include one or more of: physiological data; vital sign data (e.g., BMI, blood pressure, pain score); laboratory values ​​(e.g., A1C, eGFR, FEV1); biomarkers; medications received by the individual; and / or individual details (e.g., age, sex, smoking status, ethnicity). In some embodiments, each individual in the training dataset is associated with an embedding vector representing the individual that is based on natural language embeddings of features in the medical data associated with that individual.

[0025] In some embodiments, data regarding the individual's interactions with the healthcare system can also be included in each individual's healthcare data. Interactions can include one or more of: visits (i.e., the type of visit to the healthcare provider, e.g., emergency, inpatient, outpatient, etc.); encounters (i.e., the type of interaction the patient had, e.g., house call, imaging, etc.); and / or providers (i.e., the type of healthcare professional the patient saw, e.g., neurologist, cardiologist, etc.). Interactions can be coded with a binary flag depending on whether the individual had that type of interaction. Alternatively or additionally, the frequency of the interaction can be provided. Alternatively or additionally, the date the interaction first occurred for the individual can be provided.

[0026] The medical data may further include demographic data such as location, gender, and / or ethnicity.

[0027] Those skilled in the art will be familiar with other examples of medical data that can be used.

[0028] In some embodiments, symptoms associated with each individual are associated with the age of first occurrence of the symptom in that individual. This allows the machine learning model 120 to take into account the correlation between symptoms and age. Such associations can aid in training the machine learning model 120 to detect early-onset symptoms. A machine learning model 120 trained on such data can favor symptoms with different ages of onset between control patients and rare disease patients, and can favor early-onset / younger patients. Alternatively or additionally, symptoms can be flagged as binary present / absent. A machine learning model 120 trained on such data can favor symptoms with different prevalence between control patients and rare disease patients, and can favor an accumulation of comorbidities / older patients.

[0029] The generated training dataset 112 can be split into a first dataset 118 (also referred to herein as the “learning dataset”) used to train the machine learning model 120 and a second dataset 122 (also referred to herein as the “test dataset”) used to test the trained machine learning model 120.

[0030] During model training operations 104, a machine learning model 120 (also referred to herein as a “model”) is trained on a training dataset 118 using supervised machine learning techniques to classify individuals as either having or not having a rare disease based on medical data associated with the individuals input into the model 120. Input data 124 including medical data associated with individuals in the first dataset 118 is input into the model 120, which processes the input data to generate a proposed classification 126 for the individual, the proposed classification 126 indicating whether the model should classify the individual as having or not having a rare disease. The proposed classification 126 is output by the model 120 and compared to known classifications of the input data 124 to determine updates 128 for the parameters of the model 120. This comparison can be performed, for example, using a loss / objective function, with the training goal being to optimize the loss / objective function. The loss function can be a classification loss, such as a cross-entropy loss. This process is repeated on the training data until a threshold condition is met, for example, a threshold number of training epochs.

[0031] The machine learning model 120 can be any type of trainable model that can be used for classification. For example, the model can be a logistic regression model. The logistic regression model can have a lasso penalty term. The lasso penalty applies a penalty to some parameters / coefficients of the model that do not contribute a signal to the model. The lasso penalty allows features that do not contribute any signal to be removed from the model 120, while also providing interpretability of the parameters / coefficients of the model 120: i.e., the larger the coefficient, the more important the variable. The logistic regression can have a shrinkage parameter between 0.2 and 3. Exponential increments can be used for a threshold number of values, for example, 10 values.

[0032] Alternatively, model 120 may be a Light Gradient Boosting Model (LGBM). LGBMs attempt to improve predictive power by training a sequence of weak models (e.g., decision trees), each of which compensates for the weaknesses of its predecessor. The use of LGBMs allows for feature selection (e.g., via trees), works well with a large number of features, and accounts for symptom groups. The LGBM model may have a maximum depth of, for example, 10 to 40, e.g., 10, 20, or 30. A minimum child node weight of 0.05 to 0.25, e.g., 0.1 to 0.2, may be used. The number of child leaves may be 10 to 40, e.g., 10, 20, or 30. The number of estimators (trees) may be 50 to 350, e.g., 100, 200, or 300. An example of an LGBM is described in "LightGBM: A Highly Efficient Gradient Boosting Decision Tree" (Ke, G. et al., 2017, Advances in Neural Information Processing Systems 30, Eds Curran Associates, Inc., pp. 3146-3154).

[0033] Those skilled in the art will appreciate that other types of machine learning model 120 may alternatively be used; for example, model 120 may be a neural network, such as a recurrent neural network (RNN), or a support vector machine. RNNs may outperform other models in patient identification tasks on EHR data and may require less preprocessing than traditional methods. In the case of rare diseases, RNNs may be able to capture useful temporal patterns, such as recurrence of significant symptoms, co-occurrence of symptoms, or specific consecutive visits to a specialist.

[0034] In some embodiments, multiple different machine learning models are trained and the best performing model is selected using the area under the average precision-recall curve (AUPRC). The multiple models can include different types of models and / or the same type of model with different hyperparameters.

[0035] In some embodiments, the training data initially consisted of Z control patients for one GD patient (Y < Z). In some embodiments, X is 10, although other values may be used. In some embodiments, Y = X.

[0036] In some embodiments, to limit sample bias and improve the stability of the resulting model, k-fold cross-validation or bootstrap sampling can be used, i.e., the cross-validation is performed X times by selecting Y control patients for one GD patient from the training dataset each time. Within each bootstrap, the control patients are unique (i.e., there is no replacement). The control patients can be selected in some bootstraps (i.e., selection with replacement between bootstraps).

[0037] A good candidate model for the identification of rare diseases is expected to output continuous probabilities while being trained on a binary target, to account for age encoding of features, to function well with a mixture of binary and numerical features, and / or not necessarily have additive sensitivity to the presence of features (e.g., higher weights in the case of multiple symptoms).

[0038] To ensure the robustness of the algorithm, the distribution of the 10 best AUPRCs (one for each bootstrap) can be analyzed. The final algorithm can be a randomly chosen bootstrap of 10 choices (i.e., a training dataset with a ratio of 1 GD to 10 controls, and its associated best hyperparameters determined by cross-validation).

[0039] In some embodiments, a model is trained (104) / evaluated (106) using a subset of medical data for one or more individuals in the training dataset 112. The subset of medical data can exclude any symptoms of the identified rare disease and / or medical data collected after the diagnosis of the individual with the rare disease, i.e., can include only symptoms and / or medical data obtained before the date of diagnosis of the rare disease. In some embodiments, the subset of medical data includes only medical data collected before a predetermined time before the symptoms of the identified rare disease and / or the diagnosis of the individual with the rare disease. For example, only data prior to a predetermined number of visits prior to diagnosis (e.g., one or two visits prior to diagnosis) can be used. For individuals without a rare disease, all medical data can be used.

[0040] Following the training operation 104, the performance of the model 120 can be evaluated against a second (test) dataset 122 during a testing operation 106. Input data 130 including medical data associated with individuals in the second dataset 122 is input to the model 120, which processes the input data to generate proposed classifications 132 of the input data. The proposed classification for each individual in the test dataset 122 is compared to the known classification for that individual, and an accuracy score for the model is determined based on the results. If the model 120 meets a threshold accuracy condition with respect to the test dataset 122, the model 120 is output 110 for use in, for example, identifying individuals with a rare disease in a database comparable / compatible to that used for training. In some embodiments, the model may alternatively or additionally be output to a validation operation 108. Otherwise, the model 120 may be returned to the training operation 104 for further refinement.

[0041] In some embodiments, the testing operation 106 can include a first testing operation in which the model 120 is applied to a test dataset 122 without censoring any events during the observation period to evaluate how the model 120 performs in conditions approximating real-life applications. As an example, the model can be evaluated using a patient's / patients' complete EHR history up to that date. The testing operation 106 can alternatively or additionally include a second testing operation in which events occurring after the individual's index date (i.e., the date of rare disease diagnosis) are censored (i.e., removed from the input to the model). This allows for evaluation of whether the model can identify rare disease patients using the same amount of information as a physician prior to diagnosis.

[0042] Model validation operations 108 are used, in some embodiments, to evaluate the real-world performance of the trained model 120 in a compatible database (e.g., an independent HER system). The performance of the model 120 is evaluated using an additional dataset 134 that includes unlabeled medical data associated with multiple individuals that were not in the training dataset. Input data 136 including medical data associated with individuals in the additional dataset 134 is input to the model 120, which processes the input data to generate a proposed classification 138 of the input data. The input can be selected from a set of individuals in the additional dataset 134 who have at least a threshold number of symptoms of the rare disease. For example, the threshold number can be at least two symptoms of the rare disease.

[0043] The proposed classification for each individual in the further dataset 134 identified as having a rare disease is evaluated by a medical expert 140 to verify the diagnosis. If the model's performance meets a threshold condition, e.g., if the model's accuracy is higher than a threshold when compared to a physician's evaluation, the model is output for use 110. If not, the model 120 can be returned to training operation 104 for further refinement.

[0044] Figure 2 illustrates an example method 200 for generating a training dataset 210 used to train a machine learning model to identify individuals with a rare disease. The method 200 can be performed by a computer. The method corresponds to the dataset generation operation 102 of Figure 1. The method can operate on a medical database that includes medical data from a plurality of individuals.

[0045] Method 200 includes an initial target dataset creation operation 202, a clustering operation 204, a refinement operation 206 (also referred to as "pruning"), and a merging operation 208. Method 200 may further include an expansion operation 216, which may be performed either before or after merging operation 208, in some embodiments.

[0046] During the initial target dataset creation operation 202, a plurality of individuals identified as having a given rare disease and / or receiving a treatment associated with the rare disease (also referred to herein as "RD individuals") are identified in a medical database. From the plurality of individuals identified as having the rare disease, an initial cohort of RD individuals is selected.

[0047] One or more acceptance criteria can be applied to select a plurality of individuals with RD. The acceptance criteria can include symptom-based threshold criteria. For example, a plurality of features associated with a rare disease can be identified by examining literature related to the rare disease, and one or more of the acceptance criteria can include having a threshold number of features of the rare disease. The features can include a set of symptoms of the rare disease. The set of symptoms can be divided into subsets associated with known subtypes of the rare disease. As an example, in the case of Gaucher disease, each feature can be associated with one or more of the disease's three subtypes. A plurality of features are identified as "representative" features of the disease, i.e., characteristic of the rare disease. One or more of the features may be classified as "less representative" features, e.g., common symptoms, that are present in individuals with the rare disease but are not themselves characteristic of the disease. The threshold criteria can include having a threshold number of identified disease symptoms. The threshold number can be between one and four of the identified symptoms, e.g., at least two.

[0048] In some embodiments, the acceptance criteria may further include having at least a threshold age in the medical database, for example, having an age of at least one year.

[0049] In some embodiments, the acceptance criteria further include having at least a threshold number of independent diagnoses of rare diseases. The threshold number of independent diagnoses is at least 2. Depending on the rare disease, the threshold number may be higher.

[0050] As a specific example, a first plurality of individuals diagnosed with a rare disease are identified in a medical dataset. The identified individuals have been diagnosed with at least two of the rare diseases. A second plurality of individuals are identified who are receiving treatment related to the rare disease but have not been diagnosed with the rare disease. Individuals from the first and second plurality who meet the acceptance criteria are combined to form an initial cohort, for example, only individuals with at least two symptoms of the rare disease are included in the initial cohort.

[0051] In some embodiments, individuals can be excluded from the initial cohort based on one or more exclusion criteria. Exclusion criteria can include the individual having an inconsistent timeline (e.g., first reported date of activity after the index date; last reported date of activity before the index date; and / or date of death before the index date). Exclusion criteria can alternatively or additionally include having one or more other diseases, for example, having other diseases of a similar type (which may indicate a misdiagnosis of the subject's rare disease) or having other diseases affecting a similar lineage.

[0052] For the RD individuals, an initial target dataset 202 is generated from the medical data, and the initial target dataset 202 includes, for each RD individual, an indication of which disease-associated features are present in that RD individual. For example, each individual can be associated with a vector of binary labels, each component of the vector indicating the presence or absence of a rare disease-associated feature. Each RD individual can also be associated with demographic data for that RD individual.

[0053] The initial target dataset 202 may, in some embodiments, undergo further processing in clustering 204 and refinement operations 206 to produce a cleaner / more balanced dataset for training the machine learning model.

[0054] During clustering operation 204, an unsupervised clustering algorithm is applied to the initial target dataset to cluster the RD individuals into a number of subsets (also referred to herein as clusters). The data may be clustered into a predetermined number of subsets. The predetermined number of subsets depends on the characteristics of the rare disease and may be, for example, a fraction of the number of identified symptoms and / or the number of subtypes of the rare disease.

[0055] One example of such an unsupervised clustering algorithm is Hierarchical Agglomerative Clustering (HAC), which clusters data into a predetermined number of subsets. HAC regroups observations from the bottom up. Each individual starts with its own cluster, and the clusters are grouped together until the predetermined number of subsets is reached. To determine which ones to collect, a measure of dissimilarity is calculated between sets of observations. The dissimilarity measure is calculated using a metric and a linkage criterion. The metric is used to calculate the distance between each observation. The linkage criterion determines the distance between each set of observations. It will be understood that other unsupervised clustering algorithms may alternatively be used. In some embodiments, a p-value between clusters is calculated to assess whether the clusters could have been obtained by chance.

[0056] Once clustered into subsets, characteristic features of each subset / cluster are identified. Characteristic features can be identified based on common symptoms among RD individuals within each subset, for example, the most common symptom / symptom combination within each subset can be identified as a representative feature for that subset. Characteristic features can alternatively or additionally be based on a symptom type, for example, all individuals within a cluster may have neurological symptoms. Characteristic features can alternatively or additionally be based on demographic characteristics of individuals within a cluster, such as the individual's age, for example, "young patient with symptom X." There can be many other examples of characteristic features.

[0057] Figure 3 shows an example of a subset 300 for Gaucher disease identified by a clustering algorithm. In this example, individuals with Gaucher disease (also referred to herein as "GD individuals") are clustered into 10 clusters / subsets 302-320. Each cluster is assigned a unique feature based on the characteristics of the GD individuals within the cluster. In the illustrated example, the clusters are labeled based on the predominant symptom / symptom combination and / or patient type within the cluster: organomegaly only 302; organomegaly and dysarthria 304; secondary neurological impairment 306; Parkinsonism 308; bone loss, anemia, cytopenia, and organomegaly 310; healthy patients with a high rate of anemia 312; severe bone and respiratory disease 314; severe neurological symptoms at a young age 316; severe visceral damage and bone disease; and severe neurological impairment with mild organomegaly.

[0058] With the exception of the severe visceral and bone disease cluster, individuals within each cluster are associated with one of three types of GD. The type I cluster also correlates with age.

[0059] In some embodiments, the features identified using the clustering operation 204 can be used as feature labels for the training dataset.

[0060] Returning to FIG. 2 , following the clustering operation 204, in some embodiments, a refinement operation 206 can be performed on the initial target dataset to generate a refined dataset (also referred to herein as a “pruned” dataset). During the refinement process, one or more of the clusters are identified as having the lowest representation of rare diseases based on the medical data of the RD individuals in the subset. One or more of the RD individuals from these clusters are then discarded 212 based on their medical data, for example, if the medical data does not meet additional threshold conditions. Refining the dataset in this manner by removing individuals with little information can reduce the amount of noise in the dataset while maintaining enough information to train an accurate model.

[0061] Identifying the least representative clusters of rare diseases can include comparing the distinctive features of each cluster to a list of identified features of the disease, and identifying clusters whose distinctive features match known least representative features of rare diseases or common features of rare diseases.

[0062] For clusters identified as having the lowest representation of rare diseases, additional threshold conditions are applied to each RD individual in the cluster to determine whether to retain or discard that RD individual from the training dataset. The additional threshold conditions are more stringent than the threshold conditions used when creating the initial target dataset. RD individuals in the least representative clusters can be removed if they have fewer than a threshold number of rare disease symptoms. The threshold number may be higher than the threshold number used when creating the initial cohort. For example, if an individual is included in the initial cohort if they have at least two symptoms of the disease, individuals in the least representative clusters may be discarded if they have fewer than three symptoms. Alternatively or additionally, more stringent threshold conditions may include a requirement that at least one of the symptoms experienced by the individual be a unique symptom; for example, if an individual has only the least representative / common symptoms, data associated with that individual is discarded.

[0063] Figure 4 shows an example of a method 400 for narrowing down the Gaucher disease clusters of Figure 3. The initial cohort in this example has 610 individuals. The method includes identifying 402 one or more clusters 404 as least representative of GD, for example, by comparing distinctive features of the clusters to a list of features least representative of GD. In this example, three clusters from Figure 3 are identified as least representative: the "control" cluster; the "anemia" cluster; and the "minimal neurological symptoms" cluster.

[0064] Following identification 402, threshold conditions 406 are applied to individuals within the identified clusters 404. In this example, individuals are removed from these clusters if they have fewer than three GD symptoms or no typical / distinctive symptoms of GD.

[0065] The refined dataset 410, which excludes the removed individuals from the initial cohort, is output 408 for further processing. In this example, the refined dataset has 565 individuals, i.e., 45 individuals have been removed from the three identified clusters 404. Thus, the refined dataset contains patients with either (1) two or more features representative of GD; or (2) at least three features of GD, one of which is representative of GD. This limits the number of patients removed while maintaining a cohort with enough relevant information to train the algorithm.

[0066] Returning to Figure 2, following the refinement operation 206, or if no refinement operation is performed, following the clustering operation 204, a merging operation 208 is performed to create training data 210. During the merging operation, the pruned / refined dataset is merged with medical data from a control dataset 214. The control dataset includes medical data from individuals who have not been diagnosed with a rare disease (also referred to herein as "control individuals").

[0067] Control individuals are selected from the control dataset based on having at least a threshold number of symptoms of a rare disease. The threshold number may be between 1 and 5, for example, the threshold number may be 2.

[0068] In some embodiments, each individual in the pruned dataset is matched to multiple control individuals at a predetermined ratio to generate a balanced dataset. The predetermined ratio may depend on the prevalence of the rare disease in the general population. The predetermined ratio may be 5 to 20 control individuals, for example, 10, per RD individual. Matching may be performed based on one or more demographic characteristics of the individuals in the pruned dataset. Examples of demographic characteristics include: age (e.g., in an age range); gender; and / or data coverage (e.g., the period for which data for that individual is available).

[0069] In other words, for each RD individual, a predetermined number of control individuals are sampled from a set of control individuals with the same demographic characteristics as the RD individual. Exact matching without replacement can be used.

[0070] Figure 5 shows an example of such a matching process 500 for Gaucher disease. A pruned dataset 502 of 565 Gaucher disease patients (GD individuals) is merged 506 with control individuals selected from a dataset 504 of 10M individuals to create a training dataset 508. In this example, the GD individuals are each classified 510 into demographics based on gender (i.e., male or female), applicable age range (e.g., 0-10 years, 11-20 years, etc.), and applicable time period (0-1 year, 1-2 years, etc.). For each GD individual, a corresponding individual in the control dataset of the same demographic is identified 512 and sampled at a predetermined ratio, e.g., 1:10 514. The result is a training dataset 508 (T) containing 565 GD individuals and 5650 control individuals.

[0071] Control individuals can be excluded from inclusion in the merged dataset based on one or more additional exclusion criteria. Additional exclusion criteria can include having one or more other diseases, for example, another disease of a similar type or affecting a similar lineage. Alternatively or additionally, additional exclusion criteria can include having a different application period than the individuals in the pruned dataset.

[0072] In some embodiments, a fictitious "control index date" is created for each control individual so that both rare disease patients and control patients have the same period of data from which the algorithm can learn. The control index date can be set to the median look-back period of the rare disease population over a given application length.

[0073] 2 , in some embodiments, the method 200 further includes one or more data augmentation operations 216. The data augmentation operations 216 identify additional medical data that may be relevant to the diagnosis of the rare disease and add this additional medical data to the training dataset 210.

[0074] FIG. 6 illustrates an example method 600 for augmenting a training dataset. Prior to augmentation, each individual in the training dataset 602 is associated with an indication of which symptoms of a rare disease are present in that individual 604. For example, as shown in FIG. 6, a vector of binary labels / flags (t n ) can be associated with each individual (n), with each component of the vector corresponding to whether the feature is present in the associated individual. In the Gaucher disease example, N=69 features are selected from the literature / by experts in Gaucher disease to be the components of the vector.

[0075] Data augmentation can identify one or more additional features 606 associated with a rare disease in medical data associated with an individual that are not already included in the training data. These additional features are added to the data associated with the individual, for example, as additional components to a vector of features associated with that individual or as additional feature vectors associated with each individual.

[0076] In some embodiments, data-driven augmentation 608 can be used, which applies statistical methods to medical data associated with individuals in the control and target datasets to identify additional features present in the target dataset at significantly lower frequencies than the control dataset (or vice versa). Indications of the presence or absence of these additional features can be added to the data associated with each individual. Data-driven features are a good way to discover new features that are more prevalent within a population of interest herein. They can provide new evidence related to medical knowledge.

[0077] For example, for each feature, a chi-squared test and a Cramer's V test are performed between the rare disease cohort and the control cohort. The chi-squared test indicates whether there is a significant relationship between variables. Cramer's V is a number between 0 and 1 that indicates how strongly two variables are associated. Features are ranked according to their Cramer's V score. The decision to select a feature is then based on the highest Cramer's V value if the chi-squared test is significant.

[0078] For example, in some embodiments, the selected features are those for which the strength of association using Cramer's V coefficient between both cohorts is above a threshold (e.g., above 0.1) and the p-value using a chi-squared test is below a threshold (e.g., below 0.05) to assess sampling bias, and which are not already present.

[0079] The selected features may be coded for each individual as a binary flag indicating the presence or absence of the feature in that individual. Alternatively or additionally, the selected features may be coded for each individual as the age at which they first occurred in that individual.

[0080] Alternatively or additionally, an age differentiator 610 can be identified. One or more correlations between age and symptoms are identified in the data. These can help identify non-aging symptoms that may occur due to the rare disease. A binary variable is created for each selected symptom depending on the age of onset, e.g., whether the symptom occurs before or after the average onset in the literature for individuals without the rare disease. The selected symptoms can be symptoms that commonly occur with aging, such as neurological symptoms (e.g., Parkinson's symptoms, tremors) and / or bone conditions (e.g., osteoarthritis, osteoporosis).

[0081] Alternatively or additionally, symptom associations 612 may be identified. A symptom association is a combination of symptoms and the localization of those symptoms.

[0082] Alternatively or additionally, natural language processing (NLP) can be used to identify additional features from the literature 614. Vocabulary / tokens representing clinical terms are identified in medical literature (e.g., in a selection of medical publications from PubMED). The medical literature can be selected to include publications related to rare diseases and a selection of diseases with similar symptoms. Using natural language processing, mathematical representations of clinical terms are created, based on the context in which the terms are used in the literature. For example, each term can be represented as an embedding vector in a vector space, such that words occurring in similar contexts occupy similar positions in the vector space; i.e., clinical terms occurring in similar contexts have similar embedding vector representations. One example of a natural language processing algorithm that can be used to generate such embedding vectors is Word2Vec (see, e.g., "Distributed Representations of Words and Phrases and Their Compositionality," Mikolov et al., 2013, Adv. Neural Inf. Process. Syst., Vol. 26).

[0083] Terms that appear in similar contexts to rare diseases and / or their symptoms can be identified as additional features to include in the training dataset by taking embedding vectors that are located near (e.g., within a threshold distance of) the embedding vector for the rare disease / rare disease symptom. Cosine similarity can be used to determine the distance between embedding vectors.

[0084] In general, embedding vectors can be determined / generated for all terms related to rare diseases in the literature.

[0085] Additionally, the identified word / term embeddings can be used to generate a vectorization for each individual in the training dataset. For each individual, a vector representation of that individual can be created, for example, by averaging the embedding vectors of the features associated with that individual (e.g., taking the average of those embedding vectors). In some embodiments, the embeddings of the rare disease terms can also be subtracted from this representation to create a final vector representation for the individual. If control individuals do not have any of the rare disease features, they can be represented with a zero vector. These vector representations can be used as additional or alternative inputs to machine learning models.

[0086] When trained on a corpus of biomedical literature, word embeddings can establish relationships between concepts such as diseases, symptoms, and treatments. Word embeddings convert concepts into vectors by using the context in which the concept is mentioned within the body of text. The similarity between two vectors represents how closely related they are. Features can be extracted by calculating the distance between a symptom and a given disease and retaining the closest features (see, for example, "Learning to Identify Rare Disease Patients from Electronic Health Records," Colbaugh et al., 2018, AMIA Annu. Symp. Proceedings. AMIA Symp., vol. 2018, pp. 340-347).

[0087] The machine learning model is trained using the created training dataset. A test dataset can be created in substantially the same manner as the training dataset and used to test the performance of the machine learning model after it has been trained. The test dataset can be selected based on the same criteria as the training dataset, but with a different ratio of rare disease individuals to control individuals. The number of control individuals per rare disease individual can be greater in the test dataset than in the training dataset. For example, the test dataset can have 1,000 to 50,000 control individuals per rare disease individual, for example, a ratio of rare disease to control individuals of 1:10,000.

[0088] 7 shows a flow diagram of an example method for generating a training dataset for use in training a machine learning model to identify individuals with a rare disease. The method can be performed by a computer.

[0089] In operation 7.1, an initial dataset is received that includes medical data associated with a plurality of individuals with a rare disease. The medical data for each individual includes data indicative of a plurality of symptoms of the rare disease experienced by the individual. The medical data may further include one or more of: vital sign data (e.g., BMI, blood pressure, pain score); laboratory values ​​(e.g., A1C, eGFR, FEV); biomarkers; medications received by the individual; and / or individual details (e.g., age, sex, smoking status, ethnicity).

[0090] In operation 7.2, an unsupervised clustering method is used to identify multiple clusters of individuals in the initial dataset. The unsupervised clustering algorithm may be a hierarchical agglomerative clustering algorithm. The clustering algorithm may cluster the dataset into a predetermined number of clusters. The predetermined number may be based on known subtypes of the rare disease or may be a fraction of the known number of symptoms of the rare disease.

[0091] In operation 7.3, one or more of the clusters are identified as least representative of rare diseases based on the medical data of the individuals in the cluster. Identifying one or more of the clusters as least representative of rare diseases may include identifying, for each of the multiple clusters, a representative symptom of the rare disease based on, for example, the most common symptom or set of symptoms experienced by the individuals in each cluster and / or demographic characteristics of the individuals in each cluster. The representative symptom of each cluster may be compared to a predetermined set of known symptoms of the rare disease, the predetermined set of known symptoms including a more representative set of symptoms and a least representative set of symptoms. A cluster may be identified as least representative of rare diseases if the representative symptom of that cluster is in the least representative set of symptoms.

[0092] In operation 7.4, one or more of the individuals are removed from one or more clusters identified as least representative based on the medical data of the one or more individuals to create a pruned dataset. For each individual in the cluster identified as least representative, this may include determining whether medical data associated with that individual meets a threshold condition, and removing the individual from the identified cluster if the threshold condition is not met. The threshold condition may be based on symptoms of the rare disease, for example, having at least a threshold number of symptoms of the rare disease. The threshold condition may alternatively or additionally include having at least one symptom of the rare disease in a unique subset of symptoms of the rare disease.

[0093] In operation 7.5, the pruned dataset is combined with a control dataset including a plurality of individuals who do not have the rare disease to generate a training dataset. The control dataset can include individuals who do not have the rare disease and who have at least a threshold number of symptoms of the rare disease. Combining the pruned dataset with the control dataset can include matching a plurality of individuals in the control dataset to each individual in the pruned dataset in a predetermined ratio, e.g., 10 control individuals for each RD individual. Matching can be based on one or more demographic characteristics of the individuals in the pruned dataset.

[0094] In some embodiments, operations 7.2-7.4 are not performed, i.e., the "pruned" dataset is simply the initial dataset, possibly extended with additional data about rare disease features. In some embodiments, the initial dataset may be clustered to identify features but not pruned based on those features, i.e., operation 7.2 is performed, but operations 7.3 and 7.4 are not performed.

[0095] In some embodiments, the method may further include one or more data augmentation operations following merging the control dataset and the pruned dataset. The data augmentation operations may identify one or more potential indicators of the rare disease and incorporate them as labels into the training dataset. The augmentation may include comparing medical data associated with individuals in the pruned dataset with medical data associated with individuals in the control dataset to identify one or more potential symptoms of the rare disease, and augmenting the training dataset with the one or more potential symptoms of the rare disease. Alternatively or additionally, the data augmentation may include using natural language processing to extract one or more potential symptoms of the rare disease from literature related to the rare disease, and augmenting the training dataset with the one or more potential symptoms of the rare disease.

[0096] In some embodiments, the method includes generating an embedding vector for each of a plurality of terms related to a rare disease from a corpus of medical literature related to the rare disease using natural language processing. In some embodiments, the embedding vector for each of the plurality of terms in the corpus of medical literature is generated using a natural language processing algorithm such as Word2Vec. Each of the embedding vectors can be normalized.

[0097] These embedding vectors can then be compared using a distance metric to embedding vectors of one or more predetermined terms known to be associated with rare diseases, such as the embedding vectors of the rare disease itself and / or its known symptoms. The embedding vectors associated with the predetermined terms are generated using the same method as the embedding vectors from the corpus of medical literature. The distance metric can be any metric form that measures the difference between two vectors, such as the dot product or cosine similarity between (normalized) embedding vectors. An embedding vector for a term in the corpus of medical literature is discarded if the distance between said embedding vector and the embedding vector of the predetermined term known to be associated with rare diseases is greater than a threshold, generating a set of embedding vectors for terms associated with rare diseases.

[0098] Once the training set is generated, the method can further include generating, for each individual in the training dataset, an embedding vector representing the individual based on features associated with the individual and the embedding vectors for a plurality of rare disease-related terms. Generating an embedding vector representing the individual can include identifying one or more embedding vectors for rare disease-related terms that correspond to features of the rare disease experienced by the individual, and averaging the identified embedding vectors for the rare disease-related terms to generate an embedding vector representing the individual. In other words, the embedding of an individual is the average over the embedding vectors of the features associated with the individual. In some embodiments, the embedding vector for the rare disease term itself is subtracted from the embedding vector for the individual, i.e., the embedding vector representing the individual represents the difference between the embedding vector of the rare disease term and the average embedding vector of the features associated with the individual.

[0099] The training dataset can then be used to train a machine learning model to identify individuals with a rare disease. The machine learning model is trained using supervised learning techniques. During model training, medical data associated with the individual is input into the model, and the model provides a proposed classification indicating whether the individual has a rare disease. In some embodiments, the medical data input into the machine learning model includes an embedding vector associated with the individual. The proposed classification is compared to a known classification for the individual. The parameters of the model are updated based on the comparison.

[0100] In some embodiments, a subset of medical data associated with an individual can be input into the model. For example, for an individual with a rare disease, only that individual's pre-diagnosis medical data, e.g., up to the time of diagnosis or up to some predetermined time before diagnosis, can be input into the model. This can reduce bias associated with collecting post-diagnosis symptoms.

[0101] The trained machine learning model can be used to identify individuals who may have a rare disease. Medical data related to the individual is input into the model. In some embodiments, the medical data input into the machine learning model includes an embedding vector related to the individual. The model processes the input medical data to generate data indicating whether the individual associated with the medical data has the disease, for example, a binary classification of whether the individual has the disease or not, or a probability that the individual has the disease. The data indicating whether the individual associated with the medical data has the disease or not is output from the model.

[0102] 8 shows a schematic example of a system / apparatus for performing any of the methods described herein. The illustrated system / apparatus is an example of a computing device. Those skilled in the art will appreciate that other types of computing devices / systems may alternatively be used to implement the methods described herein, such as a distributed computing system.

[0103] The device (or system) 800 includes one or more processors 802. The one or more processors control the operation of the other components of the system / device 800. The one or more processors 802 may include, for example, a general-purpose processor. The one or more processors 802 may be single-core or multi-core devices. The one or more processors 802 may include a central processing unit (CPU) or a graphics processing unit (GPU). Alternatively, the one or more processors 802 may include specialized processing hardware, such as, for example, a RISC processor, or programmable hardware with embedded firmware. Multiple processors may be included.

[0104] The system / device includes a working or volatile memory 804. One or more processors can access the volatile memory 804 to process data and can control the storage of data in the memory. The volatile memory 804 can include any type of RAM, such as, for example, static RAM (SRAM), dynamic RAM (DRAM), or can include flash memory, such as an SD card.

[0105] The system / apparatus includes a non-volatile memory 806. The non-volatile memory 806 stores a set of operating instructions 808 in the form of computer-readable instructions for controlling the operation of the processor 802. The non-volatile memory 806 can be any type of memory, such as a read-only memory (ROM), flash memory, or magnetic drive memory.

[0106] The one or more processors 802 are configured to execute operational instructions 808 to cause the system / device to perform any of the methods described herein. The operational instructions 808 may include code (i.e., drivers) associated with hardware components of the system / device 800 and code associated with basic operations of the system / device 800. Generally, the one or more processors 802 execute one or more of the operational instructions 808 that are permanently or semi-permanently stored in non-volatile memory 806, using volatile memory 804 to temporarily store data generated during execution of the operational instructions 808.

[0107] Implementations of the methods described herein can be realized in digital electronic circuitry, integrated circuits, specially designed ASICs (application-specific integrated circuits), computer hardware, firmware, software, and / or combinations thereof. These may include a computer program product (e.g., software stored on a magnetic or optical disk, memory, programmable logic device, etc.) containing computer-readable instructions that, when executed by a computer such as that described in connection with FIG. 8, cause the computer to perform one or more of the methods described herein.

[0108] Any system features as described herein may also be provided as method features, and vice versa. As used herein, means plus function features may alternatively be expressed in terms of their corresponding structures. In particular, method aspects may be applied to system aspects, and vice versa.

[0109] Furthermore, any, some, and / or all features in one embodiment may be applied to any, some, and / or all features in any other embodiment, in any appropriate combination. It should also be understood that specific combinations of the various features described and defined in any embodiment of the present invention may be implemented and / or provided and / or used independently.

[0110] While several embodiments have been shown and described, it will be understood by those skilled in the art that changes can be made in these embodiments without departing from the principles of the present disclosure, the scope of which is defined in the claims and their equivalents.

[0111] The terms "drug" or "medicament" are used interchangeably herein to describe a pharmaceutical formulation containing one or more active pharmaceutical ingredients or pharmaceutically acceptable salts or solvates thereof, and optionally a pharmaceutically acceptable carrier. An active pharmaceutical ingredient ("API"), in its broadest sense, is a chemical structure that has a biological effect on humans or animals. In pharmacology, drugs or medications are used to treat, cure, prevent, or diagnose disease or otherwise improve physical or mental well-being. Drugs or medications can be used for a limited duration or periodically for chronic disorders.

[0112] As described below, drugs or pharmaceutical agents may contain at least one API or a combination thereof in various types of formulations for the treatment of one or more diseases. Examples of APIs include small molecules with a molecular weight of 500 Da or less, polypeptides, peptides, and proteins (e.g., hormones, growth factors, antibodies, antibody fragments, and enzymes), carbohydrates and polysaccharides, as well as nucleic acids, double-stranded or single-stranded DNA (including naked and cDNA), RNA, antisense nucleic acids, such as antisense DNA and RNA, small interfering RNA (siRNA), ribozymes, genes, and oligonucleotides. Nucleic acids can be incorporated into molecular delivery systems such as vectors, plasmids, or liposomes. Mixtures of one or more drugs are also contemplated.

[0113] The drug or agent can be contained in a primary package or "drug container" adapted for use in a drug delivery device. The drug container can be, for example, a cartridge, syringe, reservoir, or other rigid or flexible vessel configured to provide a chamber suitable for storage (e.g., short-term or long-term storage) of one or more drugs. For example, in some cases, the chamber can be designed to store the drug for at least one day (e.g., from one day to at least 30 days). In some cases, the chamber can be designed to store the drug for about one month to about two years. Storage can occur at room temperature (e.g., about 20°C) or refrigerated temperatures (e.g., from about -4°C to about 4°C). In some cases, the drug container can be or include a dual-chamber cartridge configured to separately store two or more components of a pharmaceutical formulation to be administered (e.g., an API and a diluent, or two different drugs), one in each chamber. In such cases, the two chambers of the dual-chamber cartridge can be configured to allow mixing between two or more components prior to and / or during administration to the human or animal body. For example, the two chambers can be configured to be in fluid communication with each other (e.g., via a conduit between the two chambers) and to allow mixing of the two components by a user, if desired, prior to administration. Alternatively or additionally, the two chambers can be configured to allow mixing upon administration of the components to the human or animal body.

[0114] The drugs or agents contained in the drug delivery devices described herein can be used to treat and / or prevent many different types of medical disorders. Examples of disorders include, for example, diabetes or complications associated with diabetes, such as diabetic retinopathy, and thromboembolic disorders, such as deep vein thromboembolism or pulmonary embolism. Further examples of disorders include acute coronary syndrome (ACS), angina, myocardial infarction, cancer, macular degeneration, inflammation, hay fever, atherosclerosis, and / or rheumatoid arthritis. Examples of APIs and drugs are those listed in handbooks such as Rote Liste 2014 (e.g., but not limited to, Main Group 12 (antidiabetic agents) or 86 (oncology agents)) and the Merck Index, 15th edition.

[0115] Examples of APIs for the treatment and / or prevention of type 1 or type 2 diabetes or complications associated with type 1 or type 2 diabetes include insulin, e.g., human insulin, or a human insulin analog or derivative; glucagon-like peptide (GLP-1), a GLP-1 analog or GLP-1 receptor agonist, or an analog or derivative thereof; a dipeptidyl peptidase-4 (DPP4) inhibitor; or a pharmaceutically acceptable salt or solvate thereof, or any mixture thereof. As used herein, the terms "analog" and "derivative" refer to a polypeptide having a molecular structure that is formally derivable from the structure of a naturally occurring peptide, e.g., the structure of human insulin, by deletion and / or replacement of at least one amino acid residue present in the naturally occurring peptide and / or by addition of at least one amino acid residue. The added and / or replaced amino acid residue can be either a codable amino acid residue, another naturally occurring residue, or a purely synthetic amino acid residue. Insulin analogs are also referred to as "insulin receptor ligands." In particular, the term "derivative" refers to a polypeptide having a molecular structure formally derivable from the structure of a naturally occurring peptide, for example, the molecular structure of human insulin in which one or more organic substituents (e.g., fatty acids) are attached to one or more of the amino acids. Optionally, one or more amino acids present in the naturally occurring peptide are deleted and / or replaced by other amino acids, including non-codable amino acids, or amino acids, including non-codable ones, are added to the naturally occurring peptide.

[0116] Examples of insulin analogues are Gly(A21), Arg(B31), Arg(B32) human insulin (insulin glargine); Lys(B3), Glu(B29) human insulin (insulin glulisine); Lys(B28), Pro(B29) human insulin (insulin lispro); Asp(B28) human insulin (insulin aspart); human insulin in which the proline at position B28 is replaced by Asp, Lys, Leu, Val or Ala and the Lys at position B29 may be replaced by Pro; Ala(B26) human insulin; Des(B28-B30) human insulin; Des(B27) human insulin and Des(B30) human insulin.

[0117] Examples of insulin derivatives are, for example, B29-N-myristoyl-des(B30) human insulin, Lys(B29)(N-tetradecanoyl)-des(B30) human insulin (insulin detemir, Levemir®); B29-N-palmitoyl-des(B30) human insulin; B29-N-myristoyl human insulin; B29-N-palmitoyl human insulin; B28-N-myristoylLysB28ProB29 human insulin; B28-N-palmitoyl-LysB28ProB29 human insulin; B30-N-myristoyl-ThrB29LysB30 human insulin. B30-N-palmitoyl-ThrB29LysB30 human insulin; B29-N-(N-palmitoyl-gamma-glutamyl)-des(B30) human insulin, B29-N-omega-carboxypentadecanoyl-gamma-L-glutamyl-des(B30) human insulin (insulin degludec, Tresiba®); B29-N-(N-lithocholyl-gamma-glutamyl)-des(B30) human insulin; B29-N-(ω-carboxyheptadecanoyl)-des(B30) human insulin and B29-N-(ω-carboxyheptadecanoyl) human insulin.

[0118] Examples of GLP-1, GLP-1 analogs and GLP-1 receptor agonists are, for example, lixisenatide (Lyxumia®), exenatide (exendin-4, Byetta®, Bydureon®, a 39 amino acid peptide produced by the salivary glands of the flathead monster), liraglutide (Victoza®), semaglutide, taspoglutide, albiglutide (Syncria®), dulaglutide (Trulicity®), rexendin-4, CJC-1134-PC, PB-1023, TTP-054, langrenatide / HM-11260C (efpegrenatide). , HM-15211, CM-3, GLP-1 Erigen, ORMD-0901, NN-9423, NN-9709, NN-9924, NN-9926, NN-9927, Nodexene, Viador-GLP-1, CVX-096, ZYOG-1, ZYD-1, GSK-2374697, DA-3091, MAR-701, MAR709, ZP-2929, ZP-3022, ZP -DI-70, TT-401 (Pegapamodtide), BHM-034, MOD-6030, CAM-2036, DA-15864, ARI-2651, ARI-2255, Tirzepatide (LY3298176), Bamadutide (SAR425899), Exenatide-XTEN, and Glucagon-Xten.

[0119] Examples of oligonucleotides are, for example, the cholesterol-lowering antisense therapeutic mipomersen sodium (Kynamro®) for the treatment of familial hypercholesterolemia, or RG012 for the treatment of Alport syndrome.

[0120] Examples of DPP4 inhibitors are linagliptin, vidagliptin, sitagliptin, denagliptin, saxagliptin, berberine.

[0121] Examples of hormones include pituitary or hypothalamic hormones or regulatory active peptides and their antagonists, such as gonadotropins (follitropin, lutropin, chorion gonadotropin, menotropin), somatropine (somatropin), desmopressin, terlipressin, gonadorelin, triptorelin, leuprorelin, buserelin, nafarelin, and goserelin.

[0122] Examples of polysaccharides include glycosaminoglycans, hyaluronic acid, heparin, low-molecular-weight heparin or ultra-low-molecular-weight heparin or derivatives thereof, or sulfated polysaccharides, such as the polysulfated forms of the aforementioned polysaccharides, and / or pharmaceutically acceptable salts thereof. An example of a pharmaceutically acceptable salt of polysulfated low-molecular-weight heparin is enoxaparin sodium. Examples of hyaluronic acid derivatives include Hylan G-F20 (Synvisc®) and sodium hyaluronate.

[0123] As used herein, the term "antibody" refers to an immunoglobulin molecule or an antigen-binding portion thereof. Examples of antigen-binding portions of immunoglobulin molecules include F(ab) and F(ab')2 fragments that retain antigen-binding ability. An antibody can be a polyclonal antibody, a monoclonal antibody, a recombinant antibody, a chimeric antibody, a deimmunized or humanized antibody, a fully human antibody, a non-human (e.g., murine) antibody, or a single-chain antibody. In some embodiments, an antibody has effector function and is capable of fixing complement. In some embodiments, an antibody has reduced or no binding ability to Fc receptors. For example, an antibody can be an isotype or subtype, antibody fragment, or mutant that does not support Fc receptor binding, e.g., has a mutation or deletion of the Fc receptor binding region. The term antibody also includes antigen-binding molecules based on tetravalent bispecific tandem immunoglobulins (TBTIs) and / or dual variable region antibody-like binding proteins (CODVs) with a crossover binding region orientation.

[0124] The term "fragment" or "antibody fragment" refers to a polypeptide (e.g., an antibody heavy and / or light chain polypeptide) derived from an antibody polypeptide molecule that does not include the full-length antibody polypeptide but comprises at least a portion of the full-length antibody polypeptide that is still capable of binding to antigen. Antibody fragments can include truncated portions of a full-length antibody polypeptide, but the term is not limited to such truncated fragments. Antibody fragments useful in the present invention include, for example, Fab fragments, F(ab')2 fragments, scFv (single-chain Fv) fragments, linear antibodies, monospecific or multispecific antibody fragments, e.g., bispecific, trispecific, tetraspecific, and multispecific antibodies (e.g., diabodies, triabodies, tetrabodies), monovalent or multivalent antibody fragments, e.g., bivalent, trivalent, tetravalent, and multivalent antibodies, minibodies, chelating recombinant antibodies, tribodies or bibodies, intrabodies, nanobodies, small modular immunopharmaceuticals (SMIPs), binding domain immunoglobulin fusion proteins, camelized antibodies, and VHH-containing antibodies. Additional examples of antigen-binding antibody fragments are known in the art.

[0125] The term "complementarity determining region" or "CDR" refers to short polypeptide sequences within the variable regions of both heavy and light chain polypeptides that are primarily responsible for mediating specific antigen recognition. The term "framework region" refers to amino acid sequences within the variable regions of both heavy and light chain polypeptides that are not CDR sequences and that are primarily responsible for maintaining the proper orientation of the CDR sequences to enable antigen binding. Although the framework regions themselves typically do not directly participate in antigen binding, as is known in the art, certain residues within the framework regions of a particular antibody may be directly involved in antigen binding or may affect the ability of one or more amino acids within the CDRs to interact with the antigen.

[0126] Examples of antibodies are anti-PCSK-9 mAb (e.g., alirocumab), anti-IL-6 mAb (e.g., sarilumab), and anti-IL-4 mAb (e.g., dupilumab).

[0127] Pharmaceutically acceptable salts of any of the APIs described herein are contemplated for use as drugs or medicaments in drug delivery devices. Pharmaceutically acceptable salts include, for example, acid addition salts and base salts.

[0128] Those skilled in the art will understand that modifications (addition and / or removal) of various components of the APIs, formulations, devices, methods, systems, and embodiments described herein may be made without departing from the full scope and spirit of the invention, which encompasses such modifications and any and all equivalents thereof.

[0129] Example drug delivery devices can include needle-based injection systems as described in Table 1 of Chapter 5.2 of ISO 11608-1:2014(E). As described in ISO 11608-1:2014(E), needle-based injection systems can be broadly distinguished into multi-dose container systems and single-dose container systems (with partial or full ejection). The container can be an interchangeable container or an integrated, non-interchangeable container.

[0130] As further described in ISO 11608-1:2014(E), a multi-dose container system can include a needle-based injection device with replaceable containers. In such a system, each container holds multiple doses, and the dose size can be fixed or variable (pre-set by the user). Another multi-dose container system can include a needle-based injection device with an integrated, non-replaceable container. In such a system, each container holds multiple doses, and the dose size can be fixed or variable (pre-set by the user).

[0131] As further described in ISO11608-1:2014(E), a single-dose container system can include a needle-based injection device with replaceable containers. In one example of such a system, each container holds a single dose, thereby discharging the entire deliverable volume (full discharge). In a further example, each container holds a single dose, thereby discharging a portion of the deliverable volume (partial discharge). Also as described in ISO11608-1:2014(E), a single-dose container system can include a needle-based injection device with an integrated, non-replaceable container. In one example of such a system, each container holds a single dose, thereby discharging the entire deliverable volume (full discharge). In a further example, each container holds a single dose, thereby discharging a portion of the deliverable volume (partial discharge).

Claims

1. 1. A computer-implemented method for generating a training dataset for training a machine learning model to identify individuals with a rare disease, comprising: receiving, by one or more processors, an initial dataset including medical data relating to a plurality of individuals having a rare disease, the medical data for each individual including data indicative of a plurality of symptoms of the rare disease experienced by the individual; identifying, by one or more processors, a plurality of clusters of individuals in the initial data set using an unsupervised clustering method; identifying, by one or more processors, one or more of the clusters as least representative of rare diseases based on medical data of individuals within the clusters; removing, by one or more processors, one or more of the individuals from the one or more clusters identified as least representative based on the medical data of the one or more individuals to generate a pruned dataset, wherein for each individual in the cluster identified as least representative: determining, by the one or more processors, whether medical data associated with the individual satisfies a threshold condition, the threshold condition being based on symptoms of the rare disease; removing, by one or more processors, individuals from the identified clusters if the threshold condition is not met. and combining, by one or more processors, the pruned dataset with a control dataset including a plurality of individuals who do not have the rare disease to generate a training dataset; The method comprising:

2. Identifying one or more of the clusters as least representative of rare diseases: identifying, by one or more processors, a representative symptom of the rare disease for each of the plurality of clusters; One or more processors generate representative symptoms for each cluster, comparing with a predetermined set of known symptoms of the disease, the predetermined set of known symptoms including a more representative set of symptoms and a least representative set of symptoms; identifying, by the one or more processors, a cluster as least representative of rare diseases if a representative symptom of the cluster is included in the set of least representative symptoms; The method of claim 1 , comprising:

3. 3. The method of claim 1, wherein identifying a plurality of clusters of individuals in the initial dataset comprises clustering the dataset into a predetermined number of clusters using hierarchical agglomerative clustering.

4. comparing, by one or more processors, medical data associated with individuals in the pruned dataset with medical data associated with individuals in the control dataset to identify one or more potential symptoms of the rare disease; augmenting, by one or more processors, the training dataset with one or more potential symptoms of the rare disease; 3. The method of claim 1 or 2, further comprising:

5. Extracting, by one or more processors, one or more potential symptoms of a rare disease from literature related to the rare disease using natural language processing; Expanding the training dataset with one or more potential symptoms of a rare disease; and 3. The method of claim 1 or 2, further comprising:

6. 3. The method of claim 1 or 2, wherein the control dataset comprises individuals without the rare disease who have at least a threshold number of symptoms of the rare disease.

7. 3. The method of claim 1 or 2, wherein combining the pruned dataset with the control dataset comprises matching a predetermined proportion of individuals in the control dataset to each individual in the pruned dataset, wherein the matching is based on one or more demographic characteristics of the individuals.

8. 10. A computer-implemented method for training, by one or more processors, a machine learning model to classify an individual as having a rare disease based on the individual's medical records, the method comprising: training the machine learning model using a labeled training dataset using supervised learning techniques, the training dataset being generated using the method of any one of claims 1 to 7.

9. 10. The method of claim 8, wherein the machine learning model is trained on a subset of data in the labeled dataset, the subset including, for each individual in the dataset with a rare disease, medical data collected before the individual was diagnosed with the rare disease.

10. 1. A computer-implemented method for diagnosing a disease, comprising: inputting, by one or more processors, medical data associated with the individual into a machine learning model; processing, by one or more processors, the input medical data using the machine learning model to generate data indicative of whether an individual associated with the medical data has a disease; outputting, by the one or more processors, data from the machine learning model indicating whether the individual associated with the medical data has a disease; Including, 10. The method of claim 8, wherein the machine learning model is trained using the method of claim 8 or 9.

11. 11. A system comprising one or more processors and a memory, the memory comprising computer readable code that, when executed by the one or more processors, causes the system to perform the method of any one of claims 1 to 10.

12. A computer program product comprising computer readable code which, when executed by a computing device, causes the computing device to perform the method of any one of claims 1 to 10.

Images