EP4666295A1

EP4666295A1 - Automated screening for acid sphingomyelinase deficiency

Info

Publication number: EP4666295A1
Application number: EP24713084.2A
Authority: EP
Inventors: Alexandra Dumitriu; Maja GASPARIC; Pauline Marie GUILMIN; Martin Kristian Louis MONTMERLE; Stefaan SANSEN; Neha B. SHAH
Original assignee: Genzyme Corp
Current assignee: Genzyme Corp
Priority date: 2023-02-14
Filing date: 2024-02-09
Publication date: 2025-12-24
Also published as: JP2026506658A; CN120641996A; WO2024173175A1

Abstract

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for screening a database of electronic medical records for a population of subjects to identify subjects that are predicted to have acid sphingomyelinase deficiency. In one aspect, a method comprises, for each subject in the population of subjects: obtaining, from the database of electronic medical records, a set of features characterizing the subject; processing a model input that comprises the set of features characterizing the subject using a diagnostic machine learning model, having a set of diagnostic machine learning model parameters, to generate a model output that characterizes a likelihood that the subject has acid sphingomyelinase deficiency; and classifying, based on the model output of the diagnostic machine learning model, whether the subject has acid sphingomyelinase deficiency.

Description

AUTOMATED SCREENING FOR ACID SPHINGOMYELINASE DEFICIENCY

CROSS-REFERENCE TO RELATED APPLICATIONS

[0001] This claims the benefit of priority to Provisional Application No. 63/445,649, titled “AUTOMATED SCREENING FOR ACID SPHINGOMYELINASE DEFICIENCY,” filed on February 14 2023, Provisional Application No. 63/519,957, titled “AUTOMATED SCREENING FOR ACID SPHINGOMYELINASE DEFICIENCY,” filed on August 16, 2023 and to Provisional Application No. 63/610,665, titled “AUTOMATED SCREENING FOR ACID SPHINGOMYELINASE DEFICIENCY,” filed on December 15, 2023, the contents of which are hereby incorporated by reference.

TECHNICAL FIELD

[0002] This specification relates to automated screening of a population of subjects to identify subjects having acid sphingomyelinase deficiency (ASMD).

BACKGROUND

[0003] ASMD is a rare genetic disorder that is caused by a deficiency of the enzyme acid sphingomyelinase (ASM). The disorder is a part of a group of diseases known as lysosomal storage disorders. ASMD prevents cells from breaking down sphingomyelin, a type of fat that is found in the cell membrane. As a result of this, sphingomyelin accumulates in cells and forms sphingomyelin-containing lysosomes. These lysosomes will eventually rupture and release their contents into the cell, causing damage to the cell. ASMD is a multi -systemic disease affecting various organs and is potentially fatal. Treatment for ASMD can involve enzyme replacement therapy, which involves the administration of the human recombinant to replace deficient ASM enzyme. ASMD can occur in multiple forms, e.g., ASMD type A, ASMD type A/B, and ASMD type B. Each type of ASMD can be associated with different rates of progression, ages of onset, and severity. SUMMARY

[0004] This specification describes a system implemented as computer programs on one or more computers in one or more locations that can screen a database of electronic medical records for a population of subjects to identity subjects that are predicted to have ASMD.

[0005] According to one aspect, there is provided a method performed by one or more computers, the method comprising: screening a database of electronic medical records for a population of subjects to identify subjects that are predicted to have acid sphingomyelinase deficiency, comprising, for each subject in the population of subjects: obtaining, from the database of electronic medical records, a set of features characterizing the subject; processing a model input that comprises the set of features characterizing the subject using a diagnostic machine learning model, having a set of diagnostic machine learning model parameters, to generate a model output that characterizes a likelihood that the subject has acid sphingomyelinase deficiency; and classifying, based on the model output of the diagnostic machine learning model, whether the subject has acid sphingomyelinase deficiency; and providing an output that identifies subjects, from the population of subjects, that are classified as having acid sphingomyelinase deficiency.

[0006] In some implementations, the diagnostic machine learning model has been trained, by a machine learning training technique, to determine trained values of the set of diagnostic machine learning model parameters.

[0007] In some implementations, training the diagnostic machine learning model by the machine learning training technique comprises: obtaining a set of training examples, wherein each training example comprises: (i) a training input comprising a set of features of a training subject, and (ii) a target output based on whether the training subject is designated as having acid sphingomyelinase deficiency; and training the diagnostic machine learning model on the set of training examples.

[0008] In some implementations, obtaining the set of training examples comprises: obtaining a set of positive training examples, wherein each positive training example corresponds to a subject having acid sphingomyelinase deficiency; and selecting a plurality of negative training examples that each correspond to a respective subject that does not have acid sphingomyelinase deficiency, comprising, for each positive training example: identifying a corresponding set of negative training examples based on matching one or more confounding features of the set of negative training examples with the positive training example; and selecting the set of negative training examples for training the diagnostic machine learning model.

[0009] In some implementations, the one or more confounding features comprise demographic features.

[0010] In some implementations, training the diagnostic machine learning model on the set of training examples comprises training the diagnostic machine learning model to, for each training example, process the training input of the training example to generate a model output that matches the target output of the training example.

[0011] In some implementations, obtaining the set of training examples comprises, for one or more training examples, generating the training example from one or more electronic medical records of the training subject corresponding to the training example.

[0012] In some implementations, obtaining the set of training examples comprises: generating a set of features of a subject using a generative neural network; and generating a training example that includes a training input based on the set of features generated by the generative neural network.

[0013] In some implementations, the generative neural network has been trained by operations comprising: generating an output set of features using the generative neural network; processing the output set of features, generated by the generative neural network, using a discriminator neural network to generate a discriminative score that defines a likelihood that the output set of features was generated by the generative neural network instead of being derived from one or more electronic medical records; and training the generative neural network to optimize an objective function that depends on the discriminative score generated by the discriminator neural network.

[0014] In some implementations, generating a set of features of a subject using the generative neural network comprises: instantiating initial values of the set of features of the subject as random values sampled from a probability distribution; and de-noising the set of features over a sequence of de-noising iterations, comprising, at each de-noising iteration prior to a last de-noising iteration in the sequence of de-noising iterations: processing current values of the set of features using a generative neural network to generate a de-noised version of the set of features; and providing the de-noised version of the set of features for processing at the next de-noising iteration; wherein a de-noised version of the set of features generated at the last de-noising iteration defines the set of features of the subject. [0015] In some implementations, the method further comprises: generating, for each subject in the population of subjects, a confidence score that characterizes a confidence of the diagnostic machine learning model in the classification of whether the subject has acid sphingomyelinase deficiency; selecting a proper subset of the population of subjects for diagnostic testing for acid sphingomyelinase deficiency based on the confidence scores; re-training the diagnostic machine learning model based on results of the diagnostic testing of the proper subset of the population of subjects; and re-classifying whether each subject in the population of subjects has acid sphingomyelinase deficiency using the re-trained diagnostic machine learning model.

[0016] In some implementations, selecting the proper subset of the population of subjects for diagnostic testing for acid sphingomyelinase deficiency based on the confidence scores comprises selecting a number of subjects associated with lowest confidence scores, from among the population of subjects, for diagnostic testing for acid sphingomyelinase deficiency.

[0017] In some implementations, selecting the proper subset of the population of subjects for diagnostic testing for acid sphingomyelinase deficiency based on the confidence scores comprises: selecting a number of subjects associated with the lowest confidence scores, from among subjects classified as having acid sphingomyelinase deficiency, for diagnostic testing for acid sphingomyelinase deficiency; and selecting a number of subjects associated with the lowest confidence scores, from among subjects classified as not having acid sphingomyelinase deficiency, for diagnostic testing for acid sphingomyelinase deficiency.

[0018] In some implementations, generating, for each subject in the population of subjects, a confidence score that characterizes a confidence of the diagnostic machine learning model in the classification of whether the subject has acid sphingomyelinase deficiency comprises, for each subject in the population of subjects: generating a respective model output that characterizes a likelihood that the subject has acid sphingomyelinase deficiency using each diagnostic machine learning model in an ensemble of multiple diagnostic machine learning models; and generating the confidence score based on a measure of variance of the model outputs generated the ensemble of diagnostic machine learning models for the subject.

[0019] In some implementations, the population of subjects comprises at least 10,000 subjects.

[0020] In some implementations, screening the data of electronic medical records for the population of subjects is performed in 1 minute or less. [0021] In some implementations, the method further comprises, for one or more of the subjects that are classified as having acid sphingomyelinase deficiency, administering a drug to the subject to treat the acid sphingomyelinase deficiency in the subject.

[0022] In some implementations, the drug comprises olipudase alfa or olipudase alfa-rpcp.

[0023] In some implementations, the set of features characterizing the subject comprise features derived from a sample of bodily fluid or bodily tissue of the subject.

[0024] In some implementations, the diagnostic machine learning model comprises one or more of: a decision tree model, a random forest model, a neural network model, or a support vector machine model.

[0025] In some implementations, the population of subjects is prescreened to include only subjects that have been diagnosed with one or more of: interstitial lung disease (ILD), splenomegaly, or nonalcoholic steatohepatitis (NASH).

[0026] According to another aspect, there is provided a system comprising: one or more computers; and one or more storage devices communicatively coupled to the one or more computers, wherein the one or more storage devices store instructions that, when executed by the one or more computers, cause the one or more computers to perform operations of the methods described herein.

[0027] According to another aspect, there are provided one or more non-transitory computer storage media storing instructions that when executed by one or more computers cause the one or more computers to perform operations of the methods described herein.

[0028] Throughout this specification, the term “subject” refers to a human subject, i.e., a person.

[0029] Particular embodiments of the subject matter described in this specification can be implemented so as to realize one or more of the following advantages.

[0030] ASMD often goes undiagnosed or misdiagnosed in subjects, e.g., because of the rarity of the disease (and therefore low disease awareness and knowledge) and because the disease can present with a complex array of symptoms. Many subjects with ASMD are thus unaware that they have the disease. To address this issue, the system described in this specification can automatically screen large populations of patients for ASMD, based on electronic medical record data, in order to identify patients that are predicted to have ASMD. ASMD patients identified by the system can have their diagnosis confirmed by diagnostic testing and can be treated for ASMD, e.g., using an appropriate drug. Performing diagnostic testing of every patient in a large population of patients would be infeasible. The system can automatically identify a fraction of a population of patients that are likely to have ASMD (which can be, e.g., less than 1% of the population of patients), and thus enables these patients to be precisely targeted for confirmatory diagnostic testing.

[0031] To classify whether a subject has ASMD, the system can extract a set of features characterizing the subject from the subject’s electronic medical records, and process the set of features using a diagnostic machine learning model to generate a classification output. The classification output can define, e.g., a probability that the subject has ASMD, or a hard (e.g., binary) classification predicting whether the subj ect has ASMD (as will be described in more detail below). The screening system can train the diagnostic machine learning model on a set of training examples using a machine learning training technique. The diagnostic machine learning model provides a data-driven approach for classifying whether subjects potentially have ASMD based on complex patterns and correlations in electronic medical record data well beyond what could be analyzed by a human or solely in the human mind.

[0032] The system can train the diagnostic machine learning model over a sequence of iterations to increase its prediction accuracy. At each iteration, the system can generate a respective ASMD classification and a respective confidence score for each subject in the population of subjects. The confidence score for a subject characterizes a confidence of the diagnostic machine learning model in generating the ASMD classification of the subject. The system can then flag a proper subset of the population of patients for diagnostic testing based on the confidence scores, in particular, by selecting subjects having ASMD classifications associated with the lowest confidence scores for diagnostic testing. After application of the diagnostic machine learning model and subsequent patient diagnostic testing, the system can later be refined using the results of the diagnostic testing. The system can significantly increase the prediction accuracy of the diagnostic machine learning model by training on subjects with ASMD classifications associated with low confidence scores.

[0033] ASMD is a rare disease that is predicted to affect only 1 in 250,000 people. As a result, only a very limited amount of training data is available for training the diagnostic machine learning model to perform ASMD classification. To address this issue, the system can use generative modeling techniques to generate “synthetic” training examples for training the diagnostic machine learning model (e.g., distinguished from the “real-world” training examples derived from electronic medical record data of real-world patients). Augmenting the training data used for training the diagnostic machine learning model with synthetic training examples can increase the robustness and prediction accuracy of the diagnostic machine learning model.

[0034] The details of one or more embodiments of the subject matter of this specification are set forth in the accompanying drawings and the description below. Other features, aspects, and advantages of the subject matter will become apparent from the description, the drawings, and the claims.

BRIEF DESCRIPTION OF THE DRAWINGS

[0035] FIG. 1 shows an example screening system.

[0036] FIG. 2 shows an example training system.

[0037] FIG. 3 shows an example generative system.

[0038] FIG. 4A illustrates an example of a diagnostic machine learning model.

[0039] FIG. 4B shows a table of experimental results from applying the diagnostic machine learning model of FIG. 4A on a population of subjects.

[0040] FIG. 4C provides bar charts illustrating the respective prevalence of certain features among: (i) ASMD subjects, and (ii) control (non- ASMD) subjects.

[0041] FIG. 4D illustrates an example of a diagnostic machine learning model that has been specialized through training for screening a population of patients with unexplained interstitial lung disease (ILD) for ASMD.

[0042] FIG. 4E shows a table of experimental results from applying the diagnostic machine learning model of FIG. 4D on a population of subjects having unexplained ILD.

[0043] FIG. 5 is a flow diagram of an example process for screening a database of electronic medical records for a population of subjects to identify subjects that are predicted to have ASMD. [0044] FIG. 6 shows an example process for training a diagnostic machine learning model to identify subjects with ASMD, and then using the diagnostic machine learning model to screen a population of subjects.

[0045] Like reference numbers and designations in the various drawings indicate like elements. DETAILED DESCRIPTION

[0046] FIG. 1 shows an example screening system 100. The screening system 100 is an example of a system implemented as computer programs on one or more computers in one or more locations in which the systems, components, and techniques described below are implemented.

[0047] The screening system 100 is configured to screen a database of electronic medical records 102 (EMR database 102) for a population of subjects to identify subjects 112 that are predicted to have ASMD. The population of subjects can include any appropriate number of subjects, e.g., 10,000 subjects, 100,000 subjects, or 1 million subjects.

[0048] The EMR database 102 can include electronic medical records from one or more sources, e.g., databases for each of one or more departments within a medical center, or from one or more medical centers, etc. The EMR database 102 can include one or more electronic medical records for each subject in the population of subjects.

[0049] An electronic medical record for a subject can include any appropriate data characterizing the subject. For instance, an electronic medical record for a subject can characterize clinical laboratory data obtained by analyzing a sample of blood, urine, or other bodily fluid of bodily tissue of the subject, e.g., data characterizing bilirubin levels, cholesterol levels, hemoglobin levels, platelet levels, lymphocyte levels, red blood cell levels, triglyceride levels, alanine aminotransferase levels, aspartate aminotransferase levels, etc. As another example, an electronic medical record for a subject can characterize one or more medical symptoms or medical diagnoses of the subject, e.g., diabetes, interstitial lung disease, obesity, chronic pain, chest pain, joint pain, limb pain, back pain, abnormal gas levels, weight change, confusion, hyperlipidemia, arrhythmia, sleep apnea, hepatic failure, osmosis hyponatremia, kidney failure, heart disease, dyspnea, cardiomegaly, excessive menstruation, keratosis, macular degeneration, hyperopia, cataplexy, abnormal facial features, bronchiolitis, splenomegaly, lung infection, anemia, abdominal pain, abnormal bleeding, etc. As another example, an electronic medical record for a subject can characterize one or more demographic features of the subject, e.g., the age of the subject, the gender of the subject, etc.

[0050] In some cases, the EMR database 102 can be prescreened to include medical records only for subjects having one or more characteristics (e.g., symptoms or diagnoses) associated with ASMD. For instance, the EMR database 102 can be prescreened to include medical records only for subjects diagnosed with interstitial lung disease (ILD), or with splenomegaly, or with nonalcoholic steatohepatitis (NASH), or with any combination thereof. Prescreening the EMR database 102 to include only subjects having characteristics associated with ASMD can reduce consumption of computational resources (e.g., memory and computing power) by the screening system 100. Further, prescreening the EMR database 102 can allow the screening system 100 to perform the screening using machine learning models that are specialized through training for identifying ASMD in subjects having the particular characteristics selected through the prescreening.

[0051] The screening system 100 includes a feature generation engine 104 and a diagnostic machine learning model 108, which are each described in more detail next.

[0052] The feature generation engine 104 is configured to generate a respective set of subject features 106, for each subject, based on one or more electronic medical records of the subject from the EMR database 102. The set of subject features 106 for a subject can characterize, e.g., clinical laboratory data for the subject, medical symptoms of the subject, medical diagnoses of the subject, demographic features of the subject, etc. The set of subject features for a subject can be represented in a standardized format, e.g., as a tensor (e.g., vector or matrix) of cells, where each cell includes data (e.g., numerical data, textual data, or a combination of both) associated with a predefined semantic category. For instance, one cell in a set of subject features for a subject can include a categorical (e.g., Boolean) variable defining whether the subject has been diagnosed with a particular medical condition (e.g., diabetes), while another cell can include a continuous-valued feature defining a level of an enzyme in the blood of the subject.

[0053] To generate a set of subject features 106 for a subject, the feature generation engine 104 can scan the EMR database 102 to identify one or more electronic medical records of the subject, e.g., by identifying electronic medical records that include a patient identifier associated with the subject. The feature generation engine 104 can then extract data from the electronic medical records of the subject to populate the set of subject features 106 for the subject.

[0054] In some cases, the feature generation engine 104 can format or normalize data extracted from electronic medical records of a subject as part of generating the set of subject features 106 for the subject. For instance, the feature generation engine 104 can normalize a continuous-valued feature characterizing the subject (e g., the level of an enzyme in the blood of the subject) to a predefined interval, e g., the interval [0,1] . As another example, the feature generation engine 104 can format textual data from an electronic medical record that indicates a diagnosis of a subject as a categorical value, e.g., a Boolean value indicating that the subject has received the diagnosis.

[0055] The diagnostic machine learning model 108 is configured to process a model input that includes a set of subject features 106 for a subject to generate a model output that characterizes a likelihood that the subject has ASMD. A few examples of possible model outputs of the diagnostic machine learning model 108 are described next.

[0056] In some implementations, the model output of the diagnostic machine learning model can include a hard classification that identifies the subject as being included in one diagnostic category from a set of diagnostic categories that includes: a non-ASMD category (i.e., indicating that the subject does not have ASMD), and one or more ASMD categories (i.e., indicating that the subject has ASMD). In some cases, the set of diagnostic categories can include a single ASMD category, and a subject can be classified as being included in the ASMD category if the subject has any type of ASMD, e.g., type A, type A/B, or type B. In some cases, the set of diagnostic categories can include a respective category for each type of ASMD, e.g., type A, type A/B, and type B, and a subject can be classified as being included in an ASMD category is the subject is predicted to have the corresponding type of ASMD.

[0057] In some implementations, the model output of the diagnostic machine learning model can include a soft (probabilistic) classification that defines a score distribution over a set of diagnostic categories. The set of diagnostic categories can include a non-ASMD category and one or more ASMD categories, as described above. The score for each diagnostic category can define a likelihood (probability) that the subject is included in the diagnostic category.

[0058] The diagnostic machine learning model 108 can have any appropriate machine learning model architecture that enables the diagnostic machine learning model to perform its described functions. For instance, the diagnostic machine learning model 108 can be implemented, e.g., as a neural network model, or a random forest model, or a support vector machine model, or a decision tree model, or a linear regression model, etc. In implementations, where the diagnostic machine learning model 108 is implemented as a neural network model, the diagnostic machine learning model 108 can include any appropriate types of neural network layers (e.g., fully connected layers, convolutional layers, attention layers, etc.) in any appropriate number (e.g., 5 layers, 10 layers, or 50 layers) and connected in any appropriate configuration (e.g., as a linear sequence of layers). In implementations where the diagnostic machine learning model 108 is implemented as a decision tree model, the diagnostic machine learning model 108 can include any appropriate number of vertices, and can implement any appropriate splitting function at each vertex.

[0059] The diagnostic machine learning model 108 can include a set of machine learning model parameters. For instance, for a diagnostic machine learning model 108 implemented as a neural network model, the set of machine learning model parameters can define the weights and biases of the neural network layers of the diagnostic machine learning model. As another example, for a diagnostic machine learning model 108 implemented as a decision tree, the set of machine learning model parameters can define parameters of a respective splitting function used at each vertex of the decision tree. To generate a model output, the diagnostic machine learning model 108 can process a model input in accordance with values of the set of machine learning model parameters. [0060] The screening system 100 can use a training system 200 to train the diagnostic machine learning model 108 on a set of training examples. More specifically, the training system 200 can determine trained values of the set of machine learning model parameters of the diagnostic machine learning model 108 by a machine learning training technique. An example of a training system 200 for training the diagnostic machine learning model 108 is described in more detail with reference to FIG. 2.

[0061] The screening system 100 can generate an ASMD classification 110 for a subject based on a model output generated by the diagnostic machine learning model 108 by processing a set of subject features 106 characterizing the subject. For instance, the screening system 100 can classify a subject as being predicted to have ASMD if the probability of the subject having ASMD, as defined by the diagnostic machine learning model output, exceeds a threshold value. The threshold value can be, e.g., 0.7, 0.8, 0.9, or any other appropriate threshold value.

[0062] The screening system 100 can use the feature generation engine 104 and the diagnostic machine learning model 108 to generate a respective ASMD classification 110 for each subject in the population of subjects. More specifically, for each subject, the screening system 100 can use the feature generation engine 104 to generate a set of subject features 106 based on one or more medical records of the subject from the EMR database 102. The screening system 100 can then process the set of subject features 106 of the subject using the diagnostic machine learning model 108 to generate the ASMD classification 110 for the subject. The screening system 100 can then identify each subject that is classified by the diagnostic machine learning model 108 as having ASMD (the ASMD subjects 112), and output data identifying the ASMD subjects 112. For instance, the screening system 100 can output a respective patient identifier for each subject in the population of subjects that is classified as having ASMD. The data identifying the ASMD subjects 112 can be used in any of a variety of ways, as will be described in more detail below.

[0063] The operations performed by the feature generation engine 104 and the diagnostic machine learning model 108 are automated and parallelizable, and the screening system 100 can rapidly screen a large population of subjects. For instance, in some implementations, the screening system 100 can screen a population of at least 100,000 subjects in less than 1 hour, or less than 30 minutes, or less than 10 minutes, or less than 1 minute.

[0064] In some implementations, rather than including only a single diagnostic machine learning model 108, the screening system 100 can include an ensemble of multiple diagnostic machine learning models 108. For instance, the screening system 100 can include 10 diagnostic machine learning models, or 100 diagnostic machine learning models, or 1000 diagnostic machine learning models, or any other appropriate number of diagnostic machine learning models 108. Each diagnostic machine learning model in the ensemble can have a different architecture, or different model parameter values, or both, relative to the other diagnostic machine learning models in the ensemble. Example techniques for training an ensemble of diagnostic machine learning models are described in more detail with reference to FIG. 2.

[0065] The screening system 100 can use an ensemble of diagnostic machine learning models 108 to generate an ASMD classification 110 for a subject in any of a variety of ways. For instance, the screening system 100 can process a set of subject features 106 for the subject using each diagnostic machine learning model 108 in the ensemble to generate a respective model output from each diagnostic machine learning model 108 in the ensemble. The screening system 100 can aggregate the model outputs from the diagnostic machine learning models 108 in the ensemble to generate an ensemble output, and then determine the ASMD classification 110 of the subject based on the ensemble output. For instance, each diagnostic machine learning model 108 may generate a respective probability value that defines a probability that the subject has ASMD, and the ensemble output may be a measure of central tendency (e.g., a mean, median, or mode) of the probability values generated by the ensemble of diagnostic machine learning models 108. Generating ASMD classifications using an ensemble of diagnostic machine learning models 108 can cause the ASMD classifications to be more accurate and more stable. [0066] The screening system 100 can use an ensemble of diagnostic machine learning model 108 to generate both: (i) an ASMD classification 110 for a subject, and (ii) a confidence score for the ASMD classification 110 of the subject. The confidence score for an ASMD classification can define a confidence of the screening system 100 in the accuracy of the ASMD classification. To generate a confidence score for an ASMD classification 110 of a subject, the screening system 100 can determine a measure of variance of model outputs generated for the subject by the ensemble of diagnostic machine learning models. The screening system 100 can then generate the confidence score for the ASMD classification based on the measure of variance of the model outputs. For instance, each diagnostic machine learning model 108 in the ensemble can generate a model output that defines a respective probability of the subject having ASMD, and the measure of variance can be a standard deviation of the probability values generated by the ensemble. In this example, the screening system 100 can determine the confidence score based on, e.g., an inverse of the measure of variance, such that a low variance in predictions generated by the ensemble of diagnostic machine learning models 108 indicates a high confidence, and vice versa.

[0067] The screening system 100 can predict whether a subject has ASMD based on both: (i) the ASMD classification 110 of the subject, and (ii) the confidence score for the ASMD classification 110 of the subject. For instance, in some implementations the screening system 100 designates a subject as having ASMD only if: (i) the ASMD classification 110 of the subject indicates that the subject is predicted to have ASMD, and (ii) the confidence score satisfies (e.g., exceeds) a confidence threshold. The screening system 100 can thus reduce the likelihood of false positive classifications of ASMD subjects by filtering out ASMD classifications that are associated with low confidence scores, e.g., scores that are below the confidence threshold.

[0068] In some implementations, the screening system 100 can include a set of multiple diagnostic machine learning models 108 that are each specialized for generating ASMD classifications for a particular category of subject. More specifically, prior to generating an ASMD classification for a subject, the screening system 100 can assign the subject to a patient category from a set of patient categories based on the set of subject features 106 of the subject. The patient categories can be defined in any appropriate way. For instance, a patient category can be associated with a particular symptom (e.g., chronic pain, or pulmonary manifestations such as abnormal blood gas level, abnormal lung, abnormal pulmonary test, asphyxia hypoxemia, breathing abnormalities, bronchiolitis, chest pain, symptoms of chronic obstructive pulmonary disease, dyspnea, lung infection, progressive pulmonary function impairment, pulmonary fibrosis, or respiratory failure, or a combination thereof), and the screening system 100 can be configured to assign a subject to the patient category only if the subject has the symptom. As another example, a patient category can be associated with a particular medical diagnosis (e.g., interstitial lung disease), and the screening system 100 can be configured to assign a subject to the patient category only if the subject has the medical diagnosis. As another example, a patient category can be associated with a particular demographic feature (e.g., an age), and the screening system can be configured to assign a subject to the patient category only if the value of the demographic feature for the subject satisfies a threshold.

[0069] The set of diagnostic machine learning models 108 can include a respective diagnostic machine learning model 108 corresponding to each patient category in the set of patient categories. The diagnostic machine learning model 108 for a patient category can be specialized for generating ASMD classifications for patients included in the patient category. For instance, a training system 200 can train the diagnostic machine learning model 108 for a patient category only on training data derived from subjects that are included in the patient category. Example techniques for training a diagnostic machine learning model 108 specialized for a patient category are described in more detail with reference to FIG. 2.

[0070] To generate an ASMD classification 110 for a subject, the screening system 100 can assign the subject to a patient category, and then process a set of subject features 106 of the subject using the diagnostic machine learning model corresponding to the patient category to generate a model output. The screening system 100 can then generate the ASMD classification 110 of the subject based on the model output generated by the diagnostic machine learning model 108 specialized for the patient category of the subject. Specializing a diagnostic machine learning model 108 to a particular patient category can allow the diagnostic machine learning model 108 to learn to infer implicit classification rules and criteria that are optimized for accurately classifying subjects included in the patient category.

[0071] In some implementations, for each subject in the population of subjects, the screening system can generate a respective importance score for each feature in the set of subject features 106 for the subject. The importance score for a feature can characterize the impact of the feature on the ASMD classification 110 generated for the subject using the diagnostic machine learning model 108. The screening system 100 can generate importance scores for a set of subject features using any appropriate technique. For example, the importance scores can be “SHAP” values, i.e., “Shapley additive explanations,” or “local interpretable model-agnostic explanations,” i.e., “LIME” values, that can essentially determine importance scores for features by assessing how the prediction generated by the diagnostic machine learning model changes as the features are permuted. The screening system 100 can provide importance scores for a set of subject features 106 of a subject as explainability data, e.g., that explains and provides a rationale for the ASMD classification 110 generated for the subject by the diagnostic machine learning model 108. Generating importance scores can therefore increase user confidence in the validity of ASMD classifications 110 generated by the diagnostic machine learning model 108.

[0072] Optionally, the screening system 100 can aggregate the importance scores for subject features 106 across the population of subjects to generate an aggregate importance score for each feature in the set of subject features 106. More specifically, for each subject in the population of subjects, the screening system 100 can generate an importance score distribution that defines a respective importance score for each feature in the set of subject features. The screening system 100 can aggregate the importance score distributions for the subjects in the population of subjects to generate an aggregate importance score distribution. For instance, for each feature in the set of subject features, the aggregate importance score distribution can assign an aggregate importance score to the feature that is a measure of central tendency of the importance scores assigned to the feature by each subject-specific importance score distribution. (The measure of central tendency can be, e.g., a mean, a median, or a mode). The aggregate importance score distribution can be used in any of a variety of ways. For instance, the aggregate importance score distribution can be used as explainability data to interpret the rationale used by the diagnostic machine learning model 108 to generate ASMD classifications 110 for subjects across the population of subjects.

[0073] ASMD classifications 110 generated by the screening system 100 can be used in any of a variety of applications. A few example applications of ASMD classifications 110 are described next.

[0074] In some cases, ASMD classifications 110 can provide a basis for diagnostic testing 114 of subjects in the population of subjects. For instance, for each subject that the screening system 100 classifies as having ASMD, diagnostic (clinical) testing can be performed to confirm the diagnosis. Diagnostic testing for a subject can include a blood test to measure the amount of ASM enzyme activity in the subject. A diagnosis of ASMD can be confirmed, e.g., if the test shows decreased ASM enzyme activity.

[0075] In some cases, the ASMD classifications 110 can provide a basis for treating 116 subjects that are diagnosed with ASMD, e.g., by classification by the screening system 100 and by subsequent confirmatory diagnostic testing. Treating a subject for ASMD can include administering a drug to the subject, e.g., the drug olipudase alfa or olipudase alfa-rpcp, e.g., Xenpozyme™.

[0076] In some cases, the ASMD classifications can be used to facilitate further training 118 of the diagnostic machine learning model 108. More specifically, as described above, the screening system 100 can generate a confidence score for each ASMD classification that characterizes a level of confidence of the diagnostic machine learning model 108 in the ASMD classification. The screening system 100 can use the ASMD classifications and the confidence scores to select a proper subset of the population of subjects for diagnostic testing 114 for ASMD. Diagnostic testing for ASMD can be performed on the selected subjects, and the training system 200 can use the results of the diagnostic testing to generate training examples for use in further training of the diagnostic machine learning model 108. In particular, for each subject selected for diagnostic testing for ASMD, the training system 200 can generate a corresponding training example that includes: (i) a model input including a set of subject features for the subject, and (ii) a target ASMD classification. The target ASMD classification is based on the diagnostic testing for ASMD and defines whether the subject has ASMD. The training system 200 can train the diagnostic machine learning model 108 on the training example using a machine learning training technique, as will be described in more detail with reference to FIG. 2.

[0077] The screening system 100 can designate patients for diagnostic testing for ASMD, based on the ASMD classifications and the confidence scores, to maximize the increase in prediction accuracy achieved by further training of the diagnostic machine learning model 108. For instance, the screening system 100 can designate a number of subjects having the lowest confidence scores for diagnostic testing for ASMD. As another example, the screening system 100 can designate: (i) a first number of subjects having the lowest confidence scores from among the subjects classified as having ASMD, and (ii) a second number of subjects having the lowest confidence scores from among the subjects classified as not having ASMD, for diagnostic testing for ASMD. Training the diagnostic machine learning model 108 on training examples corresponding to subjects having ASMD classifications associated with low confidence scores can cause a significant increase in the prediction accuracy of the diagnostic machine learning model 108.

[0078] The screening system 100 and the training system 200 can operate in tandem to train the diagnostic machine learning model 108 over a sequence of update iterations. At each update iteration, the screening system 100 can screen the population of subjects to generate a respective ASMD classification and confidence score for each subject in the population of subjects. A proper subset of the population of subjects can be selected for diagnostic testing based on the ASMD classifications and the confidence scores. The training system 200 can generate new training examples based on the results of the diagnostic testing for ASMD and further train the diagnostic machine learning model 108. Then the process can repeat at the next update iteration. The prediction accuracy of the diagnostic machine learning model 108 can increase at each update iteration, and thus the screening system 100 can refine the screening of the population of subjects to reduce the number of false negative and false positives.

[0079] In some implementations, the screening system 100 can be used to generate an ASMD classification for a single subject, e.g., as an alternative to or in combination with generating a respective ASMD classification for each subject in a population of subjects. That is, the screening system 100 can be used to screen individual subjects for ASMD as well as to screen populations of subjects for ASMD.

[0080] FIG. 2 shows an example training system 200. The training system 200 is an example of a system implemented as computer programs on one or more computers in one or more locations in which the systems, components, and techniques described below are implemented.

[0081] The training system 200 is configured to train the diagnostic machine learning model 108 included in the screening system 100 described with reference to FIG. 1. The diagnostic machine learning model 108 is configured to process a model input that includes a set of subject features of a subject to generate a model output that characterizes a likelihood that the subject has ASMD.

[0082] The training system 200 uses a training engine 206 to train the set of machine learning model parameters 208 of the diagnostic machine learning model 108 on a set of training examples 204. Each training example can correspond to a subject (referred to for convenience as a “training subject”) and can include: (i) a model input that includes a set of subject features characterizing the training subject, and (ii) a target ASMD classification of the training subject. For each training example 204, the training engine 206 trains the diagnostic machine learning model 108 to process the model input of the training example to generate a model output that matches the target ASMD classification of the training example. More specifically, the training engine 206 trains the diagnostic machine learning model 108, by a machine learning training technique, to optimize an objective function that measures an error between: (i) the model output generated by the diagnostic machine learning model 108 for the training subject, and (ii) the target ASMD classification of the training subject. The objective function can measure the error between a model output and a target ASMD classification in any appropriate way, e.g., as a squared error or as an absolute error.

[0083] The training engine 206 can train the diagnostic machine learning model 108 using any machine learning training technique appropriate for the architecture of the diagnostic machine learning model 108. For instance, if the diagnostic machine learning model 108 is implemented as a neural network model, then the training engine 206 can train the diagnostic machine learning model using stochastic gradient descent.

[0084] The training system 200 can obtain the training examples used for training the diagnostic machine learning model 108 in a variety of possible ways. For instance, the training system 200 can generate “positive” training examples (i.e., corresponding to training subjects having ASMD) and “negative” training examples (i.e., corresponding to training subjects that do not have ASMD) by scanning EMR databases. In particular, the training system 200 can scan an EMR database to identify one or more training subjects that have been diagnosed with ASMD (e.g., through diagnostic testing), and generate a respective positive training example corresponding to each patient identified as having ASMD. Similarly, the training system 200 can scan an EMR database and generate negative training examples based on patients that are not diagnosed with ASMD.

[0085] Optionally, the training system 200 can construct the set of training examples by leveraging the statistical technique of matching based on confounding features, e.g., demographic features such as age, gender, geographic region of residence, etc. More specifically, for each of one or more positive training examples, the training system 200 can identify one or more corresponding negative training examples that “match” the positive training example based on a matching criterion evaluated based on values of the confounding feature dimensions. The training system 200 can then select the negative training examples that are matched to the positive training example for inclusion in the set of training examples used for training the diagnostic machine learning model. The matching criterion can define that a negative training example matches a positive training example if the confounding feature values of the negative training example are identical to (or within a tolerance threshold of) the confounding feature values of the positive training example. Constructing the set of training examples using statistical matching can reduce the impact of confounding variables and enhance the robustness of the trained diagnostic machine learning model.

[0086] ASMD is predicted to affect 1 in 250,000 people and is therefore a rare disease. Generating positive training examples (i.e., corresponding to training subjects having ASMD) by scanning EMR databases may yield only a small number of positive training examples, e.g., a few hundred positive training examples, or less than one hundred. However, training the diagnostic machine learning model 108 on only a small number of positive training examples may be insufficient to allow the diagnostic machine learning model to achieve an acceptable prediction accuracy. To address this issue, the training system 200 can optionally use a generative system 300 to generate “synthetic” training examples for use in training the diagnostic machine learning model 108. An example of a generative system is described in more detail with reference to FIG. 3.

[0087] In some implementations, as described with reference to FIG. 1, the screening system 100 includes an ensemble of diagnostic machine learning models 108 that each have a different architecture, or different model parameter values, or both, relative to the other diagnostic machine learning models in the ensemble. The training engine 206 can train each diagnostic machine learning model 108 in the ensemble on the set of training examples, e.g., as described above. Optionally, the training engine 206 can train each diagnostic machine learning model 108 in the ensemble on a different proper subset of the training examples 204 in order to introduce variation between the trained parameter values of the diagnostic machine learning models 108 in the ensemble. For instance, the training engine 206 can train each diagnostic machine learning model 108 in the ensemble on a different randomly selected proper subset of the set of training examples. [0088] In some implementations, as described with reference to FIG. 1, the screening system 100 includes a set of diagnostic machine learning models 108 that are each specialized for generating ASMD classifications for a particular category of subject. For each category of the subject, the training engine 206 can train the corresponding diagnostic machine learning model 108 on only training examples based on training patients that are included in the category of subject. For instance, for a patient category that includes only patients having a particular medical diagnosis (e.g., pneumonia, anemia, or interstitial lung disease), the training engine 206 can train the corresponding diagnostic machine learning model 108 on only training subjects having the medical diagnosis.

[0089] In some implementations, as described with reference to FIG. 1, the training system 200 can train the diagnostic machine learning model 108 over a sequence of update iterations. At each update iteration, one or more “test” subjects are selected for diagnostic testing for ASMD, and a respective new training example is generated for each test subject. At each update iteration, the training system 200 can add the new training examples to the set of training examples 204, and then train the diagnostic machine learning model 108 on the set of training examples.

[0090] In some implementations, the training engine 206 pre-trains the diagnostic machine learning model 108 to perform an auxiliary prediction task prior to training the diagnostic machine learning model 108 to perform the ASMD classification task. That is, the training engine 206 can pre-train the diagnostic machine learning model 108 to perform an auxiliary prediction task and then fine tune the diagnostic machine learning model on the ASMD classification task. Pre-training the diagnostic machine learning model 108 to perform the auxiliary prediction task can provide an effective initialization for the parameter values of the diagnostic machine learning model 108, thus accelerating training of the diagnostic machine learning model on the ASMD classification task. The auxiliary prediction task can be, e.g., to process a set of subject features of a subject to classify whether the subject has a medical condition other than ASMD, e.g., diabetes, heart disease, etc. The auxiliary prediction task can be a prediction task for which significantly more training data is available than for the ASMD prediction task.

[0091] FIG. 3 shows an example generative system 300. The generative system 300 is an example of a system implemented as computer programs on one or more computers in one or more locations in which the systems, components, and techniques described below are implemented.

[0092] The generative system 300 is configured to generate a set of synthetic training examples 304, for training the diagnostic machine learning model, based on a set of real -world training examples 302. A synthetic training example refers to a training example that does not directly correspond to a real-world subject, but rather, is computationally synthesized by the generative system 300. A real-world training example refers to a training example that directly corresponds to a real -world subject. The generative system 300 can provide the synthetic training examples 304 to the training system for use in training the diagnostic machine learning model, as described with reference to FIG. 2. [0093] The generative system 300 can generate the synthetic training examples 304 in any of a variety of possible ways. A few example implementations of the generative system 300 are described next.

[0094] In some implementations, the generative system 300 can generate a synthetic training example 304 by selecting a real-world training example 302, and adding random noise to the set of subject features included in the model input of the real-world training example 302. The generative system can define the target ASMD classification of the synthetic training example 304 as being the same as the target ASMD classification of the real-world training example 302. Thus the generative system 300 can generate the synthetic training example 304 as a noisy version of the real-world training example 302. In particular, the generative system 300 can generate synthetic positive training examples 304 as noisy versions of real-world positive training examples 302. The generative system 300 can sample the random noise values added to the model input of the real-world training example 302 from an appropriate probability distribution, e.g., a standard Normal distribution. (This implementation of the generative system 300 is illustrated in FIG. 3 as “adding random noise 306”).

[0095] In some implementations, to generate a synthetic training example, the generative system 300 can condition a generative neural network 308 on a target ASMD classification and then generate a set of subject features as an output of the generative neural network. The generative system 300 can then instantiate a synthetic training example that includes: (i) a model input that includes the set of subject features generated by the generative neural network 308, and (ii) the target ASMD classification used to condition the generative neural network 308. (“Conditioning” a neural network on a target ASMD classification can refer to providing the target ASMD classification as an input to the neural network). The generative system 300 can train the generative neural network 308 to generate realistic sets of subject features, e.g., that are drawn from the same distribution as the sets of subject features of the real -world training examples. A few example techniques by which the generative system 300 can train the generative neural network 308 are described next.

[0096] In one example, the generative system 300 can train the generative neural network as a generative adversarial neural network 310. More specifically, the generative system 300 can train the generative neural network jointly with another neural network, referred to as a discriminator neural network. The generative neural network is configured to process: (i) a tensor (e.g., a vector or matrix) of random values, and (ii) a target ASMD classification, to generate a synthetic set of subject features. The discriminator neural network is configured to process: (i) a target ASMD classification, and (ii) a set of subject features, to generate a discriminative score. The discriminative score defines a likelihood that the set of subject features was generated by the generative neural network instead of being a set of subject features characterizing a real -world subject. The generative system 300 alternates between training the generative neural network and the discriminator neural network. In particular, the generative system 300 trains the generative neural network to optimize an objective function, based on discriminative scores generated by the discriminator neural network, that encourages the generative neural network to generate sets of subject features that are misclassified by the discriminator neural network as characterizing real- world subjects. The generative system 300 trains the discriminator neural network to optimize an objective function that encourages the discriminator neural network to accurately distinguish between real-world sets of subject features and synthetic sets of subject features. After training, the discriminator neural network can be discarded, and the generative neural network can be used for generating synthetic training examples for training the diagnostic machine learning model 108. [0097] In another example, the generative system 300 can train the generative neural network 308 as a diffusion model 312. More specifically, the generative system 300 can train the generative neural network to process: (i) a target ASMD classification, and (ii) a noisy version of a set of subject features of a subject having the target ASMD classification, to generate a de-noised version of the set of subject features. A “noisy version” of a set of subject features refers to a set of subject features of a subject that have been combined with random noise sampled from a probability distribution, e.g., a standard Normal distribution. A “de-noised” version of a set of subject features refers to a set of subject features where the random noise has been removed. The generative system 300 can train the generative neural network on noisy and de-noised versions of real-world training examples. After training, the generative system 300 can use the generative neural network 308 to generate synthetic training examples.

[0098] To generate a synthetic training example, the generative system 300 can select a target ASMD classification, and generate initial values of a synthetic set of subject features by sampling each feature from a probability distribution, e.g., a standard Normal distribution. The generative system 300 can then iteratively de-noise the synthetic set of subject features over a sequence of de-noising iterations. More specifically, at the first de-noising iteration, the generative system 300 can process the initial values of the synthetic set of subject features using the generative neural network 308 to generate a de-noised version of the synthetic set of subject features. The generative system 300 can then provide the de-noised version of the synthetic set of subject features for processing by the generative neural network 308 at the next de-noising iteration. The generative system 300 can thus iteratively de-noise the synthetic set of subject features over the sequence of de-noising iterations using the generative neural network 308. The generative system 300 can generate a synthetic training example that includes: (i) a model input that includes the synthetic set of subject features generated at the final de-noising iteration, and (ii) the target ASMD classification.

[0099] FIG. 4A illustrates an example of a diagnostic machine learning model, implemented as a decision tree, that can be used by the screening system to screen a population of subjects (or a single subject) for ASMD. Each node in the decision tree (other than the leaf nodes) is associated with a splitting function that defines a rule for assigning a subject to a respective child node of the node. For instance, the root node 402 is associated with a splitting function that assigns a subject to a first child node 404 if the cholesterol HDL level of the subject is below 29.16 mg/dl, and assigns a subject to a second child node 406 if the cholesterol HDL level of the subject is above 29.16 mg/dl. The decision tree classifies a subject by progressing the subject from the root node of the decision tree to a leaf node of the decision tree. At each node prior to the leaf node, the decision tree applies the splitting function associated with the node to the set of features of the subject to assign the subject to a child node. Each leaf node in the decision tree has an associated ASMD classification; for instance, leaf node 410 is associated with a classification of “control” (e.g., identifying that the subject does not have ASMD), and leaf node 408 is associated with a classification of ASMD (e.g., identifying that the subject has ASMD). The decision tree can classify a subject based on the ASMD classification of the leaf node to which the subject is assigned. The example diagnostic machine learning model in FIG. 4A is provided for illustrative purposes only, and is non-limiting. Other implementations of the diagnostic machine learning model are possible, e.g., using a larger decision tree, or using a different machine learning model architecture (e.g., a random forest or a neural network), or using an ensemble of machine learning models, etc.

[0100] FIG. 4B shows a table of experimental results from applying the diagnostic machine learning model of FIG. 4A on a population of subjects. Row 412, 414, and 416 of the table each correspond to a respective leaf node of the decision tree and show: (i) a sequence of splitting decisions applied by the decision tree to reach the leaf node; (ii) a number of ASMD subjects classified into the leaf node; and (iii) a number of control (non-ASMD) subject classified into the leaf node.

[0101] FIG. 4C provides bar charts illustrating the respective prevalence of certain features among: (i) ASMD subjects, and (ii) control (non-ASMD) subjects. It will be appreciated that certain features are significantly more prevalent among ASMD subjects than among control subjects. A diagnostic machine learning model can learn to identify and leverage combinations of such features to precisely classify whether subjects have ASMD, as described throughout this specification.

[0102] FIG. 4D illustrates an example of a diagnostic machine learning model, implemented as a decision tree, that has been specialized through training for performing ASMD screening on a population of subjects having unexplained ILD.

[0103] FIG. 4E shows a table of experimental results from applying the diagnostic machine learning model of FIG. 4D on a population of subjects having unexplained ILD.

[0104] FIG. 5 is a flow diagram of an example process 500 for screening a database of electronic medical records for a population of subjects to identify subjects that are predicted to have ASMD. For convenience, the process 500 will be described as being performed by a system of one or more computers located in one or more locations. For example, a screening system, e g., the screening system 100 of FIG. 1, appropriately programmed in accordance with this specification, can perform the process 500.

[0105] The system obtains, for each subject, a set of features characterizing the subject from the database of electronic medical records (502). The set of features characterizing the subject comprise features can include features derived from a sample of bodily fluid or bodily tissue of the subject.

[0106] The system processes, for each subject, a model input that includes the set of features characterizing the subject using a diagnostic machine learning model to generate a model output that characterizes a likelihood that the subject has ASMD (504). The diagnostic machine learning model can include one or more of: a decision tree model, a random forest model, a neural network model, or a support vector machine model. [0107] The diagnostic machine learning model may have been trained, by a machine learning training technique, to determine trained values of the set of diagnostic machine learning model parameters. In particular, the diagnostic machine learning model may have been trained on a set of training examples that each include: (i) a training input based on a set of features of a training subject, and (ii) a target output based on whether the training subject is designated as having ASMD. Some of the positive training examples, i.e., for training subjects designated as having ASMD, may be generated by a generative neural network. The generative neural network may be trained as a generative adversarial neural network, or implemented as a diffusion model, as described above with reference to FIG. 3.

[0108] The system classifies, for each subject and based on the model output of the diagnostic machine learning model for the subject, whether the subject has ASMD (506). For instance, the system can classify a subject as having ASMD if the likelihood of the subject having ASMD, as defined by the model output for the subject, satisfies a threshold.

[0109] The system provides an output that identifies subjects, from the population of subjects, that are classified as having ASMD (508). In some cases, a drug for treating ASMD (e.g., olipudase alfa or olipudase alfa-rpcp) can be administered to one or more of the subjects classified as having ASMD, e.g., after confirmatory diagnostic testing.

[0110] FIG. 6 shows an example process for training a diagnostic machine learning model to identify subjects with ASMD, and then using the diagnostic machine learning model to screen a population of study subjects. More specifically, at a training phase 602, the screening system trains a diagnostic machine learning model (in this illustration, a decision tree 610) to perform ASMD classification on a population of training subjects 606. At an application phase 604, the screening system uses the diagnostic machine learning model to screen a population of study subjects 608 for ASMD. In this example, the population of study subjects is prescreened to include only subjects having unexplained ILD and being less than or equal to 50 years of age. As a result of the screening, N — 691 study subjects (from the population of N — 35,930 study subjects) are flagged as potentially having ASMD. The subjects flagged as potentially having ASMD may undergo subsequent clinical testing (e.g., enzyme and genetic testing in leukocytes, dried blood spots, or cultured fibroblasts) and (if appropriate) treatment for ASMD, as described throughout this specification. [0111] Certain aspects of the subject matter of this specification are further described in Appendix A.

[0112] This specification uses the term “configured” in connection with systems and computer program components. For a system of one or more computers to be configured to perform particular operations or actions means that the system has installed on it software, firmware, hardware, or a combination of them that in operation cause the system to perform the operations or actions. For one or more computer programs to be configured to perform particular operations or actions means that the one or more programs include instructions that, when executed by data processing apparatus, cause the apparatus to perform the operations or actions.

[0113] Embodiments of the subject matter and the functional operations described in this specification can be implemented in digital electronic circuitry, in tangibly-embodied computer software or firmware, in computer hardware, including the structures disclosed in this specification and their structural equivalents, or in combinations of one or more of them. Embodiments of the subject matter described in this specification can be implemented as one or more computer programs, i.e., one or more modules of computer program instructions encoded on a tangible non-transitory storage medium for execution by, or to control the operation of, data processing apparatus. The computer storage medium can be a machine-readable storage device, a machine- readable storage substrate, a random or serial access memory device, or a combination of one or more of them. Alternatively or in addition, the program instructions can be encoded on an artificially-generated propagated signal, e.g., a machine-generated electrical, optical, or electromagnetic signal, that is generated to encode information for transmission to suitable receiver apparatus for execution by a data processing apparatus.

[0114] The term “data processing apparatus” refers to data processing hardware and encompasses all kinds of apparatus, devices, and machines for processing data, including by way of example a programmable processor, a computer, or multiple processors or computers. The apparatus can also be, or further include, special purpose logic circuitry, e.g., an FPGA (field programmable gate array) or an ASIC (application-specific integrated circuit). The apparatus can optionally include, in addition to hardware, code that creates an execution environment for computer programs, e.g., code that constitutes processor firmware, a protocol stack, a database management system, an operating system, or a combination of one or more of them. [0115] A computer program, which may also be referred to or described as a program, software, a software application, an app, a module, a software module, a script, or code, can be written in any form of programming language, including compiled or interpreted languages, or declarative or procedural languages; and it can be deployed in any form, including as a stand-alone program or as a module, component, subroutine, or other unit suitable for use in a computing environment. A program may, but need not, correspond to a file in a file system. A program can be stored in a portion of a file that holds other programs or data, e.g., one or more scripts stored in a markup language document, in a single file dedicated to the program in question, or in multiple coordinated files, e.g., files that store one or more modules, sub-programs, or portions of code. A computer program can be deployed to be executed on one computer or on multiple computers that are located at one site or distributed across multiple sites and interconnected by a data communication network.

[0116] In this specification the term “engine” is used broadly to refer to a software-based system, subsystem, or process that is programmed to perform one or more specific functions. Generally, an engine will be implemented as one or more software modules or components, installed on one or more computers in one or more locations. In some cases, one or more computers will be dedicated to a particular engine; in other cases, multiple engines can be installed and running on the same computer or computers.

[0117] The processes and logic flows described in this specification can be performed by one or more programmable computers executing one or more computer programs to perform functions by operating on input data and generating output. The processes and logic flows can also be performed by special purpose logic circuitry, e.g., an FPGA or an ASIC, or by a combination of special purpose logic circuitry and one or more programmed computers.

[0118] Computers suitable for the execution of a computer program can be based on general or special purpose microprocessors or both, or any other kind of central processing unit. Generally, a central processing unit will receive instructions and data from a read-only memory or a random access memory or both. The essential elements of a computer are a central processing unit for performing or executing instructions and one or more memory devices for storing instructions and data. The central processing unit and the memory can be supplemented by, or incorporated in, special purpose logic circuitry. Generally, a computer will also include, or be operatively coupled to receive data from or transfer data to, or both, one or more mass storage devices for storing data, e.g., magnetic, magneto-optical disks, or optical disks. However, a computer need not have such devices. Moreover, a computer can be embedded in another device, e.g., a mobile telephone, a personal digital assistant (PDA), a mobile audio or video player, a game console, a Global Positioning System (GPS) receiver, or a portable storage device, e.g., a universal serial bus (USB) flash drive, to name just a few.

[0119] Computer-readable media suitable for storing computer program instructions and data include all forms of non-volatile memory, media and memory devices, including by way of example semiconductor memory devices, e.g., EPROM, EEPROM, and flash memory devices; magnetic disks, e.g., internal hard disks or removable disks; magneto-optical disks; and CD-ROM and DVD-ROM disks.

[0120] To provide for interaction with a user, embodiments of the subject matter described in this specification can be implemented on a computer having a display device, e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor, for displaying information to the user and a keyboard and a pointing device, e.g., a mouse or a trackball, by which the user can provide input to the computer. Other kinds of devices can be used to provide for interaction with a user as well; for example, feedback provided to the user can be any form of sensory feedback, e.g., visual feedback, auditory feedback, or tactile feedback; and input from the user can be received in any form, including acoustic, speech, or tactile input. In addition, a computer can interact with a user by sending documents to and receiving documents from a device that is used by the user; for example, by sending web pages to a web browser on a user’s device in response to requests received from the web browser. Also, a computer can interact with a user by sending text messages or other forms of message to a personal device, e.g., a smartphone that is running a messaging application, and receiving responsive messages from the user in return.

[0121] Data processing apparatus for implementing machine learning models can also include, for example, special-purpose hardware accelerator units for processing common and computeintensive parts of machine learning training or production, i.e., inference, workloads.

[0122] Machine learning models can be implemented and deployed using a machine learning framework, e.g., a TensorFlow framework, or a Jax framework.

[0123] Embodiments of the subject matter described in this specification can be implemented in a computing system that includes a back-end component, e.g., as a data server, or that includes a middleware component, e.g., an application server, or that includes a front-end component, e.g., a client computer having a graphical user interface, a web browser, or an app through which a user can interact with an implementation of the subject matter described in this specification, or any combination of one or more such back-end, middleware, or front-end components. The components of the system can be interconnected by any form or medium of digital data communication, e.g., a communication network. Examples of communication networks include a local area network (LAN) and a wide area network (WAN), e.g., the Internet.

[0124] The computing system can include clients and servers. A client and server are generally remote from each other and typically interact through a communication network. The relationship of client and server arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other. In some embodiments, a server transmits data, e.g., an HTML page, to a user device, e.g., for purposes of displaying data to and receiving user input from a user interacting with the device, which acts as a client. Data generated at the user device, e.g., a result of the user interaction, can be received at the server from the device.

[0125] While this specification contains many specific implementation details, these should not be construed as limitations on the scope of any invention or on the scope of what may be claimed, but rather as descriptions of features that may be specific to particular embodiments of particular inventions. Certain features that are described in this specification in the context of separate embodiments can also be implemented in combination in a single embodiment. Conversely, various features that are described in the context of a single embodiment can also be implemented in multiple embodiments separately or in any suitable subcombination. Moreover, although features may be described above as acting in certain combinations and even initially be claimed as such, one or more features from a claimed combination can in some cases be excised from the combination, and the claimed combination may be directed to a subcombination or variation of a subcombination.

[0126] Similarly, while operations are depicted in the drawings and recited in the claims in a particular order, this should not be understood as requiring that such operations be performed in the particular order shown or in sequential order, or that all illustrated operations be performed, to achieve desirable results. In certain circumstances, multitasking and parallel processing may be advantageous. Moreover, the separation of various system modules and components in the embodiments described above should not be understood as requiring such separation in all embodiments, and it should be understood that the described program components and systems can generally be integrated together in a single software product or packaged into multiple software products.

[0127] Particular embodiments of the subject matter have been described. Other embodiments are within the scope of the following claims. For example, the actions recited in the claims can be performed in a different order and still achieve desirable results. As one example, the processes depicted in the accompanying figures do not necessarily require the particular order shown, or sequential order, to achieve desirable results. In some cases, multitasking and parallel processing may be advantageous.

[0128] What is claimed is:

Claims

1. A method performed by one or more computers, the method comprising: screening a database of electronic medical records for a population of subjects to identify subjects that are predicted to have acid sphingomyelinase deficiency, comprising, for each subject in the population of subjects: obtaining, from the database of electronic medical records, a set of features characterizing the subject; processing a model input that comprises the set of features characterizing the subject using a diagnostic machine learning model, having a set of diagnostic machine learning model parameters, to generate a model output that characterizes a likelihood that the subject has acid sphingomyelinase deficiency; and classifying, based on the model output of the diagnostic machine learning model, whether the subject has acid sphingomyelinase deficiency; and providing an output that identifies subjects, from the population of subjects, that are classified as having acid sphingomyelinase deficiency.

2. The method of claim 1, wherein the diagnostic machine learning model has been trained, by a machine learning training technique, to determine trained values of the set of diagnostic machine learning model parameters.

3. The method of claim 2, wherein training the diagnostic machine learning model by the machine learning training technique comprises: obtaining a set of training examples, wherein each training example comprises: (i) a training input comprising a set of features of a training subject, and (ii) a target output based on whether the training subject is designated as having acid sphingomyelinase deficiency; and training the diagnostic machine learning model on the set of training examples.

4. The method of claim 3, wherein obtaining the set of training examples comprises: obtaining a set of positive training examples, wherein each positive training example corresponds to a subject having acid sphingomyelinase deficiency; and selecting a plurality of negative training examples that each correspond to a respective subject that does not have acid sphingomyelinase deficiency, comprising, for each positive training example: identifying a corresponding set of negative training examples based on matching one or more confounding features of the set of negative training examples with the positive training example; and selecting the set of negative training examples for training the diagnostic machine learning model.

5. The method of claim 4, wherein the one or more confounding features comprise demographic features.

6. The method of any one of claims 3-5, wherein training the diagnostic machine learning model on the set of training examples comprises: training the diagnostic machine learning model to, for each training example, process the training input of the training example to generate a model output that matches the target output of the training example.

7. The method of any one of claims 3-6, wherein obtaining the set of training examples comprises, for one or more training examples: generating the training example from one or more electronic medical records of the training subject corresponding to the training example.

8. The method of any one of claims 3-7, wherein obtaining the set of training examples comprises: generating a set of features of a subject using a generative neural network; and generating a training example that includes a training input based on the set of features generated by the generative neural network.

9. The method of claim 8, wherein the generative neural network has been trained by operations comprising: generating an output set of features using the generative neural network; processing the output set of features, generated by the generative neural network, using a discriminator neural network to generate a discriminative score that defines a likelihood that the output set of features was generated by the generative neural network instead of being derived from one or more electronic medical records; and training the generative neural network to optimize an objective function that depends on the discriminative score generated by the discriminator neural network.

10. The method of claim 8, wherein generating a set of features of a subject using the generative neural network comprises: instantiating initial values of the set of features of the subject as random values sampled from a probability distribution; and de-noising the set of features over a sequence of de-noising iterations, comprising, at each de-noising iteration prior to a last de-noising iteration in the sequence of de-noising iterations: processing current values of the set of features using a generative neural network to generate a de-noised version of the set of features; and providing the de-noised version of the set of features for processing at the next de- noising iteration; wherein a de-noised version of the set of features generated at the last de-noising iteration defines the set of features of the subject.

11. The method of any preceding claim, further comprising: generating, for each subject in the population of subjects, a confidence score that characterizes a confidence of the diagnostic machine learning model in the classification of whether the subject has acid sphingomyelinase deficiency; selecting a proper subset of the population of subjects for diagnostic testing for acid sphingomyelinase deficiency based on the confidence scores; re-training the diagnostic machine learning model based on results of the diagnostic testing of the proper subset of the population of subjects; and re-classifying whether each subject in the population of subjects has acid sphingomyelinase deficiency using the re-trained diagnostic machine learning model.

12. The method of claim 11, wherein selecting the proper subset of the population of subjects for diagnostic testing for acid sphingomyelinase deficiency based on the confidence scores comprises: selecting a number of subjects associated with lowest confidence scores, from among the population of subjects, for diagnostic testing for acid sphingomyelinase deficiency.

13. The method of claim 11, wherein selecting the proper subset of the population of subjects for diagnostic testing for acid sphingomyelinase deficiency based on the confidence scores comprises: selecting a number of subjects associated with the lowest confidence scores, from among subjects classified as having acid sphingomyelinase deficiency, for diagnostic testing for acid sphingomyelinase deficiency; and selecting a number of subjects associated with the lowest confidence scores, from among subjects classified as not having acid sphingomyelinase deficiency, for diagnostic testing for acid sphingomyelinase deficiency.

14. The method of any one of claims 11-13, wherein generating, for each subject in the population of subjects, a confidence score that characterizes a confidence of the diagnostic machine learning model in the classification of whether the subject has acid sphingomyelinase deficiency comprises, for each subject in the population of subjects: generating a respective model output that characterizes a likelihood that the subject has acid sphingomyelinase deficiency using each diagnostic machine learning model in an ensemble of multiple diagnostic machine learning models; and generating the confidence score based on a measure of variance of the model outputs generated the ensemble of diagnostic machine learning models for the subject.

15. The method of any preceding claim, wherein the population of subjects comprises at least 10,000 subjects.

16. The method of any preceding claim, wherein screening the data of electronic medical records for the population of subjects is performed in 1 minute or less.

17. The method of any preceding claim, further comprising, for one or more of the subjects that are classified as having acid sphingomyelinase deficiency: administering a drug to the subject to treat the acid sphingomyelinase deficiency in the subject.

18. The method of claim 17, wherein the drug comprises olipudase alfa or olipudase alfa-rpcp.

19. The method of any preceding claim, wherein the set of features characterizing the subject comprise features derived from a sample of bodily fluid or bodily tissue of the subject.

20. The method of any preceding claim, wherein the diagnostic machine learning model comprises one or more of: a decision tree model, a random forest model, a neural network model, or a support vector machine model.

21. The method of any preceding claim, wherein the population of subjects is prescreened to include only subjects that have been diagnosed with one or more of: interstitial lung disease (ILD), splenomegaly, or nonalcoholic steatohepatitis (NASH).

22. A system comprising: one or more computers; and one or more storage devices communicatively coupled to the one or more computers, wherein the one or more storage devices store instructions that, when executed by the one or more computers, cause the one or more computers to perform operations of the respective method of any one of claims 1-21.

23. One or more non-transitory computer storage media storing instructions that when executed by one or more computers cause the one or more computers to perform operations of the respective method of any one of claims 1-21 .