Peripheral blood smear-based deep learning to predict cancer-associated thrombosis

Abstract

Presented herein are systems and methods of determining values indicating degrees of risk for venous thromboses (VTEs) in subjects. A computing system can identify a first plurality of images for a first subject at risk of VTE, each of the first plurality of images corresponding to a respective white blood cell (WBC) in a first blood sample obtained from the first subject at a first time. The computing system can provide the first plurality of images to a machine learning (ML) architecture. The computing system can generate a plurality of embedding sets, each embedding set of the plurality of embedding sets corresponding to a respective image of the first plurality of images. The computing system can determine, based on executing the ML architecture, a value indicating a degree of risk of VTE for the first subject at the time interval relative to the first time.

Description

PERIPHERAL BLOOD SMEAR-BASED DEEP LEARNING TO PREDICTCANCER-ASSOCIATED THROMBOSISCROSS-REFERENCE TO RELATED APPLICATIONS

[0001] The present application claims the benefit of and priority to U.S. Provisional Patent Application 63 / 677,234, titled “Peripheral Blood Smear-Based Deep Learning to Predict Cancer-Associated Thrombosis,” filed July 30, 2024, which is incorporated herein by reference in its entirety.BACKGROUND

[0002] A computing device can use machine learning models to process an input to generate an output.SUMMARY

[0003] Aspects of the present disclosure are related to systems and methods of determining values indicating degrees of risk for venous thromboses (VTEs) in subjects. One more processors can identify a first plurality of images for a first subject at risk of VTE, each of the first plurality of images corresponding to a respective white blood cell (WBC) in a first blood sample obtained from the first subject at a first time. The one or more processors can provide the first plurality of images to a machine learning (ML) architecture, wherein the ML architecture can be established using a plurality of examples, each of the plurality of examples comprising: (i) a respective plurality of embedding sets for each of a second plurality of images, each of the second plurality of images corresponding to a respective WBC in a second blood sample obtained from a second subject at a second time, and (ii) a label indicating one of the presence or absence of VTE at a time interval relative to the second time. The one or more processors can generate, based on executing the ML architecture, a plurality of embedding sets, each embedding set of the plurality of embedding sets corresponding to a respective image of the first plurality of images. The one or more processors can determine, based on executing the ML architecture using the plurality ofembedding sets, a value indicating a degree of risk of VTE for the first subject at the time interval relative to the first time. The one or more processors can store, using one or more data structures, an association between the first subject and the value.

[0004] In some embodiments, the one or more processors can identify the subject as a non-candidate for administration of a medication for VTE, responsive to the value not satisfying a threshold. The one or more processors can provide, for presentation via a user interface, an output identifying the subject as the non-candidate for administration of the medication, wherein subsequent to provision of the output, administration of the medication for VTE may be refrained. The one or more processors can identify the subject as a candidate for administration of a medication for VTE, responsive to the value satisfying a threshold. The one or more processors can provide, for presentation via a user interface, an output identifying the subject as the candidate for administration of the medication. Subsequent to provision of the output, the subject may be administered with the medication, wherein the medication for VTE comprises at least one of a direct oral anticoagulant (DOAC), a low molecular weight heparin (LMWH), an unfractionated heparin (UFH), a vitamin K antagonist, or a fondaparinux.

[0005] In some embodiments, the one or more processors can generate a classification indicating a risk level of a plurality of risk levels for the first subject, in accordance with a comparison between the value and a respective range for each of the plurality of risk levels. The one or more processors can provide, for presentation via a user interface, an output identifying the classification for the first subject. The one or more processors can receive a biomedical image of the first blood sample having first WBCs and first red blood cells (RBCs) obtained from the first subject at the first time, wherein the biomedical image may be acquired in accordance with peripheral blood smear (PBS). The one or more processors can identify, from the biomedical image, the first plurality of images comprising (i) a first set of WBC images corresponding to the first WBCs and (ii) a first set of RBC images corresponding to the first RBCs. The one or more processors can provide the first set of WBC images and the first set of RBC images as input to the ML architecture to generate the first plurality of embedding sets, wherein at least one of the plurality of examples comprises respective plurality of embedding sets for each of the secondplurality of images, the second plurality of images comprising (i) a second set of WBC images corresponding to second WBCs in the second blood sample, and (ii) a second set of RBC images corresponding to second RBCs in the second blood sample.

[0006] In some embodiments, the one or more processors can receive a first dataset comprising at least one of (i) a first electronic document defining a first plurality of traits associated with the first subject, (ii) a first gene profile defining a first plurality of characteristics of genes in a first sample from the first subject, or (iii) a first sequence record comprising a first plurality of sequences associated with the genes in the first sample of the first subject. The one or more processors can provide the first dataset to the ML architecture, wherein at least one of the plurality of examples comprises at least one of (i) a second electronic document defining a second plurality of traits associated with the second subject, (ii) a second gene profile defining a second plurality of characteristics of genes in a second sample from the second subject, or (iii) a second sequence record comprising a second plurality of sequences associated with the genes in the second sample of the second subject. The one or more processors can determine, based on executing the ML architecture using the first dataset, the value indicating the degree of risk of VTE for the first subject.

[0007] In some embodiments, the one or more processors can generate, based on providing the first plurality of images to the ML architecture, the first plurality of embedding sets, each embedding set of the plurality of embedding sets identifying a set of morphological features of the respective WBC of the respective image of the first plurality of images. The one or more processors can generate, based on executing the ML architecture, a plurality of attention scores for the set of morphological features, each attention score of the plurality of attention scores indicating a degree of weight of a respective morphological feature of the set of morphological features in determining the value.

[0008] The ML architecture further comprises a feature encoder configured to generate the plurality of embedding sets using the first plurality of images, each embedding set of the plurality of embedding sets corresponding to a reduced dimensional representation of therespective image of the first plurality of images, an aggregator configured to determine the value indicating the degree of risk using the plurality of embedding sets, and a classifier configured to generate a classification based on the value and a first dataset.

[0009] The VTE comprises at least one of a deep vein VTE (DVT) or a pulmonary embolism (PE), wherein the VTE affects at least one of a leg vein, an arm vein, a pelvic vein, a pulmonary artery, lobar artery, or segment artery, wherein the time interval comprises at least one of 5 days, 7 days, 10 days, 14 days, 21 days, 28 days, 1 month, 2 months, 3 months, 4 months, 5 months, 6 months, 7 months, or 8 months. The subject may be at risk of or diagnosed with cancer, wherein the cancer comprises at least one of carcinoma, sarcoma, hematopoietic cancer, adrenal cancer, bladder cancer, blood cancer, bone cancer, brain cancer, breast cancer, carcinoma, cervical cancer, colon cancer, colorectal cancer, corpus uterine cancer, ear, nose and throat (ENT) cancer, endometrial cancer, esophageal cancer, gastrointestinal cancer, head and neck cancer, Hodgkin’s disease, intestinal cancer, kidney cancer, larynx cancer, leukemia, liver cancer, lymph node cancer, lymphoma, lung cancer, melanoma, mesothelioma, myeloma, nasopharynx cancer, neuroblastoma, non- Hodgkin’s lymphoma, oral cancer, ovarian cancer, pancreatic cancer, penile cancer, pharynx cancer, prostate cancer, rectal cancer, sarcoma, seminoma, skin cancer, stomach cancer, teratoma, testicular cancer, thyroid cancer, uterine cancer, vaginal cancer, vascular tumor, or metastases thereof.BRIEF DESCRIPTION OF THE DRAWINGS|0010] The foregoing and other objects, aspects, features, and advantages of the disclosure will become more apparent and better understood by referring to the following description taken in conjunction with the accompanying drawings, in which:

[0011] FIG. 1 depicts a block diagram of a schematic to determine values indicating a degree of risk for a subject to contract venous thromboses (VTE).(0012[ FIG. 2 depicts a table indicating performance of a machine learning architecture.

[0013] FIGs. 3A-3D depict graphs of attention scores of various cellular morphology with respect to predicting VTE in 24 hours (FIG. 3 A), 7 days (FIG. 3B), 30 days (FIG. 3C), and 6 months (FIG. 3D).

[0014] FIGs. 4A-4D depict images of red blood cells (RBCs) and white blood cells (WBCs) with the highest attention scores with respect to predicting VTE in 24 hours (FIG. 4A), 7 days (FIG. 4B), 30 days (FIG. 4C), and 6 months (FIG. 4D).

[0015] FIG. 5 depicts a graph of important features as determined from training a random forest classifier.

[0016] FIG. 6 depict a graph of important features as determined from training a deeplearning based machine learning model in accordance with multiple-instance learning (MIL).

[0017] FIG. 7 depicts a block diagram of a system for determining values indicating degrees of risk for venous thromboses (VTEs) in subjects, in accordance with an illustrative embodiment.|0018] FIG. 8 depicts a block diagram of a process to train a machine learning (ML) architecture to determine a value indicating a degree of risk of VTE for a subject, in accordance with an illustrative embodiment.

[0019] FIG. 9 depicts a block diagram of a process to determine a value indicating the degree of risk of VTE for a subject based on a blood sample of the subject; in accordance with an illustrative embodiment.

[0020] FIG. 10 depicts a block diagram of a process to provide an output based on the determination of the values indicating risk of VTE, in accordance with an illustrative embodiment.

[0021] FIG. 11 depicts a flowchart of a method of training the ML architecture to determine values indicating VTE risk; in accordance with an illustrative embodiment.

[0022] FIG. 12 depicts a flowchart of a method of determining values indicating a degree of risk for VTE and providing an output identifying the subject as a candidate or non-candidate for administration of medication, in accordance with an illustrative embodiment.

[0023] FIG. 13 is a block diagram of a computing environment according to an example implementation of the present disclosure.DETAILED DESCRIPTION10024] Following below are more detailed descriptions of various concepts related to, and embodiments of, systems and methods for determining values indicating degree of risk for thrombosis in subjects. It should be appreciated that various concepts introduced above and discussed in greater detail below may be implemented in any of numerous ways, as the disclosed concepts are not limited to any particular manner of implementation. Examples of specific implementations and applications are provided primarily for illustrative purposes.A: Peripheral Blood Smear-Based Deep Learning to Predict Cancer-Associated Thrombosis|0025] Patients with cancer have an elevated risk of developing venous thromboembolism(VTE), a common source of morbidity and mortality. Neutrophil Extracellular Traps (NET’s) are thought to contribute to the hypercoagulable state of malignancy. NET’s are classified as smudge cells or artifacts using the CellaVision analyzer (CellaVision, Sweden), an automated system for performing hematology laboratory differentials. New VTE biomarkers could potentially improve risk estimation for this subpopulation of patients. It is hypothesized that cellular morphology of white blood cells (WBC’s), especially smudge cells and artifacts, captured by imaging would be predictive of VTE in individuals with cancer.(0026] The cohort included patients with solid tumors and hematological malignancies who had at least one peripheral blood smear scanned by the CellaVision analyzer. VTE events consisting of lower extremity deep vein thrombosis or pulmonary embolism were flagged using the CEDARS+PINES natural language processing platform. WBC images were obtained from aclinical CellaVision database. Observation times were discretized to intervals of 7 days to make predictions. Data was partitioned into training, validation, and test sets (68% train, 12% validation, 20% test). The area under the receiver operating characteristic curve (AUROC) was used to evaluate model performance, with an AUROC of 0.5 being consistent with random chance. The proportions of each WBC type were calculated for each slide and compared using the Mann- Whitney U test. A ResNext-based classifier was trained on 20,000 CellaVision images with 18 classes. This network can be used to generate a feature embedding for each WBC image instance in the selected cohort of patients. A gated-attention-based multiple instance learning model can be trained to predict VTE development within the designated time period, using the previously derived slide-specific WBC image embeddings. Control undersampling can be used to ensure class balance during training. Models can be trained on patients with hematological malignancies, solid tumors, and the combined cohort. Attention scores can be generated with these models, indicating the relative weight of each cell to the VTE prediction. To analyze attention scores, the proportions of different cell morphologies between the top 10% of scores and the remaining 90% among correctly identified slides can be compared using a two-proportion Z-test.

[0027] The cohort contained 53,789 patients (41,598 solid tumors, 12,191 hematological malignancies). Within 7 days, 411 patients (0.8%) developed a VTE event. In patients with solid tumors, slides from those who developed VTE had increased proportions of lymphocytes (p < 0.01). For patients with heme malignancies, slides from those who developed VTE were enriched with lymphocytes and eosinophils (p < 0.01). All the models predicting VTE at 7 days exhibited an AUROC significantly higher than 0.50. The heme malignancies cohort model yielded an AUROC of 0.72 (95% CI: 0.60-0.85), while the AUROC for the model trained on the solid tumor cohort was 0.70 (95% CI: 0.65-0.76). This is compared to a metric value of 0.68 (95% CI: 0.61- 0.74) for the combined cohort. In the model predicting VTE using the combined cohort, top attention scores showed higher proportions of smudge cells and lymphocytes (p < 0.01). When predicting VTE among patients with heme malignancies, smudge cells, eosinophils, and artifacts were the top morphologies disproportionately represented among the cells assigned top attentionscores (p < 0.01). In the group of patients with solid tumors, eosinophils, monocytes, and smudge cells were overrepresented among the cells with the highest attention scores (p < 0.01).

[0028] Information derived from peripheral white blood cell images embeddings through the use of multiple instance deep learning can help improve risk estimation models for cancer- associated VTE.B: Deep Learning Identifies Peripheral Blood Smear Morphologic Predictors of Cancer- Associated Venous Thromboembolism

[0029] Venous thromboembolism (VTE) is a major cause of morbidity and mortality in cancer patients. Currently available models aimed at predicting VTE events feature clinical, demographic, laboratory and sometimes pharmacological predictors. Morphological features of peripheral blood cells have not been used so far. The systems and methods described herein assess if deep learning-based multiple instance learning (MIL) models applied to peripheral blood smear images could be derived to make clinically useful predictions and identify morphological predictors of cancer-associated thrombosis events. This retrospective cohort study included patients with cancer. The MIL models were trained using peripheral blood smear image encodings, including both white blood cell (WBC) and red blood cell (RBC) patches. The primary outcome was incident VTE at 1, 7, 30, and 180 days.

[0030] Performance was assessed using the concordance index (C-index). Among patients included, models combining CBC data with WBC and RBC morphologic features outperformed those using CBC or MIL features alone. The highest 7-day prediction performance (C-index 0.7521) was achieved using CBC + WBC MIL features. Attention analysis confirmed neutrophil predominance (segs and bands) as a major risk driver. In conclusion, deep learning applied to peripheral blood smear images revealed morphologic features associated with VTE risk in cancer patients. This approach may facilitate early risk stratification and inform prophylactic strategies.Introduction(0031 ] Venous thromboembolism (VTE) is a common and impactful complication of cancer and the treatment. The risk of VTE varies substantially across tumor types. Anticoagulants can be administered prophylactically to decrease the risk of cancer-associated thrombosis (CAT). However, this can sometimes result in bleeding events and can incur significant costs for patients. Based on these considerations, several clinical prediction rules and other models have been derived to better estimate the risk of CAT in individual patients. The most commonly applied risk estimation tool is the Khorana score, which utilizes complete blood count (CBC) values, cancer type and body mass index to estimate VTE risk for a six-month period. While the Khorana score has contributed to improve the care of patients with cancer, recall and precision are modest, and include limited usefulness.

[0032] The most important mechanisms by which CAT develops remain unclear. One leading hypothesis includes NETosis, in which neutrophil extracellular traps (NET’s) are released due to inflammatory conditions, activating plasmatic coagulation and resulting in thrombus formation. Previous studies have reported NET formation across a range of malignancies. Based on this hypothesis, quantification of NET’s on peripheral blood smears could be leveraged to better estimate the risk of CAT. Peripheral blood smears are regularly prepared for patients with solid and hematological malignancies. Standard peripheral blood analyzers such as CellaVision often misclassify NET’s as smudge cells or artifacts, requiring manual annotation by pathologists for accurate identification.

[0033] Given these limitations, deep learning offers a promising solution for high- throughput morphological analysis. These tools have successfully been applied to bone marrow aspirates for cell classification and outcome prediction, demonstrating potential for broader clinical applications. In this study, what is desired is multiple instance learning (MIL) models to predict VTE development in patients with solid and hematologic malignancies using images of peripheral blood smears and identify predictive peripheral blood cellular morphologies associated with thrombotic risk in order to ultimately improve mechanistic understanding of cancer- associated thrombosis.MethodsCohort Selection

[0034] A list of peripheral blood smears (PBS) prepared from patients with solid tumors and hematologic malignancies was generated. The corresponding images were retrieved from the clinical database. Electronic medical records were reviewed using the CEDARS+PINE natural language processing pipeline, covering a time window from three months before to nine months after each smear. Patients with prior events or those taking anticoagulants and captured new events occurring up to six months after the smear were excluded. Patients without CBC data corresponding to a PBS or patients with missing CBC lab values were excluded. Primary cancer diagnoses were obtained from the CellaVision Cancer Registry and classified per ICD-10 code. Time periods for VTE outcome windows were discretized into 7 days, 30 days, and 6 months.Model Development

[0035] The systems and methods described herein include a multimodal, multi-encoder MIL model that integrates two ResNeXt-50 image encoders-one for RBC’s and one for WBC’s- to predict the occurrence of VTE events. To incorporate CBC laboratory values, random forest classifier that combines the MIL-generated probabilities from CellaVision images with CBC and CellaVision data was trained. The dataset was split into 80% for training, 10% for validation, and 10% for testing. To benchmark the model performance against conventional VTE risk scoring tools, trained logistic regression models using CBC-derived predictors are used. Depicted in FIG. 1 is a block diagram of a schematic to determine values indicating a degree of risk for a subject to contract venous thromboses (VTE).Statistical Analysis

[0036] To evaluate model performance, one or more processors computed the area under the receiver operating curve (AUROC) alongside a bootstrapped 95% confidence interval as shown in FIG. 2. AUROCs were calculated for both MIL and random forest models. Statisticalsignificance was defined as greater than 0.5 (e.g., greater than random chance). VTE risk probabilities were generated from the random forest models and stratified. Cumulative incidence functions per cancer diagnosis were fit based on VTE risk stratification, accounting for competing risks. Sub distribution hazard ratios were estimated using a Fine-Gray competing risks regression model.Results

[0037] The test generated and used a cohort of 53,789 patients. Among these patients, VTE events occurred infrequently across all time intervals, across 24 hours, 7 days, 30 days, and 6 months of the index PBS. These findings highlight both the rarity and delayed timing of VTE occurrence relative to the peripheral blood smear collection across the cohort.CBC values combined with deep learning-derived features predict VTE incidence across multiple time intervals.

[0038] The random forest models, which integrated MIL-derived probabilities with CBC and Cellavision features, outperformed logistic regression models that used only CBC predictors. Logistic regression models based on white blood cell (WBC), platelet, and hemoglobin values achieved AUCs of 0.60, 0.63, and 0.57 for predicting VTE within 7 days, 30 days, and 6 months, respectively.

[0039] When comparing MIL models, those using WBC-derived features consistently outperformed those using RBC-derived features across all time intervals, except at 24 hours. Integration of both RBC and WBC derived features improved performance for all time intervals. Among the random forest models, the highest performance for predicting VTE within 1 day (AUC: 0.732, 95% CI: 0.693-0.770) and 7 days (AUC: 0.751, 95% CI: 0.727-0.774) was achieved by integrating MIL-derived probabilities from RBCs along with CBC and Cellavision data. For predicting VTE within 30 days (AUC: 0.715, 95% CI: 0.701-0.729) and 6 months (AUC: 0.641, 95% CL 0.631-0.650), maximal performance was attained by integrating MIL-derived probabilities from both RBCs and WBCs, in combination with CBC and Cellavision data.Referring to FIG. 5, depicted is a graph of important features as determined from training a random forest classifier. Referring to FIG. 6, depicted is a graph of important features as determined from training a deep-learning based machine learning model in accordance with multiple-instance learning (MIL).Distinct RBC and WBC Morphologies Are Differentially Enriched Across VTE Prediction Time Points

[0040] Across all time intervals, log odds ratio analysis of the top 10% highest-attention cells revealed consistent enrichment of schistocytes, hypochromic cells, macrocytes, and echinocytes among RBC-derived features in VTE cases. This pattern was observed consistently from 24 hours up to 6 months post-PBS collection. Referring now to FIGs. 3 A-D, depicted graphs of attention scores of various cellular morphology with respect to predicting VTE in 24 hours (FIG. 3A), 7 days (FIG. 3B), 30 days (FIG. 3C), and 6 months (FIG. 3D). Referring to FIGs. 4A-D, depicted are selected images of red blood cells (RBCs) and white blood cells (WBCs) with the highest attention scores with respect to predicting VTE in 24 hours (FIG. 4A), 7 days (FIG. 4B), 30 days (FIG. 4C), and 6 months (FIG. 4D).[00411 Among WBC-derived features, band neutrophils and segmented neutrophils were enriched in VTE cases, particularly at 7 days, 30 days, and 6 months. In contrast, early myeloid precursors-including blasts and promyelocytes-showed greater enrichment at the 24-hour time point, suggesting a temporal shift in the dominant white cell morphologies associated with VTE risk.

[0042] At longer time intervals (7 and 30 days), erythroblasts and smudge cells became more prominent among enriched WBC types, indicating a possible role for nucleated red cells and cell debris in delayed VTE events. As shown in FIGs. 4A-4D, these findings suggest that while RBC morphologic abnormalities remain stable predictors across time points, the WBC features associated with VTE risk evolve over time, transitioning from immature precursors at early intervals to more differentiated and fragmented cells at later stages.

[0043] Table 1 below shows the impact of various cell morphologies in predicting VTE risk at time intervals of 7 days, 30 days, and 180 days.C: Systems and Methods for Determining Values Indicating Degrees of Risk for Venous Thromboses (VTEs) In Subjects

[0044] Subjects with cancer may be at a heightened risk of developing venous thromboembolism (VTE) (e.g., deep vein thrombosis (DVT) and pulmonary embolism (PE)), which occur when blood clots form in a deep vein such as in the legs. If not identified early, VTE can result in severe complications, prolonged hospitalization, and risk to morbidity or mortality. Predicting VTE in subjects may be a complex problem due to the multifaceted nature of the condition, as VTE is influenced by a variety of risk factors, including age, obesity, immobility, surgery, and comorbid conditions. When cancer is present in the subject, tumors in the body can release tissue factors and procoagulants further exacerbating the risk of VTE. However, differentcancers have varied associations with the increased risk in VTE. In addition, other conditions, such as age, prior injury, and co-morbidities, can affect the occurrence of VTE in the subject.

[0045] Approaches to predicting VTE may include using these risk factors. For example, Khorana scores may be used to predict likelihood of VTE in a subject, using oncological and physiological factors such as cancer type, platelet count, hemoglobin level, leukocyte count, and body mass index, among others. These factors, however, may be correlated with one another, and may not factor into all considerations that could affect the risk of VTE. As a result, these approaches may suffer from low accuracy and precision in predicting VTE in subjects. Furthermore, scores such as Khorana scores may also not be sensitive or specific enough to be used in clinical decisions with respect to VTE. This may lead to improper administration of VTE- related treatments, and thus to negative clinical outcomes. From a computing perspective, inaccurate predictions may result in wasted computing resources (e.g., processor and memory) and bandwidth consumption that could have been used for more useful processes.

[0046] To address these and other technical challenges, a machine learning (ML) model can be used to determine values indicating risk of VTE in subjects using images of white blood cells (WBCs) along with other multimodal data. The ML model may have been trained using multiple instance learning (MIL) to make the ML model more adept in capturing features that are determinants for risk for VTE. During inference, a feature encoder may take WBC images from peripheral blood smear from the subject as input data. The ML model may generate feature embeddings from the WBC images to capture morphological characteristics. For instance, these feature embeddings may include latent features in images of WBCs capturing the levels of neutrophil extracellular traps (NETs) in the subject. NETs may be implicated in promoting a hypercoagulable state in malignancy, and thus a higher risk of VTE. An aggregator of the ML model can include a self-attention layer that uses the image embeddings, based on the WBCs within the blood sample, to analyze pixels or portions to gather information associated with the image to generate final embeddings for the image. The other multimodal data may also be fed into the aggregator to generate final embeddings, which capture additional latent features that aredeterminants for risk of VTE. The embeddings may be used by the ML model to generate an output value indicating the degree of risk of VTE in the subject.

[0047] In this manner, by using the ML model to analyze WBC morphology, this approach can systematically identify potential markers associated with VTE, potentially leading to more consistent and objective assessments. The ML model can achieve high accuracy and precision in determining the degree of risk for VTE in subjects relative to other approaches. From a clinical perspective, the higher accuracy in VTE prediction may lead to more proper administrations of VTE-related medication, and therefore potentially better clinical outcomes for the subject. From a computing perspective, the use of ML model can reduce computing resources such as processors and memory by reducing the amount of misclassifications and interpretations of the blood samples. Furthermore, the use of the ML architecture can reduce latency in processing images by increasing the amount of proper classifications.

[0048] FIG. 7 depicts a block diagram of a system 100 for determining values indicating degrees of risk for venous thromboses (VTEs) in subjects. In a brief overview, the system 100 can include at least one data processing system 105, at least one imaging device 110, at least one administrative device 115, at least one sequencing device 120, and at least one database 155, among others, communicatively coupled via at least one network 125. The data processing system 105 can include at least one dataset indexer 130, at least one model trainer 135, at least one model applier 140, at least one output evaluator 145, and at least one machine learning (ML) architecture 150, among others. The ML architecture 150 may include at least one feature encoder 165, at least one aggregator 170, and at least one classifier 175, among others. Each of the components of the system 100 can be implemented using the computing system as described in Section D.

[0049] In further detail, the data processing system 105 can be any computing device comprising one or more processors coupled with memory and software capable of performing the various processes and tasks described herein. The data processing system 105 can be housed within a computing system (e.g., laptop, PC, smart device) or within a server group (e g., a data center, a branch office, or a server site), and include instructions to manage the identifying ofimages, generating a classification, and storing an association. The data processing system 105 can be in communication with the imaging device 110, administrative device 115, the sequencing device 120, and the database 155, among others.

[0050] The data processing system 105 can include or execute any number of modules, processes, components, or subcomponents to performing the various processes and tasks described herein. On the data processing system 105, the dataset indexer 130 within the data processing system 105 can receive, retrieve, or otherwise identify a dataset from the imaging device 110 and the administrative device 115. The model trainer 135 can train, establish or otherwise initialize the ML architecture 150. The model applier 140 can apply, execute, or otherwise run the ML architecture 150. The output evaluator 145 can generate, determine, or otherwise identify an output based on a classification.[00511 The ML architecture 150 can be any type of ML algorithm or model to determine a value indicating a probability of risk for a subject developing VTE based on the white blood cells (WBC’s). The ML architecture 150 can be maintained on the data processing system 105. The ML architecture 150 can be, for example, a deep learning artificial neural network (ANN) (e.g., an encoder-decoder model with a convolution neural network architecture, a transformer architecture, a diffusion model, or an encoder-classifier model (e.g., as depicted in FIG. 1)), among others. In some embodiments, the ML architecture 150 can also include, for example, a clustering algorithm (e.g., K-means clustering), a support vector machine (SVM), a Naive Bayesian classifier, a decision tree, or a random forest classifier, among others. In general, the ML architecture 150 can have an image of a WBC in any modality from a subject as an input and a determination of a value indicating a degree of risk for VTE as an output. The ML architecture 150 may have been initialized, trained, and established using training data in accordance with learning techniques (e.g., supervised or semi-supervised). The training data can include or identify images or non-image data set and a label for the subject. In some embodiments, the ML architecture 150 can be trained for a specific subject. In some embodiments, the ML architecture 150 can be trained for a group of subjects and then refined or fine-tuned for a particular subject.

[0052] The ML architecture 150 can include at least one feature encoder 165, at least one aggregator 170, and at least one classifier 175, among others. The feature encoder 165 can execute dimensionality reduction, feature extraction, or sequence modeling to generate, determine, or otherwise create a set of embeddings. The aggregator 170 can generate, determine, or otherwise identify a value indicating a degree of risk for the subject. The classifier 175 can generate, determine, or otherwise identify a classification for the value indicating the degree of risk. In some embodiments, at least a portion of the ML architecture 150 (e.g., the feature encoder 165) may be pre-trained. In some embodiments, the ML architecture 150 may be an instance of the model architecture shown in FIG. 1.]0053] The imaging device 110 can be any device capable of acquiring images of blood samples of subjects. The image can be captured via various techniques such as fluorescence microscopy, phase-contrast microscopy, bright-field microscopy, confocal microscopy, scanning electron microscopy, transmission electron microscopy, quantitative phase imaging, and automated digital microscopy, among others. Although primarily discussed herein in terms of peripheral blood smear (PBS), other imaging modalities besides those listed above may be supported by the data processing system 105. The imaging device 110 can be in communication with the data processing system 105 and the administrative device 115 to provide acquired images.

[0054] The administrative device 115 can be any device comprising one or more processors coupled with memory and software and capable of providing an output projection image. The administrative device 115 can be associated with an entity (e.g., clinician, physician, doctor) examining the subject or biomedical images from the subject. The administrative device 115 can be in communication with the data processing system 105 and the imaging device 110 to exchange data. The administrative device 115 can display images acquired from the imaging device 110 on a display.10055] The sequencing device 120 may be any device to perform genetic sequencing on a gene segment (e.g., deoxyribonucleic acid (DNA) or ribonucleic acid (RNA) sample) in a sample taken from a subject and generate sequencing datasets using the genetic sequencing. The geneticsequencing carried out may be a high throughput, massively parallel sequencing technique (sometimes herein referred to as next-generation sequencing), such as whole genome sequencing (WGS), pyrosequencing, Reversible dye-terminator sequencing, SOLiD sequencing, Ion semiconductor sequencing, and Helioscope single molecule sequencing, among others. Using the genetic sequencing data, the sequencing device 120 may generate genomic dataset profding the subject. In generating the genomic dataset, the sequencing device 120 may execute read alignment, variant calling, and gene expression quantification, among others. In some embodiments, the sequencing device 120 may be part of a genomic profiling platform, such as IMPACT, GENIE, or FoundationOne platform, among others. The sequencing device 120 may use the gene sequencing to generate a sequencing dataset. The sequencing dataset may lack identifiers for at least a portion of genes (e.g., due to differences in sequencing or acquisition protocols across the different genomic profiling platforms). The sequencing dataset may be maintained using one or more files according to a format (e.g., FASTQ, BAM, SAM, BCL, or VCF formats).[0056| The database 155 may store and maintain various resources and data associated with the data processing system 105, the imaging device 110, the administrative device 115, and the sequencing device 120, among others. The database 155 may include a database management system (DBMS) to arrange and organize the data maintained thereon. The database 155 may be in communication with the data processing system 105, the imaging device 110, the administrative device 115, and the sequencing device 120, via the network 125. While running various operations, the data processing system 105, the sequencing device 120, and the administrative device 115 may access the database 155 to retrieve identified data therefrom. The data processing system 105, the imaging device 110, the administrative device 115, and the sequencing device 120 may also write data onto the database 155 from running such operations.

[0057] FIG. 8 depicts a block diagram of a process 200 to train a machine learning (ML) architecture to determine a value indicating a degree of risk of VTE for a subject. Under the process 200, the dataset indexer 130 can receive, obtain, or otherwise extract training data 205 from the database 155. The training data 205 can be used to initialize, train, and establish the MLarchitecture 150. The training data 205 can be for a subject 210 and include an associated blood sample 215 from a vein 212 of the subject. The training data 205 can include a plurality of examples to train the ML architecture 150. Each example of the training data 205 can include one or more of a set off images 220A-N (referred to as images 220 herein) of the blood sample 215, embedding sets 225 A-N (referred to as embedding sets 225 herein), and at least one label 235 for the subject 210, among others. In some embodiments, at least one example may omit or lack the set of images 220.

[0058] The subject 210 (sometimes herein referred to as a sample subject 210 when associated with the training data 205) can be a human or animal subject, among others. The subject 210 can have, can be at risk of, can be afflicted, can be diagnosed with cancer. The cancer can be at least one of carcinoma, sarcoma, hematopoietic cancer, adrenal cancer, bladder cancer, blood cancer, bone cancer, brain cancer, breast cancer, carcinoma, cervical cancer, colon cancer, colorectal cancer, corpus uterine cancer, ear, nose and throat (ENT) cancer, endometrial cancer, esophageal cancer, gastrointestinal cancer, head and neck cancer, Hodgkin's disease, intestinal cancer, kidney cancer, larynx cancer, leukemia, liver cancer, lymph node cancer, lymphoma, lung cancer, melanoma, mesothelioma, myeloma, nasopharynx cancer, neuroblastoma, non- Hodgkin's lymphoma, oral cancer, ovarian cancer, pancreatic cancer, penile cancer, pharynx cancer, prostate cancer, rectal cancer, sarcoma, seminoma, skin cancer, stomach cancer, teratoma, testicular cancer, thyroid cancer, uterine cancer, vaginal cancer, vascular tumor, or metastases thereof, among others. In some embodiments, the subject 210 may be at risk of, can be afflicted, can be diagnosed with conditions associated with VTE, such as age (e.g., above age of 50 or 60), prior injury (e.g., injury to leg), and other co-morbidities (e.g., obesity, cardiovascular disease, diabetes, inflammatory disease, or autoimmune disease), among others. In some embodiments, the subject 210 may lack cancer or any condition.

[0059] The vein 212 of the subject 210 can be any functioning vein or artery that carries blood with the body of the subject 210. The vein 212 may be at risk of or may already affected by VTE. The vein 212 can include an arm vein, a pelvic vein, a lobar artery, an aorta, superior vena cava, inferior vena cava, pulmonary artery, pulmonary vein, brachial artery, radial artery,ulnar artery, femoral artery, popliteal artery, anterior tibial artery, posterior tibial artery, dorsalis pedis artery, external iliac artery, internal iliac artery, common carotid artery, internal carotid artery, external carotid artery, subclavian artery, axillary artery, renal artery, celiac trunk, superior mesenteric artery, inferior mesenteric artery, left gastric artery, splenic artery, hepatic artery, cephalic vein, basilic vein, median cubital vein, great saphenous vein, small saphenous vein, external jugular vein, internal jugular vein, subclavian vein, axillary vein, femoral vein, popliteal vein, common iliac vein, right segmental artery, or superior lobar artery, among others.

[0060] A physician can extract, obtain, or otherwise retrieve a blood sample 215 from the subject 210. The blood sample 215 may be acquired in accordance with a peripheral blood smear (PBS). The blood sample 215 can be extracted from peripheral veins, a central venous catheter, bone marrow, and arteries of the subject 210 for analysis. Once extracted, the physician can use or control the imaging device 110 to capture the images 220 of the blood sample 215. The blood sample 215 can include various compositions of the blood, such as, circulating tumor cells, leukemic cells, prostate-specific antigens, CA-125, Carcinoembryonic Antigen, Alphafetoprotein, CA 19-9, altered blood cell counts (e.g., low red blood cells, low white blood cells), and C-reactive Proteins, among other compositional changes (e.g., due to cancer or other medical conditions). The blood sample 215 can include red blood cells (RBCs),WBCs, plasma, platelets, enzymes, minerals, glucose, lipids, serum, or blood proteins, among others.100611 The set of images 220 can derived from at least one blood sample 215 from the subject 210. The images 220 can be acquired via the imaging device 110. Each image captured by the imaging device 110 can be stored within the training data 205 of the database 155. The images 220 can be acquired in accordance with any number of imaging techniques, such as fluorescence microscopy, phase-contrast microscopy, bright-field microscopy, confocal microscopy, scanning electron microscopy, transmission electron microscopy, quantitative phase imaging, and automated digital microscopy, among others. The set of images 220 may include at least one of WBC images or RBC images, among others. Each image 220 may correspond to at least one blood cell (e.g., WBC or RBC) in the blood sample 215. Each image 220 may correspondto at least one WBC from the blood sample 215. Each image 220 may correspond to at least one RBC from the blood sample 215.(0062] Each embedding set 225 can include a plurality of embeddings that correspond to at least one image 220 of the set of images 220. The embedding set 225 can map to the WBC and the RBC within the images. For instance, a first embedding set 225 can map WBC images and a second embedding set 225 can map to RBC images within the images 220. The embedding set 225 can be numerical representation of the images 220 that were previously generated by the feature encoder 165 of the ML architecture 150 for the subject 210. The feature encoder 165 can generate the embeddings by processing the images 220 by extracting features (e.g., edges, colors, shapes, objects), performing dimensionality reduction to condense the features into the embedding. In some embodiments, the embedding set 225 may identify a set of morphological features of the respective blood cell of the respective image of the images 220.(0063] The dataset 230 can be a non-image dataset. In some embodiments, the non-image dataset can include an electronic document (sometimes herein referred to as a report). The electronic document may identify, define, or otherwise include characteristics, information, and traits about the subject 210. The traits of the subject 210 can be provided by the subject 210, extracted from an electronic medical record, or documented by a physician. The traits can include symptoms (e.g., fatigue, weakness, bleedings), medical history, family medical history, lab results, physical traits, psychological traits (e.g., personality, coping mechanism, emotional state), behavioral traits (e.g., adherence to treatment, health seeking behavior, communication style), or social traits, among other traits. In some embodiments, the non-image dataset can be the sequencing dataset obtained from the sequencing device 120. The sequencing dataset can include the gene profile associated with the subject 210. The gene profile can define a plurality of characteristics of genes in the blood sample, or a DNA sample of the subject 210. The gene profile can further include a sequencing record that include a plurality of sequences associated with the genes of the subject 210.

[0064] The label 235 can correspond to the classification of the subject 210 based on a value from the aggregator 170. The label 235 can identify or indicate the presence (or occurrence or detection) or the absence (or lack of occurrence or detection) of VTE within a time interval relative to the time of the extraction of the blood sample 215. The time interval can be any range between 6 hours to 8 months. The time interval may be, for example, at least one of 1 day, 5 days, 7 days, 10 days, 14 days, 21 days, 28 days, 1 month, 2 months, 3 months, 4 months, 5 months, 6 months, 7 months, or 8 months, among other time intervals. In another example, the label 235 can indicate the presence of a deep vein thromboses (DVT) or a pulmonary embolism (PE) within the subject at the time interval.

[0065] In some embodiments, with a plurality of subjects 210, the dataset indexer 130 can index, sort, or otherwise organize each example within then training data 205 to correspond to each subject 210 in the plurality of subjects 210. The dataset indexer 130 may process or parse each example in the training data 205 to identify or extract the images 220, the embeddings sets 225, the non-image dataset 230, and the label 235, among others, for the corresponding sample subject 210. The dataset indexer 130 may partition, section, or otherwise divide a single image 220 into a set of patches forming the images 220. The set of patches can correspond to a matrix or a data table such that each element correspond to each patch. Each patch may correspond to a respective portion of the image 220. For example, the training data 205 may originally include a single image 220, from which the dataset indexer 130 divides to form the plurality of patches for analysis.

[0066] With the retrieval of the training data 205, the dataset indexer 130 can select, identify, or otherwise determine WBC images 220’ from the images 220. The selection can be based on the visual characteristics of the image 220 or by preprocessing each image 220. In some instances, the images 220 can include a label to identify the images 220 as RBC images and WBC images. For example, the label can be a binary value assigned to the images 220. The binary value can be a 0 or a 1, such that a 0 indicates a WBC image and a 1 indicates an RBC image. The dataset indexer 130 can preprocess the images 220 to remove noise, distortions, artifacts, and stains, among other blemishes associated with the image 220 to perform or execute contrastenhancement between the cells of the blood sample 215 and the background of the image 220. In some embodiments, the dataset indexer 130 can execute segmentation or contouring to identify at least one portion, region, or object of the image 220 to separate the WBCs from the background and non-similar cells (e.g., a WBC from a group of RBCs). The dataset indexer 130 can use thresholding, algorithms (e.g., watershed algorithm), or edge detection, among other methods to execute the segmentation. The dataset indexer 130 can select, identify, or otherwise obtain RBC images 220” from the images 220 in a similar manner to the WBC images 220’. In some cases, the dataset indexer 130 can extract, identify, or otherwise determine features (e.g., visual characteristics) from each cell within the image 220 to distinguish the WBCs from the RBCs. The features can include shape, size, color, intensity, nucleus to cytoplasm ratio, or granularity, among other distinguishing features.

[0067] The model trainer 135 can provide, feed, or otherwise input the WBC images 220’ to the feature encoder 165 of the ML architecture 150 for each example of the training data 205. The feature encoder 165 can ingest, receive, or otherwise obtain the WBC images 220’ from the model trainer 135. The ML architecture 150 can include a plurality of weights and hyperparameters to determine values indicating the degree of risk of VTE. The model trainer 135 can fine-tune, update, or otherwise modify the plurality of weights associated with the ML architecture 150. By modifying the plurality of weights, the model trainer 135 can improve the aggregator 170 to determine more accurate values indicating degrees of risk for VTE. The plurality of hyperparameters can establish the architecture, design, or configuration of the ML architecture 150 to determine values indicating a degree of risk for VTE and a classification based on the value.(0068] Within the ML architecture 150, the plurality of hyperparameters can be arranged across the feature encoder 165, the aggregator 170, and the classifier 175. The plurality of hyperparameters can include a learning rate, number of epochs, a batch size, a model architecture (e.g., deep learning convolutional neural network, ensemble network, or clustering algorithm, or any combination thereof), and regularization parameters. In some implementations, the plurality of parameters can include a plurality of model parameters. The plurality of model parameters can be variables learned from the training data 205 (e.g., one ormore examples) that establish, dictate, or otherwise define a link between the input (e.g., WBC images 220’) and the output (e.g., classification 245). The model parameters can continuously optimize during training to minimize a loss function of the ML architecture 150. The plural of model parameters can include target weights or biases (e.g., influence the output of the ML architecture 150 based on accurate targets), or loss function (e.g., compound loss), among others. This is not limited to the feature encoder 165, but the hyperparameters of the ML architecture 150 can be arranged and adjusted for the aggregator 170 and the classifier 175 in a similar manner.

[0069] The feature encoder 165 can generate, identify, or otherwise determine a plurality of embedding sets 225’A-N (referred to as embedding sets 225’ herein). The plurality of embeddings sets 225’ can be similar to the embedding sets 225. The feature encoder can ingest the WBC images 220’ and can execute a dimensionality reduction, feature extraction, or sequence modeling. In some implementations, the feature encoder 165 can include a plurality of layers to extract, retrieve, or otherwise obtain the WBC images 220’ from the model trainer 135 and the model trainer 135. The plurality of layers can include one or more hidden layers. The one or more hidden layers can transform the WBC images 220’ to extract or retrieve one or more representations of the WBC images 220’. The one or more representations can include the set of embedding sets 225’.

[0070] In some embodiments, the model trainer 135 can provide or input the plurality of embeddings sets 225 to the aggregator 170. Each embedding set of the plurality of embedding sets identifying a set of morphological features of the respective WBC of the respective image of the WBC images 220’. This may only occur during training of the ML architecture 150 to bias a loss metric toward the aggregator 170. In this manner, the model trainer 135 can focus training on the aggregator 170 if the value indicating degrees of risk is inaccurate during execution of the ML architecture 150. The administrative device 115 can provide an indication to the model trainer 135 that the degree of risk is inaccurate. Responsive to the indication, the model trainer 135 can provide the aggregator 170 with the embedding sets 225 and rather than embedding sets 225’ generated by the feature encoder 165.[00711 The feature encoder 165 can provide the plurality of embedding sets 225’ to the aggregator 170 of the ML architecture 150. The feature encoder 165 can generate a plurality of attention scores for the set of morphological features. The set of morphological features can refer to aspects of the WBC or RBS external form or structure. The set of morphological features can include size, shape, nucleus to cytoplasm ration, nuclear shape, chromatic pattern, nucleoli, cytoplasm appearance, granules, border, structures, and inclusions among other features of the cells. The attention score can be a numerical value that indicates a level of focus or importance for the aggregator 170 when determining the value indicating a degree of risk. The feature encoder 165 can generate the attention score by using linear layers to transform the embeddings sets 225’ into one or more categories (e.g., query, key, and value). The feature encoder 165 can calculate a score between the query and the key of each input position by executing a dot product of the key and the query and applying a SoftMax function thereby generating the attention score. Each attention score of the plurality of attention scores can indicate a degree of weight of a respective morphological feature of the set of morphological features in determining the value.[0072| Upon generation of the embedding sets 225’ and attention scores, the aggregator 170 can receive or obtain the embedding sets 225’ from the feature encoder 165. The aggregator 170 can determine, identify, or otherwise indicate a value 240 indicating a degree of risk of VTE. The value 240 can be a vector of numbers that include the same or a similar dimension as the embedding sets 225’. The value 240 can indicate the degree of risk of VTE for the subject 210 within the time interval. The degree of risk of VTE can indicate a likelihood or a threshold that the subject 210 can contract or be afflicted with VTE in at least one vein 212 within the subject 210. To determine the value 240, the aggregator 170 can take the average of the embedding sets 225’ or execute neural attention, pooling, bagging, boosting, stacking, majority voting, or weighted averaging (e.g., via attention scores), among other methods to determine the value 240. In some instances, the aggregator 170 can establish a plurality of predictions of values 240 for the subject. By averaging or combining the predictions, the aggregator 170 can determine the value 240.

[0073] The aggregator 170 can feed, input, or otherwise provide the value 240 to the classifier 175. Concurrently, the model trainer 135 can feed, input or otherwise provide the dataset230 to the classifier 175. The classifier 175 can receive, retrieve, or otherwise obtain the value 240 from the aggregator 170. Concurrently, the classifier 175 can receive, retrieve, or otherwise obtain the dataset 230 from the model trainer 135. The classifier 175 can receive the value 240 and the dataset 230 as an input and generate at least one classification 245 as an output. The classier 175 can include one or more input layers that receives the value 240 indicating the degree of risk and the dataset 230 to provide the value 240 and the dataset 230 to the subsequent layers of the classifier 175. In some embodiments, the classifier 175 can execute, for example, linear transformation, non-linear activation, Softmax functions, and sigmoid functions, among other functions. In some embodiments, the classifier 175 can include one or more hidden layers (e.g., fully connected layers, convolutional layers, or recurrent layers) to learn patterns and representations of the value 240 with the dataset 230 based on an activation function (e.g., ReLU, Sigmoid, or Tanh). For instance, the one or more hidden layers can form a pattern based on the value 240 and a trait of the subject 210 within the dataset 230.

[0074] Using the value 240 and the dataset 230, the classifier 175 can generate, determine, or otherwise identify the classification 245 for the subject. In some instances, the output layer of the classifier 175 can generate the classification 245. The classification 245 can indicate a risk level of a plurality of risk levels for the subject 210 in the given time interval. The plurality of risk levels can include low risk (e.g., < 1% chance of VTE), moderate risk (e.g., <3% chance of VTE), high risk (e.g., <10% change of VTE), or very high risk >10% chance of VTE). The risk level for the classification 245 can be determined in accordance with the value and the ranges of the each of the plurality of risk levels. In some instances, the classifier 175 can generate secondary or intermediary value to compare against the value 240. The intermediary value can indicate the degree of risk for the subject 210 based on the value 240 and the dataset 230. The classifier 175 may compare the intermediary value to a threshold based on the plurality of risk levels. If the value satisfies (e.g., greater than or equal to, or within the range) at least one threshold, the classifier 175 may generate the classification 245 to indicate that the subject 210 is a first risk level for at least the given time interval. Based on the risk level, the classification 245 can identify that the subject is at risk for contracted VTE.

[0075] The model trainer 135 can generate, determine, or otherwise identify a loss metric 250 to use to update the ML architecture 150. In some implementations, the model trainer 135 may generate the loss metric 250 be based on the value 240. In some implementations, the classifier 175 can determine, calculate, or otherwise generate the loss metric 250. The loss metric 250 can measure, rate, or identify how well the ML architecture 150 can determine accurate classifications 245 for the subjects 210. The model trainer 135 can measure a discrepancy between the classification 245 and the label 235. In further detail, when classifier 175 is presented with the value 240 indicating a high degree of risk, the classifier 175 can learn to generate a classification 245 indicating a risk level for the subject 210 and can calculate the loss metric 250. Conversely, when classifier 175 is presented with the value 240 indicating a low degree of risk, the classifier 175 can learn to generate a classification 245 indicating a risk level for the subject 210 and can calculate the loss metric 250.

[0076] In some embodiments, the loss metric 250 may correspond to a loss function to quantify the difference between the classification 245 and the label 235 within the training data 205 to optimize ML architecture 150 during training. The loss function can be at least one of mean squared error, binary cross-entropy loss, categorical cross-entropy loss, hinge loss, or Wasserstein loss, among others. Using the loss function for the aggregator 170 and the classifier 175, the ML architecture 150 can avoid any issues of mode collapse and vanishing gradients to allow the model trainer 135 to provide more stable training for the ML architecture 150. Therefore, the system 100 can effectively adapt to each subject 210 with relative ease, low computer resources, and stable training for the ML architecture 150.

[0077] The model trainer 135 can compare the classification 245 with the label 235 of the training data 205 to generate the loss metric 250 according to a delta from a threshold. The threshold can indicate an optimal value for the comparison. When the comparison is at the threshold, the model trainer 135 may not generate a loss metric 250 for the ML architecture 150. When the comparison deviates from the threshold (e.g., above or below), the delta can be used to generate the loss metric 250. Therefore, a high delta (e.g., large deviation from threshold) can cause the model trainer 135 to generate a high loss metric 250. Conversely, a low delta (e.g.,small deviation from the threshold) can cause the model trainer 135 to generate a low loss metric 250.

[0078] The model trainer 135 can update, further train, or otherwise tune the at least one of a plurality of weights in the ML architecture 150 using the loss metric 250. In some embodiments, the feature encoder 165 may have been pretrained to generate proper embedding sets 225, and the model trainer 135 may use the loss metric 250 to update the weights in the aggregator 170 or the classifier 175, or both. In some embodiments, the model trainer 135 may use the loss metric 250 to update the ML architecture 150 from end-to-end (e.g., including the feature encoder 165, the aggregator 170, and the classifier 175). In some embodiments, the model trainer 135 can adjust the plurality of weights so the ML architecture 150 can generate more accurate classifications 245. In some embodiments, the model trainer 135 can adjust the plurality of weights so the ML architecture 150 can generate more accurate values 240. For example, the model trainer 135 can use the loss metric 250 to compute gradients for the direction and magnitude of parameter adjustments to minimize the loss metric 250. In another instance, the model trainer 135 can use optimization algorithms such as, stochastic gradient descent or RMS prop to update model parameters iteratively. In some implementations, the direction and step size of the parameter update according to the gradients and the optimization algorithm.

[0079] FIG. 9 depicts a block diagram of a process 300 to determine values indicating the degree of risk of VTE for a subject based on a blood sample of the subject. Under the process 300, the imaging device 110 can capture or identify a plurality of images 320A-N (sometimes referred to as images 320 herein) derived from at least one blood sample 315 from the subject 310. The imaging device 110 can be at least one of a light microscope (brightfield, phase contrast), digital cell systems (e.g., CellaVision), or flow cytometers, among others. Each image captured by the imaging device 110 can be provided to the dataset indexer 130. The images 320 can be acquired in accordance with any number of imaging techniques, such as fluorescence microscopy, phasecontrast microscopy, bright-field microscopy, confocal microscopy, scanning electron microscopy, transmission electron microscopy, quantitative phase imaging, and automated digital microscopy, among others.

[0080] The administrative device 115 associated with a physician can generate, determine, or otherwise create at least one dataset 330. The dataset 330 can include an electronic document (sometimes herein referred to as a report). The electronic document that includes characteristics, information, and traits about the subject 310. The traits of the subject 310 can be provided by the subject 310, extracted from an electronic medical record, or documented by the physician. The traits can include symptoms (e.g., fatigue, weakness, bleedings), medical history, family medical history, lab results, physical traits, psychological traits (e.g., personality, coping mechanism, emotional state), behavioral traits (e.g., adherence to treatment, health seeking behavior, communication style), and social traits, among other traits. The administrative device 115 can continuously update the dataset 330 for a respective subject 310 based on future visits with the physician.

[0081] The sequencing device 120 can perform DNA extraction, sequencing, and Bioinformatic analysis for the respective subject. The output of the Bioinformatic analysis can be presented or provided to the administrative device 115. In some instances, the sequencing device 120 can generate a dataset 330 that includes gene data or a gene profile. The sequencing device 120 can provide the gene data to the administrative device 115 to append to the dataset 330 generated by the administrative device 115. The dataset 330 can be a sequencing dataset obtained from the sequencing device 120. The sequencing dataset can include the gene profile associated with the subject 310. The gene profile can define a plurality of characteristics of genes in the blood sample 215, a DNA sample of the subject 310, expression level of genes, presence of mutations or variants, among others. The gene profile can further include a sequencing record that include a plurality of sequences associated with the genes of the subject 310.

[0082] The subject 310 can be a human or animal subject, among others. The subject 310 can have, can be at risk of, can be afflicted, can be diagnosed with cancer. The cancer can be at least one of carcinoma, sarcoma, hematopoietic cancer, adrenal cancer, bladder cancer, blood cancer, bone cancer, brain cancer, breast cancer, carcinoma, cervical cancer, colon cancer, colorectal cancer, corpus uterine cancer, ear, nose and throat (ENT) cancer, endometrial cancer, esophageal cancer, gastrointestinal cancer, head and neck cancer, Hodgkin's disease, intestinalcancer, kidney cancer, larynx cancer, leukemia, liver cancer, lymph node cancer, lymphoma, lung cancer, melanoma, mesothelioma, myeloma, nasopharynx cancer, neuroblastoma, non- Hodgkin's lymphoma, oral cancer, ovarian cancer, pancreatic cancer, penile cancer, pharynx cancer, prostate cancer, rectal cancer, sarcoma, seminoma, skin cancer, stomach cancer, teratoma, testicular cancer, thyroid cancer, uterine cancer, vaginal cancer, vascular tumor, or metastases thereof, among others. In some embodiments, the subject 310 may be at risk of, can be afflicted, can be diagnosed with conditions associated with VTE, such as age (e.g., above age of 50 or 60), prior injury (e.g., injury to leg), and other co-morbidities (e.g., obesity, cardiovascular disease, diabetes, inflammatory disease, or autoimmune disease), among others. In some embodiments, the subject 310 may lack cancer or any condition.

[0083] The vein 312 of the subject 310 can be any functioning vein or artery that carries blood with the body of the subject 310. The vein 312 may be at risk of or impacted by VTE. The vein 312 can include one or more of: an aorta, superior vena cava, inferior vena cava, pulmonary artery, pulmonary vein, brachial artery, radial artery, ulnar artery, femoral artery, popliteal artery, anterior tibial artery, posterior tibial artery, dorsalis pedis artery, external iliac artery, internal iliac artery, common carotid artery, internal carotid artery, external carotid artery, subclavian artery, axillary artery, renal artery, celiac trunk, superior mesenteric artery, inferior mesenteric artery, left gastric artery, splenic artery, hepatic artery, cephalic vein, basilic vein, median cubital vein, great saphenous vein, small saphenous vein, external jugular vein, internal jugular vein, subclavian vein, axillary vein, femoral vein, popliteal vein, common iliac vein, right segmental artery, or superior lobar artery, among others.

[0084] A physician can extract, obtain, or otherwise retrieve a blood sample 315 from the subject 310. The blood sample 315 may be acquired in accordance with a peripheral blood smear (PBS). The blood sample 315 can be extracted from peripheral veins, a central venous catheter, bone marrow, and arteries of the subject 310 for analysis. Once extracted, the physician can use or control the imaging device 110 to capture the images 320 of the blood sample 315. The blood sample 315 can include various compositions of the blood, such as, circulating tumor cells, leukemic cells, prostate-specific antigens, CA-125, Carcinoembryonic Antigen, Alpha-fetoprotein, CA 19-9, altered blood cell counts (e.g., low red blood cells, low white blood cells) and C-reactive Proteins, among other compositional changes (e.g., due to cancer or other medical conditions). The blood sample 315 can include red blood cells (RBC), WBC, plasma, platelets, enzymes, minerals, glucose, lipids, serum, and blood proteins, among others.|0085| The dataset indexer 130 can receive, obtain, or otherwise extract a plurality of images 320 from the imaging device 110 based on a blood sample 215 of a subject 310. In conjunction, the administrative device 115 can provide the dataset 330 to the dataset indexer 130. The dataset indexer 130 can receive the plurality of images 320 and the dataset 330 via the network 125. The set of images 320 may include at least one of WBC images or RBC images, among others. Each image 320 may correspond to at least one blood cell (e.g., WBC or RBC) in the blood sample 315. Each image 320 may correspond to at least one WBC from the blood sample 315. Each image 320 may correspond to at least one RBC from the blood sample 315. Upon reception of the dataset 330 and the images 320, the dataset indexer 130 can store or house the dataset 330 and the images 320 at a location associated with the subject 310 within the database 155. In some embodiments, The dataset indexer 130 may partition, section, or otherwise divide a single image 320 into a set of patches forming the images 320. The set of patches can correspond to a matrix or a data table such that each element correspond to each patch. Each patch may correspond to a respective portion of the image 320. In some embodiments, the dataset indexer 130 can provide the images 320 and the dataset 330 to the model applier 140. The model applier 140 can select, identify, or otherwise indicate WBC images 320’ from the images 320. The selection can be based on the visual characteristics of the image 320 or by preprocessing each image 320. In some instances, the images 320 can include a label to identify the images 320 as RBC images and WBC images. For example, the label can be a binary value assigned to the images 320. The binary value can be a 0 or a 1, such that a 0 indicates a WBC image and a 1 indicates an RBC image.

[0086] The dataset indexer 130 can preprocess the images 320 to remove noise, distortions, artifacts, and stains, among other blemishes associated with the image 320 to perform or execute contrast enhancement between the cells of the blood sample 315 and the background of the image 320. In some embodiments, the dataset indexer 130 can execute segmentation or contouring toidentify at least one portion, region, or object of the image 320 to separate the WBCs from the background and non-similar cells (e.g., a WBC from a group of RBCs). The dataset indexer 130 can use thresholding, algorithms (e.g., watershed algorithm), or edge detection, among other methods to execute the segmentation. The dataset indexer 130 can select, identify, or otherwise indicate RBC images 320” from the images 320 in a similar manner to the WBC images 320’. In some cases, the dataset indexer 130 can extract, identify, or otherwise determine features (e.g., visual characteristics) from each cell within the image 320 to distinguish the WBCs from the RBCs. The features can include shape, size, color, intensity, nucleus to cytoplasm ratio, and granularity, among other distinguishing features.

[0087] The model applier 140 can provide, feed, or otherwise input the WBC images 320’ to the feature encoder 165 of the ML architecture 150. The feature encoder 165 can ingest, receive, or otherwise obtain the WBC images 320’ from the model applier 140. The ML architecture 150 can include a plurality of weights and hyperparameters to determine values indicating the degree of risk of VTE. The plurality of hyperparameters can establish the architecture, design, or configuration of the ML architecture 150 to determine values indicating a degree of risk for VTE and a classification based on the value.

[0088] Within the ML architecture 150, the plurality of hyperparameters can be arranged across the feature encoder 165, the aggregator 170, and the classifier 175. The plurality of hyperparameters can include a learning rate, number of epochs, a batch size, a model architecture (e.g., deep learning convolutional neural network, ensemble network, or clustering algorithm, or any combination thereof), and regularization parameters. In some implementations, the plurality of parameters can include a plurality of model parameters. The plurality of model parameters can be variables learned from the training data 205 (e.g., one or more examples) that establish, dictate, or otherwise define a link between the input (e.g., WBC images 320’) and the output (e.g., classification 345). In some embodiments, the model parameters can continuously optimize during training to minimize a loss function of the ML architecture 150. The plurality of model parameters can include target weights or biases (e.g., influence the output of the ML architecture 150 based on accurate targets), or loss function (e.g.,compound loss), among others. This is not limited to the feature encoder 165, but the hyperparameters of the ML architecture 150 can be arranged and adjusted for the aggregator 170 and the classifier 175 in a similar manner.

[0089] The feature encoder 165 can generate, identify, or otherwise determine a plurality of embedding sets 325A-N (referred to as embedding sets 325 herein). Each embedding set 325 may correspond to at least one of the WBC images 320’ or the RBC images 320”. The feature encoder 165 can ingest the WBC images 320’ (or the RBC images 320”) and can execute dimensionality reduction, feature extraction, or sequence modeling. In some embodiments, the feature encoder 165 can include a plurality of layers to extract, retrieve, or otherwise obtain the WBC images 320’ from the model applier 140 and the model applier 140. The plurality of layers can include one or more hidden layers. The one or more hidden layers can transform the WBC images 320’ to extract or retrieve one or more representations of the WBC images 320’. The one or more representations can include the set of embeddings 325. In some embodiments, the embedding set 225 may define or identify a set of morphological features of the respective blood cell of the respective image of the images 220.

[0090] The feature encoder 165 can provide the plurality of embedding sets 325 to the aggregator 170 of the ML architecture 150. The feature encoder 165 can generate a plurality of attention scores for the set of morphological features. The set of morphological features can refer to aspects of the WBC or RBC external form or structure. The set of morphological features can include size, shape, nucleus to cytoplasm ration, nuclear shape, chromatic pattern, nucleoli, cytoplasm appearance, granules, border, structures, inclusions, heme malignancies, smudge cells, eosinophils, and artifacts among other features of the cells. The attention score can be a numerical value that indicates a level of focus or importance for the aggregator 170 when determining the value indicating a degree of risk. The feature encoder 165 can generate the attention score by using linear layers to transform the embeddings sets 325 into one or more categories (e.g., query, key, and value). The feature encoder 165 can calculate a score between the query and the key of each input position by executing a dot product of the key and the query and applying a SoftMax function thereby generating the attention score. Each attention score of the plurality of attentionscores can indicate a degree of weight of a respective morphological feature of the set of morphological features in determining the value.

[0091] Upon generation of the embedding sets 325 and attention scores, the aggregator 170 can receive or obtain the embedding sets 325 from the feature encoder 165. The aggregator 170 can determine, identify, or otherwise indicate a value 340 indicating a degree of risk of VTE. The value 340 can be a vector of numbers that include the same or a similar dimension as the embedding sets 325. The value 340 can indicate the degree of risk of VTE for the subject 310 within the time interval. The degree of risk of VTE can indicate a likelihood or a threshold that the subject 310 can contract or be afflicted with VTE in at least one vein 312 within the subject 310. To determine the value 340, the aggregator 170 can take the average of the embedding sets 325 or execute neural attention, pooling, bagging, boosting, stacking, majority voting, and weighted averaging (e.g., via attention scores), among other methods to determine the value 340. In some instances, the aggregator 170 can establish a plurality of predictions of values 340 for the subject. By averaging or combining the predictions, the aggregator 170 can determine the value 340.

[0092] The aggregator 170 can feed, input, or otherwise provide the value 340 to the classifier 175. Concurrently, the model applier 140 can feed, input or otherwise provide the dataset 330 to the classifier 175. The classifier 175 can receive, retrieve, or otherwise obtain the value 340 from the aggregator 170. Concurrently, the classifier 175 can receive, retrieve, or otherwise obtain the dataset 330 from the model applier 140. The classifier 175 can receive the value 340 and the dataset 330 as an input and generate the classification 345 as an output. The classier 175 can include one or more input layers that receives the value 340 indicating the degree of risk and the dataset 330 to provide the value 340 and the dataset 330 to the subsequent layers of the classifier 175. In some embodiments, the classifier 175 can execute, for example, linear transformation, non-linear activation, Softmax functions, and sigmoid functions, among other functions. In some embodiments, the classifier 175 can include one or more hidden layers (e.g., fully connected layers, convolutional layers, recurrent layers) to learn patterns and representations of the value 340 with the dataset 330 based on an activation function (e.g., ReLU, Sigmoid, Tanh).For instance, the one or more hidden layers can form a pattern based on the value 340 and a trait of the subject 310 within the dataset 330.

[0093] Using the value 340 and the dataset 330, the classifier 175 can generate, determine, or otherwise identify a classification 345 for the subject. In some instances, the output layer of the classifier 175 can generate the classification 345. The classification 345 can indicate a risk level of a plurality of risk levels for the subject 310 in the given time interval. The time interval can be any range between 6 hours to 8 months. The time interval may be, for example, at least one of 1 day, 5 days, 7 days, 10 days, 14 days, 21 days, 28 days, 1 month, 2 months, 3 months, 4 months, 5 months, 6 months, 7 months, or 8 months, among other time intervals. The plurality of risk levels can include low risk (e.g., < 1% chance of VTE), moderate risk (e.g., <3% chance of VTE), high risk (e.g., <10% change of VTE), or very high risk >10% chance of VTE). The risk level for the classification 345 can be determined in accordance with the value and the ranges of the each of the plurality of risk levels. In some instances, the classifier 175 can generate secondary or intermediary value to compare against the value 340. The intermediary value can indicate the degree of risk for the subject 310 based on the value 340 and the dataset 330. The classifier 175 may compare the intermediary value to a threshold based on the plurality of risk levels. If the value satisfies (e.g., greater than or equal to, or within the range) at least one threshold, the classifier 175 may generate the classification 345 to indicate that the subject 310 is a first risk level for at least the given time interval. Based on the risk level, the classification 345 can identify the presence or absence of VTE at the time interval.

[0094] FIG. 10 depicts a block diagram of a process 400 to provide an output identifying the subject as a candidate or non-candidate for administration of medication. Under the process 400, the output evaluator 145 can receive, retrieve, or otherwise obtain the classification 345 from the classifier 175. Concurrently, the output evaluator 145 can receive, retrieve, or otherwise obtain the value 340 from the aggregator 170. Upon receipt, the output evaluator 145 can store, house, or otherwise save an association between the subject 305 and the value 340. The output evaluator 145 can store, house, or otherwise maintain an association between the subject 310 and theclassification 345. The association can be a link, a map, or a connection between the subject 310, the classification 345, and the value 340.

[0095] To store the association, the output evaluator 145 can generate one or more data structures. The one or more data structures can include an array, a linked list, a stack, a tree, and a hash table, among others. For example, the data structure can be a hash table where the subject 310 is the key to the hash table and the classification 345 is the value of the hash table. The hash table can include a plurality of keys (e.g., for each subject 310) mapped to a plurality of values (e.g., classification 345 of the subject 310). In another example, the data structure can be a plurality of linked lists. A first linked list can correspond to a first subject 310. Each node in the first linked list corresponds to the value 340 of the first subject 310. Concurrently, a second linked list can correspond to a second subject 310. Each node in the second linked list corresponds to the value 340 of the second subject 310. In some implementations, the output evaluator 145 can store an association among the images 320, WBC images 320’, RBC images 320”, the classification 345, the value 340, and the attention scores, among others, on the database 155.10096] In some embodiments, the output evaluator 145 may identify or determine whether the subject 310 is a candidate for medication 410 based on the value 340. The medication 410 can include at least one of a direct oral anticoagulant (DOAC), a low molecular weight heparin (LMWH), an unfractionated heparin (UFH), a vitamin K antagonist, or a fondaparinux. To identify, the output evaluator 145 can measure or compare the value 340 against a threshold. The threshold in this context can be a minimum numerical percentage that corresponds to a respective risk level (e.g., low risk, medium risk, high risk). In some instances, the output evaluator 145 can compare the intermediary value against the threshold. When the value 340 satisfies (e.g., greater than or equal to) the threshold, the output evaluator 145 can identify the subject 310 as the candidate for administration of the medication 410 for VTE. When the value 340 does not satisfy (e g., less than) the threshold, the output evaluator 145 can identify the subject 310 as a noncandidate for administration of the medication 410. For example, the value 240 can indicate that the degree of risk is less than 1% for VTE. The subject 310 can include a classification 345 of low risk, thereby the value 340 may not satisfy the threshold.

[0997] In some embodiments, the output evaluator 145 can identify or determine whether the subject 310 is a candidate for medication 410 based on the classification 345. The output evaluator 145 can identify the subject 310 is a candidate for medication 410 based on the risk level associated with the classification 345. For example, the ML architecture 150 can generate a classification 345 for a subject 310 afflicted with pancreatic cancer. The classification 345 can correspond to very high risk (e.g., >10% risk) for VTE in the subject 210. Based on the indication of very high risk, the output evaluator 145 can determine that the subject 310 is a candidate for the medication 410. Furthermore, the output evaluator 145 can direct or indicate that the subject 310 be administered with the medication 410.

[0098] The output evaluator 145 can generate, determine, or otherwise identify at least one output 405 based on at least one of the value 340 or the classification 345. The output 405 can include, for example, the classification 345 for presentation to the physician, subject 310, or lab technician, among other clinicians, or the value 340 for presentation to the same. When the subject 310 is identified as the candidate, the output evaluator 145 can generate the output 405 to indicate that the subject is to be administered with the medication 410. In some embodiments, the output 405 can be a notification to administer medication within the time interval or a notification for the subject 310 to request medical attention. When the subject 310 is identified as a non-candidate, the output evaluator 145 can generate the output 405 identifying the subject as a non-candidate for administration of a medication for VTE.

[0099] The output evaluator 145 can send, transmit, or otherwise or provide the output 405 for presentation at the user interface 415 of the administrative device 115. The output 405 can be provided as a system call (e.g., to cause the administrative device 115 to render a message), a notification, an alert, a text message, and electronic mail (e-mail), among other messages in a human-readable format. In some cases, the output 405 can include instructions to render an application for presentation on the user interface 415. The application can include the value 340, the classification 345, the dataset 330, and the images 320, among other information associated with the subject 310. The message or notification within the output 405 can indicate that the subject 310 is a candidate or a non-candidate for the medication 410.

[0100] Upon receipt, the administration device 115 can render, display, or present the output 405 on the user interface 415. The administrative device 115 can present the user interface 415 based on the instructions within the output 405. The instruction within the output 405 can cause the user interface 415 to display the notification or message associated with the output 405. The information presented via the user interface 415 may include, for example, at least one of the images 320, the WBC images 320’, the RBC images 320”, the dataset 330, the value 340, the classification 345, or the attention scores, among others. When the output 405 identified the subject as a candidate for the administration of the medication, the subject 310 can be provided or administered (e.g., by a clinician examining the subject 310 or by the subject 310) with a therapeutically effective amount of the medication 410. As a result of the administration of the medication, it is expected that the risk for VTE in the subject 310 to become ameliorated or decreased (e.g., by impeding blood coagulation or lowering the likelihood of clot formation). When the output 405 identified the subject as the non-candidate for the administration of the medication, the administration of the medication 410 may be refrained or withheld from the subject 310. In some instances, the subject 310 may be directed (e.g., by the clinician) to perform other preventative measures, such as exercise or administration of medication for other conditions in the subject 310.|0101| In this manner, the systems and methods described herein can improve the accuracy and precision in determining the degree of risk for VTE in subjects relative to semi -automated and manual techniques. The ML architecture 150 uses information from multiple datasets (e g., image and non-image data) to determine the degree of risk in the subject. The use of data from datasets to encode the parameters of the ML architecture 150 can provide the improved accuracy and precision, thereby reducing false positives or negatives, misinterpretations, and misclassifications. From a clinical perspective, the different datasets can provide for better insights as to the risk of the subject to make personalized decisions regarding the medical care of the subject to prevent the occurrence of VTE. Furthermore, in contrast to other approaches (e.g., the Khorana Score, which is based on oncological and physiological factors), this approach can leverage latent morphological features from blood cell images to determine VTE risk, with higher accuracy and precision. Sincethe ML architecture 150 may be able to achieve accurate determinations, the data processing system can generate automated alerts, when the subject is classified at a level of risk, to provide to the subject or a clinician for the subject to seek immediate medical assistance if needed.

[0102] In addition, from a computer resource perspective, the use of the integrated ML architecture to process the multimodal datasets, can allow for leveraging correlated features and reducing the reliance on redundant calculations. The use of attention may focus and allocate computer resources to the most pertinent and salient features extracted from the multimodal data, thereby increasing computing efficiency. The reduction in redundant calculations for each of the modalities can save computing resources (e.g., in terms of processing and memory) as well as time (e.g., from executing the individual models individually, or manually analyzing the blood samples). The use of the integrated ML architecture can allow for streamlined processing, minimizing repetitive cleaning or transformation tasks.

[0103] FIG. 11 depicts a flowchart of a method 500 of training the ML architecture to determine values indicating venous thromboses (VTE) risk. The method 500 can be implemented or performed by any of the components in the system 100 or system 700 as detailed herein. A computing system can retrieve training data (505). The computing system can identify embedding sets and dataset (505). The computing system can provide embedding sets and dataset to machine learning (ML) architecture (510). The computing system can determine the value indicating VTE risk (515). The computing system can generate classification (520). The computing system can compare classification with label (525). The computing system can determine loss metric based on comparison (530). The computing system can update weights of ML architecture (535).

[0104] FIG. 12 depicts a flowchart of a method 600 of determining values indicating a degree of risk for venous thromboses (VTE) and providing an output identifying the subject as a candidate or non-candidate for administration of medication. The method 600 can be implemented or performed by any of the components in the system 100 or system 700 as detailed herein. The computing system can identify white blood cell (WBC) images and a dataset (605). The computing system can provide WBC images and a dataset to machine learning (ML) architecture (610). Thecomputing system can generate embedding sets (615). The computing system can determine a value indicating VTE risk (620). The computing system can generate classification (625). If the computing system can determine that the value is greater than or equal to the threshold (630). If the value is greater than or equal to the threshold, the computing system can identify subject as a candidate (635). Otherwise, the computing system can identify subject as a non-candidate (640). The computing system can generate an output (645). The computing system can provide an output (650).D: Computing and Network Environment

[0105] Various operations described herein can be implemented on computer systems. FIG. 13 shows a simplified block diagram of a representative server system 700, client computing system 714, and network 726 usable to implement certain embodiments of the present disclosure. In various embodiments, server system 700 or similar systems can implement services or servers described herein or portions thereof. Client computing system 714 or similar systems can implement clients described herein. The system 700 described herein can be similar to the server system 700. Server system 700 can have a modular design that incorporates a number of modules 702 (e.g., blades in a blade server embodiment); while two modules 702 are shown, any number can be provided. Each module 702 can include processing unit(s) 704 and local storage 706.

[0106] Processing unit(s) 704 can include a single processor, which can have one or more cores, or multiple processors. In some embodiments, processing unit(s) 704 can include a general- purpose primary processor as well as one or more special-purpose co-processors such as graphics processors, digital signal processors, or the like. In some embodiments, some or all processing units 704 can be implemented using customized circuits, such as application specific integrated circuits (ASICs) or field programmable gate arrays (FPGAs). In some embodiments, such integrated circuits execute instructions that are stored on the circuit itself. In other embodiments, processing unit(s) 704 can execute instructions stored in local storage 706. Any type of processors in any combination can be included in processing unit(s) 704.

[0107] Local storage 706 can include volatile storage media (e.g., DRAM, SRAM, SDRAM, or the like) and / or non-volatile storage media (e.g., magnetic or optical disk, flash memory, or the like). Storage media incorporated in local storage 706 can be fixed, removable or upgradeable as desired. Local storage 706 can be physically or logically divided into various subunits, such as a system memory, a read-only memory (ROM), and a permanent storage device. The system memory can be a read-and-write memory device or a volatile read-and-write memory, such as dynamic random-access memory. The system memory can store some or all of the instructions and data that processing unit(s) 704 need at runtime. The ROM can store static data and instructions that are needed by processing unit(s) 704. The permanent storage device can be a non-volatile read-and-write memory device that can store instructions and data even when module 702 is powered down. The term “storage medium” as used herein includes any medium in which data can be stored indefinitely (subject to overwriting, electrical disturbance, power loss, or the like) and does not include carrier waves and transitory electronic signals propagating wirelessly or over wired connections.[0108| In some embodiments, local storage 706 can store one or more software programs to be executed by processing unit(s) 704, such as an operating system and / or programs implementing various server functions, such as functions of the system 700 of FIG. 13 or any other system described herein, or any other server(s) associated with system 700 or any other system described herein.[0l09| “Software” refers generally to sequences of instructions that, when executed by processing unit(s) 704, cause server system 700 (or portions thereof) to perform various operations, thus defining one or more specific machine embodiments that execute and perform the operations of the software programs. The instructions can be stored as firmware residing in read-only memory and / or program code stored in non-volatile storage media that can be read into volatile working memory for execution by processing unit(s) 704. Software can be implemented as a single program or a collection of separate programs or program modules that interact as desired. From local storage 706 (or non-local storage described below), processing unit(s) 704 can retrieveprogram instructions to execute and data to process in order to execute various operations described above.

[0110] In some server systems 700, multiple modules 702 can be interconnected via a bus or other interconnect 708, forming a local area network that supports communication between modules 702 and other components of server system 700. Interconnect 708 can be implemented using various technologies, including server racks, hubs, routers, etc.[0111 A wide area network (WAN) interface 710 can provide data communication capability between the local area network (interconnect 708) and the network 726, such as the Internet. Technologies can be used, including wired (e.g., Ethernet, IEEE 702.3 standards) and / or wireless technologies (e.g., Wi-Fi, IEEE 702.24 standards).[01.1.2J In some embodiments, local storage 706 is intended to provide working memory for processing unit(s) 704, providing fast access to programs and / or data to be processed while reducing traffic on interconnect 708. Storage for larger quantities of data can be provided on the local area network by one or more mass storage subsystems 712 that can be connected to interconnect 708. Mass storage subsystem 712 can be based on magnetic, optical, semiconductor, or other data storage media. Direct attached storage, storage area networks, network-attached storage, and the like can be used. Any data stores or other collections of data described herein as being produced, consumed, or maintained by a service or server can be stored in mass storage subsystem 712. In some embodiments, additional data storage resources may be accessible via WAN interface 710 (potentially with increased latency).

[0113] Server system 700 can operate in response to requests received via WAN interface 710. For example, one of the modules 702 can implement a supervisory function and assign discrete tasks to other modules 702 in response to received requests. Work allocation techniques can be used. As requests are processed, results can be returned to the requester via WAN interface 710. Such operation can generally be automated. Further, in some embodiments, WAN interface 710 can connect multiple server systems 700 to each other, providing scalable systems capable ofmanaging high volumes of activity. Other techniques for managing server systems and server farms (collections of server systems that cooperate) can be used, including dynamic resource allocation and reallocation.10114] Server system 700 can interact with various user-owned or user-operated devices via a wide-area network such as the Internet. An example of a user-operated device is shown in FIG. 6 as client computing system 714. Client computing system 714 can be implemented, for example, as a consumer device such as a smartphone, other mobile phone, tablet computer, wearable computing device (e.g., smart watch, eyeglasses), desktop computer, laptop computer, and so on.

[0115] For example, client computing system 714 can communicate via WAN interface 710. Client computing system 714 can include computer components such as processing unit(s) 716, storage device 718, network interface 720, user input device 722, and user output device 724. Client computing system 714 can be a computing device implemented in a variety of form factors, such as a desktop computer, laptop computer, tablet computer, smartphone, other mobile computing device, wearable computing device, or the like.

[0116] Processing unit(s) 716 and storage device 718 can be similar to processing unit(s) 704 and local storage 706 described above. Suitable devices can be selected based on the demands to be placed on client computing system 714; for example, client computing system 714 can be implemented as a “thin” client with limited processing capability or as a high-powered computing device. Client computing system 714 can be provisioned with program code executable by processing unit(s) 716 to enable various interactions with server system 700.

[0117] Network interface 720 can provide a connection to the network 726, such as a wide area network (e.g., the Internet) to which WAN interface 710 of server system 700 is also connected. In various embodiments, network interface 720 can include a wired interface (e.g., Ethernet) and / or a wireless interface implementing various RF data communication standards such as Wi-Fi, Bluetooth, or cellular data network standards (e.g., 3G, 4G, LTE, etc.).

[0118] User input device 722 can include any device (or devices) via which a user can provide signals to client computing system 714; client computing system 714 can interpret the signals as indicative of particular user requests or information. In various embodiments, user input device 722 can include any or all of a keyboard, touch pad, touch screen, mouse or other pointing device, scroll wheel, click wheel, dial, button, switch, keypad, microphone, and so on.

[0119] User output device 724 can include any device via which client computing system 714 can provide information to a user. For example, user output device 724 can include a display to present images generated by or delivered to client computing system 714. The display can incorporate various image generation technologies, e g., a liquid crystal display (LCD), lightemitting diode (LED) including organic light-emitting diodes (OLED), projection system, cathode ray tube (CRT), or the like, together with supporting electronics (e.g., digital-to-analog or analog- to-digital converters, signal processors, or the like). Some embodiments can include a device such as a touchscreen that functions as both input and output device. In some embodiments, other user output devices 724 can be provided in addition to or instead of a display. Examples include indicator lights, speakers, tactile “display” devices, printers, and so on.

[0120] Some embodiments include electronic components, such as microprocessors, storage and memory that store computer program instructions in a computer-readable storage medium. Many of the features described in this specification can be implemented as processes that are specified as a set of program instructions encoded on a computer-readable storage medium. When these program instructions are executed by one or more processing units, they cause the processing unit(s) to perform various operations indicated in the program instructions. Examples of program instructions or computer code include machine code, such as is produced by a compiler, and files including higher-level code that are executed by a computer, an electronic component, or a microprocessor using an interpreter. Through suitable programming, processing unit(s) 704 and 716 can provide various functionality for server system 700 and client computing system 714, including any of the functionality described herein as being performed by a server or client, or other functionality.[01211 It will be appreciated that server system 700 and client computing system 714 are illustrative and that variations and modifications are possible. Computer systems used in connection with embodiments of the present disclosure can have other capabilities not specifically described here. Further, while server system 700 and client computing system 714 are described with reference to particular blocks, it is to be understood that these blocks are defined for convenience of description and are not intended to imply a particular physical arrangement of component parts. For instance, different blocks can be but need not be located in the same facility, in the same server rack, or on the same motherboard. Further, the blocks need not correspond to physically distinct components. Blocks can be configured to perform various operations, e g., by programming a processor or providing appropriate control circuitry, and various blocks might or might not be reconfigurable depending on how the initial configuration is obtained. Embodiments of the present disclosure can be realized in a variety of apparatus, including electronic devices implemented using any combination of circuitry and software.

[0122] While the disclosure has been described with respect to specific embodiments, one skilled in the art will recognize that numerous modifications are possible. Embodiments of the disclosure can be realized using a variety of computer systems and communication technologies, including but not limited to the specific examples described herein. Embodiments of the present disclosure can be realized using any combination of dedicated components and / or programmable processors and / or other programmable devices. The various processes described herein can be implemented on the same processor or different processors in any combination. Where components are described as being configured to perform certain operations, such configuration can be accomplished, e.g., by designing electronic circuits to perform the operation, by programming programmable electronic circuits (such as microprocessors) to perform the operation, or any combination thereof. Further, while the embodiments described above may make reference to specific hardware and software components, those skilled in the art will appreciate that different combinations of hardware and / or software components may also be used and that particular operations described as being implemented in hardware might also be implemented in software or vice versa.

[0123] Computer programs incorporating various features of the present disclosure may be encoded and stored on various computer-readable storage media; suitable media include magnetic disk or tape, optical storage media such as compact disk (CD) or DVD (digital versatile disk), flash memory, and other non-transitory media. Computer-readable media encoded with the program code may be packaged with a compatible electronic device, or the program code may be provided separately from electronic devices (e.g., via Internet download or as a separately packaged computer-readable storage medium).

[0124] Thus, although the disclosure has been described with respect to specific embodiments, it will be appreciated that the disclosure is intended to cover all modifications and equivalents within the scope of the following claims.

Claims

WHAT IS CLAIMED IS:

1. A method of determining values indicating degrees of risk for venous thromboses (VTEs) in subjects, comprising: identifying, by one or more processors, a first plurality of images for a first subject at risk of VTE, each of the first plurality of images corresponding to a respective white blood cell (WBC) in a first blood sample obtained from the first subject at a first time; providing, by the one or more processors, the first plurality of images to a machine learning (ML) architecture, wherein the ML architecture is established using a plurality of examples, each of the plurality of examples comprising: (i) a respective plurality of embedding sets for each of a second plurality of images, each of the second plurality of images corresponding to a respective WBC in a second blood sample obtained from a second subject at a second time, and (ii) a label indicating one of presence or absence of VTE at a time interval relative to the second time; generating, by the one or more processors, based on executing the ML architecture, a plurality of embedding sets, each embedding set of the plurality of embedding sets corresponding to a respective image of the first plurality of images; determining, by the one or more processors, based on executing the ML architecture using the plurality of embedding sets, a value indicating a degree of risk of VTE for the first subject at the time interval relative to the first time; storing, by the one or more processors, using one or more data structures, an association between the first subject and the value.

2. The method of claim 1, further comprising: identifying, by the one or more processors, the subject as a non-candidate for administration of a medication for VTE, responsive to the value not satisfying a threshold; and providing, by the one or more processors, for presentation via a user interface, an output identifying the subject as the non-candidate for administration of the medication, wherein subsequent to provision of the output, administration of the medication for VTE is refrained.

3. The method of claim 1, further comprising” identifying, by the one or more processors, the subject as a candidate for administration of a medication for VTE, responsive to the value satisfying a threshold; and providing, by the one or more processors, for presentation via a user interface, an output identifying the subject as the candidate for administration of the medication.

4. The method of claim 3, wherein subsequent to provision of the output, the subject is administered with the medication, wherein the medication for VTE comprises at least one of a direct oral anticoagulant (DOAC), a low molecular weight heparin (LMWH), an unfractionated heparin (LTFH), a vitamin K antagonist, or a fondaparinux.

5. The method of claim 1, further comprising: generating, by the one or more processors, a classification indicating a risk level of a plurality of risk levels for the first subject, in accordance with a comparison between the value and a respective range for each of the plurality of risk levels; and providing, by the one or more processors, for presentation via a user interface, an output identifying the classification for the first subject.

6. The method of claim 1, further comprising receiving, by the one or more processors, a biomedical image of the first blood sample having first WBCs and first red blood cells (RBCs) obtained from the first subject at the first time, wherein the biomedical image is acquired in accordance with peripheral blood smear (PBS), wherein identifying the first plurality of images further comprises identifying, from the biomedical image, the first plurality of images comprising (i) a first set of WBC images corresponding to the first WBCs and (ii) a first set of RBC images corresponding to the first RBCs, and wherein providing the first plurality of images to the ML architecture comprises providing the first set of WBC images and the first set of RBC images as input to the ML architecture to generate the first plurality of embedding sets, wherein at least one of the plurality of examples comprises respective plurality of embedding sets for each of the second plurality ofimages, the second plurality of images comprising (i) a second set of WBC images corresponding to second WBCs in the second blood sample and (ii) a second set of RBC images corresponding to second RBCs in the second blood sample.

7. The method of claim 1, further comprising receiving, by the one or more processors, a first dataset comprising at least one of (i) a first electronic document defining a first plurality of traits associated with the first subject, (ii) a first gene profile defining a first plurality of characteristics of genes in a first sample from the first subject, or (iii) a first sequence record comprising a first plurality of sequences associated with the genes in the first sample of the first subject, wherein providing the first plurality of images to the ML architecture further comprises providing the first dataset to the ML architecture, wherein at least one of the plurality of examples comprises at least one of (i) a second electronic document defining a second plurality of traits associated with the second subject, (ii) a second gene profile defining a second plurality of characteristics of genes in a second sample from the second subject, or (iii) a second sequence record comprising a second plurality of sequences associated with the genes in the second sample of the second subject, and wherein determining the value further comprises determining, based on executing the ML architecture using the first dataset, the value indicating the degree of risk of VTE for the first subject.

8. The method of claim 1, wherein generating the first plurality of embeddings further comprises generating, based on providing the first plurality of images to the ML architecture, the first plurality of embedding sets, each embedding set of the plurality of embedding sets identifying a set of morphological features of the respective WBC of the respective image of the first plurality of images, and further comprising: generating, by the one or more processors, based on executing the ML architecture, a plurality of attention scores for the set of morphological features, each attention score of the plurality of attention scores indicating a degree of weight of a respective morphological feature of the set of morphological features in determining the value.

9. The method of claim 1, wherein the ML architecture further comprises: a feature encoder configured to generate the plurality of embedding sets using the first plurality of images, each embedding set of the plurality of embedding sets corresponding to a reduced dimensional representation of the respective image of the first plurality of images; an aggregator configured to determine the value indicating the degree of risk using the plurality of embedding sets; and a classifier configured to generate a classification based on the value and a first dataset.

10. The method of claim 1, wherein the VTE comprises at least one of a deep vein thromboses (DVT) or a pulmonary embolism (PE), wherein the VTE affects at least one of a leg vein, an arm vein, a pelvic vein, a pulmonary artery, lobar artery, or segment artery, wherein the time interval comprises at least one of 1 day, 5 days, 7 days, 10 days, 14 days, 21 days, 28 days, 1 month, 2 months, 3 months, 4 months, 5 months, 6 months, 7 months, or 8 months, and wherein the subject is at risk of or diagnosed with cancer, wherein the cancer comprises at least one of carcinoma, sarcoma, hematopoietic cancer, adrenal cancer, bladder cancer, blood cancer, bone cancer, brain cancer, breast cancer, carcinoma, cervical cancer, colon cancer, colorectal cancer, corpus uterine cancer, ear, nose and throat (ENT) cancer, endometrial cancer, esophageal cancer, gastrointestinal cancer, head and neck cancer, Hodgkin’s disease, intestinal cancer, kidney cancer, larynx cancer, leukemia, liver cancer, lymph node cancer, lymphoma, lung cancer, melanoma, mesothelioma, myeloma, nasopharynx cancer, neuroblastoma, nonHodgkin’s lymphoma, oral cancer, ovarian cancer, pancreatic cancer, penile cancer, pharynx cancer, prostate cancer, rectal cancer, sarcoma, seminoma, skin cancer, stomach cancer, teratoma, testicular cancer, thyroid cancer, uterine cancer, vaginal cancer, vascular tumor, or metastases thereof.

11. A system for determining values indicating degrees of risk for thromboses in subjects, comprising: one or more processors coupled with memory, configured to:identify a first plurality of images for a first subject at risk of VTE, each of the first plurality of images corresponding to a respective white blood cell (WBC) in a first blood sample obtained from the first subject at a first time; provide the first plurality of images to a machine learning (ML) architecture, wherein the ML architecture is established using a plurality of examples, each of the plurality of examples comprising: (i) a respective plurality of embedding sets for each of a second plurality of images, each of the second plurality of images corresponding to a respective WBC in a second blood sample obtained from a second subject at a second time, and (ii) a label indicating one of presence or absence of VTE at a time interval relative to the second time; generate, based on executing the ML architecture, a plurality of embedding sets, each embedding set of the plurality of embedding sets corresponding to a respective image of the first plurality of images; determine, based on executing the ML architecture using the plurality of embedding sets, a value indicating a degree of risk of VTE for the first subject at the time interval relative to the first time; store, using one or more data structures, an association between the first subject and the value.

12. The system of claim 11, wherein the one or more processors are further configured to: identify the subject as a non-candidate for administration of a medication for VTE, responsive to the value not satisfying a threshold; and provide, for presentation via a user interface, an output identifying the subject as the noncandidate for administration of the medication, wherein subsequent to provision of the output, administration of the medication for VTE is refrained.

13. The system of claim 11, wherein the one or more processors are further configured to identify the subject as a candidate for administration of a medication for VTE, responsive to the value satisfying a threshold; and provide, for presentation via a user interface, an output identifying the subject as the candidate for administration of the medication.

14. The system of claim 13, wherein subsequent to provision of the output, the subject is administered with the medication, wherein the medication for VTE comprises at least one of a direct oral anticoagulant (DOAC), a low molecular weight heparin (LMWH), an unfractionated heparin (UFH), a vitamin K antagonist, or a fondaparinux.

15. The system of claim 11, wherein the one or more processors are further configured to: generate a classification indicating a risk level of a plurality of risk levels for the first subject, in accordance with a comparison between the value and a respective range for each of the plurality of risk levels; and provide, for presentation via a user interface, an output identifying the classification for the first subject.

16. The system of claim 11, wherein the one or more processors are further configured to: receive a biomedical image of the first blood sample having first WBCs and first red blood cells (RBCs) obtained from the first subject at the first time, wherein the biomedical image is acquired in accordance with peripheral blood smear (PBS), identify, from the biomedical image, the first plurality of images comprising (i) a first set of WBC images corresponding to the first WBCs and (ii) a first set of RBC images corresponding to the first RBCs, and provide the first set of WBC images and the first set of RBC images as input to the ML architecture to generate the first plurality of embedding sets, wherein at least one of the plurality of examples comprises respective plurality of embedding sets for each of the second plurality of images, the second plurality of images comprising (i) a second set of WBC images corresponding to second WBCs in the second blood sample and (ii) a second set of RBC images corresponding to second RBCs in the second blood sample.

17. The system of claim 11, wherein the one or more processors are further configured to: receive a first dataset comprising at least one of (i) a first electronic document defining a first plurality of traits associated with the first subject, (ii) a first gene profile defining a first plurality of characteristics of genes in a first sample from the first subject, or (iii) a first sequencerecord comprising a first plurality of sequences associated with the genes in the first sample of the first subject, provide the first dataset to the ML architecture, wherein at least one of the plurality of examples comprises at least one of (i) a second electronic document defining a second plurality of traits associated with the second subject, (ii) a second gene profile defining a second plurality of characteristics of genes in a second sample from the second subject, or (iii) a second sequence record comprising a second plurality of sequences associated with the genes in the second sample of the second subject, and determine, based on executing the ML architecture using the first dataset, the value indicating the degree of risk of VTE for the first subject.

18. The system of claim 11, wherein the one or more processors are further configured to generate, based on providing the first plurality of images to the ML architecture, the first plurality of embedding sets, each embedding set of the plurality of embedding sets identifying a set of morphological features of the respective WBC of the respective image of the first plurality of images, and further comprising: generate, based on executing the ML architecture, a plurality of attention scores for the set of morphological features, each attention score of the plurality of attention scores indicating a degree of weight of a respective morphological feature of the set of morphological features in determining the value.

19. The system of claim 11, wherein the ML architecture further comprises: a feature encoder configured to generate the plurality of embedding sets using the first plurality of images, each embedding set of the plurality of embedding sets corresponding to a reduced dimensional representation of the respective image of the first plurality of images; an aggregator configured to determine the value indicating the degree of risk using the plurality of embedding sets; and a classifier configured to generate a classification based on the value and a first dataset.

20. The system of claim 11, wherein the VTE comprises at least one of a deep vein VTE (DVT) or a pulmonary embolism (PE), wherein the VTE affects at least one of a leg vein, an arm vein, a pelvic vein, a pulmonary artery, lobar artery, or segment artery, wherein the time interval comprises at least one of 5 days, 7 days, 10 days, 14 days, 21 days, 28 days, 1 month, 2 months, 3 months, 4 months, 5 months, 6 months, 7 months, or 8 months, and wherein the subject is at risk of or diagnosed with cancer, wherein the cancer comprises at least one of carcinoma, sarcoma, hematopoietic cancer, adrenal cancer, bladder cancer, blood cancer, bone cancer, brain cancer, breast cancer, carcinoma, cervical cancer, colon cancer, colorectal cancer, corpus uterine cancer, ear, nose and throat (ENT) cancer, endometrial cancer, esophageal cancer, gastrointestinal cancer, head and neck cancer, Hodgkin’s disease, intestinal cancer, kidney cancer, larynx cancer, leukemia, liver cancer, lymph node cancer, lymphoma, lung cancer, melanoma, mesothelioma, myeloma, nasopharynx cancer, neuroblastoma, nonHodgkin’s lymphoma, oral cancer, ovarian cancer, pancreatic cancer, penile cancer, pharynx cancer, prostate cancer, rectal cancer, sarcoma, seminoma, skin cancer, stomach cancer, teratoma, testicular cancer, thyroid cancer, uterine cancer, vaginal cancer, vascular tumor, or metastases thereof.

Images