Method and system to detect, predict and prevent outbreaks of rare pathogens

Abstract

The present disclosure relates to artificial intelligence systems and methods for detecting, predicting and preventing outbreaks of rare pathogens. Applications for the present disclosure include, but are not limited to, use cases in the medical sector and in healthcare as well as for decision making in such sectors. More specifically, the present disclosure relates to machine learning technologies that allow to predict infection probabilities of rare pathogens within a population of persons in a medical context. Counter measures can be identified, optimized, and executed to improve patient safety, trigger infection control, and manage medical resources. One aspect relates to a computer-implemented method comprising: obtaining time and location resolved information characterizing movement and interaction of the plurality of persons in the environment of connected locations during a first time period, obtaining one or more test results confirming infection of one or more persons with the first pathogen within the first time period, obtaining a machine learning model trained for estimating the infection probabilities based on the time and location resolved information, and the one or more test results, and estimating the infection probabilities for the first pathogen and the plurality of persons based on inputting the obtained time and location resolved information, and the obtained test results into the trained machine learning model.

Description

June 10, 2024NEC Laboratories Europe GmbH N173OO1WO KAU / Bmn / GfpMETHOD AND SYSTEM TO DETECT, PREDICT AND PREVENT OUTBREAKS OF RARE PATHOGENS5TECHNICAL FIELD

[0001] The present disclosure relates to artificial intelligence (Al) systems and methods for detecting, predicting and preventing outbreaks of rare pathogens causing diseases,o Applications for the present disclosure include, but are not limited to, use cases in the medical sector and in healthcare as well as for decision making in such sectors. More specifically, the present disclosure relates to graph-based machine learning (ML) technologies that allow to estimate / predict infection probabilities of pathogens within a population of persons interacting in an environment e.g., in a medical context. In5 response suitable counter measures can be identified, optimized, and automatically triggered to improve health of the population of persons e.g., by enhancing patient safety, triggering infection control, and managing resource allocation in a hospital or similar environments. 0 TECHNICAL BACKGROUND

[0002] The present disclosure relates to Al-based systems and methods for outbreak detection of bacteria and similar pathogens (e.g., prions, viruses, fungi, etc.) in a hospital or a similar environment. Typically, a challenging part is that on the one hand,5 some pathogens are rarer than others (i.e. , little knowledge, training data etc.), and on the other hand, different pathogen types can spread in different ways. Overall, the mode of transmission of a pathogen depends on several factors. For example, pathogens can spread through contact, airborne, droplet, foodborne, vectorborne, and fomite transmission, etc. The likelihood of a transmission through one of these 0 mediums / modes typically depends on the pathogen type and the number of pathogens.For example, some pathogens can be easily transmitted through the air whereas others cannot be transmitted through fomites. This behavior makes it challenging to identify the source of an outbreak as the transmission route is unclear.

[0003] Pathogens can be related in different ways such as by genetic, morphological, physiological, and ecological properties. While these aspects are important for understanding how they cause diseases and how they evolve, they also determine how they interact with their environment including how they can be transmitted.

[0004] The present disclosure addresses such problems through a novel graph-based Al technique that allows to overcome typical issues caused by rare pathogen outbreaks by leveraging inter alia information obtained from related but more common pathogens. In the following, several prior art technologies are discussed that form a part of the general background of the present disclosure.

[0005] Strategies to Prevent Healthcare-Associated Infections: A Narrative Overview as discussed in Haque M, McKimm J, Sartelli M, Dhingra S, Labricciosa FM, Islam S, Jahan D, Nusrat T, Chowdhury TS, Coccolini F, Iskandar K, Catena F, Charan J, “Strategies to Prevent Healthcare-Associated Infections: A Narrative Overview” published in Risk Management and Healthcare Policy. Available: https: / / doi.org / 1o.2147 / RMHP.S269315 (in the following Ref. [1]) found that hand and environmental hygiene with antibiotic stewardship are the principal measures that minimize healthcare-associated infections and improve treatment outcomes.

[0006] Using Artificial Intelligence in Infection Prevention as discussed in Fitzpatrick F, Doherty A, Lacey G, “ Using Artificial Intelligence in Infection Prevention” published in New Technologies and Advances in Infections Prevention. Available: https: / / d0i.0rg / 10.1007 / s40506-020-00216-7 (in the following Ref. [2]) proposes that Al offers huge potential in infection prevention and control (IPC) and explores its potential IPC benefits in epidemiology, laboratory infection diagnosis, and hand hygiene.

[0007] A Comprehensive Study of Artificial Intelligence and Machine Learning Approaches in Confronting the Coronavirus (COVID-19) Pandemic as discussed in Rahman MM, Khatun F, Uzzaman A, Sami SI, Bhuiyan MA-A, Kiong TS. Published in International Journal of Health Services. Available athttps: / / d0i.0rg / 10.1177 / 00207314211017469 (in the following Ref. [3]) addresses the link between the technologies and the epidemics as well as the potential impacts of technology in medical services with the introduction of ML, deep learning (DL), and natural language processing (NLP) tools.

[0008] An infectious disease / fever screening radar system which stratifies higher-risk patients within ten seconds using a neural network and the fuzzy grouping method as discussed in Guanghao Sun, Takemi Matsui, Yukiya Hakozaki, Shigeto Abe in Journal of Infection. Available at https: / / d0i.0rg / 10.1016 / j.jinf.2014.12.007 (in the following Ref. [4]) finds that a neural network (NNs) and fuzzy clustering method could efficiently detect higher-risk influenza patients within 10s using multiple vital signs.

[0009] Explaining Neural Matrix Factorization with Gradient Rollback as discussed in Carolin Lawrence, Timo Sztyler, Mathias Niepert. Available at https: / / arxiv.org / abs / 2010.05516 (in the following Ref. [5]) propose gradient rollback as a simple yet effective method to track the influence of training samples on the parameters of a ML model.

[0010] In the following, it is assumed that a reader skilled in the art has fundamental knowledge of Al, ML and in particular graph-based ML etc. Thus, for conciseness, terminology, and concepts such as neutral network types, and associated training algorithms that have been presented in relevant textbooks and review articles known to the skilled person are not defined and / or explained in detail herein. For example, the skilled person is assumed to know that generally, neural networks, such as graph-based neural networks (GNNs), as discussed herein, are machine learning models that employ interconnected layers of nonlinear processing units to predict an output or properties of objects based on a received input. Some neural networks include hidden layers in addition to an output layer. The output of each (hidden) layer is used as input to the next layer in the network, i.e., the next hidden layer or the output layer. Each layer of the network generates an output from a received input in accordance with current values of a respective set of network parameters (processing unit connection weights, activation function parameters, etc.). As discussed in detail in the prior art references mentioned above, some neural networks represent and processgraph structures comprising nodes connected by edges. The graphs may be multigraphs in which nodes may be connected by multiple edges.[oon] The nodes and edges of such graphs may have associated node features and edge features. These may be updated using node update functions and edge update functions, which may be implemented by conventional neural networks such as a multi-layer perceptron (MLP) or graph convolutional networks. For example, training such a GNN e.g., via supervised learning using a training set, unsupervised learning, or combinations thereof, results - inter alia - in determining the network parameters of the neural networks implementing the node and edge update functions and / or the function of a graph encoder as accurate as possible, e.g., by minimizing a loss function during training. William L. Hamilton. (2020). Graph Representation Learning. Synthesis Lectures on Artificial Intelligence and Machine Learning, Vol. 14, No. 3, available at https: / / www.cs.mcgill.ca / ~wlh / grl_book / and published in Graph Representation Learning (2020) ISBN: 978-3-031-00460-5 provides a general overview of parts of the general knowledge of the skilled person regarding GNNs and associated training algorithms.SUMMARY

[0012] Prior art technologies, such as the examples discussed above, typically face several challenges e.g., due to scarce existing data on infection dynamics of rare pathogens. For example, while there are several studies on the spread of bacteria in hospitals, most of these studies stop at investigating the main sources of outbreaks and offer concrete solutions such as handwashing (Ref. [1]) or relate to Al-based systems to prompt disinfecting and hand hygiene (Ref. [2]). Recent research focused on predicting the spread of pathogens in different countries, particularly for CO VID-19 (Ref. [3]) and others have studied the detection of infections through patient monitoring to stratify higher-risk patients (Ref. [4]) as well as cancer prediction through images and patient data. Hower such methods do no allow to reliably predict infection dynamics of rare pathogens, in particular in high-risk environments such as hospitals.

[0013] By contrast, the present disclosure relates to a dynamic method for monitoring and preventing outbreaks in healthcare situations for hospital managementand similar aims. Specifically, a first aspect of the present disclosure relates to a computer-implemented method for estimating infection probabilities for a first pathogen (e.g., a rare and dangerous virus) and a plurality of persons interacting in an environment of connected locations (e.g., a hospital building), comprising obtaining time and location resolved information characterizing movement and interaction of the plurality of persons in the environment of connected locations during a first time period, obtaining one or more test results confirming infection of one or more persons with the first pathogen within the first time period, obtaining a machine learning model trained for estimating the infection probabilities based on the time and location resolved information, and the one or more test results, and estimating the infection probabilities for the first pathogen and the plurality of persons based on inputting the obtained time and location resolved information, and the obtained test results into the trained machine learning model.

[0014] For example, as discussed in more detail below, the machine learning model may have been trained based on (i) pathogen similarity information for a plurality of pathogens comprising the first pathogen and one or more second pathogens, (ii) pathogen transmission mode information for the plurality of pathogens, (iii) connectivity information for the environment of connected locations, and (iv) a plurality of time and location resolved movement and interaction information for confirmed historic infections with one or more of the second pathogens

[0015] In other implementations, only a subset of the above-mentioned information (i) to (iv) may be used. For example, simpler systems may only use (i) the pathogen similarity information, and (iv) the plurality of time and location resolved movement and interaction information for confirmed historic infections for training of a machine learning model. Details of model training are discussed below, e.g., with reference to the drawings.

[0016] A further aspect of the present disclosure relates to a computer- implemented method for containing spread of a first pathogen among a plurality of persons interacting in an environment of connected locations comprising estimating infection probabilities for the first pathogen and the plurality of persons interacting inthe environment of connected locations in a first time period using the method discussed above and below, e.g., with reference to Fig. 1, and determining one or more counter measures for the plurality of persons interacting in the environment of connected locations based on the estimating infection probabilities. In some implementations, such a method my further comprise executing a counter measure of the determined one or more counter measures and / or outputting instructions for executing a counter measure of the determined one or more counter measures. The present disclosure also relates to corresponding computing devices and systems (see Fig. io) and computer programs.

[0017] The various aspects disclosed herein may for example be helpful for hospitals to allow them to divert patients that are suspected of carrying infectious diseases or other urgent symptoms that they are not equipped to handle, and also ensures that they can adjust and prepare in an advance for these situations by having an eagle-eye view of their system. For instance, not all hospitals have the capacity to deal with novel diseases or new variants of locally previously eradicated pathogens arriving from abroad. Instead of posing risk to the hospital staff and other patients, redirecting the infected patient to another facility that would be forewarned of the incoming patient can both (1) prevent outbreaks, be (2) lifesaving by assuring prompt and specialized care, (3) ensure a better resource management of staff and medical items.

[0018] Two key advantages provided by the aspects disclosed herein are: First, flexibility and this is insured through constant updates to graphs which allows for displaying current information about the hospital resources and the likelihood of pathogen spread at any given time. Second, the ML systems disclosed herein can operate in an explainable manner which allows for human intervention where a management staff member can decide whether an action is appropriate or not to prevent any outbreaks and offers evidence to show why graph nodes are marked as critical. Further details, possible implementations, and associated technical benefits of the aspects described above are discussed below with reference to the drawings.SHORT DESCRIPTION OF THE DRAWINGS

[0019] Various aspects of the present disclosure are described in more detail in the following by reference to the accompanying drawings. These drawings show:Fig. 1 an exemplary computer-implemented method for estimating infection probabilities for a first pathogen and a plurality of persons interacting in an environment of connected locations according to aspects disclosed herein;Fig. 2 an exemplary computer-implemented method for training a machine learning model for estimating infection probabilities for a first pathogen and a plurality of persons interacting in an environment of connected locations according to aspects disclosed herein;Fig. 3 an exemplaiy overview of a system architecture according to aspects disclosed herein;Fig. 4 an exemplary illustration of static graph components according to aspects disclosed herein;Fig. 5 an exemplaiy illustration of dynamic graph components according to aspects disclosed herein;Fig. 6 an exemplaiy illustration of a deep learning architecture for static graph components graph components according to aspects disclosed herein;Fig. 7 an exemplaiy illustration of a deep learning architecture for dynamic graph components graph components according to aspects disclosed herein;Fig. 8 an exemplary application of the trained ML systems according to aspects disclosed herein;Fig. 9 an exemplary computer-implemented method for containing spread of a first pathogen among a plurality of persons interacting in an environment of connected locations according to aspects disclosed herein;Fig. io a schematic block diagram of a computing device or system configured to execute methods according to aspects disclosed herein.DETAILED DESCRIPTION OF EXEMPLARY IMPLEMENTATIONS

[0020] In the following, some exemplary embodiments / implementations of the various aspects disclosed herein are described in more detail, with reference to the drawings. Naturally, the computing systems and apparatuses of the present disclosure may employ standard hardware components (e.g., a set of on-premises edge computing hardware and / or cloud-based computing resources connect to each other via conventional wired or wireless networking technology). In some implementations, application-specific hardware (e.g., circuitry for training a ML model and / or circuitry for executing a trained ML for pathogen outbreak prediction and prevention) may also be employed. Further, such computing hardware maybe configured to execute software instructions (e.g., retrieved from collocated or remote memory circuitry) to execute the computer-implemented methods discussed herein.

[0021] While specific feature combinations are described in the following paragraphs with respect to the exemplary embodiments of the present disclosure, it is to be understood that not all features of the discussed embodiments have to be present for realizing the disclosure, which is defined by the subject matter of the claims. The disclosed embodiments may be modified by combining certain features of one embodiment with one or more technically and functionally compatible features of other embodiments. Specifically, the skilled person will understand that features, components, processing steps and / or functional elements of one embodiment can be combined with technically compatible features, processing steps, components and / orfunctional elements of any other embodiment of the present disclosure as long as covered by the invention as specified by the appended claims.

[0022] Moreover, the various embodiments discussed herein can be implemented in hardware, software or a combination thereof. For instance, the various modules of the systems and apparatuses disclosed herein maybe implemented via application specific hardware components such as application specific integrated circuits, ASICs, and / or field programmable gate arrays, FPGAs, and / or similar components and / or application specific software modules being executed on multipurpose data and signal processing equipment such as CPUs, DSPs and / or systems on a chip, SOCs, or similar components or any combination thereof.

[0023] For instance, the various computing (sub)-systems discussed herein may be implemented, at least in part, on multi-purpose data processing equipment such as edge computing servers. Similarly, ML model training subsystems or processes discussed herein may be implemented, at least in part, on multi-purpose cloud-based data processing equipment such as a set of cloud-severs and similar technology.

[0024] Fig. 1 shows an exemplary computer-implemented method too for estimating infection probabilities for a first pathogen and a plurality of persons interacting in an environment of connected locations according to aspects disclosed herein. Step 110 comprises obtaining time and location resolved information characterizing movement and interaction of the plurality of persons in the environment of connected locations during a first time period (e.g. organized in a graph as discussed below). Step 120 comprises obtaining one or more test results confirming infection of one or more persons with the first pathogen within the first time period. Step 130 comprises obtaining a machine learning model trained (as discussed below) for estimating the infection probabilities based on the time and location resolved information, and the one or more test results.

[0025] For example, the machine learning model may have been trained based on (i) pathogen similarity information for a plurality of pathogens comprising the first pathogen and one or more second pathogens, (ii) pathogen transmission modeIO information for the plurality of pathogens, (iii) connectivity information for the environment of connected locations, and (iv) a plurality of time and location resolved movement and interaction information for confirmed historic infections with one or more of the second pathogens

[0026] Step 140 comprises estimating the infection probabilities for the first pathogen and the plurality of persons based on inputting the obtained time and location resolved information, and the obtained test results into the trained machine learning model. As discussed for the example system of Fig. 8 the method 100 may thus allow to determine infection probabilities for each person of a population of persons (e.g. patients and hospital staff) that have been interacting in a past time period. As mentioned above, some of the information (i) to (iv) can be left out when training the machine learning model. This allows to simplify training and may lead to higher hardware and energy efficiency.

[0027] In some implementations, as discussed in more detail for the example of Fig. 6, the trained machine learning model may comprise a static component trained for generating pathogen specific embeddings of the pathogen similarity information and the pathogen transmission mode information. Further the trained machine learning model may also comprise a static component trained for generating location specific embeddings of the connectivity information for the environment of connected locations.

[0028] In some implementations, the trained machine learning model may comprises a dynamic component trained for generating a set of infection probabilities for the plurality of persons at the end of the first time period based on the obtained time and location resolved information, the obtained test results and the embeddings generated by the static component of the trained machine learning model. Using the embeddings generated by the static component further simplifies training. In addition, a trained static component maybe used in different systems each comprising a dedicated dynamic component.

[0029] For example, as illustrated in Fig. 6, the trained machine learning model may comprise a graph neural network (GNN) that may comprises a static graph neural network as the static component, preferably comprising three connected static subcomponents for processing the pathogen similarity information, the pathogen transmission mode information, and the location connectivity information, respectively. Technical benefits of such an architecture are also discussed below.

[0030] In some implementations, as illustrated in the example of Fig. 6, the three connected static sub-components of the static graph neural network may comprise: a first static sub-component, comprising a first input layer configured to receive the pathogen similarity information as input, one or more first graph convolutional layers connected to the first input layer, and a first output module connected to the one or more first graph convolutional layers, preferably comprising a first pooling layer and a first dense layer, wherein the first output module outputs a representation of the pathogen similarity information in a first embedding space such that a distance measure for elements in the first embedding space provides a similarity measure for the plurality of pathogens.

[0031] The static graph neural network may further comprise a second static sub-component, comprising: a second input layer configured to receive the pathogen transmission mode information, and an output of the first static sub-component as input, one or more second graph convolutional layers connected to the second input layer, and a second output module connected to the one or more second graph convolutional layers, preferably comprising a second pooling layer and a second dense layer, wherein the second output module outputs a second representation of the pathogen similarity information and the pathogen transmission mode information in a second embedding space such that a distance measure for elements in the second embedding space provides a second similarity measure for the plurality of pathogens that is based in part on similarities of pathogens in respect of pathogen transmission behavior.

[0032] The static graph neural network may further comprise a third static subcomponent, comprising: a third input layer configured to receive the connectivityinformation for the environment, and an output of the second static sub-component as input, one or more third graph convolutional layers connected to the third input layer, and a third output module connected to the one or more third graph convolutional layers, preferably comprising a third pooling layer and a third dense layer, wherein the third output module outputs a plurality of third representations for the plurality of connected locations and the plurality of pathogens in a third embedding space.

[0033] As discussed below, such a static graph neural network may be trained to generate hospital and pathogen specific embeddings efficiently encoding pathogen and / or environment properties that influence infection dynamics while other non-relevant properties maybe filtered out or suppressed during encoding.

[0034] For example, as illustrated in the example of Fig. 4, the pathogen similarity information may be obtained in form of a pathogen similarity graph comprising a plurality of nodes for each pathogen and a plurality of edges connecting pairs of nodes and encoding a pathogen similarity measure (e.g., categorical or via a real number, e.g. a real number between o and 1) for the connected pair of pathogens. In such a configuration, the method too, may further comprise obtaining one or more of: genetic information, phylogenetic information, morphologic information, pathologic information, and ecologic information for the plurality of pathogens, and generating the pathogen similarity measure based on the obtained information for the plurality of pathogens. For example, a neural network model such a multi-layer perceptron, MLP, may be trained for generating the pathogen similarity measure.

[0035] In some implementations, as illustrated by the example of Fig. 4, the pathogen transmission mode information may also be obtained in form of a pathogen transmission mode graph comprising a first plurality of nodes for each pathogen, a second plurality of nodes for each pathogen transmission mode, and a plurality of edges connecting pairs of nodes such that each node of the first plurality of nodes is connected with one or more nodes of the second plurality of nodes but not with other nodes of the first plurality of nodes.

[0036] In some implementations, as illustrated by the examples of Fig. 5, Fig. 7 and Fig. 8, the time and location resolved information may be obtained in form of a movement graph for the plurality of persons and the environment of connected locations, wherein the movement graph may comprise a plurality of nodes for each person and a plurality of edges connecting pairs of nodes and encoding an interaction intensity measure for pairs of persons within the first time period, and, optionally, an interaction type. Further each node for each person may comprises a time and location resolved data structure characterizing behavior and movement of the person within the first time period and the environment of connected locations. This allows for modeling the movement of entities in a static environment e.g., as a sequence of vectors where these sequences encode the dynamic and static characteristics. In addition, the static characteristics may capture the meaning when entities have contact.

[0037] In addition, a transformer model may be used to capture the context and the long-range dependencies through an attention mechanism. For example, as discussed in A. Vaswani et al. “Attention Is All You Need” 31st Conference on Neural Information Processing Systems (NIPS 2017), Long Beach, CA, USA sequence transduction models may be based on recurrent or convolutional neural networks that include an encoder and a decoder. The best performing models typically also connect the encoder and decoder through an attention mechanism. More specifically, a transformer model architecture is discussed that can be trained to enable classifying of time-ordered sequences of movement vectors of patients in the context of infection probability prediction. For example, movement vectors of patients may correspond to a list of n-dimensional vectors, where n is the embedding dimension. Those n- dimensional vectors may have a specific order, i.e., they are time-ordered sequences. The task is now to classify or label this ordered sequence of n-dimensional vectors. One possibility is that the target class is also a vector, then the task is to find a mapping function from our ordered sequence of n-dimensional vectors to the target vector. Such a transformer model as it is known in the art is one solution among others to find such a mapping function.

[0038] In some implementations, the time and location resolved data structure characterizing behavior and movement of the person within the first time period andthe environment of connected locations may thus comprise a time-ordered sequence of vectors encoding dynamic and static characteristics of the behavior and movement of the person within the first time period. Further the data structure for each node of the movement graph may comprise the pathogen specific embeddings and / or the location specific embeddings for the person and the first time period. Further, the data structure of one or more nodes of the plurality of nodes may comprise an indication that the associated person was infected with the first pathogen in the first time period, wherein the indication that the associated person was infected with the first pathogen in the first time period optionally further comprises person specific information and pathogen test specific information.

[0039] For example, the data structure may essentially be a graph, where attributes of entities in the movement graph may also be represented as edges and nodes. For example, if we have a node representing person A, then the related / connected pathogen information are represented / linked has follows: "Person A" "has pathogen information" "[1, ...., 2]" (the last element is a numerical vector). To be more generic, it is also possible to model it as "Person A" "has attributes" "[[0.1, ..., 1], ..., [1, ..., 3]]", where the last element is a feature matrix. A link to the output(s) of the static component discussed above maybe implemented by a pre-processing step, where available pathogen information (e.g., similarity & transmission mode) are associated with a hospital graph (e.g., comprising rooms / locations). Further, for each person a movement graph maybe obtained, i.e., to know for a specific point in time or time window the positions of a person in the hospital. This information can be used to connect the (embedded) pathogen information with the respective person and its movement.

[0040] In some implementations, the dynamic component of the trained machine learning model may thus comprise a trained transformer model component, configured for transforming the time-ordered sequence of vectors into a vector in a vector space with dimension N, based on a learned mapping function that enables to classify the representation vector as being associated with an infection with one of the plurality of pathogens. Further, obtaining the time and location resolved information characterizing movement and interaction of the plurality of persons in the environmentof connected locations may comprise one or more of: obtaining sensor data characterizing the movement and interaction of the plurality of persons from a sensor system of the environment, and obtaining data characterizing the movement and interaction of the plurality of persons from a database of the environment.

[0041] The machine learning model described above and below with reference to Figs. 4 to 8 thus is based inter alia on a novel static and dynamic graph-based Al approach to detect the outbreak of rare pathogens in a hospital environment. Generally, this approach comprising the steps of: reading event logs for each person in the hospital through a sensor network, the hospital database, and laboratory test results; create a static graph component consisting of a three-level graph structure, create a pathogen similarity graph e.g., via accessing internal and external databases; creating a pathogen transmission graph e.g., via accessing internal and external databases; creating a hospital graph of the hospital of interest; creating a dynamic graph component consisting of a graph linked to the static graph component; creating a set of movement graphs based on historical data of the hospital of interest; training a graphbased machine learning model e.g., end-to-end, with the static and dynamic graphs, incl. learning representations encompassing the buildings’ structure combined with pathogen specific properties linked through historical reports; building data sets comprising the building structure, the characteristics of the pathogen, and the order of events to learn e.g., through a transformer model a pathogen outbreak behavior; using the trained machine learning model with a given movement graph and a certain pathogen (see Fig. 8) to classify each node in the graph whether the respective person got potentially infected; optionally a gradient-based method can be applied to compute explanations for the infection dynamics.

[0042] Some technical advantages of the methods and systems discussed herein include: (1) Usage of a three-level graph structure to capture a physical environment with corresponding biological pathogen characteristics and transmission behavior, to identify those pathogens which provide viable information for assessing other specific rare pathogens; (2) learning vector representations encompassing the physical environment combined with biological specific properties linked through historical events for a population of persons associated with the physical environment. This is forexample enabled through a three-step deep learning architecture to encapsulate stepwise the information of individual static graphs; (3) modeling the entity movement in a static environment as a sequence of vectors where these sequences encode the dynamic and static characteristics, where the static characteristics capture when entities have contact, and where a transformer model is used to capture the context and the long-range dependencies throughout an attention mechanism as discussed above.

[0043] Fig. 2 illustrates an exemplary computer-implemented method 200 for training a machine learning model as discussed herein for estimating infection probabilities for a first pathogen and a plurality of persons interacting in an environment of connected locations (e.g., in a hospital) according to aspects disclosed herein. Step 210 comprises obtaining pathogen similarity information for a plurality of pathogens comprising the first pathogen and one or more second pathogens (e.g., less rare pathogens that have at least some similarities with the first, rare, pathogen). Step 220 comprises obtaining pathogen transmission mode information for the plurality of pathogens. For example, if the first pathogen is essentially unknown, such transmission mode information for the first, rare, pathogen may also be absent and may be inferred during training the model, e.g., based on the similarity information.

[0044] Step 230 comprises obtaining connectivity information for the environment of connected locations. For example, this information may just encode whether rooms in a hospital are neighboring or may also encode distance between rooms, etc. Step 240 comprises obtaining a plurality of time and location resolved movement and interaction information for confirmed historic infections with one or more of the second pathogens. For example, historic data recorded for an outbreak of one or more of the second pathogens maybe obtained for a similar structured environment, e.g., a hospital in a different region or country. Step 250 comprises training the machine learning model based on the obtained pathogen similarity information, the obtained pathogen transmission mode information, the obtained connectivity information for the environment of connected locations, and the obtained plurality of time and location resolved movement and interaction information for confirmed historic infections with one or more of the second pathogens. Crucially, combining all these information allows to obtain a trained machine learning model thatgeneralizes to predict infection dynamics of a rare pathogen even though historic infection records for such a pathogen may not exist or may be sparse.

[0045] As discussed in a similar context above, in some implementations, the machine learning model may comprise a static neural network component and a dynamic neural network component, wherein the static neural network component may comprise a first graph neural network and may be configured to generate pathogen specific and / or location specific embeddings of the pathogen similarity information, the pathogen transmission mode information, and the connectivity information. Further, a dynamic neural network component may comprises a second graph neural network and may be configured to generate an output movement graph for the plurality of persons interacting in an environment of connected locations within a first time period, wherein the output movement graph comprises a plurality of nodes for each person and a plurality of edges connecting pairs of nodes and encoding an interaction intensity measure for pairs of persons within the first time period, and, optionally, an interaction type.

[0046] Further, each node for each person of the output movement graph may comprise an indication of an estimated infection probability for the pathogen and the person within the first time period, and, optionally, a time and location resolved data structure characterizing behavior and movement of the person within the first time period and the environment of connected locations. Training such a ML model allows to predict infection probabilities and explain infection dynamics for the first pathogen as soon as an initial infection event of the first pathogen is detected. This allows to automatically identify and execute suitable counter measures as early as possible substantially mitigating harm to patients and hospital staff alike.

[0047] In some implementations, training the static neural network component (see Fig. 6) may comprise: generating a pathogen similarity graph, a pathogen transmission mode graph and a location connectivity graph, preprocessing the pathogen similarity graph into a set of pathogen similarity batches, wherein each pathogen similarity batch comprises a set of pathogen similarity triples, and using a first contrastive loss function to train a first sub-component of the static neural networkcomponent based on the pathogen similarity batches. For example, such a triple can look like the following: ("Pathogen A", "is similar", "Pathogen B", [o.i, ..., -0.4]). Thus, the structure comprises node, relation, node, and a feature vector describing the similarity of the pathogens. The method may further comprises preprocessing the pathogen transmission mode graph into a set of pathogen transmission mode batches, wherein each pathogen transmission mode batch comprises a set of pathogen transmission mode triples, and using a second contrastive loss function to train a second sub-component of the static neural network component based on the pathogen similarity batches and, optionally, an output of the first sub-component of the static neural network component.

[0048] For example, such a triple can look like the following: ("Transmission Type A", "is possible", "Bacteria Type X", [0.1, ..., -0.4]). Here, the structure comprises node, relation, node, and a feature vector describing the likelihood that the transmissions happen. Alternative relation types may comprises "is unlikely", "very likely", etc. The method may further comprise preprocessing the location connectivity graph into a set of location connectivity batches, wherein each location connectivity batch comprises a set of location connectivity triples, and using a third contrastive loss function to train a third sub-component of the static neural network component based on the location connectivity batches and, optionally, an output of the first subcomponent of the static neural network component, and, optionally, an output of the second sub-component of the static neural network component. Here, for example, such a triple can look like the following: ("Room A", "is physically / directly connected with", "Room B"). Thus, the structure comprises node, relation, node. Two nodes are connected if you can reach Room B from Room A without entering another room / staircase / ward / etc. in between. A feature vector may also be added that may encode an effective distance between Room B and Room A.

[0049] In some implementations, training the dynamic neural network component may comprise: generating a plurality of labeled movement graphs for the plurality of time and location resolved movement and interaction information for the confirmed historic infections with the one or more of the second pathogens, wherein each node of each labeled movement graph comprises an indication whether the corresponding person was infected with the respective second pathogen within therespective first time period; and training the dynamic neural network component using a supervised learning algorithm based on the plurality of labeled movement graphs and an output of the static neural network component. For example, the dynamic neural network component may be trained to generate infection probabilities for all people in the movement graph. Further details on training of the neural network models disclosed herein are described below with reference to Figs. 6 to 8.

[0050] Fig. 3 shows an exemplary overview of a neural network system architecture according to aspects disclosed herein. In a typical configuration, a graph neural network model can be trained based on training data obtained from various databases and / or form a sensor network installed in a hospital or a similar environment. In particular, the illustrated system can use existing test result from a laboratory to understand the potential that a certain pathogen infection occurred or did already spread, i.e., did infect other patients. To achieve that, it is possible to consider the movement and interaction of people in the hospital, the transmission probability of different pathogen types via different transmission types, pathogen characteristics, and the structure of the hospital.

[0051] The hospital sensor network may be designed to capture the movement of people and their interactions with various items / locations in a detailed manner. This network may comprise cameras, microphones, temperature sensors, and other sensors strategically placed throughout the hospital. Data collection and processing are performed continuously by the sensors and connected computing devices or systems. Based on the processed sensor data, detailed event log files may be generated for each person in the hospital. These log files may contain comprehensive information about the person's movement, room occupancy, and interactions with items, locations and / or other persons. They may include timestamps and details such as which rooms the person entered, how long they stayed in each room, and which items they used or interacted with during their stay.

[0052] It is possible that the sensor network also includes mobile sensors carried by people. Further, the event logs are potentially extended or combined with the hospital database which records admissions, room assignment, etc. In addition to theselog files, laboratory results are available. Those results are for specific people describing infection events. This information is part of the input for the movement graph which is described below. The system can distinguish between static graph components (“bacteria similarity graph”, “bacteria transmission graph”, and “hospital graph”) and dynamic graph components (“movement graph”). The static graph components describe different infection relevant properties of the hospital and the pathogens. In particular, as discussed below three static partially interlinked knowledge graphs can be constructed form the static part of the obtained input information.

[0053] Fig. 4 an exemplary illustration of static graph components according to aspects disclosed herein. A pathogen similarity graph mentioned above may comprise pathogens (nodes) connected with weighted edges describing how similar they are. As discussed above pathogens may be related in several ways, including genetic, morphological, physiological, and ecological characteristics. For example, a phylogenetic distance measure may be associated with the edges in the pathogen similarity graph. The similarity may also be derived from these four characteristics, e.g., via a trained MLP to construct the graph. The similarity between pathogens is important to understand which information or training data can be exploited for rare bacteria. In particular, different pathogen types may spread in different ways where the mode of transmission of a pathogen typically depends on a few factors. For example, some bacteria can be spread more easily through the air than others but there are also those which cannot be spread through the air at all.

[0054] Besides, pathogens which are similar do not necessarily have the same probability of transmission for a particular transmission mode, e.g., because of point mutations strongly affecting transmission modes. The “pathogen transmission graph” shown in Fig. 4 allows to capture this information. Such a graph typically comprises two types of nodes: Pathogen nodes and transmission mode nodes. This graph may thus be a bipartite graph, i.e., there are only relations between a pathogen and a transmission node but not between two nodes of the same type. The “pathogen transmission graph” may be linked through the pathogen nodes to the “pathogen similarity graph” allowing to exploit, e.g., during model training correlations between pathogen similarity and transmission modes. For example, morphological properties,such as specific hull proteins of viruses are known to be strongly correlated to certain transmission modes, etc.

[0055] The third static graph component captures the structure of the environment, e.g. a hospital, i.e., connectivity information for rooms, staircases, wards, and so on. Those entities are represented by nodes, and they are connected via edges if they are next to each other, e.g., encoding whether and ow easily a person can move between them without entering another entity in between. The nodes may again be interlinked to the nodes of the bacteria transmission graph. In particular, there maybe edges from a node in the Hospital Graph to a pathogen node in the "Pathogen Transmission Graph”, e.g., if the node in the Hospital Graph was in the past, e.g., within the previous year involved in an outbreak event of the linked pathogen node. In essence, such static graphs allow to model different properties and abstraction levels (i.e., pathogen similarity, pathogen transmission, hospital structure) in a single or united graph to find and connect, through a three-level graph structure, semantically matching pathogens for inferring alternative training data for rare outbreaks.

[0056] Fig. 5 shows an exemplary illustration of dynamic graph components (also designated movement graphs herein) according to aspects disclosed herein. The dynamic graph represents a network of people (e.g. everyone in the hospital), where each person is a node, and connections between nodes indicate that they had contact with each other and how. For example, the contact relations between nodes maybe associated with a vector that represents the scope of the contact. As discussed above, the contact information may be derived from recorded log files, sensor data, etc. The scope of contact may include, e.g., if two people were in the same room for a short or a long time, or whether and how intensely they interacted with the same items, etc. This graph can be re-created or dynamically updated on demand (e.g., at the end of each day). The nodes (i.e., people) may further be associated with respective attributes (e.g., via feature vectors or similar) including age, gender, weight, medical history, test results etc. This graph can thus be used for assessing which people potentially got infected with a certain pathogen.

[0057] Details of the deep learning architecture disclosed herein are now described it in two steps: First the deep learning architecture for static graph components (Fig. 6), and subsequently for the dynamic graph components (Fig. 7). Overall, the goal of this architecture is to learn, via training, pathogen type dependent functions which return representations for all people (nodes) so that these can be reliably classified (node classification) for the respective pathogens as infected or not infected, e.g., via predicting an infection probability.

[0058] Fig. 6 is an exemplary illustration of a deep learning architecture for static graph components according to aspects disclosed herein. The deep learning model for the static graph components may processes static graphs as shown in Fig. 4 sequentially, while taking previously learned information into account. As a first step, node representations can be learned for all nodes in a pathogen similarity graph. For that purpose, the input layer of a first static sub-component may take the pathogen similarity graph as input, which comprises nodes representing different types of pathogens and weighted edges describing their similarity as discussed above. Subsequent graph convolutional layer(s) process the graph structure and extract features from nodes and edges. A (graph attention) pooling layer may aggregate information from neighboring nodes and generate a summarized representation for each node. Finally, a dense layer may apply linear transformations and non-linear activations to learn high-level representations. A contrastive loss function maybe used for training e.g., to encourage similar nodes to have embeddings that are closer together in the embedding space, while dissimilar nodes have embeddings that are farther apart, e.g., measured by a distance metric in the embedding space.

[0059] The second step takes the learned pathogen representations into account while processing a pathogen transmission graph by a second sub-component. The transmission type input layer may be connected to an embedding layer to generate embeddings for each transmission type node (e.g., categorical). Similar to before, graph convolution layers maybe used to learn node representations for the transmission type nodes. However, it is not necessary to learn new representations for the pathogen nodes in the pathogen transmission graph. Instead, the output of the first subcomponent obtained in the preceding step may be used. This may ensure (e.g., throughusing (graph attention) pooling layers) that the outcome of the second step comprises representations which encode both, pathogen similarity e.g., in respect of genetic, morphological, physiological, and ecological characteristics and pathogen similarity in respect of the transmission behavior for each pathogen.

[0060] The third step takes the output of the previous step and a hospital graph into account. This sub-component of the static neural network component illustrated in Fig. 5 may be designed such that it learns a n x m representation matrix where n is the number of nodes (in the hospital graph), and m is the number of pathogens, e.g. present in the output of the second sub-component. In other words, it is possible to learn pathogen specific representations for all entities / locations in the hospital graph. This maybe important because a common representation would not capture the distinct characteristics appropriately. For example, the outcome of the second step may be linked to nodes in the hospital graph through historical reports (e.g., records of past outbreaks of more common pathogens in the hospital).

[0061] In some implementations, there is no need to add further items to the hospital graph. In such cases, it is sufficient to model only locations / rooms and their connectivity. The reason is that other relevant information can be captured by the dynamic graph components described below. The architecture of the third subcomponent may further be similar to the first and second sub-component, i.e., it may rely on graph convolution layers and (graph attention) pooling layers to learn representations, e.g., via minimizing a contrastive loss function as known in the art. Overall, such a model architecture constitutes an holistic approach that allows to learn representations encompassing the buildings’ structure combined with pathogen specific properties (e.g., similarity and transmission behavior) linked through historical reports. By learning representations that encompasses both, it is possible to capture how these factors interact and influence each other. For example, a certain class of similar pathogens maybe more infectious in certain hospitals than in others, e.g., due to different arrangement of patient rooms with respect to typical gathering places or the emergency room etc. Further, by considering both specific properties of the physical environment (e.g., a buildings' structure) and biological aspects of the pathogens, valuable insights on infection dynamics and patterns can be extracted.

[0062] Fig. 7 shows an exemplary illustration of a deep learning architecture for dynamic graph components according to aspects disclosed herein. The illustrated deep learning model for the dynamic graph components typically may require the output of the static model and / or its sub-components and a movement graph as illustrated by the example of Fig. 5. As further illustrated in Fig. 7, the nodes of such a movement graph may also be associated with so-called documents which typically are not text documents but a feature matrix. In particular, such a feature matrix may represent a time-ordered sequence of vectors. These vectors on the one hand may represent the event history of the person (node) (see Fig. 5). Additionally, already learned pathogen and room embeddings, e.g., provided by the static neural network component discussed above may also part of the document / feature matrix, e.g., they may be linked through the event history to a person. Such a sequence of vectors may be transformed to a n-dimensional vector through a graph transformer model as known in the art and as discussed above.

[0063] For example, such a transformer model can learn a mapping function which enables to reliably classify the resulting vector, e.g., associating an infection probability to the person based on the event history and the learned embeddings of the static components. Hence, the overall task is to train an graph-based neural network model to reliably classify all nodes in a given movement graph whether they got infected with a certain pathogen. In essence, the model is trained with a set of historic movement graphs where each graph is associated with an infection with one or more pathogens. In such a training set of movement graphs a node label may thus indicate for each node whether the respective person got infected or not. The methods and system disclosed herein allow to combine this training information with the static information on pathogen similarity, transmission modes and building structure (see Fig. 4 and Fig. 5).

[0064] It is important that the static information differs from pathogen-to- pathogen type as described above. As the described deep learning architecture ensures that the learned embeddings are closer together in the embedding space, e.g., by sequentially minimizing a contrastive loss function based on the static graph components, it is possible to translate infection-related information known forcommon pathogens to detect, predict and prevent outbreaks of rare and dangerous pathogens. Overall, aspects of the present disclosure allow to model infection relevant behavior of persons in a hospital as a sequence of vectors where these sequences also encode the building structure, the characteristics of the pathogens, and the order of infection-relevant events for each person (including contact with other people and items etc.). This information is person-specific, and e.g., by exploiting the capability of the transformer model, it becomes possible to capture the context and long-range dependencies across nodes and time using the attention mechanism.

[0065] Further, the methods and systems disclosed herein also include an explainability aspect. For example, it may be important for the user of the methods and systems disclosed herein to understand why a certain prediction of an infection probability is given (i.e., why a specific person is likely to be infected or not). For this purpose, it is possible to use gradient based methods such as gradient rollback (Ref. [5]) e.g., to trace back the connected nodes in a movement graph that mainly contribute to a given prediction. This may be helpful in two aspects. First, critical nodes may be highlighted and brought to the attention of management staff to help to control a given outbreak. It is also possible to link this to the pathogen similarity and pathogen transmission graphs to recommend an appropriate course of action, such as stocking masks, quarantining wards as need, etc. This mechanism may also be helpful for the reallocation and restructuring of wards, e.g., by emphasizing locations where the infection is likely to spread from, and by ensuring a speedy reaction from the medical staff, etc.

[0066] Furthermore, the methods and systems disclosed herein are intended to reflect constantly evolving situations in various healthcare environments. This is shown by consistent updates to the dynamic graph components e.g., regarding medical resources (e.g., staff, equipment) and may be done through a combination of methods such as graph convolutional networks or graph attention networks.

[0067] Fig. 8 illustrates an exemplary application of the trained ML systems according to aspects disclosed herein. As shown in Fig. 8 a model trained as discussed in detail above allows to predict / estimate infection probabilities for a rare pathogenand a plurality of persons interacting in an environment of connected locations in a first time period based on a movement graph for the first time period and one or more positive test results for an infection event of the rare pathogen. As illustrated in Fig. 8 a possible output of such a trained model may be a graph where each node (person) is labeled with a predicted infection probability. In addition, edges may encode correlation between infection probabilities or similar derived information helpful for understanding infection dynamics.

[0068] Fig. 9 shows an exemplary a computer-implemented method 900 for containing spread of a first pathogen among a plurality of persons interacting in an environment of connected locations according to aspects disclosed herein, e.g., by using the output of a system as shown in Fig. 8. Step 910 comprises estimating infection probabilities for the first pathogen and the plurality of persons interacting in the environment of connected locations in a first time period using a method and / or system as discussed above, e.g., with reference to Fig. 1. For example, a subpopulation of persons may be identified with an infection probability larger than a given threshold value, e.g., 75%. Step 920 comprises determining one or more counter measures for the plurality of persons interacting in the environment of connected locations based on the estimating infection probabilities.

[0069] In some implementations, the method 900 may further comprise a step 930 of executing a counter measure of the determined one or more counter measures, and / or outputting instructions for executing a counter measure of the determined one or more counter measures, e.g., to an electronic infection containment system. For example, the one or more counter measures may comprise one or more of: administration of a drug for treatment of the first pathogen and / or of a vaccine against the first pathogen; restricting movement and / or interaction of the persons, controlling operation of a pathogen filter system of the environment of connected locations, controlling operation of a disinfection device or system of the environment of connected locations, controlling operation of a movement and interaction sensor system of the environment of connected locations, determining a pathogen test schedule for the plurality of persons, quarantining the environment of connected locations, outputting a pathogen infection warning to one or more of the plurality ofpersons, controlling operation of a waste management system or a waste monitoring system.

[0070] Estimating infection probabilities for the first pathogen and the plurality of persons interacting in the environment of connected locations in the first time period using one of the methods discussed above provides all of the advantages mentioned above with regards to the computer-implemented method for estimating infection probabilities (e.g., flexibility, explainability). The determination of one or more counter measures for the plurality of persons interacting in the environment of connected locations based on the estimating infection probabilities provides the advantage of dynamically preventing outbreaks. More specifically, the determination of one or more counter measures allows adjustment to and preparation for possible outbreak situations.

[0071] Execution of a counter measure of the determined one or more counter measures prevents outbreaks in a dynamic manner by adjusting to and preparing for possible outbreak scenarios. Similarly, outputting instruction for executing a counter measure provides the same advantages. In addition, outputting instructions for execution may further have the advantage of communicating a counter measure to a plurality of people further strengthening the effectiveness of the counter measure.

[0072] Administration of a drug for treatment of the first pathogen and / or of a vaccine against the first pathogen may further have the advantage of mitigating the effectiveness of the first pathogen and thus preventing further spreading of the pathogen. Similarly, restricting movement and / or interaction of the persons may decrease the geographical reach of the pathogen and further prevent the pathogen from spreading. For example, if infected Person A is restricted to stay in a specific area, the pathogen may not spread outside of the restricted area that Person A is moving in. The same logic holds for infected Person A interacting with a second Person (i.e., Person B). If such interaction is restricted, the spread of the pathogen may also be prevented. Further, controlling operation of a pathogen filter system of the environment of connected locations may prevent the pathogen from spreading through the filter system. In addition, controlling operation of a disinfection device or system of theenvironment of connected locations may further prevent the pathogen from spreading. Moreover, controlling operation of a movement and interaction sensor system of the environment of connected locations may further provide information about the outbreak of the pathogen and allow prevention of spreading the pathogen further.

[0073] Similarly, determining a pathogen test schedule for the plurality of persons may provide information about the infection rate of the specific pathogen. Similar to restricting interaction and movement, quarantining the environment of connected locations may restrict the pathogen to a specific geographic area. Further, outputting a pathogen infection warning to one or more of the plurality of persons may result in more cautious behavior. Finally, controlling operation of a waste management system or a waste monitoring system may ensure proper disposal of infected materials and / or may provide further information about the spread of the pathogen. In summary, while targeting different areas the one or more counter measures may have the purpose of preventing the pathogen from further spreading.

[0074] Fig. 10 shows a schematic block diagram of a computing device or system 1000 comprising processing circuitry 1010 operably connected to memory 1020 and a communication interface 1030. The computing device or system 1000 is configured to execute the various methods disclosed herein, e.g., by executing a computer program stored in the memory 1020 which comprises instructions for the processor 1010 to do so.

[0075] The foregoing disclosure provides illustration and description but is not intended to be exhaustive or to limit the aspects to the precise form disclosed. Modifications and variations may be made in light of the above disclosure or may be acquired from practice of the aspects. As used herein, the term component is intended to be broadly construed as hardware, firmware, or a combination of hardware and software. As used herein, a processor is implemented in hardware, firmware, or a combination of hardware and software.

[0076] It will be apparent that systems and / or methods described herein may be implemented in different forms of hardware, firmware, or a combination ofhardware and software. The actual specialized control hardware or software code used to implement these systems and / or methods is not limiting of the aspects. Thus, the operation and behavior of the systems and / or methods were described herein without reference to specific software code— it being understood that software and hardware can be designed to implement the systems and / or methods based on the description herein.

[0077] Even though particular combinations of features are recited in the claims and / or disclosed in the specification, these combinations are not intended to limit the disclosure of various aspects. In fact, many of these features maybe combined in ways not specifically recited in the claims and / or dis-closed in the specification. Although each dependent claim listed below may directly depend on only one claim, the disclosure of various aspects includes each dependent claim in combination with every other claim in the claim set. A phrase referring to “at least one of’ a list of items refers to any combination of those items, including single members. As an example, “at least one of: a, b, or c” is intended to cover a, b, c, a-b, a-c, b-c, and a-b-c, as well as any combination with multiples of the same element (e.g., a-a, a-a-a, a-a-b, a-a-c, a-b-b, a- c-c, b-b, b-b-b, b-b-c, c-c, and c-c-c or any other ordering of a, b, and c).

[0078] No element, act, or instruction used herein should be construed as critical or essential unless explicitly described as such. Also, as used herein, the articles “a” and “an” are intended to include one or more items, and maybe used interchangeably with “one or more.” Furthermore, as used herein, the terms “set” and “group” are intended to include one or more items (e.g., related items, unrelated items, a combination of related and unrelated items, and / or the like), and may be used interchange-ably with “one or more.” Where only one item is intended, the phrase “only one” or similar language is used. Also, as used herein, the terms “has,” “have,” “having,” and / or the like are intended to be open-ended terms.

[0079] As used herein, the phrase “based on” shall not be construed as a reference to a closed set of information, one or more conditions, one or more factors, or the like. In other words, the phrase “based on A” (where “A” may be information, acondition, a factor, or the like) shall be construed as “based at least on A” unless specifically recited differently.

[0080] As used herein, the term “or” is an inclusive “or” unless limiting language is used relative to the alternatives listed. For example, reference to “X being based on A or B” shall be construed as including within its scope X being based on A, X being based on B, and X being based on A and B. In this regard, reference to “X being based on A or B” refers to “at least one of A or B” or “one or more of A or B” due to “or” being inclusive. Similarly, reference to “X being based on A, B, or C” shall be construed as including within its scope X being based on A, X being based on B, X being based on C, X being based on A and B, X being based on A and C, X being based on B and C, and X being based on A, B, and C. In this regard, reference to “X being based on A, B, or C” refers to “at least one of A, B, or C” or “one or more of A, B, or C” due to “or” being inclusive. As an example of limiting language, reference to “X being based on only one of A or B” shall be construed as including within its scope X being based on A as well as X being based on B, but not X being based on A and B.

[0081] Further, process diagrams such as Fig. 1, Fig. 2 and Fig. 9 do not necessarily indicate a particular order or sequence of steps. For example, steps may also be performed in a different order or, if hardware capabilities allow it, simultaneously, without deviating from the scope of the present disclosure. For ease of reference, further implementation examples are described in the following as numbered aspects:

[0082] Aspect 1: A computer-implemented method for estimating infection probabilities for a first pathogen and a plurality of persons interacting in an environment of connected locations, comprising: obtaining time and location resolved information characterizing movement and interaction of the plurality of persons in the environment of connected locations during a first time period; obtaining one or more test results confirming infection of one or more persons with the first pathogen within the first time period; obtaining a machine learning model trained for estimating the infection probabilities based on the time and location resolved information, and the one or more test results, wherein the machine learning model was trained based on (i) pathogen similarity information for a plurality of pathogens comprising the firstpathogen and one or more second pathogens, (ii) pathogen transmission mode information for the plurality of pathogens, (iii) connectivity information for the environment of connected locations, and (iv) a plurality of time and location resolved movement and interaction information for confirmed historic infections with one or more of the second pathogens; and estimating the infection probabilities for the first pathogen and the plurality of persons based on inputting the obtained time and location resolved information, and the obtained test results into the trained machine learning model.

[0083] Aspect 2: The method of aspect 1, wherein the trained machine learning model comprises a static component trained for generating pathogen specific embeddings of the pathogen similarity information and the pathogen transmission mode information; and / or wherein the trained machine learning model comprises a static component trained for generating location specific embeddings of the connectivity information for the environment of connected locations.

[0084] Aspect 3: The method of aspect 2, wherein the trained machine learning model comprises a dynamic component trained for generating a set of infection probabilities for the plurality of persons at the end of the first time period based on the obtained time and location resolved information, the obtained test results and the embeddings generated by the static component of the trained machine learning model.

[0085] Aspect 4: The method of any of aspects 1 to 3, wherein the trained machine learning model comprises a graph neural network that comprises a static graph neural network as the static component, preferably comprising three connected static sub-components for processing the pathogen similarity information, the pathogen transmission mode information, and the location connectivity information, respectively.

[0086] Aspect 5: The method of aspect 4, wherein the three connected static sub-components of the static graph neural network comprise: a first static subcomponent, comprising a first input layer configured to receive the pathogen similarity information as input; one or more first graph convolutional layers connected to the firstinput layer; and a first output module connected to the one or more first graph convolutional layers, preferably comprising a first pooling layer and a first dense layer, wherein the first output module outputs a representation of the pathogen similarity information in a first embedding space such that a distance measure for elements in the first embedding space provides a similarity measure for the plurality of pathogens; a second static sub-component, comprising: a second input layer configured to receive the pathogen transmission mode information, and an output of the first static subcomponent as input; one or more second graph convolutional layers connected to the second input layer; and a second output module connected to the one or more second graph convolutional layers, preferably comprising a second pooling layer and a second dense layer, wherein the second output module outputs a second representation of the pathogen similarity information and the pathogen transmission mode information in a second embedding space such that a distance measure for elements in the second embedding space provides a second similarity measure for the plurality of pathogens that is based in part on similarities of pathogens in respect of pathogen transmission behavior; and a third static sub-component, comprising: a third input layer configured to receive the connectivity information for the environment, and an output of the second static sub-component as input; one or more third graph convolutional layers connected to the third input layer; and a third output module connected to the one or more third graph convolutional layers, preferably comprising a third pooling layer and a third dense layer, wherein the third output module outputs a plurality of third representations for the plurality of connected locations and the plurality of pathogens in a third embedding space.

[0087] Aspect 6: The method of any of the preceding aspects 1 to 5, wherein the pathogen similarity information is obtained in form of a pathogen similarity graph comprising a plurality of nodes for each pathogen and a plurality of edges connecting pairs of nodes and encoding a pathogen similarity measure for the connected pair of pathogens.

[0088] Aspect 7: The method of aspect 6, further comprising: obtaining one or more of: genetic information, phylogenetic information, morphologic information, pathologic information, and ecologic information for the plurality of pathogens; andgenerating the pathogen similarity measure based on the obtained information for the plurality of pathogens.

[0089] Aspect 8: The method of any of the preceding aspects 1 to 7, wherein the pathogen transmission mode information is obtained in form of a pathogen transmission mode graph comprising a first plurality of nodes for each pathogen, a second plurality of nodes for each pathogen transmission mode, and a plurality of edges connecting pairs of nodes such that each node of the first plurality of nodes is connected with one or more nodes of the second plurality of nodes but not with other nodes of the first plurality of nodes.

[0090] Aspect 9: The method of any of the preceding aspects 1 to 8, wherein the time and location resolved information is obtained in form of a movement graph for the plurality of persons and the environment of connected locations, wherein the movement graph comprises a plurality of nodes for each person and a plurality of edges connecting pairs of nodes and encoding an interaction intensity measure for pairs of persons within the first time period, and, optionally, an interaction type; wherein each node for each person comprises a time and location resolved data structure characterizing behavior and movement of the person within the first time period and the environment of connected locations.

[0091] Aspect 10: The method of aspect 9, wherein the time and location resolved data structure characterizing behavior and movement of the person within the first time period and the environment of connected locations comprises a time-ordered sequence of vectors encoding dynamic and static characteristics of the behavior and movement of the person within the first time period.

[0092] Aspect 11: The method of aspect 9 or aspect 10, wherein the data structure for each node of the movement graph comprises the pathogen specific embeddings and / or the location specific embeddings for the person and the first time period of the method of claim 2.

[0093] Aspect 12: The method of aspect 11, wherein the data structure of one or more nodes of the plurality of nodes comprises an indication that the associated person was infected with the first pathogen in the first time period.

[0094] Aspect 13: The method of aspect 12, wherein the indication that the associated person was infected with the first pathogen in the first time period further comprises person specific information and pathogen test specific information.

[0095] Aspect 14: The method of any of aspects 10 to 13, wherein the dynamic component of the trained machine learning model comprises a trained transformer model component, configured for transforming the time-ordered sequence of vectors in to a vector in a vector space with dimension N, based on a learned mapping function that enables to classify the representation vector as being associated with an infection with one of the plurality of pathogens.

[0096] Aspect 15: The method of any of aspects 1 to 14, wherein obtaining the time and location resolved information characterizing movement and interaction of the plurality of persons in the environment of connected locations comprises one or more of: obtaining sensor data characterizing the movement and interaction of the plurality of persons from a sensor system of the environment; obtaining data characterizing the movement and interaction of the plurality of persons from a database of the environment.

[0097] Aspects 16: A computer-implemented method for containing spread of a first pathogen among a plurality of persons interacting in an environment of connected locations comprising: estimating infection probabilities for the first pathogen and the plurality of persons interacting in the environment of connected locations in a first time period using the method of any of aspects 1 to 15; and determining one or more counter measures for the plurality of persons interacting in the environment of connected locations based on the estimating infection probabilities.

[0098] Aspect 17: The method of aspect 16, further comprising: executing a counter measure of the determined one or more counter measures; and / or outputtinginstructions for executing a counter measure of the determined one or more counter measures.

[0099] Aspect 18: The method of any of aspects 16 or 17, wherein the one or more counter measures comprise one or more of: administration of a drug for treatment of the first pathogen and / or of a vaccine against the first pathogen; restricting movement and / or interaction of the persons; controlling operation of a pathogen filter system of the environment of connected locations; controlling operation of a disinfection device or system of the environment of connected locations; controlling operation of a movement and interaction sensor system of the environment of connected locations; determining a pathogen test schedule for the plurality of persons; quarantining the environment of connected locations; outputting a pathogen infection warning to one or more of the plurality of persons; controlling operation of a waste management system or a waste monitoring system.

[0100] Aspect 19: Computing device or system for estimating infection probabilities for a first pathogen and a plurality of persons interacting in an environment of connected locations, comprising: means for executing the method of any of aspects 1 to 15.

[0101] Aspect 20: Apparatus or system for containing spread of a first pathogen among a plurality of persons interacting in an environment of connected locations comprising: an interface to the computing device or system of aspect 19; and means for executing the method of any of aspects 16 to 19, based on information received via the interface from the computing device or system of aspect 19.

[0102] Aspect 21: Computer program comprising instructions for executing the method of any of aspects 1 to 15, or aspects 16 to 18, when being executed on processing circuitry of a computing device or system.

[0103] Aspect 22: A computer-implemented method for training a machine learning model for estimating infection probabilities for a first pathogen and a plurality of persons interacting in an environment of connected locations, the methodcomprising: obtaining pathogen similarity information for a plurality of pathogens comprising the first pathogen and one or more second pathogens, obtaining pathogen transmission mode information for the plurality of pathogens, obtaining connectivity information for the environment of connected locations; obtaining a plurality of time and location resolved movement and interaction information for confirmed historic infections with one or more of the second pathogens; and training the machine learning model based on the obtained pathogen similarity information, the obtained pathogen transmission mode information, the obtained connectivity information for the environment of connected locations, and the obtained plurality of time and location resolved movement and interaction information for confirmed historic infections with one or more of the second pathogens.

[0104] Aspect 23: The method of aspect 22, wherein the machine learning model comprises a static neural network component and a dynamic neural network component; wherein the static neural network component comprises a first graph neural network and is configured to generate pathogen specific and / or location specific embeddings of the pathogen similarity information, the pathogen transmission mode information, and the connectivity information; and / or wherein the dynamic neural network component comprises a second graph neural network and is configured to generate an output movement graph for the plurality of persons interacting in an environment of connected locations within a first time period, wherein the output movement graph comprises a plurality of nodes for each person and a plurality of edges connecting pairs of nodes and encoding an interaction intensity measure for pairs of persons within the first time period, and, optionally, an interaction type; wherein each node for each person of the output movement graph comprises an indication of an estimated infection probability for the pathogen and the person within the first time period, and, optionally, a time and location resolved data structure characterizing behavior and movement of the person within the first time period and the environment of connected locations.

[0105] Aspect 24: The method of aspect 23, wherein training the static neural network component comprises: generating a pathogen similarity graph, a pathogen transmission mode graph and a location connectivity graph; preprocessing the pathogen similarity graph into a set of pathogen similarity batches, wherein eachpathogen similarity batch comprises a set of pathogen similarity triples; and using a first contrastive loss function to train a first sub-component of the static neural network component based on the pathogen similarity batches; and / or preprocessing the pathogen transmission mode graph into a set of pathogen transmission mode batches, wherein each pathogen transmission mode batch comprises a set of pathogen transmission mode triples; and using a second contrastive loss function to train a second sub-component of the static neural network component based on the pathogen similarity batches and, optionally, an output of the first sub-component of the static neural network component; and / or preprocessing the location connectivity graph into a set of location connectivity batches, wherein each location connectivity batch comprises a set of location connectivity triples; and using a third contrastive loss function to train a third sub-component of the static neural network component based on the location connectivity batches and, optionally, an output of the first subcomponent of the static neural network component, and, optionally, an output of the second sub-component of the static neural network component.

[0106] Aspect 25: The method of aspect 23 or aspect 24, wherein training the dynamic neural network component comprises: generating a plurality of labeled movement graphs for the plurality of time and location resolved movement and interaction information for the confirmed historic infections with the one or more of the second pathogens, wherein each node of each labeled movement graph comprises an indication whether the corresponding person was infected with the respective second pathogen within the respective first time period; and training the dynamic neural network component using a supervised learning algorithm based on the plurality of labeled movement graphs and an output of the static neural network component.

[0107] Aspect 26: Computing device or system for training a machine learning model for estimating infection probabilities for a first pathogen and a plurality of persons interacting in an environment of connected locations, comprising: means for executing the method of any of aspects 23 to 25.[ooio8] Aspect 27: Computer program comprising instructions for executing the method of any of aspects 23 to 25, when being executed on processing circuitry of a computing device or system.USE CASES AND ADDITIONAL APPLICATION SCENARIOS

[0109] Detecting outbreaks of bacteria and other pathogens in hospitals is important for effective infection control and patient safety. Timely identification of outbreaks allows healthcare providers to implement appropriate measures to contain the spread of infection, protect vulnerable patients, and maintain the overall well-being of the hospital environment. This use case highlights the significance of early detection of bacterial outbreaks in a hospital setting and its crucial role in preventing healthcare- associated infections. In a busy hospital setting, where patients with various medical conditions are treated, the risk of bacterial outbreaks is ever-present. It is crucial to note that bacteria are always present in healthcare facilities, and it is the level of concentration that determines whether an outbreak has occurred.

[0110] In a modern hospital, the implementation of a comprehensive sensor network has become increasingly common. This network allows for the continuous monitoring and recording of various environmental factors. The present disclosure provides an assessment for each person in the hospital how likely it is that the person got infected with a particular bacterium. This overall enables us to assess whether an outbreak occurred. It is possible to output a probability value including an explanation for each person in the hospital which reflects how likely it is that the person got infected with the given bacterium. To this end, the methods and systems disclosed herein processes physical entities such as networks of bacteria. Further the system records and process environmental parameter including the temperature. Independently, the output of the system can be used to automatically lock doors or to instruct cleaning machines to clean or disinfect certain areas.

[0111] In addition, the development of new drugs is increasingly necessary as humans adapt to existing antibiotics. The development of new drugs goes along with so called wet lab experiments. A wet lab is a type of laboratory in which a wide range ofexperiments are performed, including characterizing enzymes in biology and titration in chemistry. Therefore, wet lab experiments take time and are expensive and at the same time there are an enormous number of possible combinations which can be analyzed. Therefore, researchers must decide at some point whether it is worth continuing the wet-lab experiments for a particular drug or to continue with the next step. In such a scenario, Al can help to discover molecules which are structurally distinct from known antimicrobials (e.g., an antibiotic). Al can help to identify a small subset of molecules out of millions of molecules which are potentially of interest for the development of a new drug. Our invention enables us to model such complex structures (e.g., molecules) including related attributes (e.g., mass, shape, and physical properties). Naturally, the amount of information is huge and grows daily.

[0112] As input and training data any available protein, drug, disease information including DNA data amino acid sequences may be used. The methods and systems disclosed herein can be adapted to provide an assessment for each entity in the graph how likely it is that it did contribute to a particular disease. In this setting, the molecular structure replaces the structure of the hospital. This overall enables us to assess which elements to consider for the wet-lab experiments. The three-level graphs describe proteins and diseases. Proteins can cause diseases through various mechanisms, primarily when they undergo structural or functional abnormalities. The first graph describes the proteins and their similarity. The similarity includes whether they physical interact with each other, folding behavior, etc. The second graph the relation between proteins and diseases (bipartite graph). The relation described the likelihood that the protein is involved in the outbreak of the disease. The third graph describes how disease are connected, e.g., if they can occur together or if there are potentially transitions between them. The dynamic movement graphs reflect the movement of the molecules in our body. In such a use case, a probability value can be outputted including an explanation for each element in the graph which reflects how likely it is that it is involved in causing a particular disease. This allows the systems and methods disclosed herein to processes physical entities such as bacteria.Independently, the output of the system can be used to automatically control machines used for the wet-lab experiments, e.g., sequencing robots.

[0113] Thus, the present disclosure also provides corresponding methods and systems for such an application scenario and is not limited to prediction, detection and / or prevention of outbreaks of rare pathogens.

Claims

Claims1. A computer-implemented method for estimating infection probabilities for a first pathogen and a plurality of persons interacting in an environment of connected locations, comprising: obtaining time and location resolved information characterizing movement and interaction of the plurality of persons in the environment of connected locations during a first time period; obtaining one or more test results confirming infection of one or more persons with the first pathogen within the first time period; obtaining a machine learning model trained for estimating the infection probabilities based on the time and location resolved information, and the one or more test results, wherein the machine learning model was trained based on (i) pathogen similarity information for a plurality of pathogens comprising the first pathogen and one or more second pathogens, (ii) pathogen transmission mode information for the plurality of pathogens, (iii) connectivity information for the environment of connected locations, and (iv) a plurality of time and location resolved movement and interaction information for confirmed historic infections with one or more of the second pathogens; and estimating the infection probabilities for the first pathogen and the plurality of persons based on inputting the obtained time and location resolved information, and the obtained test results into the trained machine learning model.

2. The method of claim 1, wherein the trained machine learning model comprises a static component trained for generating pathogen specific embeddings of the pathogen similarity information and the pathogen transmission mode information; and / or wherein the trained machine learning model comprises a static component trained for generating location specific embeddings of the connectivity information for the environment of connected locations.

3. The method of claim 2, wherein the trained machine learning model comprises a dynamic component trained for generating a set of infection probabilities for the plurality of persons at the end of the first time period based on the obtained time and location resolved information, the obtained test results and the embeddings generated by the static component of the trained machine learning model.

4. The method of any of claims 1 to 3, wherein the trained machine learning model comprises a graph neural network that comprises a static graph neural network as the static component, preferably comprising three connected static sub-components for processing the pathogen similarity information, the pathogen transmission mode information, and the location connectivity information, respectively.

5. The method of claim 4, wherein the three connected static sub-components of the static graph neural network comprise: a first static sub-component, comprising a first input layer configured to receive the pathogen similarity information as input; one or more first graph convolutional layers connected to the first input layer; and a first output module connected to the one or more first graph convolutional layers, preferably comprising a first pooling layer and a first dense layer, wherein the first output module outputs a representation of the pathogen similarity information in a first embedding space such that a distance measure for elements in the first embedding space provides a similarity measure for the plurality of pathogens; a second static sub-component, comprising: a second input layer configured to receive the pathogen transmission mode information, and an output of the first static sub-component as input; one or more second graph convolutional layers connected to the second input layer; and a second output module connected to the one or more second graph convolutional layers, preferably comprising a second pooling layer and a second dense layer, wherein the second output module outputs a second representation of the pathogen similarity information and the pathogen transmission modeinformation in a second embedding space such that a distance measure for elements in the second embedding space provides a second similarity measure for the plurality of pathogens that is based in part on similarities of pathogens in respect of pathogen transmission behavior; and a third static sub-component, comprising: a third input layer configured to receive the connectivity information for the environment, and an output of the second static sub-component as input; one or more third graph convolutional layers connected to the third input layer; and a third output module connected to the one or more third graph convolutional layers, preferably comprising a third pooling layer and a third dense layer, wherein the third output module outputs a plurality of third representations for the plurality of connected locations and the plurality of pathogens in a third embedding space.

6. The method of any of the preceding claims 1 to 5, wherein the pathogen similarity information is obtained in form of a pathogen similarity graph comprising a plurality of nodes for each pathogen and a plurality of edges connecting pairs of nodes and encoding a pathogen similarity measure for the connected pair of pathogens, and wherein the method, optionally, further comprises obtaining one or more of: genetic information, phylogenetic information, morphologic information, pathologic information, and ecologic information for the plurality of pathogens; and generating the pathogen similarity measure based on the obtained information for the plurality of pathogens.

7. The method of any of the preceding claims 1 to 6, wherein the pathogen transmission mode information is obtained in form of a pathogen transmission mode graph comprising a first plurality of nodes for each pathogen, a second plurality of nodes for each pathogen transmission mode, and a plurality of edges connecting pairs of nodes such that each node of the first plurality of nodes is connected with one or more nodes of the second plurality of nodes but not with other nodes of the first plurality of nodes.

8. The method of any of the preceding claims 1 to 7, wherein the time and location resolved information is obtained in form of a movement graph for the plurality of persons and the environment of connected locations, wherein the movement graph comprises a plurality of nodes for each person and a plurality of edges connecting pairs of nodes and encoding an interaction intensity measure for pairs of persons within the first time period, and, optionally, an interaction type; wherein each node for each person comprises a time and location resolved data structure characterizing behavior and movement of the person within the first time period and the environment of connected locations.

9. The method of claim 8, wherein the time and location resolved data structure characterizing behavior and movement of the person within the first time period and the environment of connected locations comprises a time-ordered sequence of vectors encoding dynamic and static characteristics of the environment and / or of the behavior and movement of the person within the first time period; and / or wherein the data structure for each node of the movement graph comprises the pathogen specific embeddings and / or the location specific embeddings for the person and the first time period of the method of claim 2; and, optionally, wherein the data structure of one or more nodes of the plurality of nodes comprises an indication that the associated person was infected with the first pathogen in the first time period, wherein the indication that the associated person was infected with the first pathogen in the first time period optionally further comprises person specific information and pathogen test specific information.

10. The method of claim 9, wherein the dynamic component of the trained machine learning model comprises a trained transformer model component, configured for transforming the time-ordered sequence of vectors into a vector in a vector space with dimension N, based on a learned mapping function that enables to classify the representation vector as being associated with an infection with one of the plurality of pathogens.

11. The method of any of claims 1 to 10, wherein obtaining the time and location resolved information characterizing movement and interaction of the plurality of persons in the environment of connected locations comprises one or more of: obtaining sensor data characterizing the movement and interaction of the plurality of persons from a sensor system of the environment; obtaining data characterizing the movement and interaction of the plurality of persons from a database of the environment.

12. A computer-implemented method for containing spread of a first pathogen among a plurality of persons interacting in an environment of connected locations comprising: estimating infection probabilities for the first pathogen and the plurality of persons interacting in the environment of connected locations in a first time period using the method of any of claims 1 to 11; and determining one or more counter measures for the plurality of persons interacting in the environment of connected locations based on the estimating infection probabilities.

13. Computing device or system, comprising: processing circuitry operably connected to memory and a communication interface, wherein the computing device or system is configured for executing the method of any of claims 1 to 12.

14. Computer program comprising instructions for carrying out the method of any of claims 1 to 12, when being executed by processing circuitry of a computing device or system.

Images