System for providing a warning in respect to the state of health of a patient during intensive treatment of a malign primary disease

Abstract

The invention relates to a system (1) for providing a warning in respect of the state of health of a patient (P1...PN) during intensive treatment of a malign primary disease, the system comprising • a device (W1), which can be worn on the body of the patient (P1) and serves to measure a plurality of vital parameters and • an evaluation device (A), • wherein vital parameters are selected from a group including pulse, pulse variation, respiratory rate, temperature, electrical skin resistance, oxygen saturation, blood pressure, acceleration, changes in orientation, and activity, • the wearable device (W1) is designed such that the plurality of vital parameters are continuously recorded, • the wearable device (W1) furthermore has at least an interface (I / O) in order to be able to pass data on to the evaluation device (A), • the evaluation device (A) furthermore has at least an interface (I / O) for receiving data recorded by the wearable device (W1), • the evaluation device (A) has a processing unit (CPU), • the evaluation device (A) has a data input device (HID) and a data output device (DIS), • first received data recorded by the data output device (DIS) can be displayed in order for a user (D1...DN) to distinguish and classify regular data and irregular data, • the data input device (HID) is designed to accept the result of distinction and classification of regular data and irregular data made by the user (D1...DN), • the evaluation device (A) has a neural network (KI), and the neural network has previously been trained onto regular data in a self-supervised manner, • the evaluation device (A) evaluates second received data by means of the neural network (KI), the second received data a temporally subsequent to the first data, and the neural network classifies regular data and irregular data, and • the evaluation device (A) is also designed to inform the user by way of a warning if irregular data are found that indicate an imminent infectiological or immunological complication.

Description

[0001] System for providing an alert regarding a patient's health status during intensive treatment of an underlying malignant disease

[0002] background

[0003] Oncological diseases are being diagnosed with increasing frequency and are becoming increasingly treatable. Depending on the type of treatment, therapies can be provided on an inpatient or outpatient basis. This generally depends on the toxicity of the treatment regimen. It can also be assumed that therapies with a higher probability of severe complications (adverse events) are generally performed on an inpatient basis to ensure prompt and effective treatment should a complication arise.

[0004] However, inpatient therapy incurs significantly higher costs than outpatient therapy. These costs are due, on the one hand, to the provision of therapy spaces and, on the other hand, to the provision of appropriate staff. Furthermore, this approach results in a large number of inpatient therapy spaces being used prophylactically, meaning they are not available to patients in acute need of treatment. This is particularly problematic when there is a high demand for inpatient therapy spaces.

[0005] In situations of shortages—caused, for example, by staff shortages—patients who would previously have been treated as inpatients are now being treated on an outpatient basis, thus enabling them to receive treatment. However, due to the higher probability of serious complications, this requires close monitoring so that prompt action can be taken if a serious complication arises.

[0006] In an inpatient treatment setting, this can be well ensured through regular checkups and visits by physicians or other qualified staff. Objective parameters such as temperature, blood pressure, heart rate, and oxygen saturation, as well as blood values ​​such as leukocytes, platelets, hematocrit, erythrocyte sedimentation rate, hemoglobin, and / or CRP, are recorded. On the other hand, the experienced "medical eye" of the staff records clinical changes and can thus identify complications early.

[0007] In an outpatient setting, this cannot be achieved with the same level of intensity, as home visits by a doctor or nursing service, or a patient visit to a specialized outpatient clinic, occur less frequently and do not result in documentation of the same quality. Likewise, not all objective parameters are always recorded in care provided by an outpatient physician.

[0008] In other words, the assessment frequency by qualified personnel decreases accordingly from several times a day to, for example, three times a week, thus changing the quasi-continuous monitoring into aperiodic, punctual and thus intermittent monitoring.

[0009] This means that reliable early detection of potentially life-threatening complications is only possible to a limited extent with this model. As a result, serious complications frequently occur in outpatient oncology treatment, requiring inpatient admission. The severity of the complications is often greater than in inpatient treatment and requires more complex therapy. This often also leads to a delay in the actual oncology treatment.

[0010] The most common complications of intensive oncological therapy regimens are infections, immunologically mediated reactions, such as cytokine release syndrome (CRS), immune effector cell-associated neurotoxicity syndrome (ICANS), etc., and cardiac side effects, such as cardiac arrhythmias.

[0011] Early detection and initiation of appropriate countermeasures at the earliest possible stage are crucial for the successful and rapid management of these complications. Failure to do so can result in severe and prolonged complications, potentially endangering the patient's life and delaying the continuation of oncological therapy. This jeopardizes the cure of the underlying disease. The prior art article is "ViSiBiD: A learning model for early discovery and real-time prediction of severe clinical events using vital signs as big data," by Abdur Rahim Mohammad Forkan, Ibrahim Khalil, and Mohammed Atiquzzaman, published in Computer Networks, Volume 113, 2017, pages 244-257, ISSN 1389-1286, https: / / doi.org / 10.1016 / i.comnet.2016.12.019.The article describes a system for detecting common clinical events resulting from the worsening of chronic diseases. The approach is based on the analysis of large, publicly available databases that were not created for the purpose of targeted patient monitoring.

[0012] However, such an approach is unsuitable for patients at risk of specific complications due to a therapy. In particular, the system is unsuitable for targeted monitoring of specific, potentially life-threatening complications (SCC = serious clinical complications) that can occur during intensive chemo-immunotherapy for cancer, because these complications and the resulting individual physical reactions of the patients are diverse and cannot be diagnosed from the raw data. Rather, the system presented there represents a system for detecting general clinical events that result from the worsening of chronic diseases.

[0013] Furthermore, US patent application US 2021 / 0 290 060 A1 is known from the prior art. It describes a system and method for remote patient management and monitoring. The patient is monitored using a wireless sensor system. The system continuously monitors physiological parameters. The system triggers alarms if the patient's physiological data exceeds threshold values.

[0014] The system is therefore suitable for acute events, but not suitable for predicting serious complications so that they could be treated before an acute phase.

[0015] US patent application US 2002 / 0 338 923 A1 discloses a wearable device that can diagnose specific locations associated with electrical arrhythmias. For this purpose, individual data is collected at specific times. However, this system, which enables individualized rhythmological therapy for patients with heart disease, is not suitable for detecting inflammatory conditions.

[0016] Against this background, one object of the invention is to enable safe outpatient therapy for patients.

[0017] The object is achieved by a system according to claim 1. Further advantageous embodiments are the subject matter of the dependent claims.

[0018] The invention is explained in more detail below with reference to the figures, which show:

[0019] Fig. 1-4 each show a schematic view of aspects according to embodiments of the invention.

[0020] Detailed description of the invention

[0021] The invention will be described in more detail below with reference to the figures. It should be noted that various aspects are described, each of which can be used individually or in combination. This means that any aspect can be used with different embodiments of the invention, unless explicitly presented as a mere alternative.

[0022] Furthermore, for the sake of simplicity, reference will generally be made to only one entity in the following. Unless explicitly stated, the invention may also comprise several of the entities in question. Therefore, the use of the words "a," "an," and "another" is to be understood merely as an indication that at least one entity is used in a simple embodiment.

[0023] Where procedures are described below, the individual steps of a procedure can be arranged and / or combined in any order, unless the context explicitly indicates otherwise. Furthermore, the procedures can be combined with each other unless expressly stated otherwise. Numerical values ​​are generally not to be understood as exact values, but include a tolerance of + / - 1% to + / - 10%.

[0024] References to standards or specifications are to be understood as references to standards or specifications that are / were in effect at the time of the application and / or, if priority is claimed, at the time of the priority application. However, this does not imply a general exclusion of applicability to subsequent or replacing standards or specifications.

[0025] Embodiments of the invention are explained below with reference to the figures.

[0026] An exemplary system 1 for providing a warning regarding the health status of one or more patients PI...PN during intensive treatment of an underlying malignant disease, in particular hematological malignancies, according to embodiments of the invention, comprises a device W1...WN that can be worn on the body of each patient PI...PN for measuring a plurality of vital parameters. Each portable device W1...WN is assigned to exactly one patient PI...PN. Treatment refers in particular to treatment, e.g., using chemotherapy and / or immunotherapy and / or radiation therapy.

[0027] The assignment can be made, for example, by providing a unique identification feature, e.g., during data transmission, provision of data sets, etc. For example, a MAC address of a communication device of the respective portable device W1...WN can also be used for this purpose.

[0028] The system also has an evaluation device A.

[0029] The portable devices W1...WN are each configured to continuously record a plurality of vital parameters. Without limiting the generality, a wide variety of (digitizable) vital parameters can be recorded. In one embodiment of the invention, vital parameters are selected from a group comprising pulse, pulse variation, respiratory rate, temperature, electrical skin resistance, oxygen saturation, blood pressure, acceleration, changes in orientation, and activity. It should be noted that, for example, activity can be derived from other vital parameters such as acceleration, power, and / or location data and / or magnetic field data.

[0030] It should be noted that the portable devices W1...WN are not necessarily of the same type. Likewise, it is not absolutely necessary for all portable devices W1...WN to record the same vital parameters. However, it is desirable and advantageous if the portable devices W1...WN are of the same type and / or at least record a similar majority of vital parameters.

[0031] The portable devices W1...WN further comprise at least one I / O interface for forwarding acquired data to the evaluation device A. Likewise, the evaluation device A further comprises at least one I / O interface for receiving acquired data directly or indirectly from one or more of the portable devices W1...WN.

[0032] Examples of interfaces can be both wireless and wired. Examples of wireless interfaces include Bluetooth, DECT, Wi-Fi, and ZigBee, but this does not exclude other wireless interfaces. Examples of wired interfaces include interfaces based on electrical conductors, such as USB or Ethernet, or interfaces based on light waves, such as fiber optics, but this does not exclude other wireless interfaces.

[0033] The evaluation device A has a processing device CPU and also a data input device HID and a data output device DIS.

[0034] The evaluation device A can, for example, be a conventional, program-prepared desktop computer. The processing device (CPU) can then be an existing microprocessor and / or a graphics processor. Alternatively or additionally, ASICS, FPGAs, signal processors, etc. can also be provided as (part of) a processing device (CPU).

[0035] The data input device (HID) can be, for example, a keyboard and / or a graphics tablet and / or a pointing device, such as a mouse. Other devices serving the purpose of input can be provided alternatively or additionally.

[0036] The data output device (DIS) can be a screen. Other devices serving the output purpose can be provided alternatively or in addition.

[0037] First received data recorded by the data output device DIS can be displayed by a user Dl ... DN, e.g. by means of the data output device DIS, for the purpose of distinguishing and assigning regular data and irregular data.

[0038] Typically, such a user is trained Dl ... DN, e.g. a licensed physician.

[0039] The data input device HID is configured to accept a differentiation and assignment of regular data and irregular data by the user Dl ... DN.

[0040] Thus, the user Dl ... DN can, for example, classify data for a training phase for a neural network Kl with respect to a specific patient PI.

[0041] The evaluation device A has a neural network Kl in terms of programming, whereby the neural network Kl was previously trained self-supervised on regular data.

[0042] The evaluation device A evaluates second data received after the training phase by means of the neural network K1, wherein the second data received are temporally subsequent to the first data, wherein the neural network assigns regular data and irregular data, wherein the evaluation device A is further configured to notify a user Dl ... DN and / or the patient PI by means of a warning when irregular data is detected.

[0043] This means that by means of the invention, vital parameters of a patient PI (or several patients PI...PN) can now be recorded via a (respective) portable device W1...WN and evaluated in real time using a Kl algorithm. The current condition of the patient in question can be continuously compared with the condition that has been defined as normal for the patient in question by a user D1 ... DN. This involves a complex interplay of several vital parameters. The interplay of the vital parameters can vary depending on the time of day and the activity of the patient in question. This means that unlike the prior art, not only are limit violations of a single vital parameter detected, but the time-of-day and activity-dependent course of different vital parameters is taken into account by the evaluation unit A.

[0044] A deviation from the individual's normal state indicates a potential complication and can therefore be used as a starting point for further medical measures. Such a deviation can be used to notify a user (Dl...DN) and / or the patient (PI...PN). Notification can include an appointment at a practice or clinic, contact to clarify the condition, initiating an emergency response, etc.

[0045] Thus, medical examinations can be avoided as long as the evaluation using the individually adjusted algorithm indicates a normal, risk-free patient condition. This can also be communicated to the respective user (Dl...DN) as well as to the respective patient (PI...PN).

[0046] The invention thus eliminates numerous routine visits to a clinic or practice, while also preventing excessively long intervals between visits, which can lead to the aggravation of complications and thus to the development of serious side effects. The invention thus makes it possible to transform outpatient cancer therapy from a rigid visit schedule to a needs-based visit schedule. This also makes it possible to safely transfer certain forms of therapy (CAR-T cell therapy, autologous blood stem cell transplantation) from inpatient to outpatient care.

[0047] Only the early detection of complications enables early intervention and, accordingly, outpatient treatment is more frequent than inpatient admission.

[0048] Continuing the automated KL analysis of vital signs after the intervention also allows the physician to remotely monitor the success of the treatment and thus the progression of complications. In many cases, this makes inpatient admission for the treatment of an infectious complication obsolete, as the primary purpose of inpatient admission is close patient monitoring. This can be ensured more continuously and efficiently with the invention.

[0049] The invention thus also reduces the burden on inpatient oncology and reduces outpatient appointments. The possibility of early intervention when complications arise and the outpatient monitoring of the success of these interventions lead to further resource savings in both inpatient and outpatient settings. At the same time, they increase patient safety and enhance quality of life and treatment satisfaction, as patients can spend less time in the hospital and more time at home.

[0050] The present invention describes a system and a method for distinguishing between a normal state and the deviations from the normal state (anomalies) of a patient PI... PN.

[0051] In the system or method, an individual risk index is created for each patient. For this purpose, self-supervised learning algorithms are used to map the time series of vital signs P1_HD1 ... P1_HDN of the patient PI to a vector in a step 100 and to concentrate the information. For example, as shown in Figure 1, vital parameters P1_HD1 ... P1_HDN of the patient PI are recorded by different sensors S1 ... SN. These vital parameters originate from different time periods 1 ... N, in which the patient is healthy according to the assessment by a user D1 ... DN. These time periods can be parts of a longer (consecutive) period as well as non-consecutive time periods. This means that vital parameters from different days can also be used as time periods. Sensors can be of different types. In addition to electrical sensors, acceleration sensors, optical sensors, for example, can also be used.for measuring blood sugar, oxygen saturation, etc., thermal sensors, etc.

[0052] This time series of vital parameters P1_HS1 ... P1_HSN can then be mapped to a respective vector P1_V1 ... P1_VN in a step 100 using feature extraction by a parameterized model (e.g., a neural network based on self-supervised learning). The vectors P1_V1 ... P1_VN of the patent PI form a vector data set.

[0053] This feature extraction can be performed using self-supervised learning, whereby features that occur unchanged in close temporal proximity and / or in different vital signs are filtered out. In information theory terms, this feature filtering corresponds to the approximate calculation of mutual information between adjacent sections of the time series and / or between vital signs. The vectors P1_V1 ... P1_VN of a healthy patient form a reference system.

[0054] If another time series of vital parameters P1_DX of the patient PI is now recorded for a further period X, this can be mapped to a vector P1_VX in the same way in an analogous step 150. In this case, a vector P1_VX is again mapped to a corresponding vector P1_VX using the parameterized model (e.g., the already trained neural network).

[0055] In a further step 200, the similarity of the current vector P1_VX with vectors from the vectors P1_V1 ... P1_VN of the patient PI from the vector data set can then be determined.

[0056] In step 300, the closest vector to P1_VX in the reference system can be determined for a new time series P1_HSX to be tested. Subsequently, in step 400, it can be determined, for example, from the distance of the closest vector to the vector P1_VX to be tested, whether the patient has an abnormality in the vital data. If the distance exceeds a predetermined value, there is a risk of serious complications.

[0057] It should be noted that the vector-generating algorithm is preferably trained on data from all patients. The vector to be tested, however, is then compared with a similarity measure to all other vectors of the patient without complications. Anomaly detection is specific to each patient. The method makes it possible to interpret the same pattern as an anomaly for one patient (PI) and as normal for another patient (P2...PN).

[0058] Preferably, the invention can use the algorithms of deep self-supervised learning, which approximately calculate the transinformation in order to filter out statistically over-represented patterns in a time series and to ignore noise (more details can be found in the document "A Simple Framework for Contrastive Learning of Visual Representations" be taken).

[0059] Each of these patterns can be assigned a vector (feature vector) and vectors generating similar patterns that point in a similar direction. Each patient PI...PN is characterized by a set of vectors P1_V1 .... P1_VN in their vector data set that correspond to their normal state. Deviations from the normal state (anomalies) are revealed by vectors whose directions are significantly different from the directions of the vectors of the normal state.

[0060] This means that unlike existing methods and systems, the vector-generating algorithm is trained on the data of all patients (PI...PN), but the anomaly detection is specific to each patient. Thus, the same pattern can represent an anomaly for one patient (PI) and normal for another patient (P2). The training data for the KL algorithm was collected as part of the CoMMoD-CAST "proof of principle" study between 2018 and 2020 at the University Hospital Düsseldorf. Vital parameters were monitored in 79 patients during intensive hematological therapy, as well as after allogeneic or autologous blood stem cell transplantation, using portable devices (W1...WN). The recorded data were retrospectively correlated with the clinical course by a user (Dl...DN). A total of approximately 43,000 hours of vital parameters were recorded.

[0061] Obviously, other training data can also be used, particularly with regard to other oncological therapies. Therefore, the above training data represents only one implementation of the inventive concept with regard to a specific form of therapy. It is essential that the automated risk analysis and assessment is based on the analysis of a very large dataset of vital signs of similar patients (PI...PN) in similar clinical situations.

[0062] Complications that occur during oncological therapies lead to characteristic changes in several vital parameters, which are measured by wearable devices W1...WN or their respective sensors S1...SN. Patterns emerge that deviate from those of normal activity fluctuations. This makes it possible to identify complications based on the degree of deviation of a pattern from those of a normal state.

[0063] This degree of deviation can - possibly together with the patient's subjective perceptions ("patient reported outcomes") - help medical personnel, i.e. the users Dl ... DN, to get a comprehensive impression of the current health of the respective patient PI...PN.

[0064] In addition, with the growing number of new treatment options, the number of complications also increases, such as cytokine release syndrome (CRS) as an immunologically mediated complication of therapy with CAR-T cells or bispecific antibodies.

[0065] Unlike before, this allows for data collection over a longer period of time, meaning that the user (Dl...DN) is no longer dependent on a single, isolated impression for risk analysis. However, a significantly more timely analysis can be performed than was previously possible with conventional inpatient treatment.

[0066] In addition, the invention also allows therapy progress to be monitored.

[0067] Figure 2 shows typical communication relationships between exemplary elements of a system 1 according to the invention.

[0068] For example, a portable device W1 is assigned to a patient PI. This portable device W1 records the patient's vital parameters using sensors S1...SN (see also Figure 4). Alternatively or additionally, it can also record the patient PI's subjective sensations. This data can then be forwarded to the evaluation device A.

[0069] This can be done either by establishing a direct connection (not shown) or by establishing an indirect connection via one or more intermediate stations. The data can be transmitted periodically or aperiodically. The term "transmission" encompasses both active transmission and passive retrieval.

[0070] Figure 2 shows an example of an indirect connection. In this example, acquired data is forwarded to a mobile device M1 of the patient PI in a step 50. This can be done via a wired or wireless interface (as previously described). Without limiting the generality, the communication between the mobile device M1 and the portable device W1 can be and preferably is encrypted.

[0071] The mobile device Ml can be programmed by an application to alternatively or additionally record subjective sensations of the patient PI.

[0072] From the mobile device M1, the recorded data (including the recorded subjective sensations) can be transmitted (i.e., actively or passively) to a database DB for intermediate storage, therapy progress documentation, etc. in a step 60. Without limiting the generality, the communication between the mobile device M1 and the database DB can and preferably is encrypted.

[0073] The recorded data of the patient PI can be transmitted from the database DB to the evaluation device A (actively or passively) in a step 70. Without limiting the generality, the communication between the evaluation device A and the database DB can and preferably is encrypted.

[0074] Likewise, the data can also be transmitted (actively or passively) to a user Dl ... DN in a step 80. Without limiting the generality, the communication between the user Dl ... DN and the database DB can and preferably is encrypted.

[0075] Preferably, vectorization of the acquired data is performed in the evaluation device A as described above. In principle, however, it would also be possible to transfer this vectorization to another device, such as the mobile device M1 or, if sufficiently powerful, to the portable device W1.

[0076] The evaluation device A performs an evaluation as previously described with reference to steps 100-400. The result of the evaluation can then be made available both as information 500 to the user Dl ... DN and (alternatively or additionally) to the user PI, e.g., as a message to their mobile device Ml in step 550 and / or to their portable device W1 or by forwarding from their mobile device Ml to their portable device W1 in step 950. For example, the patient PI can be requested to seek treatment or to make contact, as previously described.

[0077] In order not to unnecessarily unsettle patients, it can also be provided that the user Dl ... DN contacts the user PI via his mobile device Ml or another means of communication such as a landline, email, etc. only after an assessment of the data from step 500 (and possibly from step 80). A risk index SSC is determined using the smallest distance (ie cosine similarity) of the vector P1_VX to be tested with all other vectors of the patient without complications.

[0078] Previously, the neural network Kl is trained within the evaluation device A. The corresponding principle is explained using Figure 3.

[0079] As before, it is assumed that data is being acquired by a portable device W1. This acquired data can then be divided into preferably equal time periods. For example, in Figure 3, the measurement curve shown under the portable device W1 is divided into seven equal time periods Pl_1 ... Pl_7.

[0080] With the help of other data, especially the patient's subjective perceptions, as well as objectively collected data on the state of health, especially clinical documentation, such as blood values, patient charts and findings, laboratory data, diagnostic results, body weight, height, the user Dl ... DN can rate the time periods as either regular or irregular. In a regular time period, no serious clinical complications occurred, whereas in an irregular time period, at least one serious clinical complication occurred.

[0081] A serious complication has occurred, for example, in embodiments of the invention if the CTCAE grade is >3. CTCAE is a classification scheme commonly used in medicine to classify adverse events. A CTCAE grade of 3 represents a medically significant but not immediately life-threatening event that indicates hospitalization or an extension of hospital treatment. The event itself is disabling or can lead to a restriction of activities of daily living and self-care. The higher the grade, the more threatening the situation.

[0082] Then, a number of regular time periods—in this case, P1_1...P1_3—are used, preferably with additional regular time periods from other patients P2...PN characterized in the same way, to train the neural network Kl under self-supervision. The trained neural network Kl determines an SCC (serious clinical complication) value.

[0083] From another of the regular sections—here, this is shown in section Pl_7—the evaluation device A can then determine a so-called null hypothesis Ho, i.e., a value distribution Pi for the patient Pi. This then provides a null hypothesis for this specific patient, which the evaluation device A can then use to calculate a statistical measure for anomaly detection.

[0084] As a control measure, the functioning of the trained neural network Kl can then be tested using other time periods – in this case, Pl_4 ... Pl_6. If functioning, this network then yields a value outside the distribution of the null hypothesis. The significance can be chosen appropriately according to clinical requirements, so that an action is only triggered when the significance is exceeded.

[0085] Further explanations of exemplary details of an embodiment of the invention are given below.

[0086] Anomaly identification is inherently a highly imbalanced binary classification problem, where normal or typical data points are very common and abnormal data points or outliers are typically rare. The distribution of possible anomalies (out-distribution) is unknown, but is assumed to be much broader than the distribution of normal data points (in-distribution). To detect anomalies, the invention pursues, for example, the strategy of finding distribution-specific features, assuming that there are sufficiently large subsets of data points that share at least some of these features. This strategy implies that normal data points are typically very close to at least one of the subsets in the feature space, while outliers are expected to be farther away.

[0087] One way to learn distribution-specific features is to augment the dataset with examples that have high variance for features that are not distribution-specific and are expected to also appear in outliers, but low variance for distribution-specific features. For example, if an in-distribution consists of images of natural objects (e.g., images of "cats," "ships," etc.), then the transformations applied to each image, such as combinations of moderate cropping and resizing, moderate color dithering, and horizontal flipping, will have a strong effect on the individual pixel values ​​(low-level features) but little effect on the object category (high-level features)—a "cat" remains a "cat."Consequently, the information shared by any two transformations of the same image (a positive pair) can be used to define the distribution-specific features. The disadvantage of this approach is that it requires a priori knowledge of which transformations can significantly shift the data points while leaving the distribution-specific features unchanged.

[0088] For the time series data used in the invention, no new data is generated; instead, two arbitrary time intervals x and x' of, for example, 1000 seconds in length, randomly selected within the same hour but separated by, for example, at least 500 seconds, are defined as a positive pair. The valid transformations are random shifts of these intervals within a given hour by, for example, at most 500 seconds. To extract the features that are invariant under these transformations, each time interval x is mapped to a d-dimensional feature vector h using a deep convolutional neural network h = fo(x) as the feature extractor. The neural network fo(x) is trained by a self-supervised contrastive learning objective that approximately maximizes the mutual information for the sampled positive pairs across all recorded hours.

[0089] The goal of self-supervised contrastive learning is to align feature vectors that share invariant information in the feature space (positive pairs) while simultaneously pushing apart feature vectors that do not share invariant information (negative pairs). Negative pairs are not explicitly generated but arise from the formation of pairs of time intervals from different hours. Let hi = fo (xi) and h = fo (x'i) be feature vectors for two randomly selected time intervals within the same hour i of the training dataset. Then (xj, x'i) is a positive pair and (xi, x'k) is a negative pair for I * k. We define the similarity between feature vectors by which is the dot product between l2-normalized feature vectors hl and fi2 (cosine similarity). In self-supervised contrastive representation learning, a loss function can be defined

[0090] (see Ting Chen, Simon Kornblith, Mohammad Norouzi, and Georey Hinton "A simple framework for contrastive learning of visual representations," published in Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Research, pages 1597-1607. PMLR, 13-18 Jul 2020), where 0 < T < 1 is a scalar temperature parameter, n is the number of randomly selected hours (minibatch size) with two randomly selected intervals of 1000s, Xi and x'i per hour. The temperature parameter was set to T = 0.07. The neural network fo was implemented using a ResNet architecture.

[0091] An exemplary neural network Kl, trained with exemplary training data, is based on a ResNet architecture with at least 24 blocks, each block having at least two convolutional layers, each block having a subsequent normalization (e.g., batch normalization) and a ReLU (Rectified Linear Unit) activation. Exemplary convolutional layers have filters of at least size 16 with a step size of 2. Each convolutional layer has at least 32 filters, which duplicate every 12 blocks. The output of the neural network Kl can be forwarded to a projection head having BN, ReLU, and a linear layer.

[0092] Details on this architecture can be found in Shenda Hong, Yanbo Xu, Alind Khare, Satria Priambada, Kevin Maher, Alaa Alji ry, Jimeng Sun, and Alexey Tumanov Holmes, "Health online model ensemble serving for deep learning models in intensive care units," published in Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 1614{1624, 2020.

[0093] The encoder fo maps inputs to a 128-dimensional feature space embedding. The outputs of this network are L-normalized, / r / || / r||, and consequently map to a unit hypersphere. Exemplary five of the exemplary 12 signals of a wearable device W1...WN for recording vital signs and physical activity have a quality index ranging from 0-100. Each of these five signals with quality index is displayed at its corresponding position in the 101-dimensional quality index vector per second. The remaining entries of this vector are set to zero. This representation results in a 5 x 101 + 7 = 512-dimensional input for each second. An Adam optimizer was used with the parameters ßll = 0.9, ß2 = 0.98 and an initial learning rate of IO -3 and a weight loss of IO -3 The model was trained with a batch size of 128 for 500 epochs.

[0094] The input features were stored at a sampling rate of 1 Hz from the portable device W1...WN. The data set of vital and activity data can be accessed with { n}^ =1 with X n E IPi DzT where N is the number of hours across all patients, D is the input dimension, and T is the number of consecutive time points within an hour. We assume T = 3000, which is less than the expected T = 3600 seconds for an hour, because interruptions can occur. Shortening T allows most consecutive time series to be preserved in the data.

[0095] From the set of feature vectors for the training examples T) train = {h m}™ =1, where K is the number of randomly selected 1000 intervals per hour, a score function can be defined to evaluate whether a particular test sample should be classified as an outlier (SCC). For a given feature vector of a test sample, h te st = fo (x te st), the cosine similarity to the nearest training example T) train used as a measure for the detection of SCC samples.

[0096] The cosine similarity based SCC score SSCC(x) is defined as

[0097] K' = 6 and the / it est Samples are random samples from the same hour. The test example x test is classified as SCC if the SCC score is above a threshold. For patient-specific evaluation, the cosine similarity is calculated with respect to the nearest example among the regular training hours of the patient being tested, rather than the regular training hours of all patients. The invention now enables safe outpatient therapy for patients, while at the same time, valuable time and space resources can be reserved for acute patients.

Claims

Claims 1. System (1) for providing a warning regarding the health status of a patient (PI...PN) during intensive treatment of an underlying malignant disease, the system comprising • a device (Wl) that can be worn on the patient’s body (PI) for measuring a plurality of vital parameters and • an evaluation device (A), • where vital parameters are selected from a group comprising pulse, pulse variation, respiratory rate, temperature, electrical skin resistance, oxygen saturation, blood pressure, acceleration, changes in orientation, and activity, • wherein the portable device (Wl) is configured to continuously record the plurality of vital parameters, • wherein the portable device (Wl) further comprises at least one interface (I / O) in order to be able to forward recorded data to the evaluation device (A), • wherein the evaluation device (A) further comprises at least one interface (I / O) in order to be able to receive recorded data from the portable device (Wl), • wherein the evaluation device (A) has a processing device (CPU), • wherein the evaluation device (A) further comprises a data input device (HID) and a data output device (DIS), • whereby first received data recorded by the data output device (DIS) can be displayed by a user (Dl ... DN) for distinguishing and assigning regular data and irregular data, • wherein the data entry device (HID) is configured to accept a distinction and assignment of regular data and irregular data by the user (Dl ... DN), • wherein the evaluation device (A) comprises a neural network (Kl), wherein the neural network was previously trained in a self-supervised manner on regular data, • wherein the evaluation device (A) evaluates second received data by means of the neural network (Kl), wherein the second received data are temporally subsequent to the first data, where the neural network assigns regular data and irregular data, • wherein the evaluation device (A) is further configured to notify a user by means of a warning when irregular data are detected which indicate an impending infectious or immunological complication.

2. System according to claim 1, characterized in that the neural network (Kl) is trained using data from a plurality of patients.

3. System according to claim 1 or 2, characterized in that data are assigned as irregular if a severe clinical complication occurs, the CTCAE grade being >3.

4. System according to one of the preceding claims, characterized in that the evaluation device (A) is configured to determine a null hypothesis from the acquired first regular data, wherein the evaluation device (A) is further configured to calculate a statistical measure for an anomaly detection from the null hypothesis thus determined.

5. System according to one of the preceding claims, characterized in that the interface (I / O) of the portable device (W) is designed as a wireless interface (I / O).

6. System according to one of the preceding claims, characterized in that the interface (I / O) of the evaluation device (A) is designed as a wireless interface (I / O).

7. System according to one of the preceding claims, characterized in that the evaluation device (A) further comprises at least one further interface (I / O) in order to be able to obtain other data relating to the patient.

8. System according to claim 7, characterized in that the other data are selected from a group comprising subjectively perceived symptoms, objectively collected data on the state of health (such as laboratory data, diagnostic results and body weight).

Images