Patient similarity

Abstract

This invention relates to patient similarity. [Solution] A computer-implemented method comprising: a step of using an ML model to predict the prognosis for a target patient based on medical history information relating to the target patient, wherein the medical history information relating to the target patient includes treatment instructions given to the target patient, and the ML model is trained with training data including medical history information relating to multiple subjects; a step of comparing modified medical history information relating to the target patient, including the medical history information relating to the target patient and the predicted prognosis, with multiple medical history instances; a step of selecting at least one medical history instance that is most similar to the modified medical history information relating to the target patient based on the comparison; and a step of outputting information indicating the predicted prognosis and the selected at least one medical history instance.

Description

Technical Field

[0001] The present invention relates to patient similarity, and in particular, to a computer-implemented method, a computer program, an information programming device, and a system.

Background Art

[0002] An accurate early diagnosis of a disease or condition is an important aspect of modern medicine. To help patients after an early disease prediction, it is useful to know how to treat the diagnosed condition or disease. Selecting a treatment for a condition or disease from among the many possible treatments that can be performed is sometimes difficult. As the amount of available medical data continues to increase, a method for using this data when selecting a treatment is desired.

Summary of the Invention

Problems to be Solved by the Invention

[0003] In view of the above, a method for predicting a prognosis is desired.

Means for Solving the Problems

[0004] The present invention is defined by the independent claims and is referred to hereinafter. Specific embodiments are defined in the dependent claims.

[0005] According to one embodiment of the first aspect, the Specified herein discloses a computer implementation method that includes the steps of: predicting a prognosis for a patient based on medical history information relating to the patient, wherein the medical history information relating to the patient includes treatment instructions given (proposed) to the patient, and the ML model is trained on training data including medical history information relating to multiple subjects; comparing modified medical history information relating to the patient, including the medical history information relating to the patient and the predicted prognosis, with multiple medical history instances, each medical history instance including medical history information from the training data and relating to a subject (and a certain time period) among the multiple subjects; selecting at least one medical history instance that is most similar to the modified medical history information relating to the patient based on the comparison; and outputting information indicating the predicted prognosis and the selected at least one medical history instance. [Brief explanation of the drawing]

[0006] For example, please refer to the attached drawing. [Figure 1] This diagram shows an overview of the methodology. [Figure 2] This is a diagram showing a machine learning model. [Figure 3] This is a diagram showing an encoder model. [Figure 4] This diagram shows the method. [Figure 5] This diagram shows the method. [Figure 6] This is a diagram of computing devices. [Modes for carrying out the invention]

[0007] The following terms may be used in this specification (this list of definitions is not exhaustive): ● EHR: Electronic Health Record. ● Preventive medicine: A medical specialty that focuses on the health of individuals and communities. The goal of preventive medicine is to promote health and well-being and to prevent disease, disability, and death. ● Diagnostic and Prognostic Prediction: The goal of diagnostic prediction is to calculate the probability that a patient currently has the disease, while prognostic prediction calculates the probability that a certain condition will occur in the future, for example, as a result of treatment. ●Clinical Decision Support System (CDSS): Used to support the decision-making process of clinicians. Clinical Decision Support (CDS) can take various forms. In our case, we focus on prognostic CDS. ● Black Box System: A "black box" is a system whose internal workings cannot be directly observed or understood from the outside. In other words, while the system's inputs and outputs are known, the internal mechanism that converts the inputs into outputs is not transparent or easily interpretable. Deep learning models are sometimes considered "black boxes." ● Machine Learning (ML): A subfield of computer science that "gives computers the ability to learn without being explicitly programmed." It explores the study and construction of algorithms that can learn from data and make predictions about that data. ● Artificial Neural Networks (ANN / NN): A branch of machine learning inspired by how biological nervous systems, such as the brain, process information. These can be used to extract patterns and detect trends that are too complex to be noticed by humans or any other computer technique. ● Deep Learning (DL): A subset of ML that uses ANNs to model data, where "deep" means that multiple layers of neural networks are stacked on top of each other. This "depth" contributes to achieving better results on a variety of problems, but makes it more difficult to reason why a particular prediction was produced.

[0008] Figure 1 is a diagram illustrating the outline of the methodology disclosed herein. In step S1, at least one treatment is proposed. In Figure 1, three treatments, treatments 12A, 12B, and 12C, are proposed. The treatments are proposed based on the EHR 11 of the “current patient,” sometimes referred to as the subject patient. Treatment proposals may be made by a clinician / physician and / or by a model designed / trained to predict treatment.

[0009] In step S2, a prognosis model (also referred to herein as the ML model) is used to predict the prognosis for at least one of the proposed treatments. In Figure 1, three outcomes 13A, 13B, and 13C are predicted, corresponding to three proposed treatments 12A, 12B, and 12C. The prognosis model predicts at least one prognosis based on the patient's EHR 11 and at least one of the proposed treatments. The prognosis model is trained (before its use) with EHR data, for example, in an unsupervised manner. As an example, the prognosis model can predict that a male patient (all his background data is recorded in his EHR) will develop symptom S1 one week later and disease D1 one month later if he receives treatment T1.

[0010] In step S3, a patient similarity model is used to find at least one medical history instance that is most similar to the at least one predicted prognosis. For example, as shown in Figure 1, several similar EHRs 14A, 14B, and 14C are selected corresponding to each of the predicted prognoses 13A, 13B, and 13C. The patient similarity model compares the medical history information and at least one predicted prognosis of the subject with EHRs 10 of several other patients / subjects. Rather than selecting EHRs, at least one medical history instance may be selected, and the medical history instance may include several events relating to the subject.

[0011] EHR 10 data from multiple subjects can be considered as multiple medical history instances. EHR 10 data from multiple subjects can be used as training data for prognostic models and / or patient similarity models.

[0012] In step S4, the clinician proposed a selected treatment 12D, which was selected from, for example, the proposed treatments 12A, 12B, and 12C, based on the predicted prognosis 13A–C and the identified similar medical history instances 14A–C. The identification of similar medical history instances helps clinicians in selecting treatment by improving confidence in the predicted prognosis and by allowing clinicians to study similar medical history instances in more detail to help them select treatment.

[0013] As described above, Figure 1 shows an overview of the methodology. The key parts of the overview are described in more detail below. This specification discloses a method that includes steps / operations corresponding to steps S2 and S3, but does not necessarily include steps S1 or S4.

[0014] Figure 2 shows a prognosis prediction model 20 implemented in a specific way.

[0015] Given clinical data and proposed treatment for a patient, Model 20 predicts the outcome (future clinical data) of the same patient after receiving the input treatment. A neural network-based implementation of the prognostic prediction model 20 can learn from unlabeled clinical data. Relevant clinical information (e.g., diagnosis, symptoms, medication, treatment, etc.) is encoded as time-stamped events. The input treatment is encoded as part of a set of time-stamped events (as if it had been administered).

[0016] In particular, Model 20 makes its predictions based on medical history information about the patient, including medical history information from records or databases (e.g., from EHRs) along with the proposed treatment. The proposed treatment is included in the medical history information as if it had already been administered.

[0017] Medical history information is input to Model 20 as a series of events "e" accompanied by time information. Time information is augmented to each event "e" to generate augmented event information "e-hat" (e with a hat). This augmentation is referred to herein as time embedding / encoding and may be thought of as similar to positional embedding, where each word in a sentence is encoded with information representing its position in the sentence, but instead of position, time (for example, time in the form of the number of days elapsed since the first event) is used. Time embedding can be implemented in several ways, for example, by encoding the time of events into a vector using sine and cosine functions and adding this vector on top of the event information "e". Time embedding in time series representations can generally be used in transformer-based architectures that use positional encoding modules. Some use the original fixed sinusoidal positional encoding, while others can learn time embedding using linear and fully connected layers, respectively. An example of time embedding for time series is Time2Vec (Non-Patent Literature 1), which was developed for supervised learning tasks. The time embedding function is indicated by reference numeral 21 in Figure 2.

Non-Patent Document 1

[0018] A series of self-attention layers 22A, 22B, 22C (collectively referred to as 22) generate a contextualized embedding (“h”) based on the augmented event information ê. The self-attention layer 22 is, for example, similar to those used in conventional transformer architectures.

[0019] The self-attention layer 22 (and optionally the time embedding function 21 as well) can be regarded as a first encoder model. The first encoder model encodes the medical history information regarding the target patient to generate a contextualized embedding h of the target.

[0020] The prognosis prediction model 20 further includes a pooling layer 24, an update predictor 26, an attention network 28, and a plurality of event predictors 29 (29A, 29B, 29C, 29D). The pooling layer 24, the update predictor 26, the attention network 28, and the plurality of event predictors 29 can be regarded as a prediction model. That is, the prognosis prediction model 20 may be considered to include the first encoder model and the prediction model as described above.

[0021] The pooling layer 24 pools the contextualized embedding to generate a single output vector for input to the update predictor 26 (where many pooling techniques can be used; an advantageous example is self-attention pooling). The update prediction unit 26 uses the output of the pooling layer 24 to predict the time of at least one next event among a series of events e that constitute the medical history information regarding the target patient. The update predictor in a specific implementation includes a neural network and uses a regression technique to predict the time of the next event.

[0022] The attention network 28 uses the contextualized embedding h and the output of the update predictor 26 to generate a vector that summarizes the information contained in the embedding h at the time predicted by the update predictor 26. This vector is output to multiple event predictors 29. Each predictor 29 predicts a single event (probability of presence / absence).

[0023] Multiple event predictors 29 predict at least one subsequent event. For example, in a particular implementation, each event predictor is a classifier (neural network) that predicts whether a certain item (event) of medical history information is positive or negative (or absent). For example, a classifier 29 may predict whether the presence of a certain symptom is positive (present in the patient), negative (determined not to be absent), or absent (present or not determined to be absent), and / or another classifier 29 may predict whether the diagnosis of a disease or condition is positive, negative, or absent. In this way, assuming that the proposed treatment has already been administered to the patient, at least one subsequent event is predicted by the prognosis model 20 (along with the time of the said at least one subsequent event). In other words, at least one prognosis is predicted.

[0024] In Figure 2, a specific number of self-attention layers 22 and event predictors 29 are given reference codes, but the prognosis model 20 may include any number of self-attention layers 20 and any number of event predictors 29. The specific implementation of the prognosis model 20 shown in Figure 2 and described above is not mandatory, and other implementations may be used. For example, other machine learning or deep learning methods / architectures such as recurrent neural networks (RNNs) or convolutional neural networks (CNNs) may be used.

[0025] Figure 3 shows the implementation of the patient similarity model 30.

[0026] The patient similarity model 20 uses vector distance (such as cosine similarity) to determine the similarity between instances of medical history information within the medical history information for several subjects. A vector representation is generated for each medical history instance, and a vector representation is generated for the modified medical history information for the target patient, where the modified medical history information includes the medical history information for the target patient along with at least one predicted prognosis (predicted by the prognosis prediction model 20).

[0027] A medical history instance is a series of events relating to a given subject. Each subject in the available data does not necessarily generate only one medical history instance. For example, if the available data contains n events relating to a given subject, one medical history instance may contain those n events, further medical history instances may contain n-1 of those events (e.g., all but the last one), further medical history instances may contain n-2 of those events (e.g., all but the last two or the first two, or all but the first and last one), and so on. For example, a medical history instance may consist of a series of consecutive events relating to a given subject.

[0028] The generation of a vector representation of a single medical history instance is shown in Figure 3. The same process is performed for multiple medical history instances (and for modified medical history information concerning the patient in question). The elements of the dotted L-shaped box 39 are the elements used to train the patient similarity model 30, which will be discussed later.

[0029] The patient similarity model 30 includes a time-embedding function 31, a series of self-attention layers 32, and a pooling layer 34. The functions of these elements are the same as those of the corresponding elements in the prognosis prediction model 20 described above with reference to Figure 2, and redundant explanations are omitted. Elements 31, 32, and 34 are sometimes referred to as the second encoder model. The architecture of the second encoder model in this implementation is the same as that of the first encoder model. However, while the first encoder network (used for prognosis prediction) is optimized during training to avoid overfitting and generalize well, the second encoder model is optimized during training to compress clinical information into smaller vectors.

[0030] Although not shown in Figure 3, the patient similarity model 30 includes a comparison function to compare a vector representation of modified medical history information for a patient with vector representations of multiple medical history instances in order to find at least one most similar medical history instance. The comparison may include calculating the cosine distance and / or another distance measure between the vector representations.

[0031] The data from which medical history instances are extracted may be the same as the training data used to train prognosis prediction models (and patient similarity models).

[0032] The prognosis prediction model 20 is trained using the model to predict a known prognosis for a given subject based on training data, and then the weights of the elements in the prognosis prediction model 20 are updated based on the result of comparing the predicted prognosis with the known prognosis. This is done over a predetermined number of iterations and / or until an error threshold is met (based on the error between the predicted prognosis and the true prognosis). Training may include not only predicting prognosis (post-treatment events) but also predicting subsequent events in general.

[0033] The patient similarity model 20, or more specifically the second encoder model, is trained using the elements in the L-shaped box 39 in Figure 3. Following the generation of a vector representation of a medical history instance from the training data, the decoder model 38 is used to generate a reconstructed medical history instance (in the form of a series of events e as shown in the figure), and the weights of the second encoder model are adjusted based on the results of a comparison between the reconstructed medical history instance and the actual (ground truth) medical history instance. In a particular implementation, the decoder model 38 includes a long-term short-term memory (LSTM network).

[0034] In some implementations, training the second encoder model involves using an auxiliary decoder model 36 that generates an auxiliary reconstructed medical history instance that additionally includes at least one predicted next event based on a vector representation. This auxiliary reconstructed medical history instance is compared with the auxiliary medical history instance from among the multiple medical history instances, which include the events of the medical history instance and at least one next event. In this case, tuning the weights of the second encoder model involves tuning based on this comparison (auxiliary comparison). In certain implementations, the auxiliary decoder model 36 includes a long short-term memory (LSTM network).

[0035] When training the prognosis prediction model 20, the first encoder model may be initialized using the weights of the second encoder model after training (but not necessarily frozen). In other words, the initial form of the first encoder model before training the prognosis prediction model 20 may be the same as the second encoder model after training. The use of an auxiliary decoder model 36 may help train the second encoder model to generate vector representations in a way that leads to predictions of the next event, which may be useful, for example, if the weights of the first encoder are initialized using the weights of the second encoder model after training.

[0036] Pooling layers 24 and 34 may use weighted combinations of contextualized embeddings, and the relevant weights may be adjusted during the training of models 20 and 30.

[0037] Figure 4 is a flowchart of the method disclosed herein. This method includes steps S30, S40, S50, and S60.

[0038] Step S30 includes predicting the prognosis of the patient. Specifically, step S30 includes using an ML model to predict the patient's prognosis based on medical history information, where the medical history information includes the treatment instructions given (proposed) to the patient, and the ML model is trained on training data containing medical history information for multiple subjects. The ML model referred to here can be considered the prognosis prediction model described above with reference to Figure 2, and any explanation therein can be applied here, and vice versa.

[0039] Step S40 includes comparing the modified medical history information with medical history instances. Specifically, step S40 includes comparing modified medical history information about the subject, including medical history information and predicted prognosis, with multiple medical history instances, each medical history instance including medical history information from training data and relating to a subject (and time period) among the multiple subjects. The comparison may include the use of the patient similarity model 30 described above with reference to Figure 3, any description thereof may be applied here, and vice versa.

[0040] Step S50 includes selecting at least one most similar medical history instance; that is, step S50 includes selecting at least one medical history instance that is most similar to the modified medical history information relating to the patient in question, based on the comparison.

[0041] Step S60 includes outputting information indicating the predicted prognosis and at least one selected medical history instance.

[0042] For example, as shown in Figure 4, according to the first embodiment, the Specified Computer Implementation Method is disclosed, which includes the steps of: predicting the prognosis for a subject based on medical history information relating to the subject using an ML model, wherein the medical history information relating to the subject includes treatment instructions given (proposed) to the subject, and the ML model is trained with training data including medical history information relating to multiple subjects; comparing modified medical history information relating to the subject, including medical history information relating to the subject and a predicted prognosis, with multiple medical history instances, wherein each medical history instance includes medical history information relating to a subject (and time period) among the multiple subjects from the training data; selecting at least one medical history instance that is most similar to the modified medical history information relating to the subject based on the comparison; and outputting information indicating the predicted prognosis and the selected at least one medical history instance.

[0043] The ML model may include a first encoder model for encoding medical history information about a target patient in order to generate a contextualized embedding of the target; and a predictive model for predicting the prognosis based on the contextualized embedding of the target.

[0044] Comparing modified medical history information for a target patient with multiple medical history instances may include the steps of: encoding each medical history instance using a second encoder model and generating a contextualized embedding for each medical history instance; generating a vector representation of each medical history instance based on the contextualized embedding for each medical history instance; encoding modified medical history information for a target patient using a second encoder model and generating a contextualized embedding for the modified medical history information for the target patient; generating a vector representation of the modified medical history information for the target patient based on the contextualized embedding for the modified medical history information for the target patient; and comparing the vector representation of the modified medical history information for the target patient with the vector representation of the medical history instances.

[0045] Comparing a vector representation of the modified medical history information for a target patient with a vector representation of the medical history instance may involve calculating the cosine distance between the vector representation of the modified medical history information for the target patient and the vector representation of the medical history instance for each of the medical history instance vector representations.

[0046] The first encoder model may include multiple self-attention layers.

[0047] The second encoder model may include multiple self-attention layers.

[0048] The first encoder model may include a first set of self-attention layers, and the second encoder model may include a second set of self-attention layers.

[0049] Encoding medical history information for a target patient may involve supplementing information about each event within the patient's medical history with the time information of that event, and then applying a self-attention layer to the supplemented information to generate a contextualized embedding of the target corresponding to each event.

[0050] The predictive model may include a first neural network for predicting the expected prognostic event time based on the contextualized embedding of the subject, and a second neural network for predicting the prognosis based on the contextualized embedding of the subject and the expected event time.

[0051] The second neural network may include multiple classifiers that can handle multiple possible events / diagnoses.

[0052] The predictive model may include an attention network for generating summary vectors based on contextualized embeddings of the target and on event times predicted by a first neural network.

[0053] Predicting the prognosis based on contextualized embeddings and predicted event times of the subject using a second neural network may include predicting the prognosis based on summary vectors.

[0054] The prediction model may include an attention network for generating a summary vector based on a contextualized embedding of the subject and on the event time predicted by a first neural network, and predicting the prognosis by a second neural network based on the contextualized embedding of the subject and the predicted event time may include predicting the prognosis based on the summary vector.

[0055] The first encoder model may be the same as the second encoder model.

[0056] The medical history information of the patient may include multiple / series of events.

[0057] Medical history information for multiple subjects may include multiple / series of events for each of the subjects.

[0058] The patient's medical history information may include information indicating any of the following about the patient: at least one diagnosis; at least one treatment; at least one report of at least one symptom; at least one physiological measurement; at least one surgery and / or surgical procedure; at least one administered therapy; at least one medical procedure; sex assigned at birth; ethnicity; at least one (germline) variant.

[0059] Medical history information for multiple subjects may include, for each of the multiple subjects, information indicating any of the following about the subject: at least one diagnosis; at least one treatment; at least one report of at least one symptom; at least one physiological measurement; at least one surgery and / or surgical procedure; at least one administered therapy; at least one medical procedure; sex assigned at birth; ethnicity; at least one (germline) variant.

[0060] Medical history information for the target patient may include information from the target patient's electronic health record.

[0061] Medical history information for multiple subjects may include information from each subject's electronic health record.

[0062] Medical history information for the target patient may include time-series data.

[0063] Medical history information for multiple subjects may include time-series data.

[0064] The computer implementation method may include outputting information that shows medical history information about a subject corresponding to at least one selected medical history instance.

[0065] The predicted prognosis may include the predicted diagnosis of a condition or disease.

[0066] A computer-implemented method may include the steps of: using an ML model to predict a further prognosis for a patient based on further medical history information relating to the patient, wherein the further medical history information relating to the patient includes instructions for further (proposed) treatments to be given to the patient; comparing further modified medical history information relating to the patient, including the further medical history information relating to the patient and the further predicted prognosis, with the plurality of medical history instances; selecting at least one further medical history instance that is most similar to the further modified medical history information relating to the patient based on the comparison; and outputting further information indicating the further predicted prognosis and the selected at least one further medical history instance.

[0067] A computer-implemented method includes the steps of: predicting a first prognosis for a patient based on first medical history information relating to the patient using an ML model, wherein the first medical history information relating to the patient includes first (proposed) treatment instructions given to the patient; comparing a first modified medical history information relating to the patient, which includes the first medical history information relating to the patient and a first predicted prognosis, with a plurality of medical history instances; selecting at least one first medical history instance that is most similar to the first modified medical history information relating to the patient based on the comparison; and predicting a second prognosis for the patient based on second medical history information relating to the patient using an ML model. The steps may include: a step in which the second medical history information relating to the patient includes a second (proposed) treatment instruction given to the patient; a step in which a second modified medical history information relating to the patient, including the second medical history information relating to the patient and the second predicted prognosis, is compared with the plurality of medical history instances; a step in which, based on the comparison, at least one second medical history instance is selected that is most similar to the second modified medical history information relating to the patient; and a step in which first information indicating a first predicted prognosis and the selected first at least one medical history instance is output, and second information indicating a second predicted prognosis and the selected second at least one medical history instance.

[0068] A second encoder model may be trained according to a second encoder model training process, which includes: encoding a medical history instance using the second encoder model to generate a contextualized embedding of the medical history instance; generating a vector representation of the medical history instance based on the contextualized embedding of the medical history instance; generating a reconstructed medical history instance based on the vector representation using a decoder model; comparing the medical history instance with the reconstructed medical history instance; and adjusting at least one weight of the second encoder model based on the comparison.

[0069] The second encoder model training process may sequentially include the steps of: encoding the medical history instances using the second encoder model to generate contextualized embeddings of the medical history instances; generating vector representations of the medical history instances based on the contextualized embeddings of the medical history instances; generating reconstructed medical history instances based on the vector representations using a decoder model; comparing the reconstructed medical history instances with the medical history instances; and adjusting at least one weight of the second encoder model based on the comparison.

[0070] A second encoder model training process may include sequentially repeating the above steps for a predetermined number of iterations and / or until an error threshold is met.

[0071] A second encoder model training process may further include the steps of: generating an auxiliary reconstructed medical history instance containing at least one predicted event using an auxiliary decoder model; and comparing the auxiliary reconstructed medical history instance with an auxiliary medical history instance from among the plurality of medical history instances containing the events of the medical history instance and at least one subsequent event, wherein adjusting at least one weight of the second encoder model includes adjusting at least one weight of the second encoder model based on the comparison of the auxiliary reconstructed medical history instance with the auxiliary medical history instance.

[0072] Decoder models may include recurrent neural networks.

[0073] Auxiliary decoder models may include recurrent neural networks.

[0074] Decoder models may include long-term short-term memory (LSTM) networks.

[0075] The auxiliary decoder model may include an LSTM network.

[0076] The ML model may be trained according to an ML model training process, which includes the steps of: using the ML model to predict the prognosis of one of the plurality of subjects based on training data including medical history information of the subject, wherein the medical history information of the subject includes treatment instructions given to the subject; comparing the predicted prognosis with the correct prognosis in the training data; and adjusting at least one weight of the ML model based on the comparison.

[0077] The ML model training process may include, for a plurality of subjects, using an ML model to predict the prognosis for one of the plurality of subjects based on training data including medical history information of the subject, wherein the medical history information of the subject includes treatment instructions given to the subject; comparing the predicted prognosis with the correct prognosis in the training data; and adjusting at least one weight of the ML model based on the comparison, these steps are repeated sequentially.

[0078] The ML model training process may include sequentially repeating these steps over a predetermined number of iterations and / or until an error threshold is met.

[0079] The computer implementation may include training the ML model according to an ML model training process (before predicting the prognosis).

[0080] The computer implementation method may include training a second encoder model according to a second encoder model training process (before comparing the modified medical history information for the target patient with multiple medical history instances).

[0081] The computer implementation may include training a second encoder model before training the ML model, and training the ML model according to the ML model training process includes initializing the weights of the first encoder model using the weights of the second encoder model (after training the second encoder model).

[0082] The computer-implemented method may include a step of training an ML model according to an ML model training process (before predicting prognosis); and a step of training a second encoder model according to a second encoder model training process (before comparing modified medical history information about the patient with multiple medical history instances), wherein the computer-implemented method includes training the second encoder model before training the ML model, and training the ML model according to the ML model training process includes initializing the weights of the first encoder model using the weights of the second encoder model (after training the second encoder model).

[0083] A second encoder model may be trained according to a second encoder model training process, the second encoder model training process comprising: using the second encoder model to encode a medical history instance to generate a contextualized embedding of the medical history instance; generating a vector representation of the medical history instance based on the contextualized embedding of the medical history instance; using a decoder model to generate a reconstructed medical history instance based on the vector representation; comparing the medical history instance with the reconstructed medical history instance; and adjusting at least one weight of the second encoder model based on the comparison, wherein the ML model is trained according to an ML model training process, the ML model training process comprising: using the ML model to predict the prognosis of one of a plurality of subjects based on training data including medical history information relating to the subject, wherein the medical history information relating to the subject includes treatment instructions given to the subject; comparing the predicted prognosis with a ground truth prognosis in the training data; and adjusting at least one weight of the ML model based on the comparison.

[0084] The computer implementation method may include a step of training a second encoder model according to a second encoder model training process; and a step of training an ML model according to an ML model training process after training the second encoder model, wherein training the ML model according to the ML model training process includes (after training the second encoder model) initializing the weights of the first encoder model using the weights of the second encoder model.

[0085] Outputting information can include displaying information on a screen.

[0086] Treatment may include drugs, therapies, and medical procedures.

[0087] A computer-implemented method includes: using at least one first processor and / or computing device to encode each medical history instance using a second encoder model and generate a vector representation of each medical history instance; using the second encoder model to encode modified medical history information and generate a vector representation of the modified medical history information; and using at least one second processor and / or computing device to compare the vector representation of the modified medical history information relating to the patient in question with the vector representations of the medical history instances, wherein the at least one first processor and / or computing device has greater processing power and / or computing load than the at least one second processor and / or computing device.

[0088] The use of the first at least one processor and / or computing device may be partially concurrent with the use of the second at least one processor and / or computing device.

[0089] According to a second embodiment, a computer program is disclosed herein that, when executed on a computer, causes the computer to perform a method. The method includes: a step of using a machine learning (ML) model to predict the prognosis for a subject based on medical history information relating to the subject, wherein the medical history information relating to the subject includes treatment instructions given (proposed) to the subject, and the ML model is trained on training data including medical history information relating to multiple subjects; a step of comparing modified medical history information relating to the subject, including medical history information relating to the subject and a predicted prognosis, with multiple medical history instances, each medical history instance including medical history information from the training data and relating to a subject (and a time period) among the multiple subjects; a step of selecting at least one medical history instance relating to the subject that is most similar to the modified medical history information relating to the subject based on the comparison; and a step of outputting information indicating the predicted prognosis and the selected at least one medical history instance.

[0090] According to a third embodiment, an information processing device is disclosed herein, comprising a memory and a processor connected to the memory. The processor is configured to perform the following steps: predict a prognosis for a patient based on medical history information relating to the patient, wherein the medical history information relating to the patient includes treatment instructions given (proposed) to the patient, and the ML model is trained on training data including medical history information relating to multiple subjects; compare modified medical history information relating to the patient, including medical history information relating to the patient and a predicted prognosis, with multiple medical history instances, each medical history instance including medical history information from the training data and relating to a subject (and a certain time period) among the multiple subjects; select at least one medical history instance relating to the patient that is most similar to the modified medical history information relating to the patient based on the comparison; and output information indicating the predicted prognosis and the selected at least one medical history instance.

[0091] According to a fourth embodiment, the Specified herein discloses a system comprising multiple computing devices configured to perform the following steps: predict a prognosis for a patient based on medical history information relating to the patient using a machine learning (ML) model, wherein the medical history information relating to the patient includes treatment instructions given (proposed) to the patient, and the ML model is trained on training data including medical history information relating to multiple subjects; compare modified medical history information relating to the patient, including medical history information relating to the patient and a predicted prognosis, with multiple medical history instances, each medical history instance including medical history information from the training data and relating to a subject (and a certain time period) among the multiple subjects; select at least one medical history instance relating to the patient that is most similar to the modified medical history information relating to the patient based on the comparison; and output information indicating the predicted prognosis and the selected at least one medical history instance.

[0092] Comparing modified medical history information for a target patient with multiple medical history instances includes: a step of encoding each medical history instance using a second encoder model to generate a contextualized embedding for each medical history instance; a step of generating a vector representation of each medical history instance based on the contextualized embedding for each medical history instance; a step of encoding modified medical history information for a target patient using a second encoder model to generate a contextualized embedding for the modified medical history information for the target patient; a step of generating a vector representation of the modified medical history information for the target patient based on the contextualized embedding for the modified medical history information for the target patient; and a step of comparing the vector representation of the modified medical history information for the target patient with the vector representations of the medical history instances. Furthermore, the plurality of computing devices include: a first processor or computing device configured to encode each medical history instance using a second encoder model, generate a vector representation of each medical history instance, encode modified medical history information using the second encoder model, and generate a vector representation of the modified medical history information; and a second processor or computing device configured to compare the vector representation of the modified medical history information relating to the patient in question with the vector representations of the medical history instances, wherein the first processor or computing device has greater processing power and / or computing load than the second processor or computing device.

[0093] The use of the first at least one processor or computing device may be partially simultaneous with the use of the second at least one processor or computing device.

[0094] Features relating to any aspect / embodiment may apply to any other aspect / embodiment.

[0095] Figure 5 is a flowchart showing a training method that can be used to train an ML model and a second encoder model. This method includes steps S11, S12, S13, S14, S15, S21, S22, S23, S24, and S25.

[0096] Step S11 includes encoding the medical history instances (from the training data) using a second encoder model to generate contextualized embeddings. Step S12 includes generating a vector representation of the medical history instances based on the contextualized embeddings. Step S13 includes generating reconstructed medical history instances using a decoder model.

[0097] Step S14 includes comparing the reconstructed medical history instance with the (original) medical history instance. Step S15 includes adjusting at least one weight of the second encoder model based on the comparison.

[0098] Steps S11-S15 may be repeated (using different medical history instances from the training data) over a given number of iterations and / or until an error threshold is met.

[0099] Step S21 includes initializing the weights of the first encoder model based on the trained second encoder model, for example, using the weights of the trained second encoder model. Step S22 includes using the ML model to predict the prognosis for a subject (based on a certain treatment) using medical history instances from the training data. Step S23 includes comparing the predicted prognosis with the ground truth prognosis (i.e., the actual prognosis from the training data). Step S24 includes adjusting at least one weight of the ML model based on the comparison.

[0100] Steps S22-S24 may be repeated (using different medical history instances from the training data) over a given number of iterations and / or until an error threshold is met.

[0101] The training method shown in Figure 5 is not mandatory, and the ML model and the second encoder network may be trained in other ways. The stages of the training method shown in Figure 5 correspond to the method stages described above with respect to Figure 4, for example, and the descriptions therein apply here, and vice versa. Stages S11-S15 may be considered as the second encoder model training process, and / or stages S22-S24 (or stages S21-S24) may be considered as the ML model training process. The methods disclosed herein include the method of Figure 5 and the method of Figure 4 thereafter. [Examples]

[0102] This section describes an example of predicting prognosis and identifying similar patients using a prognosis prediction model 20 and a patient similarity model 30. The EHR data (medical history information) of the target patients may be represented in a JSON file, and an example is shown below (not exhaustive). [Table 1] [subject, sex, visits, days, events, findings, negative, positive, drugs, procedures]

[0103] This JSON example shows some data about the patient being entered into models 20 and 30. A short example is used for brevity. The data includes the sex assigned at birth, under the key "sex". In practice, this information is included as an event along with time information indicating the time of birth (i.e., "day 0"). Similarly, other information, such as germline mutations or ethnicity, can be included in the same way.

[0104] Under the key "visits," a sequence of event sets is shown. This represents the points in time when the patient visited a doctor or hospital and several events were recorded. However, it may also include some events recorded by the device or the patient themselves. Each visit includes the key "days," which represents the number of days since the patient's birth. In this example, the patient was 27823 days old, or approximately 76 years and 3 months, on the first visit date shown above. The next visit occurs approximately 2 months (59 days) later. The patient's age is expressed in days to avoid dates altogether. The "events" key contains all events recorded on that date. Under "findings," diagnosed diseases and symptoms are listed as "positive." The "negative" set contains conditions that the doctor determined to be absent. For example, if it was found or determined that the patient did not have a fever, this would be listed under "negative." Different events are referenced using the ID recorded in the relevant database. For example, 12908 could represent "acute exacerbated systolic-diastolic heart failure." Past treatments are also recorded under "drugs" and "procedures" above.

[0105] The last item also includes the proposed procedure (insertion of a single-chamber device and several medications) in underlined italics. The above input example to the system (prognosis prediction model 20 and patient similarity model 30) generates predicted prognosis and similar patient information according to the output shown below in JSON format. [Table 2] [predictions, days, predicted, similar]

[0106] As described above, the system generates predictions for future dates represented by "days". Each predicted event is represented by a pair (probability, ID). For each predicted event date, predictions with a probability exceeding a certain threshold (e.g., 0.30) are considered. When an event exceeds the threshold, a new set of predictions may be included. From the above, it is clear that event 7587 has an increased probability for later dates. The IDs of some patients with similar records are also recorded. "Similar" patients not only have similar past records, but for example, they received this treatment and recorded a response similar to what was predicted. As the predictions progress into the future, the confidence of the predictions decreases, and the number of pieces of evidence (i.e., other similar patients) recorded decreases.

[0107] In this embodiment, the ID of a similar patient is output as information indicating at least one most similar medical history instance. Alternatively or additionally, information that specifically identifies a medical history instance, such as the ID of the most similar medical history instance (event sequence), or the event sequence itself, may be output.

[0108] By modifying the proposed treatment (underlined and italicized in the input list), different predictions and similar medical history instances can be generated, thus allowing different treatments to be compared and the most appropriate / best treatment to be selected, for example, by a clinician.

[0109] Certain aspects disclosed herein may be considered a Clinical Decision Support (CDS) system to help compare multiple candidate treatments when those treatments are applied to a specific patient described by an EHR. One useful feature is a mechanism for referencing evidence from the most similar known cases that received similar treatments.

[0110] Certain aspects disclosed herein may help provide explainability of predicted prognoses by providing information that shows similar medical history instances / subjects. Since predictions are not 100 percent accurate, and predictions made by machine learning systems may influence the treatment administered to patients, it may be useful for clinicians to verify the usefulness of the predictions.

[0111] The purpose / goal of the aspects disclosed herein is a system to assist physicians in treatment decision-making by predicting the outcomes of different candidate treatments supported by clinical evidence. The aspects disclosed herein go beyond merely providing predictions about prognosis by including a mechanism that responds to queried patient cases and also provides supporting historical clinical cases. The aspects disclosed herein may include connections to EHR databases used for both (1) training predictive models and (2) supporting evidence for predictions.

[0112] In other words, due to the complexity of the prognosis prediction task, predictions may not be guaranteed with a large margin of confidence, and therefore the final decision on treatment may remain with the clinician. Because of the responsibility of such a task, clinicians may naturally doubt the system's predictions and will not prescribe treatments they are unfamiliar with. Therefore, the usefulness of such prognosis predictions may be limited without providing evidence for the predictions (i.e., the most similar medical history instance / subject). When determining the best treatment for a patient according to the aspects disclosed herein, clinicians can study cases of similar subjects to improve their prognostic assessment.

[0113] The aspects disclosed herein may be considered general (in the sense that they are generally applicable to different treatments / outcomes), accurate, and simultaneously explainable. Conventional methodologies for prognosis prediction can use past data from other subjects as training data and then predict prognosis, but this training data is subsequently discarded, and predictions are provided without explicit evidence or explanation. The aspects disclosed herein may use the training data even after training to augment output predictions with supporting evidence.

[0114] Figure 6 is a block diagram of an information processing device 90 or computing device 90, such as a data storage server, which may be used to embody the present invention, implement some or all of the operations of a method of embodying the present invention, and perform some or all of the tasks of the apparatus of the embodiment. The computing device 90 may be used to implement any of the method steps described above, for example, steps S2 and S3 and any of steps S30 to S60 and steps S11 to S24, and / or any operations of models 20 and 30.

[0115] The computing device 90 comprises a processor 993 and memory 994. Optionally, the computing device also includes a network interface 997 for communicating with other such computing devices, for example, other computing devices in embodiments of the present invention. Optionally, the computing device also includes one or more input mechanisms, such as a keyboard and a mouse 996, and one or more display units, such as a monitor 995. These elements can facilitate user interaction. Each component is connectable to one another via a bus 992.

[0116] Memory 994 may include computer-readable media, and this term may refer to one or more media configured to carry computer-executable instructions (e.g., a centralized or distributed database and / or associated caches and servers). Computer-executable instructions may include, for example, instructions and data that are accessible by a computer (e.g., one or more processors) and cause the computer to perform one or more functions or operations. For example, computer-executable instructions may include instructions for implementing the methods disclosed herein or any method steps disclosed herein, e.g., steps S2 and S3 and any of steps S30-S60 and S11-S24, and / or any operations of models 20 and 30. Thus, the term “computer-readable storage medium” may also include any medium capable of storing, encoding, or carrying a set of instructions for machine execution and causing a machine to execute any one or more of the method steps of this disclosure. Thus, the term “computer-readable storage medium” may be interpreted to include, but not limited to, solid-state memory, optical media, and magnetic media. Such computer-readable media may include, but are not limited to, non-temporary computer-readable storage media, including random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), compact disk read-only memory (CD-ROM) or other optical disk storage, magnetic disk storage or other magnetic storage devices, and flash memory devices (e.g., solid-state memory devices).

[0117] The processor 993 is configured to control the computing device and perform processing operations, for example, by executing computer program code stored in memory 994 to implement any of the steps of the method described herein, for example, steps S2 and S3 and any of steps S30-S60 and steps S11-S24, and / or any operation of Models 20 and 30. Memory 994 stores data read and written by the processor 993, and may store medical history information and / or EHR data and / or proposed treatment information and / or information indicating at least one predicted prognosis and / or event information and / or augmented event information and / or contextualized embeddings and / or vector representations and / or reconstructed medical history instances and / or auxiliary reconstructed medical history instances and / or predictions and / or similarity information and / or any network weights as described above and / or input data and / or other data, and / or programs for performing any of the method steps described above. As referred to herein, the processor may include one or more general-purpose processing devices such as a microprocessor, a central processing unit, etc. The processor may include a complex instruction set computing (CISC) microprocessor, a reduced instruction set computing (RISC) microprocessor, a very long instruction word (VLIW) microprocessor, or a processor implementing a processor or combination of instruction sets. The processor may also include one or more dedicated processing devices such as application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), digital signal processors (DSPs), and network processors. In one or more embodiments, the processor is configured to execute the operations and instructions for performing the operations described herein. The processor 993 may be considered to include any of the modules described above, for example, either Model 20 or 30. Any operations described as being performed by a module may be performed by a computer, for example, by the processor 993, as a method.

[0118] The display unit 995 may display a representation of data stored by the computing device, such as medical history information and / or EHR data and / or suggested treatment information and / or information indicating at least one predicted prognosis and / or event information and / or augmented event information and / or contextualized embeddings and / or vector representations and / or reconstructed medical history instances and / or auxiliary reconstructed medical history instances and / or predictions and / or similarity information and / or any network weights as described above and / or the input data and / or other data and / or a GUI window and / or interactive representation and / or representation of any other output as described above, which allows the user to interact with the device 90, for example by drag-and-drop or selection dialogues, and may also display a cursor and dialog boxes and screens which allow interaction between the user and programs and data stored in the computing device. The display may be used to display information indicating at least one predicted prognosis and at least one most similar medical history instance. The input mechanism 996 may allow the user to input data and commands into the computing device, such as allowing the user to input any user input as described above.

[0119] The network interface (network I / F) 997 may be connected to a network such as the Internet and may be connected to other such computing devices via the network. The network I / F 997 may control data input / output to and from other devices via the network. Other peripheral devices such as microphones, speakers, printers, power supply units, fans, cases, scanners, and trackballs may be included in the computing device.

[0120] A method of embodying the present invention may be performed on a computing device / apparatus 90 as shown in Figure 6. Such a computing device does not need to have all the components shown in Figure 6, but may consist of a subset of those components. For example, apparatus 90 may comprise a processor 993 and memory 994 connected to the processor 993. Alternatively, apparatus 90 may comprise a processor 993, memory 994 connected to the processor 993, and a display 995. A method of embodying the present invention may be performed by a single computing device that communicates with one or more data storage servers over a network. The computing device may be data storage itself that stores at least a portion of the data.

[0121] A method of embodying the present invention may be performed by a plurality of computing devices working together. One or more of the plurality of computing devices may be a data storage server that stores at least a portion of the data. Device 90 may be one of a plurality of similarly configured computing devices, which together are configured to perform any method step described herein, for example, any of steps S2 and S3 and steps S30-S60 and steps S11-S24, and / or any operation of models 20 and 30. For example, at least one of the computing devices may encode each medical history instance using a second encoder model and generate a vector representation of each medical history instance, encode modified medical history information using the second encoder model and generate a vector representation of the modified medical history information, and at least one other of the computing devices may compare the vector representation of the modified medical history information relating to the patient in question with the vector representations of the medical history instances. The processors and / or computing devices may be heterogeneous, for example, in terms of processing power.

[0122] The present invention may be implemented in digital electronic circuits, or in computer hardware, firmware, software, or a combination thereof. The present invention may be implemented as a computer program or computer program product, i.e., a computer program tangibly embodied in a non-temporary information carrier, such as a machine-readable storage device, or in a propagating signal, for execution by or control of the operation of one or more hardware modules.

[0123] Computer programs may be in the form of standalone programs, computer program portions, or multiple computer programs, and may be written in any form of programming language, including compiled or interpreted languages, and may be deployed in any form, such as as standalone programs or as modules, components, subroutines, or other units suitable for use in a data processing environment. Computer programs may be deployed to run on one module or on multiple modules at one site, or they may be distributed across multiple sites and interconnected by a communication network.

[0124] The method steps of the present invention (for example, steps S2 and S3, steps S30 to S60, any of steps S11 to S24, and / or any of models 20 and 30) may be executed by one or more programmable processors that execute a computer program for performing the functions of the present invention by acting on input data and generating an output. The apparatus of the present invention may be implemented as programmed hardware or as a dedicated logic circuit including, for example, an FPGA (Field-Programmable Gate Array) or an ASIC (Application-Specific Integrated Circuit).

[0125] Processors suitable for executing computer programs include, for example, both general-purpose and dedicated microprocessors, and any one or more processors in any type of digital computer. Generally, a processor receives instructions and data from read-only memory or random-access memory or both. An essential element of a computer is a processor for executing instructions, coupled with one or more memory devices for storing instructions and data.

[0126] The above-described embodiments of the present invention may be used, advantageously, independently of any other embodiment, or in any feasible combination with one or more other embodiments.

[0127] This disclosure extends to the following statements:

[0128] [S1] A computer-implemented method comprising: a step of using an ML model to predict the prognosis for a target patient based on medical history information relating to the target patient, wherein the medical history information relating to the target patient includes treatment instructions given to the target patient, and the ML model is trained with training data comprising medical history information relating to multiple subjects; a step of comparing modified medical history information relating to the target patient, including the medical history information relating to the target patient and the predicted prognosis, with multiple medical history instances, wherein each medical history instance comprises medical history information relating to a subject (and a time period) among the multiple subjects from the training data; a step of selecting at least one medical history instance that is most similar to the modified medical history information relating to the target patient based on the comparison; and a step of outputting information indicating the predicted prognosis and the selected at least one medical history instance. [S2] The computer-implemented method according to Statement S1, wherein the ML model includes a first encoder model for encoding the medical history information relating to the subject patient to generate a contextualized embedding of the subject; and a predictive model for predicting the prognosis based on the contextualized embedding of the subject. [S3] A computer-implemented method according to Statement S1 or S2 for comparing the modified medical history information relating to the subject patient with the plurality of medical history instances, comprising the steps of: encoding each medical history instance using a second encoder model to generate a contextualized embedding of each medical history instance; generating a vector representation of each medical history instance based on the contextualized embedding of that medical history instance; encoding the modified medical history information relating to the subject patient using the second encoder model to generate a contextualized embedding of the modified medical history information relating to the subject patient; generating a vector representation of the modified medical history information relating to the subject patient based on the contextualized embedding of the modified medical history information relating to the subject patient; and comparing the vector representation of the modified medical history information relating to the subject patient with the vector representation of the medical history instance. [S4] The computer-implemented method according to Statement S3, wherein comparing the vector representation of the modified medical history information relating to the subject patient with the vector representation of the medical history instance includes calculating the cosine distance between the vector representation of the modified medical history information relating to the subject patient and the vector representation of the medical history instance for each of the vector representations of the medical history instance. [S5] The computer implementation method according to any one of statements S2 to S4, wherein the first encoder model includes a plurality of self-attention layers. [S6] The computer implementation method according to any one of statements S3 to S5, wherein the second encoder model includes multiple self-attention layers. [S7] The computer implementation method according to any one of statements S3 to S5, wherein the first encoder model includes a first plurality of self-attention layers, and the second encoder model includes a second plurality of self-attention layers. [S8] A computer-implemented method according to any one of statements S2 to S7, wherein encoding the medical history information relating to the subject patient includes supplementing the information relating to each event in the medical history information relating to the subject patient with time information of that event, and applying the self-attention layer to the supplemented information to generate a contextualized embedding of the subject corresponding to each event. [S9] The computer-implemented method according to any one of statements S2 to S8, wherein the prediction model includes a first neural network for predicting a predicted prognostic event time based on a contextualized embedding of the subject, and a second neural network for predicting the prognosis based on the contextualized embedding of the subject and the predicted event time. [S10] The second neural network is a computer implementation as described in Statement S9, comprising multiple classifiers corresponding to multiple possible events / diagnoses. [S11] The computer-implemented method according to any one of statements S2 to S10, wherein the prediction model includes an attention network for generating a summary vector based on the contextualized embedding of the target and the event time predicted by the first neural network. [S12] A computer-implemented method according to any one of statements S9 to S11, wherein the second neural network predicts the prognosis based on the contextualized embedding and predicted event time of the subject, and predicts the prognosis based on the summary vector. [S13] The computer-implemented method according to any one of statements S2 to S10, wherein the prediction model includes an attention network for generating a summary vector based on a contextualized embedding of the subject and an event time predicted by the first neural network, and the second neural network predicts the prognosis based on the contextualized embedding of the subject and the predicted event time, and predicts the prognosis based on the summary vector. [S14] The computer implementation method described in any one of statements S1 to S13, wherein the first encoder model is the same as the second encoder model. [S15] The medical history information relating to the aforementioned patient is provided by a computer-implemented method as described in any one of statements 1 to 14, including multiple / series of events. [S16] The medical history information relating to the plurality of subjects is a computer-implemented method according to any one of statements 1 to 15, including a plurality of / series of events relating to each of the plurality of subjects. [S17] The computer-implemented method described in any one of statements S1 to S16, wherein the medical history information relating to the subject patient includes information relating to the subject patient that indicates any of the following: at least one diagnosis; at least one treatment; at least one report of at least one symptom; at least one physiological measurement; at least one surgery and / or surgical procedure; at least one administered therapy; at least one medical procedure; sex assigned at birth; ethnicity; or at least one (germline) mutation. [S18] The computer-implemented method described in any one of statements S1 to S17, wherein the medical history information relating to the plurality of subjects includes, for each of the plurality of subjects, information relating to that subject, namely: at least one diagnosis; at least one treatment; at least one report of at least one symptom; at least one physiological measurement; at least one surgery and / or surgical procedure; at least one therapy administered; at least one medical procedure; sex assigned at birth; ethnicity; and at least one (germline) mutation. [S19] The medical history information relating to the subject patient includes information from the electronic health record relating to the subject patient, provided that the method is computer-implemented as described in any one of statements S1 to S18. [S20] The medical history information relating to the plurality of subjects is a computer-implemented method according to any one of statements S1 to S19, wherein for each of the plurality of subjects, the information is taken from an electronic health record relating to that subject. [S21] The medical history information relating to the aforementioned patient is provided by a computer-implemented method described in any one of statements S1 to S20, including time-series data. [S22] The medical history information relating to the aforementioned multiple subjects is provided by a computer-implemented method described in any one of statements S1 to S21, including time-series data. [S23] The computer-implemented method described in any one of statements S1 to S22, comprising the step of outputting information indicating medical history information relating to a subject corresponding to at least one selected medical history instance. [S24] The predicted prognosis includes the predicted diagnosis of a condition or disease, as described in any one of statements S1 through S23, by a computer-implemented method. [S25] The computer-implemented method is the computer-implemented method according to any one of statements S1 to S24, comprising the steps of: using the ML model to predict a further prognosis for the subject patient based on further medical history information relating to the subject patient, wherein the further medical history information relating to the subject patient includes instructions for further (proposed) treatment to be given to the subject patient; comparing further modified medical history information relating to the subject patient, including the further medical history information relating to the subject patient and the further predicted prognosis, with the plurality of medical history instances; selecting, based on the comparison, at least one further medical history instance that is most similar to the further modified medical history information relating to the subject patient; and outputting further information indicating the further predicted prognosis and the selected at least one further medical history instance. [S26] The computer-implemented method includes the steps of: predicting a first prognosis for a subject based on first medical history information relating to the subject, using the ML model, wherein the first medical history information relating to the subject includes a first (proposed) treatment instruction given to the subject; comparing a first modified medical history information relating to the subject, including the first medical history information relating to the subject and a first predicted prognosis, with a plurality of medical history instances; selecting at least one first medical history instance that is most similar to the first modified medical history information relating to the subject, based on the comparison; and predicting a second prognosis for the subject based on second medical history information relating to the subject, using the ML model. A computer-implemented method according to any one of statements S1 to S25, comprising: a step of comparing a second modified medical history information relating to a patient, including a second (proposed) treatment instruction given to the patient; a step of selecting at least one second medical history instance relating to the patient, including the second medical history information relating to the patient and the second predicted prognosis; a step of selecting at least one second medical history instance relating to the patient that is most similar to the second modified medical history information relating to the patient, based on the comparison; and a step of outputting first information indicating the first predicted prognosis and the selected first at least one medical history instance, and second information indicating the second predicted prognosis and the selected second at least one medical history instance. [S27] The computer-implemented method according to any one of statements S1 to S26, wherein the second encoder model is trained according to a second encoder model training process, the second encoder model training process comprising: encoding a medical history instance using the second encoder model to generate a contextualized embedding of the medical history instance; generating a vector representation of the medical history instance based on the contextualized embedding of the medical history instance; generating a reconstructed medical history instance based on the vector representation using a decoder model; comparing the medical history instance with the reconstructed medical history instance; and adjusting at least one weight of the second encoder model based on the comparison. [S28] The computer-implemented method according to Statement S27, wherein the second encoder model training process comprises the steps of: encoding a medical history instance using the second encoder model to generate a contextualized embedding of the medical history instance; generating a vector representation of the medical history instance based on the contextualized embedding of the medical history instance; generating a reconstructed medical history instance based on the vector representation using a decoder model; comparing the reconstructed medical history instance with the medical history instance; and adjusting at least one weight of the second encoder model based on the comparison, repeated sequentially for a plurality of medical history instances. [S29] The computer-implemented method described in Statement S27 or S28, wherein the second encoder model training process comprises sequentially repeating the steps described above for a predetermined number of iterations and / or until an error threshold is met. [S30] The computer-implemented method according to any one of statements S27 to S29, wherein the second encoder model training process includes: generating an auxiliary reconstructed medical history instance including at least one predicted event using an auxiliary decoder model; and comparing the auxiliary reconstructed medical history instance with an auxiliary medical history instance from among the plurality of medical history instances including the events of the medical history instance and at least one subsequent event, and adjusting at least one weight of the second encoder model includes adjusting at least one weight of the second encoder model based on the comparison of the auxiliary reconstructed medical history instance with the auxiliary medical history instance. [S31] The decoder model is a computer implementation method described in any one of statements S27 to S30, including a recurrent neural network. [S32] The auxiliary decoder model is a computer implementation method described in any one of statements S27 to S31, including a recurrent neural network. [S33] The decoder model is a computer implementation method described in any one of statements S27 to S32, including a long short-term memory (LSTM) network. [S34] The auxiliary decoder model is a computer-implemented method described in any one of statements S27 to S33, including an LSTM network. [S35] The ML model is trained according to an ML model training process, the ML model training process comprising: a step of using the ML model to predict the prognosis of one of the plurality of subjects based on training data including medical history information relating to the subject, wherein the medical history information relating to the subject includes treatment instructions given to the subject; a step of comparing the predicted prognosis with the correct prognosis in the training data; and a step of adjusting at least one weight of the ML model based on the comparison, the computer-implemented method according to any one of statements S1 to S34. [S36] The computer-implemented method according to Statement S35, wherein the ML model training process includes the steps of: predicting, for a plurality of subjects, the prognosis for one of the plurality of subjects using the ML model, based on training data including medical history information relating to the subject, wherein the medical history information relating to the subject includes treatment instructions given to the subject; comparing the predicted prognosis with the correct prognosis in the training data; and adjusting at least one weight of the ML model based on the comparison. [S37] The computer-implemented method described in Statement S35 or S36, wherein the ML model training process comprises sequentially repeating the steps described above over a predetermined number of iterations and / or until an error threshold is met. [S38] The computer-implemented method is the computer-implemented method described in any one of statements S35 to S37, which includes training the ML model according to the ML model training process (before predicting the prognosis). [S39] The computer-implemented method is the computer-implemented method described in any one of statements S27 to S38, which includes training the second encoder model according to the second encoder model training process (before comparing the modified medical history information relating to the subject patient with the plurality of medical history instances). [S40] The computer-implemented method according to Statement S39, comprising training the second encoder model before training the ML model, and training the ML model according to the ML model training process, comprising (after training the second encoder model) initializing the weights of the first encoder model using the weights of the second encoder model. [S41] The computer-implemented method, as described in any one of statements S27 to S38, comprises the steps of: training the ML model according to the ML model training process (before predicting the prognosis); and training the second encoder model according to the second encoder model training process (before comparing the modified medical history information relating to the subject patient with the plurality of medical history instances), wherein the computer-implemented method comprises training the second encoder model before training the ML model, and training the ML model according to the ML model training process comprises initializing the weights of the first encoder model using the weights of the second encoder model (after training the second encoder model). [S42] The second encoder model is trained according to a second encoder model training process, the second encoder model training process comprising: using the second encoder model to encode a medical history instance to generate a contextualized embedding of the medical history instance; generating a vector representation of the medical history instance based on the contextualized embedding of the medical history instance; using a decoder model to generate a reconstructed medical history instance based on the vector representation; comparing the medical history instance with the reconstructed medical history instance; and based on the comparison, the second encoder model A computer-implemented method according to any one of statements S1 to S26, comprising: a step of adjusting at least one weight of, the ML model being trained according to an ML model training process, the ML model training process comprising: a step of using the ML model to predict a prognosis for one of the plurality of subjects based on training data including medical history information relating to the subject, wherein the medical history information relating to the subject includes treatment instructions given to the subject; a step of comparing the predicted prognosis with a correct prognosis in the training data; and a step of adjusting at least one weight of the ML model based on the comparison. [S43] The computer-implemented method according to Statement S42 comprises the steps of: training the second encoder model according to the second encoder model training process; and, after training the second encoder model, training the ML model according to the ML model training process, wherein training the ML model according to the ML model training process includes (after training the second encoder model) initializing the weights of the first encoder model using the weights of the second encoder model. [S44] Outputting the aforementioned information includes outputting the aforementioned information on a display, as described in any one of statements S1 to S43. [S45] The treatment is a computer-implemented method as described in any one of statements S1 to S44, comprising any of drugs, therapies, and medical procedures. [S46] The computer-implemented method is the computer-implemented method according to any one of statements S1 to S45, comprising: using at least one first processor and / or computing device to encode each medical history instance using a second encoder model and generate a vector representation of each medical history instance; using the second encoder model to encode modified medical history information and generate a vector representation of the modified medical history information; and using at least one second processor and / or computing device to compare the vector representation of the modified medical history information relating to the patient in question with the vector representations of the medical history instances, wherein the at least one first processor and / or computing device has greater processing power and / or computing load than the at least one second processor and / or computing device. [S47] The computer implementation method according to Statement S46, wherein the use of the first at least one processor and / or computing device is partially simultaneous with the use of the second at least one processor and / or computing device. [S48] A computer program, when executed on a computer, causes the computer to perform a method, the method comprising: a step of using a machine learning (ML) model to predict the prognosis for a subject based on medical history information relating to the subject, wherein the medical history information relating to the subject includes treatment instructions given (proposed) to the subject, and the ML model is trained on training data comprising medical history information relating to multiple subjects; a step of comparing modified medical history information relating to the subject, including the medical history information relating to the subject and the predicted prognosis, with multiple medical history instances, each medical history instance comprising medical history information from the training data and relating to a subject (and a time period) among the multiple subjects; a step of selecting at least one medical history instance relating to the subject that is most similar to the modified medical history information based on the comparison; and a step of outputting information indicating the predicted prognosis and the selected at least one medical history instance. [S49] An information processing device comprising a memory and a processor connected to the memory, wherein the processor is configured to perform the following steps: a step of using a machine learning (ML) model to predict the prognosis for a subject patient based on medical history information relating to the subject patient, wherein the medical history information relating to the subject patient includes treatment instructions given (proposed) to the subject patient, and the ML model is trained on training data including medical history information relating to a plurality of subjects; a step of comparing modified medical history information relating to the subject patient, including the medical history information relating to the subject patient and the predicted prognosis, with a plurality of medical history instances, wherein each medical history instance includes medical history information from the training data and relates to a subject (and a certain time period) among the plurality of subjects; a step of selecting at least one medical history instance relating to the subject patient that is most similar to the modified medical history information based on the comparison; and a step of outputting information indicating the predicted prognosis and the selected at least one medical history instance. [S50] A system comprising multiple computing devices configured to perform the following steps: a step of using a machine learning (ML) model to predict the prognosis for a target patient based on medical history information relating to the target patient, wherein the medical history information relating to the target patient includes treatment instructions given (proposed) to the target patient, and the ML model is trained on training data including medical history information relating to multiple subjects; a step of comparing modified medical history information relating to the target patient, including the medical history information relating to the target patient and the predicted prognosis, with multiple medical history instances, wherein each medical history instance includes medical history information from the training data and relates to a subject (and a time period) among the multiple subjects; a step of selecting at least one medical history instance relating to the target patient that is most similar to the modified medical history information based on the comparison; and a step of outputting information indicating the predicted prognosis and the selected at least one medical history instance. [S51] Comparing the modified medical history information relating to the subject patient with the multiple medical history instances includes: the steps of: encoding each medical history instance using a second encoder model to generate a contextualized embedding for each medical history instance; generating a vector representation of each medical history instance based on the contextualized embedding for the medical history instance; encoding the modified medical history information relating to the subject patient using the second encoder model to generate a contextualized embedding for the modified medical history information relating to the subject patient; generating a vector representation of the modified medical history information relating to the subject patient based on the contextualized embedding for the modified medical history information relating to the subject patient; and comparing the vector representation of the modified medical history information relating to the subject patient with the vector representations of the medical history instances. The system according to Statement S50, wherein the plurality of computing devices include: a first processor or computing device configured to encode each medical history instance using the second encoder model, generate a vector representation of each medical history instance, encode modified medical history information using the second encoder model, and generate a vector representation of the modified medical history information; and a second processor or computing device configured to compare the vector representation of the modified medical history information relating to the patient in question with the vector representations of the medical history instances, wherein the first processor or computing device has greater processing power and / or computing load than the second processor or computing device. [S52] The system according to Statement S51, wherein the use of the first at least one processor or computing device is partially simultaneous with the use of the second at least one processor or computing device. [Explanation of symbols]

[0129] 10. EHR for all patients 11. EHR for current patients 12A, B, C Proposed Treatments 12D Selected Treatments 13A, B, C Predicted Outcomes 14A, B, C Similar past cases of EHR S1 Treatment proposal S2 Prognostic Prediction Model S3 Patient Similarity Model S4 Clinical staff propose treatment 22A, B, C Self-Attention (Transformer) Layers 24 Pooling 28 Attention 26 Update Predictor <Regression> 28 Event Predictor (Classifier) 32A, B, C Self-Attention (Transformer) Layers 34. Pooling S30 Predicting the prognosis for target patients S40 Compare the corrected medical history information with the medical history instances. S50 Select at least one most similar medical history instance. S60 Output information showing the predicted prognosis and at least one selected medical history instance. S11 Use the second encoder model to encode the medical history instance and generate context embeddings. S12 Generate a vector representation based on context embedding. The S13 decoder model is used to generate reconstructed medical history instances. S14 Compare the reconstructed medical history instance with the medical history instance. S15 Adjust the weight of at least one of the second encoder models based on the above comparison. S21 Initialize the weights of the first encoder model based on the trained second encoder model. We use the S22 ML model to predict the prognosis for subjects. S23 Compare predicted prognosis with true prognosis S24 Adjust at least one weight of the ML model based on the above comparison. 993 processor 994 memory 995 Display 996 inputs 997 Network Interface

Claims

1. A step of using a machine learning (ML) model to predict the prognosis for a target patient based on medical history information of the target patient, wherein the medical history information of the target patient includes treatment instructions given to the target patient, and the ML model is trained on training data that includes medical history information of multiple subjects; A step of comparing modified medical history information relating to the subject patient, including the medical history information and predicted prognosis relating to the subject patient, with a plurality of medical history instances, wherein each medical history instance includes medical history information relating to one of the plurality of subjects from the training data; Based on the comparison, the step of selecting at least one medical history instance that is most similar to the modified medical history information relating to the subject patient; The process includes the step of outputting information indicating a predicted prognosis and at least one selected medical history instance, The method of computer implementation.

2. The ML model includes a first encoder model for encoding the medical history information relating to the target patient and generating a contextualized embedding of the target; A predictive model for predicting the prognosis based on the contextualized embedding of the subject, The computer-implemented method according to claim 1.

3. The computer-implemented method according to claim 2, wherein the prediction model includes a first neural network for predicting a predicted prognostic event time based on a contextualized embedding of the subject, and a second neural network for predicting the prognosis based on the contextualized embedding of the subject and the predicted event time.

4. The computer-implemented method according to claim 3, wherein the second neural network includes a plurality of classifiers corresponding to a plurality of possible events.

5. The prediction model includes an attention network for generating summary vectors based on the contextualized embedding of the target and the event time predicted by the first neural network, The computer-implemented method according to claim 4, wherein the second neural network predicts the prognosis based on the contextualized embedding and predicted event time of the subject, and further predicts the prognosis based on the summary vector.

6. The computer-implemented method according to any one of claims 2 to 5, wherein the first encoder model includes a plurality of self-attention layers.

7. Comparing the modified medical history information relating to the aforementioned patient with the multiple medical history instances is as follows: The process involves: encoding each medical history instance using a second encoder model to generate a contextualized embedding for each medical history instance; The steps include: generating a vector representation of each medical history instance based on the contextualized embedding of that medical history instance; The steps include: encoding the modified medical history information relating to the patient using the second encoder model to generate a contextualized embedding of the modified medical history information relating to the patient; A step of generating a vector representation of the modified medical history information relating to the subject patient, based on the contextualized embedding of the modified medical history information relating to the subject patient; A step of comparing the vector representation of the modified medical history information relating to the subject patient with the vector representation of the medical history instance, A computer-implemented method according to any one of claims 1 to 5, including the following:

8. The computer-implemented method according to claim 7, wherein the second encoder model includes a plurality of self-attention layers.

9. The second encoder model is trained according to a second encoder model training process, which is: The steps include: encoding a medical history instance using the second encoder model to generate a contextualized embedding of the medical history instance; The steps include: generating a vector representation of the medical history instance based on the contextualized embedding of the medical history instance; A step of generating a reconstructed medical history instance based on the vector representation using a decoder model; The step of comparing the aforementioned medical history instance with a reconstructed medical history instance; The steps include adjusting at least one weight of the second encoder model based on the comparison, The aforementioned ML model is trained according to an ML model training process, which is: A step of using the ML model to predict the prognosis of one of the multiple subjects based on training data including medical history information of the subject, wherein the medical history information of the subject includes treatment instructions given to the subject; The steps include: comparing the predicted prognosis with the correct prognosis in the training data; Based on the above comparison, the step of adjusting at least one weight of the ML model, including, The computer-implemented method according to claim 7.

10. A computer-implemented method according to claim 9, comprising training the second encoder model according to the second encoder model training process, and after training the second encoder model, training the ML model according to the ML model training process, wherein training the ML model according to the ML model training process includes initializing the weights of the first encoder model using the weights of the second encoder model.

11. Using at least one first processor and / or computing device, encode each medical history instance using the second encoder model and generate a vector representation of each medical history instance; encode the modified medical history information using the second encoder model and generate a vector representation of the modified medical history information; The process includes comparing a vector representation of the modified medical history information relating to the patient with a vector representation of the medical history instances using at least one second processor and / or computing device, The computer implementation method according to claim 7, wherein the first at least one processor and / or computing device has greater processing power and / or computing load than the second at least one processor and / or computing device.

12. A computer program, when executed on a computer, causes the computer to perform a method, wherein the method is: A step of using a machine learning (ML) model to predict the prognosis for a target patient based on medical history information of the target patient, wherein the medical history information of the target patient includes treatment instructions given to the target patient, and the ML model is trained on training data that includes medical history information of multiple subjects; A step of comparing modified medical history information relating to the subject patient, including the medical history information and predicted prognosis relating to the subject patient, with multiple medical history instances, wherein each medical history instance includes medical history information from the training data and relates to a subject among the multiple subjects; Based on the comparison, the step of selecting at least one medical history instance that is most similar to the modified medical history information for the subject patient; The process includes the step of outputting information indicating a predicted prognosis and at least one selected medical history instance, Computer program.

13. An information processing apparatus comprising a memory and a processor connected to the memory, wherein the processor is: A step of using a machine learning (ML) model to predict the prognosis for a target patient based on medical history information of the target patient, wherein the medical history information of the target patient includes treatment instructions given to the target patient, and the ML model is trained on training data that includes medical history information of multiple subjects; A step of comparing modified medical history information relating to the subject patient, including the medical history information and predicted prognosis relating to the subject patient, with multiple medical history instances, wherein each medical history instance includes medical history information from the training data and relates to a subject among the multiple subjects; Based on the comparison, the step of selecting at least one medical history instance that is most similar to the modified medical history information for the subject patient; A step of outputting information showing the predicted prognosis and at least one selected medical history instance. An information processing device configured to perform the following.

14. A step of using a machine learning (ML) model to predict the prognosis for a target patient based on medical history information of the target patient, wherein the medical history information of the target patient includes treatment instructions given to the target patient, and the ML model is trained on training data that includes medical history information of multiple subjects; A step of comparing modified medical history information relating to the subject patient, including the medical history information and predicted prognosis relating to the subject patient, with multiple medical history instances, wherein each medical history instance includes medical history information from the training data and relates to a subject among the multiple subjects; Based on the comparison, the step of selecting at least one medical history instance that is most similar to the modified medical history information for the subject patient; A step of outputting information showing the predicted prognosis and at least one selected medical history instance. A system comprising multiple computing devices configured to perform a certain task.

15. Comparing the modified medical history information for the aforementioned patient with the multiple medical history instances is as follows: The process involves: encoding each medical history instance using a second encoder model to generate a contextualized embedding for each medical history instance; The steps include: generating a vector representation of each medical history instance based on the contextualized embedding of the medical history instance; The steps include: encoding the modified medical history information relating to the patient using the second encoder model to generate a contextualized embedding of the modified medical history information relating to the patient; A step of generating a vector representation of the modified medical history information relating to the subject patient, based on the contextualized embedding of the modified medical history information relating to the subject patient; The step includes comparing the vector representation of the modified medical history information relating to the subject patient with the vector representation of the medical history instances, The plurality of computing devices include at least one first processor or computing device configured to encode each medical history instance using the second encoder model, generate a vector representation of each medical history instance, encode the modified medical history information using the second encoder model, and generate a vector representation of the modified medical history information; The system includes at least one second processor or computing device configured to compare the vector representation of the modified medical history information relating to the subject patient with the vector representation of the medical history instances, The first at least one processor or computing device has greater processing power and / or computing load than the second at least one processor or computing device. The system according to claim 14.

Images