A computer-implemented method and apparatus for performing medical test value analysis

Abstract

The present invention has particular application in medical laboratory analysis, and provides a computer-implemented method for providing at least one predictive value for at least one medical laboratory parameter, the method comprising: providing (S1) at least one test value history showing the history of historical test values ​​of at least one test parameter at at least two historical time points; determining (S3) at least one test parameter feature for each of the at least one test parameter from the corresponding test value progression; determining (S5) at least one predicted value at the set prediction time point based on the trained data-based prediction model and at least one test parameter feature for each of the at least one test value transition; The present invention relates to a method comprising:

Description

Technical Field

[0001] The present invention relates to the evaluation of medical examination parameters, particularly in hematology, urinalysis, clinical chemistry Endocrinology, blood gas analysis, autoantibodies, tumor markers and the like. In particular, the present invention relates to measures for individually presenting for each patient a reference range for identifying pathological deviations.

Background Art

[0002] Background Art Test values in medical diagnosis, particularly in hematology, clinical chemistry or urinalysis, are generally evaluated by medical staff. These test values are usually collected by a medical laboratory or a similar device and provided to a doctor together with a reference range specific to these values. Here, the reference values are in most cases the "normal ranges" shown by research, for example, the interquartile range of 2.5% to 97.5% of a healthy population, that is, the range in which each value is observed in 95 out of 100 healthy people. In some cases, the reference values are further adjusted depending on gender, age, weight or other patient characteristics. Test results with values outside such reference ranges are separately marked to emphasize the deviation / pathology.

[0003] Furthermore, especially when a particularly significant value, that is, a result that means a direct danger to the patient occurs, in many laboratories, an individual process for promptly providing such information to a doctor has been realized.

[0004] Automatically structuring and standardizing the temporal progression of test parameters and taking it into account is not done during evaluation. Even if it were done, such trend analysis is performed manually, often subjectively and intuitively, by a doctor.

Summary of the Invention

Problems to be Solved by the Invention

[0005] Due to a very large number of various test parameters in recent medicine, the interactions that are mostly unclear and multiply unknown in most cases, the lack of expertise, the lack of reviewed data criteria, and the lack of statistical knowledge, it is virtually impossible for a doctor to identify pathological deviations that the above-mentioned standard methods did not detect. At the same time, for this reason, opportunities to early identify problematic changes within the standard reference ranges of corresponding test values in a patient's health condition and introduce appropriate medical measures are lost.

Means for Solving the Problem

[0006] Disclosure of the Invention According to the present invention, there is provided a method for medical test value analysis according to claim 1, and an apparatus according to other independent claims.

[0007] Further embodiments are described in each dependent claim.

[0008] According to a first aspect, there is provided a computer-implemented method for providing at least one predicted value for at least one medical test parameter, particularly applicable to medical test value analysis, comprising: providing at least one test value transition showing the transition of historical test values of at least one test parameter at at least two historical time points; calculating at least one test parameter feature for each of the at least one test parameter from the corresponding test value transition; determining at least one predicted value at a set prediction time point depending on a prediction model based on trained data and at least one test parameter feature for each of the at least one test value transition; and a method is provided.

[0009] Furthermore, the prediction model can be configured to consider, in addition to at least one test parameter feature, patient data such as age, gender, BMI, and other biological data such as build, weight and / or diagnosis, findings, treatment and / or medication.

[0010] In the normal case, test parameters, for example, hematology or urinalysis parameters, are generally statically evaluated by a physician based on the pathological values labeled by a laboratory among the test parameters. The classification as being pathological is generally performed according to defined reference limits or reference ranges. The evaluation of test values over time is not performed automatically. Therefore, in particular, problematic trends in values that have not been prominent so far, that is, values of test parameters that are within the medical reference range but have pathological implications, are likely to be overlooked.

[0011] The computer-implemented method described above provides an automatic evaluation of medical test parameter trends and presents a correspondingly adjusted reference range for each parameter that indicates the pathological deviation of the relevant test values. For this purpose, at least one test parameter feature characterizing the trend of the corresponding test parameter is extracted from each test value trend.

[0012] By computer-assisted evaluation of test value trends, all supplied test values, their correlations, and, optionally, additional patient data such as age, gender, weight, build, medical history, underlying diseases, etc. are considered using an evaluation model, the reference range for each test value is calculated based on the model, and provided in addition to the test values, thereby providing support to physicians for interpreting the test values. Thereby, medical staff can evaluate any test value not only with respect to the "normal range" such as the interquartile range of 2.5% to 97.5% of a healthy population, but also additionally based on individual reference ranges and considering the historical trends of test parameters.

[0013] Furthermore, the prediction model can be trained to provide at least one predicted test value corresponding to at least one predicted value at a set prediction time point, depending on the test parameter characteristics.

[0014] In particular, in order to determine the time point when the predicted value exceeds the set limit value, the transition of the predicted values at multiple prediction time points can be calculated, where, in particular, the time point may further define the time point of a medical intervention such as drug administration.

[0015] Alternatively, the prediction model can be trained to provide at least one predicted quantile value at a set prediction time point as at least one predicted value, depending on the test parameter characteristics. The quantile value can indicate the upper limit value or the lower limit value of the reference range for at least one test parameter at the prediction time point, where the result of the comparison between the current test value of at least one test parameter and the corresponding quantile value at the current time point is notified as the prediction time point.

[0016] For each of the test parameters, the test parameter characteristics are each of the following characteristics, namely, · The minimum value of the historical test values, · Different quantile values respectively, for example, the first quantile value, the third quantile value, and the median, · The average value of the historical test values, · The maximum value of the historical test values, · The standard deviation of the historical test values, · The period elapsed for the historical test value last detected with the current time point as the reference, · The period elapsed for the penultimate historical test value with the current time point as the reference, · The period elapsed for the oldest test value with the current time point as the reference, · The average value of the time intervals between the detection time points of the historical test values, · The newest historical test value, · The second newest historical test value, · The value of the last gradient between the second newest historical test value and the newest historical test value, · The period to the first outlier below the historical test value · The time point of the most recent outlier below the historical test value, · The number of historical test values classified as outliers, · The maximum rate of increase between two consecutively detected historical test values, · The minimum rate of decrease between two consecutive historical test values, · The estimated linear offset of the historical test value, · The estimated linear gradient rate of the historical test value, · The estimated linear prediction of the historical test value, · The number of historical test values may include one or more of them.

[0017] It is possible to configure at least one of at least one test parameter feature to depend on a set prediction time point.

[0018] According to other embodiments, the prediction model may include a deep neural network, a convolutional neural network, a recurrent neural network, a support vector machine, a random forest model, a hidden Markov chain model, or a generalized linear model.

[0019] According to other aspects, a method for training a prediction model, particularly used by the method described above, providing at least one test value transition of at least one test parameter, each showing the transition of the historical test value of at least one test parameter at at least three historical time points; calculating at least one test parameter feature for each of at least one test parameter from the corresponding at least one test value transition before the label time point; creating a training data set by forming each training data set from at least one test parameter feature for at least one test parameter and the test value of at least one test parameter at the label time point as a label; Training a data-based prediction model depending on a training data set, and A method including the same is provided.

[0020] Hereinafter, embodiments will be described in more detail with reference to the accompanying drawings.

Brief Description of the Drawings

[0021]

Figure 1

Figure 2

Figure 3a

Figure 3b

Figure 3c

Modes for Carrying Out the Invention

[0022] Description of Embodiments Figure 1 shows a conventional computer system 1 including a computer unit 2, an input device 3, and an output device 4 in the form of a monitor or the like. The computer unit 2 is used to process the patient's historical test values based on software using the processor unit 21 and provide support for evaluating the current test values. The software and the patient's test values are stored in the data memory 22 within the computer unit 2. Further, the data memory 22 stores the parameters of the prediction model described below. The software is executed by the processor unit 21 and accesses the test values stored in the data memory 22. Further, the computer unit 2 can receive and process the test values input automatically or manually for the test parameters.

[0023] Figure 2 shows a flowchart for explaining a method for test value analysis that identifies potential pathological implications that may occur in current and future test values and, if necessary, takes treatment measures based on the results of the test value analysis, thereby reducing the burden on medical staff such as doctors.

[0024] The method first contemplates providing historical test values from the data memory of the computer unit in step S1. The historical test values can be test values such as hematology, urinalysis, clinical chemistry Endocrinology, blood gas analysis, autoantibodies, tumor markers , smear diagnosis, etc. In particular, for each type of detected test parameter, a large number of test values for the test parameter can be detected. For example, in the case of a hematological blood test, the following test parameters can be determined: AST / GOT, white blood cell count, red blood cell count, hemoglobin, hematocrit, MCV, MCH, MCHC, platelets, pH (SB status), pCO2 (SB status), standard bicarbonate, O2 saturation, lactate, ionized Ca, C-reactive protein, glucose, sodium, potassium from BGA, calcium, creatinine, GFR-MDRD, urea, INR (therapeutic range), PTT, etc.

[0025] In step S2, the current test value of the test parameter at the current time is provided. This is the result of the most recent test and serves as the basis for determining the patient's current health status. The current test value can be manually input into the computer system 1 or automatically received via a communication connection. The current test value is used by the doctor for treatment decisions. In subsequent steps, the doctor should be assisted in evaluating the current test value of the test parameter by considering the historical test values and the trends characterized by these historical test values.

[0026] In step S3, first, the test parameter features are extracted from the historical test values. For each of the detected test parameters, a plurality of test parameter features that at least partially depend on the transition are extracted. The extraction is explicitly performed without incorporating the current test value at the current time. For example, in the 26 test parameters exemplified above, a plurality of test parameter features, for example, 23 test parameter features, are extracted for each. The test parameter features for each test parameter are the following features, namely, · The minimum value of the historical test values, · The first quartile value of the historical test values, · The median value of the historical test values · The average value of the historical test values, · The third quartile value of the historical test values, · The maximum value of the historical test values, · The standard deviation of the historical test values, · The period elapsed for the last detected historical test value based on the current time, · The period elapsed for the penultimate historical test value based on the current time, · The period elapsed for the oldest test value based on the current time, · The average value of the time intervals between the detection times of the historical test values, · The most recent historical test value, · The second most recent historical test value, · The value of the last gradient between the second most recent historical test value and the most recent historical test value, · The period until the first outlier below the historical test values, · The time point of the most recent outlier that is below the historical test value, · The number of historical test values classified as outliers, · The maximum rate of increase between two consecutively detected historical test values, · The minimum rate of decrease between two consecutive historical test values, · The estimated linear offset of the historical test value, · The estimated linear gradient rate of the historical test value, · The estimated linear prediction of the historical test value, · The number of historical test values may be included.

[0027] Other features can also be defined. Some of the features depend on the prediction time point at which the predicted value is to be determined.

[0028] Feature extraction generates a so-called feature matrix, in which each column represents one test parameter feature. Examples of test parameter features are the average time interval between creatinine measurements, the total number of extreme fluctuations observed in hemoglobin trends, the maximum value of sodium, etc. The rows of the feature matrix represent the patient's state at a given time point (as the sum of the patient's features).

[0029] In some machine learning models (such as neural networks), raw data can be used. In this case, feature extraction can be omitted. However, feature extraction is supported as a possibility to directly introduce medical expertise. Therefore, the 23 features described above are commonly selected. Because these features are regarded as important features in test value trends. Note that the test parameter features generated in this way make it possible to more easily interpret the subsequent results or directly generate new problems (see feature selection).

[0030] In step S4, individual features that can vary significantly both in terms of their number of digits and in the control are scaled by normalization. For example, scaling at unit intervals can be performed. Alternatively, scaling can also be done with respect to the standard normal distribution. Basically, normalization may be performed before the step of extracting features. This may be done in addition to the normalization step, or alternatively, instead of it.

[0031] In subsequent step S5, the normalized features are supplied to the prediction model. The prediction model calculates a predicted value from the time series of the inspection parameters. For this purpose, the prediction model is trained to create a predicted value depending on the inspection parameter features.

[0032] Since the current time point is considered as the prediction time point in some of the above-described inspection parameter features, prediction based on the current time point is possible. For example, the predicted value can represent the estimated value of each inspection parameter at the current time point. Thereby, for example, a doctor can confirm a deviation from a trend revealed in the historical test values, which may suggest an acute disease.

[0033] Alternatively, two separate trained prediction models can be prepared to calculate the upper quantile value and the lower quantile value for the prediction time point, for example, the 97.5% quantile and the 2.5% quantile, from the previously determined inspection parameter features. These can be used to indicate a reference range for the evaluation or interpretation of each inspection parameter. The said reference range indicates the range in which the current test value of the corresponding inspection parameter of the individual patient at the current time point should exist or the predicted range.

[0034] If a deviation from the currently determined personalized reference range for each patient occurs in one or more test parameters, then in step S6, correspondingly, for each of the observed test parameters, this can be notified, for example, by a colored indication, thereby suggesting to the doctor the corresponding abnormality.

[0035] Alternatively or additionally, multiple queries can be made to a prediction model trained for the output of the current test values in order to output the trend of one or more test parameters, the trend corridor or the trend from the reference range at a future point in time. It should be noted here that since the test parameter characteristics are partially dependent on the prediction time point, this must be taken into account for each query.

[0036] FIG. 3a shows a corresponding graph of the trend of exemplary test parameters of an exemplary patient. A personalized prediction of the test parameter trend K for the test parameter potassium is shown. For each individual patient, the trends OG, UG of the 2.5th percentile and the 97.5th percentile are predicted, which are shown by dotted lines. In this way, the model predicts the trend within this range in 95 out of 100 cases. Furthermore, certain upper and lower limits of the potassium value according to the conventional analysis method are shown by dashed lines.

[0037] FIG. 3b shows, as an alternative embodiment, a prediction of the average value towards the future and a prediction of the expected time until the average value exceeds the standard normal range (shown by dashed lines as certain upper and lower limits) at the future time point T.

[0038] FIG. 3c shows the transitions OG, UG of the personalized upper and lower limits, indicated by a dotted line as a curve representing the overall time course of the test parameters. The points in time T1, T2 at which a deviation from the range defined by the upper limit OG and the lower limit UG is identified can be identified as being pathological. A scenario is shown in which, by a standard method, only the test value of the test parameter at time T2 was identified as being pathological. The personalized method presented herein enables, also in such scenarios, an earlier identification of the pathological test value at time T1 in addition to time T2.

[0039] The prediction model can additionally be configured to take into account patient data, such as age, gender, and other biometric data, such as build, weight, etc., and / or diagnosis, findings, medication treatment (e.g., according to ICD-10 codes). In particular, age can also be appropriately taken into account at the time of prediction.

[0040] The prediction model can be trained based on a large number of patient data. For this purpose, as soon as the time series of the test parameters includes three or more points in time, the time series of the test values of the test parameters can be processed to form a training data set. In this time series, a plurality of points in time for calculating the test parameter features and the label data of the label points in time following the points in time considered at which the test values are detected can be considered. Here, the test parameter features are calculated from the test parameter features at the points in time of the test values used as label data at the label points in time. In this way, a training data set is obtained from, respectively, the test parameter features for each of the observed test parameters and the label data for each of the test parameters as the prediction values to be trained, such as the corresponding test values at the label points in time, the corresponding lower or upper percentile values at the label points in time, and optionally patient data.

[0041] Since a large number of inspection parameter features are calculated, it is possible to perform a step of selecting features before the original training method of the prediction model. For this purpose, the so-called wrapper method can be used. This means that the prediction model is applied to various subsets of all the inspection parameter features of each inspection parameter, that is, the prediction model is applied only to a specific combination of inspection parameter features. Since it is not possible to test all combinations due to the large number, the method follows the set heuristics. For example, a variant of forward selection can be used, where the best evaluated inspection parameter features are sequentially added to the currently used subset of inspection parameter features to be considered. As a result, an optimized subset of inspection parameter features for a specific type of predicted value, for example, a predicted value indicating the 2.5% quantile of the inspection parameter, is obtained. For the selection of the set heuristics for selecting the subset, for example, backward selection, random search or other so-called Monte Carlo methods, gradient methods, etc. can be used. In addition to the wrapper method, other methods for reducing the dimension, such as principal component analysis (PCA), can also be used.

[0042] To train the prediction model, label data corresponding to the desired output value is evaluated. That is, the label data is set according to, for example, the predicted value that can correspond to the lower quantile value, the upper quantile value, or the estimated value of the corresponding inspection parameter.

[0043] After a subset of inspection parameter features has been determined, a specific training data set therefrom is presented along with the corresponding predicted values. As possible prediction models, neural networks, convolutional neural networks, support vector machines, random forest models, hidden Markov chain models, generalized linear models, etc. can be used. Preferably, an implementation of SVM (support vector machine) is used, for example, with core RGF of training parameters, gamma grid from 0.001 to 10, lambda grid from 0.001 to 10, hyperparameter selection, 5-fold cross-validation, having weights, and a pinball loss function using 0.025 and 0.0975 as the lower and upper quantile values. The loss function for training the prediction model may reflect the problem raised in the method.

[0044] In the case described above, two optimization methods are correspondingly executed for each weighting of the loss functions of 0.025 and 0.975. As a result, two prediction models are obtained that predict the 2.5% quantile or the 97.5% quantile, respectively.

[0045] For testing the prediction model, the training can be applied to 80% of the available data set. For the final evaluation of the model prediction quality, the remaining 20% is used as test data. The evaluation criteria depend on the model variant used, regarding inspection value model targets, etc. Since the model training is performed independently of the test data, the quality measure obtained by the method is more robust with respect to the problem of overfitting, for example, than the evaluation by so-called cross-validation, and can be more advantageous than such evaluation.

Claims

**Claim 1** A computer-implemented method for providing at least one predicted value for at least one medical test parameter applicable to medical test value analysis, comprising: providing at least one test value transition (S1) showing the transition of historical test values of at least one test parameter at at least two historical time points; calculating at least one test parameter feature for each of the at least one test parameter from the corresponding test value transition (S3); determining at least one predicted value at a set prediction time point depending on a prediction model based on trained data and at least one test parameter feature for each of the at least one test value transition (S5); wherein the prediction model is trained to provide at least one predicted quantile value at the set prediction time point as the at least one predicted value depending on the at least one test parameter feature, the quantile value indicating an upper limit value or a lower limit value of a reference range for at least one test parameter at the prediction time point, and a result of comparison between a current test value of the at least one test parameter and the corresponding quantile value at the current time point is notified as the prediction time point. **Claim 2** The method according to claim 1, wherein the prediction model is trained to provide at least one predicted test value at the set prediction time point as the at least one predicted value depending on the at least one test parameter feature. The method according to claim 1. **Claim 3** The method according to claim 1 or 2, wherein the at least one test parameter includes at least one parameter among hematology, clinical chemistry, endocrinology, blood gas analysis, autoantibodies, tumor markers or urinalysis. The method according to claim 1 or 2. **Claim 4** The test parameter feature for each of the test parameters is each of the following features, namely: - minimum value of historical test values, - different quantile values respectively, for example, the first quantile value, the third quantile value and the median of historical test values, - average value of historical test values, - maximum value of historical test values, - standard deviation of historical test values, - period elapsed for the last detected historical test value based on the current time point, - period elapsed for the penultimate historical test value based on the current time point, - period elapsed for the oldest test value based on the current time point ・ The average value of the time intervals between the detection times of historical test values, ・ The most recent historical test value, ・ The second most recent historical test value, ・ The value of the last gradient between the second most recent historical test value and the most recent historical test value, ・ The period until the first outlier below the historical test value, ・ The time point of the most recent outlier below the historical test value, ・ The number of historical test values classified as outliers, ・ The maximum rate of increase between two consecutively detected historical test values, ・ The minimum rate of decrease between two consecutive historical test values, ・ The estimated linear offset of the historical test value, ・ The estimated linear gradient rate of the historical test value, ・ The estimated linear prediction of the historical test value, ・ The number of historical test values including one or more of The method according to any one of claims 1 to 3.

5. The prediction model further considers, in addition to the at least one test parameter feature, patient data including age, gender, and BMI, and other biological data including the transition of physique, weight, and / or diagnosis, findings, treatment, and / or medication or drug administration, The method according to any one of claims 1 to 4.

6. At least one of the at least one test parameter feature depends on the set prediction time point, The method according to any one of claims 1 to 5.

7. The prediction model includes a deep neural network, a convolutional neural network, a recurrent neural network, a support vector machine, a random forest model, a hidden Markov chain model, a generalized linear model, The method according to any one of claims 1 to 6.

8. A method for training a prediction model used by the method according to any one of claims 1 to 7, providing at least one test value transition of at least one test parameter, each showing the transition of the historical test value of at least one test parameter at at least three historical time points, calculating at least one test parameter feature for each of the at least one test parameter from the corresponding at least one test value transition before the label time point, Creating a training data set by forming each training data set from at least one test parameter feature for the at least one test parameter and the test value of the at least one test parameter at the time of the label as a label; Training a data-based prediction model depending on the training data set; A method comprising. **Claim 9** A computer program comprising instructions for causing a computer to perform each step of the method according to any one of claims 1 to 8 when the computer program is executed by the computer. **Claim 10** A machine-readable storage medium comprising instructions for causing a computer to perform each step of the method according to any one of claims 1 to 8 when executed by the computer. **Claim 11** An apparatus comprising the machine-readable storage medium according to claim 10.

Images