Self-learning input filters for medical devices

Abstract

The present invention relates to a method for monitoring the reliability of inputs in a medical device, which comprises the following steps: inputting at least one value into the medical device; calculating the occurrence probability density of the at least one value based on the input history; calculating the error probability of the input based on a priori error probability density stored in a first database and loaded from the first database and the calculated occurrence probability density; and outputting the error probability for further processing of the input.

Description

Technical Field

[0001] The present invention relates to a method for monitoring the reliability of input to a medical device and the field of medical devices, and in particular to a device for monitoring the reliability of input to a medical device. Background Art Background of the Invention

[0003] Intensive care unit staff hear multiple alarms in their daily work. There are 150 to 350 alarms per patient per day, so different devices often compete for attention with barely distinguishable sounds. As technologization increases and patient monitoring becomes more invasive, the number of alarms and false alarms in intensive care units also increases. Among other things, this can lead to desensitization and therefore endanger patients.

[0004] In this context, for example, the Joint Commission on Accreditation of Healthcare Organizations in the United States, which is responsible in particular for the accreditation of hospitals in the United States, warns that alarm fatigue or alarm exhaustion / alarm fatigue among employees can endanger patient safety. In this regard, the American Institute of Emergency Medicine has stated that such alarm fatigue is the greatest medical technology hazard in the past two years.

[0005] Alarm fatigue is the desensitization of clinical staff, especially, due to an excessive number of irrelevant alarms from medical devices. In addition to desensitization and overwork, which can lead to inappropriate responses to alarms, excessive alarms interrupt care activities, leading to errors and increased patient stress. Furthermore, frequent, especially irrelevant, alarms lead to unnecessary workload.

[0006] Reasons for irrelevant alarms are, for example: major preparations or measures for the patient, alarm thresholds not adapted to the patient or clinical situation, use of sensors of insufficient quality and service life, inadequate threshold logic, 'over-monitoring', active monitoring and / or false alarm transmission.

[0007] Prior art

[0008] Most alarms are triggered by exceeding or falling below predefined alarm thresholds. These alarm thresholds are hard-coded in the software or stored in a database and can only be changed manually with great effort. Therefore, it is not possible for clinical staff to independently and easily adjust the alarm thresholds as needed.

[0009] The disadvantage of strictly programming error thresholds is that these thresholds may not be applicable in every situation. For example, aspirin may be used in a minimum dose to prevent platelet aggregation and in a large dose to relieve pain. Therefore, the database will not detect underdosing.

[0010] An algorithm for a medical device as previously used with fixed programmed alarm thresholds is described in FIG. 3 of the following figure. Here, in step S1, at least one value is input into the medical device. In step S9, at least one value is calculated / compared with data from database 1 (a DERS database already adopted by known systems). If the result in step S9 is positive (yes), the system goes to step S12 and outputs that the value is 'OK'. If the result in step S9 is negative (no), the system goes to step S11 and outputs an error message. This error message is followed by step S14, in which the operator is asked whether the value entered in step S1 should be overwritten. If the operator decides against the new input (i.e., against overwriting the previous value) in step S13, the algorithm stops. If the operator makes a new input in step S13, the previous value that was evaluated as incorrect is overwritten and then output as 'OK'. After that, the algorithm ends / stops until a new input is made (step S15).

[0011] Previous solutions therefore have the disadvantage that strictly defined alarm thresholds when checking the correctness of the entered values ​​lead to false alarms and unrecognized operating errors, which ultimately endanger the patient and can lead to death in the worst case. Summary of the invention

[0012] The object of the present invention is to prevent the above-mentioned disadvantages and problems and to provide a method and a device for reducing false alarms, thereby improving the safety of medical devices.

[0013] This object is achieved by a method for monitoring the reliability of an input in a medical device having the features of the invention and by a device for monitoring the reliability of an input in a medical device having the features of the invention.

[0014] The present invention relates to a method for monitoring the reliability of input to a medical device. In a first step, at least one value is input into the medical device. After the input, the occurrence probability density of at least one value is calculated based on the input history. The input history provides historical data, which is stored in a first database. Subsequently, in another step, the error probability of the input is calculated according to the prior error probability density and the previously calculated occurrence probability density. The prior error probability density is stored in the first database and loaded from the database before calculating the error probability. The prior error probability density is a discrete or absolutely continuous density function with an associated probability distribution (i.e., a priori probability distribution). After the error probability of the input has been calculated, it is input to the input processing for further use, such as for dynamically and independently adjusting the alarm threshold when checking the correctness of at least one input value. In other words, this means adjusting the alarm threshold based on the outputted error probability. Here, it is preferred whether the adjustment to the alarm threshold is performed independently / automatically.

[0015] By this method, the number of error messages, in particular the most serious error messages, can be reduced, making it possible to reduce the risk of alarm fatigue. Furthermore, fewer misuses have the advantage of reducing the number of victims of misuse. In addition, the flexible and self-learning method / system described can reduce the workload of clinical staff as well as the stress of patients. In other words, the inputs are evaluated and further processed with the help of a self-learning, flexible and adaptive system (with the help of artificial intelligence). The present invention thus provides a method for dynamically and autonomously adapting alarm thresholds to situations depending on values ​​and experience.

[0016] In an additional step, it is advantageous if at least one of the entered values ​​is stored in a database and if the a priori error probability density is recalculated and the recalculated a priori error probability density is also stored in the first database.

[0017] Preferably, the probability of error is calculated using a probability distribution based on a recursive Bayesian filter. In other words, all inputs are stored in a first database in the medical device. The inputted values ​​are recorded together with related / associated inputs such as alarms, alarm confirmations, alarm descriptions, and alarm cancellations.

[0018] A (input) filter in the style of a Bayesian filter (such as a Kalman, a particle filter, a Hidden Markov Model, and / or a Dynamic Bayer Network) computes a prior probability distribution from these values. Each new input offsets this probability. Based on the resulting posterior probabilities, the algorithm evaluates the input and asserts / decides that the input is valid or raises an alarm if the input is asserted / decided to be invalid.

[0019] The Bayesian filter or Bayesian estimation method differs from other estimation methods in classical statistics in that it treats the parameters to be estimated as random variables. As a result, a probability distribution over the resulting parameter space must be specified a priori. This makes it possible to include prior information, such as in this case known results for different inputs up to the time of a new input. After specifying the prior distribution, the values / data are used to move from the prior distribution to the posterior distribution, based on which the estimate is ultimately made.

[0020] The aforementioned method provides the possibility of an algorithm with adaptive alarm thresholds.All inputs are recorded in a first database and the prior distribution density and the occurrence probability density are calculated for the values ​​according to which the aforementioned filters calculate the effective values ​​of the inputs.

[0021] The filter works as a recursive Bayesian filter type of function. Here, the posterior filter density can be recursively determined by Bayesian estimation theory. Based on Bayes' theorem, the posterior filter density can first be expanded and then decomposed to obtain the recursive determination equation:

[0022]

[0023] In the equation, the variable x k ,y k Describes the value at time k, variable X k , Y k describes the sequence of all values ​​up to time k, and the variable p describes the probability.

[0024] The denominator of the equation is a normalizing constant, so the integral over the distribution function yields 1. Therefore, only the prediction or prior distribution density p(x k |Y k-1 ) and the probability distribution density p(y k |x k ). The prior distribution density is calculated based on historical input data. These historical input data are input data that have been evaluated as correct or incorrect in the past. The occurrence probability distribution density is generated by the occurrence probability of the corresponding individual input values. For both calculations, data is available in the first database.

[0025] Therefore, when using a Bayesian based filter, it is advantageous that the filter adapts to the input behavior of its operator and is flexible to his or her changing behavior.The filter can learn permanently and will therefore continue to filter out new incorrect input types independently.

[0026] In a further step, it is preferred that the probability of occurrence is determined using a predetermined probability distribution, in particular a Gaussian distribution, based on the data of the second database, provided that no input history is available or a sufficient number of input histories are available. This is particularly advantageous at the beginning when the first database has no historical data yet. According to FIG. 3 , the second database contains data from a previous Dose Error Reduction System (DERS).

[0027] Furthermore, in a further step, all input values ​​are stored, and the probability density of occurrence of all input values ​​is calculated and the a priori error probability density is calculated based on all input values. This procedure comprises a training method according to the 'train-everything principle' (TEFT) and, due to the fact that all inputs are included in the database, it offers the great advantage that the filter adapts very quickly to new inputs and habits. However, this fact is simultaneously connected to the fact that rapidly changing the input behavior in the aforementioned manner generates a large amount of data very quickly, which can have an unexpectedly strong influence on the evaluation and can lead to errors.

[0028] In an alternative step, only the inputted values ​​that have been judged to be wrong by the operator are stored, and then the priori error probability density is recalculated based on the initial priori error probability density and the value that has been judged to be wrong. In this error training (train-on-error, TOE) method, the learning process of the filter is initiated only when an error has been made when evaluating the input. As long as the operator does not explicitly cover the error message, the filter only works on the basis of the priori probability learned at the beginning. Therefore, the filter working with this method can only slowly adapt to the newly appearing spam / errors. However, compared with the TEFT principle, the memory used for the data set is relatively less consumed in this way.

[0029] Furthermore, in an alternative step, once a predetermined amount of stored values ​​has been stored in the database for at least one of the entered values, the calculation of the probability density of occurrence and the recalculation of the a priori error probability density are interrupted. This interruption is canceled when the operator has judged the entered value to be an error. This training method, called the "train-until-mature method" (TUM), is a compromise between the two aforementioned methods TEFT and TOE. Here, the data point reaches a predetermined amount of stored data / values, i.e. a specific 'maturity level'. Therefore, the filter is not trained with these data points until the error message is covered. Only then is this data point added to the database accordingly, or the existing number is increased by one, respectively.

[0030] In another alternative step, the filter is trained with the test data until a certain threshold value is reached when calculating the error probability, which preferably corresponds to the operator's work profile. In other words, the filter is trained with the test data in this 'train-until-no-error principle' (TUNE) until a certain threshold value is reached when evaluating the input and thus corresponds to the operator's work profile. However, if this incorrectly evaluated message is an error, the filter can also be trained during use with TUNE. However, the disadvantage here is that in the case of TUNE, the time for training with the test data can become extremely long, depending on the scope of these data and the computing power.

[0031] Therefore, four training options are available to train the system.

[0032] In addition, the present invention relates to a device for monitoring the reliability of input in a medical device, comprising: a device for inputting at least one value into the medical device; a device for calculating the occurrence probability density of the at least one value based on the input history; a device for calculating the error probability of the input based on a priori error probability density stored in and loaded from a first database and the calculated occurrence probability density; and a device for outputting the error probability for further use in input processing.

[0033] Furthermore, the above-mentioned apparatus is suitable for performing all the above-mentioned steps. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The invention is explained in more detail below with the aid of preferred embodiments and with reference to the accompanying drawings.

[0035] Figure 1 A flow chart showing a method for monitoring the reliability of an input;

[0036] Figure 2 shows an overview algorithm with adaptive alarm thresholds; and

[0037] FIG. 3 shows an overview algorithm with fixed alarm thresholds from the prior art. DETAILED DESCRIPTION

[0038] In the following, embodiments of the present disclosure are described on the basis of the associated drawings. In the various drawings, identical or functionally equivalent features are provided with identical reference numerals and are not described again for the sake of convenience.

[0039] Figure 1 The sequence of the method for monitoring the reliability of an input is shown in the form of a flow chart.

[0040] First, in step S1, the operator inputs at least one value into the medical device, and in step S2, at least one value is stored in the first database 3. Then, in step S3, the occurrence probability density of at least one value is calculated based on historical data. The historical data is stored in the first database 3. In step S4, the priori error probability density is loaded from the first database 3. Based on this, the priori error probability density and the occurrence probability density calculated in step S3 are loaded, and in step S5, the error probability of the input is calculated in the computing device 2. In step S6, the error probability calculated in step S5 is output for further use in input processing. In the final step S7, the priori error probability density is recalculated, and in step S8, it is loaded into the first database 3. In other words, new data points are added to the first database 3.

[0041] exist Figure 2 In, an algorithm with adaptive alert thresholds is outlined.

[0042] In step S1, the operator enters at least one value into the medical device, and in step S2, the at least one value is stored in the first database 3 and processed in the computing device 2, which calculates the value according to Figure 1 In step S9, the result output by the calculation device 2 or the input output valid value determines whether the input value is evaluated as positive / correct (yes) or negative / false (no).

[0043] If the result output by the calculation device 2 is positive (yes), that is, no alarm is output, the program goes to step S12. In step S12, the output value is 'OK' and the algorithm terminates.

[0044] In the case where the result output by the calculation means 2 is evaluated as negative (No), the next step S11 is executed, which outputs an error. In a further step S13, the operator must decide whether to make a new input to overwrite the value in step S14 or whether not to make a new input. In the first case, if the operator decides to overwrite the value in step S14, the algorithm proceeds to step S12 and outputs 'OK' again and the algorithm terminates. In the case where the operator decides not to make a new input and therefore not to overwrite the value in step S14, step S15 will be executed, which ends / stops the algorithm.

Claims

1. A method for monitoring the reliability of an input into a medical device, comprising the steps of: inputting at least one value into the medical device (S1); Characterized in that the method further comprises the following steps: calculating an occurrence probability density of the at least one value based on an input history (S3), wherein the input history provides historical data input to the medical device; Calculating the error probability (S5) of the input based on the a priori error probability density (S4) stored in the first database (3) and loaded from the first database (3) and the calculated occurrence probability density; and The error probability is output for further use in input processing (S6) for dynamically and autonomously adjusting an alarm threshold when checking the correctness of the at least one input value.

2. The method according to claim 1, further comprising the steps of: storing the at least one entered value in the first database; Recalculating the priori error probability density; as well as The recalculated a priori error probability density is stored in the first database.

3. The method according to claim 1, further comprising the steps of: The error probability is calculated using a probability distribution based on a recursive Bayesian filter.

4. The method according to claim 1, further comprising the steps of: The occurrence probability density is obtained using a predetermined probability distribution based on data of the second database, provided that no input history is available or a sufficient amount of input history is available. The method of claim 4 , wherein the predetermined probability distribution is a Gaussian distribution.

6. The method of claim 1, further comprising the steps of: Store all entered values; as well as The occurrence probability density of all input values ​​is calculated and the a priori error probability density is calculated based on all input values.

7. The method of claim 1, further comprising the steps of: storing only the inputted values ​​which have been judged by the operator to be erroneous; as well as The a priori error probability density is recalculated based on the initial a priori error probability density and the value judged as an error by the operator.

8. The method of claim 7, further comprising the steps of: interrupting said calculation of said occurrence probability density and said recalculation of said a priori error probability density once a predetermined amount of stored values ​​has been stored in said database for said at least one input value; and When the operator has judged the input value to be erroneous, the interruption is canceled.

9. The method of claim 1, further comprising the steps of: Training is performed with test data until a certain threshold corresponding to the operator's work profile is reached when performing said calculation of said error probability.

10. An apparatus for monitoring the reliability of an input into a medical device, comprising: means for inputting at least one value into the medical device (S1); Characterized in that the device also includes: means for calculating the occurrence probability density of the at least one value based on an input history (S3), wherein the input history provides historical data input to the medical device; means for calculating the error probability (S5) of the input based on the a priori error probability density (S4) stored in the first database (3) and loaded from the first database (3) and the calculated occurrence probability density; and Means for outputting the error probability for further use in input processing (S6) for dynamically and autonomously adjusting an alarm threshold when checking the correctness of the at least one input value.

Images