Neural networks arrangement for the determination of a substance dosage to administer to a patient

p0001The present invention relates to a device for determining the amount of a substance to be administered to a patient in order to modify a physiological condition of the latter, may vary in response to administration of said substance and approach a state physiological target.

p0002The invention relates more particularly but not exclusively insulin therapy, that is to say, the administration of insulin or any product equivalent to a patient with diabetes to control their blood sugar.

p0003Insulin is usually administered intravenously as a continuous infusion.

p0004It is difficult to determine exactly the infused insulin flow which allows to reach a target blood sugar because many factors affect changes in blood sugar, such as a recent meal, the patient's physical activity, his nervous condition , etc ...

p0005The infused insulin flow is determined more or less empirically by specialized medical teams based on regular measurements of blood sugar and experience in this area.

p0006Given the difficulty to regulate blood sugar levels in a patient, competent medical teams in insulin therapy are relatively few and there is a need to have a device that would assist a medical team specializing in non-insulin treatment patients with diabetes.

p0007The invention thus relates to a device for determining the amount of a substance to be administered to a patient in order to change physiological state of the latter may vary in response to the administration of said substance and approach of a physiological state target.

p0008According to the invention, this device comprises several neural networks have been trained, from a set of experimental cases of representative data and / or clinical, to provide an indication of the recommended amount of the substance to be administered according to said status target physiological parameters specific to the patient and considered knowledge of at least one prior physiological state of the latter, each neural network is further specific to a predetermined period of patient treatment.

p0009In a particular non-limiting embodiment of the invention, said physiological condition is blood sugar and said substance to be delivered is insulin or equivalent.

p0010Thus, it has thanks to the invention of a device able to recommend the amount of insulin to be administered to achieve a target blood sugar, which can be used by personnel who have received no special training in insulin therapy, if the patient itself or that one regulating independently glucose.

p0011In a particular embodiment of the invention, each neural network is trained by presenting the network a succession of input patterns each having among its components a value representative of the blood glucose measured in a case experimental and / or clinical and imposing the network in each case as the output boss a value representative of the administered insulin flow.

p0012Alternatively, drives each neural network in the network having a succession of input patterns each having among its components a value representing the insulin flow has dministré in a case experimental and / or clinical and imposing to the network for each case output boss a representative glucose value measured leading to the administration of the insulin flow.

p0013In a particular example of implementation of the invention, a number N of the blood glucose measurements performed each day, the device comprises a first set of neural networks used for recommending an insulin flow administered during each period s extending between two consecutive measurements from the second measurement of glucose to the Nth measuring blood sugar, and a second set of neural networks used to advocate an insulin flow to administer for each subsequent period between two consecutive glucose measurements.

p0014According to a particular embodiment of the invention, the device further comprises an injection device.

p0015Other features and advantages of the present invention will become apparent from reading the detailed description that follows, a non-limiting embodiment of the invention, and on examining the attached drawing in which:<ul><li>1 shows schematically a device according to an embodiment of the invention,</li><li>Figure 2 schematically shows the general organization of the calculation means used, and</li><li>Figures 3 and 4 are timing diagrams illustrating two different phases of operation of the device.</li></ul>

p0016Is shown in Figure 1 a device 1 according to an embodiment of the invention, which is advantageously in the form of a portable battery-powered device.

p0017This device 1 comprises a CPU 2 and a number of devices connected to it, including a display 3, a keyboard 4, a card reader-writer chip 5, an alarm 6 and an injection device 7, schematically represented.

p0018Was designated under the reference means 11 for measuring the patient's blood glucose.

p0019The CPU 2 has an internal clock, two redundant microcontrollers 8.9 for device management and October 1 microcontroller programmed to advocate, as will be explained below, a dose of insulin to administer.

p0020Microcontrollers 8.9 and 10 are arranged in a loop to exchange information and monitor each other.

p0021The device 1 allows to assist unskilled personnel in insulin therapy in the choice of the dose of insulin to administer to a patient to control his blood sugar.

p0022Considering that for the next ten daily blood glucose measurements are made by sampling capillary blood at the following times:    1 hour 30, 4 h, 7 h 30, 9 h, 12 h 13 h 30 16 h, 19 h, 20 h and 23 h.

p0023It is assumed that meals are taken regularly every day at the following times:    7 h 30, 12 h and 19 h.

p0024Will designate "prandial periods" time slots beginning at time of blood glucose measurement coinciding with the start of a meal and ending at the time of the next blood glucose measurement.

p0025In the example, the prandial periods therefore correspond to the following time slots:    7: 30 am to 9 am, 12 am to 13 h 30 and 19 h to 20 h.

p0026The injection device 7 comprises a mechanism for advancing the plunger of a cartridge of insulin, and can control the insulin flow administered by intravenous infusion to the patient by acting on the piston advancing speed.

p0027Be referred to below as "base rate" a flow of insulin administered by injection device 7 out of prandial periods.

p0028is schematically represented in Figure 2 the general organization of the calculation means used.

p0029In general, there are three successive stages in the treatment of the patient, the first phase extending between the first and second blood glucose measurements, the second phase extending between the second and tenth blood glucose measurements, and the third stage starting from the eleventh measure blood sugar.

p0030During the first phase, we calculate the insulin flow advocated without involving neural networks, as will be explained below.

p0031For the second and third phases, we calculate the insulin flow advocated by involving each time a neural network.

p0032Can be distinguished at the second stage nine successive time slots each extending between two consecutive blood glucose readings.

p0033For each of these time slots is calculated the recommended insulin flow by means of a specific neural network.

p0034Given that ten blood glucose measurements each day, there must be ten neural networks for the respective time slots beginning at each blood glucose measurements.

p0035is schematically referenced in Figure 2 D_1.2, D_2.2... D_10.2 these ten neural networks.

p0036neural network D_1.2 is associated with the time period starting from the first blood glucose measurement performed every day, etc ...

p0037Divide the third phase in cycles of ten successive time slots each extending between two consecutive blood glucose measurements.

p0038For each of the ten time bands of a cycle, calculating the rate recommended insulin by means of a specific neural network.

p0039Thus cyclically uses the same ten neural network to calculate the recommended insulin flow during the third phase, which was respectively referenced D_1.3, D_2.3... D_10.3 in Figure 2.

p0040neural network D_1.3 is associated with the time period starting from the first blood glucose measurement performed every day, etc ...

p0041During the first phase of treatment, there are four cases depending on whether it is during one of the three periods prandial or not.

p0042In all three cases corresponding to a prandial period, we calculate the insulin flow advocated by a linear combination of four parameters that are:<ul><li>patient weight ratio of the square of its size</li><li>the measured blood glucose,</li><li>the patient's age,</li><li>weight of the patient.</li></ul>

p0043For the latter, corresponding to a blood glucose test that is not performed at the beginning of a meal, the insulin is calculated flow promoted by a linear combination of five parameters, which are:<ul><li>patient weight ratio of the square of its size</li><li>the measured blood glucose,</li><li>the patient's age,</li><li>weight of the patient and</li><li>blood glucose target.</li></ul>

p0044During the second phase of treatment, determining the insulin flow advocated by successively using nine of the ten neural networks D_1.2, D_2,2,..., D_10.2.

p0045neural network is subjected to a used input boss whose components:<ul><li>the last measure of blood sugar,</li><li>the penultimate measurement of blood glucose,</li><li>a Boolean indicator of a previous sugar intake,</li><li>a Boolean indicator of a point earlier injection of insulin,</li><li>the insulin flow administered after the penultimate measurement of blood glucose,</li><li>the average value of all base rates since the beginning of treatment and until the penultimate measure of blood sugar,</li><li>weight of the patient relative to the square of their height</li><li>his age and</li><li>his weight.</li></ul>

p0046During the third phase of treatment, using each of the ten neural networks D_1.3, D_2.3... D_10.3 presenting the selected network an input boss whose components:<ul><li>the last measure of blood sugar,</li><li>the penultimate measurement of blood glucose,</li><li>the penultimate measure of blood sugar,</li><li>a Boolean indicator of a previous sugar intake,</li><li>a Boolean indicator of a point earlier injection of insulin,</li><li>blood sugar measured the day before at the same time,</li><li>blood sugar measured the day before when measuring following that carried the day at the same hour,</li><li>an insulin resistance coefficient,</li><li>the insulin flow of the day at the same hour,</li><li>the average value of the basal rate since the beginning of treatment,</li><li>the previous insulin flow, and</li><li>antépénultien the insulin flow.</li></ul>

p0047The coefficient of insulin resistance is the difference between the insulin flow the day at the same hour and the product of a predetermined amount of insulin flow for the time period considered by the linear regression coefficient between all the basal rate since the beginning of treatment and the predetermined values ​​of insulin rate for the corresponding time slots.

p0048As an indication, the predetermined values ​​of insulin flow are the ten time slots respectively associated with the ten blood glucose measurements taken periodically to a multiplicative constant:    0.8; 0.8; 1.2; 1.2; 1.2; 1.6; 1.6; 1.6; 1.3; 1.3

p0049Is shown in Figures 3 and 4 are two examples of timing diagrams on which we postponed the different phases of treatment, with each of the figures on a smaller scale specifying the number of blood glucose measurements from the start of treatment.

p0050Note that, in the example of Figure 3, the first phase of treatment is not a prandial period, the second phase of treatment begins when blood glucose measurement at 4 am and the third phase treatment begins when blood glucose measurement at 1 pm 30 am.

p0051Thus, during the second phase, successively uses neural networks D_2.2... D_10.2 and in the third phase is cyclically uses neural networks D_1.3... D_10.3.

p0052It will be noted in the review of Figure 4 that the first phase of treatment is in the example a prandial period, the second phase of treatment begins when blood glucose measurement at 20 am and that the third phase of treatment begins when blood glucose measurement at 19 am two days after the start of treatment.

p0053Thus, during the second phase, successively uses neural networks D_9.2, D_10.2, D_1.2... D_7.2And in the third phase is cyclically uses neural networks D_8.3, D_9.3, D_10.3, D_1.3... D_7.3.

p0054Neural networks D_1.2... D_10.2 and D_1.3... D_10.3 are of known type Multilayer Perceptron fully connected with a hidden layer and sigmoid transfer function.

p0055The number of input cells is determined by the number of components of the input pattern, the number of hidden cells in the example described is equal to 3, and the number of output cells is 1.

p0056Learning each neural network is done from a database corresponding to experimental cases and / or clinical network by presenting a succession of input patterns with components from said experimental cases and / or clinical and imposing as output boss insulin flow administered in each case.

p0057The components of each input boss underwent before being presented to the network a type of affine transformation so as to be between -1 and +1.

p0058The component of the output boss imposed on the network is also undergoing a type of affine transformation so as to be between -1 and +1.

p0059Each neural network is driven by a back-propagation method of the gradient of the error without second order and without momentum.

p0060The number of iterations is around 10000 with a sequential adjustment weights.

p0061After learning the parameters related to each neural network D_1.2... D_10.2, D_1.3... D_10.3 are stored as value tables in a ROM that can be read by the microcontroller 10.

p0062To calculate an insulin flow advocate, microcontroller 10 load the settings for the neural network to use, which is selected depending on the time period considered, as explained above.

p0063The device 1 comprises in addition preferably at least one network of neurons autoassociator type known per se, to obtain information regarding the likelihood values ​​of the components of the input boss used to calculate the recommended insulin flow.

p0064More particularly, in the embodiment described, there is associated a neural network of the self-associator type to each neural network D_1.2, D_2.2... D_10.2, D_1.3... D_10.3.

p0065Is referenced in Figure 2 Q_1.2, ..., Q_10.2, Q_1.3, ..., Q_10.3 autoassociator type neural networks respectively associated with neural networks D_1.2... D_10.2, D_1.3, ... D_10.3.

p0066Each Q neural network_1.2, ..., Q_10.2, Q_1.3, ..., Q_10.3 was driven to reproduce output that it receives as input and can be deduced from a difference between the output and the input of such a neural network information that is likely to indicate that the values ​​of the components of input boss are vitiated by an error.

p0067We can then sound an alarm signal and display a message on the screen asking the user to verify the entered data. Alternatively, one can use the RBF neural network-type, known per se, to provide zero output if input deviates too learned data.

p0068It may happen that during a treatment, blood glucose measurements are missing, which can disturb the calculation of the recommended insulin flow which occurs, in particular for the third phase of treatment, depending on several previous measurements of glycemia.

p0069This is the case for example if blood glucose measurements can only be performed after the patient transfer from one service to another.

p0070Missing values ​​of blood glucose may be provided to the device by the medical team or, in a particular embodiment of the device, be determined by several specially trained neural network to output a blood glucose value for a number of parameters 'Entrance.

p0071The operation of device 1 is as follows.

p0072After installation of new insulin cartridge in the injection device 7, serving a catheter for connecting the injection device 7 to the patient is performed.

p0073Before use of the device 1, it is necessary to set it up on the way a number of parameters specific to the particular patient.

p0074This initial configuration can be done manually with the keyboard 4, or by introducing a pre-programmed chip card into the recorder 5.

p0075The return information to the keyboard 4 is performed in the example described by a drop down menu system, the transition from one menu to another, or the increment and decrement a value or letter s' performing with two keys, while two buttons allow passage of a data field to another and a fifth button allows the validation of a selected entry.

p0076Following each measuring glucose, we enter the device 1 in addition to the glucose value measured possibly the value of certain parameters, including the number and nature depend on the current processing phase.

p0077After the return of the requested data, the device 1 calculates and displays a recommended insulin flow.

p0078If the displayed value valid user, the CPU 2 controls the injection device 7 to provide the patient the corresponding insulin dose.

p0079The device 1 is also capable of displaying different messages about abnormal data for example.

p0080The device 1 advantageously stores all parameters related to the treatment of a patient on a smart card through the reader-recorder 5.

p0081a set of data can easily be to increase the number of experimental cases and / or clinical used for learning neural networks or use the data thus stored on the smart card to reset immediately if the device 1 fails, another device working properly.

p0082Although in the embodiment described, the blood glucose is measured discontinuously, it is possible, without departing from the scope of the invention, performing a continuous measurement of the blood sugar to improve the prediction accuracy of the insulin dose to be administered to achieve a target blood glucose.

p0083In this case, it can advantageously increase the number of time slices considered for the calculation of insulin flow advocated and jointly increase the number of associated neural networks.

p0084Can be further increased, without exceeding the scope of the invention, the number of parameters taken into account to calculate the recommended insulin flow, for example the type of diabetes affecting the patient.

p0085Of course, the invention is not limited to the embodiment which has just been described.

p0086Use may in particular a device according to the invention for administering to a patient a substance other than insulin to regulate a physiological condition other than glucose, in resulting from experimental cases and / or clinical one or more networks neurons to provide an indication of the amount of the substance to be administered to approach a target physiological state.