mias are known from European Application 0 465 241 and FIG. 5 shows an embodiment of the device according to
U.S. Pat. No. 5,280.792. the invention in the form of an active implant. The units
A neural network generally has the option of receiving within the dashed line are units contained within the implant,
several input signals and generating therefrom one or more in a housing having a size and shape adapted for implanta
output signals. Each input signal is processed and combined 5 tion in a patient. Connected to the input terminals 1 and 2 of
with the other weighted input signals. The combinations of a unipolar or bipolar lead is a pulse generator 3. shown as a
the signals are further processed in order to produce one or functional block. This block contains circuitry for generat
more output signals. The processing of the signals includes mg me, heart stimulation pulses. The block also contains
multiplication, addition and non-linear operations. A neural circuitry for interfacing with the control unit 4. The control
network may have different structures. This structure must 10 unit 4 has a overridi„g function in establishing the stimu
first be decided. The neural network then has a number of lation rate
hidden parameters (W„„. b„. weights and biases). A charac- "... , ., .
teristic for the neural network is that it must learn how to another embodiment the control functions could just as
adjust these parameters. we^ ^e placed m another unit, since this is only a matter of
The input signals shown are in the shown example are S1, organizing the functions.
S2, S3 and S4. Each input signal is then multiplied by a The device also includes an IEGM amplifier 5 for ampli
certain weight (W„. to W4t). For every input signal there are fication and filtering of the IEGM signal The amplifier 5 is
as many weights as there are summing junctions. In FIG. 3 connected to the input terminals for receiving IEGM signals
there are k summing junctions. In a summing junction all from connectable implanted leads and for amplifying the
weighted signals and a bias signal are added. The bias signal IEGM signals. The amplified signal are thereafter sent to the
is a constant necessary for the functioning of the neural 20 next block in the implantable device,
network. ^ next block ^ a ciassjfier 5 its function is to
The output signal from the summing junctions is then dassify ^ istered mGMs into different morphological
passed through a transfer function denoted * . Very often (wave } TWs b ... t0 ±e described rate
this transfer function is a non-linear function, like the frT ,r' . ., . A .. . . .
.. c .. T .. . . . , regulation method. The classifying function can be lmple
sigmoid function. In the very same way as the input signals 25 6 . , .._ J » »u
are processed, the output signals from the transfer function mente,d m several different ways, both with regard to the
blocks are now handled. Here there are new weights and underlying algorithms and the hardware used as will be
biases, vu to V^, b, to bm.. In FIG. 3 only one single output discussed below in connection with FIGS. 6 and 7.
Sout is shown. The transfer function is denoted "F". Nor- The control unit 4 interprets classification results from the
mally this last transfer function is a linear function. 30 classifier 6 and based thereon sends rate regulating signals to
The neural network used in the method according to the a control input 10 of the pulse generator 3.
invention can be somewhat different from the above The control unit 4 also controls which set of weights is to
example but the operation of the network will be the same. t,e used by the classifier 6 and handles the transfer of weights
The learning action is performed on representative input vja telemetry. The wave table is coded as several sets of
signals which are fed into the neural network in order to 35 weights.
produce an output signal or pattern, which will be compared . . , , . . . -.^^.11
r, . , „ „ , 7 ■ The active implant also has a telemetry unit 7 that allows
with a reference signal or pattern. The learning procedure is , .... 7^_ , c . ,
, ., B, .. ; .. ^ . . • , . ^. for bidirectional transfer of data, performed by repeatedly feeding these input signals to the
neural network and adjusting each parameter so that finally A" external telemetry unit 8 is also shown, which is the output signal or pattern from the neural network ^ connected to an extracorporeal device 9 in which the trainresembles the reference signal or pattern. The neural net- inS process of the neural network may be performed. The work is now adjusted. New input signals will then produce device 9 receives *** from &t active imPlant mi output signals or patterns, which will be similar to the ttams an external neural network using that data and sends signals or pattern that the process, which the neural network the weights, which constitute the underlying data for the is simulating, will produce. wave back t0 ^ implant, to be used by the classifier Neural networks are today relatively well established 6.in ""jun^on with the control unit 4 in regulating the tools for pattern classification tasks. Several different net- stimulation rate so as to be appropriate one the situation, i.e.. work types and learning algorithms may be used for solving for woridoad to which P^"118 subjected, this particular problem. The invention has been tested with 111 me ruture an external unit for training of the neural a multilayer feed forward network trained by using the 50 network may not be necessary in implementations of the back-propagation algorithm. There are also statistical meth- active implant according to the invention since the power ods equivalent to neural networks that do basically the same consumption for sending bulk data via telemetry soon may job. These methods, however, may be viewed as distinctly be balanced by the power used for training carried out in the different from neural networks. Finally, a neural network can active implant.
be converted to fuzzy logic so as to operate as a fuzzy logic 55 The once established weights may for different reasons
processor. Fuzzy logic may be more favorable than neural have to be changed by training the network all certain
networks for hardware implementation. intervals. i.e.. in case of major changes in the status of the
When the pacemaker is in use the device continuously patient's heart. These changes must not necessarily be
monitors the heart beats. This includes monitoring of the dependent on serious events occurring in the heart. Also due
waveforms. If the waveform detected by the device is not eo t0 toe natural ageing of the heart one might suspect that the
optimal, the pacing rate has to be changed either upwardly waveform table once registered may have to be changed,
or downwardly, depending on the characteristics of the This means that a new set of weights should be established
waveform. This procedure will be repeated until the device as functions of different workloads and rates. The correlation
finds an optimal waveform match. between workloads and the preferred pacing rate for the
The waveform table (wavetable) after being produced is 65 specific workload may have been changed,
stored in the implantable device and may be fine-tuned by This will be rather simple, however, since new IEGMs
the patient or by a doctor. may be registered using the already implanted leads and the