Epileptic seizure warning system

Abstract

Epileptic seizure warning system based on pattern recognition principles is embodied in a small self-contained device which can be carried in a pocket of a person subject to grand mal seizures, to monitor the person's condition and provide sufficient advance warning of any imminent seizure so that the person can take preventive medication or lie down to prevent or minimize any harmful seizure effects. Scalp electrode connections provide a suitable (brain wave) electrical signal to the device which preprocesses the signal to transform it into a predetermined format, reiteratively processes the transformed signal and extracts or detects essential features from the processed signal and decides from the detected features the condition of the person and, when a predetermined number of preseizure decisions is exceeded within a given period indicating an imminent seizure, produces both audio and visual warning signals. The warning signals can comprise an audible beeping and a flashing light. Only the initiated audible signal can be suppressed and, as long as the warning condition persists, the light continues to flash.

Description

The invention described herein was made in the course of, or under, a contract with the United States Department of Health, Education, and Welfare.

BACKGROUND OF THE INVENTION

Our present invention pertains generally to the field of warning systems. More particularly, the invention relates to an epileptic seizure warning system.

Epilepsy has been defined as a tendency in a person to recurrent attacks involving changes in the state of consciousness, motor activity or sensory phenomena, associated with indications of abnormal overactivity of at least some part of the brain at the time of an attack. Epilepsy is a medical term covering a wide variety of episodic disturbances, and any recurrent seizure pattern might be properly termed epilepsy.

Grand mal seizures or convulsions are the most common form of epilepsy and may occur at any age. About 90 percent of patients are subject to this form of seizures, either alone (60 percent) or in combination with other forms (30 percent). Petit mal epilepsy is characterized by a brief and sudden loss of consciousness, occurring predominantly in children and usually disappears after adolescence. This form of seizures occurs in about 25 percent of patients, either alone (4 percent) or in combination with other forms (21 percent). The psychomotor form of epilepsy is characterized by a clouding of consciousness for one or two minutes and may develop at any age. Psychomotor seizures occur in approximately 18 percent of patients, either alone (6 percent) or in combination with other forms (12 percent). The petit mal, psychomotor and other forms of epilepsy are relatively mild and less disruptive than the convulsive form. Overall, about 70 percent of patients have only one form of seizures and the remaining 30 percent have two or more forms.

Epileptics who are ambulatory, able to work, but still vulnerable to convulsive seizures can be greatly aided by a device which reliably predicts the occurrence of an imminent seizure at least several minutes in advance. The epileptic would then have sufficient time to administer preventive medication and allow it to take effect. Alternatively, the epilieptic could lie down or otherwise have enough time to adjust to a safe situation against the forthcoming seizure. The importance of a device which can predict with good reliability an immiment seizure derives from the fact that, in the United States alone, approximately three million persons have epilepsy and about 75,000 new cases appear annually.

Pattern recognition technology has been applied to, for example, sleep state classification. In this case, the electroencephalograph recording or electroencephalogram (EEG) was utilized for categorizing sleep periods of selected subjects. The EEG signal successively assumes several well-defined patterns throughout the course of a normal night's sleep. The procedure followed for processing this EEG data, the recognition technique used for pattern classification, and the results are discussed by S. S. Viglione on "Applications of Pattern Recognition Technology" in Chapter 4 of "Adaptive, Learning and Pattern Recognition Systems" edited by J. M. Mendel and K. S. Fu, and published by Academic Press, New York, 1970.

Of course, it is well-known that the EEG has long served as a clinical aid for the diagnosis of mental disorders and brain defects, and has been quite useful in the study of epilepsy. The hypothesis was thus formulated that definable preseizure activity may be noted from EEG data, and pattern recognition techniques can be used for detecting and classifying such activity. Certain experimental procedures and equipment have been established and tested to assure reliable acquisition of EEG and other correlated data from physically active epileptic subjects. Pattern recognition studies of the EEG data using a known iterative design technique have demonstrated that hypotheses may be easily formed by automatic methods on the characteristics of preseizure patterns in individual subjects. The results support the hypothesis that preseizure activity can be localized. The acquisition of data and pattern recognition studies are discussed by S. S. Viglione, V. A. Ordon and Frank Risch in a paper entitled "A Methodology for Detecting ongoing Changes in the EEG Prior to Clinical Seizures" presented at the 21st Western Institute on Epilepsy on Feb. 27 and 28, 1970.

SUMMARY OF THE INVENTION

Subsequent further experiments and studies produced new and useful developments in methodology and techniques, based upon pattern recognition procedures, which led to the invention and realization of an autonomous epileptic seizure warning system capable of predicting and forewarning an epileptic subject of an impending attack. Briefly, and in general terms, our invention is embodied in a small battery-supplied device which can be carried in a shirt pocket of a person subject to grand mal and other forms of epileptic seizures. The device monitors the person's condition and can provide sufficient advance warning of any imminent seizure so that the person can take preventive medication or lie down to prevent or minimize any harmful seizure effects.

The warning device broadly comprises suitable scalp electrode means for sensing an electrical EEG (brain wave) signal, a preamplifier for amplifying the EEG signal, means for preprocessing the amplified signal to transform it into a format of a plurality of frequency bands including input data therein, means for processing the input data to rectify, integrate and normalize (and additionally square where required) such input data, means for detecting certain predetermined features or properties from the processed input data, means for classifying or deciding from the detected features the condition of the person, and warning means responsive to the output of the decision means for providing both audio and visual warning signals when the person is subject to an imminent seizure. The amplified EEG signal is reiteratively sampled to determine the person's condition and, when a predetermined number of preseizure decisions is exceeded within a given immediately preceding period, the warning means is activated to produce the audio and visual warning signals. The warning device further comprises means for inhibiting the audio signal after it is initiated.

BRIEF DESCRIPTION OF THE DRAWINGS

Our invention will be more fully understood and other advantages and features thereof will become apparent from the description given below of an exemplary embodiment of the invention. The description is to be taken in conjunction with the accompanying drawings, in which:

FIG. 1 is a block diagram of an epileptic seizure warning system according to this invention;

FIG. 2 is a block diagram illustrating effective decision structure for an epileptic subject RA and which can be used in a warning system similar to that shown generally in FIG. 1;

FIG. 3 is a block diagram of effective decision structure for another epileptic subject BJ and which can be used in a warning system simplified from that shown generally in FIG. 1;

FIG. 4 is a block diagram of warning logic means which can be used in the warning system generally shown in FIG. 1;

FIG. 5 is a circuit diagram showing the scalp electrode connections and preamplification means which can be used in an epileptic seizure warning system for the subject BJ;

FIG. 6 is a circuit diagram of the preprocessing and processing means including filters, rectifiers, integrators and scaling amplifiers for the warning system of the subject BJ;

FIG. 7 is a circuit diagram showing the typical supply and stabilizing connections normally used in all of the operational amplifiers;

FIG. 8 is a circuit diagram showing the decision structure which can be used in the warning system for the subject BJ;

FIG. 9 is a graph showing plots of output voltage versus frequency for the filters used in the warning system of the subject BJ;

FIG. 10 is a circuit diagram of a voltage regulation circuit for providing a stable voltage reference used in voltage-sensitive portions of the warning system for the sujbect BJ;

FIG. 11 is a circuit diagram of a squaring circuit which can be used in the warning system shown in FIG. 1, to provide the square of the value of a normalized filter output;

FIG. 12 is a circuit diagram of an illustrative summing circuit which implements both negative and positive input weights, and can be used with suitable input modifications as the summing amplifiers in the decision structures of FIGS. 2 and 3;

FIG. 13 is a circuit diagram of the timing and control means broadly indicated in the warning system of FIG. 1; and

FIG. 14 is a circuit diagram of the warning logic means shown in block diagram form in FIG. 4.

DESCRIPTION OF THE PRESENT EMBODIMENT

In the following description and accompanying drawings of an illustrative embodiment of our invention, some specific values and types of components are disclosed. It is to be understood, of course, that such values and types of components are given as examples only and are not intended to limit the scope of this invention in any manner.

As originally conceived, the epileptic seizure warning system for use by individuals suffering primarily from grand mal epilepsy might have been configured in two different ways. The first method would have employed a subjectmounted EEG amplification and telemetry system remotely coupled to a telemetry receiver and a small digital data processor. The advantages of this technique included the use of existing, proven telemetry, and standard computer techniques, the ability to perform accurate frequency analysis through the use of the Fourier transform, and the ability to add fairly complex logical operations to the outputs of the pattern recognition systems. However, this remote processor concept implied that the subject would have to remain within range of the telemetry receiver or a telephone tie-line.

The second alternative employed a set of analog signal processors (filters, etc.) which operated on the EEG to produce approximations to the frequency parameters and decision logic used by the digital simulations of the warning systems. This approach was expected to be less accurate than the digital design and would limit the types of logical operations that could be performed on the output decisions. However, the subject would be allowed free range of movements and would not be tied to a central receiving or processing station. The processor would be completely self-contained, including battery power supply and audible alarm, so continuous monitoring by trained personnel would not be required. This alternative was selected and the following material describes a self-contained subject-mounted warning system.

FIG. 1 is a block diagram of the self-contained epileptic seizure warning system 30 implemented according to our invention. The EEG preamplification means 32 builds up the low level EEG signal obtained from scalp electrodes (not shown in this Figure) suitably attached to a subject. The amplified EEG signal is fed through a bank of bandpass filters 34. The center frequencies indicated by the filters 34 are given only as illustrative and typical examples. The output of each filter 34 is rectified and integrated over, for example, successive 15-second epochs to yield a voltage for each epoch proportional to the energy in each frequency band. Each filter output is normalized by subtracting the mean and dividing by the standard deviation of the EEG signal obtained over a 30-minute time interval of normal EEG. This information has been previously obtained and is implemented as fixed resistive values in the scaling amplifier circuits shown in FIG. 6. Rectification, integration and normalization are accomplished by means 36. Where the decision structure 38 selected uses both the normalized spectral values and the squares of the normalized values, each normalized spectral output is passed through a squaring circuit 40 as indicated in FIG. 1. The decision structure 38 is preferably of the well known error correction type.

The output of the decision structure 38 is applied to warning logic means 42 which provides both audio and visual warning signals whenever the person being monitored is subject to an imminent seizure. Timing and control means 44 produces reiterative integration in the means 36 for the successive 15-second epochs and, when a predetermined number of preseizure decisions for the epochs from the decision structure 38 to the warning logic means 42 is exceeded within an immediately preceding period of, for example, 5 minutes, warning devices of the warning logic means are activated to provide the audio and visual warning signals. The warning logic means 42 includes shift register means for counting the number of preseizure decisions from the decision structure 38. The timing and control means 44 includes clocking means for properly clocking in the decisions from the decision structure 38 into the shift register means and, for system checkout purposes only, clearing means for clearing and resetting the entire warning system. The warning logic means 42 further includes means for inhibiting the audio warning signal after it is initiated but not the visual warning signal.

FIGS. 2 and 3 are block diagrams of exemplary decision structures 46 and 48 for epileptic subjects RA and BJ, respectively. These decision structures 46 and 48 are depicted as multiple and single layer threshold logic systems. The examples shown contain three threshold logic units 50, 52 and 54 in the first logic layer for subject RA (FIG. 2) and one unit 56 for subject BJ (FIG. 3). In general, there may be any number of first-layer logic units. Each threshold logic unit (TLU) has a linear input and for subject RA, also a squared input, from each bandpass filter. Each first-layer logic unit computes a weighted sum of all its inputs and compares the sum with its predetermined threshold .theta.. If the sum is greater than the threshold, the unit is "on" (output = 1); otherwise, the unit is "off" (output = 0). Similarly, the second-layer threshold logic unit 58, where required as in FIG. 2, calculates a weighted sum of the outputs of the first-layer units 50, 52 and 54 to determine if the observed 15-second epoch were baseline (output = 0) or preseizure (output = 1). These decision devices 46 and 48 can be implemented using integrated circuit summing amplifiers 60 in conjunction with Schmitt triggers 62.

Generally, in achieving the decision structures 46 and 48, recorded EEG analog data obtained by use of a biotelemetry system and including baseline (normal) and preseizure periods of an epileptic subject are digitized and a large number of samples of 15-seconds epochs are gathered therefrom. A preseizure period is chosen to be the 10 minutes immediately prior to a seizure. A Fourier analysis is performed on all of the samples and, for example, 13 frequency bands are normally found desirable and practical to cover the frequency range of interest (such as 0 to 26 Hz.) A bank of 13 analog filters (each having an output voltage versus frequency characteristic similar to the plots 142 and 144 of FIG. 9) is then digitally simulated in a general purpose computer, with processing normalization of filter outputs and squaring of the normalized outputs. Ideally, the filter characteristic could be rectangular but this is impractical to achieve. The simulated filters provide 13 linear and 13 squared outputs (coefficients or numbers) for each sample.

Digital (computer) simulation of a decision structure is also made wherein such structure includes first and second layers of interconnected threshold logic units and selected (by estimation) to begin with two baseline channels and two preseizure channels. The logic units have arbitrarily chosen input weights. Training and test patterns are randomly selected from the processed 15-second baseline and preseizure data. A test pattern can be, for example, every fifth processed sample. Baseline data do not include any EEG data recorded approximately 1 hour prior to the EEG signs of grand mal seizure or data from 30 minutes following the start of a seizure. The test (or system evaluation) set of data consists of the data from other recorded seizures and the balance of the baseline recordings.

Computer runs are made with the numerous training patterns which are used for adjustment of the weights of the threshold logic units. The 26 coefficients or numbers of each training pattern are provided as inputs to the simulated decision structure and the weights are adjusted according to well known error correction procedure. For each possible output of the threshold logic units of the decision structure, if the decision is incorrect (or correct) for the particular input pattern, each input coefficient or number is multiplied by a small predetermined constant (c) and adding (or subtracting) the product to the corresponding channel input weight (or from the other associated channel input weight). The value +1, multiplied by the small constant, is also added to (or subtracted from) the related threshold weight. See, for example, pages 69 through 74 of "Learning Machines" by Nils J. Nilsson, McGraw-Hill Book Company, New York, 1965. Every so often during the design process, the system is tested on the separate group of test patterns. These tests involve no weight changes but the best design point for a particular error correction run set is determined by the percentage correct on the test group. Thus, the final decision structures were based upon good performance on an independent set of test patterns as well as the set of training patterns.

On successive training pattern runs, inputs (filters) are deleted based upon low weight values. The lowest weighted groups are progressively dropped until no more can be deleted without degradation of the required performance. Under this process, the final adequate decision structure 46 (FIG. 2) for subject RA was reduced to two baseline and one preseizure channels each requiring five linear and five squared input signals. On the other hand, the final adequate decision structure 48 (FIG. 3) for the subject BJ was reduced to a single channel wherein the second layer TLU became unnecessary and was omitted for further simplification of the structure. The first layer threshold logic units 50, 52 and 54 in the decision structure 46 can be considered to be feature extractors or detectors, and the second layer response unit 58 can be considered to be a decision element or classifier. In the decision structure 48, the input weights and summing amplifier 60 can be considered to be the feature detector, and the threshold .theta..sub.5 and Schmitt trigger 62 can be considered to be the decision element.

For the RA decision structure 46 in FIG. 2, all of the input signals X.sub.1 through X.sub.5 and X.sub.1.sup.2 through X.sub.5.sup.2 are applied to each of the three first-layer summing amplifiers 60 through respective weights W.sub.1 through W.sub.10, W.sub.11 through W.sub.20 and W.sub.21 through W.sub.30. The thresholds .theta..sub.1 through .theta..sub.4 are established by respective weights W.sub.31 through W.sub.34. The output signals of the three Schmitt triggers 62 are applied to the second-layer summing amplifier 60 through respective weights W.sub.35 through W.sub.37. For the BJ decision structure 48 in FIG. 3, only the input signals X.sub.1 ' and X.sub.2 ' are required and are applied through respective weights W.sub.38 and W.sub.39 to the summing amplifier 60 having a threshold .theta..sub.5 established by the weight W.sub.40. The weights are, for example, as follows.

______________________________________ W.sub.1 = -24.7 W.sub.21 = +15.1 W.sub.2 = -37.5 W.sub.22 = +10.8 W.sub.3 = -18.3 W.sub.23 = +17.3 W.sub.4 = +68.6 W.sub.24 = -51.2 W.sub.5 = +27.6 W.sub.25 = -32.7 W.sub.6 = -20.8 W.sub.26 = - 2.45 W.sub.7 = +30.6 W.sub.27 = -22.4 W.sub.8 = + 5.79 W.sub.28 = -19.4 W.sub.9 = -109.8 W.sub.29 = +70.2 W.sub.10 = -55.6 W.sub.30 = +19.2 W.sub.11 = -16.8 W.sub.31 = +36.8 W.sub.12 = -22.9 W.sub.32 = + 3.53 W.sub.13 = + 6.66 W.sub.33 = - 4.51 W.sub.14 = +77.4 W.sub.34 = - 4.96 W.sub.15 = +70.0 W.sub.35 = + 6.09 W.sub.16 = - 7.87 W.sub.36 = + 3.54 W.sub.17 = +52.5 W.sub.37 = - 1.69 W.sub.18 = +30.0 W.sub.38 = +27.241 W.sub.19 = -152.2 W.sub.39 = -91.228 W.sub.20 = -48.3 W.sub.40 = +95.884 ______________________________________

FIG. 4 is a block diagram of the warning logic means 42 showing the configuration of audible and visible alarm circuitry portion of the warning system 30 (FIG. 1). The stream of successive 15-second decisions output from the decision structure 38 is stored in a 20-bit shift register 64. A summing amplifier 66 is connected to the 20 parallel outputs of the shift register 64 and serves to count the number of preseizure decisions occuring in the preceding five minutes.

A voltage regulator 68 is provided in connection with the summing amplifier 66 to compensate for battery supply voltage variations. The output of the summing amplifier 66 is applied to a threshold comparator 70 which governs the activation of pacer 72 that energizes the audio and visual alarms 74 and 76. The output of the comparator 70 also conditions the flip-flop 78 so that the audio alarm 74 can be energized. A pushbutton switch 80 can be operated to inhibit the audio alarm 74 after it is energized by resetting the flip-flop 78. When the output of comparator 70 changes from a preseizure conditin, timer 82 operates to deactivate the pacer 72 after, for example, 5.5 minutes.

A circuit was also needed to provide signals for timing the 15-second analysis intervel, resetting the filter integrators, and clocking the decision outputs into the warning system shift register (see the timing and control block diagram, FIG. 13). The 15-second timing is provided by an integrator and Schmitt trigger. The required pulses are produced by a digital storage register and logic gates connected in a "one-shot" configuration. Additional circuitry provides capability for pushbutton resetting of the system, including clearing of the warning logic memory.

In deriving the basic configuration of the system, some of the more significant criteria were low power consumption, number of components, accuracy of analog operations, and insensitivity to battery supply voltage variations.

FIG. 5 is a circuit diagram showing the scalp electrode connections 84, 86 and 88 for EEG acquisition and the preamplification means 32 which can be used in the epileptic seizure warning system for the subject BJ. Central and occipital placement of electrodes 84 and 88 can be used with a C.sub.z reference (behind the ear) electrode 86. A commercially available EEG preamplifier 90 (BioCom Model 121 IC) was obtained for the system. This unit 90 consumes more power than desired and does not provide as high common mode rejection ratio (CMRR) as certain other preamplifiers, but it produces satisfactory EEG traces.

Gain amplifiers 92 and 94 are connected to the output of the preamplifier 90. The amplifier 92 includes a low pass filter 96. The output of the amplifier 94 is connected by a normally closed switch 98 to the bandpass filters shown in FIG. 6. A jack socket 100 is mechanically coupled to the switch 98. When the jack of a tape recorder (not shown) is plugged into the socket 100, the switch 98 is automatically opened to disconnect the preamplification means 32.

FIG. 6 is a circuit diagram of the signal preprocessing and processing means 102 and 104 of the seizure warning system for the subject BJ. The preprocessing means 102 includes filters 106 and 108 which can be operational amplifiers having bridged-T filter feedback networks 110 and 112. The processing means 104 includes rectifiers 114 and 116, integrators 118 and 120, and scaling amplifiers 122 and 124 for the two frequency bands involved in this warning system.

The rectifiers 114 and 116 can be operational amplifier circuits including diodes CR1 and CR2 which apply respective rectified signals to the integrators 118 and 120. The integrating capacitors 126 and 128 are respectively discharged by the field effect transistors Q5 and Q6 when a dump signal is provided on their gates at the end of each 15-second epoch. The scaling amplifiers 122 and 124 normalize the integrated signals from the integrators 118 and 120. The -V reference bias and its scaling resistor pairs 130 and 132 serve to subtract a predetermined mean of the energy in each frequency component and, since gain is the reciprocal of division, the gain of the operational amplifiers 122 and 124 is suitably selected to divide the mean by the standard deviation of that component for normalization.

In order to test for differences more easily between epochs or periods temporally associated with seizures, a long period referred to as a "baseline" period is chosen from each subject's EEG which (1) did not precede or follow a seizure by less than several hours, (2) did not contain a seizure, (3) was recorded under similar physical conditions as the seizure (although for the overnight records, several stages of sleep may be included), and (4) was relatively free of artifacts. The mean and standard deviation of the energy in each frequency component of the spectrum can be estimated using the energies in the numerous epochs of these long records. Normalization highlights shifts from baseline averages for a subject.

FIG. 7 is a circuit diagram showing the supply and stabilizing connections of virtually all of the operational amplifiers used in the seizure warning system for the subject BJ. The only exception is in the low frequency oscillator amplifier in FIG. 14, where the pin 8 is connected through a one megohm resistor to the output of the timer threshold comparator. All of the operational amplifiers are, for example, of the type FU5B7776393.

FIG. 8 is a circuit diagram showing the decision structure 48 which can be used in the seizure warning system for the subject BJ. A single threshold logic unit with only two inputs, one positive and one negative, is utilized. The summing amplifier 60 (FIG. 3) is embodied in an operational amplifier 134, and the Schmitt trigger 62 (threshold unit or classifier) can be constructed from two NAND gates 136 and 138. The diode CR3 protects the input to the Schmitt trigger 62 which has an inverter 140 output. The weights W.sub.38, W.sub.39 and W.sub.40 of FIG. 3 have been converted into appropriate resistances by suitable application of the pertinent scaling factors.

FIG. 9 is a graph showing plots 142 and 144 of the output voltage characteristics for the bandpass filters 106 and 108 (FIG. 6) centered at 9.6 and 25.9 Hz. Each of the filters 106 and 108 utilizes a single operational amplifier with a nulling circuit in the feedback path as previously described. This filter provided the desired sharp peak in the center of the band and dropped off sharply outside of the passband. Only one active device and a relatively small number of components are required. The circuit also produced a substantial voltage gain, reducing the requirements imposed on the EEG preamplifier 32 (FIG. 5).

To determine the energy within their passbands, the outputs of the filters 106 and 108 are subjected to suitable rectification and integration. By using a half-wave rectifier requiring only one amplifier for each filter to keep the parts count down, an approximate energy determination is made. The half-wave rectifier functioned satisfactorily; however, full-wave rectification of the filter outputs is preferred to determine the energy within the passbands more accurately. The full-wave rectifier designs can, of course, employ two operational amplifiers.

The integrators 118 and 120 are of substantially conventional design; however, considerable time was required to obtain satisfactory accuracy. It was found that the critical element was selection of a capacitor with the lowest possible dissipation factor. Capacitors fabricated with polycarbonate or polystyrene dielectrics were found to be suitable but very bulky. No smaller capacitors were found that would give satisfactory performance as the integrating component. Each integrator includes a circuit that resets the initial conditions every 15 seconds by dumping the charge on the integrator capacitor through a field effect transistor (FET). The integrator also requires very careful balancing of the amplifier input bias currents to obtain accurate integation.

FIG. 10 is a circuit diagram of a voltage regulation circuit 146 which provides a stable voltage reference for the integrators 118 and 120 (FIG. 6). The stable voltage reference is required since the integrator balance control (potentiometers 148 and 150 in FIG. 6) is sensitive to changes in battery supply voltage. This voltage reference is also used in other voltage-sensitive portions of the system (timing circuit, decision threshold logic units, and warning logic circuit). The operational amplifier 152 measures the voltage drop across transistor Q7. Since the voltage drop across transistor Q7 is stable, then the output of the amplifier 152 will be stable.

Scaling and normalization of each of the integrator outputs is accomplished by adjusting the gain and bias of the operational amplifiers 122 and 124 (FIG. 6) to reflect values of mean and standard deviation previously calculated by the computer for baseline data for each filter as previously described.

The digital simulation of the seizure warning system for subject RA required, in addition to the normalized filter output, the square of that value for each filter. Historically, analog operations involving addition and subtraction have been relatively easily accomplished using operational amplifiers. Multiplication (including squaring) and division have been much more difficult and have involved cumbersome or inaccurate equipment such as synchrodriven potentiometers, Hall effect devices (interaction between a current and a magnetic field in a semiconductor), diode function generators, and log-antilog devices. Recently, integrated circuits which are small and accurate have been developed to perform multiplication base upon the principle of variable transconductance.

FIG. 11 is a circuit diagram of a squaring circuit 154 including one of these devices, the Motorola MC 1594 which maintained an accuracy of 1/2 percent of full scale over its full range of operation. In addition, the new circuit was found to be much more accurate at low input levels, required fewer external components to trim the operating conditions, and proved to be less prone to oscillation. The circuit 154 did require special attention to prevent oscillation. Input leads were twisted and lead lengths were kept as short as possible. Decoupling capacitors were used on the power supplies and input leads. Potentiometers 156 and 158 can be adjusted to balance the circuitry to obtain a true square output, and variable resistor 160 can be used to adjust the gain of the operational amplifier 162.

The primary disadvantage of the squaring circuit is power consumption. Each circuit draws 250 milliwatts. Thus, the five squaring circuits 40 (FIG. 1) required for the RA system draw more power (1.25 watts) than the entire remainder of the seizure warning system (0.3 watt), which is composed mainly of micropower amplifiers and low-power digital logic. The relatively large battery pack needed for the RA system is required primarily to supply power to these five devices for a 10-hour period. The BJ system does not use squared inputs; consequently, power requirements for BJ are greatly reduced.

It may be noted that in implementing the decision structure 46 (FIG. 2) of subject RA, the only major problem encountered was in obtaining sufficient accuracy with teninput summing amplifiers to approximate the simulated decision threshold logic units. The goal established was 1/2 percent of full-scale accuracy for all combinations of inputs. It was impossible to test all possible inputs, but tests were conducted using many combinations of 1 volt and 0 bolt inputs, as well as varying voltage levels for individual inputs. During these tests, input weighting resistors were selected, replaced, or trimmed to achieve the desired accuracy for almost all combinations tried.

FIG. 12 is a circuit diagram of an illustrative summing circuit 164 with four inputs and which implements both negative and positive input weights. The equation for the output voltage as a function of the input voltages is indicated below the summing circuit. This summing circuit 164, with more or less inputs, can be used as the summing amplifiers 60 in the decision structure 46 of FIG. 2. The threshold voltage input is critical and can be supplied by the voltage regulation circuit 146 (FIG. 10) described earlier.

The summing amplifier 60 of the response unit 58 in FIG. 2 presented a much easier design problem than the summing amplifiers 60 of the units 50, 52 and 54. Inasmuch as the three inputs are limited to values of 1 or 0, there are only eight possible input combinations, each of which may be checked for the proper output decision. An output level of 1 (4 volts) corresponds to a preseizure decision while 0 (0 volts) indicates a baseline period.

FIG. 13 is a circuit diagram of an embodiment of the timing and control means 44 indicated in the warning system of FIG. 1 and which is used with the circuits shown in FIGS. 6 and 14. Two choices were evaluated for timing the 15-second analysis epoch. Oscillators, especially those which are crystal controlled, provide an However, means for setting timing accurately. however, the associated counting and triggering circuitry would have consumed more power than was considered desirable. The second alternative consists of integrating a constant voltage and comparing the output with a fixed threshold. When the threshold is exceeded, a chain of pulses is triggered, one of which dumps the charge on the capacitor 166 of integrator 168 to initiate the next analysis epoch. The 15-second timing is achieved by adjusting the input resistor 170 or the output threshold voltage. A Schmitt trigger 172 performs the output voltage comparison.

The integrator approach was selected primarily because of the lower power consumption; however, recent advances in timing technology, particularly crystal-controlled clocks, may make the oscillator and counting circuitry more feasible in future applications.

One-shot multivibrators are the natural choice for providing the short timing pulses. However, no commercially available integrated circuit one-shot was uncovered with a power consumption of less than 80 mw. Therefore, one-shots 174 and 176 were constructed using a pair of low-power inverters and a NAND gate with suitable capacitive coupling as shown in FIG. 13. The duration of the one-shot pulse is controlled by the value of the coupling capacitor.

The low-power flip-flop storage element 178 is required to allow the timing integrator capacitor 166 to discharge fully. At the end of the 15-second epoch, the timing trigger 172 sets the flip-flop 178 and initiates a chain of two one-shot pulses. The first pulse from one-shot 174 clocks the output of the decision unit 48 (FIG. 8) into the warning logic shift register 64. (FIG. 14). As the decision is stored, a signal due to the first pulse from the Q4 circuit dumps the filter integrator capacitors 126 and 128 (FIG. 6) in preparation for the next analysis epoch. The second pulse from one-shot 176 clears the flip-flop 178, unlocking the timing integrator 168.

Two additional one-shots 180 and 182 are included to provide a capability for clearing and resetting the entire system with an external pushbutton 184. This feature was very useful in system checkout but is not available to the subject being monitored to prevent accidental system interruption. The circuit 186 centered around transistor Q3 was provided to ensure that the system will start reliably when the power switch is turned on.

When the power is first turned on, the transistor Q3 conducts such that the base of transistor Q1 is connected to ground by way of the diode CR4. Transistor Q1 conducts and back biases diode CR5 so that field effect transistor Q2 is made conductive to assure discharge of the 15-second integrating capacitor 166. The transistor Q3 becomes nonconducting when the capacitors in circuit 186 are suitably charged, and the transistors Q1 and Q2 are also rendered nonconducting. The rise in potential of the collector of transistor Q3 further assures that the Q output of flip-flop 178 is set to a 0 (low or ground potential) output condition.

Closing of the momentary bushbutton switch 184 produces a positive output pulse from the one-shot 180. This (high) pulse is applied through inverter 188 as a clear (low) signal to the shift register 64 (FIG. 14). The high pulse is also applied to one-shot 182 which produces an inverted (low or 0) pulse on pin 10 of the NAND gate 190. Since the pin 9 of gate 190 is normally high in potential (when the capacitor 166 is charging and the Schmitt trigger 172 is not triggered), an output pulse is produced which is inverted by inverter 192 and applied to the preset pin 4 of the flip-flop 178 to produce a 0 (low) output from pin 6 and a 1 (high) output from pin 5 or the Q output. The low output on pin 6 turns on transistor Q1 to discharge the integrating capacitor 166.

The high output on pin 5 of the flip-flop 178 produces a 1 millisecond output pulse from the one-shot 174. This pulse is applied too the shift register 64 (FIG. 14) to shift everything in each register and clock in the output from the decision structure 48 (FIG. 8). The output pulse from the one-shot 174 is inverted by inverter 194 and applied to transistor Q4 and the one-shot 176. The transistor Q4 is rendered conductive so that a low (ground) dump signal is produced and applied to the gates of the field effect transistors Q5 and Q6 (FIG. 6) to discharge the capacitors 126 and 128. The one-shot 176 produces a 2 millisecond pulse which is inverted by inverter 196 and applied through diode CR6 to the gate of transistor Q1 without effect since a low signal is already being applied from pin 6 of the flip-flop 178.

The trailing edge of the low pulse from inverter 196 through diode CR6 does result in a rise on the clock input of flip-flop 178, however, so that its 0 (ground) data input sets the Q output (pin 5) to a 0. Of course, the output on pin 6 of the flip-flop 178 goes high such that the transistors Q1 and Q2 are turned off and the capacitor 166 begins charging again. After 15 seconds, the threshold of the Schmitt trigger 172 is exceeded to produce a low output signal which is applied to pin 9 of gate 190. This results in a high output signal from the gate 190, which is inverted by inverter 192 and applied to the preset input (pin 4) of the flip-flop 178 to produce a low output on its pin 6 and a high output on its pin 5 to repeat the cycle.

FIG. 14 is a circuit diagram of the warning logic means (and alarm circuitry) 42 connecting with the circuits of FIGS. 8 and 13. The warning logic is designed to satisfy four basic criteria.

1. To reduce the number of false alarms, the number of 15-second preseizure decisions occuring with a 5-minute period are counted. If a specified number of preseizure indications is exceeded (18 out of 20 for subject BJ), a warning must be initiated.

2. The warning must continue as long as this level is exceeded.

3. After the last indication of warning condition (count drops below 18 for BJ), the subject should remain in a position of safety for a period of 5.5 minutes, for example.

4. The alarm should be sufficiently annoying to alert the subject; but once he acknowledges the warning, he should be able to suppress the audible signal while a less conspicuous visual alarm would continue.

A 20-bit shift register 64 serves as the memory for the warning logic system 42 (FIG. 14). The output of the decision unit 48 (FIG. 8) is clocked into the shift register 64 at the end of each 15-second analysis epoch. The 20 parallel outputs of the shift register 64, representing the output decisions for the immediately preceding five minutes, are input to a summing amplifier 66 through equally weighted resistors 198. Thus, the output of the summing amplifier 66 is directly proportional to the number of preseizure decisions in the 20-bit memory. The number of counts required to trigger a seizure warning is controlled by the value of the threshold input resistor 200.

When the output of the summing amplifier 66 goes negative and the output of threshold comparator 70 goes positive, indicating a warning condition, two actions take place. First, the field effect transistor Q11 of timer 82 is switched on so that timing integrator 202 is set to an initial value and held there by the FET switch. The new (negative) output of the integrator 202 produces a negative output from threshold comparator 204 to activate a low frequency (about 5 Hz) oscillator or pacer 72. The pacer 72 drives transistor Q8 which controls a tiny light 76, causing it to flash. The second action is to set a flip-flop 78 which, in turn, permits gating of the output of the flashing circuit 72 to control a commercially available audible device 74 incorporated into the system. The 1 (high) data input on pin 12 of flip-flop 78 is clocked or set to the pin 9 by the rise in potential on pin 11. Gate 206 can, therefore, produce an oscillating high and low output so that the control transistor Q9 is turned on and off. Thus, when the warning is initiated, the subject is presented with a flashing light and an audible beeping.

At this point, the subject has a choice. If the subject takes no action, the alarm 74 will continue to beep as long as the light 76 is flashing; however, the subject may press a button 80 that clears the flip-flop 78, suppressing the audible signal only. This action has no effect on the operation of the flashing light 76, and the flip-flop 78 will retrigger the next time a warning condition is initiated. This feature precludes the possibility of the subject's suppressing the alarm but forgetting to turn it back on after the warning subsides.

As long as the warning condition persists, the timing integrator 202 is locked to its initial condition. When the number of preseizure counts drops below the preset threshold, the FET switch Q11 is turned off and the timer capacitor 208 is free to integrate. The rate of integration is adjusted so that 5.5 minutes later the output voltage of integrator 202 crosses the comparator 204 threshold and the comparator output goes positive, turning off the pacer 72 and flashing light 76. Should a new warning be initiated during this period, the integrator 202 will be clamped to its initial value again, the light 76 will continue to flash, and the audible alarm 74 will be retriggered. All components, with the exception of the light 76 itself, are low-power devices.

One of the design difficulties encountered was that the output voltages of the 20 parallel bits did not change in direct proportion to battery supply variations. As the supply voltage decreased, the number of counts required to trigger a warning changed. To overcome this problem, the threshold input voltage is supplied by regulator circuit 68 including a logic inverter 210 whose input is tied to ground. This provides a dummy reference voltage from the operational amplifier 212 that varies exactly as the shift register 64 output varies with battery supply changes.

The power requirement of subject RA's warning system is approximately 1.6 watts. Approximately 75 percent of this power is used to operate the five integrated circuits that provide the square of the filter outputs. The remainder of the system requires approximately 350 milliwatts. These squaring circuits 154 (FIG. 11) also require .+-. 15 volts, so the battery pack was designed to supply four nominal voltages, .+-. 5v and .+-. 15v. The power supply consists of six 450-milliampere-hour rechargeable nickel-cadmium batteries and has the capacity to provide 10 hours of continuous operation on a single charge. The battery supply is about 4 .times. 6 .times. 2 inches and weighs about 2-1/2 pounds.

The BJ system posed a completely different supply problem. Power consumption for this system is significantly lower, and the device itself can be much smaller. A single package which is small enough to fit in a shirt pocket was used to contain the system electronics, battery supply, and warning means. The system electronics requires 140 milliwatts and the flashing light in the alarm system uses an additional 100 milliwatts. Power for the BJ system is supplied by two 5.4 volt dispoable mercury cell batteries. The batteries selected each have a capacity of 1,000 milliampere-hours, which is sufficient to drive the systems for over 40 hours of continuous monitoring. They occupy a space of about 2-1/4 cubic inches and weigh less than 4 ounces.

The case for the compact BJ seizure warning system is tapered to fit comfortably in a shirt pocket. It is constructed in two sections of molded fiberglass. The circuit board, battery holders, and power switch are mounted on the back section. The front cover holds the audible alarm 74 (FIGS. 4 and 14), the EEG harness socket, and the alarm supression button 80. The flashing light 76 is mounted in the top of the back cover, where the subject may easily check the status of the alarm by a downward glance at the device. The device also contains a tiny socket 100 (FIG. 5) that allows operation of the system on tape recorder outputs for test and demonstration purposes. When the tape recorder jack is plugged in, the EEG preamplifier 32 is automatically disconnected. The case is 3-1/8 .times. 6-1/2 inches and the average depth or thickness is 1 inch. The entire system including batteries weighs 13 ounces. During tests with the subject, the system has been subjected to severe shocks and crushing loads. The case has withstood these rigors remarkably well.

The principles involved in the epileptic seizure warning system are, of course, applicable to other systems in addition to those for providing warnings of imminent epileptic seizures. Such principles can be directly applied to, for example, a cardiac seizure warning system for providing warnings of any imminent heart attacks. The principles can also be applied to automatic diagnostic devices. Thus, while an exemplary embodiment of this invention has been described above and shown in the accompanying drawings, it is to be understood that such embodiment is merely illustrative of, and not restrictive on, the broad invention and that we do not desire to be limited in the scope of our invention to the details of construction or arrangements described and shown, for obvious modifications may occur to persons having ordinary skill in the art.

Claims

We claim:

1. A warning activation system comprising:

means for providing an electrical signal characteristic of the condition of a component part of a subject;

means for preprocessing said electrical signal to transform it into a predetermined format;

means for processing said transformed signal to measure energy characteristics thereof;

means for detecting predetermined features indicative of a relatively early abnormal condition of said component part of said subject, when predisposed to occurrence of an impending and disturbing event, from said measured energy characteristics of said transformed signal;

means for deciding from said detected features the condition of said component part of said subject and providing an output of a predetermined category for said abnormal condition thereof;

means for controlling said processing means to process said transformed signal reiteratively over successive sampling epochs thereof whereby decisions on the condition of said component part of said subject can be made for said epochs by said decision means; and

means responsive to the output of said decision means for providing an activation signal in advance of the occurrence of said event when said decision means output is of said predetermined category for at least a predetermined number of said epochs within an immediately preceding predetermined period.

2. The invention as defined in claim 1 wherein said preprocessing means includes a bank of bandpass filters having different predetermined passbands to provide a plurality of signal components having resepctive predetermined bandwidths, and said processing means includes a corresponding plurality of means for measuring respective energy characteristics of said signal components, said plurality of measuring means being controlled by said controlling means, said detecting means detects predetermined features from said measured energy characteristics of said signal components and said decision means decides from said detected features the condition of said component part of said subject.

3. The invention as defined in claim 1 wherein said processing means includes rectifying, integrating and normalization means to measure said energy characteristics of said transformed signal.

4. The invention as defined in claim 1 wherein said activation signal providing means includes shift register means of a predetermined stage capacity corresponding numerically to said epochs within said predetermined period, for receiving and storing said decision means output over said successive sampling epochs therein and producing said activation signal when said decision means output is of said predetermined category for at least said predetermined number of said epochs within said predetermined period.

5. The invention as defined in claim 2 wherein said plurality of measuring means each includes rectifying, integrating and normalization means to measure said energy characteristics of said transformed signal, and said activation signal providing means includes shift register means of a predetermined stage capacity corresponding numerically to said epochs within said predetermined period, for receiving and storing said decision means output over said successive sampling epochs therein and producing said activation signal when said decision means output is of said predetermined category for at least said predetermined number of said epochs within said predetermined period.

6. An epileptic seizure warning system comprising:

means for providing an electrical signal characteristic of the brain activity of a subject;

means for preprocessing said electrical signal to transform it into a predetermined format;

means for processing said transformed signal to measure energy characteristics thereof;

means for detecting predetermined features indicative of a relatively early preseizure condition of said subject, when predisposed to an impending seizure, from said measured energy characteristics of said transformed signal;

means for deciding from said detected features the condition of said subject and providing an output of a predetermined category for said preseizure condition thereof;

means for controlling said processing means to process said transformed signal reiteratively over successive sampling epochs thereof whereby decisions on the condition of said subject can be made for said epochs by said decision means; and

means responsive to the output of said decision means for providing a preseizure warning signal in advance of the occurrence of said seizure when said decision means output is of said predetermined category for at least a predetermined number of said epochs within an immediately preceding predetermined period.

7. The invention as defined in claim 6 wherein said warning signal includes a visual signal, and said warning signal providing means further includes means for providing a simultaneous audio signal when said decision means output is of said predetermined category for at least said predetermined number of said epochs within said predetermined period.

8. The invention as defined in claim 7 wherein said warning signal providing means further includes means for electively suppressing said audio signal after said visual and audio signals are initiated, and means for maintaining said visual signal for a predetermined duration after said decision means output is no longer of said predetermined category for at least said predetermined number of said epochs within said predetermined period.

9.

A warning activation system comprising:

means for providing an electrical signal characteristic of the condition of a component part of a subject;

pattern recognition means for analyzing and classifying said characteristic signal and providing an output of a predetermined category for a relatively early abnormal condition of said component part of said subject when predisposed to occurrence of an impending and disturbing event;

means for controlling said recognition means to analyze and classify said characteristic signal reiteratively over successive sampling epochs thereof whereby classifications on the condition of said component part of said subject can be made for said epochs by said recognition means; and

means responsive to the output of said recognition means for providing an activation signal in advance of the occurrence of said event when said recognition means output is of said predetermined category for at least a predetermined number of said epochs within an immediately preceding predetermined period.

10. The invention as defined in claim 9 wherein said activation signal providing means includes shift register means of a predetermined stage capacity corresponding numerically to said epochs within said predetermined period, for receiving and storing said recognition means output over said successive sampling epochs therein and producing said activation signal when said recognition means output is of said predetermined category for at least said predetermined number of said epochs within said predetermined period.

11. A method of providing an activating control signal, which comprises the steps of:

sensing an electrical signal characteristic of the condition of a component part of a subject;

analyzing and classifying said characteristic signal by pattern recognition means and providing an output therefrom of a predetermined category for a relatively early abnormal condition of said component part of said subject when predisposed to occurrence of an impending and disturbing event;

controlling said recognition means to analyze and classify said characteristic signal reiteratively over successive sampling epochs thereof whereby classifications on the condition of said component part of said subject can be made for said epochs by said recognition means; and

producing an activation signal for energizing control means in advance of the occurrence of said event when said recognition means output is of said predetermined category for at least a predetermined number of said epochs within an immediately preceding predetermined period.

12. The invention as defined in claim 11 wherein said activation signal is produced by receiving and storing said recognition means output over said successive sampling epochs in shift register means of a predetermined stage capacity corresponding numerically to said epochs within said predetermined period, and generating said activation signal when said recognition means output is of said predetermined category for at least said predetermined number of said epochs within said predetermined period.