CA 02207330 l997-06-06 W O 96/17S47 PCTrUS9S~15172 NEURAL NETWORK METHOD AND APPARATUS FOR DISEASE, ~ INJURY AND BODILY CONDITION SCREENING OR SENSING
Technical Field The present invention relates generally to a method and apparatus for 5 screening or sensing disease states, injury sites or bodily conditions in a living org~ni~m by ~let~cting the DC biopotential of the electromagnetic field present between a reference and a plurality of test points on the living or~ni~m to measure the gradient of electrical activity which occurs as a function of biological activity.
10 Back~round Art In recent years the theo~y that measurement of the potential level of the electromagnetic field of a living org~ni.~m can be used as an accurate screening and diagnostic tool is g~inin~ greater acceptance. Many methods and devices have been developed in an attempt to implement this theory.
For example, U.S. Patent No. 4,328,809 to B.H. Hirschowitz et al. deals with a device and method for ~letÁcting the potential level of the electromagnetic field present between a reference point and a test point on a living org~ni~m. Here, a reference electrode and a test electrode provide DC signals indicative of the potential level of the electromagnetic field 20 measured between the Icfelellce point and the test point. These signals are provided to an analog-to-digital converter which generates a digital signal as a function thereof, and a processor provides an output signal indicative W O96117547 PCTrUS95/15172 of a parameter or parameters of the living org~ni~m as a function of this digital signal.
Similar biopotential measuring devices are shown by U.S. Patent Nos. 4,407,300 to Davis, and 4,557,271 and 4,557,273 to Stroller et al.
S Davis, in particular, discloses the diagnosis of cancer by measuring the electromotive forces generated between two electrodes applied to a subject.
Often, the measurement of biopotentials has been accomplished using an electrode array, with some type of multiplexing system to switch between electrodes in the array. The aforementioned Hirschowitz et al.
10 patent contemplates the use of a plurality of test electrodes, while U.S.
Patent Nos. 4,416,288 to Freeman and 4,486,835 to Bai disclose the use of measuring electrode arrays.
Unfortunately, previous methods for employing biopotentials measured at the surface of a living org~ni~m as a diagnostic tool, while 15 basically valid, are predicated upon an overly simplistic hypothesis which does not provide an effective diagnosis for many disease states. Prior methods and devices which implement them operate on the basis that a disease state is indicated by a negative polarity which occurs relative to a reference voltage obtained from another site on the body of a patient, while 20 normal or non-malignant states, in the case of cancer, are indicated by a positive polarity. Based upon this hypothesis, it follows that the detection and diagnosis of disease states can be accomplished by using one measuring electrode situated externally on or near the disease site to provide a measurement of the polarity of the signal received from the site relative to 25 that from the reference site. Where multiple measuring electrodes have been used, their outputs have merely been summed and averaged to obtain one average signal from which a polarity determination is made. This W O 96/17547 PCTnUS9StlS172 approach can be subject to major deficiencies which lead to diagnostic inaccuracy, pa~ticularly where only surface measurements are taken.
First, the polarity of diseased tissue underlying a recording electrode has been found to change over time. This fact results in a potential change 5 which confounds reliable diagnosis when only one external recording electrode is used. Additionally, the polarity of tissue as measured by skin surface recording is dependent not only upon the placement of the recording electrode, but also upon the placement of the reference electrode.
Therefore, a measured negative polarity is not necessarily indicative of 10 diseases such as cancer, since polarity at the disease site depends in part on the placement of the le~ ellce electrode.
When many electrodes are used to sense small DC biopotenti~lc at the surface of the skin, such as in screening for breast cancer, it is crucial for the sensing electrodes to be accurately placed and spaced so that two 15 adjacent electrodes are not overlapping or sensing the same test area. If these tolerances are not accurately m~int~ined, false readings are likely to result.
As disease states such as cancer progress, they produce local effects which include changes in vascularization, water content, and cell division 20 rate. These effects alter ionic concentrations which can be measured at the skin surface and within the neoplastic tissues. Other local effects, such as distortions in biologically closed electrical circuits, may occur. A key point ~ to recognize is that these effects do not occur uniforrnly around the disease site. For example, as a tumor grows and dir~~ Li~tes, it may show wide 25 variations in its vascularity, water content and cell division rate~ depending on whether ex~min~tion occurs at the core of the tumor (which may be necrotic) or at the margins of the tumor (which may contain the most W O 96/17547 PCTrUS95/15172 metabolically active cells). The tumor may not respond significantly to growth factors, while the growth factors and the enzymes produced may significantly affect the normal cells surrounding the tumor. This fact was recognized by one of the present inventors, and his U.S. Patent Nos.
4,995,383 and 5,099,844 disclose a method and ~a~dL-Is which implement the principle that important electrical indications of disease are going to be seen in the relative voltages recorded from a number of sites at and near a diseased area, and not, as previously assumed, on the direction (positive vs.
negative) of polarity.
Still, the accurate measurement of DC biopotentials for sensing or screening for disease, injury or bodily functions is very difficult to accomplish, for the DC potÁnti~l~ to be sensed are of a very low amplitude.
Due to factors such as the low DC potentials involved and the innate complexity of biological systems, the collected data si~nals tend to include a substantial amount of noise which makes accurate analysis difficult. Also.
biological systems are notorious for their complexih. nonlinearity and nonpredictability, and wide variations from the norm are not uncommon.
Thus it is necessary to develop a method and apparatus for obtaining the necessary data from the measurement of biopotentials and then to extract and analyze pertinent information which is relevant to a condition under study.
Disclosure of the Invention It is a primary object of the present invention is to provide a novel and improved method and a~pa~ s for condition screening or sensing wherein DC biopotentials from the area of a site on a living org~ni~m are -W O96117547 PCT~US95/15172 measured and then processed in a neural network which has been taught to recognize information p~ rn~ indicative of a particular condition.
A filrther object of the present invention is to provide a novel and improved method and ~dlus for disease, injury or bodily condition S screening or sensing wherein DC biopotentials are received on separate channels from a plurality of sites at and near a suspected area of disease, injury or condition change on a living org~ni~m. A maximum potential di~felcllLial is then obtained from the averages of multiple biopotential values taken over time and subsequently provided to a neural network which has been taught to recognize patterns indicative of a disease, injury or other bodily condition.
Yet a further object of the present invention is to provide a novel and improved method and al)p~Lus for disease. injury or condition screening or sensing wherein DC biopotentials are received from a plurality of measuring sensors located in the area of a suspected disease, injury or condition change site. These potentials are then specifically provided to a particular type of neural network or a combination of neural networks uniquely adapted to receive and analyze data of an identifi~kle type to provide an indication of specific conditions.
Another object of the present invention is to provide a novel and improved method and apparatus for disease~ injury or condition screening or sensing wherein dir~lelllial values are derived from DC biopotÁnti~l~
located in the area of a suspected disease. injury. or condition change site.
These potÁnti~l~ are then provided to a plurality of neural networks and the outputs of these neural networks are then used to provide an indication of specific conditions.
W O 96/17547 PCTrUS95/15172 A still further object of the present invention is to provide a novel and improved method and a~paldLus for condition sensing or sensing wherein dirr~,lel~Lial values are derived from DC biopotentials obtained from mirror image sensors located on the opposite breasts of a subject.
S Brief Description of the Drawin~s Figure 1 is a block diagram of the general apparatus of the present invention;
Figure 2 is a diagram of the breasts of a human subject which receive the sensor arrays of Figure l;
Figure 3 is a generalized sectional diagram of the electrode which can be used as a sensor for the apparatus of Figure l;
Figure 4 is a flow diagram of the measurement operation of the a~,uald~ls of Figure l;
Figure 5 is a block diagram of a probabilistic neural network used 15 with the apparatus of Figure l;
Figure 6 is a block diagram of a general regression and neural network used with the apparatus of Figure l;
Figure 7 is a block diagram of a plurality of neural networks used with the apparatus of Figure l;
W O96/17~47 PCTrUS9~/15172 Figure 8 is a plan view of a disposable skin removal unit of the present invention; and Figure 9 is a sectional view of a mechanical skin removal unit of the present invention.
Best Mode for Carryin~ Out the Invention Figure 1 discloses a basic block diagTam of an apparatus indicated generally at 10 for performing a discriminant analysis to obtain both raw data signals and dirr~,clllial signals for a pattern recognition device that then discrimin~tec between patterns to achieve disease, injury or other condition screening or sensing. For purposes of illustration, the apparatus 10 will be described in connection with methods involving the screening for, or diagnosing of breast cancer. However, it should be recognized that the method and a~)palaL~ls of the invention can be similarly employed for screening or diagnosis at other sites involving other conditions of a living human or ~nim~l. For example, the apparatus and method to be described can be used to detect disease conditions such as infection~ ischemia, spasm, arthritis or other injury, or non-disease conditions such as ovulation, labor, abnormalities of labor, and fetal distress.
In Figure 2, a human subject 12 may have a cancerous lesion 14 on one breast 16. This cancerous lesion has a core 18 and an outer zone 20 surrounding the core where various differing local effects. such as changes in vascularization, water content and cell division rate occur. The outer zone 20 will include normal cells surrounding the lesion, for these cells often exhibit a much greater biopotential effect in response to tumor growth W O96/17547 PCTrUS95/15172 than does the actual tumor. Assuming first, for purposes of discussion, that the location of the lesion 14 is not known, and the device 10 is to be used to screen the breast 16 to determine whether or not a disease condition exists, skin surface pot~nti~l~ will be measured in an area of the breast, S including the zone 20 using an electrode array 22. The device and method of this invention contemplate the use of a variety of different sensor arrays and even the use of different types of sensors within an array depending upon the intended application for which the device 10 is used. For example, in the diagnosis of clinically symptomatic breast or skin lesions, 10 the sensor array should cover various areas of the lesion as well as relatively normal tissue near the lesion site. For breast cancer screening (where patients are asymptomatic) the array should give maximum coverage of the entire breast surface. The aim in both of these cases is to measure the gradient of electrical activity which occurs as a function of the 15 underlying biological activity of the organ system. The number of sensors used in the measurement will also be a function of specific application, and breast cancer screening may require the use of as few as twelve or as many as one hundred or more sensors 24 arranged on each breast 16 and 16a.
In breast cancer detection, one sensor array 22 is used to obtain 20 symptomatic breast dirr~l~nLials from a single breast~ but two sensor arrays 22 and 22a are arranged on both breasts to obtain between the breast dirrelel.lial measurements.
For breast cancer screening, the sensors 24 and a central sensor 26 of an electrical array 22 should be mounted in a manner which permits the 25 sensors to be accurately positioned against the curved surface of the breast 16 while still m~int~inin~ uniform spacing and the position of the sensors in a predetermined pattern. The sensor array 22 and the sensor arrays 22 W O 96/17547 PCT~US95115172 and 22a are used in conjunction with at least one reference sensor 30, and all of these sensors are of a type suitable for detecting ~e potential level of the electromagnetic field present in a living org~ni~m In Figure 3, an electrode for use as the sensors 24 and 30 is shown generally, and may include a layer of silver 32 having an electrical lead 34 secured in electrical contact thcr~wiLh. In contact with the silver layer is a layer of silver chloride 37, and extending in contact with the silver chloride layer is a layer of ion conductive electrolyte gel or cream material 38 which contacts the surface of a living or~ni~m This gel or cream, as will be subsequently indicated, is important in m~king a determin~tion of the type of measurement to be taken.
The device 10 is a multi-channel device having electrical electrode leads 34 and 34a extending separately from the electrodes 24 and 26 to one or more solid state multiplexors 35. These multiplexors can, for example, be multiplexors designated as Harris Semiconductor Model Hl-546-5. Each sensor array connected to the device 10, when in use, provides a plurality of outputs to a multiplexor connected to the array, and this multiplexor switches between the electrode leads 34 and 34a during a test period to connect the analog signals on each lead seqllenti~lly to a multiplexor output to create a time division multiplexed output. By providing one or more high speed solid state multiplexors for each array, it is possible to repeatedly sample biopotenti~l~ from a large number of sensors during a test period of minim~l duration. The multiplexors 35 control whether signals will be received from only one breast 16 by means of the sensor array 22 or whether signals will be received from both breasts 16 and 16a by means of the sensor arrays 22 and 22a.
W O96/17~47 PCTrUS95/15172 The outputs from the multiplexors 35 are provided to a low pass filter assembly 36 which operates to remove undesirable high frequency AC
components which appear on the slowly varying DC voltage signal outputs provided by each of the sensors as a result of the electromagnetic field 5 measurement. The low pass filter assembly 36 may constitute one or more multiple input low pass filters of known type which separately f1lter the signals on each of the input leads 34 and 34a when the array 22a is used, and then pass each of these filtered signals in a separate channel to a multiple input analog-to-digital converter 40. Obviously, the low pass filter 10 assembly 36 could constitute an individual low pass filter for each of the specific channels represented by the leads 34 which would provide a filtering action for only that channel, and then each filtered output signal would be connected to the input of the analog-to-digital converter 40. For multiple channels, it is possible that more than one multiple input analog-15 to-digital converter will be used as the converter 40.
The analog -to-digital converter 40 converts the analog signal in each input channel to a digital signal which is provided on a separate output channel to the multiple inputs of a central processing unit 42. The central processing unit includes RAM and ROM memories 46 and 4g. Digital 20 input data from the analog-to-digital converter 40 is stored in memory and is processed by the CPU in accordance with a stored program to perform the sensing and scanning methods of the present invention.
The measurement data processed by the CPU 42 contains indications of the presence or absence of a disease or other body condition. such as a 25 tumor, but those indications may not be readily discernable from a casual inspection of the data. Tn~te~cl, analysis of the data is required. and it is imperative that this analysis yield results that are consistent and reliable.
W O 96117547 PCTrUS95115172 The collected data tends to be obscured by noise due to factors such as the low DC potenti~l~ involved and the innate complexity of biological systems.
Biological systems are notorious for their complexity, and wide variances from the norm are not uncommon.
In order to accurately analyze the collected data in spite of the noise problem and the inherent nonlin~rity of biopotential data, the present invention involves the use of at least one neural network 44 to examine the data developed by the CPU 42 and identify patterns indicative of the presence or absence of a disease or other condition. Essentially, the neural network is a processing system wherein a !sim~ te~ set of connected proc~s.sin~ elements (neurons) react to a set of wei~hted input stimuli. The output from these neurons bears a nonlinear relationship to the input signal vector, with the nature of the relationship being determined by the strength of the connections. The connection strengths between the neurons must be set to values a~up-o~liate to the problem solution, and this is done in an indirect fashion by having the network "learn" to recognize inforrnation p~tte.rns. Once this is accomplished, the network can be called upon to identify a pattern from a distorted facsimile of the same pattern.
For use with the present invention, it is desirable to employ a neural network having a learning capability which is separate from the normal network function for data analysis. This is desirable because otherwise the network would continue to learn and evolve as it is used~ making validation of the results impossible. Known networks of this type are commercially available, such as a product identified as Neuro Shell II (Backpropogation) from the Ward Systems Group, Inc.
The network is first trained to recognize disease or other condition or injury p~t~erns using data resulting from known studies. Subsequently W O 96/17547 PCTrUS95115172 during use, similar data derived by the CPU as a result of procÁ~ing is then fed to the neural network 44, and the output of the neural network is directed to an indicator device 50 which may constitute a printer, a CRT
display device, or a combination of such indicators. The indicator device 5 50 may incorporate computer technology and graphics capable of im~ in~;, in at least two dimensions, the disease or injury condition indicated by the output from the neural network. Although the neural network 44 is shown as a separate block for illustrative purposes in Figure 1, it may in fact constitute a function performed by software for the central processing unit 10 42.
The operation of the discrimin~nt analysis device 10 will be clearly understood from a brief consideration of the method steps of the invention which the device is intended to perform. When the lesion 14 or other condition has not been identified and a screening operation is performed to 15 determine whether or not a lesion or other condition is present, a screening sensor array 22 is positioned in place on the site bein~ screened with the sensors 24, 26 positioned over various diverse areas of the site. If a breast 16 is screened, the sensor array will cover either the complete breast or a substantial area thereof. The reference sensor 30 is then brought into 20 contact with the skin of the subject 12 in spaced relationship to the sensor array 22, and this reference sensor might, for example, be brought into contact with a hand of the subject. Then, the electromagnetic field between the reference sensor and each of the sensors 24. 26 is measured, converted to a digital signal and stored for processing by the central processing unit 25 42. The central processing unit controls the multiplexors 35, and the program control for the central processing unit causes a plurality of measurements to be taken over a period of time. Usually, measurements W O 96117547 PCTrUS95115172 from individual sensors are taken seqllen~i~lly and repetitively, but the measurements on all channels may be taken simultaneously and repetitively for the pre~l~L~ i..ed measurement time period.
In prior art units, a plurality of measurements have been taken over S a period of time and often from a plurality of eleckodes, but then these plural measurements are merely averaged to provide a single average output indication. In accordance with the method of the present invention, the measurement indications on each individual channel are not averaged with those ~om other charmels, but are instead kept separate and averaged by 10 channel within the central processing unit 42 at the end of the measurement period. For the duration of a single pre~letermined measurement period, for example, from sixteen measurement channels, the cenkal processor will obtain sixteen average signals indicative of the average electromagnetic field for the period between the reference sensor 30 and each of the sensors in the sensor array 22 or the sensor arrays 22 and 22a. Of course, more reference sensors can be used, although only one reference sensor 30 has been shown for purposes of illustration.
Having once obtained an average signal level indication for each channel, the results of the measurements taken at multiple sites are analyzed 20 in terms of a mathematical analysis to determine the relationships between the average signal values obtained. It has been found that the result of such an analysis is that a subset of relationships can be obtained which are indicative of the presence of a disease, injury or other body condition while a different subset might be obtained which will be indicative of the absence 25 of these factors. Although either a discrimin~nt mathematical analysis procedure or decision making logic may be designed to separately obtain and analyze the relationship between the average potential values in accordance with this invention for screening or diagnostic purposes, the discrimin~nt mathematical analysis procedure to be hereinafter described in combination with data pattern recognition is a method which appears to be effective.
An important relationship to be obtained is often the m~xhllull-voltage dirrerellLial (MVD), which is defined as the minim~m average voltage value obtained during the measurement period subtracted from the m~xhllulll average voltage value obtained for the same period where two or more sensors are recording voltages from a test area. Thus, for each predetermined measurement period, the lowest average voltage level indi~a~ion ~btaine& on ~y ~e cha~els is subt.rac~d ~o~ th~h~g~e~t average voltage level indication obtained on any of the other channels to obtain an MVD voltage level. If this MVD voltage level is above a desired level >x, then a disease, injury or other condition, such as a malignancy, could be indicated. Similarly, if the average taken over the measurement period from one channel is an abnormally low value <y, the presence of this a~bnormally low individual sensor reading (IER) could be indicative of a disease, injury or other condition. These primar.v indicators may be further analyzed in accordance with a neural network control program to be subsequently described to reduce the number of false positive diagnosis, usually cases of non-malignant hyperplasia which may be falsely identified as cancer on the basis of high MVD or low IER re~tlings.
When the device 10 is used in accordance with the method of the present invention for a screening function where a specific lesion 14 has not yet been identified. using as an example breast screening where the patient is asymptomatic, an array 22 is employed which will give maximum coverage of the entire breast surface. Then the MVD level. and possibly W O 96/17547 PCTrUS95tl~17Z
an IER level is obtained in accordance with the method previously described as well as the average values for each channel. All of this data, namely, the average values, and the MVD level is input to the neural network 44 which has been trained to discern a pattern from the data which 5 is indicative of a disease condition. This same process can be performed with sensors of various types other than surface sensors, such as needle electrodes for invasive measurement and electrodes which measure resistance or impedance.
The general overall operation of the central processing unit 42 will 10 best be understood with refelellce to the flow diagram of Figure 4. The operation of the unit 10 is started by a suitable start switch as indicated at 52 to energize the central processing unit 42, and this triggers an initiate state 54. In the initiate state, the various components of the device 10 are automatically brou~ht to an op~ldLillg mode, with for example, the indicator 15 50 being activated while various control registers for the central processing unit are reset to a desired state. Subsequently, at 56, a predetermined multiple measurement period is initiated and the digital outputs which are either generated in the processing unit 42 or those from the analog-to-digital converter 40 are read. When an analog neural network 44 is employed, the 20 analog values will be read at 56. The central processing unit may be programmed to simultaneously read all channel outputs but these channel outputs are usually seqllenti~lly read.
Once the analog or digital signals from all channels are read, an average signal for each channel is obtained at 58 for the portion of the 25 measurement period which has expired. The average or norm~li7e-1 value for each channel is obtained by summing the values obtained for that channel during the measurement period and dividing the sum by the number W O 96/17547 PCTrUS95tl5172 of measurements taken. Then, at 60, the cenkal processin ~ unit ~letermines whether the measurement period has expired and the desired number of measurements have been taken, and if not, the collection of measurement samples or values continues.
Once the measurement period has expired, the microprocessor will have obtained a final average value for each channel derived from the measurements taken during the span of the measurement period. From these average values, the highest and lowest average values obtained during the measurement period are sampled at 62, and at 64, and the lowest 10 average channel value which was sampled at 62 is subtracted from the highest average channel value to obtain a maximum voltage differential value. Then both the channel average values from 62 as well as the maximum voltage differential value from 64 are directed as inputs to the neural network at 66 which has been kained to recognize disease patterns 15 from such data. Alternatively, only the maximum voltage differential values from 64 are directed to the neural network 66.
In the neural network at 66, if a disease or other condition pattern is identified from the input MVD signal values or the MVD signal values and the average signal values, then a probably disease indication, such as 20 cancer present, is provided at 68, but if a disease pattern is not recogni7e.1, then the lack of a probable disease condition is indicated at 70. Since neural networks generally provide a probability value~ the probability of the presence or absence of a disease condition is indicated at 68 and 70, and the device may be used to distinguish high risk patients from low risk patients.
25 After the indication of the probable presence or non-presence of a disease at 68 or 70, the routine is ended at 72.
The neural network 66 may be used to reco~nize other p~tterns derived from the DC biopotential signals provided by the device 10. For example, there is a phasicity pattern which occurs in the avera~ed electrical biopot~nti~ls over time which can be sensed, and variations in this phasicity 5 pattern may be recognized as indicative of particular disease, injury or other conditions. Another pattern can be recognized in the phasicity of the multiplicity of individual electropotential values which are obtained by the device 10 prior to averaging. During a test period, individual measurements in the hundreds will be taken from each sensor for averaging, and phasicity 10 changes in these individual values provide a complex pattern which can be analyzed by the neural network. Changes in his complex phasicity pattern could be identified as indicative of certain disease, injury or other conditions. It is contemplated that a combined analysis may be made using the phasicity pattern of the biopotential signals before averaging or 15 averaged signal values after averaging to obtain a combined analysis as an indicator of the presence or absence of a specific condition.
The apparatus for condition screening or sensing l O of Figure 1 ma~
be either a digital or analog unit. For an analog unit. the analog to digital converter 40 is elimin~te~l and the output of the low pass filter 36 is sent 20 directly to the CPU 42 where the analog signals are averaged and an analog MVD is developed. These analog average and MVD signals are then provided as inputs to an analog neural network 44 rather than a digital network.
The a~palaLIls for condition screening or sensing of the present 25 invention may be used in a method for monitoring the efficacy of therapy for some diseases or injury conditions. A problem which arises in cancer treatment, for example, is assessing the efficacy of the treatment~ whether W O96/17547 PCTrUS95/15172 it be by radiation or chemotherapy. However, the electrical biopotential dirr~le.~lials resulting from some cancers tend to change in response to chemotherapy. Thus once a cancer has been identified and an initial MVD
level for that particular cancer has been computed, subsequent MVD levels 5 are taken as treatment progresses and compared to previous MVD levels to rl~t~nnine whether or not a change is occurring. Depending upon the treatment in progress, a change, or in some instances a lack of change in the MVD level will tend to indicate that a therapy treatment is positive and is having some success.
Other conditions of the human body may also be effectively monitored using the method and ~ ~aldL~lS of the present invention. For example, in females during ovulation, tissue lu~tllle occurring incident to passage of the ova to the fallopian tubes results in significantly altered biopotential levels which can be sensed. Symptoms caused by ovulation can be confused with those associated with appendicitis, but the high MVD
levels resulting from ovulation can be used to differentiate between the two.
The efficacy of a variety of therapeutic treatments, such as post menopause hormone therapy and various immunal therapies may also be monitored using the method and apparatus of the present invention.
The use of the device and method of the present invention to provide an objective test for ovulation will be extremely useful in the prevention of an unwanted pregnancy as well as an aid in assisting women who are experiencing difficulties with infertility.
Another normal bodily function associated with DC biopotentials is labor or uterine contractions related to the birth of a child. The measurement of DC biopotentials in accordance with the present invention during the occurrence of these conditions can be used to determine the effectiveness of the contractions in leading to cervical dilation and the probability of a normal vaginal delivery. Also, DC electrical biopotential measurements can be useful in distinguishing normal from abno~nal delivery patte~ns as well as in the detection of fetal distress and - 5 abnormalities of labor such as sepa~ation of the placenta prematurely or infraction of the placenta.
The empioyment of a plurality of sensors 24, 26, 30 for the measurement of bioelectric phenomena means that a nearly limitless number of dirr~lelllials can be calculated. ~arly investigations of this phenomena focused on two general classes of differentials; Symptomatic (within) breast dirrel~llials and Between-Breast Dirr. ~ llials. The Symptomatic (within) breast dirr~,~lllial was obtained using one or more sensor arrays 22 on a single breast 16 which contains a suspicious lesion in the manner described, while the Between-Breast Dirr~l~lllials were obtained using sensor arrays 22 and 22a on the breasts 16 and 16a. The Between-Breast Dirrelential obtained was the difference between the symptomatic breast differential and the asymptomatic breast dirr~ llial, the asymptomatic breast being the breast 1 6a which does not include a suspicious lesion. Once the symptomatic and asymptomatic breast differential were obtained by the central processor unit 42 using the procedure of Figure 4 to obtain an MVD
for each breast, the processor then compares the asymptomatic and synlptolllatic dirre~llials to obtain the difference therebetween as a Between-Breast dirr~lelltial.
With the advent of more precise measurement sensors for DC
biopotenti~l~, additional, more precise differentials can be calculated to reveal more about a given disease state. An example of one such dirrelelllial developed in accordance with this invention is the Mirror-Image W O96/17547 PCT~US95/15172 differential, in which differences between the corresponding, mirror image sensors are calculated; for example between the sensor placed in the upper outer quadrants of each breast. These can then be averaged to produce a more precise indicator of between-breast asymmetry. This is opposed to the S antecedent Between Breast Dirrelelllials which calculated this asymmetry as the difference between the within breast dirrelc~llials of both breasts.
Mirror-Image differential measurement is a more sensitive indicator of between breast symmetry, as it provides differentials from individual areas of both breasts rather than an overall dirrelelltial for the complete 10 breasts. With a between breast dirrei~lltial, if there is a lesion on the asymptomatic breast which was unknown, the dirrel~;nlial value is significantly affected. With Mirror Image differential measurements, the dirrelelllials obtained from areas removed from the unknown lesion on the asymptomatic breast may not be significarltly affected by the lesion.
To obtain the Mirror-Image differential, multiple measurements are taken from corresponding sensors (i.e. 26b) in the sensor arrays 22 and 22a under the control of the central processor 4'~. These multiple measurement values are averaged for each sensor by the central processor, and then a difference value is obtained from the averages for the two sensors by the 20 central processor. This process is continued until differentials are obtainedfor all mirror image sensor pairs, and then these differentials are averaged by the central processor to obtain a final mirror image differential.
The mirror image dirr~r~l"ial or the between breast differential for each test period can be provided by the central processor as a processed 25 signal value to an a~ opliate neural network of the types to be described.
Another class of variables which may improve disease detection are those which compress the differential by m~king it conform to a known W O 96/17547 PCTrUS95115172 diskibution. This is useful especially when maximum dirr~enlials are calc~ te~l from many data points, which may contain an outlier. By imposing a known distribution on the set of data points, such as the Bienayme-Tschebycheff or Cramer distribution, the effect of statistical S outliers can be ~iimini~hed.
Still another approach to gain additional information from bioelectric measurements is to weight the dirrelclllials by the distance between the sensors. Other manipulations which may reduce noise in bioelectric dirr~ ials are norm~ ing procedures, by which the range of dirr~r~ntials 10 in the symptomatic breast 16 is evaluated and constrained in terms of the range of dirre.ellLials in the opposite breast 16a.
Clinical studies have indicated that the relative effectiveness of the electrolyte gel or cream 38 used as an electroconductive medium in skin sensor electrodes 24126/30, relates to the types of variables employed in 15 disease detection. For example, if the sensor type is kept constant, more effective discrimin~tion of disease state for gel-containing sensors is afforded by employing within breast dirrclclltials, such as the maximum difference between five sensors which are located in a quadrant of the breast with a suspicious lesion. On the other hand, sensors loaded with an 20 electroconductive cream rely more on differentials between the two breasts, such as the set of mirror-image dirr~lcllLials described above. For these types of sensors, dirr.lcl,lials calculated from sensors located at some distance away from each other tend to give better diagnostic information than dirr~,le.llials calculated from sets of sensors placed more closely 25 together, such as the within ~uadrant differential found effective for gel loaded sensors. The dirr~,elllials used for disease discrimin~tion should be tuned to the type of sensor employed.
W O96/17547 PCT~US95/15172 A number of different techniques may be employed in accordance with the present invention for deconvolving bioelectric measurements recorded from living or~ ni.~m.~. The techniques described result from the complexity of data generated by devices specifically designed to record 5 bioelectric information from a plurality of points on an org~ni.cm, and/or from a plurality of measurements made over time from at least one biosensor in contact with an or~ni~m, either internally or on the skin surface. The advantages of these techniques for the intended applications are that they do not assume the shape of the distribution of data, that is, 10 they are effective for both linear and nonlinear systems. Biological systems, including bioelectric fields, which are notorious for their nonline~rity and nonpredictiveness are best analyzed using the distribution-independent techniques to be hereinafter described. Often, the technique employed is dependent upon a number of variables, such as the type of 15 biosensor used and the type and volume of data to be analyzed.
As previously indicated, in artificial neural networks. data can be processed by several layers of interacting decision points or neurons. The network "learns" to recognize patterns from input data to produce a predictive output, such as benign vs. malignant breast disease. There are 20 several varieties of artificial neural networks which have been found to have specific utility for condition sensing using the apparatus 10 of Figure 1. These networks are operative with a variety of one or more types of measurements provided by the apparatus 10. The input to these networks can be maximum voltage differentials, channel averages, between breast 25 differentials, mirror image differentials, and in some instances, raw unaveraged biopotential measurements.
W O 96tl7547 PCT~US95/15172 Referring now to Figure 5, when the apparatus 10 of Figure 1 is used to obtain the somewhat limited data derived from a subject during a single test or group of tests taken during a single test period, the neural network 44 should consist of a probabilistic neural network 74.
S Probabilistic neural n~lwolhs produce probability values ranging from 0.00 to 1.00 as to whether a given disease state exists. The probabilistic neural networks learn ~uickly and do not require large amounts of data. They function well in situations where the output is bimodal. for example, cancer vs. benign disease states. In order to produce a predictive probabilistic 10 neural network, data is divided into three sets; the le~rning set, the test set, and the production set. Typically, 80 percent of the data is used in the learning set, 10 percent in the test set, and 10 percent in the production set.
The probabilistic neural network identifies bioelectric patterns associated with benign vs. malignant disease states using the learning set. The 15 predictiveness of the probabilistic neural network is monitored and altered periodically by comparison with the test set. and the final network is then checked for predictive accuracy against the production set. A key point for developing predictive probabilistic neural networks is the distribution of cases which make up the three data sets. For biopotenti~l~, the most 20 predictive probabilistic neural networks require that a relatively large proportion of data cases be reserved for the test and production sets (at least 20 percent for each set). This is most likely due to the fact that probabilistic neural networks are highly data driven (i.e., training set probabilistic neural networks don't necessarily transfer to new data) if given 25 the large number of variables (e.g., surface electrical potential differentials) generated by multi-sensor arrays.
W O96/17547 PCTrUS95/15172 As opposed to probabilistic neural networks in which the output typically is bimodal, general regression neural networks are specifically developed to handle continuous variable outputs. Because of this, they tend to require relatively large amounts of data. An application of a general 5 regression neural network would be in identifying a set of bioelectric variables which correlate with a continuous measure of disease state or cell proliferation such as Thymidine Labeling Index, a continuous variable which ranges from 0.0 to about 15ō Similarly, backpropogation neural networks require relatively large amounts of data to generate predictive 10 patterns. They differ from other types of neural networks because of the degree of complexity between the various levels of neurons and their interconnections. Backpropogation neural networks tend to generalize well to (predict for) new sets of data. For this reason, they can be effective for bioelectric information, especially if large amounts of data are available.
In instances where the apparatus 10 of Figure 1 is used to collect differentials from a multiplicity of tests taken over an extended period of time and this data is then stored and later reentered in the central processor unit 42, the neural network 44 will be either a general regression or a backpropogation neural network as indicated at 76 in Figure 6. An 20 advantage of these neural networks 76 is their ability to identify intermediate disease states on the basis of biologic electromagnetic fields.
As opposed to artificial intelligence paradigrns, such as classification and regression trees, the neural networks 76 can provide an output on a scale of continuous values. Since disease states also lack the ~bil~ly black and 25 white nature sometimes imposed by statistical reduction, these neural networks may be a better approximator of actual disease progression. For example. the progression of a tissue or an organ system from a normal state W O 96/17547 PC~rnUS95115172 to malignant state is not saltatory, rather the tissue goes from normal - through various pre-malignant stages until frank cancer is discovered. Thus there exist "shades of grey" in the progression of certain diseases. The neural networks 76, when coupled with fuzzy logic or other model-free methods of estimation or approximation, may provide a better means to discrimin~te the normal state from early disease, or early disease from later disease.
In fuzzy systems, rules in an algo,ilhlll are defined as any number of "patches" which cover events in a nonlinear system. In terms of disease states, the events could be defined as various stages along the collthluulll from normal to malignant. The "patches" would be the bioelectric measurements which define the set of disease states in the continuum. In fuzzy systems, all of the input rules are activated simultaneously, with different weights, to define a disease state. This can result in improved clinical utility in the following way. Various cut-off values in the neural net result can be stored in the central processor unit 42 and used to make determinations regarding patient management. For example. cutoffs could be established at .25, .50, and .75. Patients with less than .25 could be relatively assured they are free of serious disease, patients between .25 and .50 could be monitored by follow-up, patients between .50 and .75 could be monitored using additional tests while patients higher than .75 could be advised to proceed directly to biopsy. In this way, cost effective triaging - of patients could occur.
Commercial versions of the probabilistic, general regression and backpropogation neural networks which can be used with this invention are available from Ward Systems Group Inc. of Frederick, Maryland~ and are designated as Neuro Shell 2, Release 2ō
W O 96/17547 PCTrUS95/15172 All of the pattern recognition strategies discussed above are approximations of real world phenomena. As such, each has its own set of advantages and disadvantages as estim~tors of actual biologic functions and disease states. Given this, the optimal approach to disease detection by S pattern recognition of bioelectric fields might lie in combining various pattern recognition strategies in novel ways. For example, artificial neural networks are reputed to be analogs of brain function. However, in recognizing p~ rn.~, the brain probably uses a combination of digital and analog strategies. By combining the outputs of several artificial neural 10 networks in a classification tree or entropy pattern, it is possible to more closely match patterns with disease states.
With reference to Figure 7, a plurality of separate neural networks 78, 80 and 82 are connected to separately process the differentials from the central processor unit 42. Depending on the volume of data input to the 15 central processor, these plural neural networks may be the networks 74 or 76. The outputs of these neural networks are then input to a processor 84 which may actually be a separate processor or a portion of the central processor unit 42. The processor 84 may include an entropy minim~x program which operates on the basis of minim~l information entropy.
20 Commercial versions which apply the entropy minim~x theory to applied pattern recognition are available from Entropy Ltd. of Lincoln, Mass. under the designation entropy minim~x. The entropy minim~x form of pattern recognition is similar mathematically to thermodynamic entropy~ hence its name. The aim of the entropy minim~x program is to identify sets of 25 variables or events which predict a state on the basis of minim~l information entropy. Entropy minim~x recognition can be understood as a process by which multivariate data is partitioned into n sets~ each of which coll~sponds to a disease state. The members of the set are all possible unique vectors (combinations) of attribute values (variables) which may be created from the atkibute list. The partitioned sets are created by finding the subset of data having the least entropy (or most predictive S pattern). Having found that subset, it removes those cases from further consideration, and moves on to the next subset with the least amount of entropy. These successive identifications of sets proceed until all the cases are accounted for. This approach lies somewhere in between classification tree analysis and artificial neural networks in that although the sets of data 10 are discrete (as in a decision tree) the weights placed on the variables are done so simultaneously (as in artificial neural networks).
Another way in which the various pattern recognition strategies could be combined would be to have the processor 84 weight the values from the three neural networks 78, 80 and 82 to obtain a final disease classification.
15 For example, the value used for a final disease classification might be the value closest to that provided by a majority of the neural networks or.
alternatively, might be an average value derived from the plural neural net~,vork outputs.
In taking DC biopotential measurements, the interface between the 20 measurement and reference electrodes and the skin of a subject is a source of high impedance and electronic noise which can cause high and unpredictable variations in the DC signal values measured. Therefore, it is important to minimi7e the noise and reduce the impedance resulting from this interface. It has been found that the presence of a relatively thick 25 cornified epithelial layer of skin under an electrode contributes to high noise and impedance values at the interface, and care must be taken to remove this layer of skin or to position the electrode in another area. For W O96tl7547 P~ '3~tl5172 example, a reference electrode or electrodes should be positioned on low impedance areas of the body such as the low-suprasternal notch or the skin over the exphoid process or the subxyphoid area.
Selective positioning ofthe measurement electrodes is more difficult, 5 as for effective screening, it is often necessary to space these electrodes subst~nti~lly equal distances apart. Consequently, removal of some of the cornified epithelid layer must be accomplished in areas where electrodes are to be positioned. To accomplish this with electrocardiograph electrodes, a small patch of fine sandpaper has been developed which sticks to the finger 10 of a technician and which can m~nll~lly be manipulated to remove a skin layer. This m~nll~l technique is not acceptable for many DC biopotential measurement processes, as for example in breast screening for cancer, where the potential for injury to sensitive skin is high.
With reference to Figure 8, a simple, throw away~ skin removal 15 device 80 is illustrated which operates effectively to remove dead surface skin without having the ability to apply a pressure sufficient to cause injury to other skin layers. This skin removal unit includes a head section 82 and a body section 84 of wood, plastic or similar substantially rigid material with the head section being joined to the body section by a neck section 86.
20 A very fine abrasive layer 88 is applied to one surface of the head section and is adapted to be rubbed against the skin of a subject by m~nu~l reciprocatory movement of the body section when it is held in the hand of a technician. The neck section is designed to break or collapse if pressure on the abrasive layer and head section exceeds a pressure which is safe for 25 the skin surface being abraded. This can be accomplished by reducing the cross sectional area of the neck section relative to that of the head and body sections or by forming the neck section of a flexible or spring material WO 96/17547 PCr/US95/15172 which flexes u~ dly when too much pressure is applied to the head section.
Alternatively, skin layer removal can be mechanically accomplished with the adjustable skin removal unit 90 of Figure 9. This unit includes a S housing 92 to rotatably mount a spring biased assembly 94 that includes a shaft 96 having an upper end 98 which is received in the open end of a chamber 100 formed in a second shaft 102. An abrasive disc support 104 is formed at the lower end of the shaft 96 to receive a disposable disc of abrasive material 106. This disc is removably secured to the disc support by any suitable means, such as a pressure sensitive adhesive which permits the disc to be removed and discarded after use. A pin 108 projecting from the shaft 96 projects into and engages the lower end of a slot 110 formed in the shaft 102. This positions the shaft 96 relative to the housing 92 so that the abrasive disc 106 extends beneath a lower support surface 112 of the housing.
The shaft 96is permi1te~1 to move longit--~lin~lly upward relative to the shaft 102 for a limited instance determined by the slot 110? and when the pin 108 reaches the uppermost limit of the slot the abrasive disc 106 will have been moved inwardly of the housing 92 above the lower support surface 112 thereof. A spring 114 mounted within the chamber 100 engages the upper end 98 of the shaft 96 and biases the shaft so as to tend to m~int~in the pin 108 in contact with the lower ~xLIe~ y of the slot 110.
The bias of this spring will determine the pressure which the abrasive disc 106 will apply to the skin of a patient, and when a preset pressure is exceeded, the shaft 96 will move upwardly against the bias of the spring 114 to reduce the pressure of the disc against the skin.
W O96/17547 PCTrUS95115172 The bias of the spring 114 may be varied by rotating a threaded shaft 116 which extends through the upper end of the shaft 102 into contact with the upper end of the spring 114. The rotation of the shaft 102 is biased by a spring 118 having one end 120 connected to the housing 92 and a second end 122 connected to the shaft 102. The shaft 102 may be rotated against the bias of the spring 118 by a handle 124 positioned externally of the housing 92 and connected to the uppermost end of the shaft 102. When the shaft 102 is rotated to increase the bias of the spring 118, a detent 126 in the handle 124 engages the end of a pivoted, spring biased trigger 128.
When the kigger 128 is pivoted out of the detent 124, the spring 118 rapidly rotates the shaft 102 which in turn rotates the shaft 96 through engagement with the pin 108. This caused the disc 106 to rotate against the skin of a subject.
The housing 92 may also contain an impedance measuring unit which measures the electrode interface impedance in the area where a portion of the skin layer has been removed by the disc 106 prior to the placement of an electrode 24. 26 or 30. This impedance measuring unit includes a DC power supply 130 connected to an electrode assembly 132 ,and an impedance measuring circuit 134. When the impedance at the interface between the electrode assembly 132 and the skin of a subject is within a low range sufficient for good operation of the electrodes in the electrode assembly 22,m the impedance measuring circuit will permit current to flow to light an indicator light 136 which may be formed by an LED.
In accordance with the present invention, the mechanical skin preparation accomplished with devices such as those shown in Figures 8 and 9 prior to the application of electrodes may be replaced by chemical skin ~i~dlion. Keratolytic agents such as salicylic acid, glycolytic acid and acetic acid cause a swelling and disruption of the cornified epithelium, which results in a reduction of the impedance in a treated area at the electrode-skin interface. For exarnple, tests have shown that the direct 5 application of a 2% solution of salicylic acid in an over the counter Salicyclic Acid Acne Preparation on Double Textured Pads distributed by Proctor & Gamble under the trademark Clearasil(~) to the female breast causes about a two-fold reduction in skin impedance. To automatically achieve an impedance reduction at the skin-electrode interface, a keratolytic 10 agent, such as salicylic acid, is added to the gel 38 in the electrode of Figure 3, which forrns the electrodes 24, 26 and 30 of Figure 1. It has been found that a keratolytic agent, such as salicylic acid, may be dissolved in the gel without adversely affecting the electrolytic action of the gel.
When the electrode is mounted on a subject with the gel in contact with the 15 skin, the keratolytic agent in the gel causes a disruption of the cornified epithelium with a resultant reduction of impedance at the electrode-skin interface. The keratolytic agent should be within the range of 1-10% of the gel-agent mixture.