• Home   /  
  • Archive by category "1"

Introduction Brain Fingerprinting Essay

Brain Fingerprinting, invented by Dr. Farewell,  is a technique that gives us a simple yes/no answer to the question: “Is a particular information stored in a person’s brain?” The question may sound simple, but the answer to the question can help differentiate between the guilty and the innocent. This is the major application of brain fingerprinting. It can be used to detect lies and find the guilty.

In India, the rice-chew-and-spit method was used in the early times to detect lies. The suspects were given a grain of rice each, and were told to chew and spit it out. The innocent could easily chew and spit the rice, but the guilty’s mouth would become dry, and so the guilty could not spit the rice out. This was used as a lie detection test in those times.

Brain fingerprinting can be viewed as a modern, and more accurate, lie detection test. To understand how it works, we must understand the basic difference between the guilty and the innocent. The guilty, since he has committed the crime, has the details of the crime in his memory. However, the innocent does not have the details in his memory. Brain fingerprinting is a test of whether the details are present in the memory or not. If they are, then the person is guilty, otherwise innocent.

Now, how does brain fingerprinting check for these details in a person’s memory? Brain fingerprinting does this by showing certain images of objects (possible murder weapons, etc) and other stimulus. It then records the response of the person’s brain on seeing these objects. EEG sensors are used to record these responses. Further, a specific waveform called the P300 gets activated (as measured in the response) only if the person is guilty; otherwise it does not. P300 is the specific response of a brain that recognizes the object/image shown, and hence, has the object/image in his memory. This helps identify the guilty. (Read more about P300 here)

However, the human brain is capable of generating as well as modifying thoughts. No matter what the truth is, the brain has the power to choose to believe otherwise. The guilty may modify his thoughts even on seeing the murder weapon (for example), and choose not to respond or think of the murder incident. The solution to this problem is to record the response of the person to the stimulus within fractions of  a second. The response is recorded before the person can modify it, or even be aware of his own thoughts.

To summarize/review, let’s hear what the founder of brain fingerprinting has to say about this amazing technique:

Abstract

Brain fingerprinting detects concealed information stored in the brain by measuring brainwave responses. We compared P300 and P300-MERMER event-related brain potentials for error rate/accuracy and statistical confidence in four field/real-life studies. 76 tests detected presence or absence of information regarding (1) real-life events including felony crimes; (2) real crimes with substantial consequences (either a judicial outcome, i.e., evidence admitted in court, or a $100,000 reward for beating the test); (3) knowledge unique to FBI agents; and (4) knowledge unique to explosives (EOD/IED) experts. With both P300 and P300-MERMER, error rate was 0 %: determinations were 100 % accurate, no false negatives or false positives; also no indeterminates. Countermeasures had no effect. Median statistical confidence for determinations was 99.9 % with P300-MERMER and 99.6 % with P300. Brain fingerprinting methods and scientific standards for laboratory and field applications are discussed. Major differences in methods that produce different results are identified. Markedly different methods in other studies have produced over 10 times higher error rates and markedly lower statistical confidences than those of these, our previous studies, and independent replications. Data support the hypothesis that accuracy, reliability, and validity depend on following the brain fingerprinting scientific standards outlined herein.

Keywords: Brain fingerprinting, P300-MERMER, P300, Event-related potential, Detection of concealed information, MERMER

Introduction and background

This paper reports four field/real-life studies using event-related potentials in the detection of concealed information. A primary purpose of these studies was to test the effectiveness of certain very specific methods for using brain responses in the detection of concealed information. We tested these specific methods in two types of field/real-life tests in detecting concealed information obtained in the course of real-life events. Two studies were specific issue tests. They used event-related brain potentials to detect concealed information regarding specific incidents in the lives of subjects, including major crimes with life-changing judicial outcomes. Two studies were specific screening or focused screening tests.1 They used event-related potentials to detect knowledge related to a particular kind of training or expertise, specifically knowledge characteristic of FBI agents and knowledge characteristic of explosives experts or bomb makers.

Another major purpose of the research reported herein is to identify the scientific principles and specific methods required to obtain valid and reliable results, an extremely low error rate, and high statistical confidence for all determinations made. We also have sought to identify the specific principles and methods required to obtain resistance to countermeasures and to minimize indeterminate outcomes while maintaining an extremely low error rate.

Due to overriding security concerns, it has not previously been possible to publish details of some of our research at the CIA, the FBI, the U.S. Navy, and elsewhere. These concerns have now been resolved, and this research can now be published. This is the fourth in a series of six recent peer-reviewed articles to be published, beginning with Farwell (2011a, 2012, 2013). We have also published two new patents (Farwell 2007, 2010) and other papers (e.g., Farwell 2011b). We hope through these publications to address some of the major issues that have arisen in the field since our original publications (Farwell and Donchin 1991; Farwell and Smith 2001), and, most importantly, to provide extensive data regarding the methods that are sufficient and necessary to produce error rates and statistical confidences excellent enough for practical application in the field.

In the original papers on the specific set of scientific methods that has come to be called “brain fingerprinting,” Farwell and Donchin (1991)2 and Farwell and Smith (2001) reported a 0 % error rate and high statistical confidence for all determinations. To make our statistical statements more conservative, and for the purpose of meaningful mathematical comparisons, in our discussions herein, we will generally exaggerate the error rate slightly and use “less than 1 %” to characterize the error rate in studies where in fact a 0 % error rate was obtained. In addition to 0 % error rate, Farwell and Smith also reported 0 % indeterminates. Replications of these methods in other laboratories, e.g., Allen and Iacono (1997), have achieved similar results (see also Iacono 2007, 2008; Iacono and Lykken 1997; Iacono and Patrick 2006; Neshige et al. 1991). These are the methods that were ruled admissible in court in the Harrington case (Erickson 2007; Farwell and Makeig 2005; Harrington v. State 2001; Roberts 2007; Farwell 2013). For a comprehensive tutorial review of all related publications in English to date, see Farwell (2012).

In the criminal justice system and in national security, error rate is of paramount importance. In the United States, error rate is specified as one of the fundamental tests of the prevailing Daubert standard for admissibility in court (Farwell and Makeig 2005; Roberts 2007). Accuracy is 100 % minus the error rate. For a discussion of these terms and their legal and scientific implications, see Farwell (2012).

Some subsequent studies have been based on fundamentally different scientific principles, have applied fundamentally different methods, and consequently have produced substantially different results. They have reported error rates approximately 10–50 times higher than those of the original brain fingerprinting studies, as well as susceptibility to countermeasures. For a comprehensive review, see Farwell (2012). For example, Rosenfeld et al. (2004) reported an overall 35 % error rate in detecting information-present subjects (65 % accuracy) without countermeasures and 67 % error rate (33 % accuracy) with countermeasures for their various techniques and conditions. Error rates of their various methods in detecting information-present subjects ranged from 8 to 46 % without countermeasures, and from 61 to 82 % with countermeasures. In a series of studies, the “complex trial protocol” (Rosenfeld et al. 2008) has produced an overall error rate of 15 % without countermeasures and 29 % with countermeasures. For a review, see Farwell (2012). (See also Farwell 2011a, b).

Farwell (2012) described in detail the brain fingerprinting methods that have produced less than 1 % error rate and high statistical confidence, and discussed the reasons that various alternative methods that have produced error rates an order of magnitude or more higher and statistical confidences substantially lower. That paper also discussed the specific methodological shortcomings that led to susceptibility to countermeasures in various alternative techniques that did not follow the methods applied in our original research and the present research.

In our view, in order to be viable for field use or any other application with major consequences, a technique must produce an overall error rate of less than 1 % in all studies and field applications, an error rate of less than 5 % in every individual study, and a record of consistently high statistical confidences for both information-present and information-absent determinations—averaging at least 90 % for information-present determinations and 90 % in the opposite direction for information-absent determinations, and preferably averaging over 95 % in the correct direction for all determinations of both types. To make a decision in a specific field case with judicial or other life-changing consequences, in our view the statistical confidence for the determination should be at least 95 %, whether it is information present or information absent. In our actual field applications, every individual determination to date has been with at least 99 % statistical confidence with the P300-MERMER.

The present four studies were designed to test the hypothesis that following the standard scientific procedures specified in Farwell and Donchin (1991), Farwell and Smith (2001), and Farwell (1992, 1994, 1995a, b, 2007, 2010, 2012) is sufficient to produce valid and reliable results, consistently less than 1 % error rates, and extremely high statistical confidence for each determination. We tested this hypothesis in demanding field conditions involving real-life crimes and life-changing consequences of the outcome of the tests, including judicial consequences such as the death penalty or life in prison, and in other cases a $100,000 reward for beating the test. In one study we also tested the hypothesis that these specific brain fingerprinting methods are unaffected by the countermeasures that have proven effective against alternative techniques (e.g., Rosenfeld et al. 2004, 2008; Mertens and Allen 2008).

In these studies we compared the performance of two data analysis methods involving different but overlapping time epochs and the corresponding event-related brain responses. One included only the P300, which has been known for over half a century as consisting of a positive peak. The other included both the P300 and a later negative peak (late negative potential or LNP) that follows the P300 in the data collected in our laboratory and in other laboratories applying the same paradigm. The P300 and the LNP together we refer to as a P300-MERMER (memory and encoding related multifaceted electroencephalographic response). The characteristics of this response are described in Farwell (1994, 1995b, 2007, 2010), in Farwell and Smith (2001), and in more detail in Farwell (2012).

Four brain fingerprinting field/real-life studies

The present report comprises four brain fingerprinting field/real-life studies. In all four studies we used brain fingerprinting to detect information obtained in the course of real-life events. Studies 1 and 2 used specific issue brain fingerprinting tests to detect specific issue information regarding real-life events, including capital crimes. Studies 3 and 4 used specific screening brain fingerprinting tests to detect real-life specific group knowledge of FBI agents and experts in bomb making, i.e., explosive ordnance disposal (EOD) and improvised explosive devices (IEDs).

In Study 1, the “CIA Real Life Study,” brain fingerprinting was used to detect concealed information regarding real-life events, including a number of felony crimes. There were, however, no significant consequences of the outcome of the tests, and consequently no substantial motivations for subjects.

In Study 2, the “Real Crime Real Consequences $100,000 Reward Study,” brain fingerprinting was used to detect information regarding real crimes. In some cases, the subjects were highly motivated because they were facing either the death penalty or life imprisonment, and the brain fingerprinting test detected presence or absence of information regarding the crime in question. In cases where there was less inherent motivation resulting from a potential judicial outcome, subjects were offered a $100,000 reward for beating the test. Except in cases where life and freedom were at stake, subjects were taught countermeasures that have previously proved effective against other, fundamentally different, non-brain fingerprinting techniques (Mertens and Allen 2008; Rosenfeld et al. 2004) but not against brain fingerprinting (see Farwell 2011a, b, 2012).

In Study 3, the “FBI Agent Study,” brain fingerprinting was used to detect information that is known to FBI agents but not to the general public, such as FBI techniques, procedures, acronyms, information learned in FBI training, etc.

In Study 4, the “Bomb Maker Study,” brain fingerprinting was used to detect information that is known to explosive ordnance disposal (EOD) and improvised explosive device (IED) experts but not to the general public.

In a fifth study, we set out to apply an alternative, non-brain fingerprinting “complex trial protocol” in detecting real-life information with some of the same subjects as Study 2. We discontinued this study for scientific and ethical reasons, as explained in the Discussion section. (For details, see Farwell 2012.)

P300 and P300-MERMER

The original brain fingerprinting research (Farwell 1992; Farwell and Donchin 1986, 1991) used the P300 component of the event-related brain potential (ERP). The P300 is a positive voltage potential maximal at the midline parietal scalp (Pz in the International 10–20 System) that peaks at 300 or more milliseconds from the onset of the eliciting event (Donchin et al. 1986; Farwell and Donchin 1988a; Miller et al. 1987; Sutton et al. 1965). Farwell and colleagues (Farwell 1994, 1995b, 2012; Farwell and Smith 2001) have shown that in the brain fingerprinting paradigm this positive peak is followed by a late negative peak (the Late Negative Potential or LNP). The two together have been termed P300-MERMER (memory and encoding related multifaceted electroencephalographic response). Both the P300 and the P300-MERMER undoubtedly have other features beyond the simple time-domain pattern that becomes apparent through the usual ERP signal-averaging procedures (Farwell 1994; Farwell and Smith 2001; see also Rapp et al., 1993). The positive–negative-peaked pattern in the time domain (or negative–positive–negative pattern if the N2 preceding the P300 is included), however, is sufficient to define the response, and is all that is necessary to detect it. This time-domain analysis is all that is used in our data analysis in the present paper. Data analysis in the present paper compares the results obtained by including only the P300 in the analysis with the results obtained by including the full P300-MERMER in the analysis. In both cases, only the usual time-domain characteristics of the signals that are used in conventional ERP analysis are considered in the computations. The only difference is the length of the epoch analyzed.

When we first discovered the P300-MERMER, specifically the late negative peak (LNP) that follows the positive P300 peak in the full P300-MERMER, our initial hypothesis was that the LNP of the MERMER was an artifact, possibly caused by the effect of the analog filters used in data collection on the return of the P300 to baseline. We soon discovered that the artifact hypothesis is not supported by the data. Experimentation (including recording without analog filters), scalp distribution (the relative amplitude at different scalp sites), and morphology (the latency and shape of the waveforms) has proven that the LNP of the P300-MERMER is not an artifact of the signal-detection or noise-reduction procedures or equipment, such as digital and analog filters or of the return of the P300 to baseline (Farwell 1994, 1995b, 2012).

To definitively test the hypothesis that the LNP of the P300-MERMER was a filter-generated artifact, we recorded without analog filters. We found that recording without analog filters did not diminish the amplitude of the LNP or change its latency, thus disproving the filter-generated artifact hypothesis.

Moreover, the data recorded with filters are also incompatible with the hypothesis that the LNP of the P300-MERMER is an artifact of any kind. The recording equipment is identical for all scalp sites and all subjects. If the LNP were an artifact of the equipment, the identical equipment would produce the same effects in different instances. The characteristics of the LNP would simply be a function of the P300. On the contrary, we found that the relative latency, amplitude, and morphology of the P300 and the LNP are very different for different subjects and for different scalp sites in the same subject. In different subjects, we found that virtually identical P300s were followed by LNPs that differed in latency by hundreds of milliseconds and differed in amplitude by a factor of two or more. In a number of cases the LNP was substantially smaller than the P300 at one channel (usually Pz) and substantially larger than the P300 at another channel (usually Fz) for the same subject in the same data set. These data are incompatible with the hypothesis that the LNP of the P300-MERMER is simply an artifact generated by some combination of the P300, the return to baseline after the P300, and the recording equipment and filters.

The positive P300 peak (or two peaks—P3a and P3b) is preceded by a negative peak, the N200, and followed by another negative peak, the LNP, producing a tri-phasic shape for the P300-MERMER. We first observed this triphasic negative–positive–negative pattern at the scalp in the early 1990s (Farwell 1994, 1995b, 2012; Farwell and Smith 2001). The same negative–positive–negative pattern has been observed in intracranial recordings in various structures (Halgren et al. 1998), including the inferior parietal lobe/supramarginal gyrus, superior temporal sulcus (Halgren et al. 1995), the amygdala and hippocampus (Halgren et al. 1986; Stapleton and Halgren 1987), dorsolateral and orbital frontal cortices, and the anterior cingulate (Baudena et al. 1995).

In short, the P300-MERMER is not simply the P300 followed by an artifact. It is a response produced in the brain that includes a negative peak following the positive P300 peak. By now virtually all of the researchers involved in detection of concealed information with brainwaves include in their computational algorithms both the positive P300 and the late negative potential (LNP) that constitutes the other major facet of the P300-MERMER (for a review, see Farwell 2012). Differences in nomenclature still exist, however. Some use the term the “amplitude of the P300” to refer to what we call the amplitude of the P300-MERMER, that is, the sum of the amplitudes of the P300 and the LNP. Computationally this is the voltage difference between the most positive point in the P300 time range and the most negative point in the LNP time range. In any case, we use all points in the entire waveform in our computations, not just the peaks, so the question of nomenclature is moot. Our data analysis matches patterns, and it does not matter what the responses are called.

The term “Brain Fingerprinting”

The term “brain fingerprinting” is based on the defining feature of matching something on the person of the suspect with something from the crime scene. Fingerprinting matches prints at the crime scene with prints on the fingers of the suspect. DNA “fingerprinting” matches biological samples from the crime scene with biological samples from the suspect. “Brain fingerprinting” matches information stored in the brain of the suspect with information from the crime scene. We use the term “brain fingerprinting” to refer to any methods, beginning with the original Farwell and Donchin (1991), Farwell (1992), and Farwell and Smith (2001) studies, that meet or exceed all or almost all the brain fingerprinting scientific standards specified below. Brain fingerprinting studies analyze the data based on either the P300-MERMER, or the P300 alone, or both (as in the present study).

Methods

Standard methods for all four studies

Previous publications (Farwell 1992, 1994, 1995a, b, 2007, 2010, 2012; Farwell and Donchin 1991; Farwell and Smith 2001) specify the standard methods we have applied in all our brain fingerprinting studies. For a detailed account of the scientific principles and specific methods, see Farwell (2012). These methods are briefly summarized below.

Three types of stimuli consisting of words or phrases are presented on a computer screen. (Pictures and auditory stimuli may also be used, but were not in these studies.) Probe stimuli contain specific information relevant to the investigated situation. The test is designed to detect the subject’s knowledge or lack of knowledge of the probes as relevant in the context of the crime or other investigated situation. (We shall generally refer to the investigated situation as a “crime,” although of course other, non-criminal situations can be investigated, as in Studies 3 and 4.) Probes have the following defining characteristics.

  1. Probes contain features of the crime or investigated situation that in the judgment of the criminal investigator the perpetrators would have experienced in committing the crime, or the subjects would have learned in the course of gaining the specific knowledge, training, or expertise investigated;

  2. Probes contain information that the subject has no way of knowing if he did not participate in the crime or other situation of interest; and

  3. Probes contain information that the subject claims not to know or to recognize as significant for any reason.

In order to test whether or not the subject recognizes the probes as significant in the context of the investigated situation, we present two additional types of stimuli. Target stimuli elicit a response that provides a standard for the subject’s brain response to known information relevant to the investigated situation. Irrelevants elicit a response that provides a standard for the subject’s response to irrelevant, unknown information.

Target stimuli present situation-relevant information that is known to be known to the subject. This information may have been revealed to the subject through news reports, interrogation, etc. In any case, the targets are disclosed to the subject before the test. The subject instructions also convey the significance of each target in the context of the investigated situation.

Irrelevant stimuli contain plausible, but incorrect, information about the crime. For a subject lacking the relevant knowledge contained in the probes, the irrelevants and probes are equally plausible as crime-relevant details. For each probe (and each target) several irrelevants are structured that contain similar but incorrect information. For example, if a probe is the murder weapon, a pistol, corresponding irrelevants could be a rifle, a shotgun, and a knife. The subject is informed of the significance of the probes in the context of the investigated situation (e.g., “the murder weapon”), but is not informed which is the correct, crime-relevant probe and which are the corresponding irrelevants.

The previous research that has produced a 0 % error rate and extremely high statistical confidence for each determination was conducted according to the following scientific standards for brain fingerprinting tests. The present studies also met these standards.

Scientific standards for brain fingerprinting tests

The following procedures comprise the scientific standards for brain fingerprinting tests (Farwell 1992, 1994, 1995a, b, 2007, 2010, 2012; Farwell and Donchin 1991; Farwell and Smith 2001; Harrington v. State 2001).

  1. Use equipment and methods for stimulus presentation, data acquisition, and data recording that are within the standards for the field of cognitive psychophysiology and event-related brain potential research. These standards are well documented elsewhere. For example, the standard procedures Farwell introduced as evidence in the Harrington case were accepted by the court, the scientific journals, and the other expert witnesses in the case. Use a recording epoch long enough to include the full P300-MERMER. For pictorial stimuli or realistic word stimuli, use at least a 1,800-ms recording epoch. (Shorter epochs may be appropriate for very simple stimuli.)

  2. Use correct electrode placement. The P300 and P300-MERMER are universally known to be maximal at the midline parietal scalp site, Pz in the standard International 10–20 System.

  3. Apply brain fingerprinting tests only when there is sufficient information that is known only to the perpetrator and investigators. Use a minimum of six probes and six targets.

  4. Use stimuli that isolate the critical variable: the subject’s knowledge or lack of knowledge of the probe stimuli as significant in the context of the investigated situation. Obtain the relevant knowledge from the criminal investigator (or for laboratory studies from the knowledge-imparting procedure such as a mock crime and/or subject training session). Divide the relevant knowledge into probe stimuli and target stimuli. Probe stimuli constitute information that has not been revealed to the subject. Target stimuli contain information that has been revealed to the subject after the crime or investigated situation.

  5. If initially there are fewer targets than probes, create more targets. Ideally, this is done by seeking additional known information from the criminal investigators. Note that targets may contain information that has been publicly disclosed. Alternatively, some potential probe stimuli can be used as targets by disclosing to the subject the specific items and their significance in the context of the investigated situation.

  6. For each probe and each target, fabricate several stimuli of the same type that are unrelated to the investigated situation. These become the irrelevant stimuli. Use stimuli that isolate the critical variable. For irrelevant stimuli, select items that would be equally plausible for a non-knowledgeable subject. The stimulus ratio is approximately one-sixth probes, one-sixth targets, and two-thirds irrelevants.

  7. Ascertain that the probes contain information that the subject has no known way of knowing, other than participation in the investigated situation. This information is provided by the criminal investigator for field studies, and results from proper information control in laboratory studies.

  8. Make certain that the subject understands the significance of the probes, and ascertain that the probes constitute only information that the subject denies knowing, as follows. Describe the significance of each probe to the subject. Show him the probe and the corresponding irrelevants, without revealing which is the probe. Ask the subject if he knows (for any non-crime-related reason) which stimulus in each group is crime-relevant. Describe the significance of the probes and targets that will appear in each test block immediately before the block.

  9. If a subject has knowledge of any probes for a reason unrelated to the investigated situation, eliminate these from the stimulus set. This provides the subject with an opportunity to disclose any knowledge of the probes that he may have for any innocent reason previously unknown to the scientist. This will prevent any non-incriminating knowledge from being included in the test.

  10. Ascertain that the subject knows the targets and their significance in the context of the investigated situation. Show him a list of the targets. Describe the significance of each target to the subject.

  11. Require an overt behavioral task that requires the subject to recognize and process every stimulus, specifically including the probe stimuli, and to prove behaviorally that he has done so on every trial. Detect the resulting brain responses. Do not depend on detecting brain responses to assigned tasks that the subject can covertly avoid doing while performing the necessary overt responses.

  12. Instruct the subjects to press one button in response to targets, and another button in response to all other stimuli. Do not instruct the subjects to “lie” or “tell the truth” in response to stimuli. Do not assign different behavioral responses or mental tasks for probe and irrelevant stimuli.

  13. In order to obtain statistically robust results for each individual case, present a sufficient number of trials of each type to obtain adequate signal-to-noise enhancement through signal averaging. Use robust signal-processing and noise-reduction techniques, including appropriate digital filters and artifact-detection algorithms. The number of trials required will vary depending on the complexity of the stimuli, and is generally more for a field case. In their seminal study, Farwell and Donchin (1991) used 144 probe trials. In the Harrington field case, Farwell used 288 probe trials (Harrington v. State 2001). In any case, use at least 100 probe trials and an equal number of targets. Present three to six unique probes in each block.

  14. Use appropriate mathematical and statistical procedures to analyze the data. Do not classify the responses according to subjective judgments. Use statistical procedures properly and reasonably. At a minimum, do not determine subjects to be in a category where the statistics applied show that the determination is more likely than not to be incorrect.

  15. Use a mathematical classification algorithm, such as bootstrapping on correlations, that isolates the critical variable by classifying the responses to the probe stimuli as being either more similar to the target responses or to the irrelevant responses. In a forensic setting, conduct two analyses: one using only the P300 (to be more certain of meeting the standard of general acceptance in the scientific community), and one using the P300-MERMER (to provide the current state of the art).

  16. Use a mathematical data-analysis algorithm that takes into account the variability across single trials, such as bootstrapping.

  17. Set a specific, reasonable statistical criterion for an information-present determination and a separate, specific, reasonable statistical criterion for an information-absent determination. Classify results that do not meet either criterion as indeterminate. Recognize that indeterminate outcome is not an error, neither a false positive nor a false negative.

  18. Restrict scientific conclusions to a determination as to whether or not a subject has the specific situation-relevant knowledge embodied in the probes stored in his brain. Recognize that brain fingerprinting detects only presence or absence of information—not guilt, honesty, lying, or any action or non-action. Do not offer scientific opinions on whether the subject is lying or whether he committed a crime or other act. Recognize that the question of guilt or innocence is a legal determination to be made by a judge and jury, not a scientific determination to be made by a scientist or a computer.

  19. Evaluate error rate/accuracy based on actual ground truth. Ground truth is the true state of what a scientific test seeks to detect. Brain fingerprinting is a method to detect information stored in a subject’s brain. Ground truth is whether the specific information tested is in fact stored in the subject’s brain. Establish ground truth with certainty through post-test interviews in laboratory experiments and in field experiments wherein subjects are cooperative. Establish ground truth insofar as possible through secondary means in real-life forensic applications with uncooperative subjects. Recognize that ground truth is the true state of what the subject in fact knows, not what the experimenter thinks the subject should know, not what the subject has done or not done, and not whether the subject is guilty, or deceptive.

  20. Make scientific determinations based on brain responses. Do not attempt to make scientific determinations based on overt behavior that can be manipulated, such as reaction time.

Error rate/accuracy standards for field applications

In the United States and many other jurisdictions, the error rate of a scientific technique is critical for admissibility as scientific evidence in court. The error rate is the percentage of determinations made that are either false negatives or false positives. In brain fingerprinting, this is the percentage of “information present” and “information absent” determinations that are false positives and false negatives respectively.

In our view, in order to be viable for field use or any other application with major consequences, a technique must produce an overall error rate of less than 1 % in all studies and field applications, an error rate of less than 5 % in every individual study, and a record of consistently high statistical confidences for both information-present and information-absent determinations—averaging at least 90 % for information-present determinations and 90 % in the opposite direction for information-absent determinations, and preferably averaging over 95 % in the correct direction for all determinations of both types. To make a decision in a specific field case with judicial or other life-changing consequences, in our view the statistical confidence for the determination should be at least 95 %, whether it is information present or information absent. In our actual field applications, every individual determination to date has been with at least 99 % statistical confidence with the P300-MERMER.

Alternative methods that do not meet the above scientific standards have generally produced error rates at least ten times higher than this standard (e.g., Rosenfeld et al. 2004, 2008). Some methods that fail to meet the standards have consistently produced statistical confidences no better than chance for information-absent determinations (e.g., Rosenfeld et al. 2004, 2008).

All four of the experiments reported here followed the above overall methods. All four experiments met the brain fingerprinting scientific standards 1–20 described above.

All subjects signed informed consent forms. All procedures were approved by Brain Fingerprinting Laboratories, Inc.’s Institutional Review Board. Specific methods for each of the four studies are described below.

Study 1: the CIA Real Life Study

The CIA Real Life Study was a specific issue test. The information detected consisted of specifics regarding particular events in the lives of the subjects. In some but not all cases these life experiences included felony crimes. All of the tests, however, were conducted in circumstances where there were no judicial consequences of the outcome of the tests. Subjects were assured of confidentiality.

In the “information present” cases, probes were words or phrases associated with an event in the subject’s life. In three of the 20 cases the subjects were “information absent,” i.e., none of the stimuli were relevant to the subject. Their probe stimuli were the probe stimuli that were relevant for another subject.

Target stimuli were also relevant to the investigated event. Target stimuli, unlike probes, were identified to the subject in the course of experimental instructions. The target items were made relevant to all subjects by naming each target stimulus, explaining its relevance to the crime or investigated situation, and instructing the subject to press a special button only in response to targets. Subjects were instructed to press a button with one thumb in response to targets, and another button with the other thumb in response to all other stimuli. The prediction was that targets would elicit a P300-MERMER in all subjects, irrelevants would not elicit a P300-MERMER, and probes would elicit a P300-MERMER only in information-present subjects.

Information for structuring the stimuli was obtained from interviews with someone familiar with each subject. Subjects knew the identity of the informant for their case, and had given permission for the information to be provided for the purpose of research. As is the case in actual criminal investigations, subjects did not discuss the events and information to be detected with the experimenter prior to the testing session, or give any indication of having participated in the events or of knowing the relevant information. However, unlike the situation in actual criminal investigations, subjects were assured that results would be kept confidential. CIA Real Life Study results were not used in any legal proceedings. (In our second field study, as described immediately below, however, brain fingerprinting results were used in criminal cases and as evidence admitted in judicial proceedings in court.)

The probe stimuli were not identified as probes to the subjects. Subjects gave no behavioral indication of knowing the information contained in the probes. All stimuli were presented on a computer monitor under computer control, according to prearranged parameters that were identical for all subjects and for all stimuli. Data analysis was conducted using a standard signal-processing and mathematical analysis procedure for all subjects.

In the “information absent” cases, none of the probe items were relevant to the subject. All of the probe items were relevant to one of the other subjects. The target items were made relevant to the subject as described above. Subjects were instructed to press one button for targets and another button for all other stimuli. These “other” stimuli constituted irrelevant and probe stimuli. Unlike the information-present subjects, however, the information-absent subjects did not recognize the probes. For them, probes were indistinguishable from irrelevants, since the subjects lacked the relevant knowledge contained in the probes.

Stimuli were constructed in groups of six: one probe, one target, and four irrelevants. For each probe stimulus there were two similar irrelevant stimuli, and for each target stimulus there were two similar irrelevant stimuli. The stimuli were structured such that each probe and its similar irrelevants were indistinguishable for a subject lacking the information that the test was structured to reveal. That is, if a given probe was an article of clothing relevant to the crime or situation under investigation, two articles of clothing irrelevant to the crime were presented; if a particular probe stimulus was a name, there were two irrelevant stimuli that were also names, and so on. Similarly, there were two irrelevant stimuli that corresponded to each target.

For each subject, there were nine unique probes, nine unique targets, and 36 unique irrelevants, a total of 54 unique stimuli. These comprised nine groups of stimuli, each consisting of one probe, one target, and four irrelevants.

Testing was divided into separate blocks. In each block the computer display presented 72 stimulus presentations or trials. Three stimulus groups were presented in each block, that is, in each block there were three unique probes, three unique targets, and 12 unique irrelevants. Each stimulus was presented four times in a block to make the total of 72 stimulus presentations per block. Stimuli were presented in random order.

Stimuli were presented for a duration of 300 ms at an inter-stimulus interval of 3,000 ms. A fixation point was presented for 1,000 ms prior to each stimulus presentation.

Trials contaminated by artifacts generated by eye movements or other muscle-generated noise were rejected on-line, and additional trials were presented so that the required number of 72 artifact-free trials was obtained. The criterion for artifact rejection was as follows: trials with a range of greater than 97.7 μV in the EOG channel were rejected. This is discussed and illustrated in more detail in Appendix 3.

After three blocks, all nine groups of stimuli had been presented once. Blocks 4–6 then presented the same stimuli as blocks 1–3 respectively. The stimuli were repeated again in blocks 6–9. Thus, each of the 54 unique stimuli appeared in three different blocks, for a total of 12 presentations of each unique stimulus, and a grand total of 648 trials. Of these, 108 were probe trials, 108 were target trials, and 432 were irrelevant trials. Subjects had a rest period of approximately 2 min between blocks.

Stimulus presentation, data acquisition, and data analysis were accomplished with a PC-based system using custom software. One monitor presented the stimuli to the subject. A second monitor presented a display to the operator. During data acquisition, the display included continuous data from four channels in real time; continually updated averages of the three trial types overplotted; artifact data including values at each channel for threshold, range, slope, and mean absolute deviation, with a change in color when a rejection criterion was exceeded; reaction time and accuracy for each trial and averaged by trial type; information on the stimulus presented for each trial; counts of total and artifact-free trials by trial type; and additional information.

Brain responses were recorded from the midline frontal, central, and parietal scalp locations (Fz, Cz, and Pz respectively, International 10–20 System) referenced to linked mastoids (behind the ear), and from a location on the forehead to track eye movements. (Eye movements generate scalp potentials that interfere with the brain potentials being recorded.) Med Associates silver–silver chloride disposable electrodes were held in place by a custom headband.

Data were digitized at 333 Hz, and resampled at 100 Hz off-line for analysis. Electroencephalograph (EEG) data were amplified at a gain of 50,000 using custom amplifiers. Electro-oculograph (EOG/eye movement) data were amplified at a gain of 10,000. Impedance did not exceed 10 kilohm. Analog filters passed signals between .1 and 30 Hz. Data were stored on disk for off-line data analysis.

The primary data analysis task in these experiments was to determine whether the responses to the probe stimuli, like the responses to the target stimuli, contained a P300-MERMER brainwave pattern. We used bootstrapping (Farwell and Donchin 1988b; Wasserman and Bockenholt 1989) to determine whether the probe responses were more similar to the target responses or to the irrelevant responses, and to compute a statistical confidence for this determination for each individual subject. The bootstrapping procedure is described in more detail in Appendix 1. Appendix 2 provides graphic illustrations of key steps in the bootstrapping procedure.

We used bootstrapping to estimate the sampling distribution of two correlations: the correlation between the average of the probe trials and the average of the irrelevant trials, and the correlation between the probe average and the target average. In our computations we used “double-centered” correlations (i.e., the grand mean for all trials of all types was subtracted from the probe, target, and irrelevant average waveforms prior to the correlation computations). If the correlation between the probe and target trials is significantly greater than the correlation between the probe and irrelevant trials, then we can conclude that the probe brain responses are more similar to the targets (where a P300-MERMER is present) than to the irrelevants (where there is no P300-MERMER). If this is the case, then we can conclude that the subject recognizes the probes as a separate, rare category—that is, of situation-relevant events—and therefore that the subject is knowledgeable regarding the investigated situation. Similarly, if the correlation between the probe and irrelevant trials is greater than the correlation between the probe and target trials, then we can conclude that the subject lacks this information.

For each subject we computed the percentage of iterations in which the probe-target correlation was greater than the probe-irrelevant correlation. This provided the bootstrap index, or statistical confidence for an information present determination. This is the statistical probability that the probes, like the targets, contain the P300-MERMER or P300 pattern of interest. The bootstrap index for an information-absent determination is the percentage of iterations where the probe-irrelevant correlation is greater than the probe-target correlation. This is the probability that an information-absent determination is correct. This is equivalent to 100 % minus the probability for the information-present determination. That is, an information-present confidence of 99 % (that is, 99 % probability that the information present determination is correct) is equivalent to an information-absent confidence of 1 % (that is, 1 % probability that an information-absent determination is correct).

A decision regarding the status (information present or information absent) of a given subject depends on comparing his/her bootstrap index with criterion levels for information-present and information-absent determinations. The a priori criteria for information-present and information-absent determinations were set at 90 and 70 % respectively. These criteria were arrived at on the basis of the results of previous research (Farwell 1992; Farwell and Donchin 1991; Farwell and Smith 2001).

Prior to analysis, data were digitally filtered using a 49-point, equal-ripple, zero-phase-shift, optimal, finite impulse response, low-pass filter with a passband cutoff frequency of 6 Hz and a stopband cutoff frequency of 8 Hz (Farwell et al. 1993). Trials with eye-movement or muscle-generated artifacts were rejected by a signal-detection algorithm prior to analysis. Trials with a range of greater than 97.7 μV in the EOG channel were rejected.

We conducted two separate analyses on each subject. One analysis used the P300-MERMER, consisting of the positive P300 peak followed by the late negative peak (LNP). A second analysis included only the positive P300 peak. The P300-MERMER epoch was defined as 300–1,800 ms after the onset of the stimulus. The P300 epoch was defined as 300–900 ms after the onset of the stimulus. For subjects with markedly shorter or longer latencies than the norm, a more precise definition was applied using the target response as a template, as follows. The P300 epoch was defined as the epoch between 300 and 900 ms where the target response was more positive than the irrelevant response. The P300-MERMER epoch was defined as the P300 epoch followed by the epoch where the target response was more negative (or less positive) than the irrelevant response.

The data analysis algorithm produced two sets of results for each subject: a determination of information present or information absent and a statistical confidence for the determination using the full P300-MERMER, and a similar determination and statistical confidence using the P300 alone. This allowed us to compare the error rate/accuracy and statistical confidence provided by the state-of-the-art P300-MERMER as compared with the more widely known and well established P300.

Brain fingerprinting is a test to detect information stored in the brain. The accuracy of any system for detecting concealed information (or anything else) can only be meaningfully evaluated in light of ground truth. Ground truth is by definition the true state of exactly what the procedure is attempting to discover. For any detection method in any science, ground truth is the factual, real-world truth regarding whether the item to be detected is actually present at the time that the detection method is applied. If one conducts a DNA test to determine whether sample A matches sample B, ground truth is whether the two samples actually do in fact represent the same DNA. (Ground truth is not, for example, whether or not the suspect is guilty of a crime.)

For brain fingerprinting, ground truth is whether or not the relevant information is stored in the subject’s brain at the time of the test. Specifically, ground truth is whether or not the subject knows the information contained in the probe stimuli at the time of the test. Ground truth is not whether the subject is guilty of a crime, whether the subject participated in a knowledge-imparting procedure such as a mock crime, or whether the experimenter (or anyone else) thinks the subject should, could, or would know the information contained in the probe stimuli if he did or did not commit a crime or for any other reason. In particular, ground truth is not what the experimenter knows, or what the experimenter thinks the subject should know, or what the subject has done or not done. Ground truth is the true state of the subject–knowledgeable (information present) regarding the information contained in the probes, or not—at the time of the test.

Ground truth was established by post-test interviews. All subjects were cooperative and were not facing adverse consequences from the outcome of the test. Therefore it was possible to establish ground truth with a high degree of certainty through post-test interviews. The significance of each probe stimulus was described to each subject in post-test interviews, and the subject was asked to identify the correct probe stimulus. Post-test interviews established that all information-present subjects knew the information contained in all the probe stimuli, and no information-absent subjects knew the information contained in any of the probe stimuli. The correctness of determinations was evaluated in light of ground truth.

Study 2: the Real Crime Real Consequences $100,000 Reward Study

The Real Crime Real Consequences $100,000 Reward Study was a specific issue study involving real-world events with real, substantial consequences. We tested brain fingerprinting on 14 subjects3 in circumstances where subjects were highly motivated by real-world consequences of the outcome of the tests. We used brain fingerprinting to detect concealed information regarding real crimes, in circumstances where the outcome of the test could produce major, life-changing consequences. Some of the subjects were suspects in criminal investigations or convicted prisoners who claimed innocence and were appealing their convictions. In some cases the subjects were facing the death penalty or life in prison, and the outcome of the brain fingerprinting test could provide legally admissible evidence relevant to the case and the ensuing consequences.

In some cases, although the crimes were real, there were no reasonably foreseeable, life-changing legal consequences of the outcome of the brain fingerprinting test. To produce a life-changing impact in cases where no judicial outcome hinged on the scientific results, we offered subjects a $100,000 reward for beating the test. Beating the test means producing a false negative result: producing an information-absent determination when the subject knew the relevant knowledge, so the correct determination would have been information present.

Such subjects were taught countermeasures that had previously proven effective against other, fundamentally different, non-brain fingerprinting tests (but not against brain fingerprinting; see Farwell 2011a, b, 2012). One countermeasure (Rosenfeld et al. 2004) involves instructing subjects to attempt to enhance responses to irrelevant stimuli. This is done by dividing the irrelevant stimuli into categories and performing a specific covert act such as wiggling the big toe in the left shoe in response to certain specific categories. An alternative countermeasure involves instructing subjects to attempt to enhance responses to target stimuli by thinking of being slapped or applying pressure to their toes in response to each target stimulus (Mertens and Allen 2008). We taught subjects one or the other of these countermeasures, using identical subject instructions to those applied in Rosenfeld et al. and Mertens and Allen.

Probe stimuli were obtained by the criminal investigators involved through the usual evidence-collection procedures involved in criminal investigations, including interviewing witnesses and accomplices, inspecting the crime scene, examining police reports and other investigative reports, reviewing court records, etc. Target stimuli were obtained through similar sources, and also through publicly available information such as news reports.

Unlike the procedures in the CIA Real Life Study, confidentiality in the Real Crime Real Consequences $100,000 Reward Study was not maintained in cases where the brain fingerprinting results were relevant to a current criminal case. Scientific reports and expert testimony were provided as appropriate in relevant judicial proceedings. Subjects were informed of this in advance, and all subjects signed informed consent forms. For the purpose of this report, however, individual subject confidentiality is maintained.

The purpose of the brain fingerprinting test in each unsolved criminal case was to determine whether or not the information contained in the probe stimuli, provided by the criminal investigator as putative features of the crime, was stored in the brain of the subject. The brain fingerprinting determinations and statistical confidence, and the resulting reports and expert witness testimony in court, addressed only the question of whether the specific crime-relevant information contained in the probes was known to the subject.

Brain fingerprinting testing, and the brain fingerprinting scientists, did not provide an opinion regarding the guilt or innocence of the subject. The brain fingerprinting test did not address, and the brain fingerprinting scientists did not opine regarding, the effectiveness of the criminal investigation, the relevance of the probes to the crime, or the probative value of the brain fingerprinting results with respect to the question of who committed the crime.

Attorneys and prosecutors on both sides did debate these matters, referring to common sense, life experience, and other sources outside the realm of science to support their contentions. These are non-scientific issues that are decided by a judge and jury based on their human judgment, life experience, and common sense. The question of guilt or innocence is decided by a judge and jury, not by a scientist or a computer. The brain fingerprinting test results simply provided the judge and jury with additional evidence that they weighed along with the other evidence in reaching their findings of fact regarding what took place at the time of the crime and their legal verdicts regarding guilt or innocence.

Data acquisition and data analysis methods were the same as described for the CIA Real Life Experiment, except for the following. The number of probes, the number of data acquisition blocks, and the timing of the tests varied depending on the individual circumstances of the various cases. To provide ample data even in adverse circumstances, all subjects were scheduled to be tested in 25 blocks of 72 trials each. In some cases tests were conducted in prisons, and modifications were necessary to meet prison scheduling requirements and other logistical issues. Some tests actually consisted of fewer blocks and fewer trials than the number originally scheduled, but in no case fewer than the methods for the CIA Real Life Study described above. The reference was linked ears. Digitizing rate was 100 Hz.

Brain fingerprinting is a test to detect information stored in the brain. As discussed above, ground truth is the true state of whatever the test is attempting to determine. With brain fingerprinting, ground truth is whether or not the subject knew the information contained in the probe stimuli at the time of the test. In field studies involving any forensic science, it is never possible to establish ground truth with absolute certainty. Our study is no exception. Neither we nor any other forensic scientist applying any forensic science test in the real world can know with absolute certainty what ground truth is.

For information-present subjects, ground truth was that the crime-relevant information contained in the probe stimuli was stored in the brain of the subject at the time of the test. Culpability for the crimes was established with a relatively high degree of certainty through confessions, corroborated in every case with judicial outcome when relevant. Confessions, particularly when combined with convictions, can establish a reasonably high degree of confidence (although not an absolute certainty) that the subject is guilty of the crime. This does not establish ground truth, however, because ground truth for brain fingerprinting is not whether the subject is guilty but whether the subject knows the information contained in the probes.

The only way to establish ground truth with absolute certainty is for the subject to correctly identify the probes in post-test interviews, without ever having been told which stimuli are the correct, crime-relevant probes. Fortunately, all our information-present subjects eventually confessed and cooperated. In post-test interviews, we described the significance of each of the probe stimuli in the context of the crime, and asked the subject to identify the correct probe stimulus. All of the information-present subjects correctly identified all of the probe stimuli. Obviously, a subject can do this only if ground truth is that he does know the crime-relevant information contained in each probe. Thus, ground truth was established with certainty for all of our information-present subjects. The correctness of determinations was evaluated in the light of real-world ground truth.

When ground truth is the absence of something, this can never be absolutely proven without some kind of assumption or circular reasoning. Just because people have looked almost everywhere and no one has ever found a pink elephant does not absolutely prove that no such thing exists.

In the case of information-absent subjects, ground truth was that the information contained in the probes was not stored in their brains at the time of the test. For our subjects, in every case someone else confessed and/or was convicted of the crime for which the probe stimuli were relevant, and sworn witness testimony, compelling physical evidence, and judicial findings of fact held that our subject was not present at the crime scene. This established with a rather high degree of certainty that the subject did not know the relevant information through participation in the crime.

The remaining possibility was that the subject somehow knew the information contained in the probes for an innocent reason. This information had never been publicly released, and court and investigative records stated that the subject had never been exposed to the information. Moreover, subjects were provided an opportunity to disclose any knowledge of the information contained in the probes that they may have found out though some unknown innocent means. The significance of each probe stimulus was clearly described to the subjects, and they were asked if they knew the correct crime-relevant information for any reason. They had an extremely high motivation to disclose any such innocently acquired knowledge if it existed, because such disclosure would provide an innocent reason why they would be found to possess crime-relevant information that would imminently be detected by the brain fingerprinting test. No subjects offered any innocent reason (or any reason) why they might know any of the information contained in any of the probes.

The probability is vanishingly small that a subject knew specific never-released information about a crime for which someone else had confessed and been convicted, when official records and judicial findings of fact established that he had no possible known way of knowing the information, and when he had opportunity and extremely high motivation to reveal such knowledge if it had been acquired through some previously unknown innocent means. Although, as in all field studies in any forensic science, ground truth cannot be established with absolute certainty, ground truth was established with a high degree of certainty for all information-absent subjects. The correctness of determinations was evaluated in the light of real-world ground truth.

Study 3: the FBI Agent Study

The FBI Agent Study was a specific screening study. The relevant information detected was information known to FBI agents but not to the general public, obtained from interviews with FBI agents. We tested 17 FBI agents (information present) and four non-agents (information absent).

Data acquisition methods were the same as for the above studies, except for the following. Stimuli were visually presented words, phrases, and acronyms that are well known to FBI agents and not to the general public. There were a total of 33 probes, 33 targets, and 132 irrelevants. Testing consisted of six blocks of 72 trials each. In each block subjects viewed either five or six sets of stimuli, each set consisting of one probe, one target, and four irrelevants.

The reference was linked mastoids. Digitizing rate was 100 Hz.

Data analysis methods were the same as in the above described studies, except for the difference in recording and analysis epoch described below.

Ground truth was whether the subject knew the FBI-relevant information contained in the probes at the time of the test. Since all subjects were fully cooperative, ground truth could be established by post-test interviews. Post-test interviews established that all information-present subjects knew all of the probes. No information-absent subjects knew any of the probes.

The FBI Agent Study was the first time brain fingerprinting was used to detect specific group knowledge rather than specific issue knowledge. This study also resulted in two other innovations, as follows.

The FBI Agent study was the first study in which we used targets that contained information relevant to the investigated situation (in this case, inside knowledge relevant to the FBI). In previous studies we had used targets that were irrelevant to the investigated situation, and were made relevant only by subject instructions that informed the subjects which stimuli were targets and required them to push one button in response to targets and another button in response to all other stimuli.

Using target stimuli which, like the probes, are known and recognized as significant for an information-present subject increases the accuracy of the system. Recall that the data analysis involves comparing the probe-target correlation with the probe-irrelevant correlation. For an information-present subject, any procedure that maximizes the probe-target correlation and/or minimizes the probe-irrelevant correlation will improve discrimination and increase accuracy.

Situation-relevant targets improve the accuracy of the system particularly when the stimuli are acronyms, as some of them were in both the FBI Agent Study and the Bomb Maker Study. In these studies, only acronyms were presented in some blocks, and only words and phrases were presented in other blocks.

Consider the case of an information-present subject when the probe stimuli are acronyms known to the subject and the targets and irrelevants are both meaningless letter strings. Both targets and probes are relevant and noteworthy, so both will elicit large P300-MERMERs. However, the information-present subjects will be able to identify the probes more quickly than the targets and irrelevants. They can immediately recognize probes as acronyms. Targets and irrelevants, by contrast, are both meaningless letter strings. When the stimulus is not an immediately recognizable probe, subjects must search through the strings to determine if they are among the target strings they have been given. Since they can recognize the probes more quickly, the latency of the P300-MERMER to the probes will tend to be less than the latency of the P300-MERMER to both targets and irrelevants. The result is that the targets will resemble the probes in P300-MERMER amplitude (i.e., large) but will resemble the irrelevants in P300-MERMER latency (i.e., long). When correlations are computed between the responses to the respective trial types, the correlation between the targets and the probes—which should be high in an information present subject because both elicit large P300-MERMERs—will tend to be attenuated because of the latency difference. The peaks will not line up. The result is that the determinations rendered by the system will tend to be less accurate.

When targets are relevant acronyms known to an information present subject, the P300-MERMER latency differences serve to improve discrimination. In this situation, an information-present subject can quickly recognize both targets and probes as known acronyms. Therefore both targets and probes will elicit P300-MERMERs with short latency (as well as high amplitude). The irrelevants, being meaningless letter strings, will be deciphered more slowly, and will elicit a longer latency response (with a very small P300-MERMER, if any). As before, both probes and targets will elicit relatively high amplitude P300-MERMERs. Thus, the correlation between targets and probes will be high because both have large P300-MERMERs of similar (short) latency. The correlation between probes and irrelevants will tend to be lower not only because the irrelevants have a small P300-MERMER (if any) but also because the P300-MERMER latencies do not match—irrelevant brain-response latencies are longer than probes. A higher correlation between targets and probes, combined with a lower correlation between probes and irrelevants, will increase the probability of a correct information-present determination.

This same phenomenon takes place, albeit to a lesser degree, when the stimuli are words or phrases. Using crime-relevant words or phrases maximizes the similarity between the probe and target stimuli, and hence the similarity between the probe and target brain responses, for an information-present subject. This maximizes the probe-target correlation if, and only if, the subject is factually information present, which increases accuracy and statistical confidence. Standard 4 includes this feature.

All of the above applies only to an information-present subject. For a subject without the relevant information, using targets that are relevant acronyms will be no different than using meaningless strings, because she will not recognize the acronyms. For an information-absent subject, all of the stimuli will be perceived as meaningless letter strings. Thus, no latency differences between trial types are to be expected. The discrimination will be made on the basis of amplitude. For such a subject, the targets are noteworthy solely because she has been instructed to press a special button only for targets. The target P300-MERMER amplitude will be greater than that for probes and irrelevants, as usual for an information-absent subject. Thus the probes will resemble the irrelevants rather than the targets.

The second innovation of the FBI Agent Study was that it was the first study in which we recorded a long enough data epoch to observe the full late negative potential (LNP) of the P300-MERMER. In early studies, P300 research had used relatively short stimuli such as clicks, tones, and single words or short phrases. P300 researchers typically presented a stimulus every 1–1.5 s. In our previous brain fingerprinting work, for example, we (Farwell 1992; Farwell and Donchin 1991) used an inter-stimulus interval of 1,500 ms.

In the FBI Agent Study we were required to use stimuli that accurately represented knowledge of the FBI. It was not practical to use only short words and short phrases. Some of the stimuli consisted of several words of several syllables each. To give the subjects time to recognize and process the stimuli, we extended the inter-stimulus interval to 3,000 ms. As is described in the Discussion section, we discovered that when the stimuli are longer phrases, which take some time for the subject to decipher, and the inter-stimulus interval is sufficiently long to display the full response, the positive P300 peak is followed by a late negative peak (LNP). The LNP has a peak latency sometimes as long as 1,500 ms, and it sometimes does not resolve to baseline until around 1,800 ms. We called this full response, including both the P300 and the LNP, a memory and encoding related multifaceted electroencephalographic response or P300-MERMER. Other features of the P300-MERMER are discussed in the Discussion section and in Farwell (1994, 1995b, 2007, 2010, 2012) and Farwell and Smith (2001).

Because when we designed the study we had not yet discovered this late negative peak (LNP) of the P300-MERMER, we recorded and analyzed a shorter data epoch than in the above described studies (which we conducted subsequently). In the FBI agent study, the data recording and analysis epoch ended at 1,250 ms after stimulus onset. In subsequent studies we have used a 1,800-ms data analysis epoch. (We also record continuous data.)

Study 4: the Bomb Maker Study

The Bomb Maker Study was a specific screening study. The information detected was information that is known to individuals with extensive experience in making, detecting, detonating, and deactivating or destroying bombs. Information present subjects were explosive ordnance disposal (EOD) and improvised explosive device (IED) experts. We obtained the relevant information that comprised the probe and target stimuli through interviews with EOD/IED experts. We tested 17 information-present subjects and two information-absent subjects. Stimuli were words, phrases, and acronyms that are known to EOD/IED experts and not to the general public.

Data acquisition methods were the same as for the above studies, except for the following.

For each subject, there were 21 unique probes, 21 unique targets, and 84 unique irrelevants, a total of 126 unique stimuli. These comprised 21 groups of stimuli, each consisting of one probe, one target, and four irrelevants.

Testing was divided into separate blocks. In each block, either four or five stimulus groups were presented, that is, in each block there were either four unique probes, four unique targets, and 16 unique irrelevants, or five probes, five targets, and 20 irrelevants. Each block consisted of 72 trials, or individual stimulus presentations, with presentations of unique stimuli repeated as necessary to reach this total. We presented 10 blocks for each subject. Thus, we presented a total of 720 trials for each subject, including 120 probe trials, 120 target trials, and 480 irrelevant trials.

In block 1 we presented the first five stimulus groups, i.e., five unique probes, five unique targets, and 20 unique irrelevants. In each of blocks 2–5 we presented four of the remaining stimulus groups. At the end of the first five blocks, each of the 21 stimulus groups had been presented in one block. The same pattern was repeated for blocks 6–10.

The reference was linked ears. Digitizing rate was 100 Hz.

Data analysis methods were the same as in the above described studies.

Ground truth was whether the subject knew the bomb-making relevant information contained in the probes at the time of the test. Since all subjects were fully cooperative, ground truth could be established by post-test interviews. Post-test interviews established that all information-present subjects knew all of the probes. No information-absent subjects knew any of the probes.

Results

Overview of results for the four brain fingerprinting studies

The target stimuli elicited a large P300-MERMER in all subjects. This is as expected, since the targets contained known, relevant information for all subjects. Also as expected, the irrelevant stimuli did not elicit a large P300-MERMER in any subjects. As predicted, the probe stimuli elicited a large P300-MERMER only in the information-present subjects, and not in the information-absent subjects. In the information-present subjects, the response to the crime-relevant (or situation-relevant) probes was similar to the response to the known targets: both contained a large P300-MERMER. In the information-absent subjects, the response to the crime-relevant (or situation-relevant) probes was similar to the response to the irrelevant stimuli: neither contained a large P300-MERMER.

The brain fingerprinting data-analysis algorithm using the P300-MERMER produced the following overall results: Error rate was 0 %. 100 % of determinations were correct. There were no false negatives and no false positives. There were also no indeterminates. As in Farwell and Donchin (1991), Farwell and Smith (2001), and all other previous brain fingerprinting research, Grier (1971) A’ was 1.0.

The criterion for an information-present determination was a 90 % statistical confidence, computed by bootstrapping. The criterion for an information-absent determination was a 70 % statistical confidence in the opposite direction. All determinations with the P300-MERMER, both information-present and information-absent, achieved over 95 % statistical confidence. In studies 1 and 2, the specific issue studies, all determinations with the P300-MERMER achieved at least 99 % confidence. The median statistical confidence for the individual determinations with the P300-MERMER in all studies was 99.9 %. The mean statistical confidence for the individual determinations with the P300-MERMER in all studies was 99.5 %.

The brain fingerprinting data-analysis algorithm using the P300 alone produced the same determinations: 100 % of determinations were correct. Error rate was 0 %. Accuracy was 100 %.4 There were no false negatives and no false positives. Also, there were no indeterminates. The median statistical confidence for the individual determinations with the P300 in all studies was 99.6 %. The mean statistical confidence for the individual determinations with the P300 in all studies was 97.9 %. All determinations with the P300, both information-present and information-absent, achieved over 90 % statistical confidence.

The P300-MERMER produced significantly higher statistical confidences than the P300 alone, p < .0001 (sign test). For most of the subjects (57 %), the statistical confidence for the P300-MERMER-based determination was higher than the statistical confidence for the P300-based determination.

Table 1

One thought on “Introduction Brain Fingerprinting Essay

Leave a comment

L'indirizzo email non verrà pubblicato. I campi obbligatori sono contrassegnati *