Brain, Vol. 123, No. 4, 770-780,
April 2000
© 2000 Oxford University Press
Visual confrontation naming and hippocampal function
A neural network study using quantitative 1H magnetic resonance spectroscopy
Epilepsy Center, Department of Neurology, University of Alabama at Birmingham, USA
Correspondence to:
Stephen M. Sawrie, PhD, UAB Epilepsy Center, 1719 6th Avenue South, CIRC 312, Birmingham, AL 35294 USA E-mail: ssawrie{at}uab.edu
 |
Abstract |
|---|
 Top Abstract IntroductionMethodsResultsDiscussionReferences |
|---|
Prior research on the relationship between visual confrontation naming and hippocampal function has been inconclusive. The present study examined this relationship using quantitative 1H magnetic resonance spectroscopy (1H-MRS) to operationalize the function of the left and right hippocampi. The 60-item Boston Naming Test (BNT) was used to measure naming. Our sample included 46 patients with medically intractable, focal mesial temporal lobe epilepsy who had been screened for all pathology other than mesial temporal sclerosis. Statistics included Pearson correlations and neural network analysis (multilayer perceptron and radial basis function). Baseline BNT performance correlated significantly with left 1H-MRS hippocampal ratios. Thirty-six per cent of the variance in baseline BNT performance was explained by a neural network model using left and right 1H-MRS ratios(creatine/N-acetylaspartate) as input. This was elevated to 49% when input from the right hippocampus was lesioned mathematically. In a second model, left 1H-MRS hippocampal ratios were modelled using measures of semantic and episodic memory as input (including the BNT). Explained variance in left 1H-MRS hippocampal ratios fell from 60.8 to 3.6% when input from BNT and another semantic memory measure was degraded mathematically. These results provide evidence that the speech-dominant hippocampus is a significant component of the overall neuroanatomical network of visual confrontation naming. Clinical and theoretical implications are explored.
temporal lobe epilepsy; naming; hippocampus; neuropsychology; neural networks; magnetic resonance spectroscopy
BNT = Boston Naming Test; COMP = Comprehension subtest of the Wechsler Adult Intelligence Scale—Revised; Cr = creatine; CVLT = California Verbal Learning Test; 1H-MRS = proton magnetic resonance spectroscopy; LM% = Wechsler Scale Logical Memory per cent retention; MLP = multilayer perceptron; MTLE = mesial temporal lobe epilepsy; NA = N-acetylated compounds; PDP = parallel distributed processing model; RBF = radial basis function; SOP = `stages of processing' model; TIR = inversion recovery delay; WAIS-R = Wechsler Adult Intelligence Scale—Revised
|
Introduction |
|---|
|
TopAbstractIntroductionMethodsResultsDiscussionReferences |
|---|
The biological neural network of visual confrontation naming has received considerable attention in recent literature. Neural substrates of naming, identified in previous literature, include various sites in the speech-dominant temporal neocortex (Ojemann et al., 1993
), the ventral visual pathway from visual to frontal cortex (Kiyosawa et al., 1996) and the left inferior temporal, superior parietal and frontal cortices (Wiggs et al., 1999). Naming ability has also been associated with the function of the dominant left hippocampus (Snyder et al., 1994; Shear et al., 1997; Davies et al., 1998; Seidenberg et al., 1998; Martin et al., 2000). However, other studies have failed to find an association between naming and hippocampal function.
The question of whether or not the speech-dominant hippocampus is involved in the biological neural network of confrontation naming has important clinical implications. For instance, there has been some support for employing intra- or extra-operative neocortical language mapping to tailor the extent of lateral resection in a left, speech-dominant anterior temporal lobectomy (Ojemann et al., 1993; Hermann et al., 1999) in order to spare naming. However, cortical mapping may be unwarranted if naming ability is associated with hippocampal rather than lateral temporal neocortical function. A second clinical consideration includes the non-invasive, preoperative identification of those at risk for postoperative declines in naming. Currently, a model exists for prediction of episodic memory change following anterior temporal lobectomy which is based on the functional adequacy of the hippocampus to be resected (Chelune et al., 1991; Chelune, 1995). As yet, however, the functional adequacy model does not predict decline on measures of confrontation naming or other measures of semantic memory.
Determination of the relationship between dominant left hippocampal pathology and confrontation naming also has important implications in the debate between competing psychobiological models of memory function. Confrontation naming has been studied traditionally in the context of semantic (versus episodic) memory. Semantic memory is thought to consist of world knowledge, linguistic rules, vocabulary and other language-dependent memories, whereas episodic memory is defined commonly as memory for events or episodes from personal experience (Tulving, 1972, 1984; Horner, 1990; MacKay et al., 1998). The prevailing neuroanatomical memory theory suggests that the mesial temporal lobe is responsible for episodic memory, whereas lateral temporal neocortex is more the domain of semantic memory (Tulving, 1985; Zola-Morgan and Squire, 1993). More recent connectionist theories view memories as patterns of connections, where the type of memory depends largely on the strength and distribution of these connections (Mesulam, 1990). In this sense, the distinction between semantic and episodic memory is defined by the strength and distribution of their connections within the temporal lobe memory system. If the dominant hippocampus is implicated in confrontation naming, this would weaken the argument for anatomically separate episodic and semantic memory systems and strengthen arguments emphasizing connectionism.
Mesial temporal lobe epilepsy (MTLE) provides perhaps the best prototype for the study of the relationship between naming and hippocampal pathology, especially when patient samples are screened for all pathology other than mesial temporal sclerosis. Table 1 presents the results of a comprehensive search of studies on hippocampal pathology and visual confrontation naming in patients with MTLE. The search was restricted to studies which included a direct, quantifiable measure of hippocampal pathology. We did not review studies which based their findings only on memory change following surgery for MTLE, since the standard anterior temporal lobectomy involves resection of both mesial temporal cortex and neighbouring temporal neocortex. We also restricted our search to studies which assessed naming with a standardized, psychometric measure.
|
The majority of studies found some type of statistically significant relationship between the size or function of the dominant left hippocampus and performance on a measure of confrontation naming (Snyder et al., 1994
; Shear et al., 1997; Davies et al., 1998; Seidenberg et al., 1998; Martin et al., 2000). Three studies were unable to find a significant relationship (Sass et al., 1992; Westerveld et al., 1996; Baxendale et al., 1998). However, two commonalities among these three studies raise methodological concerns. First, each of these three studies used neuronal or glial cell densities to quantify the status of the hippocampus. Recent studies have criticized the use of cell densities as a measure of hippocampal pathology because of difficulties in standardizing the reference volume (Coggeshall and Lekan, 1996; Van Paesschen et al., 1997). Secondly, use of cell densities requires the separation of left and right MTLE patients in statistical analysis. This significantly diminishes the range of hippocampal values, and enhances the possibility of a Type II error. There were also several methodological concerns among the studies which found a significant relationship between hippocampal function and confrontation naming. Two of these studies were abstracts and, as such, had not undergone journal-level peer review (Snyder et al., 1994; Shear et al., 1997). Two other studies reduced hippocampal function to a dichotomous variable (presence or absence of histopathologically determined mesial temporal sclerosis) (Davies et al., 1998; Seidenberg et al., 1998). Thus, the research in this area remains inconclusive.
Recent advances in neuroimaging and statistical analysis might assist in resolving these inconsistencies. Proton magnetic resonance spectroscopy (1H-MRS) has been used recently as a means of linking cerebral metabolic function with cognitive function. Gadian and colleagues used 1H-MRS to assess levels of N-acetylaspartate relative to creatine and choline (NA/Cr + choline) in the left and right mesial temporal lobes of 22 children with intractable MTLE, and found significant and selective correlations between left-sided pathology and verbal functions as well as right-sided pathology and non-verbal functions (Gadian et al., 1996). Martin and colleagues demonstrated recently the ability of 1H-MRS to detect subtle functional cognitive disturbances related to hippocampal neuronal abnormalities (Martin et al., 2000). Hugg and colleagues used 1H-MRS to demonstrate the normalization of unoperated hippocampi that had been metabolically abnormal prior to the resection of epileptogenic hippocampi (Hugg et al., 1996). Taken together, these studies suggest that 1H-MRS may be more dependent on the function (rather than the structure) of a neuroanatomical substrate. As such, this technique appears to be particularly useful in studies of brain–behaviour relationships.
Computational neural network analysis has also demonstrated increasing promise in studies of brain and cognition. Neural network modelling is a method of data analysis that encompasses a broad array of computational techniques. Common among each of these techniques are the characteristics of parallel processing, distributed storage of knowledge and response to the environment (i.e. learning). A neural network typically involves layers of nodes which are conceptually similar to neurons. An initial layer of input nodes receives data from the environment and feeds it to one or more layers of hidden nodes, which in turn transform the data computationally before sending them to a layer of target, or output, nodes. In this way, a neural network resembles the general architecture of the nervous system where information flows from receptors, through interneurons, and then to effectors. Since a pattern of input data can be considered simultaneously, the processing of those data by the neural net is said to be parallel. This feature significantly enhances the efficiency and computational power of neural nets over more traditional, sequential methods of data processing. A neural net will store `knowledge' of the solution to which it is applied in weights between the input and hidden layers, and between the hidden and output layers. In this way, the `knowledge' gained by a neural net is distributed over the entire architecture. By corresponding with the output layer, these weights can be adjusted to solve a problem maximally. It is in this way that a neural net `learns' the solution to a particular problem. This novel type of analysis has already shown promise in psychobiological studies of cognitive domains such as attention (Mesulam, 1981), executive function (Cardoso and Parks, 1998), language (Seidenberg and McClelland, 1989) and memory (Hasselmo et al., 1998).
The present investigation represents a synthesis of these two technologies (i.e. 1H-MRS and neural network analysis) in two studies examining the relationship between hippocampal pathology and confrontation naming. The first study addresses the question of whether or not the dominant left hippocampus mediates naming by modelling performance on the Boston Naming Test (BNT) (Kaplan et al., 1978) using 1H-MRS hippocampal values. The second study examines this relationship within the context of recent competing theories of memory by modelling dominant left hippocampal 1H-MRS ratios using a small battery of episodic and semantic memory tests (including BNT). We performed these studies in a sample of patients with MTLE who had been screened for all pathology other than mesial temporal sclerosis. Using this best prototype of mesial temporal pathology, we hope to address some of the inconsistencies in current literature regarding naming and hippocampal pathology.
|
Methods |
|---|
|
TopAbstractIntroductionMethodsResultsDiscussionReferences |
|---|
Subjects
The subject sample consisted of 46 patients (18 men, 28 women) evaluated for surgery at the University of Alabama at Birmingham Epilepsy Center. Mean age was 34.9 years (SD = 10.2), mean education was 12.4 years (SD = 2.3). Average age of seizure onset was 13.9 years (SD = 10.6), whereas average duration of illness was 21.1 years (SD = 12.2). All patients were screened for other disease processes such as tumours or vascular lesions (i.e. dual pathology). IAP (intracarotid amobarbitol procedure) confirmation of language dominance was not available since in our centre IAP is not carried out routinely in all candidates for surgery. However, all patients were exclusively right hand dominant. Lateralization of seizure onset was either exclusively or primarily left temporal in 32 patients and right temporal in 14 patients. To date, 34 of the 46 patients have undergone either a left (n = 24) or right (n = 10) temporal lobe resection for treatment of medically intractable MTLE. Each surgery patient had pathological confirmation of mesial temporal sclerosis. The remaining 12 patients had MRI characteristics consistent with mesial temporal sclerosis (Kuzniecky et al., 1997
). Informed consent was obtained from each patient. The study was approved by the University of Alabama at Birmingham Review Board.
1H-MRS studies at 4.1 T
1H-MRS data were acquired using a 4.1 T whole body imaging/spectroscopy system and quadrature-driven tunable matchable head coil (Vaughan et al., 1994). Sagittal scout images were acquired using a segmented (eight encodes per inversion pulse) inversion recovery gradient echo sequence (TR/TIR/TE 2500/1000/15) (Pan et al., 1995). First, a multislice set of axial images was used to locate the plane of the interhemispheric fissure. Secondly, a multislice set of sagittal images centred on the interhemispheric fissure was used to locate the obliquely axial hippocampal plane. Finally, a multislice set of obliquely doubly angulated axial images (160 s, two averages) was acquired to visualize the hippocampal plane. We used the non-selective water signal for whole-head shimming to better than 20 Hz [0.11 parts per million (p.p.m.) at 175 MHz]. Further shimming on the 1H-MRS slice achieved 10–15 Hz (0.06–0.09) p.p.m. for water. The first 20 studies were done with angulation in the anterior–posterior directions approximately parallel to the anterior–posterior commisure line. In the last 26 studies, the degree of angulation in the anterior–posterior directions was determined so as to be approximately parallel to the long axis of the hippocampus.
The first 20 1H-MRS studies were acquired using a single slice spectroscopic imaging sequence consisting of an initial adiabatic inversion pulse followed by a slice-selective excitation pulse and a broadband semi-selective pulse (Hetherington et al., 1994b). Contributions from extracerebral lipids were suppressed using an adiabatic inversion pulse followed by dephasing gradients and an inversion recovery delay (TIR) optimized to suppress lipids (265 ms). Water suppression was performed using a broad band semi-selective refocusing pulse, providing >95% refocusing efficiency over the range of 3.4–0.7 p.p.m (Hetherington et al., 1994b). Phase-encoding gradients were applied during the initial echo period. For the remaining 26 MRS studies, optimized angulated single slice spectroscopic images were acquired using a 2D PRESS (double spin-echo) pulse sequence. Suppression of water and lipid signals near the anterior sinus cavity was achieved in part by a slice-selective refocusing pulse which excluded the anterior region of severe Bo (magnetic field) susceptibility gradients. Extracerebral lipids were suppressed using the same adiabatic pulse. Dephasing and rephasing gradients (`crushers') were applied during the echo periods to eliminate the necessity for phase cycling. Phase-encoding gradients were applied before the semi-selective refocusing pulse.
Data were acquired using TR/TIR/TE of (2000/265/50) for the single spin-echo technique. The TE during the 2D PRESS sequence was 58 ms. The spectroscopic image was acquired using a field of view of 240 x 240 mm utilizing 32 x 32 phase encodes with a slice thickness of 1 cm. The SI data were zero filled to 64 x 64 and filtered using a quarter cosine filter in the spatial domain (Hetherington et al., 1994a, b). The spectral domain was processed using a convolution difference of 50 Hz to eliminate broad water components, followed by 3 Hz of exponential broadening. Two spatial and one spectral Fourier transforms were then performed to generate the 2D spectroscopic imaging.
Using the scout anatomical image, a rectangular region of interest was selected which included the midbrain, the hippocampus and portions of the temporal lobe (both anterior and posterior to the hippocampus) along with portions of the cerebellar vermis. The region encompassed ~400–600 of the zero-filled voxels. The spectra within each voxel were then corrected for Bo shifts by setting the maximum resonance intensity in the vicinity (30 Hz) of the N-acetylated compounds (NA) resonance (as predicted by the whole slice water resonance position) to 2.02 p.p.m. All voxels within the region of interest were then analysed using NMR1 (Tripos, St Louis, Mo., USA), and the linewidth, resonance area and chemical shift were determined. From these data, the creatine to N-acetylated compounds (Cr/NA) ratios were determined for all pixels.
To exclude regions from outside the head or within ventricles (where metabolite content is zero and the ratio is therefore poorly defined), the Cr/NA ratio was set to zero in voxels where the fitted chemical shifts were more than ±0.05 p.p.m. from their predicted values or linewidths were >20 Hz (typical linewidths are 6–12 Hz). The ratio maps were then interpolated to 256 x 256. Those regions showing significant increases in Cr/NA, >2 SD exceeding the normal 95% confidence, higher than any of the voxels measured from the healthy volunteers, were then highlighted and overlaid on the anatomical image (Kuzniecky et al., 1998).
The 1H-MRS data were obtained from each patient using the statistically thresholded area covering the hippocampal body and adjacent areas in the axial plane of the spectroscopic image as described above. The 1H-MRS slice primarily covered the hippocampal body and tail in the first 20 studies, with more anterior sampling in the last 26 patients. Table 2 presents the means and standard deviations of the 1H-MRS ratios for the left and right hippocampi. Greater Cr/NA ratios reflect a greater degree of metabolic dysfunction.
|
Neuropsychological measures
The 60-item BNT was used to assess visual confrontation naming (Kaplan et al., 1978
). The BNT requires the patient to recall the names of visually presented line drawings of various objects. Semantic and phonemic cues may be given if the patient is unable freely to retrieve the correct word. Scores were generated according to standard BNT protocol. The Comprehension (COMP) subtest from the Wechsler Adult Intelligence Scale—Revised (WAIS-R) (Wechsler, 1981) was chosen as a second measure of semantic memory (scaled score). COMP requires a verbal response to questions involving common-sense judgment and practical reasoning. The Wechsler Memory Scale Logical Memory per cent retention (LM%) (Wechsler, 1945) and the California Verbal Learning Test (CVLT) (Delis et al., 1987) total score were selected as measures of episodic verbal memory (raw scores). LM% measures the percentage of retained material from orally presented stories over a 30-min delay, whereas the CVLT total score measures the total number of words recalled after five administrations of a novel 16-word list. Neither age nor education correlated significantly with any of these measures, and were therefore not used as moderating variables in further analyses. Table 2
presents mean baseline performance on each measure.
Statistical analysis
Parametric analysis was performed using the Statistical Package for the Social Sciences (SPSS). Neural network analysis was performed using a separate SPSS module called Neural Connection. Pearson correlations were used to evaluate the linear association between 1H-MRS hippocampal ratios and neuropsychological measures. Two related neural network techniques, radial basis function (RBF) and multilayer perceptron (MLP), were used to derive optimal solutions for the two modelling studies. RBF capitalizes on clusters of input data in the data space. The network attempts to capture these clusters by placing hidden nodes, or centres, randomly throughout the data space. These centres are represented by a non-linear transfer function, 
, centred around the weight of the hidden node, y. The output, h, of hidden node k is represented as 
where (||x – yj ||) is the non-linear transfer of the distance between the input, x, and the hidden node weight, y. This non-linear transfer is then multiplied by the weight attached to the output node, 
, to produce an output 
. The final network solution is the sum of the outputs from the centres. (See Lowe, 1995 for a more detailed discussion of RBF.) MLP attempts to solve a problem by partitioning the data space into separable hyperplanes. The basic architecture of an MLP neural network consists of the input vector, the output layer and one or more intervening hidden layers. The nodes of these layers typically are fully connected. Weights are assigned randomly to the connections between the input layer and hidden layer, and between the hidden layer and output layer. The output of a hidden node is a non-linear transformation of the sum of the weighted inputs at that particular node. The non-linear transfer function typically is tanh or sigmoid. An advantage of MLP is its use of backpropagation, which is a technique in which the weights in the model are adjusted in order optimally to reduce the error between modelled output and target output. (See Hertz et al., 1991 for a more detailed discussion of MLP and backpropagation.)
In the first study, left and right 1H-MRS hippocampal values were used to model performance on the BNT. In the second study, the four measures of semantic or episodic memory were used to model left 1H-MRS hippocampal values. In both studies, 75% of the sample were chosen randomly as the training group, with the remaining 25% used as a validation sample (the validation sample is used in neural network analysis to monitor the training of the model in order to prevent overfitting of the data). Final models were obtained using the optimization feature of `neural connection', in which the optimal model was the one in which the root mean square error was lowest in the validation sample. Once optimal neural models were obtained, the weights within the neural net were stored in a matrix for further processing. In order to evaluate the influence of specific components of the input vectors, the matrices were degraded selectively by reducing specific connection weights by 99%. This is sometimes referred to as `lesioning' the net, a method used by Cardoso and Parks in a neural network model of executive function (Cardoso and Parks, 1998). The output of the degraded, or lesioned, network was then compared with the output from the full model in order to evaluate the influence of the degraded segment of the input vector on the final output.
|
Results |
|---|
|
TopAbstractIntroductionMethodsResultsDiscussionReferences |
|---|
Correlational analysis
Table 3
presents the Pearson correlations between left and right 1H-MRS hippocampal ratios (Figs 1 and 2 depict these relationships graphically). The left ratio correlated significantly with both measures of semantic memory. The left ratio also correlated significantly with LM%. The CVLT approached, but did not exceed, minimal statistical significance in its correlation with the left ratio. None of the measures correlated significantly with the right hippocampal ratio.
|
|
|
Neural network analysis
BNT model
RBF optimally modelled BNT performance using 1H-MRS hippocampal ratios. The optimal non-linear transfer function in the final model was radial spline, represented mathematically as
(d = distance from the data point to the centre of a hidden node). The optimal model settled on five hidden nodes. In order to evaluate the input from the left and right hippocampal ratios separately, the neural net was selectively degraded mathematically. Table 4 presents the multivariate correlations and explained variance for the full and degraded models.
|
In the validation sample, the non-degraded neural model explained 36% of the variance in BNT performance (r = 0.60, P < 0.05). When the left hippocampal input was lesioned mathematically, the multivariate correlation moved to non-significance (r = 0.52, P > 0.05). However, the multivariate correlation returned to statistical significance when the input from the right hippocampus was lesioned mathematically (r = 0.70, P < 0.05). Furthermore, explained variance in BNT performance increased to 49%.
Figure 3 presents the topographical depiction of the optimal BNT neural model. According to this model, BNT performance is mediated almost exclusively by the left hippocampus, with higher 1H-MRS left hippocampal ratios associated with lower BNT scores and lower left hippocampal ratios associated with higher BNT scores. However, the right hippocampus appears to provide a mild protective effect in its lower, healthier range.
|
1H-MRS model
MLP optimally modelled left 1H-MRS hippocampal ratios using measures of episodic and semantic memory. The best model (based on error rates in the validation sample) utilized three hidden nodes, each of which processed information using a tanh transfer function. Table 4
presents the multivariate correlations and explained variance for the full and degraded models. The full, optimal model was able to explain 60.8% of the variance in 1H-MRS left hippocampal ratios (P < 0.05). Mathematical lesioning of input from the semantic memory tests reduced explained variance to 3.6%. Degradation of episodic input had no apparent effect on the model.
Figure 4 presents a topographical depiction of 1H-MRS left hippocampal ratios as a function of the semantic memory input (BNT, COMP) from the optimal model. The region of lowest 1H-MRS ratios was associated with high scores on both semantic measures, particularly BNT. The region of highest 1H-MRS ratios was associated clearly with poor performances on both semantic memory measures.
|
|
Discussion |
|---|
|
TopAbstractIntroductionMethodsResultsDiscussionReferences |
|---|
The present study attempted to explicate the relationship between visual confrontation naming and hippocampal function, addressing some of the inconsistencies in prior research. We studied this relationship directly by modelling performance on a naming measure using quantitative indicators of hippocampal function. We then studied this relationship within the context of current psychobiological memory theory, modelling quantitative indicators of left hippocampal function with measures of episodic and semantic memory (including our measure of naming). This study had several advantages. First, we used 1H-MRS to operationalize the metabolic function of the hippocampi. Recent research has shown that 1H-MRS is particularly sensitive to neurocognitive function (Gadian et al., 1996
). A second advantage of the present study is the inclusion of patients with both left and right MTLE. Prior studies on naming and hippocampal function have limited the analysis to patients with left MTLE, which diminishes the range of values for the hippocampus, regardless of the technique. Finally, this study utilized neural networks, a multivariate statistical technique that seems well suited for studies of brain–behaviour relationships.
The majority of evidence from prior studies of naming and hippocampal pathology in MTLE suggests that there is a significant relationship between the two (see Table 1). Results from the present study support this conclusion. In our series of 46 patients with MTLE and only mesial temporal sclerosis (i.e. no other pathology), baseline performance on the BNT correlated significantly with left 1H-MRS hippocampal ratios. An optimal neural network model was able to explain 36% of the variance in BNT performance on the basis of left and right 1H-MRS hippocampal ratios. This rose to 49% when input was limited to only the left hippocampus. Both of these models were statistically significant. Figure 3 clearly depicts the striking relationship between the metabolic function of the left hippocampus and performance on the BNT. When considered as a measure of semantic memory, BNT (with COMP) was able to model left 1H-MRS hippocampal ratios just as well with as without input from measures of episodic memory. Taken together, these results clearly position the left hippocampus as an important anatomical substrate in the neural network of visual confrontation naming.
Clinical implications
This finding has potentially important clinical implications for patients with surgically remediable MTLE. Perhaps most important is the incorporation of naming into the hippocampal functional adequacy model (Chelune, 1995). The functional adequacy model was proposed initially as a model to predict episodic memory decline following anterior temporal lobectomy. Specifically, a functionally adequate dominant left hippocampus is a risk factor for significant decline in episodic memory following a left anterior temporal lobectomy. This is very similar to findings from other studies (Davies et al., 1998; Seidenberg et al., 1998) in which absence of dominant left mesial temporal sclerosis was a risk factor for significant decline in confrontation naming. Based on our findings, it would appear that this model could be extended even further to include the prediction of visual confrontation naming. Thus, confrontation naming could be predicted (at least in part) by the functional adequacy of the speech-dominant hippocampus. By including confrontation naming in the functional adequacy model, one would also predict that a preoperative, functionally adequate left hippocampus is a risk factor for a significant decline in naming following a left anterior temporal lobectomy. These recent findings (Davies et al., 1998; Seidenberg et al., 1998) support this hypothesis.
A second clinical implication involves the issue of intra- or extra-operative cortical language mapping. One recent study was unable to find a relationship between the extent of lateral temporal resection and change in naming ability following surgery (Davies et al., 1998). However, this same study found a significant relationship between hippocampal pathology and naming (see Table 1). The findings of our study, combined with those of Davies and colleagues (Davies et al., 1998) (see Table 1), suggest that cortical mapping may be an unnecessary procedure in sparing naming. However, it should be noted that our study did not include a measure of lateral temporal metabolic function. Based on the wide distribution of anatomical sites found to be associated with naming in previous studies, it is probable that both the speech-dominant hippocampus and neighbouring temporal neocortex mediate visual confrontation naming. Future studies including measures of both anatomical sites would address this issue more definitively.
Theoretical implications
The results of our study also have implications for competing psychobiological theories of memory function. The notion of discrete anatomical memory systems is encompassed by a broader model of memory referred to as the `stages of processing' model (SOP) (MacKay et al., 1998). The SOP model postulates that memory is organized into discrete anatomical units in the brain. However, recent research has begun to cast doubt on the SOP model, and particularly on the distinction between separate anatomical episodic and semantic memory systems (Horner, 1990; MacKay et al., 1998). A newer theory of memory, called the parallel distributed processing model (PDP), emphasizes the pattern of connections between structures of the mesial temporal lobe and associated neocortical sites. The PDP model is based largely on work in artificial intelligence and computational neuroscience where cognition is studied through connectionist models. Common among these models are the characteristics of parallel processing, distributed storage of knowledge and response to the environment (i.e. learning).
The finding that visual confrontation naming is related to the functional adequacy of the left hippocampus weakens the theory that there are distinct anatomical sites for episodic and semantic memory. Indeed, input from semantic memory measures seemed to be much more important than input from episodic memory measures in modelling left 1H-MRS hippocampal ratios (see Table 4 and Fig. 4). This result was somewhat surprising. Even under the PDP model, one would expect the input from the episodic memory measures to at least equal the input from the semantic memory measures in its ability to model left 1H-MRS hippocampal ratios. Sample size probably explains at least part of this finding. While our overall sample size was adequate, findings were based only on results in the validation sample, which represented only 25% of the total sample size. Such a small validation sample is particularly vulnerable to idiosyncratic sampling biases, even when they are chosen randomly. However, another probable explanation may be found in connectionist theories of memory function such as the PDP model.
PDP views the brain as a vast array of connections. Memories, therefore, are viewed as patterns of connections, where the strength of the memory depends on the strength and distribution of these connections. Neuroanatomically, the mesial temporal lobe and neighbouring neocortex are rich with connections, or information pathways (Beatty, 1995). Input can be received by the entorhinal cortex from a variety of neocortical association areas. The entorhinal cortex then feeds this input to the dentate gyrus via the perforant pathway, which in turn feeds input to the CA3 region of the hippocampus via connections known collectively as mossy fibres. Information continues to the CA1 region of the hippocampus via the Schaeffer collateral pathway, and returns to the entorhinal cortex through the subicular cortex. All of these connections are reciprocal (Squire and Zola-Morgan, 1991). Furthermore, all neurons in the hippocampus send input to other regions of the cerebral cortex (Beatty, 1995). PDP suggests that memories within this system are represented by different patterns within this array of interconnections. In this sense, the distinction between semantic and episodic memory could simply be in the strength and distribution of their connections within the temporal lobe memory system rather than separable anatomic memory sites. As stated by Mesulam (Mesulam, 1998),
`Memory consolidation appears to involve a gradual increase in the density of the matrix that binds the components of the memory to each other and to other aspects of mental content. The outcome is to increase the number of associative approaches through which the memory can be probed. The hippocampal–entorhinal cortex may well participate in the retrieval of all autobiographical and episodic memories, recent and remote, and even in the retrieval of the semantic knowledge related to arbitrary facts about the world' (p. 1027).
Thus, the PDP model suggests that semantic memories are represented by very strong and widely distributed connections within the temporal lobe memory system, whereas episodic memories are represented by much weaker and less widely distributed connections. In our admittedly limited neural network model of left 1H-MRS hippocampal metabolic ratios, input was represented only by performance on semantic and episodic memory measures. By degrading the input from the semantic memory measures, it could be said that we significantly diminished the density of our overall verbal memory `matrix'. Under this conceptualization, it is no longer surprising that mathematical degradation of the input from the semantic memory measures in our 1H-MRS model had such a significant effect on its efficacy. It is important to note that we are not concluding that episodic memory is unrelated to dominant left hippocampal function. The integrity of that relationship is well established, and was replicated in our correlation between LM% and the left 1H-MRS hippocampal ratio (see Table 3). However, it seems clear that the dominant left hippocampus is also strongly implicated in the retrieval of semantic memories, a finding that supports distribution rather than localization within the temporal lobe memory system.
Conclusions
This study has provided evidence suggesting that the speech-dominant hippocampus is a significant component of the overall biological neural network of visual confrontation naming. However, the `large-scale neurocognitive network' (Mesulam, 1990) of visual confrontation naming has yet to be examined in its entirety. Future studies should include a broader array of neuroanatomical input. Furthermore, larger sample sizes would allow for larger validation samples, which is a key component to multivariate modelling. Finally, it is our hope that investigators will continue to utilize techniques such as neural network analysis in moving beyond the assumptions of linearity that limit studies of complex brain–behaviour relationships.
|
Acknowledgments |
|---|
This study was supported in part by grants from the Epilepsy Foundation of America (principal investigator S.M.S.) and the NIH (principal investigator R.I.K., RO1 NS-33919).
|
References |
|---|
|
TopAbstractIntroductionMethodsResultsDiscussionReferences |
|---|
Baxendale SA, Van Paesschen W, Thompson PJ, Duncan JS, Harkness WF, Shorvon SD. Hippocampal cell loss and gliosis: relationship to preoperative and postoperative memory function. Neuropsychiatry Neuropsychol Behav Neurol 1998; 11: 12–21.[Web of Science][Medline]
Beatty J. Principles of behavioral neuroscience. Chicago: Brown & Benchmark; 1995.
Cardoso J, Parks RW. Neural network modeling of executive functioning with the Tower of Hanoi Test in frontal lobe-lesioned patients. In: Parks RW, Levine DS, Long DL, editors. Fundamentals of neural network modeling: neuropsychology and cognitive neuroscience. Cambridge (MA): MIT Press; 1998. p. 209–31.
Chelune GJ. Hippocampal adequacy versus functional reserve: predicting memory functions following temporal lobectomy. Arch Clin Neuropsychol 1995; 10; 413–32.[Web of Science][Medline]
Chelune GJ, Naugle RI, Luders H, Awad AI. Prediction of cognitive change as a function of preoperative ability status among temporal lobectomy patients at 6-month follow-up. Neurology 1991; 41: 399–404.
Coggeshall RE, Lekan HA. Methods for determining numbers of cells and synapses: a case for more uniform standards of review. [Review]. J Comp Neurol 1996; 364: 6–15.[Web of Science][Medline]
Davies KG, Bell BD, Bush AJ, Hermann BP, Dohan FC Jr, Jaap AS. Naming decline after left anterior temporal lobectomy correlates with pathological status of resected hippocampus. Epilepsia 1998; 39: 407–19.[Web of Science][Medline]
Delis DC, Kramer JH, Kaplan E, Ober BA. California Verbal Learning Test (research edition) manual. San Antonio (TX): Psychological Corporation; 1987.
Gadian DG, Isaacs EB, Cross JH, Connelly A, Jackson GD, King MD, et al. Lateralization of brain function in childhood revealed by magnetic resonance spectroscopy. Neurology 1996; 46: 974–7.
Hasselmo ME, Wyble BP, Stern CE. A model of human memory based on the cellular physiology of the hippocampal formation. In: Parks RW, Levine DS, Long DL, editors. Fundamentals of neural network modeling: neuropsychology and cognitive neuroscience. Cambridge (MA): MIT Press; 1998. p. 299–330.
Hermann BP, Perrine K, Chelune GJ, Barr W, Loring DW, Strauss E, et al. Visual confrontation naming following left anterior temporal lobectomy: a comparison of surgical approaches. Neuropsychology 1999; 13: 3–9.[Web of Science][Medline]
Hertz J, Krogh A, Palmer RG. Introduction to the theory of neural computation. Redwood City (CA): Addison-Wesley; 1991.
Hetherington H, Mason G, Pan J. Evaluation of cerebral gray and white matter metabolite differences by SI and 4.1T. Magn Resonance Med 1994a; 32: 565–71.
Hetherington H, Pan J, Mason G. 2D 1H SI of the human brain at 4.1T. Magn Resonance Med 1994b; 32: 530–4.
Horner MD. Psychobiological evidence for the distinction between episodic and semantic memory. [Review]. Neuropsychol Rev 1990; 1: 281–321.[Medline]
Hugg JW, Kuzniecky RI, Gilliam FG, Morawetz RB, Faught RE, Hetherington HB. Normalization of contralateral metabolic function following temporal lobectomy demonstrated by 1H magnetic resonance spectroscopic imaging. Ann Neurol 1996; 40: 236–9.[Web of Science][Medline]
Kaplan E, Goodglass H, Weintraub S. Boston Naming Test. Philadelphia: Lea and Febiger; 1978.
Kiyosawa M, Inoue C, Kawasaki T, Tokoro T, Ishii K, Ohyama M, et al. Functional neuroanatomy of visual object naming: a PET study. Graefes Arch Clin Exp Ophthalmol 1996; 234: 110–5.[Web of Science][Medline]
Kuzniecky R, Bilir E, Gilliam F, Faught E, Palmer C, Morawetz R, et al. Multimodality MRI in mesial temporal sclerosis: relative sensitivity and specificity. Neurology 1997; 49: 774–8.
Kuzniecky R, Hugg JW, Hetherington H, Butterworth E, Bilir E, Faught E, et al. Relative utility of 1H spectroscopic imaging and hippocampal volumetry in the lateralization of mesial tempoal lobe epilepsy. Neurology 51: 66–71.
Lowe D. Radial basis function networks. In: Arbib MA, editor. The handbook of brain theory and neural networks. Cambridge (MA): MIT Press; 1995. p. 779–82.
MacKay DG, Stewart R, Burke DM. H.M. revisited: relations between language comprehension, memory, and the hippocampal system. J Cogn Neurosci 1998; 10: 377–94.[Web of Science][Medline]
Martin RC, Sawrie SM, Hugg J, Gilliam F, Faught E, Kuzniecky R. Cognitive correlates of 1H spectroscopy detected hippocampal abnormalities in mesial temporal lobe epilepsy. Neurology 1999; 53: 2052–8.
Mesulam M-M. A cortical network for directed attention and unilateral neglect. [Review]. Ann Neurol 1981; 10: 309–25.[Web of Science][Medline]
Mesulam M-M. Large-scale neurocognitive networks and distributed processing for attention, language and memory. [Review]. Ann Neurol 1990; 28: 597–613.[Web of Science][Medline]
Mesulam M. From sensation to cognition. Brain 1998; 121: 1013–52.
Ojemann GA, Sutherling WW, Lesser RP, Dinner DS, Jayakar P, Saint-Hilaire JM. Cortical stimulation. In: Engel J Jr, editor. Surgical treatment of the epilepsies. 2nd edn. New York: Raven Press; 1993. p. 399–414.
Pan JW, Vaughan JT, Kuzniecky RI, Pohost GM, Hetherington HP. High resolution neuroimaging at 4.1T. Magn Resonance Imag 1995; 13: 915–21.
Sass KJ, Sass A, Weserveld M, Lecncz T, Novelly RA, Kim JH, et al. Specificity in the correlation of verbal memory and hippocampal neuron loss: dissociation of memory, language, and verbal intellectual ability. J Clin Exp Neuropsychol 1992; 14: 662–72.[Web of Science][Medline]
Seidenberg MS, McClelland JL. A distributed, developmental model of word recognition and naming. Psychol Rev 1989; 96: 523–68.[Web of Science][Medline]
Seidenberg M, Hermann B, Wyler AR, Davies K, Dohan FC Jr, Leveroni C. Neuropsychological outcome following anterior temporal lobectomy in patients with and without the syndrome of mesial temporal lobe epilepsy. Neuropsychology 1998; 12: 303–16.[Web of Science][Medline]
Shear PK, Sullivan EV, Marsh L, Morrell MJ, Lim KO, Pfefferbaum A. Hippocampal volumes correlate with nonmnemonic abilities in temporal lobe epilepsy [abstract]. J Int Neuropsychol Soc 1997; 3: 36.
Snyder PJ, Barr WB, Schaul N, Lieberman JA. Quantitative MRI correlates of naming disorders in temporal lobe epilepsy [abstract]. Epilepsia 1994; 35 Suppl 8: 102.
Squire LR, Zola-Morgan S. The medial temporal lobe memory system. [Review]. Science 1991; 253: 1380–6.
Tulving E. Episodic and semantic memory. In: Tulving E, Donaldson W, editors. Organization of memory. New York: Academic Press; 1972. p. 381–403.
Tulving E. Precis of elements of episodic memory. Behav Brain Sci 1984; 7: 223–68.
Tulving E. How many memory systems are there? Am Psychol 1985; 40: 385–98.
Van Paesschen W, Revesz T, Duncan JS, King MD, Connelly A. Quantitative neuropathology and quantitative magnetic resonance imaging of the hippocampus in temporal lobe epilepsy. Ann Neurol 1997; 42: 756–66.[Web of Science][Medline]
Vaughan J, Hetherington H, Otu J, Pan J, Pohost G. High frequency volume coils for clinical NMR imaging and spectroscopy. Magn Resonance Med 1994; 32: 206–18.
Wechsler D. A standardized memory scale for clinical use. J Psychol 1945; 19: 87–95.[Web of Science]
Wechsler D. Wechsler Adult Intelligence Scale—revised WAIS-R manual. New York: Psychological Corporation; 1981.
Westerveld M, Sass KJ, Bell J, Spencer D, Kim J. Effects of hippocampal pathology on language outcome following temporal lobectomy [abstract]. Epilepsia 1996; 37 Suppl 5: 136.
Wiggs CL, Weisberg J, Martin A. Neural correlates of semantic and episodic memory retrieval. Neuropsychologia 1999; 37: 103–18.[Web of Science][Medline]
Zola-Morgan S, Squire LR. Neuroanatomy of memory. [Review]. Annu Rev Neurosci 1993; 16: 547–63.[Web of Science][Medline]
Received April 26, 1999. Revised October 5, 1999. Accepted October 18, 1999.
CiteULike 
Connotea 
Del.icio.us What's this?
This article has been cited by other articles:
|
|
 |
|
  A. T.-D. Fonseca, E. Guedj, F-X. Alario, V. Laguitton, O. Mundler, P. Chauvel, and C. Liegeois-Chauvel Brain regions underlying word finding difficulties in temporal lobe epilepsy Brain, October 1, 2009; 132(10): 2772 - 2784. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 K. A. Josephs, J. L. Whitwell, D. S. Knopman, W. T. Hu, D. A. Stroh, M. Baker, R. Rademakers, B. F. Boeve, J. E. Parisi, G. E. Smith, et al. Abnormal TDP-43 immunoreactivity in AD modifies clinicopathologic and radiologic phenotype Neurology, May 6, 2008; 70(19_Part_2): 1850 - 1857. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
M. J. Hamberger, W. T. Seidel, R. R. Goodman, A. Williams, K. Perrine, O. Devinsky, and G. M. McKhann II Evidence for cortical reorganization of language in patients with hippocampal sclerosis Brain, November 1, 2007; 130(11): 2942 - 2950. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 L Bartha, T Benke, G Bauer, and E Trinka Interictal language functions in temporal lobe epilepsy J. Neurol. Neurosurg. Psychiatry, June 1, 2005; 76(6): 808 - 814. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 J.-F. Demonet, G. Thierry, and D. Cardebat Renewal of the Neurophysiology of Language: Functional Neuroimaging Physiol Rev, January 1, 2005; 85(1): 49 - 95. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 N. Kapur and M. D Kopelman Advanced brain imaging procedures and human memory disorder Br. Med. Bull., March 1, 2003; 65(1): 61 - 81. [Abstract] [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||