Using multivariate data reduction to predict ...

YEBEH-03625; No. of pages: 8; 4C: 2, 4, 6 Epilepsy & Behavior xxx (2013) xxx–xxx

Contents lists available at ScienceDirect

Epilepsy & Behavior journal homepage: www.elsevier.com/locate/yebeh

Using multivariate data reduction to predict postsurgery memory decline in patients with mesial temporal lobe epilepsy Marie St-Laurent a,c,d,e, Cornelia McCormick b,d,e, Mélanie Cohn c,d,e, Bratislav Mišić a,c, Irene Giannoylis d,e, Mary Pat McAndrews b,c,d,e,⁎ a

Rotman Research Institute at Baycrest, Toronto, Ontario, Canada Institute of Medical Sciences, University of Toronto, Toronto, Ontario, Canada Department of Psychology, University of Toronto, Toronto, Ontario, Canada d Krembil Neuroscience Centre, University Health Network, Toronto, Ontario, Canada e Toronto Western Research Institute, University Health Network, Toronto, Ontario, Canada b c

a r t i c l e

i n f o

Article history: Received 10 July 2013 Revised 29 August 2013 Accepted 29 September 2013 Available online xxxx Keywords: Neuropsychology Temporal lobe excision Surgical cognitive outcome Principal component analysis Bootstrapping Cross-validation

a b s t r a c t Predicting postsurgery memory decline is crucial to clinical decision-making for individuals with mesial temporal lobe epilepsy (mTLE) who are candidates for temporal lobe excisions. Extensive neuropsychological testing is critical to assess risk, but the numerous test scores it produces can make deriving a formal prediction of cognitive change quite complex. In order to benefit from the information contained in comprehensive memory assessment, we used principal component analysis (PCA) to simplify neuropsychological test scores (presurgical and pre- to postsurgical change) obtained from a cohort of 56 patients with mTLE into a few easily interpretable latent components. We next performed discriminant analyses using presurgery latent components to categorize seizure laterality and then regression analyses to assess how well presurgery latent components could predict postsurgery memory decline. Finally, we validated the predictive power of these regression models in an independent sample of 18 patients with mTLE. Principal component analysis identified three significant latent components that reflected IQ, verbal memory, and visuospatial memory, respectively. Together, the presurgery verbal and visuospatial memory components classified 80% of patients with mTLE correctly according to their seizure laterality. Furthermore, the presurgery verbal memory component predicted postsurgery verbal memory decline, while the presurgery visuospatial memory component predicted visuospatial memory decline. These regression models also predicted postsurgery memory decline successfully in the independent cohort of patients with mTLE. Our results demonstrate the value of data reduction techniques in identifying cognitive metrics that can characterize laterality of damage and risk of postoperative decline. © 2013 Elsevier Inc. All rights reserved.

1. Introduction Comprehensive neuropsychological assessment using a large number of standardized tests is typically undertaken in individuals with intractable epilepsy who are candidates for a resection of their seizure focus. These tests are used to establish a patient's clinical profile and to assess risks of postsurgical cognitive decline [1]. In patients with mesial temporal lobe epilepsy (mTLE), many of these tests focus on memory performance, which is disrupted by temporal lobe dysfunction [1,2]. When epilepsy is unilateral, left-lateralized mTLE (L-mTLE) tends to disrupt verbal memory function, and left-lateralized temporal lobe resections can lead to a decrease in verbal memory function [3–5]. On the other hand, right-lateralized mTLE (R-mTLE) can affect visual and

⁎ Corresponding author at: Neuropsychology Clinic, 4F-409, Toronto Western Hospital, University Health Network, 399 Bathurst Street, Toronto, Ontario M5T 2S8, Canada. Fax: +1 416 603 5321. E-mail address: [email protected] (M.P. McAndrews).

visuospatial memory, although the consistency with which performance is impacted by surgery is somewhat weaker [6–9]. Recently, much effort has been dedicated to the identification of reliable predictors of postsurgery memory decline in patients with mTLE. Among such predictors are patients' age at surgery and age at onset [10,11] and measures of medial temporal structural and functional integrity [12–15]. Baseline memory performance has also been identified as an important marker of risk for postsurgery memory decline: good presurgery verbal memory performance has been shown to predict significant verbal memory decline postresection in patients with L-mTLE [11,13,15–22], while individuals with R-mTLE with good baseline visuospatial memory have been shown to be at a greater risk of postsurgery visuospatial memory decline [3,12,13,23,24]. Many studies linking memory performance pre- and postsurgery typically rely on a single test score or on a few selected test scores assessed in separate analyses (for an exception, see [25]). Individual test scores can be influenced by measurement errors and by idiosyncratic factors that are unrelated to a patient's pathology. Also, parallel analyses conducted on test scores meant to capture similar cognitive functions

1525-5050/$ – see front matter © 2013 Elsevier Inc. All rights reserved. http://dx.doi.org/10.1016/j.yebeh.2013.09.043

Please cite this article as: St-Laurent M, et al, Using multivariate data reduction to predict postsurgery memory decline in patients with mesial temporal lobe epilepsy, Epilepsy Behav (2013), http://dx.doi.org/10.1016/j.yebeh.2013.09.043

2

M. St-Laurent et al. / Epilepsy & Behavior xxx (2013) xxx–xxx

(e.g., verbal memory) can lead to conflicting results that are difficult to interpret. Since neuropsychological assessments typically generate a rich collection of test scores, some of which clearly tap the same functional domains, data reduction techniques such as principal component analysis can be used to derive a smaller number of latent components that reflect these domains. Such components are more resilient to noisy fluctuations in individual test performance [26] and can be used as robust indices of memory in mathematical models predicting postsurgical change in cognitive performance. With the current analysis, we used intelligence, verbal memory, and visuospatial memory neuropsychological test scores obtained from patients with left or right mTLE tested before and after they underwent a unilateral temporal lobe resection. We applied principal component analysis (PCA) to reduce these scores into a smaller set of latent components. We then used these components to discriminate between patients with L-mTLE and R-mTLE and to predict postsurgery changes in memory performance, both in the original cohort from which the components were derived and in an independent cohort of patients with mTLE. By cross-validating our model on an independent patient sample, we provide a rigorous assessment of whether the predictive power of the model can generalize to the general population of patients with unilateral mTLE identified as surgery candidates. Our goal was to assess whether we could derive simple yet powerful indices of verbal and visuospatial memory that could be used in both clinical and research settings.

2. Methods 2.1. Subjects Neuropsychological, clinical, and demographic data were obtained from 56 patients with unilateral mTLE (28 L-mTLE, 28 R-mTLE) who had received epilepsy surgery at the Toronto Western Hospital between 1996 and 2011 (see Table 1). All patients provided informed consent for their anonymized data to be used for research purposes, according to a series of protocols approved by the University Health Network's Research

Ethics Board. Forty-three patients underwent standard, en bloc anterior temporal lobe resection including surgical removal of the temporal pole, the amygdala, the anterior portion of the hippocampus, the rhinal and lateral temporal cortex, and portions of the parahippocampal cortex, while 13 patients underwent selective amygdalohippocampectomy including surgical removal of the amygdala, the anterior portion of the hippocampus, the rhinal cortex, and portions of the parahippocampal cortex. Patients had undergone a full neuropsychological assessment both preoperatively and postoperatively. The median interval between presurgical assessment and surgery was 14 months (1st quartile = 9 months; 3rd quartile = 21 months), and the median time interval between surgery and the postsurgical assessment was 9 months (1st quartile = 7 months; 3rd quartile = 12 months). All patients had a full scale intellectual quotient score greater than 70 on the Wechsler Abbreviated Scale of Intelligence (WASI; [27]) and had levels of English fluency appropriate for testing. Additional clinical information for this patient cohort and demographic and neuropsychological data for the cross-validation patient sample (see Table S1) are available as online supporting information.

2.2. Neuropsychological tests Mean scores for the tests selected for this analysis are presented in Table 1. Intelligence indices included the verbal (VIQ) and performance (PIQ) intellectual quotient from the WASI [27]. Verbal memory test scores included the raw score on the Warrington's Word Recognition Test (WWord; [28]) and two scores from the Rey Auditory Verbal Learning Test [29]: the total learning score (over 5 learning trials; RAVLTtot) and percent retained score (RAVLT%), which corresponded to the percentage of words recalled on the 5th learning trial that were retained after a 20-minute delay. Visuospatial memory test scores included the raw score on the Warrington's Face Recognition Test (WFace; [28]), the total learning score from the Rey Visual Design Learning Test (RVDLT; [30]), and the total number of trials to reach learning criteria from the Spatial Conditional Associative Learning task (CAL). Although the latter is not a standardized test, it has long been

Table 1 Demographic data and mean neuropsychological test scores per patient group.

L–mTLE (n = 28)

R–mTLE (n = 28)

Gender

16F/12M

20F/8M

Years of education

13.8 (2.5)

13.4 (2.6)

Type of surgery

6 selective/22 standard

7 selective/21 standard

Handedness

2 left/25 right/1 mixed

4 left/24 right

Hemi. language dominance

28 left/0 right

28 left/0 right

Age of seizure onset (years)

12.3 (12.2)

17.0 (15.6)

Post–Sx

Diff. Score

Age at test (years)

Pre–Sx 37.5 (9.5)

39.9 (9.8)

n/a

38.4 (10.9)

Pre–Sx

40.7 (10.8)

Post–Sx

Diff. Score

Duration of epilepsy (years)

25.3 (16.1)

26.6 (16.2)

n/a

21.4 (12.0)

22.6 (12.1)

VIQ

96.1 (11.1)

95.3 (10.6)

–0.8 (6.0)

101.6 (11.3)

102.5 (9.9)

0.9 (7.8)

PIQ

100.4 (11.6)

105.3 (10.8)

4.9 (8.0)

96.8 (13.7)

100.4 (16.6)

3.7 (8.3) 3.1 (7.9)

n/a n/a

RAVLTtot

44.3 (7.4)

41.2 (7.7)

–3.1 (9.7)

46.1 (6.6)

49.2 (8.9)

RAVLT%

59.1 (20.2)

49.8 (25.6)

–9.3 (25.4)

80.9 (17.1)

82.2 (13.9)

WWord

45.4 (3.2)

41.7 (4.6)

–3.7 (5.2)

47.5 (1.9)

46.7 (4.2)

–0.9 (4.4)

RVDLT

36.3 (11.0)

33.1 (11.8)

–3.3 (7.4)

35.5 (11.9)

32.8 (11.3)

–2.7 (8.6)

CAL

44.1 (20.6)

45.5 (19.1)

1.4 (29.5)

50.3 (18.7)

51.3 (21.6)

WFace

40.4 (4.7)

41.3 (4.1)

0.9 (2.5)

40.1 (5.3)

37.4 (6.6)

1.3 (22.8)

1 (24.7) –2.8 (5.3)

Note: Standard deviation is between parentheses. Duration of epilepsy at postsurgery testing corresponds to the time lapse between age at seizure onset and age at surgery. Significant differences in test scores between the group with right and the group with left mTLE are shown in bold (p b .05 based on 2-tailed independent sample t-tests). Diff. Score = difference between postsurgery and presurgery test score. F = female, Hemi. = hemispheric, M = male, Sx = surgery.

Please cite this article as: St-Laurent M, et al, Using multivariate data reduction to predict postsurgery memory decline in patients with mesial temporal lobe epilepsy, Epilepsy Behav (2013), http://dx.doi.org/10.1016/j.yebeh.2013.09.043

M. St-Laurent et al. / Epilepsy & Behavior xxx (2013) xxx–xxx

used in our center as a test of spatial learning that was adapted from an experimental task used in patient populations with epilepsy and Parkinson's disease [31,32].

2.3. Statistical analysis 2.3.1. Identifying latent cognitive components Change test scores were calculated by subtracting a participant's presurgery test score from his or her postsurgery test score (a negative score indicated a decrease in performance, while a positive score indicated improvement on all measures except the CAL). Presurgery and change test scores from the different tests in Table 1 were converted into Z-scores based on the full patient distribution (n = 56). The CAL was the only measure for which a high score reflected poorer performance within the dataset. To facilitate interpretation, the signs of the CAL's presurgery and change Z-scores were flipped (Z-scores were multiplied by −1). We then performed separate principal component analyses (PCAs) on presurgical and change Z-scores using SPSS 19 (IBM SPSS Statistics). A varimax orthogonal rotation with Keiser normalization was performed on each solution to facilitate the interpretation of the results while maintaining orthogonality among the different components. Components meeting the Keiser criterion (eigenvaluesN1) were considered significant. For both the presurgery and change PCA analyses, we identified three significant components that reflected intelligence (IQ component), verbal memory (VM component), and visuospatial memory (VSM component), respectively (see the Results section). With N = 56 observations and p = eight variables, our case-tovariable ratio is 7:1. This ratio is sufficiently large to meet some [33–35] but not all [36,37] of the guidelines found in the psychometric literature. However, stable and reproducible results can be obtained with small samples [38], and good recovery of component structure largely depends on the data themselves [39]. Resampling techniques can determine the stability of an underlying component structure, and so we cross-validated the PCA results with a bootstrapping procedure to diagnose whether our results were stable and our patient sample sufficiently large. We performed two bootstrap resampling analyses (one with presurgery and one with change test scores) using MATLAB 7.12 (R2011a, MathWorks).1 One hundred thousand (100 000) bootstrapped PCAs were performed on 56 participants sampled with replacement from the original patient group: a subject was replaced in the sample immediately after selection so that each PCA was performed on different combinations of individuals sampled from the group of 56 subjects. Procrustes rotations were performed to align each bootstrapped PCA solution with the original PCA solution. For each significant component identified by PCA, 95% confidence intervals for the original solution's loadings and coefficients were derived from the distribution of loadings and coefficients obtained from the realigned bootstrapped PCA solutions. We found that the different component structures we report in the results were stable, as evidenced by tight confidence intervals around variables with high loadings. High loadings, and confidence intervals that did not cross the x-axis (see Figs. S2 and S3), indicated that test scores correlated reliably and significantly with a component. As the component structure from the original dataset could be consistently recovered across many surrogate (bootstrapped) datasets, our sample size was adequate to conduct PCA.

1 We used MATLAB because we could not perform the Procrustes rotation essential to the bootstrapping analysis in SPSS. We recreated the original SPSS solution in MATLAB, to which we aligned the 100000 bootstrapped solutions. The MATLAB and the SPSS solutions were nearly identical, with very subtle differences in loading and coefficient values, most likely due to differences in normalization algorithms. The SPSS solution is reported in the Results section, and the MATLAB solution is reported in the online supporting information (Figs. S2 and S3).

3

2.3.2. Laterality analysis To determine how well the presurgery components identified with PCA could predict mTLE laterality, we performed discriminant analyses using MYSTAT 12 (Systat Software, Inc.) using the PCA components identified in the presurgery data. We expected latent components that reflected verbal and visuospatial memory performance to be the most informative regarding seizure lateralization. 2.3.3. Predicting postsurgery memory decline To assess whether presurgery memory performance could predict material-specific postsurgery memory decline, we performed multiple regression analyses using the presurgery verbal (VM) and visuospatial (VSM) components to predict the change in VM and VSM components. We performed separate analyses on patients with right and patients with left mTLE as this approach might reflect best the clinical setting of predicting postsurgical memory decline in patients who have been already diagnosed with either left or right mTLE. Further, laterality influences the impact of mesial temporal lobe damage on cognition, and treating left and right temporal lobe resections as equivalent could obscure the relationship between presurgery performance and postsurgery decline. 2.3.4. Validating the predictive power of presurgery memory components To evaluate the generalizability of our regression models, we tested whether they could predict postsurgery change in an independent sample of patients with mTLE (R-mTLE: n = 12; L-mTLE: n = 6). New patients' presurgery and change test scores were converted into Z-scores based on the original 56 patients' distribution (CAL was flipped). Principal component analysis scores on the presurgery and change components described above were calculated for each component by summing the product of each test's Z-score and its corresponding coefficient (coefficients are listed in Table 2). Then, we used the regression models derived from the original group of 56 patients to estimate change PCA scores based on presurgery PCA scores in the new patients with mTLE (see supporting information for formulae). Finally, we performed regression analyses to predict obtained change PCA scores based on change PCA scores estimated with the regression models. Patients with left and patients with right mTLE were tested separately with different models. 3. Results 3.1. PCA analysis: identity and description of latent cognitive components Bartlett's sphericity test indicated that data reduction was appropriate for both the presurgery and the change data matrices (presurgery: χ2(28) = 89.746, p b .001; change: χ2(28) = 44.054, p = .027). The PCA on presurgery test scores revealed three significant and stable orthogonal components (eigenvalues N 1) whose rotated loadings and coefficients are presented in the top panel of Table 2 (see Fig. S1 for a visual display of PCA results and Fig. S2 for confidence intervals derived from the bootstrapping analysis). In a PCA solution for which components are orthogonal, loadings correspond to standardized correlation coefficients between the test and the component, while coefficients (or score coefficients) are unstandardized (analogous to β and B coefficients in a regression analysis, respectively). All three verbal memory test scores, RAVLTtot, RAVLT%, and WWord, loaded reliably on the first component (Presurgery VM component). Principal component analysis scores for this component were significantly lower for patients with L-mTLE than for patients with R-mTLE (Mann–Whitney U = 610.0, p b .001; Fig. 1, top left). All three visuospatial memory tests, WFace, CAL, and RVDLT, loaded reliably on the second component (Presurgery VSM component). Mean PCA scores for this component were numerically lower in the group with R-mTLE than in the group with L-mTLE, but this difference was not significant (Mann–Whitney U = 297.0, p = .12; Fig. 1, top middle). The two

Please cite this article as: St-Laurent M, et al, Using multivariate data reduction to predict postsurgery memory decline in patients with mesial temporal lobe epilepsy, Epilepsy Behav (2013), http://dx.doi.org/10.1016/j.yebeh.2013.09.043

4

M. St-Laurent et al. / Epilepsy & Behavior xxx (2013) xxx–xxx

Table 2 Rotated factor solutions for the principal component analyses performed on presurgery and change test scores.

PRE-SURGERY SCORES ANALYSIS Component 1 “Verbal memory” Factor's eigenvalue % variance explained Test scores VIQ PIQ RAVLTtot RAVLT% WWord RVDLT CAL WFace

Component 2 “Visuospatial memory”

Component 3 “Intelligence”

1.72

1.73

1.69

21.55

21.63

21.17

Loadings 0.12 –0.08 0.82 0.76 0.65 0.17 0.06 –0.02

Coefficients 0.07 –0.09 0.47 0.46 0.37 0.02 –0.04 –0.06

Loadings –0.01 0.34 0.13 –0.09 0.17 0.79 0.83 0.50

Coefficients –0.24 0.03 0.00 –0.07 –0.02 0.47 0.59 0.25

Loadings 0.85 0.79 0.06 –0.24 0.29 0.24 –0.06 0.37

Coefficients 0.59 0.46 0.00 –0.15 0.15 –0.05 –0.26 0.12

CHANGE SCORES ANALYSIS Component 1 “Verbal memory” Factor's eigenvalue % variance explained Test scores VIQ PIQ RAVLTtot RAVLT% WWord RVDLT CAL WFace

Component 2 “Visuospatial memory”

1.93 24.09 Loadings 0.19 0.00 0.75 0.70 0.70 0.29 –0.09 –0.50

Component 3 “Intelligence”

1.40 17.51 Coefficients 0.07 –0.02 0.38 0.36 0.37 0.12 –0.08 –0.29

Loadings 0.06 –0.10 0.20 0.01 –0.01 0.74 0.80 0.40

1.25 15.59 Coefficients –0.01 –0.12 0.11 –0.03 –0.03 0.53 0.58 0.28

Loadings 0.72 0.72 0.09 0.16 0.00 –0.06 0.00 0.42

Coefficients 0.57 0.59 0.04 0.10 –0.03 –0.11 –0.05 0.33

Note: Principal component analysis with varimax orthogonal rotation using Keiser normalization. The solution was forced to a three-component solution for the change scores analysis. Rotated components with eigenvalues N1 were considered significant. Values in red correspond to test scores for which the 95% confidence intervals for the bootstrapped loadings and coefficients were both reliable (did not cross 0; see Figs. S2 and S3).

Fig. 1. Distribution of PCA scores from the three significant presurgery (top) and change (bottom) latent components plotted for patients with right and patients with left mTLE, respectively. Whiskers correspond to the full range of the distribution or to 1.5 times the interquartile range, whichever is smaller. Principal component analysis scores from patients who underwent a selective and a standard anterior temporal lobe excision are identified by dark and pale circles, respectively. *p b .05.

Please cite this article as: St-Laurent M, et al, Using multivariate data reduction to predict postsurgery memory decline in patients with mesial temporal lobe epilepsy, Epilepsy Behav (2013), http://dx.doi.org/10.1016/j.yebeh.2013.09.043

M. St-Laurent et al. / Epilepsy & Behavior xxx (2013) xxx–xxx

intelligence tests, VIQ and PIQ, loaded highly on the third component (Presurgery IQ component). Principal component analysis scores did not differ between patients with R-mTLE and patients with L-mTLE (Mann–Whitney U = 431.0, p = .52; Fig. 1, top right). The PCA conducted on change scores originally revealed four significant components,2 with the two IQ scores (VIQ and PIQ) loading highly on two separate components. To facilitate interpretation, we forced the analysis to a three-component solution, which revealed three significant components analogous to the ones obtained with the presurgery analysis (eigenvalues N 1; see Table 2, bottom panel; also see Fig. S1 for a visual display of PCA results and Fig. S3 for bootstrapped confidence intervals). The verbal memory test scores loaded reliably on the first component (Change VM component). Principal component analysis scores for this component were significantly lower in patients with L-mTLE than in patients with R-mTLE (Mann–Whitney U = 591.0, p = .001; Fig. 1, bottom left), indicating steeper verbal memory decline after a left-lateralized resection. The visuospatial memory tests loaded reliably on the second component (Change VSM component), while the two intelligence tests loaded reliably on the third component (Change IQ component). Principal component analysis scores for these two components did not differ significantly between patients with RmTLE and patients with L-mTLE (Mann–Whitney U = 343.0, p = .42, Fig. 1, bottom middle; and Mann–Whitney U = 361.0, p = .61, Fig. 1, bottom right, respectively). Although the two IQ scores loaded highly on separate components in the unforced solution, both scores loaded highly and reliably on the same component in the three-component solution, indicating that the third (IQ) component reflected latent cognitive change that was common to both tests. Supporting Table S2 indicates what change PCA scores equate to in terms of change on individual neuropsychological test scores. As noted in Table 1, a minority of patients in our sample underwent a selective rather than a standard anterior temporal lobe excision. No comparison (parametric and nonparametric) between PCA scores from patients with selective and standard excisions reached significance (p N .05; data not shown), indicating no systematic difference in presurgery memory capacity or in postsurgery memory decline between patients who underwent the different surgery types. Obviously, our capacity to detect significant differences based on surgery type using statistical tests is limited by the small number of selective cases in our sample, but the overlap in the distribution of PCA scores between patients with selective and standard excisions can be assessed visually in Fig. 1. 3.2. Discriminant analysis: predicting seizure laterality Although only the VM components differed significantly between patients with left and patients with right mTLE, it is also important to know how well the presurgery latent components can determine, on their own and in combination, whether an individual is likely to have left or right mTLE. Patients' data were entered into PCAs without consideration for the laterality of their epilepsy. Here, we conducted a series of discriminant analyses to assess how well the three presurgery components (VM, VSM, and IQ) could discriminate between L-mTLE and R-mTLE. The most accurate classification (80%) was achieved when combining the VM and VSM components as predictors (Wilk's λ (2, 1, 54) = 0.72, approx. F(2, 53) = 10.307, p b .001), indicating that using both memory indices provided the best approximation of lateralization in our sample. This accuracy exceeded what was achieved by each component individually: the VM component classified 66% of cases accurately (Wilk's λ (1, 1, 54) = 0.77, approx. F(1, 54) = 15.94, p b .001), while the VSM component classified 61% of cases accurately 2 The first component loaded highly on the three verbal memory tests, and the second component loaded highly on the visuospatial memory tests. The third component loaded highly on VIQ as well as on the two Warrington tests (WWord and WFace), while the fourth component loaded highly uniquely on PIQ.

5

(Wilk's λ (1, 1, 54) = 0.95, approx. F(1, 54) = 2.97, p = .09). Adding the third component (IQ) as an additional predictor did not improve classification (77% accurate; Wilk's λ (3, 1, 54) = 0.710, approx. F(3, 52) = 7.08, p b .001), nor did IQ predict laterality on its own (55% accurate; Wilk's λ (1, 1, 54) = 0.99, approx. F(1, 54) = 0.54, p = .47), which confirms that verbal and visuospatial memory performance predicts seizure lateralization to a greater extent than IQ. Importantly, using the six normalized (Z-scored) memory test scores to predict laterality also did not improve classification accuracy (77% accurate; Wilk's λ (6, 1, 54)= 0.599, approx. F(6, 49)=5.458, p b .001), indicating that the information contained in individual test scores regarding lateralization is fully preserved in the latent memory components. 3.3. Regression analysis: predicting postsurgery memory decline To determine whether presurgery performance could predict postsurgical memory decline, we performed regression analyses on the change VM and VSM components using the presurgery VM and VSM components as predictors (see Fig. 2). In the group with L-mTLE, the change VM component was predicted by presurgery memory performance (F(2, 25) = 6.938, p = .004, adjusted R2 = 0.305). High presurgery VM scores predicted greater verbal memory decline (t(1, 25) = −2.862, p = .008, β = −.461), while high presurgery VSM scores predicted better verbal memory outcome (t(1, 25) = 2.616, p = .015, β = .421), possibly reflecting the right hemisphere's capacity to compensate for the resected left temporal lobe. In the group with R-mTLE, presurgery memory performance did not predict the change VM component (F(2, 25) = 2.383, p = .113, adjusted R2 = .093; presurgery VM predictor: t(1, 25) = −1.938, p = .064, β = −.363; presurgery VSM predictor: t(1, 25) = −0.597, p = .556, β = −.112), although there was a trend linking good baseline verbal memory to greater verbal memory decline (Fig. 2, top right). In the group with R-mTLE, the change VSM score was predicted by presurgery memory performance (F(2, 25) = 6.333, p = .006, adjusted R2 = .283). High presurgery VSM scores were a significant predictor of decline (t(1, 25) = −3.292, p = .003, β = −.547). Also, there was a trend for high presurgery VM scores to predict lesser decline (t(1, 25) = 1.984, p = .058, β = .330), possibly reflecting the left hemisphere's capacity to compensate for the resected right temporal lobe. In the group with L-mTLE, presurgery memory performance also predicted change VSM scores significantly (F(2, 25) = 3.781, p = .037, adjusted R2 = 0.171). Good presurgery VSM scores predicted greater risk of decline (t(1, 25) = −2.726, p = .012, β = −.479), while presurgery VM scores did not predict decline (t(1, 25) = −.134, p = .894, β = −.024). Overall, these results indicate that the presurgery memory components are good predictors of material-specific memory change postsurgery in mTLE. We also conducted regression analyses to which we added surgery type (selective versus standard, entered as a dummy variable) as a predictor. Surgery type was not a significant predictor of either visual or visuospatial memory decline in either the group with R-mTLE (verbal: t = 0.507, p = .617; visuospatial: t = 0.644, p = .526) or the group with L-mTLE (verbal: t = 1.153, p = .260; visuospatial: t = 0.772, p = .447). Also, adding surgery type never affected the significance (or trend toward significance) of other predictors in the different models (data not shown), indicating that it did not mediate the relationship between the presurgery and the change PCA components. 3.4. Validating the predictive power of presurgery memory components We ascertained the generalizability of our regression models' prediction power in a new cohort of patients with mTLE who were not selected to be matched to the original cohort on any specific criteria other than documented unilateral mTLE and left language dominance. In patients with L-mTLE (n = 6), estimated VM change scores predicted obtained VM change scores successfully (F(1, 5) = 47.54, p = 0.002, adjusted R2 = 0.9; Fig. 3, left). In patients with R-mTLE (n = 12),

Please cite this article as: St-Laurent M, et al, Using multivariate data reduction to predict postsurgery memory decline in patients with mesial temporal lobe epilepsy, Epilepsy Behav (2013), http://dx.doi.org/10.1016/j.yebeh.2013.09.043

6

M. St-Laurent et al. / Epilepsy & Behavior xxx (2013) xxx–xxx

Fig. 2. Pearson correlations between presurgery PCA scores and change PCA scores in patients with left and patients with right mTLE. Pre-sx = presurgery.

obtained VSM change scores were successfully predicted by estimated VSM change scores (F(1, 10) = 7.69, p = .019, adjusted R2 = 0.38; Fig. 3, right). Although these results must be interpreted with caution given that correlation coefficients can be inflated in small samples, they nonetheless validate our models' predictive power to identify surgery candidates most at risk of postsurgery memory decline in the general population of patients with mTLE. In Tables S3 and S4 (online supporting information), we report results from regression analyses conducted in our original patient sample, for which we used either individual presurgery test scores or presurgery VM and VSM PCA scores, to predict (1) change in individual test scores and (2) change in VM and VSM PCA scores. We find that individual test scores are somewhat better at predicting change on the very same test but that presurgery VM and VSM PCA scores are much better predictors of the latent change memory components. These results suggest that single test scores, while great predictors of test-specific change, lack the generalized predictive power of latent components.

4. Discussion The primary objectives of neuropsychological testing in decisionmaking about surgical candidacy for mTLE are (1) to ascertain whether the test profile is compatible with EEG and MRI evidence in indicating the focus of damage/dysfunction and (2) to predict the risk of postsurgical decline in memory following a planned temporal lobe resection. In an attempt to satisfy the second objective, neuropsychological assessments in epilepsy surgery centers typically include a large number of memory tests selected on the basis of their clinical utility in affording a composite of verbal and visuospatial memory function. Here, we have demonstrated how such data from our neuropsychological practice may best be exploited to enhance predictive capacity and, thus, enable us to best inform patients of potential risk. Using a relatively simple data reduction technique in a retrospective analysis of clinical test scores, we were able to derive verbal and visuospatial memory components that reliably discriminated between patients with left and patients

Fig. 3. Linear regression between estimated and obtained PCA scores in a new cohort of patients with L-mTLE and patients with R-mTLE.

Please cite this article as: St-Laurent M, et al, Using multivariate data reduction to predict postsurgery memory decline in patients with mesial temporal lobe epilepsy, Epilepsy Behav (2013), http://dx.doi.org/10.1016/j.yebeh.2013.09.043

M. St-Laurent et al. / Epilepsy & Behavior xxx (2013) xxx–xxx

with right mTLE and predicted material-specific memory change following temporal lobe excision. Further, the predictive model was applied to a new cohort of patients and was successful in estimating the degree of postoperative change in both patients with left and patients with right mTLE, despite a small sample of cases in each group. This provides a powerful validation of the utility of our strategy, which can be applied readily to specific measures used in other surgery centers. How does this strategy compare with others that have been used for prediction of postsurgical change? A definitive answer to that question proves more difficult than might be appreciated at first glance, as most of the relevant studies in the literature that examine the specific predictive/correlational value of neuropsychological test variables differ from ours in several ways. First, they typically include both patients with left and patients with right mTLE in the same predictive model with side of seizure focus as a variable, unlike the present analysis in which patients with left and patients with right mTLE were modeled separately. Second, many of them use reliable change indices as the discriminate variables for prediction [i.e., did the patient's score show a significant change or not?], rather than treating memory as a continuous variable as we did. Without arguing the relative merits of these approaches, we note that the predictive power of our model (i.e., amount of variance accounted for in memory change) compares favorably with findings from single or multiple neuropsychological tests considered alone or as significant predictors in combination with other clinical or imaging variables [3,12,20,23]. Our focus on these component scores as continuous variables is different from the approach of other studies using reliable change indices that provide guidelines for ascertaining whether the degree of memory change following surgery is clinically meaningful, that is, whether it exceeds what might be expected based on factors such as test–retest reliability and practice effects that influence change scores [40,41]. Our prediction equations also differ from standardized regression based change scores used to identify risk of memory impairment, as these additionally take into account factors such as regression to the mean in the prediction [42,43]. One important limitation of those approaches is that they require one to obtain appropriate normative data (i.e., test–retest data for the relevant patient population at the relevant intervals) which may be impractical in many surgical centers and may be particularly prohibitive for including new measures or test revisions. Furthermore, they may place a higher value on conservative interpretation (i.e., specificity) at the expense of detecting more modest changes (i.e., sensitivity) that may nonetheless be meaningful to the individual patient. Indeed, it is this last aspect that becomes most important as a professional issue in providing the best information to patients: the proportion of decline expected in memory performance may be more easily understood and evaluated by the patient than whether one might expect a change outside of the typical range. Nonetheless, we acknowledge that our own prediction equations do not take into account some of these factors, although our use of composites may make them less susceptible to strong biases due to individual test characteristics, and it is possible that statistical artifacts as yet unexplored may inflate or underestimate prediction. Indeed, that is the importance of utilizing these equations for novel cases as the data from our small cross-validation sample demonstrate. We and others (e.g., [44]) argue that composites derived from multiple tests of the same domain using data reduction techniques such as PCA provide more stable estimates of the underlying dimension and, therefore, a better metric for correlational or predictive analyses. While this aggregation by type of material (verbal versus visuospatial) is well established in the neuropsychology of epilepsy, we do not deny that there may be finer-grained distinctions between memory processes supported by each hemisphere and by different regions within each medial temporal lobe or imply that lateralization is wholly exclusive. Elsewhere [45], we have argued that preoperative assessment in epilepsy would benefit from adopting some of the concepts from cognitive neuroscience such as recollection versus familiarity in

7

interpreting our test profiles with respect to whether they indicate medial temporal dysfunction, and Saling [46] has made a similar argument with respect to the perirhinal cortex and associative memory. Further, the present data confirm that some aspects of preoperative performance on the ‘nonspecific’ material (e.g., visuospatial memory in patients with left mTLE) can have a significant predictive value. Note that such findings may indicate a ‘blurring’ of material specificity due to patient characteristics (e.g., subtle reorganization of memory functions with recurrent seizures) or test characteristics (e.g., putatively visuospatial tests that lend themselves to verbal encoding), the contribution of ‘functional reserve’ by the contralateral medial temporal lobe, or both [14,18,47,48]. However, the key motivation in the current study is how well neuropsychology scores on their own can predict postoperative change, and here, we find component scores more powerful than individual test scores. 5. Conclusion The implementation of a simple data reduction strategy for predicting postoperative memory decline in patients with mTLE, as we have done in our own clinical setting, can be achieved in other epilepsy surgery centers, as scores were derived from neuropsychological tests that are either commonly used or similar to others employed in centers such as ours. Thus, similar latent components can easily be derived from datasets available at other centers. The fact that latent scores can be calculated for new patients as they are assessed and that our regression equations were applied successfully to a new prospective cohort gives us considerable confidence that data reduction can be employed immediately in our clinical practice. Principal component analysis scores can serve as simple but robust measures of cognitive performance and be used as potential correlates of structural damage, functional integrity, and other disease characteristics for clinical and research purposes. Our next phase will involve adding anatomic and functional imaging metrics to our prediction algorithm, as these may enhance prediction if they capture unique sources of variance. The ultimate goal is to provide each patient and treatment team with the best possible information regarding risk of postoperative memory decline to optimize their basis for making decisions about surgery. Disclosure of conflicts of interest None of the authors had any conflicts of interest to disclose. We confirm that we have read the Journal's position on issues involved in ethical publication and affirm that this report is consistent with those guidelines. Acknowledgments This project was supported by a CIHR grant to MPM, by a studentship from the Savoy Epilepsy Foundation and a fellowship from the Katz Foundation to MSL, and by a scholarship from the German Research Foundation to CM. We would like to thank John Paul Koning, Timour Al-Khindi, and Mary Jaciw for their help and all patients for consenting to share their results. Appendix A. Supplementary data Supplementary data to this article can be found online at http://dx. doi.org/10.1016/j.yebeh.2013.09.043. References [1] Jones-Gotman M, Smith ML, Risse GL, Westerveld M, Swanson SJ, Giovagnoli AR, et al. The contribution of neuropsychology to diagnostic assessment in epilepsy. Epilepsy Behav 2010;18:3–12.

Please cite this article as: St-Laurent M, et al, Using multivariate data reduction to predict postsurgery memory decline in patients with mesial temporal lobe epilepsy, Epilepsy Behav (2013), http://dx.doi.org/10.1016/j.yebeh.2013.09.043

8

M. St-Laurent et al. / Epilepsy & Behavior xxx (2013) xxx–xxx

[2] McAndrews MP. Remote memory in temporal lobe epilepsy. In: Zeman A, Kapur N, Jones-Gotman M, editors. Epilepsy and memory. Oxford, UK: Oxford University Press; 2012. p. 227–43. [3] Helmstaedter C, Elger CE. Cognitive consequences of two-thirds anterior temporal lobectomy on verbal memory in 144 patients: a three-month follow-up study. Epilepsia 1996;37:171–80. [4] Helmstaedter C, Roeske S, Kaaden S, Elger CE, Schramm J. Hippocampal resection length and memory outcome in selective epilepsy surgery. J Neurol Neurosurg Psychiatry 2011;82:1375–81. [5] Lee T, Mackenzie RA, Walker AJ, Matheson JM, Sachdev P. Effects of left temporal lobectomy and amygdalohippocampectomy on memory. J Clin Neurosci 1997;4:314–9. [6] Gleissner U, Helmstaedter C, Elger CE. Right hippocampal contribution to visual memory: a presurgical and postsurgical study in patients with temporal lobe epilepsy. J Neurol Neurosurg Psychiatry 1998;65:665–9. [7] Gleissner U, Helmstaedter C, Schramm J, Elger CE. Memory outcome after selective amygdalohippocampectomy: a study in 140 patients with temporal lobe epilepsy. Epilepsia 2002;43:87–95. [8] Janszky J, Jokeit H, Kontopoulou K, Mertens M, Ebner A, Pohlmann-Eden B, et al. Functional MRI predicts memory performance after right mesiotemporal epilepsy surgery. Epilepsia 2005;46:244–50. [9] Rabin ML, Narayan VM, Kimberg DY, Casasanto DJ, Glosser G, Tracy JI, et al. Functional MRI predicts post-surgical memory following temporal lobectomy. Brain 2004;127:2286–98. [10] Bell BD, Davies KG, Haltiner AM, Walters GL. Intracarotid amobarbital procedure and prediction of postoperative memory in patients with left temporal lobe epilepsy and hippocampal sclerosis. Epilepsia 2000;41:992–7. [11] Binder JR, Sabsevitz DS, Swanson SJ, Hammeke TA, Raghavan M, Mueller WM. Use of preoperative functional MRI to predict verbal memory decline after temporal lobe epilepsy surgery. Epilepsia 2008;49:1377–94. [12] Bonelli SB, Powell RH, Yogarajah M, Samson RS, Symms MR, Thompson PJ, et al. Imaging memory in temporal lobe epilepsy: predicting the effects of temporal lobe resection. Brain 2010;133:1186–99. [13] Dupont S, Duron E, Samson S, Denos M, Volle E, Delmaire C, et al. Functional MR imaging or Wada test: which is the better predictor of individual postoperative memory outcome? Radiology 2010;255:128–34. [14] McCormick C, Quraan M, Cohn M, Valiante TA, McAndrews MP. Default mode network connectivity indicates episodic memory capacity in mesial temporal lobe epilepsy. Epilepsia 2013;54:809–18. [15] Wagner K, Frings L, Halsband U, Everts R, Buller A, Spreer J, et al. Hippocampal functional connectivity reflects verbal episodic memory network integrity. Neuroreport 2007;18:1719–23. [16] Baxendale S, Thompson P, Harkness W, Duncan J. Predicting memory decline following epilepsy surgery: a multivariate approach. Epilepsia 2006;47:1887–94. [17] Baxendale S, Thompson P, Harkness W, Duncan J. The role of the intracarotid amobarbital procedure in predicting verbal memory decline after temporal lobe resection. Epilepsia 2007;48:546–52. [18] Chelune GJ. Hippocampal adequacy versus functional reserve: predicting memory functions following temporal lobectomy. Arch Clin Neuropsychol 1995;10:413–32. [19] Davies KG, Bell BD, Bush AJ, Wyler AR. Prediction of verbal memory loss in individuals after anterior temporal lobectomy. Epilepsia 1998;39:820–8. [20] Elshorst N, Pohlmann-Eden B, Horstmann S, Schulz R, Woermann F, McAndrews MP. Postoperative memory prediction in left temporal lobe epilepsy: the Wada test is of no added value to preoperative neuropsychological assessment and MRI. Epilepsy Behav 2009;16:335–40. [21] Jokeit H, Ebner A, Holthausen H, Markowitsch HJ, Moch A, Pannek H, et al. Individual prediction of change in delayed recall of prose passages after left-sided anterior temporal lobectomy. Neurology 1997;49:481–7. [22] Stroup E, Langfitt J, Berg M, McDermott M, Pilcher W, Como P. Predicting verbal memory decline following anterior temporal lobectomy (ATL). Neurology 2003;60: 1266–73. [23] Harvey DJ, Naugle RI, Magleby J, Chapin JS, Najm IM, Bingaman W, et al. Relationship between presurgical memory performance on the Wechsler Memory Scale—III and

[24]

[25]

[26] [27] [28] [29]

[30] [31] [32] [33] [34] [35] [36] [37] [38] [39] [40]

[41]

[42]

[43]

[44]

[45]

[46] [47]

[48]

memory change following temporal resection for treatment of intractable epilepsy. Epilepsy Behav 2008;13:372–5. Lineweaver TT, Morris HH, Naugle RI, Najm IM, Diehl B, Bingaman W. Evaluating the contributions of state-of-the-art assessment techniques to predicting memory outcome after unilateral anterior temporal lobectomy. Epilepsia 2006;47:1895–903. Richardson MP, Strange BA, Thompson PJ, Baxendale SA, Duncan JS, Dolan RJ. Pre-operative verbal memory fMRI predicts post-operative memory decline after left temporal lobe resection. Brain 2004;127:2419–26. Lattin JM, Carroll JD, Green PE. Analyzing multivariate data. Pacific Grove, CA: Brooks/Cole — Thomson Learning; 2003. Wechsler D. Wechsler Abbreviated Scale of Intelligence. San Antonio, TX: Harcourt Assessment; 1999. Warrington EK. Recognition Memory Test: manual. Berkshire, UK: NFER-Nelson; 1984. Strauss E, Sherman EMS, Spreen O. A compendium of neuropsychological tests: administration, norms, and commentary. 3rd ed. New York: Oxford University Press; 2006. Spreen O, Strauss E. A compendium of neuropsychological tests: administration, norms and commentary. New York: Oxford University Press; 1991. Petrides M. Deficits on conditional associative-learning tasks after frontal- and temporal-lobe lesions in man. Neuropsychologia 1985;23:601–14. Taylor AE, Saint-Cyr JA, Lang AE. Memory and learning in early Parkinson's disease: evidence for a “frontal lobe syndrome”. Brain Cogn 1990;13:211–32. Cattell RB. The scientific use of factor analysis. New York: Plenum; 1978. Gorsuch RL. Factor analysis. 2nd ed. Hillsdale, NJ: Erlbaum; 1983. Kline P. Psychometrics and psychology. London: Academic Press; 1979. Everitt BS. Multivariate analysis: the need for data, and other problems. Br J Psychiatry 1975;126:237–40. Hair JFJ, Anderson RE, Tatham RL, Black WC. Multivariate data analysis. 4th ed. Saddle River, NJ: Prentice Hall; 1995. Preacher KJ, MacCallum RC. Exploratory factor analysis in behavior genetics research: factor recovery with small sample sizes. Behav Genet 2002;32:153–61. de Winter JCF, Dodou D, Wieringa PA. Exploratory factor analysis with small sample sizes. Multivar Behav Res 2009;44:147–81. Chelune GJ, Naugle RI, Luders H, Sedlak J, Awad IA. Individual change after epilepsy surgery: practice effects and base-rate information. Neuropsychology 1993;7:41–52. Hermann BP, Seidenberg M, Schoenfeld J, Peterson J, Leveroni C, Wyler AR. Empirical techniques for determining the reliability, magnitude, and pattern of neuropsychological change after epilepsy surgery. Epilepsia 1996;37:942–50. Martin R, Sawrie S, Gilliam F, Mackey M, Faught E, Knowlton R, et al. Determining reliable cognitive change after epilepsy surgery: development of reliable change indices and standardized regression-based change norms for the WMS-III and WAIS-III. Epilepsia 2002;43:1551–8. Sawrie SM, Chelune GJ, Naugle RI, Luders HO. Empirical methods for assessing meaningful neuropsychological change following epilepsy surgery. J Int Neuropsychol Soc 1996;2:556–64. Frazier TW, Youngstrom EA, Chelune GJ, Naugle RI, Lineweaver TT. Increasing the reliability of ipsative interpretations in neuropsychology: a comparison of reliable components analysis and other factor analytic methods. J Int Neuropsychol Soc 2004;10:578–89. McAndrews MP, Cohn M. Neuropsychology in temporal lobe epilepsy: influences from cognitive neuroscience and functional neuroimaging. Epilepsy Res Treat 2012;2012:925238 [13 pp.]. Saling MM. Verbal memory in mesial temporal lobe epilepsy: beyond material specificity. Brain 2009;132:570–82. Glikmann-Johnston Y, Saling MM, Chen J, Cooper KA, Beare RJ, Reutens DC. Structural and functional correlates of unilateral mesial temporal lobe spatial memory impairment. Brain 2008;131:3006–18. Powell HW, Richardson MP, Symms MR, Boulby PA, Thompson PJ, Duncan JS, et al. Reorganization of verbal and nonverbal memory in temporal lobe epilepsy due to unilateral hippocampal sclerosis. Epilepsia 2007;48:1512–25.

Please cite this article as: St-Laurent M, et al, Using multivariate data reduction to predict postsurgery memory decline in patients with mesial temporal lobe epilepsy, Epilepsy Behav (2013), http://dx.doi.org/10.1016/j.yebeh.2013.09.043