Passive localization of the central sulcus during sleep based on intracranial EEG

Abstract

We test the performance of a novel operator-independent EEG-based method for passive identification of the central sulcus (CS) and sensorimotor (SM) cortex. We studied seven patients with intractable epilepsy undergoing intracranial EEG (icEEG) monitoring, in whom CS localization was accomplished by standard methods. Our innovative approach takes advantage of intrinsic properties of the primary motor cortex (MC), which exhibits enhanced icEEG band-power and coherence across the CS. For each contact, we computed a composite power, coherence, and entropy values for activity in the high gamma band (80–115) Hz of 6–10 min of NREM sleep. Statistically transformed EEG data values that did not reach a threshold (th) were set to 0. We computed a metric M based on the transformed values and the mean Euclidian distance of each contact from contacts with Z-scores higher than 0. The last step was implemented to accentuate local network activity. The SM cortex exhibited higher EEG-band-power than non-SM cortex (P < 0.0002). There was no significant difference between the motor/premotor and sensory cortices (P < 0.47). CS was localized in all patients with 0.4 < th < 0.6. The primary hand and leg motor areas showed the highest metric values followed by the tongue motor area. Higher threshold values were specific (94%) for the anterior bank of the CS but not sensitive (42%). Intermediate threshold values achieved an acceptable trade-off (0.4: 89% specific and 70% sensitive).

Introduction

The division within the central sulcus (CS) between the sensory and motor cortices is one of the most deep-seated concepts in neuroscience. It constitutes a commencing point for cogitating structure, connectivity, function, and plasticity of the brain. Accordingly, the assessment of CS was entertained by Rolando, Jackson, Ferrier, Horsley, and Cushing and was among the primary sub-domains of Penfield’s work (Isitan et al. 2020). Many investigators described motor responses to high-frequency cortical stimulation in humans and animals. Cushing reported sensations elicited by stimulation of the postcentral gyrus of an awake patient in 1909 (Cushing 1909). High-frequency electrical cortical stimulation (ECS) may localize the sensory-motor function, critical in preserving these vital regions in epilepsy and tumoral neurosurgical patients. This standard procedure, however, is rather inefficient and is sometimes marred by limitations, including current shunting, after-discharges, and provoked seizures (Yeomans 1990). In addition, ECS of the primary motor cortex (MC) may elicit sensory responses and vice versa (Kovac et al. 2011); hence, the identification of the anatomical CS for surgical purposes is crucial. In the last few decades, there have been parallel interests to localize CS based on direct brain surface electrophysiological recordings in relation to tasks. Likewise, Woolsey et al., showed that somatosensory evoked potentials (SSEPs) exhibit topographic relationships between the body and its cerebral representation (Woolsey and Erickson 1950; Woolsey 1958). Median SSEPs have played a role in brain mapping of the sensorimotor (SM) cortices (Dinner et al. 1987; Wood et al. 1988). The procedure is possible even under general anesthesia. SSEPs may be altered, however, when there are overlapping lesions like in cortical dysplasia. Thus, the accuracy is sometimes poor. SSEPs may be unavailable, as they require constant stimulation and averaging of recorded responses and are liable to technical inadequacies. Also, SSEPs localize CS as an anatomical landmark but do not map out the functional details of the SM strip. Methods that use task-related icEEG modulations in gamma band have been explored in the past two decades (Crone et al. 1998; Alkawadri et al. 2018). Task-related gamma activation requires patients’ engagement and participation (Crone et al. 1998) and is not always possible under anesthesia or in patients with developmental delays.

We have shown a consistent presence of high-frequency oscillations (HFOs) in the peri-Rolandic and occipital cortices, and proposed the use of this pattern for passive mapping (Alkawadri et al. 2014). In this study, we implement a task- and stimulus-free method for identification of CS and SM cortices by correlating the classification to the ground truth of ECS. The method is innovative and exploits inherent electrophysiological properties of CS.

Materials and Methods

Subjects and mapping details are available in Supplementary Text, Table 1 and Figure 1A.

Table 1.

Summary of clinical and demographic data, results of presurgical workup, resections, and surgical outcomes

Pt	A/G/H	Onset (years)	Known risk factors	Typical clinical seizure	Sz Freq per month	MRI	PET	SPECT (ISAS)	fMRI (language)	Scalp video-EEG	ECOG Sampling	Resection	Follow up (months)	Outcome (Engel)
1	10/F/R	5 (5)	DD	Olfactory > Auto-motor/Autonomic	150	Wnl	L temporal	L temporal	L	L FT	L frontotemporal	L temporal	23	I
2	27/F/R	15 (12)	None	Auto-motor > rare 2nd gen	4	Old R thalamic infarct	R anterior temporal	R temporal	L	R temporo-parietal	R FTP	R temporo-parietal	3	I
3	28/F/R	16(12)	DD	Cephalic>atonic	10	L temporal heterotopia.	R Med temporal	Bi med parietal	Bi	Gen delta	Bi-FP(L > R)	-	-	-
4	53/F/R	37(16)	AVM	Sensory > GTC	4	R frontal encephalomalacia	R FT	Increased uptake R FT	L	R frontal	Bi-max R FP	R frontal	26	II
5	20/F/R	11 (9)	AVM/resection	Loud vocalization > “looking spacey”	14	L temporo-parietal hemosiderin	NA	NA	R	L temporal	L FP	L temporo-parietal	17	I
6	27/M/R	0.5(26)	DD	Autonomic>Left motor	6	Bi-thalamic T2-signal	L frontal horn	—	L	R FP	R FP	—	—	—
7	33/M/R	17(16)	trauma	Déjà vu > ictal speech	1.5	wnl	wnl	—	L	Bi-temporal	Bi-FT	—	—	—

A, age; G, gender; H, handedness; F, female; M, male; DD, developmental delay; wnl, within normal range; L, left; R, right; AVM, arteriovenous malformation; GTC, generalized tonic clonic; Med, medial; FT, frontotemporal; FP, frontoparietal; Bi, bilateral; Gen, generalized; max, maximum.

Figure 1.

(A) Functional and anatomical contact classification of icEEG contacts. (B) Whisker plots and statistical comparisons of sum power, ^* denotes an outlier.

In brief

Data Collection and EEG Data Employed in Analysis

Each patient underwent continuous icEEG recording with clinical video-EEG monitoring equipment. Subdural grids, multiple strips, and depth electrodes (AdTech Medical) were placed as required. The subdural electrode contacts were 4-mm diameter platinum-iridium disks. The depth electrode contacts were 2.4 mm. Electrodes had 10-mm center-to-center spacing. icEEG was recorded on a 256–512 channel video-icEEG long-term monitoring equipment (Natus Medical Incorporated). EEG recordings were referenced to an electrode implanted in the diploic space using an active-ground reference system. Data were collected with a hardware high-pass filter of 0.1 Hz. The analysis was performed on 6–10 min of NREM sleep, day 3 postimplantation or day 1 before seizures if the latter was recorded sooner, typically between 3 and 4 AM which was extended as needed if NREM sleep was not recorded, verified by the visual analysis of simultaneous scalp EEG recordings by board-certified neurologist. Assessment and rejection of noisy segments have been reported elsewhere (Alkawadri et al. 2014). We chose NREM sleep EEG segments for analysis for the following reasons. 1) To standardize the procedure across subjects. 2) EEG activity within the band of interest is not modulated by events which are important to control, such as, alterations in attention, movements, and eye-opening or -closure thus maintaining complete passiveness. 3) Movement and electromyogenic artifacts are minimal (Alkawadri et al. 2014).

Description of the Passive Method

Rationale: Identifying the CS involves visual skills and knowledge of anatomy and orientation. A binary classifier is expected to perform variably due to accidental out-of-network coactivation irrespective of standard performance metrics. Further, many signal parameters are partially colinear. For instance, gamma coherence in the CS represented an outlier in an otherwise moderately linearly correlated power and coherence value (Supplement A). Those features together dampen the performance of a logistic classifier and could lead to overfitting and underperformance depending on data size and validation. Here, we devise and evaluate an individualized threshold-based visual nonparametric statistical system. The system leverages electrophysiological properties of the CS, which is sampled by multiple adjacent or contiguous electrode contacts, to highlight local activations in the CS regardless of the classifier’s standard performance metrics to guide visual identification.

The metric was based on an informed expert assessment of data and known properties of the CS and SM cortex. Specifically, the method exploits inherent properties of MC, the enhanced icEEG high gamma band-power (Alkawadri et al. 2014) and the coherence across the CS (Towle et al. 1999). For each contact X, we calculated a Z-score of the following composite of power, entropy and synchrony. In order to verify this in statistical sense, we first constructed a multivariate model based on these values along with power, synchrony in other frequency bands and verified statistically significant effect of the metric involved prior to applying the visual transformation (Supplement B). The visual detection of CS by human reviewer benefits from qualitative binary system. Classical qualitative methods and validation cutoffs may contaminate visual identification; a specific binary visual system is desired 1 for anterior lip of CS and 0 all other. We devised a metric system based on the metrics validated above that enables that binary transformation according to the following:

where pX is the scaled array of sum of the root mean square normalized to the maximum value of icEEG in the high gamma band (80–115)Hz based on a filtered time-series using a fourth-order Butterworth filter for a contact X, and cX is the normalized array scalar of mean magnitude squared coherence in the same band using a 500-ms Hamming window with 5% overlap between successive windows, between contact X and all other contacts. Since gamma coherence was positive outliers in an otherwise partially correlated power/coherence data per our data screening, simple arithmetic multiplication is a rapid method to deduce interference in binary CS classification informed by data screening and findings of initial multivariate analysis (Supplement A). eX is a scaled array of a composite value of [mean (entropy X)/Variance (entropy X)] normalized to the maximum value, in the same high-frequency band and window sizes, employing 2-Hz steps and for frequencies 80–100 Hz and 4-Hz steps for frequencies 100–115 Hz. All normalized values are scaled to the maximum value within a category. We used entropy to enhance the visibility of physiologic activity, in case of vicinity to epileptic brain regions, by suppressing accidental aberrant “activations” in the latter (Mooij et al. 2016); and are constants identified from a prior logistical fitting and irrelevant due to subsequent z-transformation (Supplement B); while 0.2 for the purpose of this analysis.

We used the product of the three measures, power, entropy, and coherence to emphasize sites with consistently high values across the features (CS) and dampen the effect of 1) high power and low coherence (e.g., sensory); 2) high power, coherence, and entropy variance (e.g., epileptic tissue); or 3) low power, coherence, and entropy variance (e.g., nonfunctional other). Since we estimate the Z-score of this value, the calculations could be performed on the nonnormalized values; however, we have employed normalization in order to standardize the procedure across different recording durations outside this study, and to shield against extreme power values such as highly active epileptic contact, or accidental deflation of coherence values in case of spatial under-sampling.

Subsequently, we applied a Z-score transformation for the composite values.

Since negative values are irrelevant for 3D visualization of the results, we applied the following transformation:

Next, we calculated the proposed metric, M

M is the proposed novel marker of CS. d is the normalized mean Euclidian distance of each contact from contacts with Z_f scores greater than 0. d is not a geometrical case-specific constant, but it is rather threshold dependent (Supplement C and D). We incorporate distance to emphasize local network activity. We quantified the performance of a given threshold by 1) performance metrics according to binary results of CS classification (1 for any value > 0, and 0); 2) two reviewers classification of CS; and 3) median differences multiplied by IQR as a measure of effectiveness of the dynamic range of metric (Supplement E).

We used contacts with the highest M values as a seed for further network analysis based on infra-slow activity (ISA). Because EEG data was recorded with a 0.1-Hz high-pass filter, the covariance between electrodes is could be analyzed (Ko et al. 2011; Breshears et al. 2012). ISA has been shown to play an important role in defining the fundamental architecture and organization of cortex as it relates to both functional and pathological brain networks (Schalk et al. 2008; Baria et al. 2011) and seems to correlate with fMRI BOLD signal thus its uses have been proposed for functional mapping.

The temporal correlation of the ISA (0.1–1 Hz) over the 2-s epochs was calculated for all electrode pairs by first computing the covariance between each pair (x, y) by calculating the covariance

where x_i and y_i represent the ith of n values of the ISA from the x and y time-series.

Linear regression and detrending removed the influence of interelectrode distance to increase the sensitivity of cutoffs tested by minimizing the effect of inflation by volume conduction. In our experience and settings, the 2 s windows provide a proper trade-off between time/frequency resolutions, that is, longer time windows have resulted into attenuation of meaningful ISA correlations. There is a hardware limitation of 0.1-Hz cutoff in many clinical differential amplifiers typically/commonly employed to shield against environmental noise. This should not preclude the use of 0.1–1 Hz nonetheless in covariance analysis. Recent studies have proposed that resting state cortical correlation in ISA (<0.1 Hz) provide a mean of identifying eloquent regions (Breshears et al. 2012). This ISA has been shown to correlate with resting state spontaneous BOLD fluctuations used in functional mapping (Biswal et al. 1995; Cordes et al. 2000; He et al. 2008) which in turn correlate with task-related localizing findings.

Visualization of Electrodes and M Values in the MRI Space

We used preimplantation, postimplantation MRI studies, and postimplantation computerized tomography (CT) scans. Each electrode contact was coregistered with a preimplantation 3D triangular mesh model of each subject’s cortex using the coregistration tools available in BioImage Suite software (v3.0). For a realistic visualization of M values, and to help guide visual localization of CS, we constrained the visualization to a mesh of the cortex (Mesh and brain template available in Brainstorm [http://neuroimage.usc.edu/brainstorm] mesh was downsampled to 10,000 vertices). We used the nearest neighbor, linear interpolation to 100 vertices around the center of each electrode contact. Two blinded reviewers marked the anatomical landmark based on the relation to nonzero M values correlating with the anterior lip of the CS in the test dataset [see supplementary text for details of the expert review process].

IRB Statement

The study received intuitional board approval that was in effect throughout the study period. This was a retrospective study and patients did not provide consent for this analysis.

Statistical Analysis

Receiver operating characteristic curves were calculated to estimate the effective cutoffs for a reliable localization of CS. AUC > 0.7 was pre-defined as acceptable. Parametric statistics were applied on the transformed logarithm of the sum of power which was transformed back to absolute values for the visualization of results after proper thresholding. Bonferroni correction for multiple comparisons was implemented when applicable.

Logistical nominal regression: A logistic regression model was used to estimate the log odds of the probability that a channel was CS or SM function as a linear function of predictor variables screened by data visualization and multivariate analysis: . The logistical regression was performed accordingly; cross-validation was performed on randomly selected 0.3 portions of data (Supplement B,F,I).

Support Vector Machine: Support vector machine (SVM) is a supervised machine learning method which is considered as one of the most popular yet robust and accurate classifiers, best suited for binary outcomes. It finds an optimal decision boundary in the hyperplane which results in the maximum margin between the nearest training data of different classes. In this study, we used a radial-basis kernel function and 10-fold cross-validation. Since K-fold cross-validation may provide optimistic estimates, we also performed holdback validation with 0.3 portion of data and three random seeds. In addition, since SVM underperforms with unbalanced data, we assigned weight of 10 for every contact located at the anterior lip of CS. We employed tuning design and chose the model with least among of misclassification from 30 models. Models were appraised according to validation misclassification rate. We employed C cost of misclassification 1 and gamma parameter 0.1 (Supplement G and J).

Statistical testing for significance was performed using JMP (versions 9 and 16; SAS Institute Inc.).

Results

The median threshold for SM mapping was 2.5 mA and range 1.5–7 mA. Please refer to Fig. 1A for the summary of SM mapping results. None of the patients underwent resections within 1 cm of the areas identified as functional by the procedure. CS-other ratio: 1:7.7, and function-other ratio: 1:2.35. The SM cortex showed higher EEG-band-power than non-SM cortex (medians normalized power 0.22 [IQR 0.11–0.93] vs. 0.12 [IQR 0.08–0.24] P < 0.0002). There was no difference between the motor/premotor and sensory cortices (medians normalized power 0.22 [IQR 0.11–0.42] vs. 0.20 [IQR 0.1–0.37] P < 0.47); however, the hand and leg motor areas exhibited higher average band-power than other noneloquent cortex (median normalized power 0.32, 0.42, respectively, P < 0.0001) (Fig. 1B).

A trend of a higher mean band-power in the primary hand motor and tongue motor areas compared with other somatosensory functions was observed, but this was not significant (Fig. 3E). There was a significant difference in ISA correlation values intra-SM versus extra-SM cortex (0.853 vs. 0.245, P < 0.0001).

Figure 3.

(A) schematic illustration of features involved in analysis and composite product M at threshold 0.4 in the source space. Analyses in the manuscript were performed contact-wise on the convexity grid only. The red line marks the anatomical CS. (B) Schematic representation [Electrode Space] of the metric M. The size represents the value M, whereas the color represents the normalized power within subject. Note that power is not adequate to classify sensory versus motor (see E). Blue line: CS as per SSEPs, red line: anatomical CS based on intra-operative image and 3D reconstruction. Value th = 0.6. (C) Intra-operative picture of the same grid shown in A, the dotted line marks the anatomical CS confirmed by SSEPs. (D) Results of ECS in same patient. Red: Motor, Green: Sensory, circle with colorful quadrant: hand/arm, circle with full color face (F), tongue (T), or glottal (Th). (E) Power within 80–115 Hz band recorded during 6–10 min of NREM sleep averaged from seven patients.

In all cases, there was an agreement between the two blinded reviewers on the correct localization of CS based on a composite images of M values registered with MRI. Thresholds 0.4 < th < 0.6, localized CS in all subjects and achieved highest scaled median values, th = 0.4 for metric M and class CS = Anterior lip of anatomical CS (Fig. 3 and Supplement Fig. 1 illustrating the performance of metrics and its individualized components, see also Supplement E and H).

The primary hand and leg motor areas exhibited the highest M values followed by the tongue motor area (Fig. 2F). The most effective th cut-off (≥1) achieved 94% specificity for the anterior bank of the CS but low sensitivity 42%. Intermediate threshold values achieved an acceptable trade-off (0.4: 89% specific and 70% sensitive) (Fig. 2C,D,G, illustrating classification performance at different thresholds). Importantly, intermediate threshold enabled CS localization by visual review in 100% of instances (Supplement E). That range was indeed most practical for visual identification of the anterior edge of CS (Fig. 3 and Supplementary Fig. 1). The other parameters employed in M calculation did not achieve acceptable performance alone (Fig. 2A,B,H). Although multivariate and SVM-based methods achieved better performance in single contact classification, they underperformed M in visual human-aided assessment (Supplement G and H). As expected, there was a modest decrease in performance when holdback validation employed in place of K-fold (Supplement E and F). Weighted K-fold cross-validation performed the best among other methods when results reviewed by human reviewers. A cut-off 0.79 for correlation coefficient of ISA-based map of a seed with highest M value was the most effective in the classification of contacts with function (vs. no-function) when ECS was employed as the gold standard sensitivity 79% and specificity 72% (92% sensitive if seed placed in all possible positive M contacts, instead of highest M). There has been moderate improvement with employing multivariate and SVM models. However, it is still not adequate to replace standard of care in surgical sense (Supplement I–J).

Figure 2.

Probability density plots of normalized power according anatomical-functional correlates of electrodes contacts (A, B, H), M value, and two thresholds −0.4 (C) and 0.4 (D), notice better electrode classification accomplished by latter. Statistical comparisons of normalized power per category function (E) and per M value (F), and pair-wise comparisons of M values between CS and other electrodes at different thresholds (G). Note that best normalized median differences accomplished at values equal or higher than th 0.4. CS: Contacts at the anterior lip of the CS.

Discussion

We implemented a rapid procedure for task-free and stimulation free localization of the CS during sleep that exploits intrinsic electrophysiological properties of the primary motor strip. The method did not need participation from subjects or delivery of stimuli and was based on passive monitoring of background electrocorticography (ECoG). Furthermore, network analysis was operator-independent, whether regarding the seed identification or network analysis, which, to our knowledge, had not been carried out before. The method made two practical priori assumptions: The CS and regions around it are sampled, and the occipital regions are not included in the analysis. In order for the method to function, non-SM cortex should be sampled, which is the case when using a standard grid.

Two reviewers identified CS in all cases. The metric M was specific for the anterior lip of CS, making it ideal for identifying this crucial anatomical landmark. It is plausible to consider its use in patients who cannot take part in standard ECS mapping, when stimulation is not available, or in the operating room.

The current method bypasses common problems with icEEG studies where the analysis is often performed at the group level, whereas case-wise results are most relevant for clinical applicability. This most evident in the subtle advantage in the performance of our method based on within subject thresholding versus the highest performing K-fold group-wise validated weighted SVM model. Also, the supervised nonparametric training is more suitable for human EEG data analysis than parametric statistics commonly employed in direct time-series and spectral analyses (Maris and Oostenveld 2007). This highlights the potential of developing models applicable on individual subjects and nonparametric thresholding methods in crucial practice such as brain surgery compared with group-wise parametric statistics that may not be always translatable on the case-to-case basis.

Previous case-wise analysis led to an estimate that 3% of modern digitized point-point icEEG data are used clinically (Alkawadri 2019). This gap underscores the excellent potential for future translational research, at an age of unprecedented exponential advances in computational power and time-series and machine learning methods, which are increasingly more likely to reveal findings that are a challenge for visual analysis. The current study is a proof of the concept for the benefits of free-running ECoG and supervised nonparametric statistical analysis. Infralow activity network-based methods, did not achieve adequate specificity to replace ECS for mapping beyond mere localization of CS, for example, with the face-hand margin neurosurgical delineation. The amplifiers’ low-frequency cutoffs might have contributed to this limitation, and future studies on histopathological and cellular substrates with lower frequencies are desirable.

Our method takes advantage of the known intrinsic properties of the CS; markers based on power and coherence are likely to show consistent activation in epileptic brain regions, hence the incorporation of entropy and distance in analyses. Problems related to volume conduction are less relevant in HFOs arising from discrete neuronal populations within several mm³ at a time while the typical inter-electrode distance is in the order of cm (Grenier et al. 2001). Detrending should account for concerns with volume conduction in the ISA band still not providing adequate specificity for mapping the SM strip in its entirety. Our data are screened with group-wise coherence analysis to eliminate possible inflation due to contaminated references (Zaveri et al. 2000). Also, reference electrical contamination is hardware-suppressed by the active-ground reference (Xu et al. 2011). Given the band of interest and the direct proximity to sources, cortex-constrained geometrical interpolation of M/power values into MRI space is adequate for data visualization (David et al. 2011; Alkawadri et al. 2014). Although noninvasive means of CS localization may provide acceptable performance, the invasive confirmation remains the gold standard when the latter indicated on a clinical basis (Towle et al. 2003). Recording of SSEPs requires an integrity of median-spine-cortical axis and competent technical recording. The phase reversal of somatosensory-evoked potentials may not be successful in one tenth of cases, due to an underlying pathology or other anatomic consideration (Cedzich et al. 1996). High-resolution EEG cast insights into suboptimal specificity for a procedure often performed on a small number of electrodes (Van de Wassenberg et al. 2008). On the other hand, metric M localized the CS in 100% of cases and provided excellent spatial resolution, and also it was successful in cases with vicinity to the pathologic brain regions and do not require patient, technicians, or expert engagement besides the standard setup.

In the future, prospective evaluation can be investigated against standard practices with a variable leniency in localization standards. We supplement our manuscript with python codes of highest performing models for validation. This can be potentially employed in mapping epileptic and functional brain networks alike. Our findings call for further investigation into the maturating behaviors, in infants and toddlers, where motor responses with standard ECS are limited (Duchowny and Jayakar 1993). Also, future work may help identify means of passive localization of other functions such as language.

Conclusion

The central sulcus and the sensorimotor brain regions exhibit intrinsic electrophysiological characteristics that enable the passive identification of this vital brain structure independent of tasks and stimulation during sleep. The findings invite for prospective evaluation against standard practices. Future studies may evaluate applicability in other surgical-vital sites such as language areas and may extrapolate the findings to applications in the operating room under anesthesia.

Notes

The authors wish to thank Rebecca Khozein DOM, MS, REEG/EPT, RPSGT, RNCST, and Tamara Wing REEG for help in EEG data acquisition. The authors would like to acknowledge the reviewer for the insightful comments that enhanced on the presentation of the manuscript. Conflict of Interest: None of the authors have conflict of interest to disclose.

• Dr Alkawadri assumes full responsibility for the data, the analyses and interpretation, and the conduct of the research; that the author has full access to all of the data; and that the author has the right to publish any and all data separate and apart from any sponsor.

• Indication that the Methods section includes a statement that an IRB or regional review board has approved the use of humans for this study.

• No personal communication cited in the article.

Funding

American Epilepsy Society (award #412064); National Center for Advancing Translational Science (NCATS) (Grant No. KL2TR000140 ), a component of the National Institute of Health (NIH) and the C.G. Swebilius Trust (all to R.A.). Note, the content of the manuscript is solely the responsibility of the authors and do not represent the official view of NIH.