Functional prefrontal reorganization accompanies learning-associated refinements in surgery: A manifold embedding approach

Abstract

The prefrontal cortex (PFC) is known to be vital for acquisition of visuomotor skills, but its role in the attainment of complex technical skills which comprise both perceptual and motor components, such as those associated with surgery, remains poorly understood. We hypothesized that the prefrontal response to a surgical knot-tying task would be highly dependent on technical expertise, and that activation would wane in the context of learning success following extended practice. The present series of experiments investigated this issue, using functional Near Infrared Spectroscopy (fNIRS) and dexterity analysis to compare the PFC responses and technical skill of expert and novice surgeons performing a surgical knot-tying task in a block design experiment. Applying a data-embedding technique known as Isomap and Earth Mover's Distance (EMD) analysis, marked differences in cortical hemodynamic responses between expert and novice surgeons have been found. To determine whether refinement in technical skill was associated with reduced PFC demands, a second experiment assessed the impact of pre- and post-training on the PFC responses in novices. Significant improvements (p < 0.01) were observed in all performance parameters following training. Smaller EMD distances were observed between expert surgeons and novices following training, suggesting an evolving pattern of cortical responses. A random effect model demonstrated a statistically significant decrease in relative changes of total hemoglobin (ΔHbT) [coefficient = −3.825, standard error (s.e.) = 0.8353, z = −4.58, p < 0.001] and oxygenated hemoglobin (ΔHbO₂) [coefficient = −4.6815, s.e = 0.6781, z = −6.90, p < 0.001] and a significant increase in deoxygenated hemoglobin (ΔHHb) [coefficient = 0.8192, s.e = 0.3034, z = 2.66, p < 0.01] across training. The results indicate that learning-related refinements in technical performance are mediated by temporal reductions in prefrontal activation.

• Surgical training
• skills assessment
• motor skills learning
• manifold embedding

Introduction

Learning to perform technical maneuvers in surgery requires explicit timing, visuospatial orientation, visuospatial working memory (VSWM), fine bimanual dexterity, attention and concentration. In this regard, surgical procedures can be viewed as comprising a set of complex skills. Successful performance of a surgical procedure represents an extremely demanding visuomotor and cognitive challenge. Studying the surgeon's brain may provide a useful model for investigating complex perceptual-motor interactions. However, analyses of surgical skills have focused on technical performance, dexterity [1], [2] and, more recently, visual search behaviors [3]. There are only isolated reports of functional brain measurements [4], [5], possibly as a result of the practical constraints of traditional neuroimaging environments. Complex motor skills that require subjects to be ambulant or that demand realistic settings are too challenging to study using functional Magnetic Resonance Imaging (fMRI). Complex everyday tasks such as walking [6], running [7], and apple peeling [8] have been evaluated successfully using functional Near Infrared Spectroscopy (fNIRS), a non-invasive optical neuroimaging technique for monitoring hemodynamic responses to brain activations.

Cortical hemodynamic change is interpreted as originating from volumes of cortex (channels) between light emitters and detectors, as illustrated in Figure 1. Cortical activation is inferred by the presence of task-induced increases in oxygenated hemoglobin (HbO₂) and decreases in deoxygenated hemoglobin (HHb) [9]. Multi-channel fNIRS data is complex; its dimension d may be given by d = c · u · h · s, where c is the number of channels, u is the number of users or subjects, h is the number of hemoglobin species, and s is the number of observations. Although linear dimensionality reduction techniques may assist visualization of fNIRS data, brain behavior is intrinsically nonlinear and dynamic [10]. Therefore, techniques such as Principal Component Analysis (PCA) and Multi-Dimensional Scaling (MDS) may have problems capturing nonlinearities of in vivo brain data. Methods that can recover the intrinsic dimensionality and non-linear structure of the dataset, such as manifold embedding [11], may improve visualization of complex fNIRS data [12].

Figure 1. Schematic illustration demonstrating a surface NIR light emitter coupled to a low-light-sensitive detector. The cortical hemodynamic change from a banana-shaped volume of tissue between the emitter and detector is probed non-invasively. Application of a series of emitters and detectors results in hemodynamic recordings from multiple cortical sites (channels) simultaneously.

In previous work, we used manifold embedding to identify expertise-related differences in prefrontal processing across a limited number of bimanual knot-tying trials [12], and demonstrated that within-trial fluctuations in prefrontal hemodynamics appeared to confer transient gains in technical performance [5]. It is hypothesized that activation of generic attention and control areas, such as the prefrontal cortex (PFC) and anterior cingulate cortex (ACC), is especially important in novice subjects. These regions are known to be crucial in the initial “cognitive” phases of complex motor skill learning for tasks that have no direct relevance to surgery [13], [14]. Following extended practice, investigators have observed attenuation of PFC and ACC activations along with expanded representations of the practiced skill in the primary and supplementary motor cortices and parietal cortex [15–18]. However, the patterns of functional activation associated with extended deliberate practice of a technical surgical task have not been investigated. Currently, it remains uncertain whether PFC attenuation accompanies learning success in this context.

Our aim is to evaluate technical expertise development in surgery through cross-sectional and serial evaluations of cortical brain function in novice and expert surgeons. In this paper, we present data from two closely allied investigations. First, to evaluate differences in cortical behavior between expert and novice surgeons, 62 subjects with varying levels of surgical experience performed a knot-tying task comprising five cycles of task repetitions. Second, a group of 19 surgical novices were evaluated on the same knot-tying task at two time points temporally spaced across a week of deliberate practice. In both investigations, prefrontal activation was evaluated using fNIRS and visualized using manifold embedding. Our research questions were as follows:

• Does prefrontal behavior differ between performers based upon their expertise in a surgical task?

• Does reorganization of prefrontal behavior accompany learning-associated refinements in technical performance following practice and instruction?

• Does manifold embedding meaningfully enhance the visualization and understanding of complex fNIRS data?

Materials and methods

Ethics

The Local Regional Ethics Committee (LREC) approved all aspects of the study protocol in July 2005. All investigations were conducted at the Department of Biosurgery and Surgical Technology (St. Mary's Hospital Campus), Imperial College London. Written informed consent was obtained from each participant prior to enrolment. Participants also completed a questionnaire regarding their previous surgical experience and knot-tying preferences and an Edinburgh Handedness Inventory [19].

Subjects

Study (A) involved 62 right-handed healthy male subjects, including 19 consultant surgeons (mean age ± SD = 46.3 ± 7.6 years), 21 surgical registrars (mean age ± SD = 31.3 ± 4.1 years), and 22 medical students (mean age ± SD = 21.9 ± 1.9 years), all of whom were recruited from Imperial College London. Study (B) involved 19 right-handed healthy male medical students (mean age ± SD = 20.2 ± 0.7 years). Subjects recruited for study (B) were not involved in the first investigation. Medical students were recruited provided they had no prior knowledge of surgical knot-tying. Subjects with a family or previous history of a neuropyschiatric disorder were excluded (n = 1). Subjects were asked to refrain from drinking alcohol for 24 hours prior to the start of each study period.

Knot-tying task paradigm

In both investigations, the task under evaluation consisted of four throws of a hand-tied surgical reef knot performed using 2/0 Polysorb suture material on a bench-top knot-tying trainer (Ethicon Ltd., Somerville, NJ). The task was introduced in a block design manner. Blocks consisted of baseline rest (30 seconds), trial (self-paced surgical reef knot) and post-trial rest (20 seconds) periods. During rest periods, subjects sat silently on a chair in front of the desk with their eyes focused solely on the knot-tying trainer and their hands placed on the table at motor rest. Following a verbal prompt (“start”) from the investigator, subjects formulated four throws of a hand-tied surgical reef knot, self-paced on the knot-tying trainer from the seated resting position. As soon as the last throw of the knot had been laid, subjects returned to the resting position (“rest”). Trial blocks were repeated five times to increase the signal-to-noise ratio.

Training episodes and practice schedule

In both studies, medical students were trained in a single one-hour session to perform hand-tied surgical reef knots, as taught in the Basic Surgical Skills (BSS) course, using the same knot-tying trainer used for the fNIRS experiment. Students received instruction on a one-to-one basis from a surgeon trainer. The trainer demonstrated how to formulate four throws of a surgical reef knot, first without and then with commentary. The trainer then performed the maneuvers again under instruction from the participant. The three sub-steps involved in each throw were then demonstrated again, first with and then without commentary. Participants were required to perform each sub-step four times. A period of 15 minutes was subsequently allocated for practice of the task, with feedback provided by the trainer where necessary. At the end of the practice session, each subject had to successfully form four throws of a hand-tied surgical reef knot using 2/0 Polysorb in order to proceed to the neuroimaging evaluation. No subjects were excluded at this stage, and all medical students progressed to fNIRS evaluation.

The medical student cohort recruited for study (B) then underwent further practice and instruction dispersed over five consecutive days. On each day, subjects attended a group training session lasting approximately 40 minutes. During these sessions, subjects repeatedly practiced the knot-tying technique on identical knot-tying trainers in the bench-top environment. Expert trainers spent 10 minutes with each subject, offering advice and feedback on individual knot-tying technique. Subjects were instructed not to practice outside of these training sessions. In total, each candidate practiced the knot-tying technique for approximately 3 hours and 40 minutes. At the end of the week, a second fNIRS evaluation (session 2) was conducted in an identical manner to the first evaluation (session 1).

Functional neuroimaging, behavioral performance and systemic monitoring

Optical measurements were acquired using a commercially available 24-channel optical topography system (ETG-4000, Hitachi Medical Co., Tokyo, Japan). A configuration comprising two 3 × 3 probes was used. Each probe comprised five optical fiber sources emitting NIR light (from laser diodes) at 690 and 830 nm coupled to four avalanche photodiode detectors, each of which detects the reflected light of its neighboring surrounding emitters. Light from each laser diode was modulated at a unique frequency for each wavelength and channel, enabling the separation of detected signals according to location, by 48 lock-in amplifiers locked to the modulation frequency of the light source. Detected light intensity data were sampled at 10 Hz, digitized and transformed according to their wavelength and location. Optodes were held within a flexible plastic helmet (inter-optode distance = 30 mm) affixed to the participant's head with a surgical bandage (Surgifix, Cisterna d’Asti, Italy). As illustrated in Figure 2, optodes were placed over the prefrontal cortices, positioned according to the original International 10–20 system of electrode placement [20]. Concerning the right hemisphere, optodes were positioned such that the lowermost medial emitter was centered on Fp2, the lowermost lateral on F8, and the central emitter on F4. Optode positioning on the left hemisphere mirrored that on the right. Specifically, the lowermost medial emitter was centered on Fp1, the lowermost lateral on F7, and the central emitter on F3.

Figure 2. The experimental set-up. Schematic illustration of a study subject engaged in a knot-tying drill. The electromagnetic pulse emitter is shown on the table in front of the knot-tying jig and motion-tracking sensors are visible on the dorsum of the subject's hands. The figure illustrates the 3 × 3 arrangement of NIRS optodes and the relative positions of NIR emitters (red) and detectors (blue). Optode location is obtained by transferring topographic data from a representative subject to a 3D cortical surface image of a high-resolution T1-weighted MRI image. The two subplots illustrate the location of NIR channels (shaded circles) on the cortical surfaces. The approximate International 10-10 landmarks are highlighted (yellow boxes). [Color version available online.]

The position on the participant's head contacting the emitting and receiving fibers, as well as four base points (the nasion, inion, and bilateral external auditory meati), was measured using a 3D electromagnetic probe positioning digitizer. Topographic data obtained using an optical tracking system (PATRIOT, Polhemus, Colchester, VT) from representative cases (n = 3) were transferred to a 3D MRI image using a 3D composite display unit (Hitachi Medical Co., Tokyo, Japan) to enhance the appreciation of the cortical anatomy underlying each channel. Probe positions were projected onto the cortical surface of a 3D maximum probability atlas [21] to determine the cortical structures underlying each measuring position.

Knot-tying dexterity was monitored using the Imperial College Surgical Assessment Device (ICSAD) [22], which consists of a commercially available electromagnetic field generating device (Isotrak II, Polhemus, Colchester, VT) and two sensors that are attached to the dorsum of the subject's hands in a standardized manner. Cartesian movement data in coordinates was converted into dexterity data such as the time taken and the number of movements required to complete the task, as well as the distance or “pathlength” traveled by each hand. Data was acquired at 20 Hz with pre-processing for noise removal. All trials were videotaped for retrospective viewing to ensure that the knot-tying technique correctly matched that during training. Inter-group comparisons of knot-tying performance were analyzed using the Kruskal-Wallis and Mann-Whitney U tests for non-parametric data. Technical performance data prior to and following practice (session 1 versus session 2) were analyzed using the Wilcoxon Sign Rank Test for non-parametric related comparisons. Statistical significance was set at the α < 1% level.

fNIRS data processing

All optical data was processed using the functional Optical Signal Analysis program (fOSA, University College London, UK) [23]. Relative changes in light intensities were converted to changes in hemoglobin (HbO₂, HHb, and their sum, total hemoglobin, HbT), applying the modified Beer Lambert Law [24], [25]. Data was baseline corrected, decimated to 1 Hz, and detrended to remove system drift and unrelated physiological signals. To overcome variations in knot-tying durations in fNIRS data, temporal standardization was performed using the “resample” function in Matlab. Final datasets comprised 20 baseline rest values, 37 trial values, and 20 post-trial rest values for each channel and hemoglobin species. A 74-D feature space F was constructed, using resampled HbO₂ and HHb trial values. This data was then reduced to an embedded space P using the fixed references Isomap algorithm FR-Isomap [26], with between-group differences being quantified using Earth Mover's Distance (EMD) [27]. Study (A) data were used as references to construct the embedded space.

FR-Isomap manifold embedding technique

The aim of data embedding techniques is to project high-dimensional data into a lower-dimensional representation, whilst maintaining to a certain degree the intrinsic topology of the input space. High-dimensional fNIRS data is presumed to lie on (or near) a lower-dimensional topological space with an intrinsically smaller number of degrees of freedom. In this sense, fNIRS data fits into the mathematical abstraction of a manifold. Isomap [11] is a nonlinear data embedding technique with the ability to learn a broad class of non-linear manifolds. This technique ensures global optimality and computational efficiency, and guarantees asymptotic convergence. Isomap has been shown to retain the intrinsic dimensionality of data with known structure, but has yet to be applied to complex fNIRS data where the underlying structure is unknown. The fact that Isomap is able to represent the global structure of a dataset within a single coordinate system is particularly useful for group-wise comparisons.

The three stages of the Isomap embedding [11] begin with the construction of a nearest neighbor graph in the -dimensional feature space F. This step can be considered as having two sub-steps: (a) Construct a weighted complete graph of the points in the experiment space with the weights being the distances between points i and j based on any metric. (b) Prune the complete graph, so that only a certain number of neighbors remain adjacent. The two standard approaches are to connect each point with its K nearest neighbors, i.e., keep the K edges with the smallest distances, or to connect all points within a radius ε, i.e. keep all edges for which . These neighborhood relations are represented in a weighted graph G over the data points, with edges of weight between neighboring points. In the second step, Isomap reconstructs the complete graph again. This time the weights are calculated as estimates of the geodesic distances between all pairs of points on the manifold M by computing their shortest path distances in graph G. There are a number of possible algorithms for computing the shortest paths in a weighted graph. The open-source version of Isomap [28] uses either Floyd's or Dijkstra algorithms to compute the shortest paths. The reconstructed complete graph can be represented by its distances matrix D_G, which contains the shortest path distances between all pairs of points in G. Finally, by applying classical multidimensional scaling (cMDS) to matrix D_G, Isomap constructs the embedding in an n-dimensional Euclidean space P that best preserves the intrinsic geometry. In cMDS, the projected output configuration is calculated by minimizing a cost function C in Equation (1) to optimize the new distances in P space.where D_P is the matrix of Euclidean distances [29].

A number of variations on the original Isomap algorithm have been developed to overcome intrinsic limitations such as noise [30], to enhance computational efficiency [26], and to enable time-series embedding [31] and multiple manifolding [32]. The FR-Isomap (fixed reference Isomap), described by Lekadir et al. [26], allows consistent dimensionality reduction across samples in serial datasets. The overall flow of FR-Isomap is represented in Figure 3. FR-Isomap allows the reduction of computational requirements by selecting a representative subset of points and performing embedding on this subset. It also allows the insertion of new points into the embedded space as they become available, avoiding execution of the embedding algorithm again. Although data from the two investigations have been acquired separately (as outlined above), FR-Isomap allows the projection of data from both experiments into the same space. FR-Isomap ensures consistent embedding, allowing comparisons between the results. Since both the experimental design and data processing were conducted with an identical protocol, data from both study (A) and study (B) can be considered to belong to the same feature space F.

Figure 3. Schematic illustration demonstrating the application and flow of the FR-Isomap data embedding technique.

FR-Isomap involves embedding of training data, which can efficiently represent the variability within large datasets. In this sense, it selects a number of representative points which act as reference points. The way in which FR-Isomap selects the reference points is flexible, as long as the reference subset is evenly distributed on the manifold. In the current analysis, we use study (A) as the reference data. Other non-reference data are mapped using the knowledge of the coordinates of both reference and non-reference data in feature space F and the coordinates of reference data in the embedded space.

With FR-Isomap, Sammon's nonlinear mapping criteria [33] can be used as a measure of the mapping error between the distances in the original and mapped spaces:By minimizing the mapping error, it is possible to calculate the coordinates in the embedded space for other points iteratively, as described by Lekadir et al. [26]:where are the coordinates in the embedded space in the m^th iteration, α is the step size, and E_s(m) is the mapping error after the m^th iteration (given in Equation (2)) which uses , the distance in the embedded space P after the m^th iteration. Since the reference points are evenly distributed on the manifold and the topology is preserved, the reference point which is the nearest neighbor to in feature space F can be used as a starting point for the iterative algorithm.

Earth Mover's Distance analysis

Once the original high-dimensional data is embedded into a common reference space, it is possible to quantify differences in cortical excitation patterns between groups. To this end, EMD is used. EMD is based on the transportation problem [27] and evaluates the dissimilarity between two multidimensional distributions. Each distribution being composed of a set of points. Each one with an associated weight w. This is referred to as the weighted signature of the distribution on the manifold. The EMD between two signatures is the minimum amount of “work” needed to transform one signature into another. The work required is the proportion of weight being moved multiplied by the distance between the old and new locations. Consider two distributions X = {(x₁, w_x1), (x₂, w_x2), …, (x_m, w_xm)} and Y = {(y₁, w_y1), (y₂, w_y2), …, (y_n, w_yn)}. It is possible to define a cost matrix C such that c_ij (known as ground distance) represents the cost of transforming a point i = 1 … m from distribution X into a point j = 1 … n from distribution Y.

Mathematically, EMD is defined as:where each f_ij is a flow representing the amount of weight moved from i to j.

EMD aims to find a flow set f = {f_ij} subject to the following constraints:that minimizes function .

Weights w_i were assigned inversely proportionally to the size of the distribution in number of points so that Σw_i = 1. We calculate the ground distances c_ij as the Euclidean distances in the embedded space P, i.e., . Since the ground distances are induced by an L₂ norm, EMD is a lower bounded metric [27].

Random effect model

We examined the effect of session (before training = 0; after training = 1) and channel (1–24) in the change of the dependent variables (ΔHbO₂, ΔHHb, ΔHbT) using a random effect model. The random effect model is the same as a standard multiple regression model, with the addition of cluster-level random effect u_j. The random effect analysis was conducted with Intercooled Stata (v8.0 for Windows, Stata Corporation, College Station, TX).

Results

Behavioral results

Statistical comparisons of dexterity data between novices and trained surgeons recruited for study (A) are shown in Table I. As expected, novices performed significantly worse than trained surgeons (consultants and registrars) on all performance parameters. No significant difference in performance was observed between consultants and registrars for any performance parameter. As illustrated in Table II, following a week of deliberate practice, a significant improvement in technical knot-tying ability was observed in study (B) novices.

Table I.  Comparison of knot-tying performance according to expertise. Data are presented as median [range]. Significant p-values ≤ 0.001 are in bold.

Parameter	Consultants	Registrars	P	Consultants	Novices	P	Registrars	Novices	P
Time (sec)	12.1 [7.0 to 23.7]	12.3 [7.9 to 23.6]	0.499	12.1 [7.0 to 23.7]	45.5 [21.0 to 149.0]	0.000	12.3 [7.9 to 23.6]	45.5 [21.0 to 149.0]	0.000
TPL	5.2 [2.9 to 9.3]	4.7 [1.9 to 10.4]	0.289	5.2 [2.9 to 9.3]	15.0 [5.5 to 12.6]	0.000	4.7 [1.9 to 10.4]	15.0 [5.5 to 12.6]	0.000
TM	13.0 [2.9 to 9.3]	13.0 [6.0 to 23.0]	0.566	13.0 [2.9 to 9.3]	29.0 [10.0 to 91.0]	0.000	13.0 [6.0 to 23.0]	29.0 [10.0 to 91.0]	0.000

Table II.  Comparison of knot-tying performance in medical students before and after a week of practice. Data are presented as median [range]. Significant p-values ≤ 0.001 (Wilcoxon Sign Rank test) are in bold.

Parameter	1st session	2nd session	P-value
Time	40.0 [18.5 to 257]	17.3 [8.9 to 32.5]	0.000
Total pathlength	9.3 [3.4 to 111]	4.5 [2.9 to 8.5]	0.000
Total movement	21.0 [7.0 to 51.2]	14.0 [9.0 to 20.0]	0.000

fNIRS prefrontal activation

FR-Isomap is capable of retaining the perceptually meaningful low-dimensional structure of the fNIRS dataset, as shown by the “elbow” on the graph of residual variance against dimensionality (Figure 4a). Additionally, the ability of FR-Isomap to resolve meaningful global coordinates is highlighted in Figure 4b, where each component of the embedding correlates well with one degree of freedom of the underlying data [left to right: HbO₂ (r² = 0.96); up and down: HHb (r² = 0.62)]. Average hemoglobin data in the high-dimensional space were correlated with coordinates in the low-dimensional space to provide r² values.

Figure 4. (a) Graph depicting residual variance of the fixed references training set in FR-Isomap for fNIRS data from expert and novice surgeons. (b) Diagram of the manifold embedding results for study (A), illustrating the distribution of channel responses and how it varies along the first two principal dimensions for all subjects studied. Locations of the original signals (a–f) are marked on the embedded space, demonstrating the intrinsic trend captured by the embedding technique. (Originally published in reference [12]. Reprinted with permission from Springer Science and Business Media.) (c) The intensity plot of change in total hemoglobin (HbT) demonstrates regions of the embedding most likely to represent brain activation. [Color version available online.]

Consultants and registrars clustered tightly around the origin of the two components in the embedded space P, whereas greater dispersion of data points was observed for novices in study (A). FR-Isomap places no a priori assumption regarding patterns most indicative of brain activation, yet is capable of producing a coordinate system in which classical patterns of activation (task-induced increases in HbO₂ and decreases in HHb) can be readily identified from the original input signals. In the current example, HbO₂ was observed to rise to the left, and HHb to fall toward the apex of the embedded space. Hemodynamic patterns consistent with brain activation may therefore be found toward the top left of the embedding, a region consisting primarily of novice data. The intensity map displayed in Figure 4c subdivides regions of the embedding according to the change in HbT. It can be seen that the greatest intensity change in HbT is located predominantly to the top left of the embedded space.

Regarding novice data from study (B), the arrangement of data in P appears less dispersed for the second versus the first session (i.e., following a week of deliberate practice). Figure 5 highlights the telescoping of data points between the first and second sessions. Interestingly, following a week of practice, novice data are mapped closer in proximity to the location of the experts, i.e., around the origin of the two components, a region of less CBF intensity change. In general, increasing expertise is associated with smaller cluster size in the embedded space (Figure 6). Figure 7a and 7b illustrate between-group EMD results. For study (A), greater similarity in the distribution of data points was observed between consultants and registrars than between novices and either of the former two groups. For study (B), the investigation of practice-dependent neuroplasticity, hemodynamic responses in the novice group were more similar (i.e., with lower EMD values) to those of trained surgeons following a week of practice at the task, and more dissimilar to the novices in study (A) who had not undergone further instruction and practice.

Figure 5. Manifold embedding results depicting practice-dependent reorganization of cortical behavior in surgical novices. Data represents 74-D trial HbO₂ and HHb values for a given channel and subject. Trajectories illustrate the transition between first-session data (pink crosses) and second-session data (green circles).

Figure 6. Using the k-means algorithm, the central point (centroid) for each group was located in the embedded space. The average distance between each centroid and all other points for that group was calculated (colored spheres), and the distance to the farthest point (cross) is shown as a solid line. [Color version available online.]

Figure 7. (a) Between-group Earth Mover's Distance (EMD) results for study (A); and (b) comparison between studies (A) and (B). [Color version available online.]

To determine whether this change in prefrontal hemodynamic responses was associated with reduced PFC activation, a random effect analysis was conducted to examine the effect of session (before training = 0; after training = 1) and channels (1–24) in the change of the dependent variables (ΔHbO₂, ΔHHb, ΔHbT), where, for a given channel, ΔHb = (ΔHb task-ΔHb rest). The regression output for the random effect model in Table III shows the estimated between-subject standard deviation (σ_u) and within-subject standard deviation (σ_e) for each measured outcome variable of interest. Statistically significant reductions in both ΔHbO₂ (coefficient = −4.6815, z = −6.90, P < 0.001) and ΔHbT (coefficient = −3.8252, z = −4.58, P < 0.001), and significant increases in ΔHHb (coefficient = 0.8192, z = 2.66, P < 0.01) were observed across practice sessions.

Table III.  Results of a random effect model for learning-related changes in (a) ΔHbO₂, (b) ΔHHb and (c) ΔHbT in surgical novices. s.e = standard error. Significant p-values < 0.01 are in bold, and significant values < 0.001 are in bold italics.

		Coefficient	s.e.	z	P > |z|	95% CI
(a) ΔHbO2
	Session	−4.6815	0.6781	−6.90	<0.001	−6.0101 to −3.3524
	Channel	−0.1199	0.0471	−2.54	<0.01	−0.2123 to −0.0275
	Constant	15.2682	1.7404	8.77	<0.001	11.8569 to 18.6794
	σu	5.3868
	σe	8.7670
	rho (fraction of variance due to uj)	0.27407
(b) ΔHHb
	Session	0.8192	0.3074	2.66	<0.01	0.2167 to 1.4217
	Channel	−0.0728	0.0215	−3.39	<0.001	−0.1148 to −0.0307
	Constant	1.4402	0.6783	2.12	0.034	0.1109 to 2.7696
	σu	1.6899
	σe	3.9842
	rho (fraction of variance due to uj)	0.1525
(c) ΔHbT
	Session	−3.8252	0.8353	−4.58	<0.001	−5.4624 to −2.1881
	Channel	−0.1924	0.5813	−3.31	<0.001	−0.3064 to −0.0785
	Constant	16.6534	2.0240	8.23	<0.001	12.6863 to 20.6204
	σu	5.8764
	σe	10.8153
	rho (fraction of variance due to uj)	0.2279

Discussion

In this paper, a novel approach to understanding and quantifying complex high-dimensional fNIRS data has been proposed. It has been shown that the intrinsic dimensionality of the data is recoverable by FR-Isomap based upon the patterns of chromophore change. This technique places no a priori assumptions upon the patterns most representative of cortical activation. However, the arrangement of data in the embedded space is intuitively meaningful, enabling typical hemodynamic patterns of brain activation to be easily identified. As well as enhancing visualization, differences in hemodynamic behavior can be easily quantified in the low-dimensional space. The technique enables simultaneous analysis of both hemoglobin species, reducing the likelihood of a Type I error, by assigning more weight to one or other hemoglobin species.

Using this approach, we have demonstrated that prefrontal hemodynamic responses vary depending upon technical expertise in the task. Similar prefrontal hemodynamic responses have been demonstrated in registrars and consultants, in whom technical performance could not be discriminated. In contrast, distinctly different patterns of cortical hemodynamics and significantly poorer knot-tying performance were observed in surgical novices. The location of novice data in the embedded space strongly suggests that these findings are related to differences in the degree of prefrontal brain activation required to perform the task.

Interestingly, we observed an evolving pattern of prefrontal hemodynamic behavior in medical students following more detailed instruction on the task. This transition in cortical behavior is associated with significant longitudinal refinements in technical knot-tying performance. Patterns of prefrontal hemodynamic behavior more closely resemble the distribution of trained surgeons following a week of deliberate practice. The results of the random effect model strongly support a reduction in PFC activation across sessions, with significantly less ΔHbO₂ and ΔHbT following extended practice. These findings support the idea that prefrontal plasticity accompanies the acquisition of a surgically relevant task, and are in line with studies of visuomotor learning [15], [34], [35], force field learning [36], pursuit rotor [37], bimanual coordination [16], [17] and non-surgical knot-tying [38], all of which demonstrate the importance of PFC activation in early learning but not after practice. One theory is that a set of “scaffolding” regions are recruited to support novel demands during unrefined performance [39]. These regions include areas of generic attention and control, such as the PFC and ACC. As motor skills are performed with increasing fluidity, activation foci are relegated from rostral regions, resulting in a redistribution of the activation map. One might link this to the scaffold falling away [13], [39].

The shift in cortical hemodynamic behavior observed across extended practice in the current analysis may be explained by neurophysiological “pruning” of attentional and control areas common to most practice-related studies [13]. “Attention to action” was likely to be prominent during initial trials, a factor known to lead to enhanced prefrontal activation compared to unattended performance [40], [41]. Alternatively, the results may be attributed to distinct cognitive functions of the PFC that are important in the novel phases of procedural learning, such as visuospatial working memory [42], error detection [43] or response monitoring. It should be noted that the current study was not designed to discriminate between these processes. Instead, we hope to characterize expertise and learning-related changes in PFC behavior using an embedding framework with much wider neuroimaging applications. Our results demonstrate that the prefrontal hemodynamic response to open surgical knot-tying is highly dependent upon technical ability. Following extended training, surgical novices execute the task more efficiently with an evolving pattern of prefrontal activation that more closely resembles that of trained surgeons. It is likely that novices have progressed from a highly attention-demanding early learning phase, where online monitoring confers performance gains, to a stage where knot-tying can be performed with reduced attentional demands.

Limitations of fNIRS technology

The results presented in this paper show a promising application of fNIRS for assessing functional prefrontal reorganization associated with learning. However, it should be noted that there are certain limitations associated with fNIRS technology, such as the need to overcome the effect of attenuation of light owing to intense scatter of photons in biological tissues. The spatial resolution is often no better than the source-detector separation, typically no more than 30 mm, although the resolution can be improved with the use of overlapping measurements from neighboring optodes [44]. Finally, fNIRS signals are limited by the depth of penetration of the infrared light through the skull and are dampened by nuisance tissues (scalp, hair follicles, skull, etc.), such that measurements are limited to a banana-shaped volume on the cortical surface. Technological advances, including time-domain spectroscopy to improve penetration depth, may assist applications designed to advance understanding of more complex visuomotor behaviors.

Acknowledgments

The authors would like to thank Dr. Karim Lekadir for his support with the implementation of the FR-Isomap algorithm and Mr. James Kinross for his help in creating the accompanying illustrations. The authors extend their thanks to their collaborators at the Biomedical Research Optics Laboratory (BORL), University College London, Dr Peck Hui Koh, Dr Clare E. Elwell and Professor David T. Delpy, for their ongoing advice and support. We are grateful to all the surgeons and medical students who participated in this series of investigations.