Inside a Programmer's Brain: Dissociating Code and Language Comprehension Networks Using a Novel fMRI‐ESM Protocol

ABSTRACT

Code comprehension and natural language share structural similarities but engage partially distinct neural systems. While language predominantly recruits the classical perisylvian language network, code comprehension has been linked to the multiple‐demand system, associated with mathematical thinking and problem solving. We report the case of a professional software engineer with an IDH‐mutant astrocytoma in the left middle frontal gyrus, a region where these networks converge. To address the potential risk of impairing computer programming abilities during surgery, we implemented a personalized protocol combining fMRI and intraoperative electrical stimulation mapping (ESM), extending the standard language assessment with a Code comprehension task. Intraoperative mapping identified sites in the left inferior frontal cortex that were functionally relevant to the C ode comprehension task but were not identified during standard object‐naming mapping. Complete tumor resection was achieved while preserving these critical regions, with the patient showing intact post‐operative language and programming performances. This single case study suggests that individualized Code comprehension mapping may identify functionally relevant sites not detected by standard object‐naming mapping, supporting the potential clinical value of tailored fMRI–ESM protocols for preserving higher‐order cognitive functions in neurosurgical patients.

Personalized intraoperative mapping using programming tasks enabled the identification and preservation of code‐related eloquent cortex, supporting postoperative recovery of professional programming abilities.

1. Introduction

Maximal safe resection of gliomas located in eloquent brain regions requires balancing oncological benefit with the preservation of functions essential for patients' autonomy and quality of life. Awake surgery with intraoperative electrical stimulation mapping (ESM) is the gold‐standard approach for identifying and preserving sensorimotor and language functions (Bu et al. 2021; Collée et al. 2023; Hamer et al. 2012; G. Ojemann et al. 2008; G. Ojemann and Mateer 1979; Penfield 1958). However, standard mapping protocols may not fully capture higher‐order cognitive abilities relevant to an individual patient's daily life or occupation (Duffau 2021; Ritaccio et al. 2018; Riva et al. 2016; Rofes and Miceli 2014). This has motivated growing interest in designing personalized functional mapping protocols aimed at preserving specific cognitive abilities with direct impact on the patient's postoperative recovery (Cepero‐Escribano et al. 2024; Duffau 2021; Duffau et al. 2008; Gomez‐Andres et al. 2022; Hamberger 2011; Martín‐Fernández et al. 2025; G. A. Ojemann 1983; Penfield and Roberts 2014; Puglisi et al. 2019; Riva et al. 2016; Rofes et al. 2017; Rojas et al. 2021; Sierpowska et al. 2015, 2018; So and Alwaki 2018; Wager et al. 2013). This body of work illustrates that task‐specific eloquent sites may not always be captured by standard language or motor mapping, reinforcing the need for individualized protocols.

In this study, we present the case of a professional software engineer with an IDH‐mutant astrocytoma located in the left middle frontal gyrus [MFG, Brodmann Area (BA) 6] for whom preserving programming ability was directly relevant to postoperative quality of life and professional career. We therefore designed a personalized programming‐specific protocol targeting a crucial component of programming: code comprehension.

Code comprehension refers to the cognitive process of understanding the semantics and execution flow of a computer program to construct an accurate mental model of its behavior. In experimental settings, code comprehension is commonly operationalized as the ability to correctly predict the output, state changes, or behavior produced by a code snippet (Ivanova et al. 2020; Peitek et al. 2020, 2021; Siegmund et al. 2017; Siegmund, Kästner, Apel, et al. 2014). This component was selected as the primary target because it is central to understanding the logic, structure, and intent of source code, thereby supporting efficient and effective problem solving (LaToza et al. 2006; Lientz et al. 1978; Peitek et al. 2018; Standish 1984). Following Peitek et al. (2020) paradigm, we also implemented a Syntax error detection task, which is similar to the Code comprehension task but does not require the understanding of the program. Together, these tasks were intended to approximate the patient's profession‐specific cognitive demands in a clinically meaningful way.

Importantly, the relevance of code comprehension extends beyond its practical role in software development. From a cognitive perspective, it constitutes a complex higher‐order process that recruits multiple interacting functions. Code comprehension is not equivalent to natural language reading or simple visual inspection of source code. Rather, it requires the integration of multiple cognitive operations, including abstract reasoning, performance monitoring (control data flow), working memory, semantic retrieval, silent reading, problem solving, and inferring the program's intended behavior (Lientz et al. 1978; Peitek et al. 2018; Renumol et al. 2010; Siegmund, Kästner, Apel, et al. 2014; Standish 1984; Wang et al. 2006).

Neuroimaging studies further support this view as previous fMRI (Floyd et al. 2017; Ikutani et al. 2021; Ivanova et al. 2020; Krueger et al. 2020; Liu et al. 2020; Peitek et al. 2020, 2021; Sharafi et al. 2021; Siegmund, Kästner, Apel, et al. 2014), EEG (Crk et al. 2016; Kosti et al. 2018; Lee et al. 2016), and fNIRS studies (Ikutani and Uwano 2014; Nakagawa et al. 2014; Sharafi et al. 2021) have associated code comprehension with activity in frontal, temporal, and parietal regions, including the MFG (BA6), inferior frontal gyrus (IFG, BAs 44 & 47), middle temporal gyrus (MTG, BA 21), and inferior parietal lobule (IPL, BA 40) (Siegmund, Kästner, Apel, et al. 2014). This cognitive and anatomical profile has often been interpreted as reflecting recruitment of the multiple‐demand system, a frontoparietal network involved in executive control, reasoning, mathematical processing, and problem solving (Amalric and Dehaene 2019; Assem et al. 2020; Duncan 2010; Fedorenko et al. 2013; Shashidhara et al. 2024). At the same time, code comprehension may partially overlap with language‐related regions, particularly in the left inferior frontal cortex (Fedorenko et al. 2013; Floyd et al. 2017; Ivanova et al. 2020; Liu et al. 2020; Sharafi et al. 2021).

This issue was particularly relevant in the present case because the tumor was located in the left MFG with a minor extension toward the pars opercularis (BA 44) and triangularis (BA 45) of the IFG. These regions lie near cortical territories implicated in language production (Broca's Area; Crank and Fox 2002; Flinker et al. 2015; Tate et al. 2014) and code comprehension (B. Becker et al. 2013; M. Becker et al. 2020; Friederici 2015; Homae et al. 2002; Ishkhanyan et al. 2020; Nissim et al. 2017; Rivas‐Fernández et al. 2021; Schuhmann et al. 2012; Siegmund, Kästner, Apel, et al. 2014; Villar‐Rodríguez et al. 2024). Moreover, previous studies suggest that the neural distinction between code comprehension and language processing may be reduced in experienced programmers, together with a lower cognitive load during code‐related tasks (Crk et al. 2016; Floyd et al. 2017; Ikutani et al. 2021; Lee et al. 2016; Peitek et al. 2020; Sharafi et al. 2021; Siegmund et al. 2017; Siegmund, Kästner, Liebig, et al. 2014). Therefore, while standard language mapping could provide clinically relevant information, it might be insufficient to fully characterize the patient's programming‐relevant functional anatomy. This anatomical and occupational context motivated the integration of preoperative fMRI, neuropsychological assessment, and intraoperative ESM using programming‐specific tasks.

Here, we report a clinical proof‐of‐concept demonstrating the feasibility and utility of individualized fMRI–ESM mapping for preserving a profession‐specific higher cognitive function during awake glioma surgery. Specifically, we developed a personalized programming‐specific protocol for a professional software engineer undergoing resection of a left frontal glioma. We aimed to determine whether a programming‐specific task could identify functionally relevant cortical and subcortical sites not captured by standard language mapping alone, and whether this information could guide surgical decision‐making while supporting postoperative return to work. This case illustrates how individualized mapping protocols may extend the clinical scope of awake surgery beyond standard language and sensorimotor domains toward the preservation of patient‐specific cognitive abilities.

2. Methods

2.1. Clinical Information

The patient was a 41‐years‐old, right‐handed male diagnosed with an Astrocytoma, IDH‐mutant; grade 2 (WHO 2021) (Louis et al. 2021) in the left MFG with a minor extension to the pars opercularis (BA44) and triangularis (BA45) (Figure 1). Due to the tumor's location and considering the potential cognitive sequelae, particularly in language, programming performance, and overall quality of life (Friederici 2015; Homae et al. 2002; Villar‐Rodríguez et al. 2024), the patient was selected for an awake ESM protocol including both programming and language assessment. The patient was a trilingual speaker (L1: Spanish, L2: Catalan, L3: English), and preoperative fMRI showed bilateral language lateralization, with language‐related activation involving both hemispheres. Additionally, he had been working as a computer engineer for 18 years and he expressed specific concern regarding potential postoperative impairment of his programming abilities, which could impair his daily functioning and quality of life. In response, we designed a tailored multimodal protocol combining fMRI and ESM to localize eloquent areas for code comprehension and, therefore, preserve their function after the surgery.

FIGURE 1.

Preoperative T1‐weighted MRI showing the location of the tumor in the left inferior frontal cortex.

The study was approved by the Ethics Committee of the University Hospital of Bellvitge (ref PR075/19) and conducted in accordance with the Declaration of Helsinki. Written informed consent was obtained from the participant. The conditions of our ethics approval do not permit public archiving of anonymized study data. Readers seeking access to the data should contact the lead author (XC‐C or AR‐F). Access may be granted to named individuals upon reasonable request and subject to approval in accordance with ethical procedures governing the reuse of sensitive data. Scripts for stimuli presentation and data analysis are available at https://osf.io/db4sn/?view_only=63e37d4b094f492aaa9a96004f1b479a. We report how we determined our sample size, all data exclusions, all inclusion/exclusion criteria, whether inclusion/exclusion criteria were established prior to data analysis, all manipulations, and all measures in the study. No part of the study procedures or analysis plans was pre‐registered prior to the research being conducted.

2.2. Neuropsychological Assessment

A comprehensive neuropsychological evaluation was performed at baseline (4 weeks before surgery), during the acute (1 week after surgery) and the early postoperative stage (3 months after surgery). Neuropsychological assessment included evaluations of executive functions, language, and memory (for further details see Gasa‐Roqué et al. (2024), Sierpowska et al. (2018) and Section S1).

2.3. Behavioral Protocol

The patient underwent behavioral programming evaluation at two preoperative baseline sessions, the acute, and the early postoperative stages. Similarly to previous studies (Kosti et al. 2018; Peitek et al. 2018, 2020; Siegmund, Kästner, Apel, et al. 2014), the evaluation included two different tasks: Code comprehension and Syntax error detection (see designs in Figure 2). The Code comprehension task was administered at both baseline sessions, whereas the Syntax error detection task was administered only at the first preoperative baseline. Both tasks were administered using a similar procedure: a fixation cross was followed by a short snippet of code with a question at the bottom of the screen. A two‐alternative forced choice answer (yes/no) was presented, and no feedback about task performance was provided. Code production was not evaluated due to intraoperative constraints and the potential introduction of motor confounds, which could have produced false positives unrelated to higher‐level programming cognition.

FIGURE 2.

Behavioral assessment of programming. The task at the top evaluates the performance during code comprehension, while the task at the bottom assesses the ability to detect syntax errors.

First, we administered the Code comprehension task, where the question asked for the value of the variable “result” (e.g., Q: result == “Hello world”? A: yes/no, see Figure 2). Then, the participant performed the Syntax error detection task, which asked whether there was a syntax error in the code snippet (e.g., Q: Is there any error? A: yes/no, see Figure 2). The Code comprehension and Syntax error detection conditions were derived from the same underlying Java snippet templates and matched in visual format, length, and programming constructs. The main difference between tasks was the cognitive operation required: code understanding and output prediction in the Code comprehension task versus detection of syntactic violations in the Syntax error detection task, which does not require understanding the code. This design aimed to minimize confounds related to visual layout, code format, or general lexical properties of the stimuli. In both tasks, the patient was instructed to answer correctly as fast as possible, and performance and reaction times (RTs) were captured for every item.

In total, 48 Java snippet templates were used (Table 1), each instantiated in four variations by changing variable values while preserving the underlying code structure. For Code comprehension, each variation had one correct and one incorrect output question, with one selected at random; thus, “yes” and “no” responses were not strictly balanced. For Syntax error detection, the same variations were used to create one syntactically correct and one incorrect version, the latter containing a minor Java syntax alteration, such as an incorrect delimiter. One version was selected at random, so correct and incorrect trials were likewise not strictly balanced (see Figure S1).

TABLE 1.

General programming constructs used to build the code stimuli. For each construct, the table shows the specific cases implemented in the task. Each specific case included one or more code examples.

General construct	Particular case
Flux control	If
	If/else
	If/else if/else if/else
	Switch/case
	Switch/case/default
	Switch/case (no break)
Loops	For (arithmetical operation)
	For (string concatenation)
	Do while
	While
	While (infinite)
Operators	Array access
	Array clone
	Array copy
	Array sum
	Casting
	Increment
	Comparisons
	String comparison
	String insertion
	String copy
	String find
	String count
	String sum
	Method call
Strings	Concatenation
	Replacement

These snippets included several programming concepts and structures such as variable assignment, mathematical operations, conditionals, loops, and simple functions, and they were inspired by the examples provided in a programming textbook (Eckel 2006). The code stimuli were designed and implemented by X.O. and X.C.C., both professors of Computer Science. In addition, their format, length, and style were modeled after the code snippets used by Peitek et al. (2020). Their adequacy for the experimental task was verified through expert review by X.O. and X.C.‐C., who checked their syntactic correctness, conceptual clarity, and consistency in structure and difficulty across snippets.

To familiarize the patient with the paradigm, the Code comprehension task was administered twice at baseline. The observed accuracy and RTs at baseline 2 confirmed that the task was suitable for intraoperative application, justifying its inclusion in the protocol. All sessions were supervised by A.G.‐R., a neuropsychologist from the University Hospital of Bellvitge. The same setup was employed during the acute evaluation and early postoperative stages. Task instructions were provided both orally and in written form at the beginning of each session, and the patient confirmed full comprehension.

2.4. Behavioral Data Analysis

To assess longitudinal changes in programming‐task performance, accuracy and response times were analyzed separately for the Code comprehension and Syntax error detection tasks because they included different session structures. For Code comprehension, Session comprised four levels: baseline 1, baseline 2, acute postoperative, and early postoperative, with baseline 1 as the reference level. For Syntax error detection, Session comprised three levels: baseline, acute postoperative, and early postoperative, with baseline as the reference level.

For accuracy, generalized linear mixed‐effects models (GLMMs) with a binomial error distribution and logit link function were used. Trial‐level accuracy was coded as a binary outcome, with incorrect responses coded as 0 and correct responses coded as 1. Because each trial therefore had only two possible outcomes, accuracy was modeled using a Bernoulli distribution as a function of Session:

$${\text{Accuracy}}_{\mathit{ij}}~\text{Bernoulli}\left(p_{\mathit{ij}}\right)$$

$$\text{logit}\left(p_{\mathit{ij}}\right)={\beta }_0+{\beta }_1{\text{Session}}_{\mathit{ij}}+u_{0j}$$

where $p_{\mathit{ij}}$ is the probability that trial $i$ for item $j$ was answered correctly, ${\beta }_0$ is the intercept, ${\beta }_1$ represents the fixed‐effect coefficients for Session, and $u_{0j}$ is the random intercept for item.

For response times, linear mixed‐effects models (LMMs) were used, with RT as the dependent variable and Session as a fixed effect:

$$R{T}_{\mathit{ij}}={\beta }_0+{\beta }_1{\text{Session}}_{\mathit{ij}}+u_{0j}+{ϵ}_{\mathit{ij}}$$

where $u_{0j}$ represents item‐level random variability and ${ϵ}_{\mathit{ij}}$ is the residual error term. Items were included as random intercepts to account for variability.

Omnibus effects of Session were assessed using Type‐II Wald chi‐square tests. The χ² statistics reported in the Results correspond to these omnibus tests. In Tables (S2–S5), z statistics correspond to Wald tests of individual GLMM coefficients, whereas t statistics correspond to individual LMM coefficients relative to the reference session. These coefficient‐level p‐values were not adjusted for multiple comparisons. Pairwise comparisons between session levels were additionally conducted using estimated marginal means, with p‐values adjusted using the false discovery rate procedure (p_FDR). All statistical analyses were conducted in R version 4.3.0.

2.5. fMRI Protocol

At baseline, a neuroimaging protocol was designed, including structural (T1‐weighted and Diffusion Tensor Imaging) and functional MRI scans (See Section S2). Structural MRI scans were acquired to visualize the tumor location and the surrounding edema (Figure 1). In addition, a programming fMRI assessment was performed to evaluate the spatial overlap between the lesion and functional regions. This information served as an additional reference for preoperative functional mapping and allowed us to predict to what extent the resection of the tumor and perilesional tissue might involve functional areas and hence lead to relevant impairments.

The programming fMRI experiment employed a block design with four conditions (see Figure 3): two programming‐related conditions (Code comprehension and Syntax error detection), one Language comprehension condition, and a Resting state condition as baseline.

FIGURE 3.

fMRI block‐design protocol for programming‐related task mapping. Three active task conditions were presented using a color‐coded frame: Blue for Code comprehension, red for Syntax error detection, and yellow for Language comprehension. A fixation cross on a black background was used as the Resting state condition. Each block lasted 30 s. During the first 25 s, the participant viewed the stimulus without the response prompt. During the final 5 s, the same stimulus remained on screen and a response prompt appeared, indicated in the figure by the dashed‐line box. In the Code comprehension condition, participants judged whether a proposed code output was correct; in the Syntax Error Detection condition, they judged whether the code contained a syntax error; and in the Language comprehension condition, they judged whether the proposed response to the written‐language problem was correct. Blocks were presented in pseudorandomized series of four conditions, with the Resting state block always presented first. In total, 17 series of four blocks were administered.

To assess programming abilities, we focused on the patient's capacity for Code comprehension. Similarly to the behavioral design, in this condition, the patient was asked to read a code snippet and determine whether the given output was correct. As a control, the Syntax error detection condition involved identifying syntactic violations (e.g., punctuation, order of elements) without requiring semantic understanding of the code. This contrast allowed us to isolate neural activity specifically related to code comprehension (Fedorenko et al. 2019; Ivanova et al. 2020).

To examine the potential overlap between code and language processing networks, a Language comprehension condition was included in which the patient followed structured instructions involving string manipulation and conditional logic (e.g., selecting specific letters from a word and applying rules to decide what to print). This task emphasized step‐by‐step reasoning and outcome evaluation, functionally resembling Code comprehension (Ivanova et al. 2020). Finally, a Resting state condition served as baseline.

Each block lasted 30 s. During the first 25 s, the stimulus was presented without a response prompt. During the final 5 s, the same stimulus remained on screen and a response prompt appeared. In the Code comprehension condition, the patient judged whether a proposed code output was correct; in the Syntax error detection condition, he judged whether the code contained a syntax error; and in the Language comprehension condition, he judged whether the proposed response to the written‐language problem was correct. Active task conditions were visually indicated by a color‐coded frame surrounding the stimulus: blue for Code comprehension, red for Syntax error detection, and yellow for Language comprehension. A fixation cross on a black background was used for the Resting state condition. Conditions were presented in pseudorandomized series of four blocks, with Resting state always presented first in each series. Each condition was repeated 17 times, resulting in a total scan duration of 2040 s (34 min). Importantly, the pseudorandom order was designed to maximize the interval between presenting snippets for Code comprehension and Syntax error detection, with a preference for displaying them first in the Code comprehension condition (Peitek et al. 2018).

2.6. fMRI Data Analysis

We established the following contrasts of interest: Code comprehension > Resting state, Code comprehension > Syntax error detection, Code comprehension > Language comprehension, and Language comprehension > Code comprehension. No spatial normalization was performed in the fMRI analysis, as the study involved a single subject, and pathological tissue could alter neurovascular coupling [particularly near gliomas and meningiomas (Krings et al. 2001)], potentially impairing normalization accuracy (see all details in Section S2).

Finally, we compared the results in the contrast Code comprehension > Language comprehension with the group results in a comparable contrast reported by Ivanova and colleagues (Ivanova et al. 2020).

2.7. ESM Protocol

Awake brain surgery was performed by senior neurosurgeons (A.G. and A.F.‐C.) at the Department of Neurosurgery of University Hospital of Bellvitge, following established procedures (Fernández‐Coello et al. 2016). A left fronto‐temporal craniotomy was carried out, and a baseline evaluation for each task was performed before stimulation. ESM was performed using an Ojemann cortical stimulator (Radionics, INC.). Stimulation consisted of biphasic current pulses delivered at 60 Hz, with a single‐pulse phase duration of 1 ms. The whole exposed surface was tested. Electrical stimulation was first performed over the primary motor cortex, starting at 1 mA while the patient performed an overt counting task. If no disruption of speech production or motor response was observed, the stimulation intensity was increased in 0.5 mA increments until a reproducible response was elicited. In this case, stimulation of the opercular region of the motor strip induced speech arrest, likely reflecting interference with the motor components of speech production involving the facial, tongue, and laryngeal musculature. Sensorimotor mapping was subsequently performed at 2.2 mA, and language and c ode comprehension mapping were performed at 3 mA. Each stimulation train lasted 3 s.

Language mapping was then performed using an object naming task at the cortical level. Before stimulation, a language baseline was obtained, and only items correctly named within 3 s were subsequently used during mapping.

Code comprehension mapping was performed next. A task‐specific baseline was first administered, and only items consistently answered correctly within 7 s during this baseline and immediately before resection in the operating room were used for stimulation mapping. For this task, the stimulation train was extended to 5 s based on the patient's response time. Task stimuli were presented on a laptop, and real‐time performance feedback was displayed on a second monitor for the neurosurgeon.

Finally, a baseline assessment of the Syntax error detection task was attempted. However, this task was not administered during intraoperative mapping because the patient showed an error rate > 30% during baseline. He reported that the task required excessive detail processing, making it unfeasible in the intraoperative setting.

Code comprehension was also administered during subcortical mapping throughout resection with the same item set as in cortical mapping to minimize task difficulty and reduce the likelihood of attributing item‐related errors to stimulation. Subcortical stimulation was performed using the CUSA stimulation probe.

For both language and Code comprehension mapping, a site was classified as functionally relevant if the patient failed to complete the task in at least two out of three non‐consecutive stimulations (Fernandez‐Coello et al. 2013; So and Alwaki 2018).

Eloquent regions were identified by the neurosurgeons, and tags were placed on the corresponding cortical sites. After surgery, these tags were co‐localized with fMRI/MRI data for visualization (see below). The whole procedure was video recorded.

The programming‐specific intraoperative protocol lasted 25 min 52 s, including 6 min 31 s of intraoperative training and 19 min 21 s of cortical mapping. Subcortical mapping was integrated into the approximately one‐hour resection and was emphasized when approaching functionally relevant white matter pathways.

2.8. ESM‐fMRI Anatomical Matching

Preoperative anatomical and functional images were co‐registered in the native space before the fMRI–ESM comparison. Therefore, fMRI activations could be directly visualized over the corresponding anatomical MRI. For the localization of intraoperative ESM sites on the preoperative anatomical volume, we used the 3D rendering visualization in MRIcroGL (Rorden 2025), using the contrast‐enhanced T1‐weighted image to identify cortical anatomy and superficial venous landmarks. ESM sites observed in the surgical field were manually matched to the best corresponding anatomical locations on the 3D‐rendered brain surface by comparing gyri, sulci, and venous patterns. Once the corresponding region was identified, the location was selected in MRIcroGL, and the voxel coordinates displayed by the software were used to create point markers at those exact coordinates. These points were then used to visualize the spatial relationship between ESM sites and fMRI activations. The resulting matching was confirmed by a multidisciplinary expert panel from the Functional Neuro‐Oncology team at the University Hospital of Bellvitge.

3. Results

3.1. Neuropsychological Assessment

At baseline, the patient showed mild impairment on the object naming task in L2, along with low scores in L3, probably due to limited language proficiency rather than language impairment. Additionally, the patient showed an impairment in immediate verbal memory recall. Executive functions and the other evaluated language domains were within the normal range. After surgery, acute evaluation revealed a decline in mental flexibility, verbal fluency, and verbal memory (learning and delayed recall), with an improvement in the immediate recall. At the early‐stage postoperative evaluation, impaired scores at the acute stage improved, with a cognitive and language profile similar to that observed at baseline. At this stage, only mild difficulties in verbal comprehension were reported, mainly associated with the most complex items. This pattern suggests an executive function effect rather than a true verbal comprehension deficit per se, which should be considered in conjunction with the rest of the executive tasks (preserved). All cognitive and language scores are detailed in Table S1.

3.2. Behavioral Evaluation

No significant omnibus effect of Session on accuracy was detected for either Code comprehension (χ ²(3) = 6.46, p = 0.091) or Syntax error detection (χ ²(2) = 1.00, p = 0.608; Figure 4A). Session significantly affected response times in both tasks (Code comprehension: χ ²(3) = 176.74, p < 0.001; Syntax error detection: χ ²(2) = 6.44, p = 0.040; Figure 4B), with shorter response times in later sessions. However, this reduction should be interpreted cautiously because repeated exposure to the tasks may have introduced practice effects.

FIGURE 4.

Behavioral results across sessions. For Code comprehension, baseline 1 and baseline 2 were administered eight and four days before surgery, respectively. For Syntax error detection, only one preoperative baseline assessment was administered, corresponding to baseline 1 of the Code comprehension protocol. Acute and early‐stage postoperative sessions were done one week and three months after surgery, respectively. A. Task performance across sessions for the two behavioral tasks: Code comprehension (red) and Syntax error detection (blue). B. Average reaction time across sessions. Error bars indicate standard errors.

FDR‐adjusted pairwise comparisons (see Table S7) showed that, for Code comprehension, response times were significantly shorter at baseline 2, acute postoperative, and early postoperative sessions than at baseline 1 (all p_FDR < 0.001). Response times were also shorter at the acute postoperative session than at baseline 2 (p_FDR < 0.001) and at the early postoperative session than at baseline 2 (p_FDR = 0.019). The difference between the acute and early postoperative sessions was not significant (p_FDR = 0.068).

For Syntax error detection (see Table S8), response times were significantly shorter at the early postoperative session than at baseline (p_FDR = 0.039), whereas the baseline–acute postoperative and acute–early postoperative comparisons were not significant (p_FDR = 0.209 and 0.316, respectively).

No statistically significant postoperative decline in accuracy was detected, although equivalence was not formally tested. Considered together with the neuropsychological assessment, clinical evaluation, and the patient's return to professional programming, the overall findings were consistent with preserved postoperative programming performance (see Table S1).

3.3. fMRI Evaluation

When contrasted to Resting state, Code comprehension elicited bilateral frontal activations (Figure 5A), with prominent clusters in the right MFG and superior segments of the IFG. In the left hemisphere, activation was confined to perilesional frontal regions, including anterior IFG and dorsal MFG near the superior frontal gyrus. Additional bilateral activity was detected in the fusiform gyri, superior cerebellum, supplementary motor area, and the right dorsomedial thalamus.

FIGURE 5.

Brain regions involved in Code comprehension. (A) Code comprehension > Resting state t‐contrasts. (B) Code comprehension > Syntax error detection t‐contrast. (C) Code comprehension > Language comprehension. (D) Language comprehension > Code comprehension. Regions belonging to significant clusters are labeled. BA, Brodmann area; dmT, dorsomedial thalamus; FFG, fusiform gyrus; IFG, inferior frontal gyrus; IPL, inferior parietal lobule; LG, lingual gyrus; MFG, middle frontal gyrus; MTG, middle temporal gyrus; PCC, posterior cingulate cortex; rMOG, right middle occipital gyrus; SMA, supplementary motor area; SPL, superior parietal lobule; STS, superior temporal sulcus; STG, superior temporal gyrus; TPJ, temporo‐parietal junction.

A similar pattern of bilateral activation was observed for the contrast Code comprehension > Syntax error detection, particularly in the right IFG and MFG, and in the left MFG near the precentral gyrus (Figure 5B). Enhanced activation was also evident in superior parietal regions, especially the precuneus, bilaterally, but more pronounced in the left hemisphere.

Compared with Language comprehension, Code comprehension produced only limited frontal activation, restricted to a small cluster in the left MFG. However, it elicited stronger and more extensive activity in bilateral parietal regions including the precuneus and inferior parietal lobules (IPL), as well as in bilateral middle temporal gyri. Right hemisphere activations include right posterior cingulate cortex, right middle occipital and lingual gyri (Figure 5C; for details, see Section S3 and Table S6).

Language comprehension > Code comprehension showed bilateral activation in the IFG, superior temporal sulci (STS), middle temporal gyri (MTG) and temporo‐parietal junction (TPJ) (Figure 5D).

Lastly, to evaluate the overlap of our results with prior group‐level findings, we compared the patient's activation pattern with group‐level data from Ivanova et al. (2020), who examined a similar contrast using Python‐based code tasks. Overlaying their published t‐map (Code comprehension > Language comprehension) with our subject's normalized activation map revealed substantial overlap in bilateral superior parietal regions (purple regions in Figure 6), including the precuneus, as well as in the middle temporal and lingual gyri.

FIGURE 6.

Code comprehension > Language comprehension t‐contrast overlaid on a comparable, group‐level t‐contrast from Ivanova et al. (2020).

3.4. ESM Evaluation

The whole exposed area was evaluated through intraoperative ESM. In the motor cortex, we identified two eloquent sites (red tags in Figure 7) in the frontal operculum, where stimulation induced speech arrest. An additional somatosensory point was located in the parietal operculum, eliciting right‐sided tongue dysesthesia (blue tag in Figure 7). No language eloquent sites were found during language mapping despite consistent fMRI activation in the pars triangularis during the language assessment in L1, 2 and 3.

FIGURE 7.

Electrical stimulation mapping results and fMRI–ESM overlay. (A) Schematic representation of the ESM results on a brain template. (B) Intraoperative view before tumor resection showing the cortical ESM sites. Single‐error points indicate stimulation sites that elicited an error but did not meet the criterion for a positive eloquent site, defined as at least two errors across three stimulations. (C) ESM functional sites overlaid on the fMRI activation map for the Code comprehension > Language comprehension contrast, shown at an exploratory uncorrected threshold of p < 0.001 for visualization purposes. The programming‐positive ESM sites overlapped with regions showing greater activation for Code comprehension than for Language comprehension. The approximate craniotomy area is indicated by the light green outline. (D) Intraoperative view after tumor resection, showing preservation of the eloquent sites.

During the Code comprehension mapping, a total of 126 Code comprehension items were administered, yielding 104 correct responses, 16 incorrect responses, and 6 no responses. By contrast, it revealed two eloquent sites (green tags in Figure 7). The first (green tag #1) was located in the anterior left IFG, near the pars triangularis (BA45), near the tumor margin. The second (green tag #2) was found at the junction of the IFG and MFG, near the inferior frontal sulcus, corresponding most likely to BA46.

Furthermore, eloquent sites were revealed during the Code comprehension subcortical mapping. Specifically, positive responses were observed when approaching the terminal fibers of the arcuate fasciculus, consisting of a transient inability to perform the task during stimulation. These fibers were therefore considered functionally relevant and preserved, although no physical tags were placed at subcortical sites for later spatial registration. During this phase, 108 items were administered, yielding 97 correct responses, 9 incorrect responses, and 2 no responses. The patient remained collaborative throughout the procedure, with no clear signs of fatigue or intraoperative decline in task performance.

4. Discussion

This study demonstrates the feasibility and clinical value of an individualized fMRI–ESM protocol designed to preserve a profession‐specific higher cognitive function during awake glioma surgery. In a professional software engineer, individualized code‐comprehension mapping enabled the identification of task‐relevant sites that were not detected during standard language mapping using an object‐naming task. By integrating behavioral assessment, preoperative fMRI, and intraoperative ESM, the surgical team tailored the resection to the patient's occupational needs, preserving postoperative programming ability.

Intraoperative mapping revealed two cortical sites that were functionally relevant to the code‐comprehension task: one in the left IFG (pars triangularis, BA 45), and another in the IFG/MFG junction (BA 46; see Figure 7). Although these regions are strongly associated with language processing (Vigneau et al. 2006), stimulation of these sites disrupted performance during the code‐comprehension task, whereas no errors were observed at the same sites during object‐naming mapping. This directly influenced the surgical strategy: if only the standard object‐naming task used in this patient had been administered, the absence of language‐positive responses could have led to an underestimation of the functional relevance of the exposed frontal cortex for this patient's professional abilities. Thus, programming‐specific mapping provided clinically actionable information by defining functional boundaries that informed a more conservative strategy near these cortical and subcortical regions.

The absence of language‐positive ESM sites should nevertheless be interpreted cautiously. It may reflect tumor‐related functional reorganization through perilesional, ipsilateral, or contralateral compensatory mechanisms, as frequently observed in slow‐growing gliomas (Desmurget et al. 2007; Garrett et al. 2012), a possibility consistent with the bilateral language activation observed in fMRI. Limited cortical exposure during surgery may also have contributed to the lack of language‐positive responses.

These findings are consistent with prior fMRI studies showing that code comprehension recruits frontal and parietal regions associated with the multiple‐demand system (Fedorenko et al. 2013; Ivanova et al. 2020; Shashidhara et al. 2024). This should not be interpreted as evidence of code‐exclusive activity, but rather as stronger recruitment of these regions during Code comprehension compared with Language comprehension. The overlap between programming‐positive ESM sites and left inferior frontal regions suggests that code comprehension may rely on cortical territories adjacent to, or partially shared with, language‐related networks, while still requiring additional executive and rule‐based processing demands.

The multiple‐demand system is thought to manage attentional focus by identifying relevant pieces of information (Blank et al. 2014; Duncan 2013; Fedorenko et al. 2013; Reverberi et al. 2012; Shashidhara et al. 2024). Within this network, the IFG has been implicated in attentional shifts, maintenance of coherent chains of thought, logical reasoning, and rule selection (Duncan 2013; Holyoak and Lu 2021; Reverberi et al. 2012). These processes are relevant for both language and code comprehension, but may be differentially recruited depending on task demands and individual expertise. In programming, the IFG/MFG region may support the selection and integration of rules, tracking of code structure, and maintenance of intermediate representations required to infer program behavior (Crank and Fox 2002; Duncan 2010; Gupta and Tranel 2012; Holyoak and Lu 2021; Reverberi et al. 2012; Vigneau et al. 2006).

From a clinical perspective, this case supports the incorporation of individualized cognitive tasks into awake neurosurgery when standard mapping protocols may not capture functions essential to the patient's professional life. Conventional intraoperative tasks are designed to preserve domain‐general sensorimotor and language functions (Bu et al. 2021; Mahon et al. 2019; Szelényi et al. 2010), but may be insufficient for expertise‐dependent abilities such as programming, music, calculation, or other individualized higher‐order skills. Importantly, fMRI was used to identify candidate regions for intraoperative exploration, whereas functional boundaries and surgical decisions were determined exclusively on the basis of ESM responses. The patient's return to professional programming within a few months after surgery supports the ecological relevance of this patient‐centered approach.

Several limitations should be acknowledged. First, this is a single‐case study, and the observed functional organization cannot be generalized to all programmers or patients with tumors in similar locations. Programming expertise, task strategy, tumor location, and neuroplastic reorganization may all influence individual activation and ESM patterns. Second, fMRI interpretation near brain tumors should be made cautiously, as neurovascular alterations may affect the BOLD signal (Krings et al. 2001; Wüstenberg et al. 2003). Third, although MRI and fMRI data were co‐registered, ESM sites were manually localized on the 3D anatomical render using cortical and venous landmarks; therefore, fMRI‐ESM correspondence should be considered an anatomically guided approximation, potentially affected by brain shift, cortical deformation, and view differences.

Importantly, the interpretation of task dissociation is limited by the intraoperative control tasks. Language mapping was primarily based on object naming, which probes only a subset of language processes, and the Syntax error detection task could not be administered during ESM because the patient did not achieve sufficient baseline performance. Therefore, the present findings do not demonstrate code‐exclusive cortex or a general functional dissociation between code and language. Rather, they show that, in this patient, stimulation of certain sites disrupted code‐comprehension performance, whereas those sites were not identified during the object‐naming task used intraoperatively.

Moreover, code comprehension was operationalized as output prediction from short code snippets. Although this captures a relevant component of programming cognition, real‐world code comprehension also involves understanding program intent, control and data flow, variable roles, broader contextual structure, and higher‐level semantic categorization (Ikutani et al. 2021; Peitek et al. 2021). In addition, the present univariate activation contrasts identify regions showing differential task‐related activity but do not reveal the specific information represented within those regions. Future studies should therefore incorporate multivariate decoding and representational similarity analyses to examine whether algorithmic, syntactic, or semantic distinctions are encoded at the population level. Repeated preoperative and intraoperative assessments may also have introduced learning effects, and the use of passive monitoring rather than active patient feedback may have limited the detection of subtle deficits (Lucas et al. 2020). Future protocols should therefore include complementary tasks probing additional aspects of programming. The exclusion of the Syntax error detection task also highlights the need to design intraoperative control tasks that are robust to the visual, ergonomic, and time constraints of awake surgery. Future work should also develop multitask protocols and systematic classifications of intraoperative errors to refine patient‐specific cognitive preservation strategies (Sierpowska et al. 2017).

In conclusion, this study provides intraoperative evidence that programming‐relevant cortical sites may not be identified by standard language mapping alone. More broadly, it supports the use of personalized fMRI–ESM protocols to preserve profession‐specific higher cognitive functions in neurosurgical patients, extending functional preservation beyond standard sensorimotor and language domains.

5. Conclusions

This case represents a proof‐of‐concept for a personalized fMRI–ESM protocol aimed at preserving programming‐related abilities in a neurosurgical context. By integrating behavioral assessment, neuroimaging, and intraoperative mapping, we identified cortical sites in the left inferior frontal cortex that were functionally relevant to code comprehension but were not detected during standard object‐naming mapping.

This approach supported complete tumor resection while preserving postoperative programming performance. The findings provide intraoperative evidence of task‐dependent functional differentiation between Code comprehension and the language task used in this patient, while not implying code‐exclusive cortex or a general dissociation between code and language processing. More broadly, they highlight the potential value of tailored mapping protocols for identifying patient‐specific cognitive functions that may not be captured by standard intraoperative assessment.

Author Contributions

Conceptualization: X.C.‐C., A.R.‐F., A.G., A.G.‐R., X.O. Data curation: X.C.‐C., A.G.‐R., A.F.‐C., A.G., P.N.‐B., A.C. Formal analysis: X.C.‐C., P.L.‐C., V.C.‐E., A.G.‐R. Funding acquisition: A.R.‐F., X.C.‐C., P.L.‐C., X.O. Investigation: X.C.‐C., A.R.‐F., A.G.‐R., A.G., A.F.‐C., V.C.‐E. Methodology: X.C.‐C., A.R.‐F., A.G.‐R., A.G., A.F.‐C., X.O., P.L.‐C., V.C.‐E. Project administration: A.R.‐F. Resources: X.C.‐C., X.O., A.G., A.R.‐F., A.F.‐C., A.G.‐R., P.L.‐C., V.C.‐E. Software: X.C.‐C. Supervision: A.R.‐F. Visualization: V.C.‐E., P.L.‐C., X.C.‐C. Writing – original draft: X.C.‐C., V.C.‐E., P.L.‐C. Writing – review and editing: all authors.

Funding

This work was supported by Xim Cerda‐Company and Patricia Leon‐Cabrera were awarded a Margarita Salas grant funded by Ministerio de Universidades and the European Union‐NextGeneration EU, and Victor Cepero‐Escribano was supported by the AGAUR Grant no. 200034TG1 (Investige program)—funded by the European Union, Next Generation EU [Ref: BDNS664047] and the pre‐doctoral program AGAUR‐FL grant (2025 FI‐3 00065) Joan Oró of the Department of Research and Universities of the Generalitat de Catalunya and the cofunding of the Social Plus European Fund. The Cognition and Brain Plasticity Group is funded by Generalitat de Catalunya 2021‐SGR‐1099. Finally, this publication is part of the R+D+I grants PID2020‐118254‐I00 and PID2023‐151083NA‐I00 financed by MCIN/AEI/10.13039/501100011033 and by ERDF/EU. Furthermore, this project was partially funded by the Agencia de Gestió d’Ajuts Universitaris i de Recerca (AGAUR) through 2021‐SGR‐01470 and the CERCA Programme/Generalitat de Catalunya.

Ethics Statement

My article reports human subjects. Recruitment meets scientific requirements & HBMs expectation of inclusivity.

Supporting information

Data S1: hbm70634‐sup‐0001‐supinfo.docx.

Table S1: Cognitive results from neuropsychological evaluation administered before (baseline) and after surgery (acute and early post‐operative stage) and behavioral performance on the programming tasks. L1: first language acquired (native language), Spanish. L2: second language acquired, Catalan. L3: third language acquired, English. SS: scalar score. TMT: Trail Making Test; TB‐r: Test Barcelona‐revised. PPT: Pyramids and Palms Tree Test. SPT: Semanic Pairs Test. BNT: Boston Naming Test. RAVLT: Rey Auditory Verbal Learning Test. Acc: accuracy. RT: response time.

Table S2: Summary of GLMM: effect of accuracy on Code comprehension. The reported z statistics and p‐values correspond to individual model coefficients relative to the reference session and are not FDR‐adjusted.

Table S3: Summary of GLMM: effect of accuracy on Syntax error detection. The reported z statistics and p‐values correspond to individual model coefficients relative to the reference session and are not FDR‐adjusted.

Table S4: Summary of LMM: effect of reaction time on Code comprehension. The reported t statistics and p‐values correspond to individual model coefficients relative to the reference session and are not FDR‐adjusted.

Table S5: Summary of LMM: effect of reaction time on Syntax error detection. The reported t statistics and p‐values correspond to individual model coefficients relative to the reference session and are not FDR‐adjusted.

Table S6: Summary of whole‐brain fMRI results in the Code comprehension task (pFWE < 0.000001) for the cluster level; Coordinates in native space; BA, Brodmann area; FFG, fusiform gyrus; LG, lingual gyrus; dmT, dorsomedial thalamus; IFG, inferior frontal gyrus; MFG, middle frontal gyrus; MTG, middle temporal gyrus; IPL, inferior parietal lobule However, it elicited stronger and more extensive activity in bilateral parietal regions including the precuneus and inferior parietal lobules (IPL), as well as in bilateral middle temporal gyri. Right hemisphere activations include rPCC, right posterior cingulate cortex; rMOG, right middle occipital gyrus; and LG, lingual gyrus (Figure 5C; for details, see Section S3 and Table S6). SPL, superior parietal lobule; SMA, supplementary motor area; DLPFC, dorsolateral prefrontal cortex.

Table S7: FDR‐adjusted pairwise comparisons of reaction times for Code comprehension.

Table S8: FDR‐adjusted pairwise comparisons of reaction times for Syntax error detection.

Figure S1: Representative stimuli used in the programming tasks. A. Example of the Code comprehension task. B. Variations of the same code template generated by changing variable values and the proposed distractor. C. Syntactically correct code snippet. D. Syntactically incorrect code snippet containing a mismatched delimiter in the conditional statement (else if {i == 20}).

Data Availability Statement

The data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions.