Therapeutic Monitoring of Post-COVID-19 Cognitive Impairment Through Novel Brain Function Assessment

Abstract

COVID-19 infection is often accompanied by psychological symptoms, which may persist long after the end of the infection (long COVID). The symptoms include fatigue, cognitive impairment, and anxiety. The reason for these long-term effects is currently unclear. Therapeutic approaches have included cognitive rehabilitation therapy, physical activity, and serotonin reuptake inhibitors (SSRIs) if depression co-exists. The neuropsychological evaluation of subjects with suspected cognitive issues is essential for the correct diagnosis. Most of the COVID-19 studies used the Montreal Cognitive Assessment (MoCA) or the Mini Mental State Examination (MMSE). However, MoCA scores can be confusing if not interpreted correctly. For this reason, we have developed an original technique to map cognitive domains and motor performance on various brain areas in COVID-19 patients aiming at improving the follow-up of long-COVID-19 symptoms. To this end, we retrospectively reanalyzed data from a cohort of 40 patients hospitalized for COVID-19 without requiring intubation or hemodialysis. Cognitive function was tested during hospitalization and six months after. Global cognitive function and cognitive domains were retrieved using MoCA tests. Laboratory data were retrieved regarding kidney function, electrolytes, acid–base, blood pressure, TC score, and P/F ratio. The dimensionality of cognitive functions was represented over cortical brain structures using a transformation matrix derived from fMRI data from the literature and the Cerebroviz mapping tool. Memory function was linearly dependent on the P/F ratio. We also used the UMAP method to reduce the dimensionality of the data and represent them in low-dimensional space. Six months after hospitalization, no cases of severe cognitive deficit persisted, and the number of moderate cognitive deficits reduced from 14% to 4%. Most cognitive domains (visuospatial abilities, executive functions, attention, working memory, spatial–temporal orientation) improved over time, except for long-term memory and language skills, which remained reduced or slightly decreased. The Cerebroviz algorithm helps to visualize which brain regions might be involved in the process. Many patients with COVID-19 continue to suffer from a subclinical cognitive deficit, particularly in the memory and language domains. Cerebroviz’s representation of the results provides a new tool for visually representing the data.

1. Introduction

Among the most famous global epidemics in human history, the SARS-CoV-2 virus (COVID-19) that emerged in 2019 is notable for the worldwide reaction to it and for the wealth of scientific data immediately available [1]. At the beginning, lung disorders caused by COVID-19 were the focus of attention because of the high mortality; however, evidence accumulated since then indicates alterations in memory and other brain functions. Furthermore, as the number of patients increased and as time has passed, it has become evident that behavioral symptoms may persist long after the end of the infection.

Many patients who have suffered from a COVID-19 infection experience long-term illness, commonly referred to as “long COVID”, which includes memory issues. The most common symptoms include persistent breathlessness, fatigue, and coughing. Other reported symptoms include chest pain, palpitations, neurological and cognitive deficits, skin rashes, and gastrointestinal dysfunction.

During hospitalization, more than half of COVID-19 patients present with nonspecific neurological symptoms [1,2], and several studies have shown that, after discharge, some patients exhibit abnormal results in neuropsychological tests [3]. In one study, 86% of participants stated that cognitive dysfunction and/or memory impairment affected their ability to work, making cognitive impairment one of the most frequently reported symptoms of long COVID [4]. This is known as “brain fog”, that is, difficulty in focusing/sustaining attention and memory issues. What differentiates post-acute COVID-19 and chronic post-COVID-19 (long COVID) is the duration of symptoms. In the first case, the symptoms last more than three weeks; in the second case, they last more than 12 weeks [5]. The mechanism by which COVID-19 infection induces long-term behavioral effects is unclear, though it may be due to sustained neuroinflammation, e.g., through mediators like IL-1beta [6] and PED-PEA [7].

Therapeutic strategies include cognitive rehabilitation therapy [8], noninvasive brain stimulation [9], physical activity [10], and SSRIs [11].

The neuropsychological evaluation of subjects with suspected cognitive issues is essential for a correct diagnosis. Therefore, it is necessary to use screening tests that allow a rough indication of the main cognitive activities in a few minutes. Most of the COVID-19 studies used the Montreal Cognitive Assessment (MoCA) or the Mini Mental State Examination (MMSE) as quick cognitive screens to evaluate cognitive functions. The MoCA assesses short-term memory, visuospatial abilities, executive functions, attention, concentration and working memory, language, and orientation to time and place. Scores relating to memory, attention, and executive functions significantly influenced the under-normality threshold global scores, according to subitem scores analyses [12]. The sensitivity MoCA for detecting mild cognitive impairment (MCI) is 90%, compared to 18% for other leading cognitive screening tools such as the MMSE [12].

The psychometric structure of the MoCA is still matter of debate. While some data confirm the face validity of the total MoCA score to identify cognitive deficit, and that many of the cognitive domains in the MoCA scores are strongly correlated, it is also clear that the total score of MoCA or MMSE is a poor representation of cognitive functions, which are a multidimensional space consisting of separate cognitive domains.

Research using functional magnetic imaging (fMRI) provided information on brain alterations after COVID-19 infection [13], such as (1) the altered function of brain areas connected with memory and thinking-related function, such as the hippocampus, and (2) neuroinflammation with modified blood–brain barrier functions and vascular modifications [14]. The modifications in blood flow and oxygen supply to the brain are hypothesized to underlie the cognitive issues experienced after COVID-19.

Cognitive impairment is strongly associated with increased mortality. Therefore, identifying cognitive impairment at an early stage has become an increasingly important challenge to physicians; however, MoCA scores can be confusing if not interpreted correctly. For this reason, in this article, we have developed an original technique to map cognitive domains and motor performance onto various brain areas. This map could help clinicians to obtain a clearer view of the injured areas related to the cognitive domains included in the MoCA. This original technique is created with Cerebroviz, an R package designed to streamline publication-quality anatomical visualizations of spatiotemporal data in the brain [15].

Here, we reanalyze the available fMRI data based on each of the tests used in the MoCA and try to create a representation of the brain using MoCA scores.

This new technique was applied to MoCA tests administered to COVID-19 patients during infection and 6 months after recovery. The attempt to derive brain imaging data from behavioral data is an emerging research effort: for example, neural representations in the supramarginal gyrus have been predicted from kinematics [16].

Justification of the Approach

This approach provides a heuristic and clinically oriented visualization rather than direct neuroanatomical evidence. The next step will be to validate this computational mapping with individual neuroimaging datasets.

The clinically oriented approach is based on the compositional theory of behavior and brain activity. Behavior can be considered a trajectory in the high-dimensional space of body states (e.g., joint angles, velocities, muscle activations) [17]:

$$x(t) \in R^{\mathrm{K}}$$

where K is the dimensionality of the motor state.

Let behavior evolve according to dynamics:

dx/dt = F (x, u, η)

where u = external input, η = internal noise. Although behavior is highly chaotic [18,19,20] and unpredictable overall [21], there are constants in it, which are evident from the fact we can describe it with natural language and the segmentation of actions using machine learning [22]. This is due to the presence of attractors [23,24], that is, stable patterns of behavior (e.g., walking, running, grasping). A trajectory of x(t) tends to visit the neighbors of these attractors (A_k) repeatedly:

$$\underset{t\to \infty }{\mathit{lim}}x(t) \to {\mathrm{A}}_{\mathrm{k}}$$

This is why trajectories can be segmented into pieces associated with each attractor. We will call this process of finding a coarse-grained (in time and space) representation of a fundamentally continuous behavioral process “tokenization” and the resulting primitives “behavioral tokens”, for similarity with Large Language Models [25].

Specifically, tokenization occurs by partitioning x(t) into episodes [t_i, t_i + 1, …] and then assigning each segment to a token p_k using a certain criterion (e.g., similarity, k-means):

$$\tau \mathrm{:} x(t) \mapsto p_k$$

where p_k is a behavioral token (e.g., “walk,” “grasp”). This process helps compare behaviors among people.

Thus, behavior becomes a symbolic sequence of tokens:

$$B(t) = \{p_{k1}, p_{k2, \dots }\}$$

The major problem with this approach is that there are many possible tokenization systems that can describe behavior equally well.

However, if the system has recurrent attractors {A_k}, tokens may correspond to low-dimensional dynamical attractors of the system, which can be defined as follows:

p_k ≡ trajectory segments where x(t) → A_k

This gives a dynamical justification for tokens: they correspond to stable, recurring patterns of movement.

Now consider neural activity n_r(t) at time t in a region r.

Suppose that each token p_k is associated with a neural signature across M regions. Then, the mapping from tokens to neural space can be expressed as

$$\mathrm{\Phi }: p_k \mapsto r_k$$

or, more generally,

$$n(t) \approx \sum w_k(t) r_k$$

where w_k(t) are the activation weights of each token at time t.

There are three major reasons why the representation of behavior is advantageous using brain signatures:

(1)

Brain signatures can be directly read using fMRI or, more conveniently, fNIR. This allows for the independent confirmation of the data, avoiding circular reasoning in behavioral tests. Indeed, the fMRI BOLD signal (or the fNIR signal) is an emerging property of the brain, indexing an average of synaptic and dendritic activity and not neuronal spikes [26,27].

(2)

There are an infinite number of ways to discretize behavior into a relatively small set of scores for different “cognitive domains”. By mapping the anchoring behavioral scores into brain regions, we pick out a canonical coordinate system imposed by neurobiology, which removes the arbitrariness of the decomposition of behavior. The new representation is physiologically grounded, may identify subtle differences among subjects (e.g., two patients may have an “executive dysfunction” score but different brain region mapping), and, potentially, might indicate the best neurostimulation programs or pharmacological interventions. Furthermore, spatial maps align naturally with genetics, connectomics, lesion data, and electrophysiology, which may provide further prediction ability in the mapping.

(3)

Spatial maps may improve behavioral prediction: one of the “aims” of behavioral studies is to predict the future behavior or the next behavioral trajectory, given the previous states:

x(t + 1) ~ P(x(t + 1)|x(t), x(t − 1),…, x(t − L))

Rather than predicting new states based only on behavior, it is possible to assume that the brain encodes the “latent dynamics” of tokens:

x(t + 1) ~ P(x(t + 1)|r(t))

The neural representation is a lossy but sufficient compression of behavioral history. Depending on how lossy the representation of behavior is in the data gathered from the brain (fMRI and fNIR are inherently noisy and limited in spatial and temporal resolution), in the optimal situation, behavior can be reconstructed entirely from the brain state through a decoding function D:

x(t) = D(r(t)) + noise

However, in more realistic conditions, the behavior x(t) provides continuity and fine-grained detail, the brain state r(t) provides latent context (intentions, goals, internal dynamics), and together they improve the prediction of the next behavioral point:

x(t + 1) = D(x(t), r(t)) + η

More generally, inferring the neural generators from behavior is conceptually similar to inferring brain source localization from EEG data (e.g., LORETA): neural generators map to brain regions, which can be related to anatomy, function, and pathology, and compress complex temporal data into fewer, spatially meaningful sources.

2. Materials and Methods

This retrospective study included all patients hospitalized for COVID-19 between 2020 and 2023 at the University of Campania, who had complete Montreal Cognitive Assessment (MoCA) items, Geriatric Depression Test (GDS) scores, and blood parameter data (detailed in

Table 1. Clinical data of the population. Data represent mean ± SD (median ± IQR when indicated).

Variable	Value
Age (years)	58 ± 19 (median ± IQR)
SBP (mmHg)	133 ± 18
DBP (mmHg)	76 ± 10
Urea (mg/dL)	44 ± 24
Creatinine (mg/dL)	1 ± 0.56
eGFR (mL/min)	8 0 ± 31
Uric Acid (mg/dL)	4.5 ± 1.3
Na (mEq/L)	137 ± 2.1
K (mEq/L)	4.1 ± 0.4
CRP (fold increase above normal value)	13 ± 12
Hb (g/dL)	13 ± 1.9
HCO3−	26 ± 3.2
CT Score	6 ± 4
P/F Ratio	288 ± 96

), and who attended a follow-up outpatient visit six months post-hospitalization. The study was approved by the Ethics Committee of the University of Campania L. Vanvitelli (Approval Code: 10117/2020, Approval Date: 6 May 2020).

The exclusion criteria were patients requiring intubation/CPAP, advanced age (>65 years), relevant comorbid conditions, oncological patients, or patients requiring hemodialysis. Patients with abnormal GDS scores were excluded from the study. Further analyses with larger samples are necessary to clarify the role of depression and anxiety in post-COVID-19 cognitive impairment.

To represent cognitive domains in terms of the brain regions involved, a database of fMRI data of brain activity during the resolution of the tests used in the MoCA was created. For each test (such as the Trail test or the clock drawing test), the brain regions activated were recorded. When more than one fMRI study reported the activation of the same brain region during a specific MoCA task, this connection was considered highly plausible. We then designed a transformation matrix with the MoCA cognitive domains in rows (language, naming, visuospatial, memory, abstraction, orienting, attention) and, in columns, the different brain regions activated in fMRI studies. Every cognitive domain–brain region pair with a highly plausible connection was denoted as 1 in the matrix.

To transform cognitive domains into brain regions, it was then necessary to multiply the MoCA by the transformation matrix. The activity of the resulting brain regions was then represented as a brain drawing using the “Cerebroviz” library in the R programming language. Thus, we created a transformation matrix containing the cognitive domains of the MoCA and the brain areas corresponding to the cognitive domains activated whenever a subject completes the MoCA test. This transformation matrix was loaded into the R program through Cerebroviz to determine the representation of anatomical brain areas activated during the MoCA.

Descriptive statistics were expressed as mean ± SD. Age was expressed as median ± IQR. Comparisons over time were performed using paired t-tests (or Wilcoxon signed-rank test when non-normality was present). Correlations were assessed with Pearson or Spearman coefficients as appropriate. A p-value < 0.05 was considered statistically significant.

3. Results

Table 1 presents the clinical parameters of the patients under study at the time of the COVID-19 hospitalization. None of the patients analyzed reported memory loss before COVID-19 infection; however, after hospitalization, patients complained of memory lapses and lack of clarity. The findings showed that several persistent symptoms can remain long after acute SARS-CoV-2 infection. One year later, the results show an improvement in cognitive functions (pre-COVID-19 MoCA score average 20.94, post-COVID-19 22.15). Memory and executive functions are the most impaired cognitive domains in patients with long COVID. Confirming this, clock drawing and deferred memory recall are used as subtests. In the first case, planning, conceptualization, and symbolic representation are assessed by asking the patient to draw the dial of a clock and correctly position the numbers. The patient is “attracted” to the strong source of stimulus rather than an appropriate response that involves a more complex operation. For example, the clock hands may be absent or point towards “10” and “11” instead of “2”. This type of stimulus-related error can also be evaluated as a conceptual error. Future studies need to thoroughly investigate COVID-19-related changes in executive functions.

Of 24 hospitalized COVID-19 patients, 61% had mild cognitive impairment (MoCA 18–25), 13% had moderate cognitive impairment (MoCA 10–17), and 2.7% had severe cognitive impairment (MoCA < 10). In addition, 13.9% showed increased creatinine. The cognitive deficit was not related to the increase in creatinine, but was associated with reduced hemoglobin (12 ± 2 g/dL vs. 14 ± 1 g/dL), hypouricemia, and hypobicarbonatemia when considering executive functions. In the patients subjected to follow-up, 50% showed mild cognitive deficit, 4% moderate deficit, and 4% severe deficit.

The analysis of cognitive domains in these patients is reported in Figure 1. Figure 1 also shows a representation of the cognitive domains as activity in different brain regions. It is clear that this type of representation is more immediate and possibly more clinically meaningful.

4. Discussion

The present study confirms that cognitive impairment related to COVID-19 behavioral alterations can persist for six months following hospitalization, particularly affecting memory and executive functions, without evidence of progressive deterioration. This persistence, in the absence of worsening, suggests a stable but incomplete recovery trajectory, consistent with other longitudinal studies on post-acute sequelae of SARS-CoV-2 infection [28,29]. Due to the retrospective nature of the study, it is difficult to formally establish a causal relationship between COVID-19 and the onset of cognitive dysfunction.

A clinically interesting aspect of this study is the identification of specific biological correlates—namely, reduced hemoglobin, bicarbonate, and uric acid concentrations—associated with greater impairment in executive functions.

Indeed, our analysis suggests that reduced hemoglobin, hypobicarbonatemia, and hypouricemia may be associated with greater executive dysfunction. Reduced hemoglobin may impair cerebral oxygenation, while hypobicarbonatemia reflects acid–base disturbances that could affect neuronal function. The role of uric acid in dopaminergic dysfunction has been already reported elsewhere [30,31,32]. Low uric acid, an endogenous antioxidant, may increase susceptibility to oxidative damage in neural tissues [32,33,34,35]. Low bicarbonate has been already reported to be linked to cognitive dysfunction in CKD [36].

These factors suggest that post-COVID-19 cognitive dysfunction may, in part, be a consequence of systemic metabolic stress, acting on vulnerable neural circuits. The hypothesis that dopaminergic dysfunction may underlie the observed cognitive patterns provides a plausible unifying mechanism [37]. Dopamine is crucial for executive function, working memory, and cognitive flexibility—domains that were most affected in the study cohort. Inflammatory cytokines elevated in COVID-19, such as IL-6 and TNF-α, have been shown to interfere with dopamine synthesis and receptor function [38]. Additionally, neuroinflammatory changes and microglial activation seen in neuropathological studies of COVID-19 support a potential dopaminergic substrate [39]. However, direct neurochemical evidence remains limited and warrants further investigation.

Equally important is the methodological advancement demonstrated in the study: the use of UMAP dimensionality reduction and Cerebroviz-based brain mapping for the representation of cognitive screening data. Traditional cognitive screening instruments are limited in that they condense a complex array of cognitive functions into a single metric, potentially obscuring meaningful patterns of impairment [40]. In order to become a valuable diagnostic tool, these procedures will need extensive studies in sensitivity and specificity components.

This approach has practical implications for therapy. Cognitive rehabilitation strategies—such as domain-targeted cognitive training, executive function exercises, and compensatory strategy coaching—can be more precisely tailored to the specific deficits identified in each patient. For example, patients with impairments in planning and symbolic reasoning, as evidenced by poor clock drawing performance, may benefit from interventions focused on problem-solving and visuospatial tasks. Furthermore, the association of cognitive deficits with metabolic markers supports a combined therapeutic model where cognitive therapies are paired with the medical management of anemia, electrolyte imbalances, or oxidative stress. Interdisciplinary care involving neurology, rehabilitation, nephrology, and nutrition may offer synergistic benefits.

Limitations of the Study

The authors acknowledge several limitations in the present study. The retrospective design and relatively small sample size reduce the statistical power and the possibility of broad generalization. Furthermore, the absence of a control group (e.g., non-COVID-19 patients matched for age and comorbidities) prevents us from fully disentangling COVID-19-related effects from other confounding factors. However, this study was primarily intended as a methodological proof-of-concept. Future prospective studies with larger cohorts are warranted to confirm our findings.

Furthermore, relying solely on the MoCA limits the depth of neuropsychological characterization. However, a major problem when using larger neuropsychological batteries is that the fMRI data for each test is not always available. Furthermore, at least in the setting of infectious diseases such as COVID-19, the use of time-expensive tests would be impractical if not impossible. Indeed, the COVID-19 pandemic resulted in longer waiting lists for neuropsychological assessment because of the closure of many neuropsychological clinics. Therefore, shorter batteries would be desirable in this scenario [41].

Larger neuropsychological batteries may also increase the risk of inter-test or inter-observer variability. The MoCA was chosen due to its wide use in post-COVID-19 reports, as well as its practicality and sensitivity in detecting mild cognitive impairment in clinical settings, especially during the pandemic context.

Finally, the variability in age within our sample may represent a potential confounder, particularly in relation to uric acid, hemoglobin, and bicarbonate levels. Nonetheless, our findings are consistent with previous literature, suggesting that any age-related effect was likely limited in this cohort. Elevated CRP levels could also indicate acute infection, which may influence cerebral perfusion and contribute to neuroinflammation. Even under these circumstances, however, the usefulness of brain mapping representations in these patients remains evident. Our study represents a preliminary observational effort that introduces a novel methodological approach, while further large-scale prospective studies are needed to establish sensitivity, specificity, and clinical utility.