The Effects of Secondary Motor and Cognitive Tasks on Gait Depend on Functional Walking Ability in Non-Traumatic Neurological Patients: A Feasibility Pilot Study

Abstract

Adaptive locomotion requires the integration of cognitive and motor processes and is challenged in neurological disorders. Dual-task (DT) training may improve cognitive–motor coordination, but its feasibility across heterogeneous clinical populations is uncertain. This pilot study aimed to understand if the effects of a secondary motor or cognitive task added to a walking task depend on the functional walking abilities of the subjects. We enrolled 30 participants with neurological disorders not related to traumatic events, 5 for each one of the following groups: healthy young subjects (HeY), healthy control subjects (HeC), subjects with stroke (ictus, IC), Parkinson’s disease (PD), multiple sclerosis (MS), and Long-COVID sequelae (LC). Spatiotemporal gait parameters were recorded using a wearable inertial magnetic unit, and subjective workload was assessed with the visual analog scale (VAS) and NASA-Task Load Index. Regression models revealed strong baseline–DT coupling for stride duration (slopes 1.11–1.37; R² 0.85–0.97), stride length (slopes 0.93–0.94; R² 0.86–0.93), walking speed (slopes 0.87–0.98; R² 0.78–0.93), and gait ratio (stance/swing, slopes 0.38–0.60; R² 0.21–0.52). Mixed-effects analyses identified significant group effects for walking speed (F(5) = 7.218, p < 0.001), stride length (F(5) = 4.834, p = 0.001), gait cycle duration (F(5) = 5.630–5.664, p < 0.001), Walking Quality (F(5) = 4.340–4.373, p = 0.001), and propulsion index (F(5) = 5.668–6.843, p < 0.001). The incongruent DT condition was the most sensitive in differentiating clinical groups. NASA-TLX indicated higher perceived workload in IC and MS compared with non-clinical groups. The protocol was completed by all participants without adverse events, supporting the feasibility of the procedure in this pilot sample. Its predictable scaling across baseline gait metrics supports its use as a personalized rehabilitation tool for diverse neurological populations. (ClinicalTrials.gov NCT07254377).

1. Introduction

Adaptive locomotion, the ability to modify gait in response to environmental, cognitive, or physiological challenges, is essential for functional mobility across a patient’s lifespan [1,2]. Far from being an automatic behavior, human walking requires continuous integration of multisensory information, anticipatory adjustments, and online motor corrections to cope with everyday challenges such as uneven terrain, obstacles, or dynamic environments [3,4]. A substantial body of evidence shows that these adaptive processes rely heavily on attentional and executive resources, highlighting the close interaction between cognitive and motor systems during locomotion [5,6]. This cognitive–motor interplay becomes particularly evident when individuals must handle additional cognitive demands while walking. In daily life, locomotion can be considered as an attention-demanding task that requires individuals to continuously adapt their movements to personal goals and environmental demands [6,7]. The capacity to manage such simultaneous demands is referred to as dual-tasking (DT) ability. In experimental settings, DT paradigms usually combine a motor task, for instance, gait, with a simultaneous cognitive task such as verbal fluency, serial subtraction, or working-memory exercises [5,8,9].

The change in performance between single-task and dual-task conditions reflects dual-task interference (DTI), which arises from the competition for shared neural resources [10,11]. The degree of DTI depends on the nature of the concurrent tasks and how individuals spontaneously allocate their attentional resources [8,12], as the limited capacity of such resources often leads to decrements in one or both tasks. Consequently, DTI during a motor–cognitive task may foster changes in gait parameters such as gait speed, stride length, stance time, and turn velocity, as well as changes in cognitive performance [13,14]. Such variation in performance can be quantified through the dual-task cost (DTC) [15].

While task-dependent DTI decrements can also occur in non-clinical populations, these effects tend to be more pronounced in neurological populations [16,17]. As gait reflects higher-order cognitive control [18], dual-task gait performance has emerged as a valuable predictor for identifying sub-clinical signs of neurological impairment [5,19]. At a neurophysiological level, functional neuroimaging studies using fNIRS showed an increment in the activation of the prefrontal cortex (PFC) during DT in the aging population, as well as in individuals with reduced neural efficacy, such as those affected by vascular, neuroinflammatory, or neurodegenerative diseases [20,21]. Indeed, when cognitive load exceeds processing capacity, automatic motor control becomes inefficient, resulting in slower gait, higher variability, and reduced rhythmicity [12], thus a diminished DT performance. Such alteration can result in detrimental effects on individuals’ functional independence, as reflected by diminished Barthel Index scores [16].

For instance, stroke patients with lower functional independence in the activities of daily living tend to recruit greater PFC resources to maintain gait performance when automatic control is impaired [22]. Consequently, under cognitive load, DTC is typically reflected by slower gait speed, increased asymmetry, reduced adaptability, and decreased cognitive scores [16,23]. Similarly, Castelli et al. [15] and Kahraman et al. [24] reported that elevated DTI in people with multiple sclerosis (MS) negatively impacts quality of life, reducing participation in community and leisure activities, as well as limiting employability.

Nonetheless, dual-task exercises can be used in rehabilitation to promote the recovery of motor and cognitive functions. Concurrent cognitive–motor training programs have shown promise in improving dynamic motor control and cognitive efficiency [25,26]. Moreover, robust evidence in favor of DT training comes from studies involving people with Parkinson’s disease (PD), for which significant improvements in walking performance have already been shown during DT protocols [27].

A lot of studies support the claim that DT training benefits gait rehabilitation in different populations such as cerebral ictus (i.e., stroke, IC) [9], MS [28], PD [27], and individuals exhibiting symptoms following SARS-CoV-2 infection [29].

Neurological disorders in which both cognitive and motor functioning could be altered are the most interesting to study with a dual-task approach. We focused our approach on the most frequent primary neurological disorders with an intrinsic biological etiology (stroke, Parkinson’s disease, and multiple sclerosis). These disorders are different from other diseases that are secondary neurological damage resulting from external traumatic events (such as spinal cord injuries or traumatic brain injuries).

Given the aforementioned findings, we included heterogeneous rehabilitation-relevant groups with the above-mentioned primary neurological disorders or with Long-COVID sequelae, that is, an emerging disease with slight neurological deficits (not resulting from a traumatic event), to span a clinically realistic range of functional walking ability and cognitive-motor vulnerability.

The aim of this study was not to perform definitive comparisons between diagnoses, but to investigate whether dual-task interference scales with baseline locomotor capacity across different etiologies. Specifically, we hypothesized that dual-task conditions modify spatiotemporal gait parameters depending on functional level and type of secondary task performed.

Because functional walking ability spans a wide range across rehabilitation populations, two non-clinical age groups were included to provide reference anchors at different functional levels rather than to investigate age-related effects per se. The younger group represents a high-functioning reference, whereas the older non-clinical group provides an age-matched control for clinical participants.

Within this framework, the proposed protocol was designed to provide insight into cognitive–motor integration across functionally meaningful reference points, thereby supporting its potential applications in neurorehabilitation contexts [30,31].

2. Materials and Methods

This experimental study was conducted at the Santa Lucia Hospital of Rome, Italy, in the period from October 2024 to September 2025, in full compliance with the Declaration of Helsinki for research involving humans. The Territorial Ethical Committee of Regione Lazio Area 5 (protocol N. 185/SL/24) approved the experimental procedures. Patients and caregivers were informed about the study procedures prior to enrolment in the study, and patients provided written informed consent.

2.1. Participants

A research team formed by neurologists, psychologists, and a biomedical engineer collected data on 30 participants, subdivided into subgroups according to their functional levels.

Participants classified as non-clinical (“healthy”) were community-dwelling individuals with no history of neurological, musculoskeletal, or cardiorespiratory conditions known to affect gait, as verified through clinical interview and medical history screening.

Two healthy non-clinical groups were enrolled: an older group (HeC) was considered the primary control group for clinical comparisons, while the younger group (HeY) was included as a high-functioning reference group; four groups were patients needing rehabilitation for respiratory problems (patients with Long-COVID sequelae, LC), for progressive neurological disorders affecting the central nervous system (Parkinson’s disease, PD, and multiple sclerosis, MS), and for cerebrovascular accident (stroke, IC). To be eligible for enrollment, patients needed to meet the following criteria:

-

Clinical diagnosis of stroke, PD, or MS. Regarding PD, only individuals with the idiopathic form of the disease were included. Concerning MS, the McDonald diagnostic criteria, revised in 2017 [32], must be met; the age must be between 25 and 65 years.

-

For patients with LC, prolonged morbidity and documented neuromotor and/or respiratory complications attributable to prior diagnosed SARS-CoV-2 infection were required, implying the need for rehabilitation [29].

-

Regardless of the neurological condition, patients must be able to walk independently and without rest for at least 6 min.

Participants were assigned to groups based on clinical diagnosis or non-clinical status using a convenience sampling approach; no random allocation was performed. Given the pilot nature of the study and small sample size, group characteristics are reported descriptively in

Table A1. Demographic characteristics of all participants. Participants whose identifier number (ID) contains HeY belong to the group of the healthy non-clinical young, and HeC is for healthy non-clinical control groups; LC is for patients diagnosed with Long-COVID; MS is for multiple sclerosis; PD is Parkinson’s disease; IC is for stroke groups. For Sex, M is masculine, and F is feminine: MoCA is for Montreal Cognitive Assessment.

ID	Age (Years)	Sex	MoCA	Bartel Index	Height (cm)	Weight (kg)
HeY1	20	M	29	100	180	85
HeY2	19	F	30	100	160	58
HeY3	21	F	30	100	179	85
HeY4	19	M	30	100	187	75
HeY5	21	F	28	100	173	70
HeC1	62	M	30	100	171	64
HeC2	41	F	30	100	174	65
HeC3	60	M	29	100	179	87
HeC4	78	M	29	100	156	68
HeC5	36	M	28	100	169	62
LC1	84	F	28	100	156	55
LC2	66	F	30	100	167	63
LC3	81	F	30	96	156	44
LC4	28	M	23	100	175	95
LC5	69	F	30	100	172	98
MS1	56	F	28	100	189	87
MS2	31	F	28	100	170	87
MS3	78	F	30	97	181	100
MS4	49	F	30	96	160	78
MS5	50	M	23	100	165	55
PD1	76	M	30	100	164	82
PD2	65	M	25	100	160	54
PD3	71	M	28	97	160	52
PD4	63	F	29	98	170	61
PD5	62	F	28	91	175	74
IC1	40	F	21	20	180	80
IC2	26	M	25	39	165	56
IC3	51	M	29	24	160	51
IC4	43	F	25	32	185	82
IC5	70	M	25	98	158	53

. All patients were examined for motor symptoms with disease-specific clinical scales and the Barthel Index to measure independence in daily living activities. Furthermore, all participants were tested with the Montreal Cognitive Assessment (MoCA) [33], a cognitive screening tool. The full assessment protocol is further described in the next paragraphs.

2.2. Procedure

The study had a cross-sectional design, and participants visited our laboratory for a single session for gait recording and demographic data collection. Two landmarks were placed on the floor of our laboratory (a long, quiet room without windows) to mark the start and endpoints, and participants were asked to walk barefoot in an unobstructed 20 m pathway at a comfortable speed. We recorded a total of four trials combining two levels of motor and cognitive difficulty. The first (baseline) trial allowed participants to walk at a self-selected comfortable speed (single-motor task); the second required them to walk holding a glass full of water (dual-motor task: Dual-M task). In the third and fourth trials, participants were recorded while walking wearing headphones with acoustic stimuli and performing an auditory version of the Stroop task in congruent (congruent motor, CM task) and incongruent (incongruent motor, IM task) conditions, as described in the next paragraph. The order of conditions progressed from the least to the most demanding task to ensure safety and feasibility in clinical populations; rest periods were allowed between trials to minimize fatigue.

Experimental Dual Task

We created an experimental auditory stimulus based on the Green and Barber gender Stroop effect [34]. We used a vocal recorder to collect audio from a 49-year-old man and a 33-year-old woman. To elicit the Stroop effect, we created two lists of words that either matched (congruent, i.e., the word “woman” spoken by the woman’s voice) or mismatched (incongruent, i.e., the word “man” spoken by the woman’s voice) the speaker’s gender. Four samples of neutral words were also created for familiarization with the task before performing the real experimental procedure.

Audacity^® software (version 3.4.7-64 bit, Audacityteam.org) was used to cut the digital audio and to create the final recording in which the pace of words is one every 2 s (i.e., a frequency of 0.5 Hz). This frequency was chosen considering that the speed of a comfortable gait is about 1.2–1.4 m/s and corresponds to a cadence of approximately 100–120 steps per minute [35]. Therefore, a relaxed speech pace was used to avoid interfering with the natural rhythm of walking, allowing any observed effects to be attributed mainly to the Stroop task.

2.3. Gait Analysis

Gait data were acquired by means of a wearable Inertial Measurement Unit (IMU, G-Walk, BTS, Padua) endowed with a triaxial accelerometer, a triaxial gyroscope, and a magnetometer. The device was worn with an ergonomic waist belt at the level of the sacral vertebrae S1–S2, connected to a portable computer via Bluetooth, as in a previous study [35]. The sample frequency of recording was 100 Hz. While walking, this sensor records lower trunk accelerations and angular velocities (respectively, along and around anterior–posterior, latero-lateral, and craniocaudal directions), estimating from these signals (and from the given path length) the gait spatiotemporal parameters. We analyzed the following spatiotemporal gait parameters: walking speed (WS, m/s), stride length (SL, m), cadence (Cad, steps/min), number of performed steps (steps, n), and stride duration (SD, s). For both cycles, we recorded the gait cycle duration (CD, s), the percentage duration of the first phase of double support (DS, % cycle), stance phase (% cycle), swing phase (% cycle), and the propulsion index (inclination of the line following the rising edge of the acceleration pattern). Additionally, we extracted the following gait quality indices [36]: the Symmetry Index (% symmetry between the anterior/posterior acceleration curve during the right and left cycles) and the Walking Quality Index (a composite measure that evaluates gait efficiency, stability, and symmetry for each cycle). All parameters were computed using standard definitions provided by the manufacturer and were consistent with previous validation studies [35,36].

2.4. Clinical Assessment

We assessed motor symptoms associated with the specific disease for each clinical group. Patients with stroke were assessed using the National Institutes of Health Stroke Scale (NIHSS) [37] to evaluate and quantify the severity of neurological deficits, with scores ranging from 0 (no symptoms) to 42 (severe impairments). Parkinson’s disease was assessed with the Unified Parkinson’s Disease Rating Scale (UPDRS) [38], multiple sclerosis was assessed using the Expanded Disability Status Scale (EDSS) [39] and, given the absence of a validated scale specific for Long-COVID sequelae, a visual analogical scale (VAS) for the self-assessment of the reported symptoms was developed according to the literature [40]; a detailed description is reported in

Table A3. Clinical assessment for patients with Long COVID symptoms. Symptom domain scores and total scores across the five Long COVID profiles (LC1–LC5). Values represent the perceived severity of fatigue, respiratory difficulties, chest pain, cognitive complaints, muscle and joint pain, gastrointestinal symptoms, anxiety/depressive symptoms, and sleep disturbances for each profile.

ID	Perceived Fatigue	Respiratory Difficulties	Chest Pain	Cognitive Complaints	Muscle and Joint Pain	Gastrointestinal Symptoms	Anxiety/ Depressive Symptoms	Sleep Disturbances	Total Score
LC1	9.7	4.7	0.3	2.0	4.3	0.3	4.6	4.1	29.9
LC2	4.5	9.7	1.7	0.1	7.4	2.2	8.9	0.5	35.1
LC3	6.7	5.8	0.5	4.1	0.7	9.2	4.9	0.8	32.6
LC4	7.0	4.8	6.8	1.3	9.8	0.0	1.9	6.9	38.5
LC5	3.5	3.6	5.3	0.0	3.4	1.4	1.9	0.5	19.7

. In the present study, functional walking ability is operationally defined as baseline locomotor performance during single-task walking, quantified through spatiotemporal gait parameters (e.g., walking speed, stride length, and stride duration) and supported by measures of functional independence (Barthel Index).

2.5. Acceptability and Feasibility Assessment

We collected data on the acceptability of the Dual Task paradigm using a VAS ranging from 0 (completely not agree) to 5 (totally agree) to assess the difficulty, discomfort, potential implications for rehabilitation, and enjoyment associated with the trial. Feasibility of experimental DT was also assessed using the NASA-Task Load Index (NASA-TLX) [41] for the multidimensional subjective assessment rates of the perceived workload associated with the task. Patients completed this scale at the end of each trial session, making an evaluation of the task according to mental, physical, and temporal demand, as well as effort and frustration. Each domain is scored on a straight line that goes from 0 to 10 points (with 10 as the highest perceived workload), and patients indicate the point that corresponds to their evaluation. Feasibility and safety were assessed by successful completion of all experimental conditions without adverse events, while tolerability was descriptively evaluated using VAS acceptability ratings and NASA-TLX workload scores, without applying predefined clinical thresholds.

2.6. Statistical Analysis

Statistical analysis was performed using MATLAB 2024b (MathWorks, 2024) and version 23 of the IBM SPSS statistical software. The Kolmogorov–Smirnov test revealed a normal distribution of all gait variables and both demographic and anthropometric data (p > 0.05), while a nonparametric assumption was made for VAS and NASA-TLX measurements (p < 0.001). All data were reported as mean and standard deviation for parametric distribution, while median and interquartile ranges (IQR) were adopted for non-parametric data. The t-test was used to assess baseline differences between the clinical and the older control group (HeC). Gait parameters collected during this research were analyzed relative to each participant’s baseline values, rather than normalized to anthropometric measures, to emphasize functional scaling of interference effects.

We analyzed the potential dual-task benefits on different levels of impairments, using a linear regression analysis between motor baseline values (independent variable) and the corresponding values obtained under each experimental condition (dependent variable). The model provided the slope (a), the intercept (b), the two-tailed p-value for the slope coefficient, and the coefficient of determination (R²) that can be interpreted according to the linear fit method [42] for the equation of the linear fit:

y = a × x + b

(1)

where x is the baseline in the single motor walking task, and y is the different conditions of the dual tasks.

A Linear Mixed Analysis of Variance (LM-ANOVA) was implemented with the built-in MATLAB function fitlme, setting as fixed effects the variable group (Group), the categorical characterization of each exercise in terms of difficulty of the task (Diff), the cognitive load (Cog), and their interaction.

For each dependent variable (DV), a series of linear mixed-effects models was fitted to the data using the MATLAB fitlme function. Each DV was then modeled separately according to the following general formula:

$$\begin{gathered} {DV}_{ij}={\beta }_0+{\beta }_1\left({Group}_i\right)+{\beta }_2\left({Diff}_j\right)+{\beta }_3\left({Cog}_j\right)+{\beta }_4({Group}_i\times {Diff}_j) \\ +{\beta }_5({Group}_i\times {DiffCog}_j)+u_i+{\epsilon }_{ij} \end{gathered}$$

(2)

In this specification, Group, Diff, and Cog were included as fixed effects, along with their two-way interactions, while participants were entered as a random effect to account for inter-individual variability. This structure allowed the model to estimate population-level effects of group, task difficulty, and cognitive covariates while controlling for repeated measurements within participants. The effect size for fixed effects and their interactions was estimated as partial eta-squares (ηp²). Both ηp² and R² are interpreted according to Cohen’s (1988) [43] guideline, and assume values of 0.01, 0.06, and 0.14 as small, medium, and large effects, respectively. Pairwise post hoc comparisons between groups were conducted on the estimated fixed effects. For each DV, all possible pairwise contrasts between groups were computed using custom contrast matrices constructed from the model coefficients [44], using the MATLAB coefTest function.

All post hoc comparisons were adjusted according to Bland and Altman (1995) [45], and the Bonferroni correction was applied to the p-value as follows: $p_{corr}=min(p\times k,1)$, where p_corr is the corrected alpha level, p is the original value, and k represents the number of possible comparisons. For all the tests, the significance was set at 0.05.

To address the limited statistical power associated with the small sample size, we estimated the robustness of LM-ANOVA implementing four bootstrap procedures commonly adopted in multilevel modeling: a cluster bootstrap resampling participants to preserve within-subject dependence, a residual bootstrap resampling model residuals, a parametric bootstrap simulating data from the fitted LME under Gaussian random effects and errors, and a wild bootstrap, which is recommended for heteroskedastic and small-sample settings because it preserves key features of the data, that standard resampling methods may distort, by resampling residuals with random weights maintaining the original design matrix and replicating the conditional variance structure of the errors [46,47]. For each dependent variable, 2000 bootstrap replications were generated, and the model was re-fitted to each sample to obtain percentile and studentized confidence intervals, bootstrap standard errors, bias estimates, sign-consistency measures, and bootstrap-based p-values. The older control group (HeC) was specified as the reference level, so all group coefficients represent deviations relative to this baseline. To summarize the robustness of fixed effects, coefficients were classified as robust when ≥75% of bootstrap methods yielded p < 0.05, confidence intervals excluded zero, and sign-consistency was ≥0.90; otherwise, they were classified as fragile or non-evidential. Similar bootstrap-based assessments of LME reliability in small-sample clinical and behavioral studies have been reported in recent methodological work [48,49].

3. Results

3.1. Gait Results

Demographic characteristics and functional locomotor capacities of all groups are shown in

Table 1. Demographic characteristics of groups. N is the number of participants enrolled; F stands for feminine genre; MoCa is the score obtained at the Montreal Cognitive Assessment; HeY and HeC denote the healthy young and the older control group (age-matched with patients), respectively; LC stands for patients with Long-COVID symptoms; PD, MS and IC denote patients diagnosed with Parkinson’s disease, multiple sclerosis and cerebral ictus, respectively. WS is the walking speed. p-values are the result of t-test statistics as a comparison of each clinical group with the matched-age control group (HeC).

Group	N	Age (Years)	Height (cm)	Weight (kg)	Sex (F)	MoCa	Chronicity (Years)	Bartel Index	Baseline WS (m/s)
HeY	5	20 ± 1	169.6 ± 12.2	64.4 ± 15.3	60%	29 ± 1	0 ± 0	100 ± 0	1.2 ± 0.1
HeC	5	55 ± 17	175.8 ± 10.1	74.6 ± 11.3	20%	29 ± 1	0 ± 0	100 ± 0	1.0 ± 0.1
LC	5	65 ± 22 (p = 0.53)	165.2 ± 8.9 (p = 0.16)	71.0 ± 24.3 (p = 0.80)	80%	28 ± 3	3 ± 2	99 ± 3	1.0 ± 0.1
MS	5	53 ± 17 (p = 0.77)	165.8 ± 6.6 (p = 0.09)	64.6 ± 13.0 (p = 0.19)	80%	28 ± 3	9 ± 6	99 ± 2	1.0 ± 0.3
PD	5	67 ± 6 (p = 0.17)	173.0 ± 11.9 (p = 0.70)	81.4 ± 16.7 (p = 0.40)	40%	28 ± 2	6 ± 5	97 ± 4	0.9 ± 0.3
IC	5	46 ± 16 (p = 0.47)	169.8 ± 8.6 (p = 0.46)	69.2 ± 10.2 (p = 0.32)	40%	25 ± 3	2 ± 3	89 ± 5	0.6 ± 0.2

and

Table 2. Linear regression results of the relationship between baseline gait parameters (x) and the corresponding condition-related change (y), modeled as y = a·x + b, where y = condition − baseline. The table reports slope (a), intercept (b), p-value, and coefficient of determination (R²). Higher slopes indicate stronger dependence of the condition-related change on baseline gait magnitude.

Condition	Gait Parameter	Slope a	Intercept b	p	R2
Dual_M	SD	1.114	−1.717	<0.001	0.97
	SL	0.935	0.050	<0.001	0.93
	WS	0.982	−0.004	<0.001	0.93
	GR	0.589	0.665	<0.001	0.52
CM	SD	1.366	−6.001	<0.001	0.85
	SL	0.934	−0.007	<0.001	0.86
	WS	0.866	0.022	<0.001	0.78
	GR	0.382	1.042	0.010	0.22
IM	SD	1.269	−3.732	<0.001	0.88
	SL	0.943	0.012	<0.001	0.86
	WS	0.888	0.030	<0.001	0.82
	GR	0.531	0.806	0.012	0.21

. Clinical groups (LC, MS, PD, and IC) did not report statistical differences in terms of age, height, and weight with respect to the HeC group, as shown by p-values. Participants with Long-COVID (LC) sequelae reported a moderate subjective symptom burden, with a median global score of 40.8 (IQR: 24.7–48.1). Within this group, perceived fatigue and respiratory difficulties emerged as the most prominent symptoms, indicating that reduced energy levels and breathing-related complaints represented the main residual limitations. In the multiple sclerosis (MS) group, disability levels were generally low, with a median Expanded Disability Status Scale (EDSS) score of 4 (IQR: 1–6.5), consistent with mild-to-moderate functional impairment and preserved ambulation. Participants with Parkinson’s disease (PD) showed a wide variability in motor symptom severity, as reflected by a median Unified Parkinson’s Disease Rating Scale (UPDRS) score of 14 (IQR: 9–59), suggesting heterogeneous motor involvement within this small cohort. Finally, individuals with stroke (IC) presented with low residual neurological deficits, with a median NIH Stroke Scale score of 2 (IQR: 1–4), indicating predominantly mild post-stroke impairment at the time of assessment.

All enrolled participants completed all four walking conditions without protocol deviations, resulting in no missing data. Accordingly, analyses were performed on the full dataset following an intention-to-treat approach.

The regression analysis of three main spatiotemporal gait parameters and the gait ratio index (GR, stance/swing) with respect to the functional walking ability reported on the x-axis in terms of baseline values of the gait parameters recorded during the single-walking task, in the three different DT conditions, is shown in Figure 1, with the equations reported in Table 2. A slope close to one and an intercept different from zero mean that the dual task had a general effect on that parameter; conversely, a slope different from one indicates that the effect of the secondary task mainly depended on the baseline functional walking ability.

The stride duration (SD, upper left panel) parameter yielded slopes greater than one for all conditions (Dual-M: 1.114, CM: 1.336, IM: 1.269) with excellent goodness-of-fit (R² = 0.85–0.97, all p < 0.001), indicating that this parameter is proportionally more increased for subjects with a reduced functional walking ability. At the same time, there was a reduction in stride length (SL, upper right panel), with slopes lower than 1 (Dual-M: 0.935, CM: 0.934, IM: 0.943), with strong model fits (R² = 0.86–0.93, all p < 0.001), suggesting that dual-task performance scaled consistently with baseline gait capacity; walking speed (WS, lower left panel) also demonstrated robust linear associations, with slopes ranging from 0.866 to 0.982 and R² values between 0.78 and 0.93 (all p < 0.001), indicating that baseline speed remained a strong predictor of dual-task inference across conditions. Finally, GR consistently exhibited substantially lower slopes (0.382–0.589) and reduced R² values (0.21–0.52), reflecting a more complex dependency on baseline walking ability. Across conditions, GR intercepts remained relatively stable, and, in the Dual-M condition, the mean value was close to the theoretical value of the golden ratio (≈1.618). Notably, while changes in SD, SL, and WS increased proportionally with baseline performance, GR showed limited modulation across functional levels and task conditions, with the smallest explained variance observed in the CM and IM conditions.

Table 3. Motor baseline of gait parameters reported for each group as mean ± standard deviation. HeY and HeC denote the healthy young and matched-age control groups, respectively. LC: Long COVID, PD: Parkinson’s disease, MS: multiple sclerosis, and IC: cerebral ictus groups. p-values are the results of the mixed ANOVA and are reported in bold when values meet the statistical significance set at 0.05.

Parameter	Side	HeY	HeC	LC	MS	PD	IC	Group Difference
WS (m/s)	Global	1.2 ± 0.1	1.0 ± 0.1	1.0 ± 0.1	1.0 ± 0.3	0.9 ± 0.3	0.6 ± 0.2	p < 0.001
SL (m)		1.3 ± 0.1	1.2 ± 0.1	1.1 ± 0.1	1.1 ± 0.3	1.1 ± 0.2	0.8 ± 0.2	p = 0.001
Cad (steps/min)		111.6 ± 4.9	105.4 ± 13.7	112.5 ± 5.3	104.1 ± 13.6	105.7 ± 18.9	85.4 ± 10.9	p < 0.001
Steps (n)		26.2 ± 3.7	32.2 ± 3.0	33.2 ± 4.7	35.8 ± 12.5	33.2 ± 10.0	49.0 ± 12.8	p = 0.003
SD (s)		18.2 ± 1.9	22.8 ± 2.1	22.2 ± 2.3	26.3 ± 11.4	25.3 ± 7.2	41.8 ± 14.6	p < 0.001
Symmetry Index (%)		97.7 ± 1.0	93.1 ± 4.7	92.6 ± 4.0	92.0 ± 7.3	87.5 ± 7.4	72.5 ± 19.1	p < 0.001
CD (s)	right	1.1 ± 0.0	1.2 ± 0.1	1.1 ± 0.1	1.2 ± 0.2	1.2 ± 0.3	1.5 ± 0.2	p < 0.001
	left	1.1 ± 0.0	1.2 ± 0.1	1.1 ± 0.0	1.2 ± 0.2	1.2 ± 0.3	1.5 ± 0.2	p < 0.001
DS (% cycle)	right	12.5 ± 1.1	12.6 ± 2.5	9.9 ± 3.2	11.7 ± 2.5	12.0 ± 4.2	10.6 ± 1.1	p = 0.332
	left	11.5 ± 1.0	11.8 ± 1.0	10.4 ± 2.5	12.8 ± 1.9	11.5 ± 2.6	10.1 ± 1.8	p = 0.067
Stance (% cycle)	right	61.2 ± 2.1	63.1 ± 2.4	58.7 ± 4.3	60.0 ± 5.1	60.0 ± 5.1	60.5 ± 8.6	p = 0.748
	left	62.4 ± 1.0	61.2 ± 1.3	61.9 ± 1.9	62.8 ± 3.3	62.8 ± 3.3	60.2 ± 10.7	p = 0.721
Swing (% cycle)	right	38.8 ± 2.1	36.9 ± 2.4	41.3 ± 4.3	40.0 ± 5.1	40.0 ± 5.1	39.5 ± 8.6	p = 0.748
	left	37.6 ± 1.0	38.8 ± 1.3	38.1 ± 1.9	37.2 ± 3.3	37.2 ± 3.3	39.8 ± 10.7	p = 0.721
Walking Quality Index (%)	right	96.9 ± 3.6	93.8 ± 4.7	93.3 ± 5.2	93.6 ± 4.7	91.6 ± 3.9	86.4 ± 8.0	p = 0.001
	left	95.2 ± 2.1	97.5 ± 2.5	96.2 ± 3.9	92.3 ± 5.2	91.5 ± 2.9	84.2 ± 11.9	p = 0.001
Propulsion Index (%)	right	8.3 ± 1.2	5.3 ± 1.0	5.3 ± 1.7	6.8 ± 1.5	5.2 ± 1.7	4.5 ± 1.9	p < 0.001
	left	8.2 ± 1.1	6.0 ± 1.1	4.8 ± 1.6	6.2 ± 2.1	5.2 ± 1.7	4.2 ± 1.7	p < 0.001

reports the gait parameters in a single motor task for each group of subjects. Significant differences were observed in several gait parameters. Analysis of variance showed a group effect for global parameters as follows: walking speed (WS, F(5)= 7.218, p < 0.001, ηp² = 0.26) stride length (SL, F(5)= 4.834, p = 0.001, ηp² = 0.19), cadence (Cad, F(5)= 5.265, p < 0.001, ηp² = 0.21), number of processed steps (Steps, F(5)= 3.827, p = 0.003, ηp² = 0.16), stride duration (SD, F(5)= 5.630, p < 0.001, ηp² = 0.22). The stroke and PD groups exhibited slower speeds than the LC and MS groups, which reported values closer to the HeC group, with HeY being the fastest. The LC group had mean values for SL (1.1 ± 0.1 m) and number of steps (33.2 ± 4.7) comparable to PD and MS groups, but demonstrated a more stable pattern, showing less variability from the mean. In fact, cadence (112.5 ± 5.3 steps/min) and SD (22.2 ± 2.3 s) were much closer to the HeC group’s mean value.

Indeed, group effects emerged for several gait parameters on both sides. In particular, large effects were observed for gait cycle duration (CD, left, F(5) = 5.664, p < 0.001, ηp² = 0.22; right, F(5) = 5.630, p < 0.001, ηp² = 0.22), Walking Quality (left, F(5)= 4.373, p = 0.001, ηp² = 0.18; right, F(5) = 4.340, p = 0.001, ηp² = 0.18) and propulsion index (left, F(5) = 6.843, p < 0.001, ηp² = 0.25; right, F(5) = 5.668, p < 0.001, ηp² = 0.22).

No significant effect was found for task difficulty. However, the cognitive load factor showed several significant effects. A medium effect size was observed for stride duration (F(1) = 4273, p = 0.041, ηp² = 0.04), number of steps (F(1)= 4582, p = 0.035, ηp² = 0.04), walking speed (F(1)= 8869, p = 0.004, ηp² = 0.08), and the right-side propulsion index (F(1) = 4113, p = 0.045, ηp² = 0.04). A large effect was found for stride length (F(1) = 17,827, p < 0.001, ηp² = 0.15).

Regarding interactions, a significant group × task difficulty interaction with medium effect size was found for the Symmetry Index (F(5) = 2761, p = 0.022, ηp² = 0.12) and the right-side Walking Quality Index (F(5) = 2465, p = 0.038, ηp² = 0.11). Additionally, a significant group × cognitive load interaction emerged for duration (F(5) = 9039, p < 0.001, ηp² = 0.31) and stride length (F(5) = 3091, p = 0.012, ηp² = 0.13), showing large and medium effects, respectively.

Post hoc comparison of groups x conditions is reported in

Table A2. Significant between-group differences in gait variables across task conditions. The table presents corrected p-values for pairwise group comparisons (IC vs. HeY, IC vs. LC, IC vs. PD, IC vs. CTR, IC vs. SM, PD vs. HeY) during Baseline, Dual-M, CM, and IM tasks. Reported gait variables include stride duration (SD), stride length (SL), walking speed (WS), cadence (Cad), cycle duration (CD), and propulsion index. Only comparisons reaching statistical significance after multiple-comparison correction are included.

Task	Group Comparison	Gait Variable	p-Value (Corrected)
Baseline	IC vs. HeY	SD	0.026
		WS	0.024
Dual-M	IC vs. HeY	SD	0.003
		SL	0.034
		Steps	0.017
		WS	0.009
		Propulsion index (left)	0.006
	PD vs. HeY	Propulsion index (left)	0.034
CM	IC vs. HeY	SD	0.019
		CD (right)	0.012
		CD (left)	0.019
	IC vs. LC	CD (right)	0.035
		CD (left)	0.045
IM	IC vs. CTR	SD	0.027
	IC vs. LC	SD	0.025
		CD (right)	0.029
		CD (left)	0.020
	IC vs. PD	CD (right)	0.037
		CD (left)	0.039
	IC vs. SM	SD	0.028
		CD (right)	0.039
		CD (left)	0.035
	IC vs. HeY	Cad	0.036
		SD	0.002
		SL	0.015
		WS	0.002
		CD (right)	0.010
		CD (left)	0.011
		Propulsion index (left)	0.004

of Appendix A.

3.2. Bootstrap-Based Robustness Analysis

Across all models, we examined a total of 324 fixed-effect coefficients, reflecting the full set of group contrasts, task effects (Diff, Cog), and their interactions estimated in the mixed-effects model. About 35 variables (11%) were classified as robust, 5 (2%) as fragile, and the remaining 87% showed no evidence of a stable effect. Among the robust effects, several refer to the IC group, which exhibited consistent differences relative to the HeC across multiple gait variables, as shown in Figure 2. The cognitive factor (Cog) produced robust main effects on step length and walking speed and contributed to robust group × cognitive interactions (notably within the IC and MS groups), indicating that dual-task demands modulate gait performance in a consistent and group-specific manner.

In contrast, the few coefficients classified as fragile reached nominal significance in the asymptotic LME analysis but did not meet the bootstrap stability criteria; these included, for example, the IC group effect on left propulsion and the Cog effects on stride duration, steps, and the right Walking Quality Index. For the remaining coefficients, bootstrap distributions were centered near zero or highly variable, consistent with the limited precision expected in a pilot study with small group sizes.

3.3. Usability Assessment

Acceptability of the DT procedure was assessed through VAS measurement. The Kruskal–Wallis test for independent samples reported differences across groups for discomfort associated with the cognitive–motor dual tasks (p = 0.027) and complexity (p = 0.030) associated with the motor dual task, graphically reported in Figure 3.

Cognitive load, graphically reported in Figure 4, associated with the DT, assessed using the NASA-TLX index, showed differences across groups in the cognitive domain (p = 0.011) for both cognitive and motor dual tasks (p = 0.018), physical demand (p = 0.006), and effort (p = 0.042) associated with the motor dual task.

Wilcoxon signed-rank test was used as a post hoc test to assess differences in group comparison, showing that none of the contrasts survived the Bonferroni correction for 15 multiple comparisons (α = 0.05/15 ≈ 0.0033).

4. Discussion

This study aimed to investigate whether the effects of adding a secondary task (motor, cognitive, with congruent stimuli, or cognitive incongruent stimuli) to walking may depend on individuals’ functional walking ability across non-clinical and neurological populations, adopting a functional rather than disease-specific perspective.

The main findings indicate that dual-task interference scales strongly with baseline gait performance, with temporal parameters, such as stride duration, being the most sensitive to cognitive–motor interference. Spatial parameters such as stride length remained comparatively stable, while walking speed showed intermediate, baseline-dependent modulation. Finally, among the experimental conditions, the incongruent cognitive dual task was the most effective in differentiating functional levels across groups. Individuals with longer baseline durations exhibit larger timing disruptions, especially when cognitive demands, with a higher impact than a secondary motor task, are introduced.

Such results align with previous evidence showing that temporal gait control relies heavily on executive and attentional resources, making it more vulnerable to interference than spatial parameters [5,50]. Moreover, the high goodness-of-fit (R² = 0.85–0.97) indicates that baseline temporal rhythm is a robust determinant of temporal performance stability, even under competing task demands.

In contrast, stride length displayed slopes close to one and with an intercept close to zero among the three different conditions, suggesting that gait spatial parameters remain relatively stable and show limited variability compared with temporal ones. This supports the idea that stride amplitude is largely governed by biomechanical and peripheral sensorimotor constraints [51,52] and is therefore relatively resilient to cognitive load. Prior studies have similarly reported that while cadence and stride timing readily shift under interference, stride length remains comparatively stable and reflective of an individual’s habitual locomotor pattern [6,53].

Walking speed showed slopes below one, but an intercept close to zero. This latter result suggests that the alteration of walking speed mainly depended on the functional level of the subjects, but still correlated strongly with baseline values. This indicates that healthy young adults who normally walk faster tend to slow down less under dual-task conditions than individuals affected by some functional deficits. This result is consistent with the presence of a motor “reserve” or automaticity buffer mitigating interference-related decrements [54,55]. The robust predictive accuracy of the regression models (R² = 0.78–0.93) further aligns with evidence identifying walking speed as a global, integrative marker of motor robustness [56] and is strongly linked to both cognitive functioning and gait automaticity [54].

Finally, the linear fit was poor for the ratio between stance and swing phases (R² = 0.21–0.93), mainly the values were in the range between 1.5 and 1.7, regardless of the conditions, and close to the value of the golden ratio (1.618), according to the idea that it is a pivotal value for locomotion [57]. Regarding the pace of walking, it should be highlighted that cognitive tasks were administered via regularly paced word lists (~0.5 Hz), and participants were exposed to a quasi-periodic auditory stream capable of acting as an external temporal attractor [8]. Human locomotion is known to spontaneously synchronize with rhythmic auditory stimuli, even in the absence of explicit instructions, due to intrinsic auditory–motor coupling mechanisms [58,59]. Such entrainment selectively affects temporal gait parameters while leaving spatial parameters largely unaffected, which offers a complementary explanation for why stride duration exhibited amplified interference effects, whereas stride length remained stable. Indeed, the regression analyses revealed that healthy young adults, older adults, and clinical participants all followed a highly similar baseline–dual-task performance. Rather than reflecting a lack of sensitivity of the protocol, this pattern indicates that the dual-task interference elicited by our procedure primarily engages shared sensorimotor and cognitive control mechanisms that are common across functionally heterogeneous clinical pathologies. These mechanisms appear to be mainly modulated by the individuals’ functional walking level, rather than by disease-specific features. Importantly, the expression of dual-task interference varies depending on the specific spatiotemporal gait parameter considered, suggesting that different parameters capture distinct aspects of cognitive–motor integration. From a rehabilitation standpoint, this is a desirable property: the task elicits consistent and quantifiable perturbations in gait performance, yet its effects are directly linked to each individual’s baseline locomotor capacity. Such behavior suggests that the dual-task paradigm can be standardized and personalized according to a patient’s functional level, enabling a controlled manipulation of the cognitive–motor load.

Although the clinical groups differ in etiology and nosological classification, they were examined here as rehabilitation-relevant conditions potentially affecting gait and cognitive–motor integration. To account for group functional differences, we performed a mixed-effects analysis showing clear group-related differences in several gait parameters. As revealed in the post hoc analyses, the stroke group showed the most pronounced deviations across gait cycle duration, stride timing, propulsion, and Walking Quality, consistent with previous literature on gait impairments following stroke [60,61]. The Parkinson’s group (PD) also exhibited dual-task-related effects, particularly in temporal and quality-related metrics [62], in line with evidence that this disorder affects fronto-striatal circuits critical for maintaining gait rhythm and attentional flexibility [5,63]. By contrast, patients with multiple sclerosis (MS) displayed moderate sensitivity to cognitive load. Although the group differences were not as pronounced as in stroke, this trend is coherent with studies showing that individuals with MS often experience an increased attentional cost of walking and greater gait variability, even during relatively simple cognitive tasks [64]. Finally, the Long-COVID group (LC) showed subtle decrements in gait quality and stability, aligning with recent evidence about fatigue-related motor and coordination deficits in individuals with LC [65,66].

Across all groups, cognitive load emerged as the factor producing the most consistent dual-task effect, influencing stride duration, stride length, walking speed, and the number of steps. Some parameters showed bilateral alterations, which typically reflect more persistent biomechanical constraints rather than momentary, task-induced fluctuations. Subjective and objective measures of dual-task performance capture complementary, non-redundant aspects of cognitive–motor interaction; accordingly, the lack of a clear association between subjective workload ratings (NASA-TLX) and objective gait parameter changes likely reflects a dissociation between perceived effort and motor performance rather than an absence of dual-task interference. Subjective workload scales capture conscious appraisal of task demands and effort, whereas spatiotemporal gait parameters primarily reflect automatic or semi-automatic motor control processes that may be modulated without explicit awareness and that rely on partially distinct neural mechanisms and timescales [41,54]. Such dissociation has been repeatedly reported in dual-task paradigms, where individuals may preserve performance through compensatory attentional strategies despite increased internal effort, or conversely exhibit measurable motor alterations without reporting elevated workload [67,68]. This phenomenon is particularly relevant in clinical populations. After a stroke, gait control often relies on increased attentional supervision despite limited changes in perceived workload [5,22]. In Parkinson’s disease, reduced gait automaticity may lead to objective dual-task costs even when subjective effort remains modest [54,62]. Similarly, individuals with multiple sclerosis frequently exhibit increased cognitive–motor interference and gait variability under dual-task conditions, while subjective workload does not always scale proportionally with motor deterioration [12,64]. In Long-COVID, emerging evidence suggests fatigue-related motor inefficiency and altered sensorimotor integration, which may manifest as subtle gait changes not consistently mirrored by perceived workload [65,66]. Overall, this pattern fits well with previous studies showing that cognitive demands can disrupt gait automaticity in favor of a more adaptive behavior relying heavily on attentional and executive resources [5,6,8]. The marked impact of cognitive load on stride length, in particular, is consistent with evidence that spatial aspects of gait become more sensitive to interference when attentional resources are divided [53].

Indeed, the fact that the incongruent motor dual task (IM) was the most sensitive condition for detecting group differences is consistent with literature showing that incongruent or competing motor demands produce the strongest breakdown of automatic locomotor patterns [69,70].

One of the most important limitations of this study is the small number of participants enrolled in each group. Each clinical condition included in this pilot study encompasses heterogeneous subtypes (e.g., stroke etiology, Parkinson’s phenotypes, Long-COVID presentations), and the small sample size did not allow stratification by subtype, age, or sex. Accordingly, the study was not designed nor powered to test age- or sex-related effects, and sex distribution was not controlled for inferential comparisons. Participants should therefore be considered representative only of individuals with similar functional walking abilities, and the present findings cannot be generalized to all subtypes within each condition but should instead be regarded as hypothesis-generating. However, we analyzed all the subjects with respect to their ability level, and the bootstrap analysis strengthens the reported interpretations of results, confirming them. Even if only about 11% of fixed effects were classified as robust, which is expected in pilot studies with small, heterogeneous groups, the stable effects predominantly involved the IC group and the cognitive load factor. The remaining fragile or non-robust effects likely reflect limited precision rather than the true absence of dual-task impact [71,72].

When spatiotemporal gait parameters were divided by left and right sides, controlled by the right- and left-brain hemispheres, respectively, we found that statistically significant differences were similar between the two sides.

Finally, regarding the usability of the dual-task protocol we designed, all groups expressed positive evaluations. Specifically, for both motor and cognitive tests, healthy adults, both young and older, consistently reported low levels of cognitive, physical, and temporal demand, as well as low effort and frustration, indicating that the tasks were feasible and did not impose excessive demands on non-clinical populations. As expected, the reported workload in individuals with motor disabilities was higher, particularly in the stroke and multiple sclerosis groups, who showed higher cognitive, physical, and effort demands than the other groups in both dual-task conditions, accompanied by greater self-reported frustration and perception of temporal dominance at the NASA-TLX assessment. This pattern suggests that the simultaneous management of motor and/or cognitive demands imposes a significant subjective burden on this group, reflecting difficulties in coordinating movement, maintaining temporal control, and sustaining performance under additional attention load. Participants with MS also reported higher levels of frustration as measured by the NASA-TLX frustration domain, though remaining within the low-tolerance range, across both dual-task categories. Overall, the combined workload profiles indicate that the two dual-task paradigms elicited measurable gait interference while remaining feasible and tolerable across groups.

Nevertheless, several limitations of this pilot study must be acknowledged. The most important limitation is the small sample size of each group of participants, combined with the heterogeneity between and within groups. It implies high risks of bias and limits the generalizability of the results, limiting the precision of group comparisons, as reflected by the limited robustness of some fixed-effect estimates. In addition, the quasi-periodic auditory pacing embedded in the cognitive task may have induced unintended entrainment effects on gait timing.

5. Conclusions

Despite the limitations of the study described above, the present findings suggest that the effects of adding a secondary task, as performed in dual-task protocols, mainly affect temporal parameters, without altering the stance-to-swing proportionality. Cognitive–motor interference acts through shared sensorimotor mechanisms that generalize across diagnostic categories, while still amplifying each individual’s pre-existing vulnerabilities.

Temporal gait parameters seem to emerge as the most sensitive to interference, highlighting the central role of executive and attentional resources in maintaining gait rhythm, whereas spatial parameters seem to be comparatively robust, reflecting underlying biomechanical constraints. This suggests that interventions aimed at strengthening temporal rhythms, such as rhythmic auditory stimulation, attentional cueing, or executive-function training [73,74], may be particularly effective in improving dual-task walking performance. Likewise, the relative stability of spatial parameters highlights the possibility of using stride length as a reliable target for monitoring progression or treatment response.

The possibility of exposing individualized interference profiles suggests that dual-task gait could be used not only as a general principle of rehabilitation but also as a means of tailoring it to specific patient needs, thus enhancing automaticity and/or stability, reducing attentional cost, or addressing fatigue according to the functional ability of each patient. As such, the dual-task approach could be a promising basis for the development of personalized, mechanism-driven interventions in which motor and cognitive demands are systematically integrated in a controlled and optimizable manner for neurorehabilitation.