The participants consisted of 16 older adults ranging in age from 50–90 years. Data was collected at the Speech Physiology and Neurobiology of Aging and Dementia (SPaN-AD) Lab as part of an ongoing longitudinal study. All participants in the study held an education level of at least 16 years. All participants were native speakers of American English and exhibited normal cognition as assessed by a neuropsychologist. The cognitive status of the participants was examined by a comprehensive array of standard neuropsychological tests that assess attention, processing speed, learning, memory, language, and executive function. Details of the neuropsychological assessments are provided by Eshghi et al. (51). In addition, all participants scored 0 on the Clinical Dementia Rating (CDR) scale global score and less than 11 on the Geriatric Depression Scale (GDS). All participants exhibited normal, age-appropriate hearing status in at least one ear as verified by a pure tone audiometer (Earscan 3, Micro Audiometric Corp.) and normal visual acuity (52). None of the participants showed a clinical diagnosis of MCI, any neurological condition, or were using any psychoactive drugs. Participants were genotyped with respect to the APOE-ɛ4 gene and were put into two groups: (1) carriers of APOE-ɛ4 who possess heterozygous ɛ3–ɛ4 alleles or homozygous ɛ4–ɛ4 alleles and (2) noncarriers of APOE-ɛ4 who possess homozygous ɛ3–ɛ3 alleles. Characteristics of the participants in each group are provided in Table 1. The study was reviewed and approved by the MGB Institutional Review Board.

Lower lip movements were recorded during sentence production using three-dimensional electromagnetic articulography (EMA) (Wave; Northern Digital, Inc.). The EMA system tracked a small (2 mm) electromagnetic sensor affixed to the midline of the lower lip’s vermilion border, using a sampling rate of 100 samples per second. Participants performed color-word interference speech tasks while seated in a supported chair. Stimuli were displayed on a wall-mounted screen positioned approximately six feet in front of them. Audio recordings were obtained using a Williams Sound MIC094 head-mounted microphone positioned approximately 5 cm from the left corner of the lips. The signal was routed through a Focusrite Scarlett 2i2 audio interface and digitized at a sampling rate of 10,000 Hz. Audio and EMA kinematic data were recorded synchronously to ensure precise temporal alignment. Participants were instructed to read sentences at their natural speaking rate and loudness.

The speech samples consisted of a block of four sentences containing color words, with the font color of each word being incongruent with its semantic meaning (see Table 2). All sentences had comparable syntactic structures, each consisting of 14 words and 18 syllables. To align with the study’s focus on lip movement properties, the sentences contained a high proportion of bilabial consonants.

Participants were presented with eight blocks of color-word interference sentences. Within each block, the same four sentences were presented in a randomized order. The font color for each color word was also randomly changed across blocks. However, the font color never matched the semantic meaning of the written word in any of the sentences. Two of the four sentences were designated for analysis, while the remaining two served as foil sentences. Data from these foil sentences were collected but not included in the final analyses. Foil sentences were included to reduce predictability and prevent participants from anticipating target sentences, thereby promoting natural speech patterns and improving the reliability of kinematic and acoustic measurements. The speech stimuli were adapted from MacPherson (53), with adjustments made to align with the objectives of the current study.

Participants were instructed to read each sentence while disregarding the conflicting font color, requiring them to suppress the influence of incongruent visual information. This aspect of the task imposed a heightened cognitive load compared to non-interfered parts of the sentences. The increased cognitive demand specifically engaged key cognitive domains, including inhibition (overcoming the automatic tendency to process the font color), selective attention (focusing on the written word while filtering out irrelevant visual distractions), and working memory (maintaining task rules and resolving conflicts between competing stimuli), all of which are susceptible to impairment in MCI and AD (54, 55).

SMASH software (56) was used to extract lip kinematic features from the three-dimensional lip movement time series recorded during the production of the entire target sentences. Lip movement duration (s), lip average speed of movement (mm/s), and lip range of movement (mm³) were calculated from three distinct sentence segments: the segment preceding the color-interference task (pre-interference), the segment corresponding to the color-interference task (during-interference), and the segment following the color-interference task (post-interference). By assessing kinematic measures of lip movements during these segments, we aimed to gain insights into the anticipatory effects (pre-task), direct effects (during task), and subsequent effects (post-task) of the heightened cognitive load imposed on the motor system.

Additionally, difference measures, including Δ-movement duration, Δ-average speed, and Δ-range of movement, were calculated to assess the impact of the interference task. Difference measures were computed by subtracting the kinematic values of the during-interference segment (i.e., the segment containing color interference) from those of the pre- and post-interference segments. These measures are hereafter referred to as ∆_During-Pre and ∆_During-Post. These measures provide valuable insights into how individual speakers adjust their motor planning and execution in response to the varying cognitive demands of each sentence segment.

Descriptive statistics were computed for each speech kinematic feature within sentence segments. Independent t-tests were performed to assess group differences. To adjust for multiple comparisons across 15 features, a Bonferroni-corrected threshold (adjusted p-value = 0.003) was applied to control the family-wise type I error rate. In addition, to quantify the magnitude and direction of these differences, standardized effect sizes were calculated using both Cohen’s d and Hedges’ g, with the latter adjusting for potential small sample bias to provide a more accurate estimate. Although the primary objective was to classify APOE-ε4 carriers (ε4⁺) and non-carriers (ε4⁻) using machine learning models trained on kinematic speech features, we also report descriptive and inferential statistics to enhance interpretability and contextual insight into group-level patterns. To evaluate the adequacy of our sample size for detecting group-level differences in kinematic speech features between APOE-ε4 carriers and non-carriers, we conducted post hoc power analyses using the observed effect sizes (Cohen’s d) for each feature comparison. For each analysis, the target statistical power was set to 0.80, and the alpha level was set to 0.05, corresponding to conventional thresholds for reliable detection of true effects. These post hoc estimates provide an approximate indication of the sample size required to detect significant differences observed in this study, thereby informing the design of future, adequately powered investigations and enhancing methodological transparency. Given the limited sample, findings from the machine learning classification are reported as exploratory and hypothesis-generating, highlighting potential group-level patterns that merit independent validation in larger cohorts. All statistical analyses were performed in MATLAB (Version 24.2, R2024b; The MathWorks Inc., Natick, MA, United States).

The purpose of this study was to identify lip kinematic features recorded during color-word interference speech tasks that can accurately classify APOE-ε4⁺ and APOE-ε4⁻ individuals. Three sets of different features were extracted to capture the dynamic characteristics of lip movement across task conditions. Specifically, lip movement duration, average speed of movement, and range of movement features were calculated (1) across the pre-, during-, and post-interference segments of the color-interference task, (2) on difference values from segments preceding the interference task (∆_During-Pre), and (3) on difference values from segments following the interference task (∆_During-Post), yielding a total of 15 features. These 15 features are listed in Table 3. Classification of APOE-ε4 carriers (ε4⁺) versus non-carriers (ε4⁻) was performed using a support vector machine (SVM) classifier with a degree-2 polynomial kernel. An exhaustive search across all 32,767 possible combinations of 15 extracted speech kinematic features was systematically conducted, with each combination evaluated using leave-one-out cross-validation. Feature sets were ranked according to overall classification accuracy, and the optimal set comprised: (1) lip movement duration during the pre-interference segment, (2) average lip speed during the interference segment, and (3) the change in lip movement range from pre- to during-interference segments. This specific combination was selected solely on the basis of its superior performance in the exhaustive search, ensuring robust and discriminative representation of task-related variation. No further model tuning or subjective filtering was performed, and all analytical procedures were implemented in MATLAB (Version 24.2, R2024b; The MathWorks Inc., Natick, MA, United States).

Altered articulatory movement duration observed during the pre-interference segment reflects anticipatory motor adjustments and resource allocation in preparation for upcoming cognitive interference. This has been attributed to proactive modulation of speech motor planning, where speakers change (i.e., typically slow down) articulatory movements to buffer against predicted instability and to facilitate error monitoring as cognitive demands increase (62, 63). Predictive feedforward mechanisms within the motor system further contribute by prolonging movement duration to accommodate potential error margins under cognitive stress (64). Consistent with this interpretation, our descriptive data showed that pre-interference lip movement duration was moderately shorter in ε4⁺ individuals (d = −0.53), suggesting altered anticipatory motor strategies in this group. The inclusion of this feature in the final classification model underscores the potential diagnostic utility of pre-task articulatory dynamics in identifying APOE-ε4 status.

Similarly, prior research suggests that individuals under cognitive interference often adopt more conservative articulation strategies, such as slower and larger articulatory excursions, to maintain intelligibility and motor stability (53, 65, 66). In our study, the ∆_During-Pre difference in movement range emerged as a key feature, suggesting that changes in articulatory flexibility under cognitive load are informative for group classification. Although the machine learning model does not directly quantify the direction of group-level changes, our descriptive analysis revealed a smaller absolute ∆_During-Pre difference in lip movement range in the ε4⁺ group (0.205 mm³) compared to the ε4⁻ group (0.609 mm³), with a moderate effect size (d = −0.371). This reduced dynamic modulation may indicate diminished neuromotor adaptability to cognitive demands in ε4⁺ individuals.

Altered articulatory speed during the interference condition of a cognitive-motor task has been linked to increased susceptibility to cognitive-motor conflict, reflecting difficulty in maintaining speech precision while resolving competing demands. Prior research suggests that when cognitive control resources are heavily taxed, neural resources are reallocated from motor execution to processes such as conflict monitoring and inhibition, resulting in slower and less precise articulatory movements (67–70). These effects are especially pronounced in at-risk populations, including older adults and individuals with executive dysfunction, who tend to exhibit greater slowing and variability in movement during cognitive interference (71), further highlighting the close interplay between executive function and speech motor control (67–70). This pattern aligns with findings from dual-task gait studies, where APOE-ε4 carriers show slower walking speeds and greater reductions in step length when simultaneously engaged in executive function tasks (72). Although these findings focus on gross motor behavior, they illustrate a broader compensatory mechanism in which motor systems prioritize stability, often through slower or more deliberate movements, under increased cognitive load. In the present study, average lip movement speed during the interference segment did not differ significantly between groups (d = −0.18), but it contributed substantially to classification accuracy, reinforcing its relevance as a speech timing feature linked to cognitive load sensitivity.

Given that the two groups were matched by age, it is unlikely that the observed differences in speech kinematics are attributable to aging effects. However, the interaction between sex and APOE-ε4 status (as discussed above) may still substantially influence these kinematic outcomes. The trends observed in this study may therefore reflect not only genotype-related vulnerability, but also distinct sex-specific patterns in motor and cognitive adaptation that warrant further investigation. Future research should aim to disentangle the independent and interactive effects of genotype and sex on anticipatory and adaptive speech motor strategies.

These findings suggest the hypothesis that subtle speech motor differences in cognitively intact APOE-ε4 carriers may be detectable via articulatory kinematic measures, particularly under mild cognitive load. The model’s cross-validated classification accuracy of 87.5%, with corresponding precision (88.9%), sensitivity (88.9%), and specificity (85.7%), indicates that speech-derived features could potentially identify ε4⁺ individuals. Importantly, these metrics reflect the methodological approach in a limited dataset and must be interpreted as exploratory rather than confirming clinical utility.

Unlike traditional neuropsychological assessments that focus on cognitive decline, speech kinematics offer a behaviorally grounded, low-burden, and non-invasive window into early neuromotor changes, especially those affecting articulatory timing and precision. These features show promise as early indicators of genetic risk and may augment existing multimodal approaches to stratification, screening, and individualized risk monitoring. While speech kinematics show promise as a low-burden and scalable approach for risk stratification, the present study is underpowered to make clinical claims. Establishing clinical utility will require: (1) replication in larger and demographically diverse samples, (2) independent validation cohorts, (3) longitudinal tracking to determine predictive value for cognitive decline, and (4) comparison with established diagnostic methods and biomarkers. Until such work is completed, the present findings should be viewed as preliminary and hypothesis-generating.

Importantly, our results suggest that APOE-ε4 carriers with intact cognition may already exhibit early motor abnormalities, reinforcing the idea that motor deficits can precede overt cognitive decline. This aligns with previous research showing that individuals with the APOE-ε4 allele experience accelerated motor decline independent of detectable cognitive impairment (6, 25). The underlying mechanisms likely involve multiple neurobiological pathways. APOE-ε4 is strongly associated with increased amyloid-β deposition and tau pathology, which disrupt neural circuits critical for motor control (73–75). Additionally, synaptic loss and dysfunction in motor regions may impair the fine motor coordination necessary for speech production (73, 75). Neuroinflammation and cerebrovascular dysfunction, both characteristic features of APOE-ε4 carriers, could further contribute to early motor deficits (76–78). While our findings provide indirect behavioral evidence of such disruptions, future work integrating speech kinematics with imaging and molecular markers will be necessary to clarify the mechanistic basis of these early articulatory changes.

Several limitations of this study should be acknowledged. First, the small sample size limits statistical power and reduces the ability to detect significant group differences or reliably assess the directionality of changes in individual kinematic features. This also increases the risk of overfitting in the machine learning model, restricts the broader interpretation of clinical implications, and limits generalizability. Therefore, the results should be viewed as hypothesis-generating that require replication and validation in larger, independent samples. Second, the cross-sectional design precludes evaluation of whether speech kinematic differences in APOE-ε4 carriers evolve longitudinally or predict future cognitive decline. Third, the absence of neuroimaging or other biological markers restricts mechanistic interpretation, making it difficult to determine whether the observed speech motor changes are linked to specific brain network alterations associated with APOE-ε4. Fourth, the lack of an MCI comparison group limits the ability to differentiate whether the articulatory changes observed are specific to preclinical AD risk or reflect broader patterns of cognitive aging. The unbalanced gender distribution within the ε4⁺ group represents a major confound. Given well-documented sex differences in APOE-ε4 effects, this imbalance reduces both generalizability and the ability to clarify potential sex contributions to speech kinematic changes. Our results may therefore reflect a combination of genotype and sex-related effects. Future studies should ensure sex-balanced recruitment and incorporate stratified analyses to disentangle sex-specific influences on speech motor outcomes in APOE-ε4 carriers. Finally, although the two groups were age-matched in our study, minimizing confounding due to general age-related motor changes, the observed differences in lip kinematics could nonetheless reflect broader motor control variability rather than AD-specific neural pathways. However, without corroborating neuroimaging or biomarker data, our findings should be interpreted as preliminary behavioral trends. Future research should incorporate multimodal designs integrating speech kinematics, neuroimaging, and fluid biomarkers to clarify the neurobiological underpinnings of these effects.