Community-based assisted screening for mild cognitive impairment using gait and handwriting kinematic parameters analysis

Abstract

BACKGROUND

Older adults with mild cognitive impairment (MCI) often show motor dysfunction, including slower gait and impaired handwriting. While gait and handwriting parameters are promising for MCI screening, their combined potential to distinguish MCI from cognitively normal adults is unclear.

AIM

To assess gait and handwriting differences and their potential for screening MCI in older adults.

METHODS

Ninety-five participants, including 34 with MCI and 61 cognitively normal controls, were assessed for gait using the GAITRite^® system and handwriting with a dot-matrix pen. Five machine learning models were developed to assess the discriminative power of gait and handwriting data for MCI screening.

RESULTS

Compared to the cognitively normal group, the MCI group had slower gait velocity (Z = -2.911, P = 0.004), shorter stride and step lengths (t = -3.005, P = 0.003; t = 2.863, P = 0.005), and longer cycle, standing, and double support times (t = -2.274, P = 0.025; t = -2.376, P = 0.018; t = -2.717, P = 0.007). They also had reduced cadence (t = 2.060, P = 0.042) and increased double support time variability (Z = -2.614, P = 0.009). In handwriting, the MCI group showed lower average pressure (all tasks: Z = -2.135, P = 0.033) and decreased accuracy (graphic task: Z = -2.447, P = 0.014; Chinese character task: Z = -3.078, P = 0.002). In the graphic task, they demonstrated longer time in air (Z = -2.865, P = 0.004), reduced X-axis maximum velocities (Z = -3.237, P = 0.001), and lower accelerations (X-axis: Z = -2.880, P = 0.004; Y-axis: Z = -1.987, P = 0.047) and maximum accelerations (X-axis: Z = -3.998, P < 0.001; Y-axis: Z = -2.050, P = 0.040). The multimodal analysis achieved the highest accuracy (74.4%) with the Gradient Boosting Classifier.

CONCLUSION

Integrating gait and handwriting kinematics parameters provides a viable method for distinguishing MCI, potentially supporting large-scale screening, especially in resource-limited settings.

Key Words: Mild cognitive impairment; Early detection; Digital health; Gait; Handwriting

Core Tip: This study introduces sensitive biomarkers for assisted screening of mild cognitive impairment (MCI) by integrating gait and handwriting kinematic parameters. The findings demonstrate that the combination of gait analysis and handwriting tasks achieves a detection accuracy of 74.44% for MCI, outperforming single-task assessments. These parameters could complement traditional tool like the Montreal cognitive assessment by reducing reliance on subjective assessments and minimizing cultural, language, and educational biases. This research establishes a foundation for translating these precision biomarkers into cost-effective, portable, and scalable wearable technologies, which could enable large-scale, rapid screening of high-risk individuals with MCI in community settings.

INTRODUCTION

Mild cognitive impairment (MCI) represents an intermediate stage between normal aging and dementia, without substantial impact on daily activities or meeting the criteria for dementia diagnostic[1]. The global prevalence of MCI over the age of 60 is currently 17.3%, with a higher prevalence of 26.48% reported in China[2,3]. Estimates suggest that 34% of individuals with MCI progress to dementia[4], while 27.57% experience a reversion to normal cognitive function[5]. However, 99.2% of older adults with MCI are unaware of their condition, and 97.2% remain undiagnosed and untreated, highlighting a critical gap in early detection and intervention, likely due to insufficient public awareness and the lack of screening tools with greater accessibility[6]. Current MCI screening tools primarily rely on scales, such as the Mini-Mental State Examination (MMSE) and Montreal cognitive assessment (MoCA). However, the specialized nature, complexity, extended administration time, and requirement for professional assessment have limited their widespread use, particularly in primary care settings[7,8]. Therefore, there is a pressing need for the development of more accessible and practical screening solutions.

Cortical function decline due to specific brain atrophy affects both cognitive and motor functions[9]. Older adults with MCI exhibit slower gait speeds, shorter stride and step lengths, and longer standing and double support times compared with cognitively normal individuals[9,10]. Previous studies have shown that a decline gait speed often precedes cognitive decline, indicating its potential as a biomarker for cognitive impairment[9,11]. Given its objectivity and practicality, gait testing may be a valuable tool for cognitive impairment screening[12]. Currently, machine learning analysis of gait data alone achieves an accuracy of approximately 70% in distinguishing older adults with MCI from those with normal cognition[13]. Lin et al[14] found that combining gait and eye movement features improved the area under the receiver operating characteristic curve (AUC) to 0.74. These findings indicate that while gait kinematic parameters are valuable for early MCI screening, and a multimodal approach offers greater accuracy than gait data alone.

Additionally, motor dysfunction can also manifest as upper extremity fine motor deficits[15]. Writing, a complex process involving fine motor skills, kinesthetic components, and multiple cognitive domains[16], is often affected in older adults with MCI. Compared with cognitively normal individuals, those with MCI exhibit slower writing speed, smaller characters, longer pauses between strokes, reduced pressure, and decreased writing accuracy[17,18]. However, due to cultural differences and varying writing habits, the English alphabet and Chinese characters differ in writing requirements and complexity. Among older adults with MCI, the control of stroke positions in Chinese character writing is notably impaired[17], with significantly higher average absolute speed, stroke pressure, and in-air length tortuosity, particularly during the writing of Chinese personal names[19]. These observations underscore the importance of considering cultural differences when developing handwriting assessment tasks. Notably, previous studies have primarily used handwriting kinematic parameters analysis to differentiate Alzheimer's disease from cognitively normal older adults and is less frequently applied in those with MCI, with reported accuracy ranging from 45-83% in identifying MCI[18,20].

While single-modality analysis of gait and or handwriting kinematic parameters has proven effective in distinguishing older adults with MCI from cognitively normal individuals, the potential enhancement of screening accuracy through the integration of these modalities remains to be fully elucidated. Including both gait and handwriting kinematic parameters provides a more thorough assessment of motor function, ensures that both gross and fine motor skills are considered, and helps capture the full extent of motor impairment or improvement. In addition, machine learning-based methods have shown promise as valuable tools for cognitive impairment screening[21]. Building on the diagnostic relevance of gait and handwriting kinematic parameters, this study is the first time to apply machine learning algorithms for an integrated analysis of these multimodal data in screening older adults with MCI. The aim of this study is to develop an objective, accessible, and accurate screening tool. Two hypotheses are proposed: (1) Gait and handwriting kinematic parameters differ significantly between older adults with MCI and cognitively normal individuals; and (2) A multimodal approach combining gait and handwriting kinematic parameters offers superior accuracy for MCI screening compared to unimodal methods.

MATERIALS AND METHODS

Study design and ethical considerations

A cross-sectional design was used in this study to examine differences in gait and handwriting kinematics and evaluate their potential for MCI screening among older adults in Huzhou city, Zhejiang Province, China, with data collected from August 2024 to October 2024. This study was approved by the Institutional Review Board of the third affiliated hospital of Huzhou university (2024-474) in accordance with the guidelines of the Declaration of Helsinki and registered at Chictr.org.cn (ChiCTR 2400089057). All eligible participants voluntarily signed the written informed consent after being fully informed.

Participants

Community-dwelling adults aged 60 years or older were recruited from senior centers and healthcare centers in 2 randomly selected districts of Huzhou, Zhejiang Province, China. Recruitment leaflets were distributed at these centres, and local healthcare providers referred potential participants. Individuals who showed interest were invited for an in-person interview to screen for eligibility by three trained staff, and a trained neurologist-psychiatrist examined the potential participants to provide the final diagnosis of MCI.

The inclusion criteria for the gait and handwriting tasks were: (1) Age ≥ 60 years old; (2) Elementary school level and above; (3) Absence of self-reported visual or auditory impairment; (4) Right-handed dominance[17]; and (5) Able to make an informed consent. The exclusion criteria included: (1) A history of neurological, psychiatric and other severe medical issues that may affect brain function; (2) Taking any medications in the past 6 months which may cause impaired or improved cognitive performance (e.g., memantine and antipsychotic)[22]; (3) Inability to walk independently or walking independently with the aid of auxiliary tools; (4) Unable to perform handwriting tasks due to hand movement disorders; and (5) Recent or upcoming mobility-impairing medical procedures (e.g., hemodialysis and elective surgery). Participants could withdraw if they developed a serious illness or requested to leave the study.

The operational criteria for MCI screening included[23]: (1) A reported decline in cognitive function over the past year by either the participant or an informant; (2) Objective cognitive impairment inconsistent with age and education, assessed by the Chinese MoCA-basic (MoCA-B) with cutoffs of 18 for elementary, 21 for secondary, and 23 for tertiary education[24]; (3) Intact activities of daily living (ADL), evaluated by physical self-maintenance with six items and instrumental activities with eight items, with a total ADL score ≤ 16 indicating intact ADL; and (4) Absence of dementia, determined by community physicians using DSM-IV criteria. This assessment considered family/medical history, chief complaints, and the MMSE scores with thresholds of ≥ 17 for illiterate, ≥ 20 for primary, and ≥ 24 for secondary education and above. Individuals without signs of MCI were classified as cognitively normal (CN group). During the initial eligibility assessment, 243 potential participants with independent walking and writing ability were recruited. The second round included assessments of functional status in ADL, general cognitive function, medical health and medication history. A total of 95 eligible participants were included in this study.

Assessment and data collection

Demographic characteristics: Demographic characteristics included age, sex, education level, occupation, marital status, residence, monthly income, region of residence, smoking and alcohol use, general cognitive function, physical activity, and multimorbidity.

Gait and handwriting kinematic parameters

The dataset comprised three unimodal types: Gait kinematic parameters, graphic kinematic, and Chinese character writing kinematic parameters. Multimodal data were subsequently synthesized by integrating gait with graphic and Chinese character handwriting data.

Measurement tools: Gait kinematic parameters were collected using the GAITRite^® system (GAITRite platinum plus classic, CIR systems Inc., Havertown, PA, United States), a portable instrumented walkway with built-in pressure sensor (as shown Figure 1A). The walkway measures 7 meters in length and 0.9 meters in width, of which the effective length is 6.1 meters and the width is 0.61 meters, containing over 18000 sensors with a sampling rate of 120 Hz. As participant walks, the system continuously recorded spatiotemporal gait data in real-time, including velocity, cadence, step length, cycle time, stride length, stance time, double support time, swing time, and step time. The GAITRite^® system is factory-calibrated and undergoes regular maintenance as recommended by the manufacturer’s guidelines to ensure data accuracy and reliability. Its validity and reliability for measuring spatial and temporal gait parameters have been demonstrated in previous studies[25,26]. Handwriting kinematic parameters were recorded using dot-matrix digital pens and dot-matrix paper (TSTUDY, China), which provide coordinate parameter information for movement trajectory (Figure 1B) and have been validated for reliable handwriting analysis[27,28]. The device was also factory-calibrated and maintained according to the manufacturer’s guidelines. The digital pen is equipped with built-in pressure sensor and camera that is activated upon contact with the dot-matrix paper, enabling the real-time recording of writing pressure and speed. Additionally, the tool mimics traditional pen-and-paper writing, matching participants’ writing habits. Studies have shown that declines in geometric copying and handwriting skills correspond with cognitive decline, making these tasks useful for screening cognitive impairment in older adults; and notably, the pentagon and cube copying tasks demonstrating greater sensitivity to cognitive change than other graphic tasks[29]. Given the participants’ familiarity with Chinese characters, two graphic handwriting tasks and four Chinese character handwriting tasks were set in this study (Supplementary Table 1).

Figure 1 Measurement tool. A: Measurement tool for gait parameters; B: Measurement tools for handwriting parameters.

Measure methods: Gait kinematic parameters were collected in a quiet, well-lit environment with participants wearing comfortable footwear. Start and end markers were placed 1 meter from the walkway to allow acceleration and deceleration. The researcher briefed participants to ensure their comprehension before walking tests. To minimize emotional influences, participants walked the walkway 3-5 times at their usual pace, with the mean of the three most consistent trials used for final analysis[30]. Handwriting kinematic parameters were recorded with digital pens, with participants completing two copying graphic handwriting tasks: Pentagons and a cube, as shown in Tasks 1-2 (Supplementary Table 1). Task 1 was scored up to 3 points, Task 2 had a maximum of 4 points, resulting in a total graphic handwriting task score of 7 points. Another handwriting task involved writing Chinese characters, which is designed to identify cognitive deficits and motor dysfunctions in individuals with neuropsychiatric disorders[19]. Based on the task setting of previous studies and taking into account the writing habits and writing features of Chinese older adults[19,31], four Chinese character handwriting tasks were set up, including signature (Task 3), memorizing and writing three words (Task 4), naming three pictures (Task 5), and copying a sentence (Task 6). The content and scoring methods for handwriting tasks were in Supplementary Table 1. Handwriting data collection was also conducted in a quiet, well-lit setting, and participants were briefed on the task to ensure understanding before writing in their usual style.

Parameters extraction: A total of 14 gait kinematic parameters were extracted, including 1 item in velocity (velocity), 8 in the time (ambulation time, step time, cycle time, swing time, stance time, double support time, swing time variability, double support time variability), 4 in the length [step length, stride length, heel-to-heel (HH) base support, stride length variability], and 1 item in other (cadence). Twenty-six handwriting kinematic parameters were extracted: 1 item in pressure (average pressure), 13 in velocity (average velocity, average velocity in X/Y-axis, maximum velocity in X/Y-axis, minimum velocity in X/Y-axis, average acceleration in X/Y-axis, maximum acceleration in X/Y-axis, minimum acceleration in X/Y-axis), 2 in time (time in air, single stroke time), 6 in length (single stroke length, X/Y-axis stroke lengths, total single stroke length, total X/Y-axis stroke lengths), and 4 items in others (tilt angle, number of times velocity slows down, number of times acceleration slows down, task score). Supplementary Tables 2 and 3 for parameter definitions.

Statistical analysis

Participant demographic differences, along with gait and handwriting kinematic parameters between groups, were analyzed using appropriate statistical tests based on variable type and distribution. Categorical variables were analyzed using χ² tests. For continuous variables, normality was assessed using the Shapiro-Wilk test. Normally distributed data were analyzed using independent samples t-tests, while non-normally distributed data were analyzed using Wilcoxon rank-sum tests. Statistical significance was set at P < 0.05.

To evaluate the classification accuracy between MCI and CN, five classification models were used, including logistic regression (LR), eXtreme Gradient Boosting (XGB), random forest (RF), gradient boosting classifier (GBC), and Gaussian Naive Bayes (Gaussian NB). Default hyperparameters were applied, with manual adjustments as needed. Variance thresholding and sequential feature selection algorithms were implemented to identify the optimal gait and handwriting parameters subsets. Model performance was assessed using ten-fold cross-validation, with evaluation metrics including accuracy, sensitivity, specificity, positive predictive value, negative predictive value, receiver operating characteristic curve, AUC, brier score, and decision curve analysis (DCA). Statistical analyses were conducted using SPSS 27.0, and PyCharm Community Edition 2022.2 × 64.

RESULTS

Participant demographics characteristics

Ninety-five participants were recruited, including 34 with MCI and 61 cognitively normal (selection process in Figure 2). Significant differences were found in MMSE and MoCA-B scores, while other characteristics were comparable. Table 1[32,33] summarized the demographics characteristics of all participants.

Figure 2 Study flow chart. MCI: Mild cognitive impairment; ROC: Receiver operating characteristic; AUC: Area under the receiver operating characteristic curve; DCA: Decision curve analysis; PPV: Positive predictive value; NPV: Negative predictive value.

Table 1 Participant demographics characteristics between groups, mean ± SD/n (%).

Basic characteristics	MCI group (n = 34)	CN group (n = 61)	t/χ2/Z	P value
Age	84.26 ± 7.85	80.26 ± 11.02	-1.868	0.065
Sex			0.623	0.430
Male	8 (23.53)	19 (31.15)
Female	26 (76.47)	42 (68.85)
Education level			-0.953	0.340
Primary school	10 (29.41)	23 (37.70)
Junior school	10 (29.41)	17 (27.87)
High School	9 (26.47)	16 (26.23)
College and above	5 (14.71)	5 (8.20)
Occupation1			0.411	0.522
Brain work	25 (73.53)	41 (67.21)
Physical work	9 (26.47)	20 (32.79)
Marital status			0.012	0.914
Single (divorced, widowed)	16 (47.06)	28 (45.90)
Married	18 (52.94)	33 (54.10)
Residence			2.473	0.116
Living alone	10 (29.41)	28 (73.68)
Living with spouse/children	24 (70.59)	33 (54.10)
Monthly income			-0.625	0.532
Low (≤ 1000)	1 (2.94)	0 (0.00)
Middle (1000-3000)	1 (2.94)	2 (3.28)
High (≥ 3000)	32 (94.12)	59 (96.72)
Region of residence			1.212	0.271
Urban	32 (94.12)	53 (86.89)
Rural	2 (5.88)	8 (13.11)
Physical activity2			-1.456	0.145
Low	19 (55.88)	25 (40.98)
Middle	11 (32.35)	24 (39.34)
High	4 (11.76)	12 (19.67)
Family history of Dementia	2 (5.88)	3 (4.92)	0.041	0.840
Current smoker	0 (0.00)	2 (3.28)	1.139	0.286
Alcohol Consumption	0 (0.00)	6 (9.84)	3.570	0.059
Multimorbidity3	14 (41.18)	26 (42.62)	0.019	0.891
MMSE	25.03 ± 2.04	27.07 ± 2.12	4.549	< 0.001
MoCA-B	17.35 ± 2.89	23.23 ± 2.53	10.318	< 0.001

¹Occupation: Brain work include civil servant, teacher, healthcare professional, accountant and police officer[32]; Physical work include farmer, laborer, employee and tailor.

²Use of the International Physical Activity Questionnaire-Short form questionnaire[33].

³Multimorbidity: ≥ 2 comorbidities.

MCI: Mild cognitive impairment; CN: Cognitively normal; MMSE: Mini-Mental State Examination; MoCA-B: Montreal cognitive assessment basic.

Comparison of gait and handwriting kinematic parameters between groups

Comparisons of kinematic parameters between the MCI and CN groups during gait and handwriting tasks revealed significant differences in gait parameters, including velocity, cadence, step length, cycle time, stride length, stance time, double support time, and double support time variability (all P < 0.05); In handwriting tasks involving graphic tasks, significant differences were found in average pressure, time in air, maximum velocity in the x and Y-axis, average acceleration in the X-axis, maximum acceleration in the x and Y-axis, and task score (all P < 0.05); in Chinese character handwriting tasks, both average pressure and task score were statistically significant between the two groups (all P < 0.05). The above results showed that compared with the CN older adults, individuals with MCI exhibited slower walking speeds, shorter stride lengths, and reduced step lengths, alongside longer standing and cycle times when walking. In handwriting tasks, they demonstrated slower writing speeds, lower pressure, extended thinking times, and an increased number of writing errors. Detailed task-related kinematic parameters were provided in Tables 2, 3, and 4.

Table 2 Comparison of gait kinematic parameters between groups, mean ± SD/M (QR).

Parameters	MCI group (n = 34)	CN group (n = 61)	t/Z	P value
Ambulation time	25.08 ± 7.55	22.77 ± 7.77	-1.402	0.164
Velocity	84.55 (23.83)	93.30 (25.35)	-2.911	0.004
Cadence	106.50 ± 11.51	110.91 ± 9.05	2.060	0.042
Step time	0.57 (0.08)	0.54 (0.06)	-1.825	0.068
Step length	47.92 ± 6.42	52.56 ± 8.12	2.863	0.005
Cycle time	1.14 ± 0.12	1.09 ± 0.09	-2.274	0.025
Stride length	95.48 ± 13.35	105.52 ± 16.71	3.005	0.003
HH base support	11.24 (4.51)	10.72 (4.30)	-1.475	0.140
Swing time	0.40 ± 0.04	0.40 ± 0.03	-0.451	0.653
Stance time	0.73 (0.13)	0.69 (0.07)	-2.376	0.018
Double support time	0.32 (0.09)	0.28 (0.08)	-2.717	0.007
Stride length variability	3.25 (1.43)	3.13 (1.10)	-0.679	0.497
Swing time variability	0.02 (0.01)	0.02 (0.01)	-1.658	0.097
Double support time variability	0.03 (0.01)	0.02 (0.01)	-2.614	0.009

MCI: Mild cognitive impairment; CN: Cognitively normal.

Table 3 Comparison of graphic tasks kinematic parameters between groups, M (QR).

Parameters	Handwriting with graphic tasks
	MCI group (n = 34)	CN group (n = 61)	t/Z	P value
Tilt angle	-42.07 (22.51)	-39.79 (19.18)	-0.823	0.411
Average pressure	593.05 (258.14)	666.84 (172.65)	-2.135	0.033
Time in air	25920.00 (18039.00)	17468.00 (11452.00)	-2.865	0.004
Single stroke time	1526.07 (1304.64)	1605.43 (1058.23)	-0.116	0.907
Single stroke length	13.23 (7.44)	14.84 (7.15)	-1.623	0.105
X-axis stroke length	8.22 (4.41)	9.34 (4.69)	-1.669	0.095
Y-axis stroke length	8.32 (4.63)	9.61 (5.13)	-1.304	0.192
Total single stroke length	255.93 (71.32)	260.29 (77.99)	-0.155	0.877
Total X-axis stroke length, mean (SD)	167.93 ± 36.92	168.48 ± 35.30	0.072	0.943
Total Y-axis stroke length	157.47 (50.29)	157.95 (47.77)	-0.738	0.461
Average velocity	10.40 (4.93)	11.12 (5.53)	-1.623	0.105
Average velocity in X-axis	6.91 (2.92)	8.12 (3.46)	-1.933	0.053
Average velocity in Y-axis	7.59 (3.34)	7.71 (3.57)	-1.102	0.270
Maximum velocity in X-axis	29.07 (9.53)	36.22 (16.09)	-3.237	0.001
Minimum velocity in X-axis	0.14 (0.32)	0.08 (0.25)	-0.927	0.354
Maximum velocity in Y-axis	33.27 (16.84)	38.45 (17.21)	-1.723	0.085
Minimum velocity in Y-axis	0.13 (0.26)	0.11 (0.25)	-0.617	0.537
Number of times velocity slows down	26.33 (22.87)	28.95 (23.08)	-0.078	0.938
Average acceleration in X-axis	0.21 (0.07)	0.25 (0.09)	-2.880	0.004
Average acceleration in Y-axis	0.23 (0.11)	0.26 (0.08)	-1.987	0.047
Maximum acceleration in X-axis	1.07 (0.32)	1.38 (0.55)	-3.998	< 0.001
Minimum acceleration in X-axis	0.00 (0.01)	0.00 (0.01)	-0.163	0.870
Maximum acceleration in Y-axis	1.27 (0.64)	1.45 (0.65)	-2.050	0.040
Minimum acceleration in Y-axis	0.00 (0.01)	0.00 (0.01)	-0.513	0.608
Number of times acceleration slows down	26.19 (24.49)	30.21 (24.99)	-0.039	0.969
Task score	5.00 (3.00)	6.00 (2.00)	-2.447	0.014

MCI: Mild cognitive impairment; CN: Cognitively normal.

Table 4 Comparison of Chinese character tasks kinematic parameters between groups, mean ± SD/M (QR).

Parameters	Handwriting with Chinese character tasks
	MCI group (n = 34)	CN group (n = 61)	t/Z	P value
Tilt angle	-57.08 (9.06)	-54.93 (8.81)	-0.668	0.504
Average pressure	499.05 (213.36)	582.68 (233.95)	-2.135	0.033
Time in air	36336.00 (25185.00)	29228.00 (21512.00)	-1.394	0.163
Single stroke time	352.1 (152.30)	356.65 (160.66)	-0.303	0.762
Single stroke length	8.93 (3.87)	9.09 (5.10)	-0.753	0.451
X-axis stroke length	5.35 (2.53)	5.20 (3.10)	-0.738	0.461
Y-axis stroke length	6.02 (2.66)	6.26 (3.69)	-0.706	0.480
Total single stroke length	791.40 ± 151.25	829.42 ± 175.50	1.062	0.291
Total X-axis stroke length	455.63 ± 99.02	478.62 ± 99.95	1.078	0.284
Total Y-axis stroke length	521.39 (132.81)	557.07 (181.17)	-0.831	0.406
Average velocity	25.93 ± 8.70	27.69 ± 9.31	0.907	0.367
Average velocity in X-axis	17.29 ± 6.17	18.55 ± 6.47	0.919	0.360
Average velocity in Y-axis	21.28 ± 7.44	22.41 ± 7.89	0.680	0.498
Maximum velocity in X-axis	51.57 ± 17.45	55.05 ± 16.51	0.964	0.338
Minimum velocity in X-axis	1.47 (1.63)	1.58 (1.32)	-0.109	0.913
Maximum velocity in Y-axis	61.86 ± 20.04	65.30 ± 20.87	0.781	0.437
Minimum velocity in Y-axis	2.02 (1.55)	1.63 (1.27)	-0.606	0.545
Number of times velocity slows down	6.22 (2.97)	6.07 (3.19)	-0.008	0.994
Average acceleration in X-axis	0.59 (0.30)	0.61 (0.34)	-0.660	0.509
Average acceleration in Y-axis	0.67 (0.45)	0.72 (0.35)	-0.481	0.630
Maximum acceleration in X-axis	1.64 (0.93)	1.78 (0.99)	-1.196	0.232
Minimum acceleration in X-axis	0.08 (0.11)	0.07 (0.09)	-0.124	0.901
Maximum acceleration in Y-axis	1.97 (0.80)	2.08 (1.13)	-1.079	0.281
Minimum acceleration in Y-axis	0.11 (0.14)	0.10 (0.11)	-0.109	0.913
Number of times acceleration slows down	5.60 (3.11)	5.37 (3.07)	-0.093	0.926
Task score	20.00 (0.00)	20.00 (1.00)	-3.078	0.002

MCI: Mild cognitive impairment; CN: Cognitively normal.

Classification analysis

The unimodal classification task consists of two components: Kinematic parameters from the gait task and the handwriting task, which includes both graphics and Chinese character handwriting tasks. Considering that the graphic and Chinese character handwriting tasks have large number and consistent parameter categories, they are prone to overfitting[34], the multimodal classification task incorporates combinations of kinematic parameters from the gait and graphics handwriting tasks, as well as the gait and Chinese character handwriting tasks.

Unimodal analysis of kinematic parameters in gait tasks: All gait kinematic parameters were included in the unimodal classification model. In the gait tasks, the optimal gait parameter subsets differed across models: The LR model included velocity, step length, and cycle time; the XGB model included cadence and cycle time; the RF model included ambulation time, cadence, and cycle time; the GBC model included cadence, step time, and cycle time; and the Gaussian NB model included ambulation time, velocity, cadence, and step length. The GBC and Gaussian NB models demonstrated the highest classification performance, with an accuracy of 69.44% and an AUC of 0.68 (Table 5, Figure 3A). DCA showed that the RF and Gaussian NB models offered a greater net benefit in the testing cohort at a threshold probability range of 0.2 to 0.7. The DCA was shown in Supplementary Figure 1.

Figure 3 Receiver operating characteristic analysis for unimodal and multimodal classification in gait and handwriting tasks. A: Gait tasks; B: Graphic handwriting tasks; C: Chinese character handwriting tasks; D: Gait and graphics handwriting tasks; E: Gait and Chinese character handwriting tasks.

Table 5 Classification power of kinematic parameters in gait and handwriting tasks between groups.

	LR (%)	XGB (%)	RF (%)	GBC (%)	Gaussian NB (%)
Gait tasks
Accuracy	66.33	69.33	63.00	69.44	69.44
Sensitivity	29.17	49.83	44.50	54.00	50.67
Specificity	91.87	85.44	82.06	82.88	85.02
Positive predictive value	53.33	65.33	49.17	64.17	64.00
Negative predictive value	68.00	73.74	71.94	73.53	72.90
AUC	0.61	0.68	0.63	0.68	0.68
Brier	0.34	0.31	0.37	0.31	0.31
Graphics handwriting tasks
Accuracy	69.56	72.44	70.44	60.78	73.44
Sensitivity	54.17	64.00	57.33	43.67	66.00
Specificity	86.51	81.21	81.81	74.41	83.27
Positive predictive value	66.67	72.00	69.00	43.33	71.67
Negative predictive value	74.82	75.82	73.19	68.51	78.76
AUC	0.70	0.73	0.70	0.59	0.75
Brier	0.30	0.28	0.30	0.39	0.27
Chinese character handwriting tasks
Accuracy	63.22	63.00	64.44	55.78	68.44
Sensitivity	18.67	26.83	59.33	25.33	20.67
Specificity	90.98	84.88	74.45	76.95	92.40
Positive predictive value	31.67	34.17	53.50	39.17	36.67
Negative predictive value	66.32	67.99	74.51	65.13	68.37
AUC	0.55	0.56	0.67	0.51	0.57
Brier	0.37	0.37	0.36	0.42	0.32
Gait and graphics handwriting tasks
Accuracy	69.67	69.44	71.67	74.44	73.67
Sensitivity	48.67	57.67	46.67	59.00	68.50
Specificity	87.78	76.90	90.10	88.21	83.17
Positive predictive value	66.67	57.50	55.83	77.50	68.33
Negative predictive value	72.66	72.74	72.63	75.73	80.29
AUC	0.68	0.67	0.68	0.74	0.76
Brier	0.30	0.31	0.28	0.26	0.26
Gait and Chinese character handwriting tasks
Accuracy	62.44	68.44	67.44	60.33	69.78
Sensitivity	42.50	44.33	37.00	39.50	41.50
specificity	80.66	85.71	87.52	76.39	85.04
Positive predictive value	47.50	56.67	58.17	44.83	57.50
Negative predictive value	69.86	71.77	71.81	67.96	71.92
AUC	0.62	0.65	0.62	0.58	0.63
Brier	0.38	0.32	0.33	0.40	0.30

AUC: Area under the receiver operating characteristic curve; LR: Logistic regression; XGB: EXtreme Gradient Boosting; RF: Random forest; GBC: Gradient boosting classifier; Gaussian NB: Gaussian Naive Bayes.

Unimodal analysis of kinematic parameters in graphic handwriting tasks: All graphic handwriting kinematic parameters were incorporated into the unimodal classification model. For the graphics handwriting tasks, the optimal handwriting parameter subsets differed by model: The LR model selected seven items (time in air, total X-axis stroke length, maximum velocity in the X and Y-axis, minimum velocity in the X and Y-axis, and average acceleration in the Y-axis); the XGB model selected four items (X-axis stroke length, total single stroke length, minimum velocity in the X-axis, and average acceleration in the Y-axis), the RF model identified six items (average pressure, total single stroke length, minimum velocity in the X-axis, average acceleration in both X and Y-axis, and maximum acceleration in the X-axis), the GBC model included 13 items (tilt angle, time in air, single stroke time, single stroke length, X-axis stroke length, total single stroke length, total X-axis stroke length, average velocity, average velocity in the Y-axis, minimum velocity in the X-axis, number of times velocity slows down, average acceleration in the Y-axis, and minimum acceleration in the X-axis); and the Gaussian NB model incorporated five items (time in air, Y-axis stroke length, total Y-axis stroke length, average velocity, and average acceleration in the Y-axis). The classification results showed that the Gaussian NB model had the highest classification performance (accuracy = 73.44%, AUC = 0.75) (Table 5, Figure 3B). DCA indicated that the Gaussian NB model provided greater net benefit at threshold probabilities of 0.2 to 0.7 in the testing cohort. The DCA was shown in Supplementary Figure 1.

Unimodal analysis of kinematic parameters in Chinese character handwriting tasks: All Chinese character handwriting kinematic parameters were incorporated into the model. In the Chinese character handwriting tasks, the optimal handwriting parameters subsets varied across models, with 14 items identified for LR (average pressure, time in air, single stroke time, single stroke length, X and Y-axis stroke length, average velocity, average velocity in the X-axis, minimum velocity in the X-axis, number of times velocity slows down, average acceleration in the x and Y-axis, and maximum acceleration in the x and Y-axis), 10 for XGB (tilt angle, time in air, X and Y-axis stroke length, total Y-axis stroke length, average velocity, minimum velocity in the X-axis, average acceleration in the X-axis, and minimum acceleration in the X and Y-axis), two for RF (minimum velocity in the Y-axis and minimum acceleration in the Y-axis), three for GBC (Y-axis stroke length, minimum velocity in the Y-axis, and minimum acceleration in the Y-axis), and five for Gaussian NB (single stroke length, X-axis stroke length, total Y-axis stroke length, minimum velocity in the X-axis, and minimum acceleration in the Y-axis). The Gaussian NB model achieved the highest classification performance (accuracy = 68.44%, AUC = 0.57) (Table 5, Figure 3C). DCA showed that the GBC and Gaussian NB models provided a greater net benefit at threshold probabilities of 0.3 to 0.5 in the testing cohort. The DCA was presented in Supplementary Figure 1.

Multimodal analysis of kinematic parameters in gait and graphic handwriting tasks: In gait and graphic handwriting tasks, the optimal multimodal parameters subsets identified varied by model, with eight items selected for LR (step time, step length, stride length, stance time, single stroke time, total single stroke length, total X-axis stroke length, and maximum velocity in the X-axis), nine for XGB (velocity, stance time, stride length variability, double support time variability, average pressure, total X-axis stroke length, average velocity in the X and Y-axis, and maximum velocity in the X-axis), eight for RF (ambulation time, velocity, stride length variability, average pressure, total X-axis stroke length, average velocity in the Y-axis, and maximum velocity in the X and Y-axis), six for GBC (velocity, cycle time, double support time, total X-axis stroke length, and maximum velocity in the x and Y-axis), and 14 for Gaussian NB (velocity, cadence, step time, step length, stride length, HH base support, double support time, time in air, single stroke time, total X and Y-axis stroke length, average velocity in the Y-axis, maximum velocity in the X-axis, and minimum velocity in the Y-axis). The GBC model attained the highest classification performance (accuracy = 74.44%, AUC = 0.74) (Table 5, Figure 3D). Additionally, the gait and graphics handwriting tasks also had the lowest brier score among all classification tasks at 0.26. DCA showed that the Gaussian NB model yielded a greater net benefit at threshold probabilities between 0.2 and 0.7 in the testing cohort. The DCA was provided in Supplementary Figure 1. In addition, a MCI screener was developed based on gait and handwriting parameters (Supplementary Figure 2).

Multimodal analysis of kinematic parameters in gait and Chinese character handwriting tasks: In gait and Chinese character handwriting tasks, the multimodal parameters subsets included 17 items for LR (velocity, step length, step length, cycle time, stride length, HH base support, stance time, double support time, swing time variability, double support time variability, average pressure, time in air, total Y-axis stroke length, average velocity in the Y-axis, maximum velocity in the X and Y-axis, and minimum velocity in the X-axis), four for XGB (cadence, step length, single stroke time, and number of times velocity slows down), three for RF (cycle time, stance time, and average velocity), four for GBC (cycle time, stride length variability, total Y-axis stroke length, and maximum velocity in the Y-axis), four for Gaussian NB (ambulation time, velocity, swing time, and number of times velocity slows down). The classification results showed that the Gaussian NB model achieved the highest classification performance (accuracy = 69.78%, AUC = 0.63) (Table 5, Figure 3E). DCA showed that the RF and Gaussian NB models offered greater net benefit in the testing cohort at threshold probabilities of 0.2 to 0.6. The DCA was presented in Supplementary Figure 1.

DISCUSSION

This study is the first to introduce a novel screening tool for MCI by integrating gait and handwriting kinematic parameter analyses. Eight gait parameters, eight graphics parameters and two Chinese character parameters were identified as significantly differentiating older adults with MCI and from cognitively normal controls. Utilizing machine learning classification models, the combined gait and graphic handwriting tasks achieved a 74.44% accuracy, surpassing single tasks and the combination of gait and Chinese character handwriting tasks, underscoring the superior efficacy of the integrated approach.

Notably, the gait measurement equipment in this study was non-invasive, requiring no sensors to be worn by participants, thereby ensuring that the data collected were both objective and unbiased. Moreover, this study expanded upon previously identified gait parameter by incorporating a more extensive analysis of 14 parameters, including cadence, HH base support, and double support time variability, providing a more comprehensive evaluation of gait data. The findings of this study indicated that older adults with MCI exhibit slower gait speeds compared to cognitively normal older adults, aligning with the majority of prior studies[35,36]. Given that gait kinematic parameters are linked to cognitive functions in specific brain regions, slower gait speed may be indicative of declines in executive function, processing speed, and visuospatial abilities[36,37]. Furthermore, the association between gait slowing and cognitive impairment is supported by a shared neural substrate, including a smaller right hippocampus, with cognitive decline being linked to hippocampal volume reduction[38]. Additionally, this study found increased double support time variability in older adults with MCI, consistent with the findings of Jayakody et al[37]. Increased double support time variability may be related to declining memory functions. Moreover, older adults with MCI demonstrate higher cadence, shorter stride and step lengths, and longer cycle, standing and double support times, which may indicate decreased balance[39]. Decreased balance may be attributed to diminished visual function. Previous studies have shown reduced gray matter volume in the right periaqueductal area, indicating the periaqueductal gyrus is vital for the visual network[9]. While cognitively normal older adults maintain balance through visual feedback, cognitive decline in MCI may impair this function[40]. Additionally, balance decline may be associated with hippocampal dysfunction, given the hippocampus’s role in sensorimotor integration during spatial navigation[41,42]. Although stride time variability did not significantly differ between the two groups in this study, older adults with MCI still exhibited higher stride time variability than cognitively normal controls. This could be due to stride time variability being more sensitive under dual-task gait assessments[43]. Future studies should consider dual-task gait kinematics measurements to capture more distinguishing indicators.

The dot-matrix digital pen and paper used in this study preserved the advantages of traditional pen-and-paper writing, providing a more natural environment for participants while also incorporating the advantages of electronic recording, enabling real-time capture of spatial and temporal data during the writing process[16,31]. Our study observed that older adults with MCI had prolonged in-air time during writing, indicating a need for additional thinking time, which may be linked to declines in cognitive function and visuospatial ability[17,18]. Additionally, the slowing of writing speed, especially in the X- and Y-axes, showed varying degrees of reduced speed and acceleration, consistent with findings from previous studies[19,44]. Fine-motor studies suggest that horizontal and vertical movements are primarily executed through wrist and finger movements, making X- and Y-axis stroke speeds may be more reflective of fine-motor impairment[17,19]. Although Kawa et al[44] reported that older adults with MCI tended to write in larger font sizes, both the present study and Yu et al[17] found no significant difference in writing length between the groups, potentially due to differences in the writing tasks employed. In fact, conclusions regarding changes in writing pressure among older adults with MCI remain inconsistent. While Zhào et al[19] found higher writing pressure in MCI, Yu et al[17] observed no significant changes, and the present study found that older adults with MCI exhibited lower writing pressure, consistent with the findings of Werner et al[18]. This variation thought to be a manifestation of fine-motor control deficits[17-19,45]. A possible inference is that writing pressure may be related to cognitive function, with groups with poorer cognitive functioning having lower writing pressure[18]. Furthermore, the present study assessed writing correctness and found that older adults with MCI had significantly lower writing scores than cognitively normal older adults, supporting the conclusion of cognitive decline in this population.

In the unimodal categorization results, the graphic handwriting task demonstrated highest accuracy. This finding is likely because graphic handwriting tasks, particularly replicating pentagons and squares, are more challenging and thus more sensitive to cognitive decline[27]. In contrast, the signature and Chinese character copying tasks were relatively less demanding on cognitive function[16,29,46]. In addition, text writing, being more frequently used in daily life with a higher degree of automation and lower cognitive demands compared to graphic replication, may have lower discriminative power in identifying MCI among older adults with some education[16,46]. Combining gait and handwriting kinematic parameters in this study, the accuracy of the multimodal classification results was improved to 74.4%. MCI-related impairments in multiple cognitive domains often manifest as subtle changes in motor performance. Gait abnormalities may be an early marker of cognitive decline, as gait performance is often linked to executive function, attention, and working memory[9,11,36,37]. Similarly, handwriting changes also provide additional insights into severity of cognitive impairment, involving fine-motor skills, and visuospatial processing[18,45]. Combining assessments from gait and handwriting tasks allows for a comprehensive evaluation of motor-cognitive interactions, which is crucial for capturing a wider spectrum of cognitive impairments and may potentially improve classification accuracy. This finding is also supported by Mc Ardle et al[47], who demonstrated that combining multiple cognitive markers improved MCI identification compared to single-task markers.

This study introduces a novel approach to MCI-assisted screening by integrating gait and handwriting kinematic parameters, offering important clinical implications. The objective, quantifiable nature of kinematics analysis enhances screening accuracy over traditional cognitive tests, enabling a standardized evaluation of cognitive function that minimizes the influence of educational or linguistic factors[48]. This combined approach may provide a sensitive and comprehensive method for early detection by engaging multiple cognitive domains, thereby improving the precision in identifying individuals at risk of cognitive decline. While the present analysis focused on individuals with confirmed MCI and cognitively normal controls, this approach may also hold promise for identifying elevated MCI risk in broader populations without prior cognitive screening. Validation of this method in diverse settings is essential to confirm its predictive utility and to support its integration into clinical and public health protocols aimed at early intervention. Additionally, portable technologies such as cranio- corpo- graphy (CCG) have been demonstrated to sensitively detect subclinical gait disturbances via balance assessments[49]. The present study utilized the GAITRite® system for comprehensive gait analysis, CCG may offer complementary insights, particularly in settings requiring cost-effective and portable solutions. Integrating these technologies in future study could enhance screening sensitivity and broaden applicability across diverse settings.

This study has several limitations that must be considered: (1) The study’s focus on Huzhou with a relatively small sample size limits the generalizability of the findings. Moreover, potential marginal effects due to differences in age distribution between groups may have impacted the findings. Future research with larger, more diverse populations is warranted to confirm these results; (2) This study only included older adults with MCI without distinguishing specific subtypes (e.g., amnesic MCI and non-memory MCI). Future research should differentiate these subtypes to better understand how they uniquely affect gait and handwriting kinematics, which could enhance the precision and effectiveness of community-based screening efforts; (3) The inclusion of only right-hand dominant participants limits the generalizability of the findings to individuals with left or mixed-hand dominance. Future studies should include participants with varying hand dominances to improve both generalizability and clinical relevance; (4) Gait kinematic parameters was assessed using a single-task paradigm in the present study; however, incorporating dual-task gait measurements in future research may provide a more comprehensive understanding of motor and cognitive dysfunction in older adults with MCI; (5) Although the multimodal approach achieved 74.4% classification accuracy, the absence of explicit memory-related assessments, integral to MCI diagnosis, may have constrained the sensitivity. In addition, the analysis in this study relied solely on behavioral data, which may limit its ability to fully capture the complexity of cognitive impairment. Future studies should consider incorporating memory tasks as well as methods such as neuroimaging to enhance predictive performance and explore mechanisms underlying the relationship between cognitive function and behavioral changes; and (6) This study identified key gait and handwriting kinematic parameters associated with MCI, but definitive threshold scores for individual parameters were not determined due to the primary focus on evaluating overall discriminative performance. Establishing such thresholds will require further validation with larger, more diverse samples and inclusion of populations without prior cognitive screening in the future.

CONCLUSION

This study introduces a promising approach for effectively distinguishing older adults with MCI by combining gait and handwriting kinematic parameters, demonstrating potential for early detection. The integration of motor-cognitive assessments may improve classification accuracy and provide a practical, scalable tool for community-based MCI screening. Further validation in diverse populations is needed to enhance early intervention strategies for MCI.

ACKNOWLEDGEMENTS

The authors would like to thank all participants in this research.