Machine learning reveals spatial orientation, BMI and balance as leading drivers of gross motor coordination in 9–10-year-olds

Table of Contents

1. Key Highlights
2. Introduction
3. Why gross motor coordination at 9–10 years matters
4. How the study measured coordination, fitness and cognition
5. Why the researchers used machine learning (and SHAP) instead of only linear models
6. Machine learning outperforms linear regression — but not by an extreme margin
7. The top predictors: spatial orientation, BMI, balance and lower-limb power
8. Nonlinear relationships: what SHAP dependence plots reveal
9. Interpreting the weaker role of executive function
10. Translating findings into school, sport and clinical practice
11. Screening workflow for schools and practitioners
12. Limitations of the evidence and where future research should go
13. Why interpretability matters: SHAP bridges prediction and action
14. Practical recommendations for teachers, coaches and clinicians
15. Policy and curriculum implications
16. Closing reflections
17. FAQ

Key Highlights

• A random forest model outperformed traditional linear regression in predicting gross motor coordination (GMC) among 167 children aged 9–10, with spatial orientation, BMI and balance emerging as the strongest contributors.
• SHAP-based interpretation uncovered nonlinear relationships for spatial orientation, eyes-closed static balance and BMI, highlighting thresholds and interaction effects that linear models miss.
• Practical targets for early intervention include spatial-orientation training, rhythm-based drills, balance work (especially with reduced visual input), lower-limb strength development and healthy weight management.

Introduction

Gross motor coordination (GMC) underpins the everyday movements children use to run, jump, balance and navigate space. For 6–12 year olds, the period when foundational movement skills consolidate, GMC influences not only physical activity participation but also social and academic functioning. Persistent coordination deficits in childhood often translate into reduced physical activity and poorer health trajectories over time.

Most prior work has relied on linear statistics to identify correlates of coordination. Those methods assume simple additive effects and can miss threshold effects, interactions and other nonlinear patterns that characterize neuromotor development. A team of researchers working with 167 children from three public primary schools in Shanghai combined detailed motor testing with machine learning and interpretable-modeling techniques to map which attributes matter most to GMC at ages 9–10. Their approach moved beyond correlation tables to build predictive models, compare algorithmic performance, and use SHAP values to explain how individual predictors influence GMC across their observed ranges.

The results point to actionable, evidence-based targets for school programs, coaches and clinicians: spatial orientation and balance—especially when vision is removed—along with lower-limb power and body composition are central to coordination competence. This article examines the study’s methods, key findings and implications, and explains how educators and practitioners can translate those findings into screening and intervention strategies.

Why gross motor coordination at 9–10 years matters

Gross motor coordination—assessed here with the Körperkoordinationstest für Kinder (KTK)—is not a narrow athletic metric. It reflects the efficiency of interactions among musculoskeletal, sensory and central nervous systems. Children who score poorly on coordination tests are more likely to avoid sports, struggle with self-care tasks, and experience social friction with peers. Public-health data show that between roughly 9% and 28% of 10–12-year-olds present motor coordination difficulties or disorders, a prevalence large enough to justify systematic screening in schools.

The window between roughly 6 and 12 years is critical. Motor patterns formed during this time scaffold later physical literacy and long-term activity habits. Improving GMC before adolescence therefore has potential downstream benefits for fitness, mental health and academic engagement. The Shanghai study focused on 9–10-year-olds, a cohort situated in the middle of this sensitive period and therefore suitable for both screening and targeted training.

How the study measured coordination, fitness and cognition

The study recruited 167 children (85 boys, 82 girls) and administered a battery spanning four domains: gross motor coordination, physical fitness, basic coordination capacities (BCC), and executive function.

• Gross motor coordination: Assessed with the KTK battery. The KTK includes four subtests—walking backwards on beams of decreasing width, jumping sideways for speed, hopping for height (single-leg), and moving sideways between platforms. Raw scores were converted to an age- and sex-adjusted Motor Quotient (MQ), the primary outcome.
• Physical fitness: Items adapted from the Youth Fitness International Test (YFIT) and evidence-based additions: 50-m sprint (speed/anaerobic power), standing long jump (lower-limb explosive strength), 1-minute sit-ups (muscular endurance), sit-and-reach (flexibility), grip strength (upper-body strength proxy), and 1-minute rope-skipping (coordination and aerobic capacity).
• Basic coordination capacities: Six tasks measured kinesthetic differentiation, spatial orientation, dynamic balance (Y Balance Test), static balance with eyes open and eyes closed, rhythm ability (rhythmic sprint vs best 30-m sprint), and simple motor reaction time.
• Executive function (EF): Reported by parents using the Behavior Rating Inventory of Executive Function (BRIEF), which yields indices such as Behavioral Regulation, Metacognition and a Global Executive Composite.

Test procedures were standardized, with warm-ups, familiarization trials and spacing of assessments to reduce fatigue. The KTK Motor Quotient provided a single, comparable measure of GMC for model training.

Why the researchers used machine learning (and SHAP) instead of only linear models

Traditional regression models are valuable but limited when predictors interact, follow nonlinear patterns or possess complex distributions. GMC development depends on biomechanics, sensory integration and cognitive control that do not add up neatly. Machine learning (ML) algorithms can capture interactions and nonlinearities without pre-specifying interaction terms.

The team compared four ML models with ordinary least squares multiple linear regression:

• Random Forest Regression (RFR)
• Extreme Gradient Boosting (XGBoost)
• Light Gradient Boosting Machine (LGBM)
• Support Vector Regression (SVR)

To make the best-performing ML model interpretable, the researchers used SHapley Additive exPlanations (SHAP). SHAP assigns each feature a contribution to the predicted value for individual cases and summarizes global importance. This approach gives two advantages: higher predictive accuracy from flexible models, and human-readable explanations of which variables drive predictions and how their effects vary across values.

Model inputs were preprocessed with checks for multicollinearity, standardization, one-hot encoding for gender, and recursive feature elimination (RFE) to retain the 15 most informative variables. The data were split 80/20 for training vs testing, and hyperparameters were tuned using randomized search with 5-fold cross-validation.

Machine learning outperforms linear regression — but not by an extreme margin

Model performance was compared using R^2, mean absolute error (MAE), and root mean squared error (RMSE). Key results:

• Multiple linear regression: R^2 = 0.558 (unadjusted); adjusted R^2 = 0.465; MAE = 4.782; RMSE = 6.128.
• Random Forest Regression (best ML model): R^2 = 0.533; MAE = 4.850; RMSE = 6.075.
• XGBoost: R^2 = 0.430; MAE = 5.233; RMSE = 6.710.
• LGBM: R^2 = 0.432; MAE = 5.525; RMSE = 6.695.
• SVR: R^2 = 0.489; MAE = 4.682; RMSE = 6.360.

The RFR model offered improved explanatory power over most ML competitors and slightly better prediction error than linear regression on RMSE. Cross-validation of the RFR showed some variance (R^2 = 0.353 ± 0.155) but generally good generalization. The relative closeness of results emphasizes that while linear regression detects several important relationships, nonparametric models reveal additional patterns—especially nonlinearities and interactions—that matter for nuanced understanding and individual-level prediction.

The top predictors: spatial orientation, BMI, balance and lower-limb power

SHAP analysis of the Random Forest identified the top contributors to GMC. The ten most important features, in descending order, were:

1. Spatial orientation ability
2. Body mass index (BMI)
3. Dynamic balance (Y Balance composite)
4. Standing long jump (lower-limb explosive strength)
5. Eyes-closed static balance
6. Rope skipping (coordination/endurance)
7. Sit-and-reach (flexibility)
8. Eyes-open static balance
9. Rhythm ability (difference between rhythmic sprint and best 30-m sprint)
10. 50-m sprint time (speed)

Spatial orientation and BMI had substantially higher importance scores than other predictors, emphasizing both perceptual-motor processing and body composition as central determinants of coordination at this age.

Detailed takeaways:

• Spatial orientation: Children better at navigating numbered routes and responding to spatial cues had higher GMC. This measure captures the capacity to relate the body to external space and adjust locomotor patterns accordingly.
• BMI: Not a simple linear predictor. Both under- and overweight patterns can affect movement efficiency; excess adiposity adds mechanical load and alters movement economy, while very low body mass can reflect insufficient muscle mass to stabilize complex movements.
• Balance: Both dynamic and static balance—particularly eyes-closed static balance—featured prominently. Eyes-closed balance depends more heavily on proprioceptive and vestibular systems; its importance suggests coordination depends critically on nonvisual sensory integration.
• Lower-limb power: Standing long jump stood out as a proxy for muscular power and propulsion, attributes needed for efficient hopping, jumping and rapid lateral transfers.

Executive function measures from the BRIEF did not rank highly as direct predictors. The researchers note measurement limitations—parent-report questionnaires may not capture online, movement-related cognitive control—and suggest objective EF tasks for future work.

Nonlinear relationships: what SHAP dependence plots reveal

SHAP produces dependence plots that show how the contribution of a predictor varies across its observed range. The Shanghai study’s dependence plots for the top five features revealed distinct patterns.

Spatial orientation: The association with GMC is nonlinear. A certain threshold of spatial orientation ability yielded large positive SHAP values—above which GMC increased markedly. Below that threshold small changes had little effect. That suggests interventions that move a child past a minimum spatial-orientation competency could yield outsized gains in coordination.

BMI: Exhibited a U-shaped or otherwise nonlinear pattern consistent with prior literature. Moderate BMI values associated with optimal GMC, while both low and high extremes predicted lower MQ. This pattern indicates interventions should aim for healthy, not merely lower, body mass and supports paired strategies combining skill work with nutrition and activity that promote appropriate growth.

Eyes-closed static balance: Demonstrated a nonlinear effect as well—improvements provided large GMC gains up to a point, after which returns diminished. This pattern aligns with a ceiling effect for balance tasks: once basic proprioceptive control is established, additional static-balance gains transfer less into composite coordination metrics than earlier gains do.

Dynamic balance and standing long jump showed more linear positive relationships across the observed ranges: steady improvements in these areas consistently related to higher GMC.

Understanding these shapes matters for program design. For example, if small gains in spatial orientation can produce large GMC benefits for low-performing children, targeted navigation and perceptual training may be more efficient than long-term strength work for some students.

Interpreting the weaker role of executive function

Executive function (EF) is theoretically relevant to motor control: planning, inhibition and working memory contribute to sequencing movements, adapting actions under changing conditions, and resisting distractors. In this study, however, EF—as measured by the BRIEF parent-report—had minimal predictive value.

Three points explain the finding:

1. Assessment modality: BRIEF captures caregiver perceptions across everyday contexts. It is not a lab-based behavioral task that measures cognitive control during movement. Behavior ratings and task-based measures often diverge.
2. Task specificity: EF contributions are more apparent when tasks are novel, cognitively demanding, or require rapid adjustments. The KTK and fitness battery assess motor competence but do not systematically tax complex cognitive control in the same way as dual-task gait or decision-heavy sports drills.
3. Developmental expression: At 9–10 years, cognition and motor systems are still maturing; some EF–motor relationships may emerge or strengthen later, or require more sensitive measurement to detect.

The authors recommend future studies pair behavioral EF assessments (e.g., n-back, go/no-go) or neurophysiological measures (fNIRS, EEG) with motor testing to better capture how cognitive control supports coordination.

Translating findings into school, sport and clinical practice

The study yields clear, actionable directions for interventions and screening at the school or community level. Four practical strategies follow from the predictors identified.

1. Screen for GMC and spatial orientation early and regularly
   • Use an easy-to-administer screening battery like select KTK items or abbreviated coordination screens to flag children below normative Motor Quotient thresholds.
   • Follow up low scorers with brief spatial-orientation tasks (e.g., route-following or numbered-cone sprints) and simple balance assessments (eyes-open and eyes-closed single-leg stance) to guide targeted programming.
2. Integrate spatial-orientation training into PE and recess
   • Maze navigation, small orienteering games and directional cueing drills train allocentric and egocentric spatial maps while keeping sessions game-like and engaging.
   • Example: a 10-minute “treasure run” where children follow sequential cues through cones or numbered stations, integrating speed and spatial decision-making.
3. Emphasize balance training that challenges nonvisual systems
   • Progress from eyes-open to eyes-closed balance holds, single-leg stances on soft surfaces, and dynamic reach tasks (Y Balance variants).
   • Incorporate balance into warm-ups with 2–3 short exercises, 3× per week; evidence shows early proprioceptive gains transfer into improved coordination.
4. Build lower-limb power and rhythmic coordination
   • Standing long jump and rhythm ability both contribute to GMC. Plyometric play (hop-and-stick, partner jumping games), jump-to-target drills and rhythmic activities like rope skipping and music-based locomotor sequences develop power and timing.
   • Practical program: 2 weekly sessions of plyometric-focused play (10–15 minutes) plus daily short rope-skipping breaks integrated into classroom transitions.
5. Address BMI through integrated physical activity and healthy eating
   • The nonlinear association underscores that extremes of body composition can impair coordination. School-based physical-activity promotion (active recess, skill-oriented PE), combined with nutrition education, supports healthy growth.
   • Interventions that increase active play and reduce sedentary time improve movement economy and can indirectly boost coordination.
6. Use individualized, ability-graded progressions
   • SHAP plots show thresholds and diminishing returns. Design progressions that aim first to move children past lower competency thresholds (e.g., basic spatial-orientation competence, minimal proprioceptive control) before expecting gains from higher-level strength or speed work.

Real-world examples and programs:

• The Daily Mile and similar short daily runs increase overall activity but can be augmented with navigation or skipping stations to address spatial and rhythm skills.
• Orienteering activities used by Scout groups or clubs are inexpensive ways to develop spatial judgment and dynamic decision-making under movement.
• School-based rhythm programs, such as jump-rope curricula or rhythmic gymnastics adaptations, can be incorporated into PE with minimal equipment.

Screening workflow for schools and practitioners

A pragmatic screening and intervention workflow follows from the study:

1. Universal screen (termly): brief KTK subset + 30-second single-leg balance (eyes open and closed) + 30-second rope skipping count. Identify bottom 15–20% for follow-up.
2. Diagnostic follow-up (within 2 weeks): full KTK MQ, spatial orientation course (numbered-cone run), standing long jump, and BMI.
3. Individualized plan (6–8 weeks): if spatial orientation score low → focus on navigation and rhythm drills 2–3×/week; if balance deficient (esp. eyes-closed) → proprioceptive balance progressions; if standing long jump low → plyometric and strength play; if BMI extreme → coordinated nutrition and activity plan with family engagement.
4. Reassessment: after 8 weeks measure MQ and key subtests to track progress and adjust.

Program intensity need not be onerous. Short, frequent, enjoyable activities embedded in PE and recess (10–20 minutes per session) show better adherence and sustained gains than long, infrequent drills.

Limitations of the evidence and where future research should go

The study provides a sophisticated cross-sectional analysis but has limitations that shape interpretation and future priorities.

• Cross-sectional design: Associations cannot establish causality. Longitudinal and intervention studies are necessary to determine whether improving spatial orientation or eyes-closed balance causes gains in GMC.
• Sample and generalizability: The cohort derived from three public schools in Shanghai. Cultural, curricular and growth-pattern differences mean findings require replication in other regions and socio-economic contexts.
• Executive function measurement: Relying solely on parent-report (BRIEF) likely underestimates EF contributions. Objective in-task cognitive assessments and neuroimaging modalities (fNIRS, EEG) would clarify how cognitive control supports coordination.
• Sample size and power: Although 167 participants provided enough data for the reported models, larger samples would refine confidence intervals and enable modeling of subgroup effects (sex, maturation stage, socioeconomic status).
• Contextual and psychosocial variables: The study did not incorporate variables such as physical-activity history, screen time, family support, or access to play spaces. These ecological factors can moderate both BMI and motor development.

Future research directions:

• Longitudinal cohorts following children from early childhood through adolescence to map causal paths linking fitness, spatial skills and GMC.
• Randomized trials testing targeted interventions (spatial-orientation games, eyes-closed balance training, combined power-rhythm curriculums) with GMC as a primary endpoint.
• Integration of wearable sensors during tasks to capture movement variability and task-specific executive demands.
• Larger multi-site studies including diverse socio-economic and cultural contexts to test generalizability and inform policy-level school programming.

Why interpretability matters: SHAP bridges prediction and action

Machine learning models can excel at prediction, but without interpretable outputs they offer limited practical guidance. SHAP values turn black-box outputs into actionable insights: they show not only which predictors matter globally, but how a predictor’s contribution changes across its observed values for individual children.

For practitioners this matters: if SHAP indicates a child’s low MQ is driven primarily by poor spatial orientation, a coach can prioritize navigation and rhythm tasks rather than generalized conditioning. Conversely, if BMI and low plyometric power are the main drivers, the plan shifts toward balanced nutrition and strength/power play. This individualized focus aligns with precision public-health goals and maximizes the efficiency of school-based programs that operate under time and resource constraints.

Practical recommendations for teachers, coaches and clinicians

• Prioritize simple screening tools: a short KTK subset, single-leg eyes-closed balance, and a standing long jump replicate key predictors with minimal equipment.
• Design short, frequent activities: 10–15 minute playful drills embedded into PE or recess produce consistent exposure and skill gains.
• Combine rhythm and spatial training: rope-skipping circuits, music-based movement sequences and orienteering games simultaneously train timing, spatial mapping and lower-limb power.
• Balance progression: begin with eyes-open tasks and progress to eyes-closed and unstable-surface conditions to strengthen proprioceptive control.
• Monitor BMI as part of coordinated care: when BMI is outside healthy ranges, combine motor skill training with family-oriented nutrition and activity interventions.
• Use individualized plans guided by data: interpret student profiles (MQ, spatial tasks, jump, balance) to prioritize interventions that target the dominant deficits for each child.

A sample 8-week microprogram for a school implementing findings:

• Week structure: 3 PE sessions/week + 3 active-recess mini-sessions
• PE session (30–40 min): 10-min dynamic warm-up; 10–15 min spatial-orientation/skill circuit (numbered cones, treasure runs, small-team orienteering); 6–8 min plyometric play (hop sequences, bounding); 5 min cool-down with balance holds (eyes open → closed).
• Active-recess mini-session (5–10 min): jump-rope groups, rhythm relay races.
• Family handout: short at-home activities (maze walks, dance sequences, rope skipping) and healthy-snack suggestions.

Policy and curriculum implications

• Embed coordination screening in routine school health checks at key ages (e.g., entry to primary school, mid-primary years).
• Support PE teacher training to deliver spatial, rhythm and proprioceptive activities without specialized equipment.
• Promote cross-sector partnerships: combine school programs with community sport clubs and public-health nutrition resources to address BMI and activity access.
• Allocate modest resources for measurement tools (jump mats, cones, stopwatches) rather than high-cost equipment; effective programs rely primarily on structured activities and instructor guidance.

Closing reflections

Predictive modeling combined with interpretable machine learning clarifies which motor and physiological attributes most strongly drive gross motor coordination during a formative developmental window. Spatial orientation, body composition and multisensory balance control stand out as high-yield targets. Because several of these factors respond to short, low-cost, game-based interventions, schools and community programs can act quickly to screen and support at-risk children. Replication across settings and longitudinal trials will sharpen causal inferences and refine program designs, but present evidence already supports a pragmatic pivot: add navigational games, rhythm-based drills and proprioceptive balance progressions to standard PE, and monitor progress with brief periodic assessments.

FAQ

Q: What exactly is gross motor coordination and how is it measured here?
A: Gross motor coordination refers to the ability to perform large-muscle movements with control, precision and efficiency. In the study summarized here, it was measured with the Körperkoordinationstest für Kinder (KTK), which includes four subtests—walking backwards on progressively narrower beams, jumping sideways, hopping for height (single-leg), and moving sideways between platforms. Results convert to a Motor Quotient (MQ) adjusted for age and sex.

Q: Which factors most strongly predict GMC at age 9–10?
A: Spatial orientation emerged as the single strongest predictor, followed by BMI, dynamic balance, standing long jump (lower-limb power) and eyes-closed static balance. Rhythm ability, rope skipping, flexibility and 50-m sprint also contributed but were less influential.

Q: Why did the researchers use machine learning instead of only using regression?
A: Machine learning models capture nonlinear relationships and interactions that linear regression can miss. The researchers compared multiple algorithms and used SHAP values to make the best-performing model interpretable, revealing predictors and threshold effects that a linear model might obscure.

Q: What does it mean that BMI and spatial orientation have nonlinear effects?
A: Nonlinear effects mean the relationship is not a straight-line increase or decrease. For BMI, coordination was best at moderate values and worse at both low and high extremes (a U-shaped pattern). For spatial orientation, passing a certain competency threshold yielded disproportionately larger GMC gains. These shapes suggest targeted interventions can be particularly effective for low-performing children.

Q: Should schools screen all children for coordination problems?
A: Periodic screening is advisable, especially during the 6–12-year window. Short, low-cost screens (a subset of KTK items, single-leg balance with eyes closed, and standing long jump) can identify children who would benefit from a focused intervention.

Q: What types of interventions are most likely to help children with low GMC?
A: Game-based spatial-orientation drills (orienteering, numbered-route games), rhythm activities (rope skipping, rhythmic running), balance progressions including eyes-closed work, and lower-limb plyometric play are recommended. For children with BMI outside healthy ranges, integrate activity with nutrition and family engagement.

Q: How should executive function be measured if it’s relevant to motor skills?
A: Behavior-rating scales like the BRIEF are useful for large-scale screening but may miss in-task cognitive control. Objective behavioral tasks (e.g., go/no-go, n-back, set-shifting tasks) and neuroimaging or neurophysiological tools (fNIRS, EEG) provide more direct measures of cognitive processes engaged during motor tasks.

Q: What are the study’s main limitations?
A: The study is cross-sectional, preventing causal inference; the sample came from three schools in Shanghai, so generalizability is limited; executive function relied on parent-report; and ecological variables like activity history or socioeconomic status were not included. Future longitudinal and intervention studies across diverse settings are needed.

Q: How can practitioners use SHAP outputs in practice?
A: SHAP identifies which features drive a child’s predicted MQ and how those features' effects change with value. Practitioners can prioritize interventions based on the dominant SHAP drivers for each child—e.g., spatial training for children whose low MQ is driven primarily by spatial-orientation deficits.

Q: Where can I find the study data or code if I want to replicate the analyses?
A: The original project made data and code available in supplemental files and deposited the code in a public repository (Zenodo). Practitioners and researchers should consult the original study’s supplemental materials for exact resources and preprocessing steps.