Infant movement classification through pressure distribution analysis

The study investigates whether a non-intrusive pressure sensing mat can objectively classify infant general movement patterns, focusing on differentiating typical fidgety movements (present during 9–20 weeks post-term) from pre-fidgety (writhing) movements. Motivated by the high clinical value of Prechtl’s General Movements Assessment (GMA) and the need for scalable, objective tools, the authors target automated, privacy-preserving alternatives to human visual assessment. Existing automated approaches largely rely on vision-based systems or wearable sensors, each with practical or privacy limitations. The research tests the hypothesis that pressure distribution dynamics, capturing spatiotemporal force changes during infant motion, can discriminate age-specific movement patterns and serve as a viable basis for clinical applications and large-scale data sharing.

Prior work has established GMA as an effective early diagnostic tool for cerebral palsy and as a biomarker for later cognitive, language, and motor outcomes. Automated GMA research has focused on vision-based methods (RGB/RGB-D) and wearable sensors, but vision methods may face privacy and occlusion issues, while wearables can be intrusive and potentially alter behavior. Pressure-sensitive mats have been used for monitoring posture, balance, and infant sleep/motor behaviors, and force plates have shown associations between variability and fidgety movements and differences in center-of-pressure features between normal and abnormal FMs. However, prior pressure-based studies did not apply machine learning for GM classification. The authors position pressure mats as non-intrusive, low-setup, privacy-preserving sensors that directly capture dynamic force distribution, potentially complementing video/IMU modalities.

Design: Proof-of-concept study using a validated, expert-annotated dataset from a prospective longitudinal cohort of typically developing infants. Ethical approvals were obtained; parents provided informed consent. Participants: 51 initially recruited; final N=45 (23 females) after exclusions for medical condition (n=1) and incomplete recordings (n=5). Inclusion criteria included term birth, uneventful neonatal course, appropriate birth weight, typical sensory development, and parental background criteria. Data acquisition: Seven biweekly lab sessions from 4–16 weeks post-term (T1–T7): T1 28±2d, T2 42±2d, T3 56±2d, T4 70±2d, T5 84±2d, T6 98±2d, T7 112±2d. For GMA, the pre-fidgety period (T1) and fidgety period (T5–T7) were analyzed. Sensors included synchronized RGB/RGB-D video, IMUs, and a pressure mat. Pressure mat: Conformat (Tekscan) with 1024 sensors arranged 32×32 (471.4×471.4 mm²), 8-bit frames at 100 Hz. Annotation and snippet selection: From 19,451 total 5 s RGB snippets across ages, 2,800 were randomly sampled (1,400 from T1; 1,400 from T5–T7). Two blinded expert GMA assessors independently labeled each as FM+, FM−, or not assessable. Interrater agreement for FM+/FM− was Cohen’s κ=0.97; intra-rater κ=0.85 and κ=0.95. Excluded: discrepant FM labels (N=24) and any ‘not assessable’ by either assessor (N=990). Remaining 1,784 video snippets had identical labels: FM+ (N=956) and FM− (N=828); 1,776 had synchronized pressure data (FM+=948; FM−=828) used for ML. Feature extraction (motion encoding): For each 5 s snippet (500 frames at 100 Hz), each frame is a 32×32 pressure grid. The region [rows 1–29, cols 4–29] (29×26) is cropped to exclude zero regions. The cropped grid is split into top (12×26; shoulders/head) and bottom (17×26; hips) areas. For each frame and each area, compute center-of-pressure coordinates (x, y) and average pressure p; apply a moving average filter (window 5 frames). Normalize x, y jointly across top/bottom and p across areas to [0,1] to reduce size/weight bias. Final input per snippet: six time series (x_t, y_t, p_t, x_b, y_b, p_b) of length 500 (shape 500×6). Models:

• SVM (scikit-learn): Statistical features (mean, SD) of signals and optionally their derivatives, yielding 12 (no derivatives) or 24 (with derivatives) features. Kernels: RBF and polynomial (degrees 1–3). Hyperparameters tuned on validation set: C∈{0.1,1,10,100,1000}, gamma∈{0.01,0.1,1,10,100}.
• Feed-forward networks (FFN/MLP): Inputs of 12 or 24 statistical features; architectures with one or two fully connected layers (e.g., 100 or 200 units).
• CNN: Input is full 500×6 signals; architectures with 1–3 temporal convolution layers followed by average pooling and one or two FC layers; learned features.
• LSTM: Input 500×6; architectures with one or two LSTM layers followed by FC layers; learned features. Implementation and training: Neural networks implemented in TensorFlow/Keras; Adam optimizer, binary cross-entropy, learning rate 0.001, batch size 4; early stopping (validation split 1/6; patience 10). For each fold, networks trained 20 times; the model with lowest validation loss evaluated on test set. SVMs trained across full hyperparameter grid; best validation accuracy model evaluated on test set. Evaluation: 5-fold cross-validation with subject-wise splits: each fold uses snippets from 36 infants for training (further split 5/6 train, 1/6 validation) and 9 infants for testing. Across folds, training sets averaged 662 FM− and 758 FM+ snippets; test sets averaged 166 FM− and 190 FM+. Metrics: sensitivity (TPR), specificity (TNR), and balanced accuracy (BA). Mean, 95% CI across folds, and pairwise t-tests (p<0.05) used for comparisons. Data and code released on Zenodo.

• Dataset: 1,776 five-second pressure snippets (FM+=948; FM−=828) from 45 infants; 5-fold subject-wise CV.
• Signal characteristics: FM− showed lower-frequency, larger-amplitude local patterns; FM+ showed higher-frequency, smaller-amplitude patterns in the extracted x, y, p signals.
• SVM (with statistical features): • Without derivatives: BA=69.13–71.49%. • With derivatives: BA=73.87–76.15% (best S2.P1=76.15%); improvement vs without derivatives not significant (p=0.0776).
• FFN (with statistical features): • Best BA=75.57% (adding derivative features and increasing neurons); adding a second FC layer did not help (73.58%).
• CNN (learned features): • Single conv layer best BA=77.46%; increasing filters/sizes did not consistently improve. • Deeper CNNs yielded best performance; top model C3F2 (three conv layers, two FC) achieved BA=81.43% (95% CI [78.00%, 84.86%]).
• LSTM (learned features): BA=66.93–69.04%; best L1F2=69.04%; adding layers/neurons did not improve.
• Comparative statistics: CNN outperformed FFN (p=0.0343) and LSTM (p=0.0008); difference vs SVM not significant (p=0.0911). For the best CNN, sensitivity≈86% and specificity≈76%, balanced accuracy≈81%.
• Relative to vision-based methods on different datasets, pressure-mat performance was somewhat lower than reported RGB/RGB-D approaches (e.g., BA up to 88–100% in prior works), but demonstrates strong feasibility for non-intrusive, privacy-preserving sensing.

The findings support the hypothesis that spatiotemporal pressure distributions can differentiate typical age-specific infant general movement patterns. The best-performing CNN achieved balanced accuracy of about 81% with sensitivity 86% and specificity 76% for FM+ vs FM− classification, indicating that learned temporal convolutional features from center-of-pressure and pressure signals capture salient differences between fidgety and pre-fidgety movements. This addresses the need for objective, scalable, and non-intrusive assessment methods aligned with GMA principles. Compared to vision-based systems, pressure sensing offers advantages in privacy (no identifiable imagery), ease of setup, and direct motion-force information, making it promising for routine clinical screening and multicenter data sharing. Lower performance relative to some video-based methods likely reflects device sensitivity (mat not tailored for infants), the inability to capture limb movements off the mat, the use of relatively short 5 s snippets and a subset of the available data, and the focus on typical developmental patterns (FM− here represents normal writhing, not pathological absence). These insights suggest that improved hardware, longer recordings, larger datasets, and optimized temporal CNN architectures could further enhance accuracy. Pressure sensing can also complement video/IMU modalities in multimodal frameworks to improve robustness, reduce occlusions, and facilitate broader clinical implementation of AI-driven GMA.

This proof-of-concept demonstrates that a non-intrusive pressure sensing mat combined with machine learning—especially CNNs—can classify typical infant fidgety versus pre-fidgety movements with balanced accuracy around 81%. The approach is practical, privacy-preserving, and conducive to large-scale data acquisition and sharing, offering a compelling foundation for clinical translation. Future work should: (1) employ infant-optimized pressure hardware, (2) expand datasets (including atypical/pathological cases) and recording durations, (3) explore advanced temporal CNNs and multimodal fusion with video/IMU data, and (4) build shared, high-quality annotated repositories to benchmark methods and accelerate progress toward scalable, objective GMA tools.

• Hardware not specifically designed for infant motion tracking, potentially limiting sensitivity.
• Pressure mats cannot directly capture movements of limbs lifted off the surface; motion is inferred indirectly via force redistribution.
• Only a subset of available data was annotated and used; short 5 s snippets may limit model context.
• Dataset comprised typically developing infants; FM− corresponds to normal writhing, not pathological absence, limiting conclusions on clinical detection of atypical patterns.
• Cross-study comparisons are indirect due to differing datasets and annotation protocols.
• LSTM architectures underperformed; alternative temporal models were not exhaustively explored within this study.