Infant movement classification through pressure distribution analysis

The assessment of infant movements is crucial for the early detection of neuromotor disorders. Traditional methods, such as the Prechtl General Movements Assessment (GMA), rely on visual observation by trained experts, which is time-consuming, expensive, and susceptible to inter-rater variability. This study addresses the need for an objective, automated, and easily accessible method for assessing infant movements. The research question is whether pressure data from a non-intrusive pressure sensing mat can accurately classify different types of infant general movements, specifically distinguishing between fidgety movements (characteristic of typical development during a specific age range) and writhing movements (observed earlier in development). This has important implications for early diagnosis and intervention for conditions such as cerebral palsy, given the established link between general movement quality and neurological outcomes. The study aims to provide a proof-of-concept evaluation of this new method, comparing its performance with existing methods. The importance of this study lies in its potential to improve the accessibility and efficiency of infant neuromotor assessment on a larger scale, facilitating early detection and intervention for potential developmental disorders.

Existing literature highlights the importance of general movements in early infant development as a predictor of neurological outcomes. The GMA, a gold standard method, relies on visual gestalt perception by highly trained assessors, limiting scalability and objectivity. Several studies have explored automated alternatives, predominantly focusing on vision-based approaches using video analysis and computer vision techniques. These methods have shown promising results but often require complex setups, meticulous calibration, and careful data annotation. Other approaches use sensor-based methods such as IMUs, which can be cumbersome and potentially disruptive to infant behavior. While pressure sensing mats have been utilized in various applications, including infant sleep and posture assessment, their application in classifying infant general movements has been limited, with only a few studies hinting at its potential. This study leverages the advantages of pressure sensing, such as non-invasiveness and ease of use, to potentially overcome the limitations of existing methods. The existing literature provides a foundation for this study, demonstrating both the clinical need for improved assessment methods and the potential benefits of pressure sensing technology.

This prospective study utilized data from a previously established, validated dataset of 51 typically developing infants (26 female) aged 4-16 weeks post-term. Five infants were excluded due to incomplete recordings, resulting in a final sample of 45 infants (23 female). Data were collected across seven biweekly sessions (T1-T7), employing a multimodal approach that included RGB and RGB-D video, IMUs, and a Tekscan Conformat pressure sensing mat with 1024 sensors. The pressure mat recorded pressure changes at 100Hz, creating 32x32 pressure images. The study focused on data from T1 (4 weeks) representing the pre-fidgety period (writhing movements) and T5-T7 (12-16 weeks) representing the fidgety period (fidgety movements). 5-second video snippets (previously annotated in a prior study) were used; 1400 from T1 and 1400 from T5-T7. Two expert assessors independently annotated these snippets as FM+ (fidgety movements present), FM- (fidgety movements absent), or "not assessable". Snippets with discrepant labels or marked as "not assessable" were excluded, resulting in 1776 snippets (948 FM+, 828 FM-) with corresponding pressure mat data. Feature extraction involved cropping the pressure grid, splitting it into top and bottom areas, computing center of pressure (CoP) coordinates (x,y) and average pressure (p) for each area, applying a moving average filter, and normalizing values. Four machine learning models were compared: Support Vector Machines (SVM), Feed-Forward Neural Networks (FFN), Convolutional Neural Networks (CNN), and Long Short-Term Memory networks (LSTM). For SVM and FFN, manually defined statistical features (mean and standard deviation) were used; for CNN and LSTM, raw signals served as input. A 5-fold cross-validation was implemented to evaluate model performance. Balanced accuracy, sensitivity, and specificity were employed as performance measures.

The CNN model achieved the highest average classification accuracy (81.4%), significantly outperforming the other models (SVM: 76.15%, FFN: 75.57%, LSTM: 69.04%). Specifically, the CNN demonstrated 86% sensitivity (true positive rate) and 76% specificity (true negative rate) in distinguishing between FM+ and FM-. Simpler models like SVM and FFN, utilizing manually extracted features, achieved moderate accuracy (up to 76%), indicating that learned features from the CNN architecture were crucial for improved performance. Analyzing signal examples revealed that FM- showed local, low-frequency patterns with larger amplitudes, while FM+ exhibited high-frequency patterns with smaller amplitudes. The results show the potential of pressure sensing for differentiating between typical infant movement patterns at different developmental stages.

This study demonstrates the feasibility of using pressure sensing for classifying infant general movements. The high accuracy achieved by the CNN model highlights the potential for this approach as an objective and non-intrusive method for automated GMA. While the accuracy is slightly lower than some vision-based methods, the pressure sensing approach offers several advantages: non-invasiveness, ease of use, cost-effectiveness, and straightforward data sharing (pseudonymization of data is easily achieved), which are crucial for wide-scale clinical implementation. The superior performance of CNNs suggests that learning relevant features directly from the data is more effective than relying on manually defined features. Future studies should explore improved pressure mats specifically designed for infant motion capture, and investigate the classification accuracy using longer recordings. Furthermore, extending the study to include infants with atypical movement patterns will be critical for assessing the clinical utility of this method in detecting neuromotor disorders. The current study focused on typical development; future work needs to address the discrimination between typical and atypical patterns, aiming for high sensitivity and specificity in clinical settings.

This proof-of-concept study demonstrates the potential of pressure sensing mats for classifying infant general movements, achieving a high classification accuracy (81.4%) with a CNN model. This non-intrusive approach offers several advantages over existing methods, paving the way for objective and large-scale neuromotor assessment. Future research should focus on optimizing the pressure sensing technology, expanding the dataset to include infants with atypical movements, and exploring other machine learning architectures to further improve classification accuracy. The development of this technique could significantly improve early detection of neuromotor disorders, facilitating timely intervention.

The current study used a subset of data from a larger dataset, potentially underrepresenting the full range of infant movement variability. The pressure mat used was not specifically designed for infant movement tracking, limiting its sensitivity compared to full-body tracking methods. The study was limited to a typically developing cohort; future research needs to include infants with atypical movements to validate the clinical utility of this technique for detecting neuromotor disorders. The use of 5-second video snippets for annotation might have resulted in missing some subtle movement characteristics, and longer recordings might further improve classification accuracy.