Pain is a sensory warning signal generated by the nervous system, typically triggered when the body detects potential injury or danger. This vital function plays an essential role in medical diagnosis. However, not all individuals can communicate their pain—such as elderly patients with cognitive impairments or young children. In these cases, automatic pain recognition can provide timely, accurate detection, potentially reducing suffering and enhancing recovery outcomes. Neural networks have shown promise in detecting pain by analyzing indicators such as facial expressions, vocal cues, and physiological data. However, building robust models remains challenging due to limited annotated data and the inherent variability in pain expression across individuals. To address these challenges, we propose a multimodal data augmentation approach that generates synthetic training data by randomly shuffling modalities. This technique enhances dataset diversity by creating novel combinations of input data, enabling the neural network to learn from a broader spectrum of pain expressions and patient contexts. With random modality shuffling, the network gains resilience to variations in data quality and availability, allowing it to identify pain even when certain modalities are altered. This paper examines the effectiveness of modality-shuffling augmentation in improving the performance, robustness, and generalization of neural network models for automatic pain recognition.