Comparative Validation of a 3D Synthetic Data Approach for At-risk Patient Monitoring with Computer Vision: An Analysis of Frameworks and Strategies

Training Computer Vision (CV) models for patient monitoring in clinical settings, such as rehabilitation beds, is often hindered by the impracticality of collecting real-world image data for at-risk scenarios. This study proposes and validates a methodology based on generating a synthetic dataset using 3D modeling. A virtual environment replicating a Reconfigurable Assistive Technology Platform (RATP) was created in Blender software, and animated human models were used to generate 960 images depicting 'safe' (sleeping) and 'at-risk' (fallen) states. A series of eight experiments was carried out to systematically evaluate eight distinct model architectures (MAs), comparing two development frameworks (TensorFlow/Keras and PyTorch), two training strategies (Transfer Learning with MobileNetV2 and training from scratch), and the effect of an image preprocessing pipeline. The results revealed a significant performance disparity between frameworks when evaluated on a benchmark dataset. The best-performing model, a PyTorch-based Transfer Learning architecture (MA3), achieved 95.8% precision in detecting at-risk states, substantially outperforming its TensorFlow counterpart (MA1) at 58.7%. Furthermore, the effect of preprocessing proved to be highly context-dependent, improving performance in one framework while severely degrading it in the other. This work demonstrates the viability of using synthetic data for this application but concludes that model performance is critically sensitive not only to the learning strategy but also to the underlying software framework and preprocessing choices, highlighting the need for careful experimental environment configuration.