Medical AI That Works

Bimodal Fall Detection System

A novel bimodal deep learning framework for elderly fall detection that addresses the high false-positive rates of single-modality systems. The architecture combines a memory-based LSTM autoencoder for inertial abnormality detection (waist-mounted IMU) with an attention-based Transformer network for visual pose assessment (RGB camera), unified through late fusion at the decision level. Tested with custom dataset of simulated falls and routine activities under dim lighting conditions, including complete darkness.

3D Respiratory Monitoring with Xbox Cameras

A Computerized Clinical Decision Support System (CCDSS) using dual Xbox Kinect cameras at opposite corners of the patient's bed to maximize spatial coverage. Calculates respiration rate and tidal volume from 3D thoraco-abdominal surface reconstruction.

• 3D Point Cloud Fusion: Two camera views aligned via 2D markers
• Triangular Mesh: Surface reconstruction with reference plane
• Octree Volume: Recursive decomposition for volume calculation
• Signal Analysis: Frequency → RR, amplitude → tidal volume

Validated at Sainte-Justine ICU on mannequin + real patients. Approved by hospital equipment inspection. High accuracy vs gold standard ventilator (10ml-500ml range).