ramyasp64/MediSense

Healthcare AI prototype: vital signs anomaly detection for low-resource settings via transfer learning from industrial sensors # MediSense: Vital Signs Anomaly Detection for Low-Resource Healthcare Settings **Status:** Active Research · Universität Koblenz, Germany · 2025–2026 **Domain:** Healthcare AI · Transfer Learning · Medical Signal Processing **Author:** Ramya Subramanian Porselva Bharathi --- ## Research Question Millions of patients in low-resource clinical settings (community health posts, rural hospitals, field clinics) go unmonitored because accurate vital signs monitoring requires either expensive hardware or large quantities of labelled medical training data. **MediSense challenges both constraints.** The core insight: the anomaly structure that appears in factory sensor degradation (*normal → degraded → failure*) is structurally identical to patient vital sign deterioration. A model trained to detect bearing failures learns representations that generalise to cardiac and respiratory emergencies, with remarkably little medical fine-tuning data. This research bridges **industrial quality control** and **healthcare AI** to serve populations who need it most. Through rigorous technical validation and a clear pathway to deployment, MediSense demonstrates that affordable, accurate vital signs monitoring for low-resource settings is achievable. --- ## System Architecture ``` INDUSTRIAL PRE-TRAINING MEDICAL FINE-TUNING ───────────────────── ───────────────────── NASA Bearing Vibration ─┐ BIDMC Waveform PPG ─┐ CMAPSS Turbofan Data ─┼─► Autoencoder CapnoBase PPG ─┼─► Fine-tuned Model SKAB Hydraulic Pump ─┘ (Encoder MIMIC-III Waveform ─┘ (Encoder frozen, frozen after decoder retrained) pre-training) │ ▼ ┌─────────────────────────────────┐ │ DUAL DETECTION PIPELINE │ │ │ │ Level 1: Sensor Malfunction │ │ Level 2: Patient Anomaly │ │ │ │ ↓ │ │ CLINICAL ALERT E …