VIPra-24/PredictiveCare-FM-6
0
PredictiveCare-FM
Foundation Models for Predictive Healthcare
(A Multimodal AI and Open Hardware Approach)
Developed as part of IITGN SRIP 2026.
About
This is a research prototype designed to define the workflow, architecture, and pipeline for a full-scale multimodal healthcare AI system.
The prototype accepts three types of patient inputs simultaneously:
- 🔬 Medical Images — Leukemia, Diabetic Retinopathy
- 📊 Tabular Clinical Data — Fever, Anaemia
- 🎙️ Respiratory Audio — COVID-19, Asthma
About the Current Accuracy
The current fusion model accuracy is ~62.5%. This is primarily because the sub-models were trained on heterogeneous datasets from different sources and patient populations — not from a single unified clinical cohort.
Accuracy is expected to improve significantly once the system is retrained on unified, locally collected clinical data where all modalities are recorded from the same patient visit.
Disclaimer
- This prototype is built solely to establish and validate the multimodal pipeline
- It is not intended for clinical use
- All datasets used are publicly available (Kaggle)
- This is a stepping stone toward the actual model which will be built on real clinical data
Summer Research Internship Programme 2026
Indian Institute of Technology Gandhinagar