Secure TinyML for ICU: A New Approach
The goal is to keep smart medical sensors small and safe while letting them learn from data across many hospitals.
In a busy ICU, each patient’s data is unique and the devices can’t handle heavy software. The new plan uses TinyML—tiny machine‑learning models that fit on simple chips like the ESP32.
Digital Twin at the Gateway
Instead of building a full patient simulator, each device:
- Stores recent vital signs, trends, and simple risk alerts.
- Creates a lightweight “digital twin” that lives at the hospital’s gateway.
This lets each device adjust to its own patient while still sharing useful patterns with other hospitals.
Decision‑Tree Ensembles & Performance‑Based Filtering
- Models are decision‑tree ensembles: light, fast, and explainable.
- Trees can’t be averaged like neural nets.
The system merges them by testing each local model against a trusted data set and keeping only the best ones. - This “performance‑based filtering” protects against attackers who might try to poison data by mislabeling critical cases as normal.
Testing results:
- On a synthetic ICU dataset split across three hospitals, filtering kept overall accuracy high and preserved strong recall for critical cases.
Robust Security
- Every model update is encrypted with a post‑quantum key scheme (ML‑KEM‑512) and an AES‑256-GCM cipher.
- Encryption adds just 0.09 ms per update, so devices remain fast.
Superior Resilience to Attacks
Compared with other federated learning methods such as FedAvg or Krum:
- The approach shows better resilience against label‑flipping attacks.
- On a real‑world critical‑care dataset from PhysioNet, the system achieved over 98 % accuracy and near‑perfect area‑under‑curve metrics.
Conclusion
TinyML, combined with a simple digital twin and smart filtering, can make ICU monitoring safer and more reliable—even on limited hardware. The work remains a proof‑of‑concept and has not yet been tested in live clinical settings.
