Cross-Modal Federated TinyML for MCU-based Internet of Medical Things.

Area of Science:

• Healthcare technology
• Machine learning
• Internet of Medical Things

Background:

• Machine learning (ML) and Internet of Medical Things (IoMT) offer advancements in healthcare, particularly for remote patient monitoring.
• Challenges include data latency, privacy concerns, integrating heterogeneous data, and the limited processing power of wearable devices.
• Traditional Federated Learning (FL) struggles with the multimodal data common in healthcare scenarios.

Purpose of the Study:

• To propose a Cross-Modal Federated TinyML (TinyCFL) solution for resource-constrained microcontroller unit (MCU)-based medical devices.
• To address the limitations of traditional FL in handling multimodal data for healthcare applications.
• To enable efficient, private, and low-latency remote patient monitoring using distributed data fusion.

Main Methods:

• Developed a Cross-Modal Federated TinyML (TinyCFL) implementation for MCU-based medical devices.
• Employed an intermediate multimodal distributed data fusion approach for feature extraction and fusion.
• Utilized the UP-Fall detection dataset across various distribution scenarios (balanced, unbalanced, Participant-Wise).

Main Results:

• Demonstrated the feasibility of TinyCFL for distributed IoT-edge scenarios.
• Successfully implemented a prototype on resource-constrained MCUs.
• Showcased the ability to process and fuse data from different modalities for cross-modal reasoning.

Conclusions:

• TinyCFL effectively enables collaborative model training on distributed, resource-constrained IoMT devices.
• The proposed approach enhances data privacy, reduces latency, and improves energy efficiency in remote patient monitoring.
• This work provides a viable solution for complex, multimodal healthcare data analysis on edge devices.