All projects

EUCLID / RESEARCH PROJECTS

EU Project

FLAMENCO

2023 - 2023

Completed

Overview

FLAMENCO brought privacy-preserving AI to one of the most sensitive settings there is: the early detection of communication disorders in children. An existing software suite records how children respond to an animated game, together with heart-rate readings from a smartwatch, and produces a risk indicator for learning and communication difficulties such as autism and intellectual disability.

Instead of pooling these records on one server, FLAMENCO trains its models with Federated Learning: each clinician keeps their own cases (451 cases across 5 clinicians in our study) and only model updates travel to the server. Autoencoder and variational-autoencoder models learn what typical development looks like and flag cases that are hard to reconstruct as potential deficiencies. Fully Homomorphic Encryption protects the aggregation step, client-selection strategies (including our novel IntelliSampler) leave out clients with corrupted or missing data, and advanced aggregators (FedAvg, FedNova, FedAvgM, FedAdam, FedYogi and more) cope with imbalanced data.

The system communicates over MQTT, so it plugs directly into TERMINET’s IoT infrastructure, and comes with a web application for clinicians. The project also released an open dataset for federated learning research on communication-skill deficiencies (Zenodo, 2024), and its results are described in “Federated Anomaly Detection for Early-Stage Diagnosis of Autism Spectrum Disorders using Serious Game Data” (2024).