[Paper Review] FedMood: Federated Learning on Mobile Health Data for Mood Detection

Xiaohang Xu, Hao Peng+4 more|arXiv (Cornell University)|Feb 6, 2021

Mental Health Research Topics31 references26 citations

TL;DR

FedMood presents a federated multi-view learning framework for mood/depression detection using mobile keystroke and accelerometer data, with late fusion architectures and experiments on IID and non-IID data distributions demonstrating performance gains and distribution sensitivity.

ABSTRACT

Depression is one of the most common mental illness problems, and the symptoms shown by patients are not consistent, making it difficult to diagnose in the process of clinical practice and pathological research. Although researchers hope that artificial intelligence can contribute to the diagnosis and treatment of depression, the traditional centralized machine learning needs to aggregate patient data, and the data privacy of patients with mental illness needs to be strictly confidential, which hinders machine learning algorithms clinical application. To solve the problem of privacy of the medical history of patients with depression, we implement federated learning to analyze and diagnose depression. First, we propose a general multi-view federated learning framework using multi-source data, which can extend any traditional machine learning model to support federated learning across different institutions or parties. Secondly, we adopt late fusion methods to solve the problem of inconsistent time series of multi-view data. Finally, we compare the federated framework with other cooperative learning frameworks in performance and discuss the related results.

Motivation & Objective

• Motivate privacy-preserving mood/depression detection using mobile data without centralizing sensitive health information.
• Extend traditional models to federated multi-view learning across institutions or parties.
• Develop and evaluate late fusion strategies to handle asynchronous multi-view time series data.
• Assess performance under IID and non-IID data distributions and compare federated vs. local training and centralized baselines.

Proposed method

• Proposes a general multi-view federated learning framework that can extend traditional models to federated settings across institutions.
• Implements late fusion to align and integrate heterogeneous time-series data from multiple views.
• Evaluates three fusion strategies: Fully Connected layer, Factorization Machine layer, and Multi-view Machine layer (DMVM/DFM/DFM).
• Adopts Google-style FedAvg federated training with local updates and weighted aggregation across clients.
• Introduces privacy considerations including discussion of SMC, differential privacy, and homomorphic encryption (contextual overview).
• Uses mobile-device data (keystroke metadata and accelerometer) collected via a custom virtual keyboard for mood prediction.

Experimental results

Research questions

• RQ1 How does federated multi-view learning perform on mobile health data for mood detection compared to local and centralized approaches?
• RQ2 What is the impact of data distribution (IID vs non-IID) and varying client data amounts on model accuracy in FedMood?
• RQ3 Do different late-fusion strategies (Fully Connected, FM, Multi-view Machine) differ in performance under federated settings?
• RQ4 How does the number of participating parties and data volume per party affect federated learning convergence and accuracy?

Key findings

• In IID settings, federated approaches (FedAvg and CIIL) achieve notable accuracy gains over local training and other baselines, with best-case around 85% observed under CIIL with DMVM in the IID experiments.
• In non-IID settings, prediction accuracy declines due to extreme distribution, with CDS often outperforming some federated variants; however, CIIL and certain fusion variants still show robust gains over purely local training.
• Increasing the number of participants with fixed per-client data can improve or stabilize accuracy for federated models, though gains depend on fusion method and data distribution.
• Increasing local data per client generally boosts local and federated performance, with FedAvg offering additional gains especially when per-client data is moderate (e.g., data in the hundreds to thousands).
• The experiments demonstrate that multi-view fusion (DMVM/DFM) can outperform a simple DNN fusion strategy in several IID scenarios, but non-IID conditions can erode these advantages.