PaperlayerView exact revision on arXiv

Source-linked AI summary

Uncertainty-Aware Federated Learning for Infant Movement Analysis

Edmond S. L. Ho

arXiv:2609.31463v1cs.LGcs.AIcs.CV

TL;DR

Clinical privacy and governance constraints limit centralized training on infant-movement data. The paper introduces federated fidgety-movement classification with MC-Dropout predictive uncertainty informing client aggregation. Federated models outperform independent local models and approach centralized performance.

  • Problem

    Most automated GMA studies use centralized training, while clinical privacy, ethical, and governance requirements often restrict sharing infant movement recordings across institutions.

    Source: §I · p. 1

  • Method

    The framework uses MC-Dropout to estimate client-level predictive uncertainty and predictive entropy to modulate client contributions during federated aggregation.

    Source: §I · p. 2, B. Client - Uncertainty Estimation · p. 3, 2) Uncertainty-Aware FedAvg (UA-FedAvg) · p. 5

  • Results

    Federated approaches improved accuracy by over 7% compared with No Federation and, in their best configurations, were less than 1% different in accuracy from the centralized baseline.

    Source: 1) Aggregated Evaluation · p. 6

  • Takeaways & Limitations

    The model-agnostic UA-FedAvg strategy may extend to other federated applications where predictive uncertainty can be estimated and used to guide aggregation.

    Source: §VII · p. 6

    The dataset contains recordings from typically developing infants, so results may not generalize to higher-risk clinical populations, including infants later diagnosed with CP.

    Source: §VII · p. 6

Abstract

from arXiv · show

Infant movement analysis provides valuable biomarkers for the early identification of neurodevelopmental disorders. Recent advances in deep learning have enabled automated analysis of infant movements from video-derived skeletal representations, achieving performance comparable to expert assessment for tasks such as General Movement Assessment (GMA). However, most existing approaches rely on centralized training, requiring data from multiple institutions to be collected and stored at a single site. Such assumptions are often impractical in clinical settings due to privacy, governance, and data-sharing constraints. To address these challenges, we present, to the best of our knowledge, the first federated learning framework for automated infant movement analysis and General Movement Assessment using skeletal motion data. As a clinically relevant use case, the proposed framework is evaluated on fidgety movement classification. To quantify model confidence, Monte Carlo (MC) Dropout is employed to estimate predictive uncertainty during inference. Building upon this, we propose an Uncertainty-Aware Federated Averaging (UA-FedAvg) strategy that incorporates predictive entropy derived from MC-Dropout into the federated aggregation process, enabling client contributions to be adjusted according to their predictive uncertainty. Experiments were conducted using a cross-subject evaluation protocol under a three-client federated learning setting. Results demonstrate that federated learning substantially improves classification performance compared with independently trained local models while achieving performance approaching that of centralized training. Furthermore, UA-FedAvg and its variant incorporating validation loss generally outperform conventional FedAvg across the evaluated data-split configurations.

I. INTRODUCTION

Automated GMA commonly classifies fidgety movements, but centralized training conflicts with clinical data-sharing constraints. This work introduces federated infant movement classification that uses model-prediction uncertainty to adjust client contributions.

A. Automated General Movement Assessment

  • Persistent Cramped-Synchronized movements and absent fidgety movements are reported as strong predictors of later Cerebral Palsy, and many automated GMA studies distinguish FM+ from FM-,.Automated GMA approaches include handcrafted kinematic features and end-to-end learning from skeletal sequences using CNNs or GCNs –,, –.Source: §I · p. 1, A. Automated General Movement Assessment · p. 2

I. INTRODUCTION

  • Privacy, ethical, and governance requirements often restrict cross-institution sharing of infant movement recordings, while site-level data may also differ in distribution.Federated learning enables collaborative training without exchanging raw data, aggregating model updates instead.Source: §I · p. 1
  • Unlike entropy-based client-selection methods based on dataset statistics or label distributions, the proposed method derives uncertainty from model predictions and uses it in aggregation.The work evaluates whether predictive-uncertainty-informed aggregation can improve federated infant movement classification under heterogeneous client data distributions.Source: §I · p. 2
  • Federated learning for automated GMA and uncertainty-aware federated learning in this setting are presented as firsts, with uncertainty estimated from model predictions to modulate client contributions.The framework aims to preserve data locality while accounting for differences in client model confidence.Source: §I · p. 2

B. Federated Learning for Healthcare AI

Federated learning supports healthcare model training without direct exchange of patient data, although institutional data heterogeneity poses challenges. This study formulates binary fidgety-movement classification from skeletal motion and evaluates it on a dataset of infant recordings.

B. Federated Learning for Healthcare AI

  • Healthcare federated learning has been applied across several tasks, but non-IID institutional data may cause client drift, slower convergence, and degraded FedAvg performance.FedProx, SCAFFOLD, and adaptive methods such as FedAdam and FedYogi are among strategies proposed to address these issues.Source: B. Federated Learning for Healthcare AI · p. 2

A. Problem Formulation

  • The task is binary classification of skeletal motion sequences, mapping each sequence to a label for fidgety-movement presence or absence.Recent automated GMA commonly uses 2D skeletal pose sequences extracted from video for classification,,,.Source: A. Problem Formulation · p. 2

B. Dataset

  • The dataset contains recordings from 45 typically developing infants, with 1,683 five-second clips: 943 FM+ and 740 FM- clips.Two senior GM Assessors independently annotated each clip for fidgety-movement presence or absence.Source: B. Dataset · pp. 2–3
  • Each clip is represented by 250 frames and 60 features: coordinates and temporal velocities for 15 retained body joints.The study uses the 2D skeletal modality from a multimodal dataset; pose sequences were extracted from RGB video using ViTPose.Source: B. Dataset · p. 3

IV. FEDERATED LEARNING PIPELINE

The federated setup simulates clients with distinct subject-level datasets and trains each client's skeletal-motion classifier locally. Because conventional classification metrics do not measure prediction confidence, the framework estimates uncertainty and uses it to guide client weighting.

IV. FEDERATED LEARNING PIPELINE

  • The setup divides data by subject identity among three clients, simulating sites that cannot access one another's data.The dataset's 45 subjects motivated dividing the data into three subsets to balance client count and local training data.Source: §IV · p. 3

A. Client - Automated GMA model

  • Each client trains a CNN classifier adapted from, using that work's optimal hyperparameters, while MC-Dropout uncertainty guides client weighting during aggregation.Predictive uncertainty reflects uncertainty from model limitations and ambiguous or noisy observations; lower uncertainty indicates greater confidence.Source: A. Client - Automated GMA model · p. 3, B. Client - Uncertainty Estimation · p. 3

B. Client - Uncertainty Estimation

  • Conventional accuracy, sensitivity, and specificity do not measure prediction confidence, and deep networks can be highly confident on unseen or out-of-distribution samples.This concern is relevant when client data distributions differ across cohorts, recording conditions, and demographics.Source: B. Client - Uncertainty Estimation · p. 3

1) Predictive Uncertainty Estimation Using MC-Dropout:

MC-Dropout estimates sample-level predictive uncertainty by repeating stochastic forward passes with dropout active. At the client level, mean predictive entropy on local validation data serves as an uncertainty score, with lower entropy corresponding to larger aggregation weights.

1) Predictive Uncertainty Estimation Using MC-Dropout:

  • During inference, MC-Dropout keeps dropout layers active and performs multiple stochastic forward passes to estimate predictive uncertainty.Predictive variance across these passes measures prediction variability; greater variance indicates greater uncertainty, while lower variance suggests more consistent predictions.Source: 1) Predictive Uncertainty Estimation… · p. 3

2) Predictive Uncertainty of Client Models:

  • Predictive entropy measures uncertainty in the predictive distribution: concentrated distributions have lower entropy and indicate greater confidence, while dispersed distributions have higher entropy.Repeated stochastic forward passes provide the predictions used to estimate this distribution.Source: 2) Predictive Uncertainty of Client… · p. 4
  • Mean predictive entropy on each client's local validation dataset provides a client-level uncertainty score, with lower entropy corresponding to larger aggregation weights.Client scores can vary because local datasets differ in size, diversity, and underlying distribution.Source: 2) Predictive Uncertainty of Client… · p. 4

C. Server - Federated Aggregation

The server coordinates federated training by distributing global parameters and aggregating client updates. Clients keep their datasets local while training and returning model parameters and validation metrics.

C. Server - Federated Aggregation

  • Each client trains an identical CNN from the current global model using its private dataset, then returns updated parameters and validation metrics.Training is conducted across three distributed clients without sharing raw infant movement data.Source: C. Server - Federated Aggregation · p. 4
  • Client-side uncertainty estimation uses MC-Dropout predictions on validation samples to compute predictive entropy and a client uncertainty score.The algorithm enables dropout for repeated stochastic predictions, then computes predictive means and entropy before estimating client uncertainty.Source: C. Server - Federated Aggregation · p. 4
  • The server aggregates received model updates to produce the next global model, using one of two aggregation approaches.The supplied passage identifies two approaches but does not specify their details here.Source: C. Server - Federated Aggregation · p. 4

1) Standard FedAvg:

The standard FedAvg approach updates the global model by averaging client model parameters with weights based on each client's training-sample count.

1) Standard FedAvg:

  • Standard FedAvg updates the global model using a weighted average of locally trained client model parameters.The number of clients is three in this study; each client’s training-sample count and local parameters are specified for the aggregation.Source: 1) Standard FedAvg · p. 4
  • Each client's aggregation weight is based on its number of training samples.The passage identifies n_i as the training-sample count for client i.Source: 1) Standard FedAvg · p. 4
  • The study uses three clients in the standard FedAvg setup.K denotes the number of clients.Source: 1) Standard FedAvg · p. 4

2) Uncertainty-Aware FedAvg (UA-FedAvg):

UA-FedAvg uses predictive entropy to inform client aggregation weights, giving greater influence to updates with lower predictive uncertainty. A variant also incorporates validation loss when calculating client weights.

2) Uncertainty-Aware FedAvg (UA-FedAvg):

  • During each round, clients train local CNNs, evaluate them on local validation data, compute mean predictive entropy, and send model updates and metrics to the server.The server computes uncertainty-aware weights and aggregates the client models.Source: 2) Uncertainty-Aware FedAvg (UA-FedAvg) · p. 5
  • UA-FedAvg w/ vLoss incorporates both predictive uncertainty and model performance into client aggregation weights.Validation loss is denoted L_i, and α controls the influence of the uncertainty and validation-loss terms; α was set to 0.25 in this study.Source: 2) Uncertainty-Aware FedAvg (UA-FedAvg) · p. 5
  • Client weights assign greater influence to updates with lower predictive uncertainty.The approach uses each client's mean predictive entropy estimated by MC-Dropout on its validation dataset.Source: 2) Uncertainty-Aware FedAvg (UA-FedAvg) · p. 5

D. Global Model Selection

The server selects the final global model using aggregated validation loss rather than automatically choosing the last communication round.

D. Global Model Selection

  • After each federated round, clients evaluate the received global model on their local validation sets.The server aggregates those validation losses using a weighted average.Source: D. Global Model Selection · p. 5
  • The server aggregates client validation losses using a weighted average.This aggregate is used to identify the best global checkpoint.Source: D. Global Model Selection · p. 5
  • The final model is the global checkpoint associated with the minimum aggregated validation loss observed during training.This selection avoids relying solely on the final communication round.Source: D. Global Model Selection · p. 5

V. EXPERIMENTS

The experiments use cross-subject validation and three client-distribution configurations to assess federated infant movement classification. Each infant’s motion segments remain assigned to one client, and the CNN follows the prior architecture and training procedure.

A. Experimental Settings

  • Nine-fold cross-subject validation used 36 infants for development and evaluation, reserving the other nine exclusively for hyperparameter optimisation.The 36 infants contributed 1,346 five-second clips; each fold trained on 32 infants and tested on four, with an 87.5%/12.5% training-validation split.Source: A. Experimental Settings · p. 5
  • Three client splits covered balanced training, consistent imbalance, and distribution shift, assigning all motion segments from each infant to one client.The settings were 111-211, 211-211, and 211-112, respectively.Source: A. Experimental Settings · p. 5
  • The CNN was reimplemented in PyTorch while preserving the original architecture and training procedure.Experiments used the hyperparameter configuration of the best-performing model reported in.Source: A. Experimental Settings · p. 5

B. Implementation details

Federated approaches substantially outperformed independent local models and approached centralized accuracy. Uncertainty-aware aggregation generally improved on FedAvg, with results varying across client splits and configurations.

B. Implementation details

  • MC-Dropout used 50 stochastic forward passes during training and 100 during inference, with dropout rate 0.3.Unless otherwise stated, experiments used 100 communication rounds, two local epochs, Adam at a 0.001 learning rate, and batch size four.Source: B. Implementation details · pp. 5–6

1) Aggregated Evaluation:

  • Federated approaches improved accuracy by over 7% versus No Federation, and their best configurations came within less than 1% of Centralized accuracy.UA-FedAvg w/ vLoss led accuracy, sensitivity, and F1 in 111-211 and 211-211; UA-FedAvg led all four classification metrics in 211-112.Source: 1) Aggregated Evaluation · p. 6
  • Performance gains over FedAvg were modest, but uncertainty-aware aggregation provided consistent benefits under certain data-partitioning schemes.Source: 1) Aggregated Evaluation · p. 6

2) Client-Level Evaluation:

  • UA-FedAvg generally achieved higher accuracy and F1 than FedAvg, except for Client #1 in 211-211; benefits were greater for clients with lower train-to-test ratios.The examples identified were Client #1 in 111-211 and Client #3 in 211-112, both with a 1:2 ratio.Source: 2) Client-Level Evaluation · p. 6
  • UA-FedAvg generally had lower median predictive uncertainty, while No Federation had the lowest average and standard deviation for most client configurations.The relationship between predictive uncertainty and model reliability in federated settings requires further investigation.Source: 2) Client-Level Evaluation · p. 6

VII. CONCLUSION AND FUTURE WORK

The paper concludes that federated learning can support infant movement analysis, with performance approaching centralized training and uncertainty-aware variants generally competitive with FedAvg. It identifies cohort scope and uncertainty-reliability assumptions as limitations and directions for further evaluation.

VII. CONCLUSION AND FUTURE WORK

  • Federated learning substantially improved classification over independent local models while approaching centralized performance; uncertainty-aware variants generally improved on FedAvg in most settings.The authors present this as the first federated learning framework for infant movement analysis and suggest predictive entropy may help guide aggregation.Source: §VII · p. 6
  • The strategy is model-agnostic and requires only a client-level uncertainty estimate, so it may extend to other federated applications where uncertainty can guide aggregation.Source: §VII · p. 6
  • The dataset contains recordings from typically developing infants, so results may not generalize to higher-risk clinical populations, including infants later diagnosed with CP.Future work will evaluate the framework on more diverse cohorts collected across multiple institutions.Source: §VII · p. 6
  • Aggregation assumes lower predictive entropy indicates more reliable client predictions, although low-entropy models may be overconfident and wrong under distribution shift.Planned work includes calibration-based measures, alternative uncertainty metrics, sample-level information, comparisons with other federated methods, and more rigorous statistical analysis.Source: §VII · pp. 6–7
Loading 2609.31463v1…