Federated AI Screens Depression and Anxiety While Shielding Patient Privacy

A new open-access study in Discover Computing presents PP-XMDA, a framework that detects depression and anxiety from text, audio, and visual information while keeping raw participant data on local devices and explaining, in auditable form, how each modality contributes to a prediction. Across three benchmark corpora, the strongest depression-detection result reaches an F1 score of 0.806 and an area under the ROC curve of 0.873, achieved without ever centralizing sensitive interview recordings.

The system’s designers argue that privacy exposure and limited interpretability are the two biggest obstacles to multimodal mental-health screening. Centralized multimodal models can perform well but require aggregating data that may reveal identity, health status, and personal context. Privacy-preserving approaches, in turn, often skip explanation quality, and explanation-oriented systems rarely evaluate formal privacy guarantees. PP-XMDA bundles five linked components into one pipeline: participant-level local processing, federated optimization with secure aggregation, differentially private gradient updates, availability-aware attention fusion, and multi-level attribution paired with uncertainty estimates.

Under federated learning, each client, such as a hospital or research institution, trains locally and shares only protected parameter updates. Secure aggregation combines secret sharing and pairwise masking so the server observes only the aggregate rather than any individual update. Differential privacy adds calibrated Gaussian noise to clipped gradients, with privacy loss tracked by a Rényi differential privacy accountant. The authors evaluated privacy budgets ranging from epsilon 0.5 to 8 and selected epsilon 4.0 as the largest reduction in attack success that kept the relative F1 loss below a predefined 2 percent threshold.

The privacy-utility measurements come from a shadow-model membership-inference attack using posterior confidence, prediction entropy, and per-sample loss, with identical training procedures, thresholds, and class priors across all configurations. Federated learning without differential privacy already lowered the reported attack success metric from 0.624 to 0.542 while retaining F1 of 0.806. Adding differential privacy at epsilon 4.0 cut the metric to 0.352 and attribute-inference accuracy to 0.339, at a cost of F1 dropping from 0.806 to 0.794, a 1.49 percent relative loss. The authors stress that the metric is interpreted only as a comparative attack outcome under the fixed protocol, not as proximity to a random-guessing benchmark.

The experimental evaluation covers three datasets: MMDA with 1,025 de-identified interview records labeled through professional HAMD and HAMA assessment, MMPsy with 1,000 anonymized adolescent records carrying PHQ-9 and GAD-7 scores, and DAIC-WOZ with 189 interview sessions labeled by PHQ-8 under an official 107/35/47 partition. On MMDA, PP-XMDA improved depression F1 by 2.4 percent and AUC by 2.6 percent over the strongest fusion baseline, with two-sided Welch tests across five random seeds confirming the differences. On MMPsy, an audio-text configuration reached F1 of 0.804, and external validation on DAIC-WOZ yielded F1 of 0.771, a 4.3 percent relative reduction the authors attribute to population and protocol shifts rather than to privacy noise, which was unchanged.

Robustness to missing inputs is built into the architecture through a modality availability mask and modality dropout during training. Under clinic-level exclusion tests, removing visual input changed depression F1 only from 0.806 to 0.798, removing audio lowered it to 0.778, and removing text lowered it to 0.749. When only one modality remained, text-only inference reached F1 of 0.774, compared with 0.713 for audio and 0.695 for visual input alone. Cross-dataset transfer experiments produced F1 between 0.742 and 0.785, with the largest drop occurring when a model trained on the smaller DAIC-WOZ set was moved to MMDA, a pattern the authors read primarily as domain shift.

Explanation is handled at three levels. Attention weights over available modalities serve as contribution indicators, with text receiving the largest weights, roughly 0.44 to 0.53 across datasets, followed by audio and then visual signals. Feature-level saliency and Integrated Gradients attributions identify influential inputs, and temporal explanations highlight high-attribution speech or facial segments. Crucially, the authors do not treat attention as a faithful causal account; faithfulness is tested separately through deletion, insertion, and perturbation-stability measures, where explanation stability rose from 0.841 to 0.892 with the regularization term active. Monte Carlo dropout over 50 forward passes supplies uncertainty estimates, and outputs exceeding fixed thresholds are routed for qualified review rather than issued as final determinations.

The authors are explicit about limits. Physiological and behavioral encoders exist only as architectural extension pathways and contribute nothing to the reported results, since the evaluated corpora lack such measurements. Federated optimization ran as a controlled logical-client simulation without wide-area latency, bandwidth limits, packet loss, client dropouts, or institution-specific non-IID effects, so scalability to real multi-site deployment remains a separate engineering question. Subgroup auditing used only demographic fields supplied with the source data, finding maximum F1 differences below 0.006 across available gender, age, and education groups, a result that characterizes the evaluated samples and is not generalized beyond them. Regression for symptom scores reduced MAE to 3.76 for HAMD and 3.85 for HAMA, values the authors treat as predictive-error measures rather than evidence of clinical equivalence.

The framework’s outputs are positioned as model-auditing and decision-support information, not validated clinical diagnoses. The discussion lists potential harms including false-positive escalation, false-negative under-referral, subgroup-specific error amplification, overreliance on automated explanations, and leakage from improperly handled updates, and it calls for qualified human oversight, uncertainty-based escalation, subgroup monitoring, access-controlled audit logs, periodic recalibration, and a clear prohibition of punitive or autonomous decisions based solely on model output. Clinician-facing validation assessing explanation usefulness, trust calibration, and the effect of uncertainty flags is named as the required next stage, alongside external validation in multilingual populations and communication-constrained federation. The implementation, privacy-accounting settings, and participant-level split manifests are available from the corresponding author on reasonable request, with the underlying datasets accessible through their original publication channels.

Subject of Research: Privacy-preserving federated multimodal machine learning for detecting depression and anxiety

Article Title: Privacy-preserving and explainable multimodal detection of depression and anxiety

Article References: Alharbi, L. A. (2026). Privacy-preserving and explainable multimodal detection of depression and anxiety. Discover Computing, 29(1), Article 822. https://doi.org/10.1007/s10791-026-10758-x

Image Credits: AI Generated

DOI: 10.1007/s10791-026-10758-x

Keywords: federated learning, differential privacy, multimodal machine learning, depression detection, anxiety detection, explainable AI, membership inference, mental health screening, uncertainty estimation, attention fusion, DAIC-WOZ, decision support