Dementia, particularly Alzheimer’s disease (AD), poses a growing global health challenge among the aging population. In the United States alone, over 6.9 million individuals aged 65 and older are estimated to be living with AD, a number projected to nearly double to 13.8 million by 2060 (Alzheimer's Association, 2024). Notably, 10.9% of this demographic is affected by the disease (Alzheimer's Association, 2024), underscoring its increasing impact.

Accurate diagnosis and effective treatment are complicated by the multifaceted nature of dementia, which is influenced by demographic, environmental, genetic, and biological factors. Artificial intelligence has emerged as a promising tool for cognitive-impairment screening, with machine learning (ML) and natural language processing applied to neuroimaging, electronic health records, speech, and other digital biomarkers. These methods show strong predictive performance and potential to improve diagnostic accuracy and efficiency, though concerns remain about misdiagnosis, confidentiality, and the psychological burden of screening (Wurtz et al., 2025; Arya et al., 2023). Among them, automated speech analysis has emerged as a non-invasive tool for detecting early cognitive decline (Hajjar et al., 2023). Language impairments associated with dementia manifest as difficulties in both speech production and comprehension (Bucks et al., 2000). However, large-scale studies leveraging speech data face significant obstacles, including concerns about patient privacy, speech de-identification, data sharing, and the need for collaboration across multiple research centers.

Federated learning (FL) has emerged as a promising ML approach for addressing these challenges by enabling multi-institutional data analysis while preserving patient confidentiality. Unlike traditional centralized methods, FL allows decentralized model training, ensuring sensitive information remains local (Kairouz et al., 2021; Hardy et al., 2017). This framework is particularly well-suited for privacy-sensitive domains such as healthcare, facilitating collaborative research while mitigating data security risks.

Although FL has shown promise in various healthcare applications (Teo et al., 2024), its use for speech-based dementia detection remains scarce and underdeveloped. Traditional centralized approaches face limitations due to restricted data access, ethical concerns, and biased representation, while variability in language, demographics, and recording conditions undermines generalization (García Gutiérrez et al., 2024; Sharafeldeen et al., 2025). FL offers a potential solution to these challenges, yet it introduces its own complexities. A key issue is data heterogeneity, where variations in dataset size and class distribution across institutions can negatively impact model fairness and overall performance (Hardy et al., 2017).

To our knowledge, until now there exists only a single prior work examining speech-based AD diagnosis within FL frameworks, which demonstrates its feasibility for AD detection from speech, with decentralized models achieving competitive performance while preserving privacy. However, these efforts rely on small, homogeneous datasets, and the impact of real-world heterogeneity on accuracy and fairness remains largely unexamined (Meerza et al., 2022).

To address this limitation, our study builds upon and significantly extends previous research by conducting a comprehensive evaluation of FL for dementia detection. Leveraging a substantially larger and more diverse dataset, we systematically evaluate model performance under various realistic scenarios, particularly emphasizing the challenges posed by data heterogeneity. Specifically, we investigate how data heterogeneity influences predictive accuracy and robustness. Our approach employs acoustic speech features within a multi-layer perceptron (MLP) neural network framework to distinguish between cognitively unimpaired (CU) and cognitively impaired (CI) individuals.

By examining the interplay between data variability and FL efficacy on a broader scale, this research provides deeper insights into optimizing FL models for real-world applications, ensuring both robust performance and equitable outcomes across diverse datasets.

Data from 2,239 participants were analyzed: 221 individuals who were CU, showing no objective cognitive or functional impairment (CDR = 0) (Jessen et al., 2014) and 2,018 participants who exhibited CI, including patients with mild cognitive impairment (MCI) (n = 1,219, CDR = 0.5) (Petersen, 2004) and dementia due to AD (n = 799, CDR > 0.5) (McKhann et al., 2011). Table 1 shows the clinical and sociodemographic characteristics of the sample used for this study.

Model performance was evaluated across three simulated FL scenarios designed to assess the effects of data imbalance and heterogeneity (Table 3). In scenario 1, where institutions contributed equally and class distributions were identical, local models achieved comparable balanced accuracies (0.71 and 0.73). Under FL, performance improved modestly: balanced accuracy increased to 0.73 and 0.79, and sensitivity rose from 0.69 to 0.82 at Node 1 and from 0.71 to 0.83 at Node 2, reflecting better detection of CI cases. Precision remained consistently high (>0.95) across all models. Specificity was stable at Node 2 but declined slightly at Node 1 (0.73 to 0.64). Importantly, FL enhanced overall discriminative ability, with AUC rising from 0.76 to 0.81 at Node 1 and from 0.82 to 0.87 at Node 2, and F1-score increasing from 0.80 to 0.88 at Node 1 and from 0.82 to 0.90 at Node 2 (Table 4).

In scenario 2, with a strong imbalance in dataset size, with Node 1 contributing only 10% of the total data (Table 3), the local model at that node struggled to identify CI cases, with a sensitivity of just 0.02 and a balanced accuracy of 0.51. This reflects the model’s inability to generalize with such limited training data. However, when applying FL, performance at Node 1 improved dramatically: balanced accuracy rose to 0.80 and sensitivity to 0.93, F1-score to 0.94, and AUC to 0.89, showing that FL training enabled robust CI detection. At Node 2, which had a much larger dataset, high performance was maintained under FL (balanced accuracy 0.70, F1-score 0.92, AUC 0.84), underscoring the scalability of the approach and its ability to distribute benefits fairly across nodes of unequal size. Notably, specificity decreased at both nodes (from 1.00 to 0.67 at Node 1 and from 0.77 to 0.51 at Node 2), reflecting a trade-off where improved CI detection came at the expense of more CU misclassifications.

In scenario 3, class distribution was imbalanced across institutions, with Node 1 having 97% CI cases and only 3% CU cases, while Node 2 had a more representative mix (Table 3). Local models performed similarly to scenario 1, with balanced accuracies of 0.70 (Node 1) and 0.74 (Node 2). Under FL, Node 1 maintained similar balanced accuracy (0.71) but with shifted trade-offs: sensitivity decreased from 0.71 to 0.57 and F1-score from 0.82 to 0.73, while specificity improved from 0.69 to 0.85, suggesting that FL training allowed the model to better identify CU cases despite their scarcity. Node 2 showed stable balanced accuracy (0.72) and slightly higher AUC (0.84), indicating that federated training conferred benefits, particularly in recognizing minority-class CU samples without loss of overall classification power.

The complete confusion matrices for each scenario, including local and FL models, are provided in the Supplementary Table S1.

In this study, an MLP model was trained using speech data and developed using a federated network to predict CI, specifically distinguishing between cognitively healthy individuals, and patients with MCI and dementia. FL was tested to confirm its ability to enable collaboration among multiple institutions while safeguarding patient privacy. This decentralized approach is particularly advantageous in healthcare, where privacy is a significant concern, as it facilitates multi-site model training without requiring the exchange of raw data (Zhang et al., 2024).

The choice of MLP reflects the trade-offs inherent in FL. While more complex architectures (e.g., CNNs, RNNs, Transformers) may achieve higher performance in centralized settings, they are computationally demanding and more prone to divergence under heterogeneous data. MLPs, in contrast, are lightweight, stable, and communication-efficient, aligning well with structured acoustic features such as the 176 descriptors used here. Empirically, this balance proved effective: the MLP achieved strong AUC and F1-scores across scenarios, while FL improved sensitivity in low-data nodes and enhanced generalization under class imbalance. Thus, the MLP offered the most pragmatic compromise between expressive power, robustness, and scalability (Zhang et al., 2024).

Equally important was the choice of aggregation strategy. The combination of FedAvg and IDA proved particularly suitable, as it consistently improved performance under challenging conditions. In Scenario 2, where the smaller institution held only 10% of the data, local training nearly failed (sensitivity = 0.02), yet the federated model raised sensitivity to 0.93 and AUC to 0.89, showing clear benefits for underrepresented nodes. In Scenario 3 with severe class imbalance, this combination also improved specificity (0.69 to 0.85) and AUC (0.79 to 0.81), enhancing recognition of CU cases. FedAvg offered scalability and efficiency, while IDA mitigated heterogeneity, together providing a balanced and pragmatic solution (Meerza et al., 2022). Other strategies, such as FedProx or Scaffold, address non-IID conditions by adding proximal terms or variance-reducing corrections, but they require additional hyperparameter tuning and may increase communication or computation costs, making them less practical for early-stage clinical deployment. Similarly, clustered FL can personalize models for subgroups, but at the expense of interpretability and cross-site comparability, which are essential in healthcare. Against this backdrop, FedAvg combined with IDA offered the best balance between simplicity, robustness, and feasibility (Qi et al., 2023).

Overall, the federated MLP consistently achieved AUC values above 80%, demonstrating reliable discrimination between CU and CI. Its greatest impact emerged under scarcity and imbalance: in Scenario 2, balanced accuracy rose from 0.51 to 0.80 and sensitivity from 0.02 to 0.93, while in Scenario 3, FL improved specificity and AUC, aiding recognition of minority CU cases. By contrast, gains in balanced settings (Scenario 1) were modest, indicating that FL’s strength lies less in boosting absolute accuracy than in ensuring fair and robust performance across heterogeneous institutions. These results highlight FL as a privacy-preserving framework that promotes equity, allowing smaller or skewed sites to contribute meaningfully without being disadvantaged.

Taken together, the findings confirm that the decentralized nature of FL provides a scalable and privacy-preserving solution for CI detection, enabling robust model development without direct data sharing (Meerza et al., 2022). By addressing dataset size inequality and class imbalance, FL supports fairer predictions across diverse populations and is particularly valuable for detecting subtle cognitive changes in early disease stages (Mateus et al., 2023). Beyond modest performance improvements, its benefits extend to equity across institutions, compliance with privacy regulations, enhanced security through decentralization, and scalability across heterogeneous clinical environments.

However, several limitations should be acknowledged. First, although the cohort included 2,239 participants, all data came from a single institution. The simulated disparities between two virtual sites with harmonized preprocessing cannot replicate true cross-site heterogeneity, where differences in devices, demographics, diagnostic protocols, and annotation quality are more pronounced. Real-world clinical settings also introduce variability in speech tasks, legal constraints, and annotation standards, which may affect both convergence and fairness. These factors highlight the need for validation beyond controlled simulations.

Second, while FL demonstrates strong potential for collaborative and privacy-preserving modeling, it is not immune to security vulnerabilities. Risks such as model inversion attacks, membership inference, and data leakage could compromise patient confidentiality or model integrity. Mitigating these risks will require safeguards including differential privacy, secure aggregation, and adversarial robustness strategies, alongside standardized monitoring and quality control, to ensure that FL systems remain safe and trustworthy in clinical deployment (Vasa et al., 2024).

Future work should prioritize multi-institutional pilots to evaluate scalability across larger networks (e.g., 5–10 sites) using diverse devices, protocols, and languages, thereby testing cross-linguistic generalizability. These studies should compare aggregation strategies (e.g., FedAvg vs. adaptive methods) and investigate personalization, domain adaptation, and harmonization frameworks to balance local adaptation with global performance. Evaluation must extend beyond AUC and F1-scores to include fairness, calibration, and node-specific robustness, ensuring that benefits are not driven solely by data-rich sites. Expanding FL to geographically distributed cohorts and incorporating multimodal data sources, such as cognitive assessments, neuroimaging, and electronic health records, will further enhance diagnostic accuracy and clinical applicability. To achieve this, key steps include: (1) establishing multi-institutional pilots under real-world conditions, (2) developing standardized protocols for data harmonization and evaluation across sites, (3) creating interoperable, privacy-compliant FL infrastructure aligned with regulations, and (4) building decision-support tools to integrate FL outputs into clinical workflows. Establishing harmonized speech protocols and secure, standardized infrastructure will be essential to move beyond experimental validation and enable safe deployment in healthcare systems (Dhade and Shirke, 2024).

Overall, this study highlights the promise of FL as a scalable, secure, and effective approach for early detection of cognitive impairment. By enabling privacy-preserving collaboration across institutions, FL addresses key challenges in digital health research and opens new avenues for developing accessible and equitable diagnostic tools for neurodegenerative diseases.