FedOcw: optimized federated learning for cross-lingual speech-based Parkinson’s disease detection

To evaluate the effectiveness of our proposed federated learning framework for speech-based Parkinson’s disease detection, we utilized five multilingual datasets, incorporating Spanish, Italian, Chinese, Czech, and English speech samples. These datasets vary in recording conditions, linguistic structure, and phonetic tasks, providing a diverse and heterogeneous training environment that closely resembles real-world clinical scenarios.

Dataset-1 (Spanish), sourced from the PC-GITA repository⁴⁰, comprises speech recordings from 100 individuals, including 50 Parkinson’s disease (PD) patients and 50 healthy controls (HCs). All recordings were conducted in professional soundproof booths at 44.1 kHz sampling frequency with 16-bit resolution. The PD participants, aged 33 to 81 years, were evaluated in the ON state by three expert phoneticians. Speech samples included:

1. (1)

   Sustained vowels: Three repetitions of /a/, /i/, /e/, /o/, and /u/.

2. (2)

   Isolated words: /blusa/, /petaka/, /apto/, /campana/, /llueve/, /reina/, /braso/, and /viaje/.

3. (3)

   Sentence reading: Simple (/laura/, /loslibros/, /luisa/, etc.) and complex (/preocupado/, /juan/, etc.) structures.

4. (4)

   Spontaneous speech: Monologues (~44.86 s on average).

These tasks were designed to capture phonation, articulation, and prosody impairments, which are critical for detecting Parkinsonian dysarthria.

Dataset-2 (Italian) was originally developed to assess speech intelligibility in PD patients using automatic speech recognition systems⁴¹. This dataset⁴² includes 28 PD patients and 37 HCs, featuring recordings of:

1. (1)

   Phonemically balanced text reading (twice, with a 30-s pause).

2. (2)

   Repetitions of syllables (/pa/ and /ta/ for 5 s each).

3. (3)

   Sustained vowels (/a/, /i/, /e/, /o/, /u/).

4. (4)

   Phonemically balanced word and phrase reading.

Recordings were conducted in low-noise, echo-free environments, with microphones placed 15–25 cm from the speaker’s lips. All PD participants were receiving antiparkinsonian treatment.

Dataset-3 (Chinese), obtained from the GYENNO SCIENCE Parkinson’s Disease Research Center³⁷, consists of 30 PD patients and 15 HCs, aged 37 to 75 years. Speech tasks included:

1. (1)

   Sustained vowels (/a/ and /e/).

2. (2)

   Short sentence reading (e.g., /si shi si zhi shi shi zi/).

Speech samples were recorded using smartphones positioned 10 cm from the speaker’s mouth. All PD participants were assessed by two neurologists and recorded in the ON state.

Dataset-4 (Czech) was designed to differentiate idiopathic Parkinson’s disease from other parkinsonian syndromes via prolonged vowel analysis⁴³. The dataset includes 22 PD patients, alongside 21 patients with multiple system atrophy, 18 with progressive supranuclear palsy, and 22 HCs. For this study, we utilized data from PD patients and HCs only.

Recordings were performed using a headset condenser microphone (5 cm from the lips) at 48 kHz sampling frequency with 16-bit resolution. Participants were instructed to sustain vowels (/A/ and /I/) in a modal voice for as long and steadily as possible.

Dataset-5 (English), the MDVR-KCL dataset (Mobile Device Voice Recordings at King’s College London)⁴⁴, was developed to explore non-invasive Parkinson’s disease monitoring through smartphone-based voice analysis⁴⁵. This dataset includes 16 PD patients and 21 HCs, recorded using a Motorola Moto G4 smartphone at 44.1 kHz sampling frequency with 16-bit resolution.

Table 1 provides a summary of the demographic, clinical, and recording characteristics of participants, including the distribution of PD and HC groups by gender, age, disease severity, recording conditions, and speech tasks across Dataset-1 to Dataset-5. These datasets provide a comprehensive and multilingual foundation for evaluating federated learning models in Parkinson’s disease detection.

To evaluate the effectiveness of the proposed federated learning approach on multilingual speech data, we define five experimental scenarios that integrate five datasets with varying language distributions and client allocations:

Scenario A (Fig. 1): Speech data from Dataset-1 (Spanish) and Dataset-2 (Italian) is distributed across eight clients (C0–C7), with an uneven distribution of Parkinson’s disease (PD) cases and healthy controls (HCs). Clients C0–C3 are assigned data from Dataset-1 (Spanish), while Clients C4–C7 receive data from Dataset-2 (Italian).

Scenario B (Fig. 1): Speech data from Dataset-1 (Spanish) and Dataset-3 (Chinese) is allocated to seven clients (C0–C6). Clients C0–C3 are assigned data from Dataset-1 (Spanish), while Clients C4–C6 are assigned data from Dataset-3 (Chinese).

Scenario C (Fig. 1): Speech data from Dataset-2 (Italian) and Dataset-3 (Chinese) is used, with seven clients (C0–C6). Clients C0–C3 receive data from Dataset-2 (Italian), and Clients C4–C6 are assigned data from Dataset-3 (Chinese).

Scenario D (Fig. 1): Speech data from Dataset-1 (Spanish), Dataset-2 (Italian), and Dataset-4 (Czech) is used. Clients C0–C3 are assigned Dataset-1 (Spanish), C4–C7 receive data from Dataset-2 (Italian), C8 is allocated Dataset-4 (Czech).

Scenario E (Fig. 1): All five datasets are incorporated for a comprehensive multilingual evaluation. Clients C0–C3 are assigned Dataset-1 (Spanish), C4–C7 receive Dataset-2 (Italian), C8–C10 are allocated Dataset-3 (Chinese), C11 is assigned Dataset-4 (Czech), and C12 receives Dataset-5 (English).

This experimental setup enables a comprehensive evaluation of the federated model’s generalization across linguistically diverse datasets. Each client was assigned speech samples from its respective dataset, which included a variety of task types such as sustained vowels, sentence reading, and spontaneous speech. A single model was trained per client using the entire local training dataset, without further partitioning based on individual speech tasks. This approach reflects real-world deployment conditions in federated learning, where heterogeneity in assessment protocols and data characteristics is common across different clinical sites.

To promote robust generalization, all speech samples were partitioned into non-overlapping training and test sets. Training data remained strictly localized on each edge client, while evaluation was independently performed on each client’s separate test node. Importantly, no speaker overlap existed across clients’ training sets, strengthening the model’s ability to generalize across languages and participants. For final evaluation, each client was tested on its corresponding test node using local testing samples, providing a comprehensive assessment of the model’s cross-lingual performance.

Regarding the choice of languages for the bilingual experiments, Spanish, Italian, and Chinese were prioritized due to the availability of well-balanced datasets with large sample sizes and a diverse set of speech tasks. These characteristics provided a robust and heterogeneous foundation for evaluating cross-lingual generalization. In contrast, the English and Czech datasets, while valuable, had comparatively smaller sample sizes and fewer speech tasks, limiting their suitability for the bilingual scenarios. Instead, English and Czech were incorporated in Scenarios D and E to further explore the impact of increasing language diversity on model performance.

Figure 1 provides a circular visualization of client distributions, including sample sizes, case-control ratios (shown as bar plots), and the number of participants (indicated in brackets). Percentages represent each client’s relative contribution to the overall training dataset.

In Fig. 1, the box plots present the evaluation results over 100 rounds of federated aggregation, capturing performance across accuracy, F1-score, and Matthews correlation coefficient (Mcc). The mathematical formulations for these metrics are detailed in Eqs. (1)–(5).

Here, TP, TN, FP, and FN denote the numbers of true positives, true negatives, false positives, and false negatives, respectively. Sensitivity and specifity quantify the model’s ability to correctly identify positive and negative cases. The F1-score represents the harmonic mean of sensitivity and specifity, providing a balanced assessment of classification performance. Mcc measures the overall quality of binary classifications, ranging from −1 to +1, where +1 indicates perfect prediction, −1 signifies total disagreement between predictions and actual labels, and 0 reflects performance equivalent to random guessing.

Tables 2–6 summarize the average performance across 100 aggregation rounds for Scenarios A, B, C, D, and E with the best-performing federated learning methods highlighted in bold. Alongside federated learning approaches, the tables also report the average results for clients trained and evaluated on their isolated local datasets (Local) and the outcomes of centralized learning for comparison.

To ensure a fair and meaningful comparison, all baseline methods were carefully tuned and evaluated under consistent experimental conditions. For FedProx, we explored the proximal term coefficient μ ∈ {0.01, 0.1, 1.0} and selected the value μ = 0.1 that achieved the best performance in each experimental setting. For SCAFFOLD, we followed the standard configuration, with control variates updated at the end of every local training round. The standard implementation of FedNova was used without modification. To maintain comparability across methods, all experiments used the same optimization settings: Adam optimizer with a learning rate of 0.001, 10 local epochs per round, and a training batch size of 8.

As shown in Tables 2–6, FedOcw consistently outperforms conventional FL methods across all evaluated scenarios (A–E) in accuracy, F1-score, specifity, sensitivity, and Mcc, demonstrating superior stability and training effectiveness. It not only surpasses FedAvg, FedProx, Scaffold, and FedNova, but also outperforms centralized learning across all key metrics. These findings highlight the advantages of federated models in privacy-preserving and heterogeneous learning environments.

FedOcw’s adaptability to linguistic diversity is evident across all scenarios. In the Spanish–Italian setting (Scenario A), it achieves the highest accuracy (74.81%) and Mcc (0.502), demonstrating effective knowledge transfer between related languages. In the Spanish–Chinese scenario (Scenario B), the model maintains strong performance with 67.85% accuracy and an Mcc of 0.288, though the increased linguistic divergence presents convergence challenges. In the Italian–Chinese setting (Scenario C), FedOcw achieves high specifity (84.19%) and sensitivity (83.44%), indicating a balanced classification approach. In the trilingual scenario (Scenario D), it maintains top performance with the highest accuracy (72.53%), F1-score (69.8%), and Mcc (0.465). Even in the most heterogeneous multilingual scenario (Scenario E), the model sustains robust performance, achieving 72.63%accuracy and an Mcc of 0.435, highlighting its robustness and ability to generalize across linguistic domains.

Table 7 reports the p values for the mean Accuracy, F1-score, and Mcc metrics across all five scenarios, evaluating the statistical significance of differences between our proposed federated model (FedOcw) and alternative methods.

Table 7 shows that FedOcw achieves statistically significant improvements over individual learning (Local), alternative federated learning methods, and centralized learning in key performance metrics, including accuracy, F1-score, specifity, and Mcc. However, sensitivity is an exception, where it performs comparably to FedProx (p = 0.0525), indicating no significant difference. Compared to centralized learning, FedOcw demonstrates significant advantages across all evaluation metrics, reinforcing its effectiveness in diverse settings. Notably, FedOcw outperforms FedAvg with strong statistical significance in specifity (p = 0.0003) and Mcc (p = 0.0011). However, the lack of statistical significance in some comparisons (p > 0.05) for sensitivity with FedProx) suggests that certain methods may still be competitive in specific aspects. These findings highlight FedOcw’s robustness in handling heterogeneous and multilingual datasets, reinforcing its potential for broader cross-linguistic and clinical applications.

To better understand the model’s behavior across multilingual settings, we conducted a language-wise accuracy analysis of client models in Scenarios A–E. Figure 2 presents the individual accuracy scores for each language (Spanish, Italian, Chinese, Czech, and English), highlighting the specific contributions of each client group to the overall federated learning performance.

As shown in Fig. 2, the Italian client consistently achieves the highest accuracy across scenarios in which it is present, reaching up to 94% in Scenario C and 91.6% in Scenario A. This suggests that the Italian dataset may contain more consistent or discriminative speech features for Parkinson’s detection, possibly due to better recording conditions, more clearly defined task protocols, or less intra-class variability. In contrast, Spanish and Chinese clients show more variable performance, with Chinese accuracy rising from 63.14% in Scenario B to 67.8% in Scenario C, depending on the pairing.

The performance gap between Scenario B (Spanish–Chinese) and Scenario C (Italian–Chinese) is particularly informative. While both involve cross-lingual collaboration with Chinese data, Scenario C significantly outperforms Scenario B. This may be attributed to greater similarity in task structure or feature distribution between Italian and Chinese datasets, leading to more effective model generalization. Alternatively, the Spanish dataset may differ more substantially in prosody, phonetic structure, or participant characteristics, making knowledge transfer more challenging.

A similar trend is observed when comparing Scenario A (Spanish–Italian) and Scenario D (Spanish–Italian–Czech). In Scenario A, a large performance gap exists between Italian (91.6%) and Spanish (58.01%), suggesting unbalanced contributions and potential dominance of the Italian dataset during model aggregation. However, when Czech is added in Scenario D, the gap narrows: Italian performance drops slightly to 83.87%, while Spanish improves to 63.52%, and Czech reaches 63.24%. This shift indicates that adding a third, linguistically distinct client introduces more diversity into the training process, which likely promotes better generalization across heterogeneous clients.

In Scenario E, where five languages are present, performance becomes more balanced across clients with different languages, though Italian still maintain relatively strong accuracy. This suggests that FedOcw is able to preserve generalization even under high linguistic and distributional heterogeneity.

To evaluate the global stability and efficiency of the federated learning framework, Fig. 3 presents the training loss convergence of various federated learning models across five evaluation scenarios (A, B, C, D, and E). The models compared include FedAvg, FedProx, Scaffold, FedNova, and the proposed FedOcw. The x-axis denotes the number of communication rounds, while the y-axis represents the average training loss across local clients. A lower training loss over time indicates improved convergence and model stability.

As shown in Fig. 3, FedOcw consistently achieves the lowest training loss across all scenarios, demonstrating superior convergence stability and effectiveness. In Scenario A, it stabilizes at a loss of ~0.3, while other models exhibit significant fluctuations, indicating sensitivity to data heterogeneity. Scenario B follows a similar pattern, with FedOcw maintaining low and stable training loss, whereas FedAvg and FedNova experience sharp oscillations, leading to poor convergence. In Scenario C, FedOcw again outperforms all models, stabilizing around 0.2, while the other methods struggle to converge, with increasing training loss over rounds, reflecting poor adaptation to the scenario. Similar trends are observed in Scenario D and E, where FedOcw demonstrates the best stability, while FedAvg, FedProx, and FedNova continue to show erratic loss patterns. These findings underscore FedOcw’s robustness in addressing non-IID data challenges, offering enhanced convergence stability and adaptability across diverse multilingual datasets. The observed training loss trends further highlight its resilience in handling complex learning environments, making it a promising candidate for real-world federated learning applications.

To examine the impact of the weighting strategy on individual clients during the federated learning process, we analyze client model C0 across five scenarios (A, B, C, D, and E) as case studies, focusing on the optimized client weights assigned by FedOcw. Table 8 presents the sample standard deviation (STDEV.S) over 100 rounds for the optimized weights of local clients when updating client model C0 in the five scenarios, considering various layer parameters of the deep learning model.

As presented in Table 8, the weights assigned to the Time-Distributed 2D-CNN layer exhibit the highest variability across aggregation rounds, underscoring their critical role in shaping the deep learning model’s performance. A similar trend is observed across other client models, indicating the central influence of this layer in the federated learning process. Given this, we focus on the Time-Distributed 2D-CNN layer for a more in-depth analysis of how the weighting strategy impacts individual clients during training.

Figure 4 shows the adjacency matrix of the weights assigned to the Time-Distributed 2D-CNN layer across five scenarios (A, B, C, D, and E). The y-axis represents the clients receiving updates, with each row corresponding to the aggregate weights assigned to the local clients. The weights are averaged over 100 rounds. The color bar visually indicates the weight values, emphasizing the relative importance of each client’s input space to the target client receiving updates.

As shown in Fig. 4, FedOcw does not confine weight assignment to clients within the same language group across all scenarios. Instead, updates are exchanged between clients from different linguistic backgrounds, demonstrating that the model enables cross-lingual knowledge transfer without imposing language-based isolation. The weight distribution remains relatively balanced, ensuring that model updates are equitably shared, allowing each client to both contribute to and benefit from diverse sources. Additionally, certain clients receive higher-weighted updates, suggesting that the personalization strategy enhances model performance by dynamically prioritizing influential clients. Importantly, these higher-weighted assignments do not consistently correspond to a specific language group, reinforcing the model’s adaptability.

To better understand the dynamics behind these weight assignments, we examined Scenario E to determine whether the most influential clients, defined as those consistently receiving higher weights, correlate with dataset-specific attributes such as training sample size, class distribution, or speech task diversity. Table 9 presents hypothetical examples illustrating this analysis.

As shown in Table 9, The analysis of FedOcw’s weight assignment strategy in Scenario E reveals that client influence is not determined solely by dataset size or task diversity. While one might expect larger or more diverse datasets to receive higher weights, FedOcw instead appears to prioritize clients with balanced class distributions, as these tend to contribute more reliable and generalizable updates. Notably, the Czech client (C11) with relatively small dataset and limited task diversity receives one of the highest weight assignments, suggesting that FedOcw values the informativeness and alignment of updates over raw data quantity. This indicates that FedOcw adopts a nuanced aggregation strategy that promotes fairness and generalization by emphasizing the quality and complementary value of each client’s contribution rather than relying on size or frequency alone.