This study’s database is based on the mPower Parkinson study database obtained from the mPower Public Research Portal [11,12]. mPower is a Parkinson’s disease clinical study performed only through a mobile application interface that consists of seven tasks that each participant must complete, three survey questionnaires (a demographic survey, the Unified-Parkinson’s Disease Rating Scale survey, and the Parkinson’s Disease-Questionnaire-8), and four activities (the memory task, the tapping task, the voice task, and the walking task). In this study, we were only interested in two tasks (the voice activity task and the demographic survey). Our original dataset was created in three stages: acquiring raw audio recordings, selecting valid participants, and finally extracting audio features. Python and SQL were used with the Synapse client to collect the voice recordings [13]. The demographic survey was then used to differentiate between participants with Parkinson’s disease (PWPs) and healthy controls (HCs).

Feature extraction is critical in ML and pattern recognition systems [15], mainly when dealing with audio data. Because audio signals are non-stationary, feature extraction is performed on a frame basis by dividing the signals into short frames [16]. All the recordings were divided into 25-ms windows with a 10-ms step size (usually between 20 and 40 ms [16]) and sampled at 44.1 kHz. Using the Surfboard [17] library, we extracted 123 audio features. This Python library is built with state-of-the-art audio analysis packages and provides an easy-to-use API to extract several audio features applicable in the medical domain. Surfboard allows the calculation of important audio feature statistics to obtain a feature-rich dataset. Table 2 summarizes the extracted features and showcases the structure of our final dataset.

Before starting the modeling phase, we divided the dataset into two sets: a training set and a holdout set. The training set was used to train the pipelines and evaluate their performance using subject-wise cross-validation (CV), while the holdout set was used to verify the models’ generalizability on unseen data. For data partitioning and using the subject-wise division approach [18], we allocated 80% of the dataset to the training set and 20% to the holdout set. This division guarantees that the subjects (or participants) and their relative recordings in the training set are independent of the subjects and their relative recordings in the holdout set. Table 3 details the distribution of participants in both sets.

In this step, we applied two forms of preprocessing [19] to the dataset: data imputation and data scaling. Data imputation replaces missing values in a feature vector with the mean, the median, or the most frequent value of the same feature vector; this step is required because missing values can cause problems for ML algorithms. Data scaling standardizes the dataset, as it contains various features with varying scales. This step ensures that all features contribute equally to the learning process. Table 4 presents more details.

Feature selection is an essential concept in ML. Having a proper subset of features reduces the complexity of a model, enables it to be trained faster, and makes it easier to interpret. Researchers use various feature selection techniques [20], such as filter, wrapper, and embedded methods. In this work, we used two embedded methods. The best subset of features was selected using a meta-transformer based on feature importance, with Lasso or ElasticNet [21] as the primary estimator.

To develop our models, we used pipelines. A pipeline automates an ML workflow by dividing it into reusable independent modular parts: in our case, data imputation, data scaling, feature selection, and then classification. In this study, we created two pipelines: a gradient boosting classifier (GBC) pipeline with Lasso (gbcpl) and a GBC pipeline with ElasticNet (gbcpen); their hyperparameters were tuned using the randomized search technique [22] (Table 3). To correctly estimate the performance of the pipelines, we used a nested CV approach, where two subject-wise CV techniques were grouped to form two loops, an inner loop and an outer loop, as shown in Figure 1. In the outer loop, we used the stratified-group k-fold CV technique with k = 10. In each loop, the training set is divided into 10 parts or folds; nine parts are used to train the pipelines, and the remaining part is used for validation. The inner loop divides the nine parts into five using stratified-group k-fold CV with k = 5. Four parts are used to fit several pipelines with random hyperparameters. The remaining part is used to select the best pipeline with specific hyperparameters by maximizing a scoring metric (in our case, recall). When the best pipeline is found in the inner loop, it is used by the outer loop to estimate its performance on the validation set. This step is repeated 10 times, and the performance of the pipelines is reported as the mean (standard deviation) for each performance metric. This technique is time-consuming, considering the time needed to find optimal hyperparameters by the inner loop. However, it has been demonstrated to produce accurate results and reduces overfitting when dealing with small datasets [23]. To assess the performance of the pipelines, we calculated three performance measures: accuracy, recall, and the F1-score.

The client-side desktop application is a Windows application developed entirely using Python and Kivy [24], an opensource, cross-platform graphical framework for natural user interface development. Our application has six screens with three main features: recording audio data, extracting audio features and communicating with the server via HTTP requests (Figures 2, 3A). Voice recording is done using PyAudio [25], which provides Python bindings for PortAudio, a cross-platform audio I/O library used to record and play audio on desktop operating systems such as Windows, Linux, and macOS. Each recorded file is sampled at 44.1 kHz and saved in .wav format with a 10-second duration. Feature extraction is performed using Surfboard; we can extract audio features either from the recorded .wav file or from a test .wav file that is available on the computer (for instance, using test recordings from the holdout set). Extracted features are then sent to the server via HTTP requests in the JSON format to a prediction endpoint. This endpoint is developed using a REST API, where each HTTP request includes a header with user credentials in the form of a token that is generated for the user when successfully logged in to the server. Only authenticated users to the server can use the desktop application. An HTTP response is returned to the application in JSON format when the server successfully authorizes the user and finishes processing the request with the extracted features by a selected ML pipeline. We packaged our application into a Windows executable file using PyInstaller [26] and Inno [27] for ease of use.

The server-side web application is developed entirely using Django [28], an open-source, high-level Python web framework that encourages rapid development and clean, pragmatic design. The server’s database is based on PostgreSQL, an open-source relational database management system. The ML pipelines were trained using Scikit-learn [29], saved using Joblib, and accessed via a prediction endpoint developed using Django Rest Framework. When a user posts a request to the server with the extracted audio features, those features are deserialized by the REST API and stored in the database to be processed by the selected ML pipeline. When processing is done, the response of the ML pipeline is stored in the database and then serialized by the REST API in a JSON format to be used by the desktop application. The web application also provides a mechanism for user management through token authentication. Its initial design is developed using HTML, CSS, and Bootstrap and served using WhiteNoise. We used Gunicorn for deployment, a pure-Python WSGI HTTP server. The entire project was then containerized using Docker in an Ubuntu image and deployed to Heroku, a platform-as-a-service cloud provider. Our web application is accessible via the following link: https://pdpredict.herokuapp.com, and its architecture is presented in Figure 3B.