The PCompanion project set out to develop a patient‑centric screening and monitoring system for Parkinson’s disease, with a particular focus on detecting rapid‑eye‑movement (REM) sleep‑behavior disorder (RBD) through automated analysis of polysomnographic data. In the predictive analytics sub‑project, researchers extracted REM‑phase segments from overnight recordings of a small cohort of patients and applied seven supervised machine‑learning algorithms—AdaBoost.M1, a neural network, Naïve Bayes, a perceptron, random forest, support‑vector machine, and a baseline model—to classify RBD versus non‑RBD episodes. A four‑fold cross‑validation scheme was used, ensuring that each patient’s data appeared in only one fold at a time. The resulting performance metrics, averaged over the four folds, are summarized below: AdaBoost.M1 achieved a specificity of 0.711 and a sensitivity of 0.512, yielding an F1‑score of 0.177 and an area under the ROC curve (AUC) of 0.612. The neural network performed similarly with a specificity of 0.702, sensitivity 0.583, F1 of 0.190, and AUC 0.643. Naïve Bayes produced a specificity of 0.700, sensitivity 0.576, F1 0.216, and AUC 0.701. The perceptron and random forest models showed higher sensitivities (0.718 and 0.702, respectively) but lower specificities (0.572 and 0.470), resulting in F1‑scores around 0.19 and AUCs between 0.586 and 0.645. The support‑vector machine delivered the highest specificity (0.730) but a lower sensitivity (0.565), with an F1 of 0.189 and an AUC of 0.647. Across all algorithms, none achieved simultaneously high specificity and sensitivity; the F1‑scores remained below 0.22 and the AUC values hovered around 0.6–0.7. These modest results are attributed primarily to the limited size of the dataset and the severe class imbalance between RBD and non‑RBD episodes. The project explored various sampling strategies to mitigate this imbalance, yet the performance gains were marginal, underscoring the need for larger, more balanced data collections. Future work will involve acquiring additional recordings from diverse sources, refining the labeling process by having at least three independent experts annotate each episode, and using the consensus labels as the target variable for training more robust models. The predictive analytics engine, currently available as a Windows executable, has already facilitated the analysis of the study data and will require validation in a larger clinical trial before it can be deployed in routine practice. The accompanying web application, developed in the app and infrastructure sub‑project, is fully functional and has received positive feedback from patients, indicating that it can be further tailored to user needs.

Collaboration has been central to the project’s progress. The core partners include Fraunhofer Institute for Software and Systems Technology (ISST), the Institute for Occupational Safety (IAW), and the University Hospital RWTH Aachen (UKA). Fraunhofer ISST led the development of the predictive analytics software and the data‑processing pipeline, while IAW contributed expertise in data sampling and statistical validation. UKA supplied the clinical data, provided expert annotations, and offered clinical oversight to ensure that the system aligns with patient care requirements. The project was funded through a German research grant awarded to the consortium, with the funding body acting as the grant recipient. The collaboration spanned several years, during which the team coordinated data collection, algorithm development, and iterative refinement of the user interface. The joint effort has produced a prototype system that demonstrates the feasibility of automated RBD detection and establishes a foundation for future large‑scale validation studies aimed at improving early Parkinson’s disease diagnosis and patient management.