Implementation Design and Validation Results of a Chronic Wound Management System Powered by Beyond 5G Networking

andreea.ancuta.corici, kalpana.chaudhary, marius-iulian.corici, benny.haeusler, hemant.zope, [email protected] martin.hocquel, simone.kuntz, anne.deter, nils.lahmann @charite.de Implementation Design and Validation Results of a Chronic Wound Management System Powered by Beyond 5G Networking Andreea Ancuta Corici Martin Hocquel-Hans Kalpana Chaudhary Simone Kuntz Marius Corici Benny Häusler Hemant Zope Anne Deter Anne Grohnert Nils Lahmann December 1, 2025 Abstract In the context of healthcare for chronic wound patients, especially those living in remote areas, alternative solutions for care at home are needed to avoid transportation and care for patients in hospitals. Based on the requirements of the clinical team, this paper proposes a solution that employs image processing to assist the caring personnel in objectively evaluating the wound as well as bio-printing of necessary gel-based patches for treating it. For reliable and secure connectivity, the solution leverages nomadic micro-networks based on technologies like 5G and beyond. The paper also dives into the key design aspects, considering aspects such as access control and interoperability. Regarding the image processing service, an evaluation of three promising wound surface detection algorithms using publicly available datasets is provided. The paper also includes lessons learned from the do-it-yourself bio-printer. Chronic Wounds Nomadic Network Telehealth 1 Introduction Although there is no common unique definition of chronic wounds, often called “hard-to-heal” wounds [1] are commonly defined as “wounds that have not proceeded through an orderly and timely reparation to produce anatomic and functional integrity after 6 weeks [2] to 3 months” [3]. Generally, wounds can be classified into different types: venous/vascular ulcers, diabetic ulcers, pressure ulcers, and ischemic wounds [4]. Wounds may seriously affect the quality of life in those affected [5] and can mean considerable costs for society [6]. A systematic review and meta-analysis of observational studies of the prevalence of chronic wounds in the general population, done by Martinegro and colleagues in 2019, found a pooled prevalence of 2.21 per 1000 population, and for chronic leg ulcers, the prevalence was estimated at 1.51 per 1000 population [7]. State-of-the-art therapy is to apply a well-fitting hydrogel dressing in chronic skin wounds in general [8] and especially to diabetic foot ulcers [9]. The solution proposed in this paper utilizes Non-Public Networking (NPN) to enable secure and reliable connectivity for nurses visiting the patient in their home environment by employing beyond 5G nomadic node connectivity and routing as well as supporting multiple backbone connectivity types (satellite and public mobile operator when available) (see Fig. 1). Thus, the nurse can collect and store highresolution 2D and 3D scans of the wounds, leading to high-quality wound documentation. Also, via the increased available bandwidth, compared to 2G/3G connectivity available in remote areas in Germany, 1 receiving medical advice via multimedia calls with the wound expert doctor working at the clinic can increase the potential for accurate assessment and decision-making in treating the wound at home. In recent years, non-invasive techniques for facilitating wound assessment, including artificial intelligence, have been progressively developed [10]. The work presented here also evaluates a set of algorithms for wound detection via image processing as a first step in implementing a solution for automatic wound surface computing. The research also stretches towards personalized patches based on hydrogels, which are currently used in treating chronic wounds and can be printed by making use of bioprinters. Figure 1: Hydrogel Patch Use Case Overview This paper focused on the ICT solution for such a use case. In Section 2, a brief requirements analysis is presented, leading to the description of the system’s distributed architecture in Section 3. Section 4 emphasizes the mobile core network features. Section 5 presents the medical services design and implementation aspects, Section 6 covers the interfacing design and lessons learned from integrating the bio-printer, and Section 7 describes the wound detection design and validation as a first step towards precise automatic surface computation. Conclusions and future work are provided in Section 8. 2 Requirements Analysis During workshops for requirements gathering and analysis, together with the medical team, a set of requirements was identified. Regarding the regulation for interfacing the hospital information system (HIS), it is required by the German digitalization law to support HL7 Fast Healthcare Interoperability Resources (FHIR) [11] as a data exchange protocol. FHIR is a protocol for specifying the data scheme for creating, updating, and deleting operations on FHIR resources like Patient, Encounter, and Questionnaire. The next phase of the interface is to be supported by the 1st of July 2025 and includes resources for vital parameters, medication, and general data exchange, with type codes for data like scans and videos defined by the Integrating Healthcare Enterprise (IHE) Germany [12, 13]. From the medical process point of view, the most important functional requirement is the employment of AI for wound analysis to compute the surface, volume, and depth of the wound using 2D and 3D wound segmentation without the employment of conventional methods like a ruler. For this, the demanded accuracy is at least 90 percent, and the image processing results are used for both documentation and comparative visualization and for generating a matching hydrogel patch. Furthermore, the wound expert should be able to edit the extracted wound model to create a better matching patch, as the wound has to be covered only 80 percent of the volume to allow for the best healing chance. 2 The nurse attending patients from remote areas expects to have a private network available on-site for connecting to the clinic where the wound expert resides. The following identified ICT features of the private area of key interest: • High availability by employing also satellite connection as a backhaul • High bandwidth and low latency for video consultations and high-resolution scanning data transfer • Minimal delay and secure communication between the devices (phone to bio-printer), which in the context of beyond 5G networks, derives as a local breakout. • Encrypted storage of the medical data and granular access control for different types of roles and dynamic consent-based association between nurses and patients As an overall requirement, the system has to be available as a prototype ready for trials in less than one and a half years and with a minimal hardware budget. 3 System Architecture Our proposed system is based on deploying a nomadic 5G network, customized to provide communication services to a specific medical use case with its own set of applications and services. The key aspect is that the use case is independent of the mobile network infrastructure at the patients’ location, providing additional security and reliability compared to using a shared commercial infrastructure, as well as providing extensive availability. As depicted in Figure 2, the network edge is traveling together with the nurse personnel, continuously providing a local 5G network at the required use case location. On the nurse side, the mobile phone application Nurse Care App allows for secure and interoperable wound documentation and visualization. An open-source mobile app allows for Over-the-Top (OTT) multimedia calls with the wound expert at the clinic. An external application is used to build the 3D model of the affected area to be analyzed using Gaussian splatting [14]. Figure 2: System Functional Architecture 3 In the cloud infrastructure of the medical headquarters, the medical data information is stored in an FHIR Store that uses an interface to an Access Control Server to inquire about the role attributed to a certain user or application. The Wound Analytics processes the 2D and 3D data transferred by the Nurse Care App, produces measurement results of the detected wound area, and extracts the 3D wound model that is used to print the hydrogel patch. The results of the Wound Analytics are automatically stored in the FHIR Store, and the wound expert can visualize them inside the Admin and Care Portal. After a close bio-printer market analysis, the team decided to employ a cost-effective do-it-yourself prototype of a bio-printer that receives geometric code (G-code) commands (currently part of the ISO 6983-1 standard [15]) from the 3D Bio-printing Service, described in section 6, and running on a Raspberry Pi. The printer is then connected to the private network via an RJ-45 to a Customer Premises Equipment (CPE) for fast prototyping compared to interfacing to a 5G modem. 4 Mobile Core Network Features From a 5G network perspective, the essential aspects include the deployment of a nomadic base station, which can provide connectivity at the use case location, as well as edge core network functionality facilitating the authentication, authorization, mobility, and end-to-end sessions between devices connected at the edge and the central infrastructure [19]. The edge and the central core network are interconnected over a best-effort access network available at the momentary location of the nurse’s vehicle, which may be a public 4G or 5G network when in coverage range or a satellite network [20]. This backhaul connectivity is used Over-The-Top (OTT), using a separate authentication and authorization, able to carry both the control and the data plane of the end-to-end network across an end-to-end encrypted data path. For being able to maintain reliable communication across this backhaul, a Software-Defined Wide Area Network (SD-WAN), specifically modified for nomadic networks, is used. This enables the dynamic establishment of backhaul connections to the central location according to the momentarily available transport networks, as well as the traffic steering across these transport networks to assure appropriate end-to-end delivery. For this, the edge base station, which is located in the nurse’s vehicle and provides local connectivity, is connected to an edge core network infrastructure that takes over part of the previously centralized core network. For being able to establish local data paths, the edge core network includes a User Plane Function (UPF), which offloads the data traffic needed for the communication between the local applications and devices, relieving the transport network from carrying it to the central entities and back [22]. Additionally, the edge core network may include functionality for local mobility and local session management, in case the nurse’s vehicle includes more than one base station. However, as in most cases, this network is relatively small; it is better to maintain most of the control plane in the central location to facilitate a simpler service with less distributed functionality [21]. This enables easier network management as well as extensive privacy, as the network authentication and authorization information does not have to be shared with the edge node and thus can not be tampered with. 5 Medical Services For application development, the Flutter programming language was selected for compiling the applications for mobile and web environments as well. For 3D model rendering and user interaction, the opensource library three.js [16] was integrated into both the mobile-running Nurse Care App and the Admin and Care Portal. To secure access to the FHIR resources, a Web security mechanism was implemented using an interceptor in the FHIR Server. The interceptor checks that the requests contain a valid authorization token and applies access control to the FHIR resources using role-based policies defined for Patient, Nurse, Doctor, 4 and Study coordinator roles. The authorization token is to be acquired via OpenID-Connect protocol [17] from the Access Control Server running an instance of the Keycloak toolkit [18] (see Fig. 3). Also, important to mention is that users, e.g., patients, are first provisioned via the Admin and Care Portal by interfacing an Account Management Service API bound to the Access Control Server and if successful, also provisioned on the Patient Data Storage Server 6 Bio Printing Interfacing The development of an affordable bio-printing solution required adapting commercially available 3D printing technology to handle gel-based materials. For this purpose, an Ender 3 V2 off-the-shelf 3D printer was converted to a bioprinter by replacing the standard Bowden extruder with a custom syringe extruder. The modification followed the ”Enderstruder” design from [24], providing a cost-effective alternative to commercial bioprinters while maintaining sufficient precision for wound patch creation. Several challenges were encountered during the hardware modification process: • The custom 3D-printed components initially manufactured using Polylactic Acid (PLA) filament exhibited insufficient structural integrity and failed during assembly. Subsequent iterations using Acrylonitrile Butadiene Styrene (ABS) material provided the necessary stability for the syringe mounting system. • Firmware modifications presented additional technical hurdles. The display controller firmware required customization, as initial flash attempts failed with the firmware from the Enderstruder repository. These issues resulted from different Printed Circuit Board (PCB) revisions and could be fixed by adding a missing configuration file and renaming the firmware to include the directory of the SD card. Figure 3: Enabling Access Control for Resources in FHIR Store 5 6.1 Communication Adaptor OctoPrint [25] was selected as the primary control software platform for several compelling reasons. Its open-source nature, extensive printer compatibility, minimal resource requirements, and ability to run on a Raspberry Pi 4 made it ideal for our nomadic network application. The platform’s large community support facilitated troubleshooting during development. While alternative platforms such as UltiMaker Cura [26] and Pronterface [27] were evaluated during development, OctoPrint’s remote control capabilities through web interfaces aligned better with our requirements for nurse-operated systems in home care environments. It offers a RESTful API to monitor and manage printer status and jobs and therefore qualifies perfectly for integration into the wound management system. The interfacing between the Bio-printing Service and the Octoprint is via REST API, and both components run on the Raspberry Pi. The printer itself is connected via USB to the Raspberry Pi and establishes a serial connection to receive gCode. 6.2 Validation While the hardware and software infrastructure have been established, the system is currently in the calibration phase. Test prints with actual hydrogels have not yet been performed, as precise calibration is essential to ensure the structural integrity and dimensional accuracy of the printed patches. The next phase will involve systematic testing of different cannula sizes as well as hydrogel compositions and the evaluation of printability, structural stability post-printing, and adherence characteristics to determine optimal parameters for wound patch production. 7 Wound Detection This section presents an evaluation of the performance and comparative analysis of publicly accessible machine learning algorithms for wound segmentation using a publicly available dataset and standard evaluation metrics. 7.1 Evaluation Metrics The performance of segmentation algorithms is evaluated using established evaluation metrics, including Dice coefficient, Precision, Recall, IoU, Segmentation time, F1 score [28], Photographic Wound Assessment Tool (PWAT)[29]. An analysis of the results is then conducted to determine the most effective approach. Dice Coefficient: Measures how well the predictions overlap with the actual wound. It is similar to IOU but gives more weight to well-matched regions. Precision: It is indicative of the accuracy of segmentation. More specifically, it is calculated as the percentage of correctly segmented pixels in the segmentation. Recall: Segmentation accuracy is also shown by recall. It’s measured as the percentage of correctly segmented pixels in the ground truth. Intersection over Union (IoU): Measures the overlap between the predicted segmentation and the ground truth mask. Calculated as the area of overlap divided by the area of union. F1 score: The harmonic mean of precision and recall. Segmentation Time: The time taken to process the image and produce a segmentation. Photographic Wound Assessment Tool: PWAT is a scoring system used by clinicians to evaluate wound healing based on images. 6 7.2 Considered Algorithms According to studies on wound segmentation, deep learning techniques utilizing Convolutional Neural Networks (CNN) are highly favored by researchers in the field. The wound detection algorithms considered were Deepskin [30], FCN [10], and WS[31]. These algorithms were then used to make predictions on the same dataset, referred to as the ’wound seg dataset’ [33] and Foot Ulcer Segmentation Challenge(FUSC). [32]. Deepskin is an open-source Python package specifically designed for the analysis of wound images acquired via a smartphone camera. The package proposes a fully automated pipeline for wound image processing that uses the deep learning UNet model, which is trained on a large Deepskin dataset. Additionally, the PWAT score is automatically estimated, and a Peri wound segmentation mask is produced by applying morphological operators to the wound mask. The next study is entitled FCN (Fully Conventional Neural Network) and was based on the MobileVNet2 model architecture, with training conducted on the FUSC dataset. Various architectures were tested, including U-Net, MobileNetV2, Mask-RCNN, SegNet, and VGG16, but it was found that MobileVNet2 yielded the best results. The model used in their work needed more than 1000 training epochs, drastically adding to the computational time of the model training. The third algorithm was developed based on the UNet model architecture. The algorithm was obtained from Kaggle and retrained on the Data-wound seg dataset. 7.2.1 Data processing Data processing steps were applied before and after model training to ensure consistency. For this, all three algorithms used data augmentation to increase the variety of training data by applying different transformations to the existing data to properly manage the images and desired transformations. FCN also used normalization to ensure consistent input to the models, improving convergence and performance, and some post-processing techniques such as noise removal to improve clarity and accuracy, hole filling to fill in gaps or missing parts, ensuring continuity, and bilinear up-sampling to increase the resolution of images by interpolating pixel values in a bilinear manner, resulting in smoother images. The WS algorithm used data generator pre-processing techniques, which generate batches of data during training and incorporate data augmentation and normalization. 7.3 Results and Interpretations The FCN model was re-trained on the FUSC dataset for 1000 epochs, and the WS model was re-trained on the wound seg dataset for 100 epochs with early stopping to achieve reproducibility of the results. Then, the trained models were tested on the wound seg dataset to monitor the generalization of these models and their performance on different images. Results of prediction on the wound seg dataset are shown in Figure 4. Table 1: Comparison of results achieved by Deepskin, FCN, and WS on the woundseg dataset. Model Deepskin FCN WS Dice 0.62 0.87 0.81 Precision 0.74 0.90 0.84 Recall 0.62 0.84 0.87 IoU 0.53 0.78 0.71 F1 0.62 0.87 0.52 Segmentation Time 2.3 s 0.63 s 0.36 s 7 Figure 4: The results of the algorithms considered: (a) the segmentation results of Deepskin (b) the results of FCN (c) the results of WS In the performance measures, the FCN model’s precision was rated highest among other models at 0.90, indicating that the predicted segmentation contains the highest percentage of true positives. FCN also achieved the highest recall, meaning the model successfully captures most positive instances and minimizes the number of false negatives, and the highest Dice, meaning the model performs well in identifying the relevant areas. On the other hand, the WS model proved to be the second highest with less segmentation time. While the evaluation metrics of Deepskin on the wound seg dataset performed average results when compared, it was claimed that the prediction results on Deepskin dataset proved to be the best when compared to other datasets. Overall, the results show that the FCN model achieves the highest accuracy and requires the longest segmentation time. WS provides a good balance between accuracy and segmentation time, making it a strong candidate for the segmentation process. 8 Conclusion and Future Work This paper proposed a solution for chronic wound care in remote areas, with reliable connectivity and secure data transfer, as well as wound detection, as key enablers. The testbed is currently extended to8 wards automatic measurements collection by employing the Experiment Life-Cycle Management platform (ELCM [34]), with one instance residing in the laboratory for enabling further features testing and validation and another instance for first validation at the medical research team premises with wound models and then transferred to a nursing home where chronic wound patients reside. On the socio-economic side, a focus group session with nursing stakeholders has validated the expectations that AI and ICT features will increase the technological adoption as well as the attractiveness of wound care nursing in the countryside. Trials and pilots with wound experts and patients are planned for the next phase of the project for validation of the scanning, fast documentation storage, assistance via automatic computation, and comparative visualization, as well as an optimized network [23]. 8.0.1 Acknowledgements The presented activities are part of the 6G-PATH project, which received funding from the Smart Networks and Services Joint Undertaking (SNS JU) under the European Union’s Horizon Europe research and innovation program under Grant Agreement No 101139172. References [1] Cazander, G., Pritchard, D. I., Nigam, Y., Jung, W., and Nibbering, P. H. (2013). Multiple actions of Lucilia sericata larvae in hard-to-heal wounds: larval secretions contain molecules that accelerate wound healing, reduce chronic inflammation and inhibit bacterial infection. Bioessays, 35(12), 10831092. 10.1002/bies.201300071 [2] da Silva, L. P., Reis, R. L., Correlo, V. M., and Marques, A. P. (2019). Hydrogel-Based Strategies to Advance Therapies for Chronic Skin Wounds. Annu Rev Biomed Eng, 21, 145-169. 10.1146/annurevbioeng-060418-052422 [3] Dubhashi, S. P., and Sindwani, R. D. (2015). A Comparative Study of Honey and Phenytoin Dressings for Chronic Wounds. Indian J Surg, 77(Suppl 3), 1209-1213. 10.1007/s12262-015-1251-6 [4] Frykberg, R. G., and Banks, J. (2015). Challenges in the Treatment of Chronic Wounds. Adv Wound Care (New Rochelle), 4(9), 560-582. 10.1089/wound.2015.0635 [5] Irfan-Maqsood, M. (2018). Classification of wounds: know before research and clinical practice. Journal of Genes and Cells, 4(1), 1-4. [6] Martinengo, L., Olsson, M., Bajpai, R., Soljak, M., Upton, Z., Schmidtchen, A., Car, J., and Jarbrink, K. (2019). Prevalence of chronic wounds in the general population: systematic review and meta-analysis of observational studies. Ann Epidemiol, 29, 8-15. https://doi.org/10.1016/j.annepidem.2018.10.005 [7] Natarajan, J., Joseph, M. A., Al Alawi, R., Al Bulushi, T., Al Alawi, I., Al Junaibi, S. M., Thanka, A. N., Al Balushi, L. D., Al Ismaili, I. S., Shummo, M., and Al Nabhani, S. S. T. (2024). A domain specific health-related quality of life of omani patients living with chronic wounds. J Tissue Viability, 33(3), 393-398. https://doi.org/10.1016/j.jtv.2024.05.004 [8] Ruiz, P. B. O., and Lima, A. F. C. (2022). Average direct costs of outpatient, hospital, and home care provided to patients with chronic wounds. Rev Esc Enferm USP, 56, e20220295. https://- doi.org/10.1590/1980-220X-REEUSP-2022-0295en 9