The clinic should not stop working because the internet did.
Close to one billion people worldwide are served by health facilities that either lack electricity entirely or have an unreliable supply. In those settings, digital health tools that depend on cloud connectivity are not digital health tools at all. They are promises that fail on the day they matter most.
This contribution props this problem as an engineering question: what would it take to run an AI-powered screening service inside a clinic that has no reliable internet, intermittent power, and no on-site IT staff? Can such system exist where a nurse could use during a routine visit, with results written into a patient record that persists locally and synchronizes to a hospital when a connection eventually becomes available?
Everything runs on one device.
Collaborators form the University of Puerto Rico at Mayagüez built an offline-first edge telemedicine platform that consolidates a local electronic health record, clinical services, and modular on-device screening microservices on a single NVIDIA Jetson Orin Nano. A clinician connects to it over local Wi-Fi through a web app. Patient registration, record retrieval, image capture, AI screening, and result storage all happen locally. When connectivity returns, queued records synchronize to external systems through an HL7 FHIR gateway. When it does not, the clinic keeps working.
The platform treats screening services as plug-in modules. Each one follows the same contract: acquire a sample, run on-device inference, and write a structured result back into the patient record. The anemia module is the first, but the architecture is designed so that additional services (dermatology, wound monitoring, other point-of-care tools) can be added without changing the core EHR or the user interface.
Anemia screening from a photograph of a fingernail.
To stress-test the platform with a real clinical workload, we implemented non-invasive anemia screening from fingernail photographs. An INT8-quantized YOLOv8n detector extracts nail regions from the image in about 22 milliseconds. Lightweight color features are computed from each region and fed into a supervised regressor that estimates hemoglobin concentration in about 34 milliseconds. The entire pipeline, from image capture to a stored result in the patient record, runs fully on-device.
We evaluated six different training strategies across eight classical regression models to understand how data balancing and augmentation choices affect not just overall accuracy but the specific failure mode that matters most in screening: missed anemia cases. The best numeric performance came from severity-balanced SVR (RMSE of 2.048 g/dL) and remark-balanced Gradient Boosting (RMSE of 2.091 g/dL). But the more important finding was in the error distributions.
The metric that matters is not the one on the leaderboard.
When we looked at prediction errors restricted to truly anemic patients, a clear pattern emerged. Models trained on the raw unbalanced data consistently overestimated hemoglobin in anemic individuals, pushing their predictions above the screening threshold and generating false negatives. Balanced training strategies shifted those errors back toward zero and reduced the worst-case overestimates. Augmentation alone did not fix the problem and in some cases made it worse.
This is the kind of insight that aggregate metrics like RMSE can hide. A model with a lower overall error can still be more dangerous in a screening context if its errors are concentrated in the direction that causes missed cases. Choosing a training strategy is not a statistical convenience. It is a policy decision about which failure mode you are willing to accept.
Industry-Academic Partnerships in Computational Sciences & Engineering
At Capicú, we work on making AI systems run reliably on the hardware that is actually available in the places where they are needed. We are incredibly grateful for Dr. Wilfredo Lugo Beauchamp and the Edge Computing Groups efforts towards building a makerspace for next-gen leaders in computing sciences and engineering. If you work on deploying AI in environments where infrastructure is the constraint rather than model performance, we would like to hear from you.
S.