<![CDATA[Epidemiological surveillance in the Democratic Republic of the Congo (DRC) suffers from reporting delays and limited digital infrastructure, while online French-language news provides a complementary real-time signal that current systems exploit poorly. We design, deploy, and rigorously evaluate an end-to-end natural-language processing (NLP) pipeline that integrates targeted web scraping of Congolese online media, sentence-level binary classification of epidemic content with a fine-tuned CamemBERT transformer, zero-shot biomedical named-entity recognition (CamemBERT-bio-GLiNER) restricted to disease, location and date, and an alerting dashboard built on a Django/Celery stack. The classifier was fine-tuned on a hybrid corpus of 11,433 sentences combining 1,433 manually annotated real news sentences and 10,000 template-generated synthetic sentences, and is benchmarked against two classical baselines (TF-IDF combined with Logistic Regression and Linear SVM) on an independent, manually annotated test set of 997 sentences (341 epidemic, 656 non-epidemic) constructed from a second scraping campaign performed three months later. We report precision, recall, F1, PR-AUC and ROC-AUC with 1,000-iteration bootstrap 95% confidence intervals. CamemBERT reaches F1 = 0.754 [0.717-0.787] and PR-AUC = 0.699 [0.644-0.756] for the epidemic class, while the Linear SVM baseline reaches F1 = 0.858 ± 0.037 and PR-AUC = 0.926 ± 0.024 in 5-fold stratified cross-validation, outperforming the transformer, a result we attribute to the dominance of synthetic data in the training corpus. A single-batch operational run of the full pipeline on MediaCongo processed 30 articles and 501 sentences in 37.5 s on a single GPU, producing 43 alerts that correctly captured the May 2026 Ebola Bundibugyo outbreak in Ituri. The system, the external benchmark, and all evaluation scripts are released as open source.]]>
Copyrights © 2026
