An explainable CNN–BiLSTM–Random Forest framework for epidemic warning-signal and disease-status classification

Abstract

This study evaluates an explainable ensemble in which Convolutional Neural Network and Bidirectional Long Short-Term Memory feature extractors feed a Random Forest classifier, applied under one architectural template to two independent tasks: labelling Ebola-related tweets as symptom-bearing warning signals and labelling structured coronavirus disease 2019 (COVID-19) patient records as positive or negative for severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2). The tasks use disjoint data on different diseases, are never fused, and model no forecasting horizon. Results come from five-fold stratified cross-validation repeated twice, with preprocessing and decision thresholds fitted inside each training partition and metrics reported per class. On the social media task, the framework attains a positive-class sensitivity of 0.918, precision of 0.975, and F1-score of 0.946, exceeding a term-frequency baseline by 0.024 in F1-score. Two ablations qualify this: raising the tokenizer vocabulary from 100 to 5,000 terms and training the extractors end-to-end account for the entire gain; the recurrent branch and the sentiment feature add nothing thereafter. On the clinical task, no configuration improves on a Random Forest over the laboratory variables alone, but the operating point matters more than the model: reporting at 90% specificity rather than at a probability cut of 0.5 raises sensitivity from 0.310 to 0.711, and to 0.803 at 80% specificity. Discrimination falls to 0.66 in patients without a complete blood count, so the instrument is a post-haematology triage aid rather than a screening tool. Shapley Additive Explanations rank leukocytes and platelets highest but cover only 14% of the model's attribution mass.